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) { }
369 #endif /* CONFIG_GROUP_SCHED */
371 /* CFS-related fields in a runqueue */
373 struct load_weight load;
374 unsigned long nr_running;
379 struct rb_root tasks_timeline;
380 struct rb_node *rb_leftmost;
382 struct list_head tasks;
383 struct list_head *balance_iterator;
386 * 'curr' points to currently running entity on this cfs_rq.
387 * It is set to NULL otherwise (i.e when none are currently running).
389 struct sched_entity *curr, *next;
391 unsigned long nr_spread_over;
393 #ifdef CONFIG_FAIR_GROUP_SCHED
394 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
397 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
398 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
399 * (like users, containers etc.)
401 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
402 * list is used during load balance.
404 struct list_head leaf_cfs_rq_list;
405 struct task_group *tg; /* group that "owns" this runqueue */
409 /* Real-Time classes' related field in a runqueue: */
411 struct rt_prio_array active;
412 unsigned long rt_nr_running;
413 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
414 int highest_prio; /* highest queued rt task prio */
417 unsigned long rt_nr_migratory;
423 /* Nests inside the rq lock: */
424 spinlock_t rt_runtime_lock;
426 #ifdef CONFIG_RT_GROUP_SCHED
427 unsigned long rt_nr_boosted;
430 struct list_head leaf_rt_rq_list;
431 struct task_group *tg;
432 struct sched_rt_entity *rt_se;
439 * We add the notion of a root-domain which will be used to define per-domain
440 * variables. Each exclusive cpuset essentially defines an island domain by
441 * fully partitioning the member cpus from any other cpuset. Whenever a new
442 * exclusive cpuset is created, we also create and attach a new root-domain
452 * The "RT overload" flag: it gets set if a CPU has more than
453 * one runnable RT task.
458 struct cpupri cpupri;
463 * By default the system creates a single root-domain with all cpus as
464 * members (mimicking the global state we have today).
466 static struct root_domain def_root_domain;
471 * This is the main, per-CPU runqueue data structure.
473 * Locking rule: those places that want to lock multiple runqueues
474 * (such as the load balancing or the thread migration code), lock
475 * acquire operations must be ordered by ascending &runqueue.
482 * nr_running and cpu_load should be in the same cacheline because
483 * remote CPUs use both these fields when doing load calculation.
485 unsigned long nr_running;
486 #define CPU_LOAD_IDX_MAX 5
487 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
488 unsigned char idle_at_tick;
490 unsigned long last_tick_seen;
491 unsigned char in_nohz_recently;
493 /* capture load from *all* tasks on this cpu: */
494 struct load_weight load;
495 unsigned long nr_load_updates;
501 #ifdef CONFIG_FAIR_GROUP_SCHED
502 /* list of leaf cfs_rq on this cpu: */
503 struct list_head leaf_cfs_rq_list;
505 #ifdef CONFIG_RT_GROUP_SCHED
506 struct list_head leaf_rt_rq_list;
510 * This is part of a global counter where only the total sum
511 * over all CPUs matters. A task can increase this counter on
512 * one CPU and if it got migrated afterwards it may decrease
513 * it on another CPU. Always updated under the runqueue lock:
515 unsigned long nr_uninterruptible;
517 struct task_struct *curr, *idle;
518 unsigned long next_balance;
519 struct mm_struct *prev_mm;
526 struct root_domain *rd;
527 struct sched_domain *sd;
529 /* For active balancing */
532 /* cpu of this runqueue: */
536 struct task_struct *migration_thread;
537 struct list_head migration_queue;
540 #ifdef CONFIG_SCHED_HRTICK
541 unsigned long hrtick_flags;
542 ktime_t hrtick_expire;
543 struct hrtimer hrtick_timer;
546 #ifdef CONFIG_SCHEDSTATS
548 struct sched_info rq_sched_info;
550 /* sys_sched_yield() stats */
551 unsigned int yld_exp_empty;
552 unsigned int yld_act_empty;
553 unsigned int yld_both_empty;
554 unsigned int yld_count;
556 /* schedule() stats */
557 unsigned int sched_switch;
558 unsigned int sched_count;
559 unsigned int sched_goidle;
561 /* try_to_wake_up() stats */
562 unsigned int ttwu_count;
563 unsigned int ttwu_local;
566 unsigned int bkl_count;
568 struct lock_class_key rq_lock_key;
571 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
573 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
575 rq->curr->sched_class->check_preempt_curr(rq, p);
578 static inline int cpu_of(struct rq *rq)
588 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
589 * See detach_destroy_domains: synchronize_sched for details.
591 * The domain tree of any CPU may only be accessed from within
592 * preempt-disabled sections.
594 #define for_each_domain(cpu, __sd) \
595 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
597 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
598 #define this_rq() (&__get_cpu_var(runqueues))
599 #define task_rq(p) cpu_rq(task_cpu(p))
600 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
602 static inline void update_rq_clock(struct rq *rq)
604 rq->clock = sched_clock_cpu(cpu_of(rq));
608 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
610 #ifdef CONFIG_SCHED_DEBUG
611 # define const_debug __read_mostly
613 # define const_debug static const
617 * Debugging: various feature bits
620 #define SCHED_FEAT(name, enabled) \
621 __SCHED_FEAT_##name ,
624 #include "sched_features.h"
629 #define SCHED_FEAT(name, enabled) \
630 (1UL << __SCHED_FEAT_##name) * enabled |
632 const_debug unsigned int sysctl_sched_features =
633 #include "sched_features.h"
638 #ifdef CONFIG_SCHED_DEBUG
639 #define SCHED_FEAT(name, enabled) \
642 static __read_mostly char *sched_feat_names[] = {
643 #include "sched_features.h"
649 static int sched_feat_open(struct inode *inode, struct file *filp)
651 filp->private_data = inode->i_private;
656 sched_feat_read(struct file *filp, char __user *ubuf,
657 size_t cnt, loff_t *ppos)
664 for (i = 0; sched_feat_names[i]; i++) {
665 len += strlen(sched_feat_names[i]);
669 buf = kmalloc(len + 2, GFP_KERNEL);
673 for (i = 0; sched_feat_names[i]; i++) {
674 if (sysctl_sched_features & (1UL << i))
675 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
677 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
680 r += sprintf(buf + r, "\n");
681 WARN_ON(r >= len + 2);
683 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
691 sched_feat_write(struct file *filp, const char __user *ubuf,
692 size_t cnt, loff_t *ppos)
702 if (copy_from_user(&buf, ubuf, cnt))
707 if (strncmp(buf, "NO_", 3) == 0) {
712 for (i = 0; sched_feat_names[i]; i++) {
713 int len = strlen(sched_feat_names[i]);
715 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
717 sysctl_sched_features &= ~(1UL << i);
719 sysctl_sched_features |= (1UL << i);
724 if (!sched_feat_names[i])
732 static struct file_operations sched_feat_fops = {
733 .open = sched_feat_open,
734 .read = sched_feat_read,
735 .write = sched_feat_write,
738 static __init int sched_init_debug(void)
740 debugfs_create_file("sched_features", 0644, NULL, NULL,
745 late_initcall(sched_init_debug);
749 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
752 * Number of tasks to iterate in a single balance run.
753 * Limited because this is done with IRQs disabled.
755 const_debug unsigned int sysctl_sched_nr_migrate = 32;
758 * period over which we measure -rt task cpu usage in us.
761 unsigned int sysctl_sched_rt_period = 1000000;
763 static __read_mostly int scheduler_running;
766 * part of the period that we allow rt tasks to run in us.
769 int sysctl_sched_rt_runtime = 950000;
771 static inline u64 global_rt_period(void)
773 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
776 static inline u64 global_rt_runtime(void)
778 if (sysctl_sched_rt_period < 0)
781 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
784 unsigned long long time_sync_thresh = 100000;
786 static DEFINE_PER_CPU(unsigned long long, time_offset);
787 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
790 * Global lock which we take every now and then to synchronize
791 * the CPUs time. This method is not warp-safe, but it's good
792 * enough to synchronize slowly diverging time sources and thus
793 * it's good enough for tracing:
795 static DEFINE_SPINLOCK(time_sync_lock);
796 static unsigned long long prev_global_time;
798 static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
801 * We want this inlined, to not get tracer function calls
802 * in this critical section:
804 spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
805 __raw_spin_lock(&time_sync_lock.raw_lock);
807 if (time < prev_global_time) {
808 per_cpu(time_offset, cpu) += prev_global_time - time;
809 time = prev_global_time;
811 prev_global_time = time;
814 __raw_spin_unlock(&time_sync_lock.raw_lock);
815 spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
820 static unsigned long long __cpu_clock(int cpu)
822 unsigned long long now;
825 * Only call sched_clock() if the scheduler has already been
826 * initialized (some code might call cpu_clock() very early):
828 if (unlikely(!scheduler_running))
831 now = sched_clock_cpu(cpu);
837 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
838 * clock constructed from sched_clock():
840 unsigned long long cpu_clock(int cpu)
842 unsigned long long prev_cpu_time, time, delta_time;
845 local_irq_save(flags);
846 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
847 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
848 delta_time = time-prev_cpu_time;
850 if (unlikely(delta_time > time_sync_thresh)) {
851 time = __sync_cpu_clock(time, cpu);
852 per_cpu(prev_cpu_time, cpu) = time;
854 local_irq_restore(flags);
858 EXPORT_SYMBOL_GPL(cpu_clock);
860 #ifndef prepare_arch_switch
861 # define prepare_arch_switch(next) do { } while (0)
863 #ifndef finish_arch_switch
864 # define finish_arch_switch(prev) do { } while (0)
867 static inline int task_current(struct rq *rq, struct task_struct *p)
869 return rq->curr == p;
872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
873 static inline int task_running(struct rq *rq, struct task_struct *p)
875 return task_current(rq, p);
878 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
882 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
884 #ifdef CONFIG_DEBUG_SPINLOCK
885 /* this is a valid case when another task releases the spinlock */
886 rq->lock.owner = current;
889 * If we are tracking spinlock dependencies then we have to
890 * fix up the runqueue lock - which gets 'carried over' from
893 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
895 spin_unlock_irq(&rq->lock);
898 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
899 static inline int task_running(struct rq *rq, struct task_struct *p)
904 return task_current(rq, p);
908 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
918 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 spin_unlock_irq(&rq->lock);
921 spin_unlock(&rq->lock);
925 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
929 * After ->oncpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
936 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
943 * __task_rq_lock - lock the runqueue a given task resides on.
944 * Must be called interrupts disabled.
946 static inline struct rq *__task_rq_lock(struct task_struct *p)
950 struct rq *rq = task_rq(p);
951 spin_lock(&rq->lock);
952 if (likely(rq == task_rq(p)))
954 spin_unlock(&rq->lock);
959 * task_rq_lock - lock the runqueue a given task resides on and disable
960 * interrupts. Note the ordering: we can safely lookup the task_rq without
961 * explicitly disabling preemption.
963 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
969 local_irq_save(*flags);
971 spin_lock(&rq->lock);
972 if (likely(rq == task_rq(p)))
974 spin_unlock_irqrestore(&rq->lock, *flags);
978 static void __task_rq_unlock(struct rq *rq)
981 spin_unlock(&rq->lock);
984 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
987 spin_unlock_irqrestore(&rq->lock, *flags);
991 * this_rq_lock - lock this runqueue and disable interrupts.
993 static struct rq *this_rq_lock(void)
1000 spin_lock(&rq->lock);
1005 static void __resched_task(struct task_struct *p, int tif_bit);
1007 static inline void resched_task(struct task_struct *p)
1009 __resched_task(p, TIF_NEED_RESCHED);
1012 #ifdef CONFIG_SCHED_HRTICK
1014 * Use HR-timers to deliver accurate preemption points.
1016 * Its all a bit involved since we cannot program an hrt while holding the
1017 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1020 * When we get rescheduled we reprogram the hrtick_timer outside of the
1023 static inline void resched_hrt(struct task_struct *p)
1025 __resched_task(p, TIF_HRTICK_RESCHED);
1028 static inline void resched_rq(struct rq *rq)
1030 unsigned long flags;
1032 spin_lock_irqsave(&rq->lock, flags);
1033 resched_task(rq->curr);
1034 spin_unlock_irqrestore(&rq->lock, flags);
1038 HRTICK_SET, /* re-programm hrtick_timer */
1039 HRTICK_RESET, /* not a new slice */
1040 HRTICK_BLOCK, /* stop hrtick operations */
1045 * - enabled by features
1046 * - hrtimer is actually high res
1048 static inline int hrtick_enabled(struct rq *rq)
1050 if (!sched_feat(HRTICK))
1052 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1054 return hrtimer_is_hres_active(&rq->hrtick_timer);
1058 * Called to set the hrtick timer state.
1060 * called with rq->lock held and irqs disabled
1062 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1064 assert_spin_locked(&rq->lock);
1067 * preempt at: now + delay
1070 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1072 * indicate we need to program the timer
1074 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1076 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1079 * New slices are called from the schedule path and don't need a
1080 * forced reschedule.
1083 resched_hrt(rq->curr);
1086 static void hrtick_clear(struct rq *rq)
1088 if (hrtimer_active(&rq->hrtick_timer))
1089 hrtimer_cancel(&rq->hrtick_timer);
1093 * Update the timer from the possible pending state.
1095 static void hrtick_set(struct rq *rq)
1099 unsigned long flags;
1101 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1103 spin_lock_irqsave(&rq->lock, flags);
1104 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1105 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1106 time = rq->hrtick_expire;
1107 clear_thread_flag(TIF_HRTICK_RESCHED);
1108 spin_unlock_irqrestore(&rq->lock, flags);
1111 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1112 if (reset && !hrtimer_active(&rq->hrtick_timer))
1119 * High-resolution timer tick.
1120 * Runs from hardirq context with interrupts disabled.
1122 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1124 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1126 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1128 spin_lock(&rq->lock);
1129 update_rq_clock(rq);
1130 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1131 spin_unlock(&rq->lock);
1133 return HRTIMER_NORESTART;
1137 static void hotplug_hrtick_disable(int cpu)
1139 struct rq *rq = cpu_rq(cpu);
1140 unsigned long flags;
1142 spin_lock_irqsave(&rq->lock, flags);
1143 rq->hrtick_flags = 0;
1144 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1145 spin_unlock_irqrestore(&rq->lock, flags);
1150 static void hotplug_hrtick_enable(int cpu)
1152 struct rq *rq = cpu_rq(cpu);
1153 unsigned long flags;
1155 spin_lock_irqsave(&rq->lock, flags);
1156 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1157 spin_unlock_irqrestore(&rq->lock, flags);
1161 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1163 int cpu = (int)(long)hcpu;
1166 case CPU_UP_CANCELED:
1167 case CPU_UP_CANCELED_FROZEN:
1168 case CPU_DOWN_PREPARE:
1169 case CPU_DOWN_PREPARE_FROZEN:
1171 case CPU_DEAD_FROZEN:
1172 hotplug_hrtick_disable(cpu);
1175 case CPU_UP_PREPARE:
1176 case CPU_UP_PREPARE_FROZEN:
1177 case CPU_DOWN_FAILED:
1178 case CPU_DOWN_FAILED_FROZEN:
1180 case CPU_ONLINE_FROZEN:
1181 hotplug_hrtick_enable(cpu);
1188 static void init_hrtick(void)
1190 hotcpu_notifier(hotplug_hrtick, 0);
1192 #endif /* CONFIG_SMP */
1194 static void init_rq_hrtick(struct rq *rq)
1196 rq->hrtick_flags = 0;
1197 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1198 rq->hrtick_timer.function = hrtick;
1199 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1202 void hrtick_resched(void)
1205 unsigned long flags;
1207 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1210 local_irq_save(flags);
1211 rq = cpu_rq(smp_processor_id());
1213 local_irq_restore(flags);
1216 static inline void hrtick_clear(struct rq *rq)
1220 static inline void hrtick_set(struct rq *rq)
1224 static inline void init_rq_hrtick(struct rq *rq)
1228 void hrtick_resched(void)
1232 static inline void init_hrtick(void)
1238 * resched_task - mark a task 'to be rescheduled now'.
1240 * On UP this means the setting of the need_resched flag, on SMP it
1241 * might also involve a cross-CPU call to trigger the scheduler on
1246 #ifndef tsk_is_polling
1247 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1250 static void __resched_task(struct task_struct *p, int tif_bit)
1254 assert_spin_locked(&task_rq(p)->lock);
1256 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1259 set_tsk_thread_flag(p, tif_bit);
1262 if (cpu == smp_processor_id())
1265 /* NEED_RESCHED must be visible before we test polling */
1267 if (!tsk_is_polling(p))
1268 smp_send_reschedule(cpu);
1271 static void resched_cpu(int cpu)
1273 struct rq *rq = cpu_rq(cpu);
1274 unsigned long flags;
1276 if (!spin_trylock_irqsave(&rq->lock, flags))
1278 resched_task(cpu_curr(cpu));
1279 spin_unlock_irqrestore(&rq->lock, flags);
1284 * When add_timer_on() enqueues a timer into the timer wheel of an
1285 * idle CPU then this timer might expire before the next timer event
1286 * which is scheduled to wake up that CPU. In case of a completely
1287 * idle system the next event might even be infinite time into the
1288 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1289 * leaves the inner idle loop so the newly added timer is taken into
1290 * account when the CPU goes back to idle and evaluates the timer
1291 * wheel for the next timer event.
1293 void wake_up_idle_cpu(int cpu)
1295 struct rq *rq = cpu_rq(cpu);
1297 if (cpu == smp_processor_id())
1301 * This is safe, as this function is called with the timer
1302 * wheel base lock of (cpu) held. When the CPU is on the way
1303 * to idle and has not yet set rq->curr to idle then it will
1304 * be serialized on the timer wheel base lock and take the new
1305 * timer into account automatically.
1307 if (rq->curr != rq->idle)
1311 * We can set TIF_RESCHED on the idle task of the other CPU
1312 * lockless. The worst case is that the other CPU runs the
1313 * idle task through an additional NOOP schedule()
1315 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1317 /* NEED_RESCHED must be visible before we test polling */
1319 if (!tsk_is_polling(rq->idle))
1320 smp_send_reschedule(cpu);
1322 #endif /* CONFIG_NO_HZ */
1324 #else /* !CONFIG_SMP */
1325 static void __resched_task(struct task_struct *p, int tif_bit)
1327 assert_spin_locked(&task_rq(p)->lock);
1328 set_tsk_thread_flag(p, tif_bit);
1330 #endif /* CONFIG_SMP */
1332 #if BITS_PER_LONG == 32
1333 # define WMULT_CONST (~0UL)
1335 # define WMULT_CONST (1UL << 32)
1338 #define WMULT_SHIFT 32
1341 * Shift right and round:
1343 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1346 * delta *= weight / lw
1348 static unsigned long
1349 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1350 struct load_weight *lw)
1354 if (!lw->inv_weight) {
1355 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1358 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1362 tmp = (u64)delta_exec * weight;
1364 * Check whether we'd overflow the 64-bit multiplication:
1366 if (unlikely(tmp > WMULT_CONST))
1367 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1370 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1372 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1375 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1381 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1388 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1389 * of tasks with abnormal "nice" values across CPUs the contribution that
1390 * each task makes to its run queue's load is weighted according to its
1391 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1392 * scaled version of the new time slice allocation that they receive on time
1396 #define WEIGHT_IDLEPRIO 2
1397 #define WMULT_IDLEPRIO (1 << 31)
1400 * Nice levels are multiplicative, with a gentle 10% change for every
1401 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1402 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1403 * that remained on nice 0.
1405 * The "10% effect" is relative and cumulative: from _any_ nice level,
1406 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1407 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1408 * If a task goes up by ~10% and another task goes down by ~10% then
1409 * the relative distance between them is ~25%.)
1411 static const int prio_to_weight[40] = {
1412 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1413 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1414 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1415 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1416 /* 0 */ 1024, 820, 655, 526, 423,
1417 /* 5 */ 335, 272, 215, 172, 137,
1418 /* 10 */ 110, 87, 70, 56, 45,
1419 /* 15 */ 36, 29, 23, 18, 15,
1423 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1425 * In cases where the weight does not change often, we can use the
1426 * precalculated inverse to speed up arithmetics by turning divisions
1427 * into multiplications:
1429 static const u32 prio_to_wmult[40] = {
1430 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1431 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1432 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1433 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1434 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1435 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1436 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1437 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1440 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1443 * runqueue iterator, to support SMP load-balancing between different
1444 * scheduling classes, without having to expose their internal data
1445 * structures to the load-balancing proper:
1447 struct rq_iterator {
1449 struct task_struct *(*start)(void *);
1450 struct task_struct *(*next)(void *);
1454 static unsigned long
1455 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1456 unsigned long max_load_move, struct sched_domain *sd,
1457 enum cpu_idle_type idle, int *all_pinned,
1458 int *this_best_prio, struct rq_iterator *iterator);
1461 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1462 struct sched_domain *sd, enum cpu_idle_type idle,
1463 struct rq_iterator *iterator);
1466 #ifdef CONFIG_CGROUP_CPUACCT
1467 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1469 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1472 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1474 update_load_add(&rq->load, load);
1477 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1479 update_load_sub(&rq->load, load);
1483 static unsigned long source_load(int cpu, int type);
1484 static unsigned long target_load(int cpu, int type);
1485 static unsigned long cpu_avg_load_per_task(int cpu);
1486 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1489 #include "sched_stats.h"
1490 #include "sched_idletask.c"
1491 #include "sched_fair.c"
1492 #include "sched_rt.c"
1493 #ifdef CONFIG_SCHED_DEBUG
1494 # include "sched_debug.c"
1497 #define sched_class_highest (&rt_sched_class)
1498 #define for_each_class(class) \
1499 for (class = sched_class_highest; class; class = class->next)
1501 static inline void inc_load(struct rq *rq, const struct task_struct *p)
1503 update_load_add(&rq->load, p->se.load.weight);
1506 static inline void dec_load(struct rq *rq, const struct task_struct *p)
1508 update_load_sub(&rq->load, p->se.load.weight);
1511 static void inc_nr_running(struct task_struct *p, struct rq *rq)
1517 static void dec_nr_running(struct task_struct *p, struct rq *rq)
1523 static void set_load_weight(struct task_struct *p)
1525 if (task_has_rt_policy(p)) {
1526 p->se.load.weight = prio_to_weight[0] * 2;
1527 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1532 * SCHED_IDLE tasks get minimal weight:
1534 if (p->policy == SCHED_IDLE) {
1535 p->se.load.weight = WEIGHT_IDLEPRIO;
1536 p->se.load.inv_weight = WMULT_IDLEPRIO;
1540 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1541 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1544 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1546 sched_info_queued(p);
1547 p->sched_class->enqueue_task(rq, p, wakeup);
1551 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1553 p->sched_class->dequeue_task(rq, p, sleep);
1558 * __normal_prio - return the priority that is based on the static prio
1560 static inline int __normal_prio(struct task_struct *p)
1562 return p->static_prio;
1566 * Calculate the expected normal priority: i.e. priority
1567 * without taking RT-inheritance into account. Might be
1568 * boosted by interactivity modifiers. Changes upon fork,
1569 * setprio syscalls, and whenever the interactivity
1570 * estimator recalculates.
1572 static inline int normal_prio(struct task_struct *p)
1576 if (task_has_rt_policy(p))
1577 prio = MAX_RT_PRIO-1 - p->rt_priority;
1579 prio = __normal_prio(p);
1584 * Calculate the current priority, i.e. the priority
1585 * taken into account by the scheduler. This value might
1586 * be boosted by RT tasks, or might be boosted by
1587 * interactivity modifiers. Will be RT if the task got
1588 * RT-boosted. If not then it returns p->normal_prio.
1590 static int effective_prio(struct task_struct *p)
1592 p->normal_prio = normal_prio(p);
1594 * If we are RT tasks or we were boosted to RT priority,
1595 * keep the priority unchanged. Otherwise, update priority
1596 * to the normal priority:
1598 if (!rt_prio(p->prio))
1599 return p->normal_prio;
1604 * activate_task - move a task to the runqueue.
1606 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1608 if (task_contributes_to_load(p))
1609 rq->nr_uninterruptible--;
1611 enqueue_task(rq, p, wakeup);
1612 inc_nr_running(p, rq);
1616 * deactivate_task - remove a task from the runqueue.
1618 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1620 if (task_contributes_to_load(p))
1621 rq->nr_uninterruptible++;
1623 dequeue_task(rq, p, sleep);
1624 dec_nr_running(p, rq);
1628 * task_curr - is this task currently executing on a CPU?
1629 * @p: the task in question.
1631 inline int task_curr(const struct task_struct *p)
1633 return cpu_curr(task_cpu(p)) == p;
1636 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1638 set_task_rq(p, cpu);
1641 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1642 * successfuly executed on another CPU. We must ensure that updates of
1643 * per-task data have been completed by this moment.
1646 task_thread_info(p)->cpu = cpu;
1650 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1651 const struct sched_class *prev_class,
1652 int oldprio, int running)
1654 if (prev_class != p->sched_class) {
1655 if (prev_class->switched_from)
1656 prev_class->switched_from(rq, p, running);
1657 p->sched_class->switched_to(rq, p, running);
1659 p->sched_class->prio_changed(rq, p, oldprio, running);
1664 /* Used instead of source_load when we know the type == 0 */
1665 static unsigned long weighted_cpuload(const int cpu)
1667 return cpu_rq(cpu)->load.weight;
1671 * Is this task likely cache-hot:
1674 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1679 * Buddy candidates are cache hot:
1681 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1684 if (p->sched_class != &fair_sched_class)
1687 if (sysctl_sched_migration_cost == -1)
1689 if (sysctl_sched_migration_cost == 0)
1692 delta = now - p->se.exec_start;
1694 return delta < (s64)sysctl_sched_migration_cost;
1698 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1700 int old_cpu = task_cpu(p);
1701 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1702 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1703 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1706 clock_offset = old_rq->clock - new_rq->clock;
1708 #ifdef CONFIG_SCHEDSTATS
1709 if (p->se.wait_start)
1710 p->se.wait_start -= clock_offset;
1711 if (p->se.sleep_start)
1712 p->se.sleep_start -= clock_offset;
1713 if (p->se.block_start)
1714 p->se.block_start -= clock_offset;
1715 if (old_cpu != new_cpu) {
1716 schedstat_inc(p, se.nr_migrations);
1717 if (task_hot(p, old_rq->clock, NULL))
1718 schedstat_inc(p, se.nr_forced2_migrations);
1721 p->se.vruntime -= old_cfsrq->min_vruntime -
1722 new_cfsrq->min_vruntime;
1724 __set_task_cpu(p, new_cpu);
1727 struct migration_req {
1728 struct list_head list;
1730 struct task_struct *task;
1733 struct completion done;
1737 * The task's runqueue lock must be held.
1738 * Returns true if you have to wait for migration thread.
1741 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1743 struct rq *rq = task_rq(p);
1746 * If the task is not on a runqueue (and not running), then
1747 * it is sufficient to simply update the task's cpu field.
1749 if (!p->se.on_rq && !task_running(rq, p)) {
1750 set_task_cpu(p, dest_cpu);
1754 init_completion(&req->done);
1756 req->dest_cpu = dest_cpu;
1757 list_add(&req->list, &rq->migration_queue);
1763 * wait_task_inactive - wait for a thread to unschedule.
1765 * The caller must ensure that the task *will* unschedule sometime soon,
1766 * else this function might spin for a *long* time. This function can't
1767 * be called with interrupts off, or it may introduce deadlock with
1768 * smp_call_function() if an IPI is sent by the same process we are
1769 * waiting to become inactive.
1771 void wait_task_inactive(struct task_struct *p)
1773 unsigned long flags;
1779 * We do the initial early heuristics without holding
1780 * any task-queue locks at all. We'll only try to get
1781 * the runqueue lock when things look like they will
1787 * If the task is actively running on another CPU
1788 * still, just relax and busy-wait without holding
1791 * NOTE! Since we don't hold any locks, it's not
1792 * even sure that "rq" stays as the right runqueue!
1793 * But we don't care, since "task_running()" will
1794 * return false if the runqueue has changed and p
1795 * is actually now running somewhere else!
1797 while (task_running(rq, p))
1801 * Ok, time to look more closely! We need the rq
1802 * lock now, to be *sure*. If we're wrong, we'll
1803 * just go back and repeat.
1805 rq = task_rq_lock(p, &flags);
1806 running = task_running(rq, p);
1807 on_rq = p->se.on_rq;
1808 task_rq_unlock(rq, &flags);
1811 * Was it really running after all now that we
1812 * checked with the proper locks actually held?
1814 * Oops. Go back and try again..
1816 if (unlikely(running)) {
1822 * It's not enough that it's not actively running,
1823 * it must be off the runqueue _entirely_, and not
1826 * So if it wa still runnable (but just not actively
1827 * running right now), it's preempted, and we should
1828 * yield - it could be a while.
1830 if (unlikely(on_rq)) {
1831 schedule_timeout_uninterruptible(1);
1836 * Ahh, all good. It wasn't running, and it wasn't
1837 * runnable, which means that it will never become
1838 * running in the future either. We're all done!
1845 * kick_process - kick a running thread to enter/exit the kernel
1846 * @p: the to-be-kicked thread
1848 * Cause a process which is running on another CPU to enter
1849 * kernel-mode, without any delay. (to get signals handled.)
1851 * NOTE: this function doesnt have to take the runqueue lock,
1852 * because all it wants to ensure is that the remote task enters
1853 * the kernel. If the IPI races and the task has been migrated
1854 * to another CPU then no harm is done and the purpose has been
1857 void kick_process(struct task_struct *p)
1863 if ((cpu != smp_processor_id()) && task_curr(p))
1864 smp_send_reschedule(cpu);
1869 * Return a low guess at the load of a migration-source cpu weighted
1870 * according to the scheduling class and "nice" value.
1872 * We want to under-estimate the load of migration sources, to
1873 * balance conservatively.
1875 static unsigned long source_load(int cpu, int type)
1877 struct rq *rq = cpu_rq(cpu);
1878 unsigned long total = weighted_cpuload(cpu);
1883 return min(rq->cpu_load[type-1], total);
1887 * Return a high guess at the load of a migration-target cpu weighted
1888 * according to the scheduling class and "nice" value.
1890 static unsigned long target_load(int cpu, int type)
1892 struct rq *rq = cpu_rq(cpu);
1893 unsigned long total = weighted_cpuload(cpu);
1898 return max(rq->cpu_load[type-1], total);
1902 * Return the average load per task on the cpu's run queue
1904 static unsigned long cpu_avg_load_per_task(int cpu)
1906 struct rq *rq = cpu_rq(cpu);
1907 unsigned long total = weighted_cpuload(cpu);
1908 unsigned long n = rq->nr_running;
1910 return n ? total / n : SCHED_LOAD_SCALE;
1914 * find_idlest_group finds and returns the least busy CPU group within the
1917 static struct sched_group *
1918 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
1920 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
1921 unsigned long min_load = ULONG_MAX, this_load = 0;
1922 int load_idx = sd->forkexec_idx;
1923 int imbalance = 100 + (sd->imbalance_pct-100)/2;
1926 unsigned long load, avg_load;
1930 /* Skip over this group if it has no CPUs allowed */
1931 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
1934 local_group = cpu_isset(this_cpu, group->cpumask);
1936 /* Tally up the load of all CPUs in the group */
1939 for_each_cpu_mask(i, group->cpumask) {
1940 /* Bias balancing toward cpus of our domain */
1942 load = source_load(i, load_idx);
1944 load = target_load(i, load_idx);
1949 /* Adjust by relative CPU power of the group */
1950 avg_load = sg_div_cpu_power(group,
1951 avg_load * SCHED_LOAD_SCALE);
1954 this_load = avg_load;
1956 } else if (avg_load < min_load) {
1957 min_load = avg_load;
1960 } while (group = group->next, group != sd->groups);
1962 if (!idlest || 100*this_load < imbalance*min_load)
1968 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1971 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
1974 unsigned long load, min_load = ULONG_MAX;
1978 /* Traverse only the allowed CPUs */
1979 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
1981 for_each_cpu_mask(i, *tmp) {
1982 load = weighted_cpuload(i);
1984 if (load < min_load || (load == min_load && i == this_cpu)) {
1994 * sched_balance_self: balance the current task (running on cpu) in domains
1995 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1998 * Balance, ie. select the least loaded group.
2000 * Returns the target CPU number, or the same CPU if no balancing is needed.
2002 * preempt must be disabled.
2004 static int sched_balance_self(int cpu, int flag)
2006 struct task_struct *t = current;
2007 struct sched_domain *tmp, *sd = NULL;
2009 for_each_domain(cpu, tmp) {
2011 * If power savings logic is enabled for a domain, stop there.
2013 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2015 if (tmp->flags & flag)
2020 cpumask_t span, tmpmask;
2021 struct sched_group *group;
2022 int new_cpu, weight;
2024 if (!(sd->flags & flag)) {
2030 group = find_idlest_group(sd, t, cpu);
2036 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2037 if (new_cpu == -1 || new_cpu == cpu) {
2038 /* Now try balancing at a lower domain level of cpu */
2043 /* Now try balancing at a lower domain level of new_cpu */
2046 weight = cpus_weight(span);
2047 for_each_domain(cpu, tmp) {
2048 if (weight <= cpus_weight(tmp->span))
2050 if (tmp->flags & flag)
2053 /* while loop will break here if sd == NULL */
2059 #endif /* CONFIG_SMP */
2062 * try_to_wake_up - wake up a thread
2063 * @p: the to-be-woken-up thread
2064 * @state: the mask of task states that can be woken
2065 * @sync: do a synchronous wakeup?
2067 * Put it on the run-queue if it's not already there. The "current"
2068 * thread is always on the run-queue (except when the actual
2069 * re-schedule is in progress), and as such you're allowed to do
2070 * the simpler "current->state = TASK_RUNNING" to mark yourself
2071 * runnable without the overhead of this.
2073 * returns failure only if the task is already active.
2075 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2077 int cpu, orig_cpu, this_cpu, success = 0;
2078 unsigned long flags;
2082 if (!sched_feat(SYNC_WAKEUPS))
2086 rq = task_rq_lock(p, &flags);
2087 old_state = p->state;
2088 if (!(old_state & state))
2096 this_cpu = smp_processor_id();
2099 if (unlikely(task_running(rq, p)))
2102 cpu = p->sched_class->select_task_rq(p, sync);
2103 if (cpu != orig_cpu) {
2104 set_task_cpu(p, cpu);
2105 task_rq_unlock(rq, &flags);
2106 /* might preempt at this point */
2107 rq = task_rq_lock(p, &flags);
2108 old_state = p->state;
2109 if (!(old_state & state))
2114 this_cpu = smp_processor_id();
2118 #ifdef CONFIG_SCHEDSTATS
2119 schedstat_inc(rq, ttwu_count);
2120 if (cpu == this_cpu)
2121 schedstat_inc(rq, ttwu_local);
2123 struct sched_domain *sd;
2124 for_each_domain(this_cpu, sd) {
2125 if (cpu_isset(cpu, sd->span)) {
2126 schedstat_inc(sd, ttwu_wake_remote);
2131 #endif /* CONFIG_SCHEDSTATS */
2134 #endif /* CONFIG_SMP */
2135 schedstat_inc(p, se.nr_wakeups);
2137 schedstat_inc(p, se.nr_wakeups_sync);
2138 if (orig_cpu != cpu)
2139 schedstat_inc(p, se.nr_wakeups_migrate);
2140 if (cpu == this_cpu)
2141 schedstat_inc(p, se.nr_wakeups_local);
2143 schedstat_inc(p, se.nr_wakeups_remote);
2144 update_rq_clock(rq);
2145 activate_task(rq, p, 1);
2149 check_preempt_curr(rq, p);
2151 p->state = TASK_RUNNING;
2153 if (p->sched_class->task_wake_up)
2154 p->sched_class->task_wake_up(rq, p);
2157 task_rq_unlock(rq, &flags);
2162 int wake_up_process(struct task_struct *p)
2164 return try_to_wake_up(p, TASK_ALL, 0);
2166 EXPORT_SYMBOL(wake_up_process);
2168 int wake_up_state(struct task_struct *p, unsigned int state)
2170 return try_to_wake_up(p, state, 0);
2174 * Perform scheduler related setup for a newly forked process p.
2175 * p is forked by current.
2177 * __sched_fork() is basic setup used by init_idle() too:
2179 static void __sched_fork(struct task_struct *p)
2181 p->se.exec_start = 0;
2182 p->se.sum_exec_runtime = 0;
2183 p->se.prev_sum_exec_runtime = 0;
2184 p->se.last_wakeup = 0;
2185 p->se.avg_overlap = 0;
2187 #ifdef CONFIG_SCHEDSTATS
2188 p->se.wait_start = 0;
2189 p->se.sum_sleep_runtime = 0;
2190 p->se.sleep_start = 0;
2191 p->se.block_start = 0;
2192 p->se.sleep_max = 0;
2193 p->se.block_max = 0;
2195 p->se.slice_max = 0;
2199 INIT_LIST_HEAD(&p->rt.run_list);
2201 INIT_LIST_HEAD(&p->se.group_node);
2203 #ifdef CONFIG_PREEMPT_NOTIFIERS
2204 INIT_HLIST_HEAD(&p->preempt_notifiers);
2208 * We mark the process as running here, but have not actually
2209 * inserted it onto the runqueue yet. This guarantees that
2210 * nobody will actually run it, and a signal or other external
2211 * event cannot wake it up and insert it on the runqueue either.
2213 p->state = TASK_RUNNING;
2217 * fork()/clone()-time setup:
2219 void sched_fork(struct task_struct *p, int clone_flags)
2221 int cpu = get_cpu();
2226 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2228 set_task_cpu(p, cpu);
2231 * Make sure we do not leak PI boosting priority to the child:
2233 p->prio = current->normal_prio;
2234 if (!rt_prio(p->prio))
2235 p->sched_class = &fair_sched_class;
2237 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2238 if (likely(sched_info_on()))
2239 memset(&p->sched_info, 0, sizeof(p->sched_info));
2241 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2244 #ifdef CONFIG_PREEMPT
2245 /* Want to start with kernel preemption disabled. */
2246 task_thread_info(p)->preempt_count = 1;
2252 * wake_up_new_task - wake up a newly created task for the first time.
2254 * This function will do some initial scheduler statistics housekeeping
2255 * that must be done for every newly created context, then puts the task
2256 * on the runqueue and wakes it.
2258 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2260 unsigned long flags;
2263 rq = task_rq_lock(p, &flags);
2264 BUG_ON(p->state != TASK_RUNNING);
2265 update_rq_clock(rq);
2267 p->prio = effective_prio(p);
2269 if (!p->sched_class->task_new || !current->se.on_rq) {
2270 activate_task(rq, p, 0);
2273 * Let the scheduling class do new task startup
2274 * management (if any):
2276 p->sched_class->task_new(rq, p);
2277 inc_nr_running(p, rq);
2279 check_preempt_curr(rq, p);
2281 if (p->sched_class->task_wake_up)
2282 p->sched_class->task_wake_up(rq, p);
2284 task_rq_unlock(rq, &flags);
2287 #ifdef CONFIG_PREEMPT_NOTIFIERS
2290 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2291 * @notifier: notifier struct to register
2293 void preempt_notifier_register(struct preempt_notifier *notifier)
2295 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2297 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2300 * preempt_notifier_unregister - no longer interested in preemption notifications
2301 * @notifier: notifier struct to unregister
2303 * This is safe to call from within a preemption notifier.
2305 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2307 hlist_del(¬ifier->link);
2309 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2311 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2313 struct preempt_notifier *notifier;
2314 struct hlist_node *node;
2316 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2317 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2321 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2322 struct task_struct *next)
2324 struct preempt_notifier *notifier;
2325 struct hlist_node *node;
2327 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2328 notifier->ops->sched_out(notifier, next);
2331 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2333 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2338 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2339 struct task_struct *next)
2343 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2346 * prepare_task_switch - prepare to switch tasks
2347 * @rq: the runqueue preparing to switch
2348 * @prev: the current task that is being switched out
2349 * @next: the task we are going to switch to.
2351 * This is called with the rq lock held and interrupts off. It must
2352 * be paired with a subsequent finish_task_switch after the context
2355 * prepare_task_switch sets up locking and calls architecture specific
2359 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2360 struct task_struct *next)
2362 fire_sched_out_preempt_notifiers(prev, next);
2363 prepare_lock_switch(rq, next);
2364 prepare_arch_switch(next);
2368 * finish_task_switch - clean up after a task-switch
2369 * @rq: runqueue associated with task-switch
2370 * @prev: the thread we just switched away from.
2372 * finish_task_switch must be called after the context switch, paired
2373 * with a prepare_task_switch call before the context switch.
2374 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2375 * and do any other architecture-specific cleanup actions.
2377 * Note that we may have delayed dropping an mm in context_switch(). If
2378 * so, we finish that here outside of the runqueue lock. (Doing it
2379 * with the lock held can cause deadlocks; see schedule() for
2382 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2383 __releases(rq->lock)
2385 struct mm_struct *mm = rq->prev_mm;
2391 * A task struct has one reference for the use as "current".
2392 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2393 * schedule one last time. The schedule call will never return, and
2394 * the scheduled task must drop that reference.
2395 * The test for TASK_DEAD must occur while the runqueue locks are
2396 * still held, otherwise prev could be scheduled on another cpu, die
2397 * there before we look at prev->state, and then the reference would
2399 * Manfred Spraul <manfred@colorfullife.com>
2401 prev_state = prev->state;
2402 finish_arch_switch(prev);
2403 finish_lock_switch(rq, prev);
2405 if (current->sched_class->post_schedule)
2406 current->sched_class->post_schedule(rq);
2409 fire_sched_in_preempt_notifiers(current);
2412 if (unlikely(prev_state == TASK_DEAD)) {
2414 * Remove function-return probe instances associated with this
2415 * task and put them back on the free list.
2417 kprobe_flush_task(prev);
2418 put_task_struct(prev);
2423 * schedule_tail - first thing a freshly forked thread must call.
2424 * @prev: the thread we just switched away from.
2426 asmlinkage void schedule_tail(struct task_struct *prev)
2427 __releases(rq->lock)
2429 struct rq *rq = this_rq();
2431 finish_task_switch(rq, prev);
2432 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2433 /* In this case, finish_task_switch does not reenable preemption */
2436 if (current->set_child_tid)
2437 put_user(task_pid_vnr(current), current->set_child_tid);
2441 * context_switch - switch to the new MM and the new
2442 * thread's register state.
2445 context_switch(struct rq *rq, struct task_struct *prev,
2446 struct task_struct *next)
2448 struct mm_struct *mm, *oldmm;
2450 prepare_task_switch(rq, prev, next);
2452 oldmm = prev->active_mm;
2454 * For paravirt, this is coupled with an exit in switch_to to
2455 * combine the page table reload and the switch backend into
2458 arch_enter_lazy_cpu_mode();
2460 if (unlikely(!mm)) {
2461 next->active_mm = oldmm;
2462 atomic_inc(&oldmm->mm_count);
2463 enter_lazy_tlb(oldmm, next);
2465 switch_mm(oldmm, mm, next);
2467 if (unlikely(!prev->mm)) {
2468 prev->active_mm = NULL;
2469 rq->prev_mm = oldmm;
2472 * Since the runqueue lock will be released by the next
2473 * task (which is an invalid locking op but in the case
2474 * of the scheduler it's an obvious special-case), so we
2475 * do an early lockdep release here:
2477 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2478 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2481 /* Here we just switch the register state and the stack. */
2482 switch_to(prev, next, prev);
2486 * this_rq must be evaluated again because prev may have moved
2487 * CPUs since it called schedule(), thus the 'rq' on its stack
2488 * frame will be invalid.
2490 finish_task_switch(this_rq(), prev);
2494 * nr_running, nr_uninterruptible and nr_context_switches:
2496 * externally visible scheduler statistics: current number of runnable
2497 * threads, current number of uninterruptible-sleeping threads, total
2498 * number of context switches performed since bootup.
2500 unsigned long nr_running(void)
2502 unsigned long i, sum = 0;
2504 for_each_online_cpu(i)
2505 sum += cpu_rq(i)->nr_running;
2510 unsigned long nr_uninterruptible(void)
2512 unsigned long i, sum = 0;
2514 for_each_possible_cpu(i)
2515 sum += cpu_rq(i)->nr_uninterruptible;
2518 * Since we read the counters lockless, it might be slightly
2519 * inaccurate. Do not allow it to go below zero though:
2521 if (unlikely((long)sum < 0))
2527 unsigned long long nr_context_switches(void)
2530 unsigned long long sum = 0;
2532 for_each_possible_cpu(i)
2533 sum += cpu_rq(i)->nr_switches;
2538 unsigned long nr_iowait(void)
2540 unsigned long i, sum = 0;
2542 for_each_possible_cpu(i)
2543 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2548 unsigned long nr_active(void)
2550 unsigned long i, running = 0, uninterruptible = 0;
2552 for_each_online_cpu(i) {
2553 running += cpu_rq(i)->nr_running;
2554 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2557 if (unlikely((long)uninterruptible < 0))
2558 uninterruptible = 0;
2560 return running + uninterruptible;
2564 * Update rq->cpu_load[] statistics. This function is usually called every
2565 * scheduler tick (TICK_NSEC).
2567 static void update_cpu_load(struct rq *this_rq)
2569 unsigned long this_load = this_rq->load.weight;
2572 this_rq->nr_load_updates++;
2574 /* Update our load: */
2575 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2576 unsigned long old_load, new_load;
2578 /* scale is effectively 1 << i now, and >> i divides by scale */
2580 old_load = this_rq->cpu_load[i];
2581 new_load = this_load;
2583 * Round up the averaging division if load is increasing. This
2584 * prevents us from getting stuck on 9 if the load is 10, for
2587 if (new_load > old_load)
2588 new_load += scale-1;
2589 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2596 * double_rq_lock - safely lock two runqueues
2598 * Note this does not disable interrupts like task_rq_lock,
2599 * you need to do so manually before calling.
2601 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2602 __acquires(rq1->lock)
2603 __acquires(rq2->lock)
2605 BUG_ON(!irqs_disabled());
2607 spin_lock(&rq1->lock);
2608 __acquire(rq2->lock); /* Fake it out ;) */
2611 spin_lock(&rq1->lock);
2612 spin_lock(&rq2->lock);
2614 spin_lock(&rq2->lock);
2615 spin_lock(&rq1->lock);
2618 update_rq_clock(rq1);
2619 update_rq_clock(rq2);
2623 * double_rq_unlock - safely unlock two runqueues
2625 * Note this does not restore interrupts like task_rq_unlock,
2626 * you need to do so manually after calling.
2628 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2629 __releases(rq1->lock)
2630 __releases(rq2->lock)
2632 spin_unlock(&rq1->lock);
2634 spin_unlock(&rq2->lock);
2636 __release(rq2->lock);
2640 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2642 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2643 __releases(this_rq->lock)
2644 __acquires(busiest->lock)
2645 __acquires(this_rq->lock)
2649 if (unlikely(!irqs_disabled())) {
2650 /* printk() doesn't work good under rq->lock */
2651 spin_unlock(&this_rq->lock);
2654 if (unlikely(!spin_trylock(&busiest->lock))) {
2655 if (busiest < this_rq) {
2656 spin_unlock(&this_rq->lock);
2657 spin_lock(&busiest->lock);
2658 spin_lock(&this_rq->lock);
2661 spin_lock(&busiest->lock);
2667 * If dest_cpu is allowed for this process, migrate the task to it.
2668 * This is accomplished by forcing the cpu_allowed mask to only
2669 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2670 * the cpu_allowed mask is restored.
2672 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2674 struct migration_req req;
2675 unsigned long flags;
2678 rq = task_rq_lock(p, &flags);
2679 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2680 || unlikely(cpu_is_offline(dest_cpu)))
2683 /* force the process onto the specified CPU */
2684 if (migrate_task(p, dest_cpu, &req)) {
2685 /* Need to wait for migration thread (might exit: take ref). */
2686 struct task_struct *mt = rq->migration_thread;
2688 get_task_struct(mt);
2689 task_rq_unlock(rq, &flags);
2690 wake_up_process(mt);
2691 put_task_struct(mt);
2692 wait_for_completion(&req.done);
2697 task_rq_unlock(rq, &flags);
2701 * sched_exec - execve() is a valuable balancing opportunity, because at
2702 * this point the task has the smallest effective memory and cache footprint.
2704 void sched_exec(void)
2706 int new_cpu, this_cpu = get_cpu();
2707 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2709 if (new_cpu != this_cpu)
2710 sched_migrate_task(current, new_cpu);
2714 * pull_task - move a task from a remote runqueue to the local runqueue.
2715 * Both runqueues must be locked.
2717 static void pull_task(struct rq *src_rq, struct task_struct *p,
2718 struct rq *this_rq, int this_cpu)
2720 deactivate_task(src_rq, p, 0);
2721 set_task_cpu(p, this_cpu);
2722 activate_task(this_rq, p, 0);
2724 * Note that idle threads have a prio of MAX_PRIO, for this test
2725 * to be always true for them.
2727 check_preempt_curr(this_rq, p);
2731 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2734 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2735 struct sched_domain *sd, enum cpu_idle_type idle,
2739 * We do not migrate tasks that are:
2740 * 1) running (obviously), or
2741 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2742 * 3) are cache-hot on their current CPU.
2744 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2745 schedstat_inc(p, se.nr_failed_migrations_affine);
2750 if (task_running(rq, p)) {
2751 schedstat_inc(p, se.nr_failed_migrations_running);
2756 * Aggressive migration if:
2757 * 1) task is cache cold, or
2758 * 2) too many balance attempts have failed.
2761 if (!task_hot(p, rq->clock, sd) ||
2762 sd->nr_balance_failed > sd->cache_nice_tries) {
2763 #ifdef CONFIG_SCHEDSTATS
2764 if (task_hot(p, rq->clock, sd)) {
2765 schedstat_inc(sd, lb_hot_gained[idle]);
2766 schedstat_inc(p, se.nr_forced_migrations);
2772 if (task_hot(p, rq->clock, sd)) {
2773 schedstat_inc(p, se.nr_failed_migrations_hot);
2779 static unsigned long
2780 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2781 unsigned long max_load_move, struct sched_domain *sd,
2782 enum cpu_idle_type idle, int *all_pinned,
2783 int *this_best_prio, struct rq_iterator *iterator)
2785 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2786 struct task_struct *p;
2787 long rem_load_move = max_load_move;
2789 if (max_load_move == 0)
2795 * Start the load-balancing iterator:
2797 p = iterator->start(iterator->arg);
2799 if (!p || loops++ > sysctl_sched_nr_migrate)
2802 * To help distribute high priority tasks across CPUs we don't
2803 * skip a task if it will be the highest priority task (i.e. smallest
2804 * prio value) on its new queue regardless of its load weight
2806 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2807 SCHED_LOAD_SCALE_FUZZ;
2808 if ((skip_for_load && p->prio >= *this_best_prio) ||
2809 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2810 p = iterator->next(iterator->arg);
2814 pull_task(busiest, p, this_rq, this_cpu);
2816 rem_load_move -= p->se.load.weight;
2819 * We only want to steal up to the prescribed amount of weighted load.
2821 if (rem_load_move > 0) {
2822 if (p->prio < *this_best_prio)
2823 *this_best_prio = p->prio;
2824 p = iterator->next(iterator->arg);
2829 * Right now, this is one of only two places pull_task() is called,
2830 * so we can safely collect pull_task() stats here rather than
2831 * inside pull_task().
2833 schedstat_add(sd, lb_gained[idle], pulled);
2836 *all_pinned = pinned;
2838 return max_load_move - rem_load_move;
2842 * move_tasks tries to move up to max_load_move weighted load from busiest to
2843 * this_rq, as part of a balancing operation within domain "sd".
2844 * Returns 1 if successful and 0 otherwise.
2846 * Called with both runqueues locked.
2848 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2849 unsigned long max_load_move,
2850 struct sched_domain *sd, enum cpu_idle_type idle,
2853 const struct sched_class *class = sched_class_highest;
2854 unsigned long total_load_moved = 0;
2855 int this_best_prio = this_rq->curr->prio;
2859 class->load_balance(this_rq, this_cpu, busiest,
2860 max_load_move - total_load_moved,
2861 sd, idle, all_pinned, &this_best_prio);
2862 class = class->next;
2863 } while (class && max_load_move > total_load_moved);
2865 return total_load_moved > 0;
2869 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2870 struct sched_domain *sd, enum cpu_idle_type idle,
2871 struct rq_iterator *iterator)
2873 struct task_struct *p = iterator->start(iterator->arg);
2877 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2878 pull_task(busiest, p, this_rq, this_cpu);
2880 * Right now, this is only the second place pull_task()
2881 * is called, so we can safely collect pull_task()
2882 * stats here rather than inside pull_task().
2884 schedstat_inc(sd, lb_gained[idle]);
2888 p = iterator->next(iterator->arg);
2895 * move_one_task tries to move exactly one task from busiest to this_rq, as
2896 * part of active balancing operations within "domain".
2897 * Returns 1 if successful and 0 otherwise.
2899 * Called with both runqueues locked.
2901 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2902 struct sched_domain *sd, enum cpu_idle_type idle)
2904 const struct sched_class *class;
2906 for (class = sched_class_highest; class; class = class->next)
2907 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
2914 * find_busiest_group finds and returns the busiest CPU group within the
2915 * domain. It calculates and returns the amount of weighted load which
2916 * should be moved to restore balance via the imbalance parameter.
2918 static struct sched_group *
2919 find_busiest_group(struct sched_domain *sd, int this_cpu,
2920 unsigned long *imbalance, enum cpu_idle_type idle,
2921 int *sd_idle, const cpumask_t *cpus, int *balance)
2923 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
2924 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
2925 unsigned long max_pull;
2926 unsigned long busiest_load_per_task, busiest_nr_running;
2927 unsigned long this_load_per_task, this_nr_running;
2928 int load_idx, group_imb = 0;
2929 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2930 int power_savings_balance = 1;
2931 unsigned long leader_nr_running = 0, min_load_per_task = 0;
2932 unsigned long min_nr_running = ULONG_MAX;
2933 struct sched_group *group_min = NULL, *group_leader = NULL;
2936 max_load = this_load = total_load = total_pwr = 0;
2937 busiest_load_per_task = busiest_nr_running = 0;
2938 this_load_per_task = this_nr_running = 0;
2939 if (idle == CPU_NOT_IDLE)
2940 load_idx = sd->busy_idx;
2941 else if (idle == CPU_NEWLY_IDLE)
2942 load_idx = sd->newidle_idx;
2944 load_idx = sd->idle_idx;
2947 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
2950 int __group_imb = 0;
2951 unsigned int balance_cpu = -1, first_idle_cpu = 0;
2952 unsigned long sum_nr_running, sum_weighted_load;
2954 local_group = cpu_isset(this_cpu, group->cpumask);
2957 balance_cpu = first_cpu(group->cpumask);
2959 /* Tally up the load of all CPUs in the group */
2960 sum_weighted_load = sum_nr_running = avg_load = 0;
2962 min_cpu_load = ~0UL;
2964 for_each_cpu_mask(i, group->cpumask) {
2967 if (!cpu_isset(i, *cpus))
2972 if (*sd_idle && rq->nr_running)
2975 /* Bias balancing toward cpus of our domain */
2977 if (idle_cpu(i) && !first_idle_cpu) {
2982 load = target_load(i, load_idx);
2984 load = source_load(i, load_idx);
2985 if (load > max_cpu_load)
2986 max_cpu_load = load;
2987 if (min_cpu_load > load)
2988 min_cpu_load = load;
2992 sum_nr_running += rq->nr_running;
2993 sum_weighted_load += weighted_cpuload(i);
2997 * First idle cpu or the first cpu(busiest) in this sched group
2998 * is eligible for doing load balancing at this and above
2999 * domains. In the newly idle case, we will allow all the cpu's
3000 * to do the newly idle load balance.
3002 if (idle != CPU_NEWLY_IDLE && local_group &&
3003 balance_cpu != this_cpu && balance) {
3008 total_load += avg_load;
3009 total_pwr += group->__cpu_power;
3011 /* Adjust by relative CPU power of the group */
3012 avg_load = sg_div_cpu_power(group,
3013 avg_load * SCHED_LOAD_SCALE);
3015 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3018 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3021 this_load = avg_load;
3023 this_nr_running = sum_nr_running;
3024 this_load_per_task = sum_weighted_load;
3025 } else if (avg_load > max_load &&
3026 (sum_nr_running > group_capacity || __group_imb)) {
3027 max_load = avg_load;
3029 busiest_nr_running = sum_nr_running;
3030 busiest_load_per_task = sum_weighted_load;
3031 group_imb = __group_imb;
3034 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3036 * Busy processors will not participate in power savings
3039 if (idle == CPU_NOT_IDLE ||
3040 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3044 * If the local group is idle or completely loaded
3045 * no need to do power savings balance at this domain
3047 if (local_group && (this_nr_running >= group_capacity ||
3049 power_savings_balance = 0;
3052 * If a group is already running at full capacity or idle,
3053 * don't include that group in power savings calculations
3055 if (!power_savings_balance || sum_nr_running >= group_capacity
3060 * Calculate the group which has the least non-idle load.
3061 * This is the group from where we need to pick up the load
3064 if ((sum_nr_running < min_nr_running) ||
3065 (sum_nr_running == min_nr_running &&
3066 first_cpu(group->cpumask) <
3067 first_cpu(group_min->cpumask))) {
3069 min_nr_running = sum_nr_running;
3070 min_load_per_task = sum_weighted_load /
3075 * Calculate the group which is almost near its
3076 * capacity but still has some space to pick up some load
3077 * from other group and save more power
3079 if (sum_nr_running <= group_capacity - 1) {
3080 if (sum_nr_running > leader_nr_running ||
3081 (sum_nr_running == leader_nr_running &&
3082 first_cpu(group->cpumask) >
3083 first_cpu(group_leader->cpumask))) {
3084 group_leader = group;
3085 leader_nr_running = sum_nr_running;
3090 group = group->next;
3091 } while (group != sd->groups);
3093 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3096 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3098 if (this_load >= avg_load ||
3099 100*max_load <= sd->imbalance_pct*this_load)
3102 busiest_load_per_task /= busiest_nr_running;
3104 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3107 * We're trying to get all the cpus to the average_load, so we don't
3108 * want to push ourselves above the average load, nor do we wish to
3109 * reduce the max loaded cpu below the average load, as either of these
3110 * actions would just result in more rebalancing later, and ping-pong
3111 * tasks around. Thus we look for the minimum possible imbalance.
3112 * Negative imbalances (*we* are more loaded than anyone else) will
3113 * be counted as no imbalance for these purposes -- we can't fix that
3114 * by pulling tasks to us. Be careful of negative numbers as they'll
3115 * appear as very large values with unsigned longs.
3117 if (max_load <= busiest_load_per_task)
3121 * In the presence of smp nice balancing, certain scenarios can have
3122 * max load less than avg load(as we skip the groups at or below
3123 * its cpu_power, while calculating max_load..)
3125 if (max_load < avg_load) {
3127 goto small_imbalance;
3130 /* Don't want to pull so many tasks that a group would go idle */
3131 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3133 /* How much load to actually move to equalise the imbalance */
3134 *imbalance = min(max_pull * busiest->__cpu_power,
3135 (avg_load - this_load) * this->__cpu_power)
3139 * if *imbalance is less than the average load per runnable task
3140 * there is no gaurantee that any tasks will be moved so we'll have
3141 * a think about bumping its value to force at least one task to be
3144 if (*imbalance < busiest_load_per_task) {
3145 unsigned long tmp, pwr_now, pwr_move;
3149 pwr_move = pwr_now = 0;
3151 if (this_nr_running) {
3152 this_load_per_task /= this_nr_running;
3153 if (busiest_load_per_task > this_load_per_task)
3156 this_load_per_task = SCHED_LOAD_SCALE;
3158 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3159 busiest_load_per_task * imbn) {
3160 *imbalance = busiest_load_per_task;
3165 * OK, we don't have enough imbalance to justify moving tasks,
3166 * however we may be able to increase total CPU power used by
3170 pwr_now += busiest->__cpu_power *
3171 min(busiest_load_per_task, max_load);
3172 pwr_now += this->__cpu_power *
3173 min(this_load_per_task, this_load);
3174 pwr_now /= SCHED_LOAD_SCALE;
3176 /* Amount of load we'd subtract */
3177 tmp = sg_div_cpu_power(busiest,
3178 busiest_load_per_task * SCHED_LOAD_SCALE);
3180 pwr_move += busiest->__cpu_power *
3181 min(busiest_load_per_task, max_load - tmp);
3183 /* Amount of load we'd add */
3184 if (max_load * busiest->__cpu_power <
3185 busiest_load_per_task * SCHED_LOAD_SCALE)
3186 tmp = sg_div_cpu_power(this,
3187 max_load * busiest->__cpu_power);
3189 tmp = sg_div_cpu_power(this,
3190 busiest_load_per_task * SCHED_LOAD_SCALE);
3191 pwr_move += this->__cpu_power *
3192 min(this_load_per_task, this_load + tmp);
3193 pwr_move /= SCHED_LOAD_SCALE;
3195 /* Move if we gain throughput */
3196 if (pwr_move > pwr_now)
3197 *imbalance = busiest_load_per_task;
3203 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3204 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3207 if (this == group_leader && group_leader != group_min) {
3208 *imbalance = min_load_per_task;
3218 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3221 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3222 unsigned long imbalance, const cpumask_t *cpus)
3224 struct rq *busiest = NULL, *rq;
3225 unsigned long max_load = 0;
3228 for_each_cpu_mask(i, group->cpumask) {
3231 if (!cpu_isset(i, *cpus))
3235 wl = weighted_cpuload(i);
3237 if (rq->nr_running == 1 && wl > imbalance)
3240 if (wl > max_load) {
3250 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3251 * so long as it is large enough.
3253 #define MAX_PINNED_INTERVAL 512
3256 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3257 * tasks if there is an imbalance.
3259 static int load_balance(int this_cpu, struct rq *this_rq,
3260 struct sched_domain *sd, enum cpu_idle_type idle,
3261 int *balance, cpumask_t *cpus)
3263 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3264 struct sched_group *group;
3265 unsigned long imbalance;
3267 unsigned long flags;
3272 * When power savings policy is enabled for the parent domain, idle
3273 * sibling can pick up load irrespective of busy siblings. In this case,
3274 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3275 * portraying it as CPU_NOT_IDLE.
3277 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3278 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3281 schedstat_inc(sd, lb_count[idle]);
3284 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3291 schedstat_inc(sd, lb_nobusyg[idle]);
3295 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3297 schedstat_inc(sd, lb_nobusyq[idle]);
3301 BUG_ON(busiest == this_rq);
3303 schedstat_add(sd, lb_imbalance[idle], imbalance);
3306 if (busiest->nr_running > 1) {
3308 * Attempt to move tasks. If find_busiest_group has found
3309 * an imbalance but busiest->nr_running <= 1, the group is
3310 * still unbalanced. ld_moved simply stays zero, so it is
3311 * correctly treated as an imbalance.
3313 local_irq_save(flags);
3314 double_rq_lock(this_rq, busiest);
3315 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3316 imbalance, sd, idle, &all_pinned);
3317 double_rq_unlock(this_rq, busiest);
3318 local_irq_restore(flags);
3321 * some other cpu did the load balance for us.
3323 if (ld_moved && this_cpu != smp_processor_id())
3324 resched_cpu(this_cpu);
3326 /* All tasks on this runqueue were pinned by CPU affinity */
3327 if (unlikely(all_pinned)) {
3328 cpu_clear(cpu_of(busiest), *cpus);
3329 if (!cpus_empty(*cpus))
3336 schedstat_inc(sd, lb_failed[idle]);
3337 sd->nr_balance_failed++;
3339 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3341 spin_lock_irqsave(&busiest->lock, flags);
3343 /* don't kick the migration_thread, if the curr
3344 * task on busiest cpu can't be moved to this_cpu
3346 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3347 spin_unlock_irqrestore(&busiest->lock, flags);
3349 goto out_one_pinned;
3352 if (!busiest->active_balance) {
3353 busiest->active_balance = 1;
3354 busiest->push_cpu = this_cpu;
3357 spin_unlock_irqrestore(&busiest->lock, flags);
3359 wake_up_process(busiest->migration_thread);
3362 * We've kicked active balancing, reset the failure
3365 sd->nr_balance_failed = sd->cache_nice_tries+1;
3368 sd->nr_balance_failed = 0;
3370 if (likely(!active_balance)) {
3371 /* We were unbalanced, so reset the balancing interval */
3372 sd->balance_interval = sd->min_interval;
3375 * If we've begun active balancing, start to back off. This
3376 * case may not be covered by the all_pinned logic if there
3377 * is only 1 task on the busy runqueue (because we don't call
3380 if (sd->balance_interval < sd->max_interval)
3381 sd->balance_interval *= 2;
3384 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3385 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3390 schedstat_inc(sd, lb_balanced[idle]);
3392 sd->nr_balance_failed = 0;
3395 /* tune up the balancing interval */
3396 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3397 (sd->balance_interval < sd->max_interval))
3398 sd->balance_interval *= 2;
3400 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3401 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3407 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3408 * tasks if there is an imbalance.
3410 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3411 * this_rq is locked.
3414 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3417 struct sched_group *group;
3418 struct rq *busiest = NULL;
3419 unsigned long imbalance;
3427 * When power savings policy is enabled for the parent domain, idle
3428 * sibling can pick up load irrespective of busy siblings. In this case,
3429 * let the state of idle sibling percolate up as IDLE, instead of
3430 * portraying it as CPU_NOT_IDLE.
3432 if (sd->flags & SD_SHARE_CPUPOWER &&
3433 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3436 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3438 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3439 &sd_idle, cpus, NULL);
3441 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3445 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3447 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3451 BUG_ON(busiest == this_rq);
3453 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3456 if (busiest->nr_running > 1) {
3457 /* Attempt to move tasks */
3458 double_lock_balance(this_rq, busiest);
3459 /* this_rq->clock is already updated */
3460 update_rq_clock(busiest);
3461 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3462 imbalance, sd, CPU_NEWLY_IDLE,
3464 spin_unlock(&busiest->lock);
3466 if (unlikely(all_pinned)) {
3467 cpu_clear(cpu_of(busiest), *cpus);
3468 if (!cpus_empty(*cpus))
3474 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3475 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3476 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3479 sd->nr_balance_failed = 0;
3484 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3485 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3486 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3488 sd->nr_balance_failed = 0;
3494 * idle_balance is called by schedule() if this_cpu is about to become
3495 * idle. Attempts to pull tasks from other CPUs.
3497 static void idle_balance(int this_cpu, struct rq *this_rq)
3499 struct sched_domain *sd;
3500 int pulled_task = -1;
3501 unsigned long next_balance = jiffies + HZ;
3504 for_each_domain(this_cpu, sd) {
3505 unsigned long interval;
3507 if (!(sd->flags & SD_LOAD_BALANCE))
3510 if (sd->flags & SD_BALANCE_NEWIDLE)
3511 /* If we've pulled tasks over stop searching: */
3512 pulled_task = load_balance_newidle(this_cpu, this_rq,
3515 interval = msecs_to_jiffies(sd->balance_interval);
3516 if (time_after(next_balance, sd->last_balance + interval))
3517 next_balance = sd->last_balance + interval;
3521 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3523 * We are going idle. next_balance may be set based on
3524 * a busy processor. So reset next_balance.
3526 this_rq->next_balance = next_balance;
3531 * active_load_balance is run by migration threads. It pushes running tasks
3532 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3533 * running on each physical CPU where possible, and avoids physical /
3534 * logical imbalances.
3536 * Called with busiest_rq locked.
3538 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3540 int target_cpu = busiest_rq->push_cpu;
3541 struct sched_domain *sd;
3542 struct rq *target_rq;
3544 /* Is there any task to move? */
3545 if (busiest_rq->nr_running <= 1)
3548 target_rq = cpu_rq(target_cpu);
3551 * This condition is "impossible", if it occurs
3552 * we need to fix it. Originally reported by
3553 * Bjorn Helgaas on a 128-cpu setup.
3555 BUG_ON(busiest_rq == target_rq);
3557 /* move a task from busiest_rq to target_rq */
3558 double_lock_balance(busiest_rq, target_rq);
3559 update_rq_clock(busiest_rq);
3560 update_rq_clock(target_rq);
3562 /* Search for an sd spanning us and the target CPU. */
3563 for_each_domain(target_cpu, sd) {
3564 if ((sd->flags & SD_LOAD_BALANCE) &&
3565 cpu_isset(busiest_cpu, sd->span))
3570 schedstat_inc(sd, alb_count);
3572 if (move_one_task(target_rq, target_cpu, busiest_rq,
3574 schedstat_inc(sd, alb_pushed);
3576 schedstat_inc(sd, alb_failed);
3578 spin_unlock(&target_rq->lock);
3583 atomic_t load_balancer;
3585 } nohz ____cacheline_aligned = {
3586 .load_balancer = ATOMIC_INIT(-1),
3587 .cpu_mask = CPU_MASK_NONE,
3591 * This routine will try to nominate the ilb (idle load balancing)
3592 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3593 * load balancing on behalf of all those cpus. If all the cpus in the system
3594 * go into this tickless mode, then there will be no ilb owner (as there is
3595 * no need for one) and all the cpus will sleep till the next wakeup event
3598 * For the ilb owner, tick is not stopped. And this tick will be used
3599 * for idle load balancing. ilb owner will still be part of
3602 * While stopping the tick, this cpu will become the ilb owner if there
3603 * is no other owner. And will be the owner till that cpu becomes busy
3604 * or if all cpus in the system stop their ticks at which point
3605 * there is no need for ilb owner.
3607 * When the ilb owner becomes busy, it nominates another owner, during the
3608 * next busy scheduler_tick()
3610 int select_nohz_load_balancer(int stop_tick)
3612 int cpu = smp_processor_id();
3615 cpu_set(cpu, nohz.cpu_mask);
3616 cpu_rq(cpu)->in_nohz_recently = 1;
3619 * If we are going offline and still the leader, give up!
3621 if (cpu_is_offline(cpu) &&
3622 atomic_read(&nohz.load_balancer) == cpu) {
3623 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3628 /* time for ilb owner also to sleep */
3629 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3630 if (atomic_read(&nohz.load_balancer) == cpu)
3631 atomic_set(&nohz.load_balancer, -1);
3635 if (atomic_read(&nohz.load_balancer) == -1) {
3636 /* make me the ilb owner */
3637 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3639 } else if (atomic_read(&nohz.load_balancer) == cpu)
3642 if (!cpu_isset(cpu, nohz.cpu_mask))
3645 cpu_clear(cpu, nohz.cpu_mask);
3647 if (atomic_read(&nohz.load_balancer) == cpu)
3648 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3655 static DEFINE_SPINLOCK(balancing);
3658 * It checks each scheduling domain to see if it is due to be balanced,
3659 * and initiates a balancing operation if so.
3661 * Balancing parameters are set up in arch_init_sched_domains.
3663 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3666 struct rq *rq = cpu_rq(cpu);
3667 unsigned long interval;
3668 struct sched_domain *sd;
3669 /* Earliest time when we have to do rebalance again */
3670 unsigned long next_balance = jiffies + 60*HZ;
3671 int update_next_balance = 0;
3675 for_each_domain(cpu, sd) {
3676 if (!(sd->flags & SD_LOAD_BALANCE))
3679 interval = sd->balance_interval;
3680 if (idle != CPU_IDLE)
3681 interval *= sd->busy_factor;
3683 /* scale ms to jiffies */
3684 interval = msecs_to_jiffies(interval);
3685 if (unlikely(!interval))
3687 if (interval > HZ*NR_CPUS/10)
3688 interval = HZ*NR_CPUS/10;
3690 need_serialize = sd->flags & SD_SERIALIZE;
3692 if (need_serialize) {
3693 if (!spin_trylock(&balancing))
3697 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3698 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3700 * We've pulled tasks over so either we're no
3701 * longer idle, or one of our SMT siblings is
3704 idle = CPU_NOT_IDLE;
3706 sd->last_balance = jiffies;
3709 spin_unlock(&balancing);
3711 if (time_after(next_balance, sd->last_balance + interval)) {
3712 next_balance = sd->last_balance + interval;
3713 update_next_balance = 1;
3717 * Stop the load balance at this level. There is another
3718 * CPU in our sched group which is doing load balancing more
3726 * next_balance will be updated only when there is a need.
3727 * When the cpu is attached to null domain for ex, it will not be
3730 if (likely(update_next_balance))
3731 rq->next_balance = next_balance;
3735 * run_rebalance_domains is triggered when needed from the scheduler tick.
3736 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3737 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3739 static void run_rebalance_domains(struct softirq_action *h)
3741 int this_cpu = smp_processor_id();
3742 struct rq *this_rq = cpu_rq(this_cpu);
3743 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3744 CPU_IDLE : CPU_NOT_IDLE;
3746 rebalance_domains(this_cpu, idle);
3750 * If this cpu is the owner for idle load balancing, then do the
3751 * balancing on behalf of the other idle cpus whose ticks are
3754 if (this_rq->idle_at_tick &&
3755 atomic_read(&nohz.load_balancer) == this_cpu) {
3756 cpumask_t cpus = nohz.cpu_mask;
3760 cpu_clear(this_cpu, cpus);
3761 for_each_cpu_mask(balance_cpu, cpus) {
3763 * If this cpu gets work to do, stop the load balancing
3764 * work being done for other cpus. Next load
3765 * balancing owner will pick it up.
3770 rebalance_domains(balance_cpu, CPU_IDLE);
3772 rq = cpu_rq(balance_cpu);
3773 if (time_after(this_rq->next_balance, rq->next_balance))
3774 this_rq->next_balance = rq->next_balance;
3781 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3783 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3784 * idle load balancing owner or decide to stop the periodic load balancing,
3785 * if the whole system is idle.
3787 static inline void trigger_load_balance(struct rq *rq, int cpu)
3791 * If we were in the nohz mode recently and busy at the current
3792 * scheduler tick, then check if we need to nominate new idle
3795 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3796 rq->in_nohz_recently = 0;
3798 if (atomic_read(&nohz.load_balancer) == cpu) {
3799 cpu_clear(cpu, nohz.cpu_mask);
3800 atomic_set(&nohz.load_balancer, -1);
3803 if (atomic_read(&nohz.load_balancer) == -1) {
3805 * simple selection for now: Nominate the
3806 * first cpu in the nohz list to be the next
3809 * TBD: Traverse the sched domains and nominate
3810 * the nearest cpu in the nohz.cpu_mask.
3812 int ilb = first_cpu(nohz.cpu_mask);
3814 if (ilb < nr_cpu_ids)
3820 * If this cpu is idle and doing idle load balancing for all the
3821 * cpus with ticks stopped, is it time for that to stop?
3823 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3824 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3830 * If this cpu is idle and the idle load balancing is done by
3831 * someone else, then no need raise the SCHED_SOFTIRQ
3833 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3834 cpu_isset(cpu, nohz.cpu_mask))
3837 if (time_after_eq(jiffies, rq->next_balance))
3838 raise_softirq(SCHED_SOFTIRQ);
3841 #else /* CONFIG_SMP */
3844 * on UP we do not need to balance between CPUs:
3846 static inline void idle_balance(int cpu, struct rq *rq)
3852 DEFINE_PER_CPU(struct kernel_stat, kstat);
3854 EXPORT_PER_CPU_SYMBOL(kstat);
3857 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3858 * that have not yet been banked in case the task is currently running.
3860 unsigned long long task_sched_runtime(struct task_struct *p)
3862 unsigned long flags;
3866 rq = task_rq_lock(p, &flags);
3867 ns = p->se.sum_exec_runtime;
3868 if (task_current(rq, p)) {
3869 update_rq_clock(rq);
3870 delta_exec = rq->clock - p->se.exec_start;
3871 if ((s64)delta_exec > 0)
3874 task_rq_unlock(rq, &flags);
3880 * Account user cpu time to a process.
3881 * @p: the process that the cpu time gets accounted to
3882 * @cputime: the cpu time spent in user space since the last update
3884 void account_user_time(struct task_struct *p, cputime_t cputime)
3886 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3889 p->utime = cputime_add(p->utime, cputime);
3891 /* Add user time to cpustat. */
3892 tmp = cputime_to_cputime64(cputime);
3893 if (TASK_NICE(p) > 0)
3894 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3896 cpustat->user = cputime64_add(cpustat->user, tmp);
3900 * Account guest cpu time to a process.
3901 * @p: the process that the cpu time gets accounted to
3902 * @cputime: the cpu time spent in virtual machine since the last update
3904 static void account_guest_time(struct task_struct *p, cputime_t cputime)
3907 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3909 tmp = cputime_to_cputime64(cputime);
3911 p->utime = cputime_add(p->utime, cputime);
3912 p->gtime = cputime_add(p->gtime, cputime);
3914 cpustat->user = cputime64_add(cpustat->user, tmp);
3915 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3919 * Account scaled user cpu time to a process.
3920 * @p: the process that the cpu time gets accounted to
3921 * @cputime: the cpu time spent in user space since the last update
3923 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
3925 p->utimescaled = cputime_add(p->utimescaled, cputime);
3929 * Account system cpu time to a process.
3930 * @p: the process that the cpu time gets accounted to
3931 * @hardirq_offset: the offset to subtract from hardirq_count()
3932 * @cputime: the cpu time spent in kernel space since the last update
3934 void account_system_time(struct task_struct *p, int hardirq_offset,
3937 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3938 struct rq *rq = this_rq();
3941 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3942 account_guest_time(p, cputime);
3946 p->stime = cputime_add(p->stime, cputime);
3948 /* Add system time to cpustat. */
3949 tmp = cputime_to_cputime64(cputime);
3950 if (hardirq_count() - hardirq_offset)
3951 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3952 else if (softirq_count())
3953 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3954 else if (p != rq->idle)
3955 cpustat->system = cputime64_add(cpustat->system, tmp);
3956 else if (atomic_read(&rq->nr_iowait) > 0)
3957 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3959 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3960 /* Account for system time used */
3961 acct_update_integrals(p);
3965 * Account scaled system cpu time to a process.
3966 * @p: the process that the cpu time gets accounted to
3967 * @hardirq_offset: the offset to subtract from hardirq_count()
3968 * @cputime: the cpu time spent in kernel space since the last update
3970 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
3972 p->stimescaled = cputime_add(p->stimescaled, cputime);
3976 * Account for involuntary wait time.
3977 * @p: the process from which the cpu time has been stolen
3978 * @steal: the cpu time spent in involuntary wait
3980 void account_steal_time(struct task_struct *p, cputime_t steal)
3982 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3983 cputime64_t tmp = cputime_to_cputime64(steal);
3984 struct rq *rq = this_rq();
3986 if (p == rq->idle) {
3987 p->stime = cputime_add(p->stime, steal);
3988 if (atomic_read(&rq->nr_iowait) > 0)
3989 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
3991 cpustat->idle = cputime64_add(cpustat->idle, tmp);
3993 cpustat->steal = cputime64_add(cpustat->steal, tmp);
3997 * This function gets called by the timer code, with HZ frequency.
3998 * We call it with interrupts disabled.
4000 * It also gets called by the fork code, when changing the parent's
4003 void scheduler_tick(void)
4005 int cpu = smp_processor_id();
4006 struct rq *rq = cpu_rq(cpu);
4007 struct task_struct *curr = rq->curr;
4011 spin_lock(&rq->lock);
4012 update_rq_clock(rq);
4013 update_cpu_load(rq);
4014 curr->sched_class->task_tick(rq, curr, 0);
4015 spin_unlock(&rq->lock);
4018 rq->idle_at_tick = idle_cpu(cpu);
4019 trigger_load_balance(rq, cpu);
4023 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4025 void __kprobes add_preempt_count(int val)
4030 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4032 preempt_count() += val;
4034 * Spinlock count overflowing soon?
4036 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4039 EXPORT_SYMBOL(add_preempt_count);
4041 void __kprobes sub_preempt_count(int val)
4046 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4049 * Is the spinlock portion underflowing?
4051 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4052 !(preempt_count() & PREEMPT_MASK)))
4055 preempt_count() -= val;
4057 EXPORT_SYMBOL(sub_preempt_count);
4062 * Print scheduling while atomic bug:
4064 static noinline void __schedule_bug(struct task_struct *prev)
4066 struct pt_regs *regs = get_irq_regs();
4068 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4069 prev->comm, prev->pid, preempt_count());
4071 debug_show_held_locks(prev);
4073 if (irqs_disabled())
4074 print_irqtrace_events(prev);
4083 * Various schedule()-time debugging checks and statistics:
4085 static inline void schedule_debug(struct task_struct *prev)
4088 * Test if we are atomic. Since do_exit() needs to call into
4089 * schedule() atomically, we ignore that path for now.
4090 * Otherwise, whine if we are scheduling when we should not be.
4092 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4093 __schedule_bug(prev);
4095 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4097 schedstat_inc(this_rq(), sched_count);
4098 #ifdef CONFIG_SCHEDSTATS
4099 if (unlikely(prev->lock_depth >= 0)) {
4100 schedstat_inc(this_rq(), bkl_count);
4101 schedstat_inc(prev, sched_info.bkl_count);
4107 * Pick up the highest-prio task:
4109 static inline struct task_struct *
4110 pick_next_task(struct rq *rq, struct task_struct *prev)
4112 const struct sched_class *class;
4113 struct task_struct *p;
4116 * Optimization: we know that if all tasks are in
4117 * the fair class we can call that function directly:
4119 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4120 p = fair_sched_class.pick_next_task(rq);
4125 class = sched_class_highest;
4127 p = class->pick_next_task(rq);
4131 * Will never be NULL as the idle class always
4132 * returns a non-NULL p:
4134 class = class->next;
4139 * schedule() is the main scheduler function.
4141 asmlinkage void __sched schedule(void)
4143 struct task_struct *prev, *next;
4144 unsigned long *switch_count;
4146 int cpu, hrtick = sched_feat(HRTICK);
4150 cpu = smp_processor_id();
4154 switch_count = &prev->nivcsw;
4156 release_kernel_lock(prev);
4157 need_resched_nonpreemptible:
4159 schedule_debug(prev);
4165 * Do the rq-clock update outside the rq lock:
4167 local_irq_disable();
4168 update_rq_clock(rq);
4169 spin_lock(&rq->lock);
4170 clear_tsk_need_resched(prev);
4172 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4173 if (unlikely(signal_pending_state(prev->state, prev)))
4174 prev->state = TASK_RUNNING;
4176 deactivate_task(rq, prev, 1);
4177 switch_count = &prev->nvcsw;
4181 if (prev->sched_class->pre_schedule)
4182 prev->sched_class->pre_schedule(rq, prev);
4185 if (unlikely(!rq->nr_running))
4186 idle_balance(cpu, rq);
4188 prev->sched_class->put_prev_task(rq, prev);
4189 next = pick_next_task(rq, prev);
4191 if (likely(prev != next)) {
4192 sched_info_switch(prev, next);
4198 context_switch(rq, prev, next); /* unlocks the rq */
4200 * the context switch might have flipped the stack from under
4201 * us, hence refresh the local variables.
4203 cpu = smp_processor_id();
4206 spin_unlock_irq(&rq->lock);
4211 if (unlikely(reacquire_kernel_lock(current) < 0))
4212 goto need_resched_nonpreemptible;
4214 preempt_enable_no_resched();
4215 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4218 EXPORT_SYMBOL(schedule);
4220 #ifdef CONFIG_PREEMPT
4222 * this is the entry point to schedule() from in-kernel preemption
4223 * off of preempt_enable. Kernel preemptions off return from interrupt
4224 * occur there and call schedule directly.
4226 asmlinkage void __sched preempt_schedule(void)
4228 struct thread_info *ti = current_thread_info();
4231 * If there is a non-zero preempt_count or interrupts are disabled,
4232 * we do not want to preempt the current task. Just return..
4234 if (likely(ti->preempt_count || irqs_disabled()))
4238 add_preempt_count(PREEMPT_ACTIVE);
4240 sub_preempt_count(PREEMPT_ACTIVE);
4243 * Check again in case we missed a preemption opportunity
4244 * between schedule and now.
4247 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4249 EXPORT_SYMBOL(preempt_schedule);
4252 * this is the entry point to schedule() from kernel preemption
4253 * off of irq context.
4254 * Note, that this is called and return with irqs disabled. This will
4255 * protect us against recursive calling from irq.
4257 asmlinkage void __sched preempt_schedule_irq(void)
4259 struct thread_info *ti = current_thread_info();
4261 /* Catch callers which need to be fixed */
4262 BUG_ON(ti->preempt_count || !irqs_disabled());
4265 add_preempt_count(PREEMPT_ACTIVE);
4268 local_irq_disable();
4269 sub_preempt_count(PREEMPT_ACTIVE);
4272 * Check again in case we missed a preemption opportunity
4273 * between schedule and now.
4276 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4279 #endif /* CONFIG_PREEMPT */
4281 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4284 return try_to_wake_up(curr->private, mode, sync);
4286 EXPORT_SYMBOL(default_wake_function);
4289 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4290 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4291 * number) then we wake all the non-exclusive tasks and one exclusive task.
4293 * There are circumstances in which we can try to wake a task which has already
4294 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4295 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4297 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4298 int nr_exclusive, int sync, void *key)
4300 wait_queue_t *curr, *next;
4302 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4303 unsigned flags = curr->flags;
4305 if (curr->func(curr, mode, sync, key) &&
4306 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4312 * __wake_up - wake up threads blocked on a waitqueue.
4314 * @mode: which threads
4315 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4316 * @key: is directly passed to the wakeup function
4318 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4319 int nr_exclusive, void *key)
4321 unsigned long flags;
4323 spin_lock_irqsave(&q->lock, flags);
4324 __wake_up_common(q, mode, nr_exclusive, 0, key);
4325 spin_unlock_irqrestore(&q->lock, flags);
4327 EXPORT_SYMBOL(__wake_up);
4330 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4332 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4334 __wake_up_common(q, mode, 1, 0, NULL);
4338 * __wake_up_sync - wake up threads blocked on a waitqueue.
4340 * @mode: which threads
4341 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4343 * The sync wakeup differs that the waker knows that it will schedule
4344 * away soon, so while the target thread will be woken up, it will not
4345 * be migrated to another CPU - ie. the two threads are 'synchronized'
4346 * with each other. This can prevent needless bouncing between CPUs.
4348 * On UP it can prevent extra preemption.
4351 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4353 unsigned long flags;
4359 if (unlikely(!nr_exclusive))
4362 spin_lock_irqsave(&q->lock, flags);
4363 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4364 spin_unlock_irqrestore(&q->lock, flags);
4366 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4368 void complete(struct completion *x)
4370 unsigned long flags;
4372 spin_lock_irqsave(&x->wait.lock, flags);
4374 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4375 spin_unlock_irqrestore(&x->wait.lock, flags);
4377 EXPORT_SYMBOL(complete);
4379 void complete_all(struct completion *x)
4381 unsigned long flags;
4383 spin_lock_irqsave(&x->wait.lock, flags);
4384 x->done += UINT_MAX/2;
4385 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4386 spin_unlock_irqrestore(&x->wait.lock, flags);
4388 EXPORT_SYMBOL(complete_all);
4390 static inline long __sched
4391 do_wait_for_common(struct completion *x, long timeout, int state)
4394 DECLARE_WAITQUEUE(wait, current);
4396 wait.flags |= WQ_FLAG_EXCLUSIVE;
4397 __add_wait_queue_tail(&x->wait, &wait);
4399 if ((state == TASK_INTERRUPTIBLE &&
4400 signal_pending(current)) ||
4401 (state == TASK_KILLABLE &&
4402 fatal_signal_pending(current))) {
4403 timeout = -ERESTARTSYS;
4406 __set_current_state(state);
4407 spin_unlock_irq(&x->wait.lock);
4408 timeout = schedule_timeout(timeout);
4409 spin_lock_irq(&x->wait.lock);
4410 } while (!x->done && timeout);
4411 __remove_wait_queue(&x->wait, &wait);
4416 return timeout ?: 1;
4420 wait_for_common(struct completion *x, long timeout, int state)
4424 spin_lock_irq(&x->wait.lock);
4425 timeout = do_wait_for_common(x, timeout, state);
4426 spin_unlock_irq(&x->wait.lock);
4430 void __sched wait_for_completion(struct completion *x)
4432 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4434 EXPORT_SYMBOL(wait_for_completion);
4436 unsigned long __sched
4437 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4439 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4441 EXPORT_SYMBOL(wait_for_completion_timeout);
4443 int __sched wait_for_completion_interruptible(struct completion *x)
4445 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4446 if (t == -ERESTARTSYS)
4450 EXPORT_SYMBOL(wait_for_completion_interruptible);
4452 unsigned long __sched
4453 wait_for_completion_interruptible_timeout(struct completion *x,
4454 unsigned long timeout)
4456 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4458 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4460 int __sched wait_for_completion_killable(struct completion *x)
4462 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4463 if (t == -ERESTARTSYS)
4467 EXPORT_SYMBOL(wait_for_completion_killable);
4470 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4472 unsigned long flags;
4475 init_waitqueue_entry(&wait, current);
4477 __set_current_state(state);
4479 spin_lock_irqsave(&q->lock, flags);
4480 __add_wait_queue(q, &wait);
4481 spin_unlock(&q->lock);
4482 timeout = schedule_timeout(timeout);
4483 spin_lock_irq(&q->lock);
4484 __remove_wait_queue(q, &wait);
4485 spin_unlock_irqrestore(&q->lock, flags);
4490 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4492 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4494 EXPORT_SYMBOL(interruptible_sleep_on);
4497 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4499 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4501 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4503 void __sched sleep_on(wait_queue_head_t *q)
4505 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4507 EXPORT_SYMBOL(sleep_on);
4509 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4511 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4513 EXPORT_SYMBOL(sleep_on_timeout);
4515 #ifdef CONFIG_RT_MUTEXES
4518 * rt_mutex_setprio - set the current priority of a task
4520 * @prio: prio value (kernel-internal form)
4522 * This function changes the 'effective' priority of a task. It does
4523 * not touch ->normal_prio like __setscheduler().
4525 * Used by the rt_mutex code to implement priority inheritance logic.
4527 void rt_mutex_setprio(struct task_struct *p, int prio)
4529 unsigned long flags;
4530 int oldprio, on_rq, running;
4532 const struct sched_class *prev_class = p->sched_class;
4534 BUG_ON(prio < 0 || prio > MAX_PRIO);
4536 rq = task_rq_lock(p, &flags);
4537 update_rq_clock(rq);
4540 on_rq = p->se.on_rq;
4541 running = task_current(rq, p);
4543 dequeue_task(rq, p, 0);
4545 p->sched_class->put_prev_task(rq, p);
4548 p->sched_class = &rt_sched_class;
4550 p->sched_class = &fair_sched_class;
4555 p->sched_class->set_curr_task(rq);
4557 enqueue_task(rq, p, 0);
4559 check_class_changed(rq, p, prev_class, oldprio, running);
4561 task_rq_unlock(rq, &flags);
4566 void set_user_nice(struct task_struct *p, long nice)
4568 int old_prio, delta, on_rq;
4569 unsigned long flags;
4572 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4575 * We have to be careful, if called from sys_setpriority(),
4576 * the task might be in the middle of scheduling on another CPU.
4578 rq = task_rq_lock(p, &flags);
4579 update_rq_clock(rq);
4581 * The RT priorities are set via sched_setscheduler(), but we still
4582 * allow the 'normal' nice value to be set - but as expected
4583 * it wont have any effect on scheduling until the task is
4584 * SCHED_FIFO/SCHED_RR:
4586 if (task_has_rt_policy(p)) {
4587 p->static_prio = NICE_TO_PRIO(nice);
4590 on_rq = p->se.on_rq;
4592 dequeue_task(rq, p, 0);
4596 p->static_prio = NICE_TO_PRIO(nice);
4599 p->prio = effective_prio(p);
4600 delta = p->prio - old_prio;
4603 enqueue_task(rq, p, 0);
4606 * If the task increased its priority or is running and
4607 * lowered its priority, then reschedule its CPU:
4609 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4610 resched_task(rq->curr);
4613 task_rq_unlock(rq, &flags);
4615 EXPORT_SYMBOL(set_user_nice);
4618 * can_nice - check if a task can reduce its nice value
4622 int can_nice(const struct task_struct *p, const int nice)
4624 /* convert nice value [19,-20] to rlimit style value [1,40] */
4625 int nice_rlim = 20 - nice;
4627 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4628 capable(CAP_SYS_NICE));
4631 #ifdef __ARCH_WANT_SYS_NICE
4634 * sys_nice - change the priority of the current process.
4635 * @increment: priority increment
4637 * sys_setpriority is a more generic, but much slower function that
4638 * does similar things.
4640 asmlinkage long sys_nice(int increment)
4645 * Setpriority might change our priority at the same moment.
4646 * We don't have to worry. Conceptually one call occurs first
4647 * and we have a single winner.
4649 if (increment < -40)
4654 nice = PRIO_TO_NICE(current->static_prio) + increment;
4660 if (increment < 0 && !can_nice(current, nice))
4663 retval = security_task_setnice(current, nice);
4667 set_user_nice(current, nice);
4674 * task_prio - return the priority value of a given task.
4675 * @p: the task in question.
4677 * This is the priority value as seen by users in /proc.
4678 * RT tasks are offset by -200. Normal tasks are centered
4679 * around 0, value goes from -16 to +15.
4681 int task_prio(const struct task_struct *p)
4683 return p->prio - MAX_RT_PRIO;
4687 * task_nice - return the nice value of a given task.
4688 * @p: the task in question.
4690 int task_nice(const struct task_struct *p)
4692 return TASK_NICE(p);
4694 EXPORT_SYMBOL(task_nice);
4697 * idle_cpu - is a given cpu idle currently?
4698 * @cpu: the processor in question.
4700 int idle_cpu(int cpu)
4702 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4706 * idle_task - return the idle task for a given cpu.
4707 * @cpu: the processor in question.
4709 struct task_struct *idle_task(int cpu)
4711 return cpu_rq(cpu)->idle;
4715 * find_process_by_pid - find a process with a matching PID value.
4716 * @pid: the pid in question.
4718 static struct task_struct *find_process_by_pid(pid_t pid)
4720 return pid ? find_task_by_vpid(pid) : current;
4723 /* Actually do priority change: must hold rq lock. */
4725 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4727 BUG_ON(p->se.on_rq);
4730 switch (p->policy) {
4734 p->sched_class = &fair_sched_class;
4738 p->sched_class = &rt_sched_class;
4742 p->rt_priority = prio;
4743 p->normal_prio = normal_prio(p);
4744 /* we are holding p->pi_lock already */
4745 p->prio = rt_mutex_getprio(p);
4750 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4751 * @p: the task in question.
4752 * @policy: new policy.
4753 * @param: structure containing the new RT priority.
4755 * NOTE that the task may be already dead.
4757 int sched_setscheduler(struct task_struct *p, int policy,
4758 struct sched_param *param)
4760 int retval, oldprio, oldpolicy = -1, on_rq, running;
4761 unsigned long flags;
4762 const struct sched_class *prev_class = p->sched_class;
4765 /* may grab non-irq protected spin_locks */
4766 BUG_ON(in_interrupt());
4768 /* double check policy once rq lock held */
4770 policy = oldpolicy = p->policy;
4771 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4772 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4773 policy != SCHED_IDLE)
4776 * Valid priorities for SCHED_FIFO and SCHED_RR are
4777 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4778 * SCHED_BATCH and SCHED_IDLE is 0.
4780 if (param->sched_priority < 0 ||
4781 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4782 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4784 if (rt_policy(policy) != (param->sched_priority != 0))
4788 * Allow unprivileged RT tasks to decrease priority:
4790 if (!capable(CAP_SYS_NICE)) {
4791 if (rt_policy(policy)) {
4792 unsigned long rlim_rtprio;
4794 if (!lock_task_sighand(p, &flags))
4796 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4797 unlock_task_sighand(p, &flags);
4799 /* can't set/change the rt policy */
4800 if (policy != p->policy && !rlim_rtprio)
4803 /* can't increase priority */
4804 if (param->sched_priority > p->rt_priority &&
4805 param->sched_priority > rlim_rtprio)
4809 * Like positive nice levels, dont allow tasks to
4810 * move out of SCHED_IDLE either:
4812 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4815 /* can't change other user's priorities */
4816 if ((current->euid != p->euid) &&
4817 (current->euid != p->uid))
4821 #ifdef CONFIG_RT_GROUP_SCHED
4823 * Do not allow realtime tasks into groups that have no runtime
4826 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4830 retval = security_task_setscheduler(p, policy, param);
4834 * make sure no PI-waiters arrive (or leave) while we are
4835 * changing the priority of the task:
4837 spin_lock_irqsave(&p->pi_lock, flags);
4839 * To be able to change p->policy safely, the apropriate
4840 * runqueue lock must be held.
4842 rq = __task_rq_lock(p);
4843 /* recheck policy now with rq lock held */
4844 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4845 policy = oldpolicy = -1;
4846 __task_rq_unlock(rq);
4847 spin_unlock_irqrestore(&p->pi_lock, flags);
4850 update_rq_clock(rq);
4851 on_rq = p->se.on_rq;
4852 running = task_current(rq, p);
4854 deactivate_task(rq, p, 0);
4856 p->sched_class->put_prev_task(rq, p);
4859 __setscheduler(rq, p, policy, param->sched_priority);
4862 p->sched_class->set_curr_task(rq);
4864 activate_task(rq, p, 0);
4866 check_class_changed(rq, p, prev_class, oldprio, running);
4868 __task_rq_unlock(rq);
4869 spin_unlock_irqrestore(&p->pi_lock, flags);
4871 rt_mutex_adjust_pi(p);
4875 EXPORT_SYMBOL_GPL(sched_setscheduler);
4878 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4880 struct sched_param lparam;
4881 struct task_struct *p;
4884 if (!param || pid < 0)
4886 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4891 p = find_process_by_pid(pid);
4893 retval = sched_setscheduler(p, policy, &lparam);
4900 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4901 * @pid: the pid in question.
4902 * @policy: new policy.
4903 * @param: structure containing the new RT priority.
4906 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4908 /* negative values for policy are not valid */
4912 return do_sched_setscheduler(pid, policy, param);
4916 * sys_sched_setparam - set/change the RT priority of a thread
4917 * @pid: the pid in question.
4918 * @param: structure containing the new RT priority.
4920 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
4922 return do_sched_setscheduler(pid, -1, param);
4926 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4927 * @pid: the pid in question.
4929 asmlinkage long sys_sched_getscheduler(pid_t pid)
4931 struct task_struct *p;
4938 read_lock(&tasklist_lock);
4939 p = find_process_by_pid(pid);
4941 retval = security_task_getscheduler(p);
4945 read_unlock(&tasklist_lock);
4950 * sys_sched_getscheduler - get the RT priority of a thread
4951 * @pid: the pid in question.
4952 * @param: structure containing the RT priority.
4954 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
4956 struct sched_param lp;
4957 struct task_struct *p;
4960 if (!param || pid < 0)
4963 read_lock(&tasklist_lock);
4964 p = find_process_by_pid(pid);
4969 retval = security_task_getscheduler(p);
4973 lp.sched_priority = p->rt_priority;
4974 read_unlock(&tasklist_lock);
4977 * This one might sleep, we cannot do it with a spinlock held ...
4979 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4984 read_unlock(&tasklist_lock);
4988 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
4990 cpumask_t cpus_allowed;
4991 cpumask_t new_mask = *in_mask;
4992 struct task_struct *p;
4996 read_lock(&tasklist_lock);
4998 p = find_process_by_pid(pid);
5000 read_unlock(&tasklist_lock);
5006 * It is not safe to call set_cpus_allowed with the
5007 * tasklist_lock held. We will bump the task_struct's
5008 * usage count and then drop tasklist_lock.
5011 read_unlock(&tasklist_lock);
5014 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5015 !capable(CAP_SYS_NICE))
5018 retval = security_task_setscheduler(p, 0, NULL);
5022 cpuset_cpus_allowed(p, &cpus_allowed);
5023 cpus_and(new_mask, new_mask, cpus_allowed);
5025 retval = set_cpus_allowed_ptr(p, &new_mask);
5028 cpuset_cpus_allowed(p, &cpus_allowed);
5029 if (!cpus_subset(new_mask, cpus_allowed)) {
5031 * We must have raced with a concurrent cpuset
5032 * update. Just reset the cpus_allowed to the
5033 * cpuset's cpus_allowed
5035 new_mask = cpus_allowed;
5045 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5046 cpumask_t *new_mask)
5048 if (len < sizeof(cpumask_t)) {
5049 memset(new_mask, 0, sizeof(cpumask_t));
5050 } else if (len > sizeof(cpumask_t)) {
5051 len = sizeof(cpumask_t);
5053 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5057 * sys_sched_setaffinity - set the cpu affinity of a process
5058 * @pid: pid of the process
5059 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5060 * @user_mask_ptr: user-space pointer to the new cpu mask
5062 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5063 unsigned long __user *user_mask_ptr)
5068 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5072 return sched_setaffinity(pid, &new_mask);
5075 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5077 struct task_struct *p;
5081 read_lock(&tasklist_lock);
5084 p = find_process_by_pid(pid);
5088 retval = security_task_getscheduler(p);
5092 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5095 read_unlock(&tasklist_lock);
5102 * sys_sched_getaffinity - get the cpu affinity of a process
5103 * @pid: pid of the process
5104 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5105 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5107 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5108 unsigned long __user *user_mask_ptr)
5113 if (len < sizeof(cpumask_t))
5116 ret = sched_getaffinity(pid, &mask);
5120 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5123 return sizeof(cpumask_t);
5127 * sys_sched_yield - yield the current processor to other threads.
5129 * This function yields the current CPU to other tasks. If there are no
5130 * other threads running on this CPU then this function will return.
5132 asmlinkage long sys_sched_yield(void)
5134 struct rq *rq = this_rq_lock();
5136 schedstat_inc(rq, yld_count);
5137 current->sched_class->yield_task(rq);
5140 * Since we are going to call schedule() anyway, there's
5141 * no need to preempt or enable interrupts:
5143 __release(rq->lock);
5144 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5145 _raw_spin_unlock(&rq->lock);
5146 preempt_enable_no_resched();
5153 static void __cond_resched(void)
5155 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5156 __might_sleep(__FILE__, __LINE__);
5159 * The BKS might be reacquired before we have dropped
5160 * PREEMPT_ACTIVE, which could trigger a second
5161 * cond_resched() call.
5164 add_preempt_count(PREEMPT_ACTIVE);
5166 sub_preempt_count(PREEMPT_ACTIVE);
5167 } while (need_resched());
5170 int __sched _cond_resched(void)
5172 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5173 system_state == SYSTEM_RUNNING) {
5179 EXPORT_SYMBOL(_cond_resched);
5182 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5183 * call schedule, and on return reacquire the lock.
5185 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5186 * operations here to prevent schedule() from being called twice (once via
5187 * spin_unlock(), once by hand).
5189 int cond_resched_lock(spinlock_t *lock)
5191 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5194 if (spin_needbreak(lock) || resched) {
5196 if (resched && need_resched())
5205 EXPORT_SYMBOL(cond_resched_lock);
5207 int __sched cond_resched_softirq(void)
5209 BUG_ON(!in_softirq());
5211 if (need_resched() && system_state == SYSTEM_RUNNING) {
5219 EXPORT_SYMBOL(cond_resched_softirq);
5222 * yield - yield the current processor to other threads.
5224 * This is a shortcut for kernel-space yielding - it marks the
5225 * thread runnable and calls sys_sched_yield().
5227 void __sched yield(void)
5229 set_current_state(TASK_RUNNING);
5232 EXPORT_SYMBOL(yield);
5235 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5236 * that process accounting knows that this is a task in IO wait state.
5238 * But don't do that if it is a deliberate, throttling IO wait (this task
5239 * has set its backing_dev_info: the queue against which it should throttle)
5241 void __sched io_schedule(void)
5243 struct rq *rq = &__raw_get_cpu_var(runqueues);
5245 delayacct_blkio_start();
5246 atomic_inc(&rq->nr_iowait);
5248 atomic_dec(&rq->nr_iowait);
5249 delayacct_blkio_end();
5251 EXPORT_SYMBOL(io_schedule);
5253 long __sched io_schedule_timeout(long timeout)
5255 struct rq *rq = &__raw_get_cpu_var(runqueues);
5258 delayacct_blkio_start();
5259 atomic_inc(&rq->nr_iowait);
5260 ret = schedule_timeout(timeout);
5261 atomic_dec(&rq->nr_iowait);
5262 delayacct_blkio_end();
5267 * sys_sched_get_priority_max - return maximum RT priority.
5268 * @policy: scheduling class.
5270 * this syscall returns the maximum rt_priority that can be used
5271 * by a given scheduling class.
5273 asmlinkage long sys_sched_get_priority_max(int policy)
5280 ret = MAX_USER_RT_PRIO-1;
5292 * sys_sched_get_priority_min - return minimum RT priority.
5293 * @policy: scheduling class.
5295 * this syscall returns the minimum rt_priority that can be used
5296 * by a given scheduling class.
5298 asmlinkage long sys_sched_get_priority_min(int policy)
5316 * sys_sched_rr_get_interval - return the default timeslice of a process.
5317 * @pid: pid of the process.
5318 * @interval: userspace pointer to the timeslice value.
5320 * this syscall writes the default timeslice value of a given process
5321 * into the user-space timespec buffer. A value of '0' means infinity.
5324 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5326 struct task_struct *p;
5327 unsigned int time_slice;
5335 read_lock(&tasklist_lock);
5336 p = find_process_by_pid(pid);
5340 retval = security_task_getscheduler(p);
5345 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5346 * tasks that are on an otherwise idle runqueue:
5349 if (p->policy == SCHED_RR) {
5350 time_slice = DEF_TIMESLICE;
5351 } else if (p->policy != SCHED_FIFO) {
5352 struct sched_entity *se = &p->se;
5353 unsigned long flags;
5356 rq = task_rq_lock(p, &flags);
5357 if (rq->cfs.load.weight)
5358 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5359 task_rq_unlock(rq, &flags);
5361 read_unlock(&tasklist_lock);
5362 jiffies_to_timespec(time_slice, &t);
5363 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5367 read_unlock(&tasklist_lock);
5371 static const char stat_nam[] = "RSDTtZX";
5373 void sched_show_task(struct task_struct *p)
5375 unsigned long free = 0;
5378 state = p->state ? __ffs(p->state) + 1 : 0;
5379 printk(KERN_INFO "%-13.13s %c", p->comm,
5380 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5381 #if BITS_PER_LONG == 32
5382 if (state == TASK_RUNNING)
5383 printk(KERN_CONT " running ");
5385 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5387 if (state == TASK_RUNNING)
5388 printk(KERN_CONT " running task ");
5390 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5392 #ifdef CONFIG_DEBUG_STACK_USAGE
5394 unsigned long *n = end_of_stack(p);
5397 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5400 printk(KERN_CONT "%5lu %5d %6d\n", free,
5401 task_pid_nr(p), task_pid_nr(p->real_parent));
5403 show_stack(p, NULL);
5406 void show_state_filter(unsigned long state_filter)
5408 struct task_struct *g, *p;
5410 #if BITS_PER_LONG == 32
5412 " task PC stack pid father\n");
5415 " task PC stack pid father\n");
5417 read_lock(&tasklist_lock);
5418 do_each_thread(g, p) {
5420 * reset the NMI-timeout, listing all files on a slow
5421 * console might take alot of time:
5423 touch_nmi_watchdog();
5424 if (!state_filter || (p->state & state_filter))
5426 } while_each_thread(g, p);
5428 touch_all_softlockup_watchdogs();
5430 #ifdef CONFIG_SCHED_DEBUG
5431 sysrq_sched_debug_show();
5433 read_unlock(&tasklist_lock);
5435 * Only show locks if all tasks are dumped:
5437 if (state_filter == -1)
5438 debug_show_all_locks();
5441 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5443 idle->sched_class = &idle_sched_class;
5447 * init_idle - set up an idle thread for a given CPU
5448 * @idle: task in question
5449 * @cpu: cpu the idle task belongs to
5451 * NOTE: this function does not set the idle thread's NEED_RESCHED
5452 * flag, to make booting more robust.
5454 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5456 struct rq *rq = cpu_rq(cpu);
5457 unsigned long flags;
5460 idle->se.exec_start = sched_clock();
5462 idle->prio = idle->normal_prio = MAX_PRIO;
5463 idle->cpus_allowed = cpumask_of_cpu(cpu);
5464 __set_task_cpu(idle, cpu);
5466 spin_lock_irqsave(&rq->lock, flags);
5467 rq->curr = rq->idle = idle;
5468 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5471 spin_unlock_irqrestore(&rq->lock, flags);
5473 /* Set the preempt count _outside_ the spinlocks! */
5474 #if defined(CONFIG_PREEMPT)
5475 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5477 task_thread_info(idle)->preempt_count = 0;
5480 * The idle tasks have their own, simple scheduling class:
5482 idle->sched_class = &idle_sched_class;
5486 * In a system that switches off the HZ timer nohz_cpu_mask
5487 * indicates which cpus entered this state. This is used
5488 * in the rcu update to wait only for active cpus. For system
5489 * which do not switch off the HZ timer nohz_cpu_mask should
5490 * always be CPU_MASK_NONE.
5492 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5495 * Increase the granularity value when there are more CPUs,
5496 * because with more CPUs the 'effective latency' as visible
5497 * to users decreases. But the relationship is not linear,
5498 * so pick a second-best guess by going with the log2 of the
5501 * This idea comes from the SD scheduler of Con Kolivas:
5503 static inline void sched_init_granularity(void)
5505 unsigned int factor = 1 + ilog2(num_online_cpus());
5506 const unsigned long limit = 200000000;
5508 sysctl_sched_min_granularity *= factor;
5509 if (sysctl_sched_min_granularity > limit)
5510 sysctl_sched_min_granularity = limit;
5512 sysctl_sched_latency *= factor;
5513 if (sysctl_sched_latency > limit)
5514 sysctl_sched_latency = limit;
5516 sysctl_sched_wakeup_granularity *= factor;
5521 * This is how migration works:
5523 * 1) we queue a struct migration_req structure in the source CPU's
5524 * runqueue and wake up that CPU's migration thread.
5525 * 2) we down() the locked semaphore => thread blocks.
5526 * 3) migration thread wakes up (implicitly it forces the migrated
5527 * thread off the CPU)
5528 * 4) it gets the migration request and checks whether the migrated
5529 * task is still in the wrong runqueue.
5530 * 5) if it's in the wrong runqueue then the migration thread removes
5531 * it and puts it into the right queue.
5532 * 6) migration thread up()s the semaphore.
5533 * 7) we wake up and the migration is done.
5537 * Change a given task's CPU affinity. Migrate the thread to a
5538 * proper CPU and schedule it away if the CPU it's executing on
5539 * is removed from the allowed bitmask.
5541 * NOTE: the caller must have a valid reference to the task, the
5542 * task must not exit() & deallocate itself prematurely. The
5543 * call is not atomic; no spinlocks may be held.
5545 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5547 struct migration_req req;
5548 unsigned long flags;
5552 rq = task_rq_lock(p, &flags);
5553 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5558 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5559 !cpus_equal(p->cpus_allowed, *new_mask))) {
5564 if (p->sched_class->set_cpus_allowed)
5565 p->sched_class->set_cpus_allowed(p, new_mask);
5567 p->cpus_allowed = *new_mask;
5568 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5571 /* Can the task run on the task's current CPU? If so, we're done */
5572 if (cpu_isset(task_cpu(p), *new_mask))
5575 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5576 /* Need help from migration thread: drop lock and wait. */
5577 task_rq_unlock(rq, &flags);
5578 wake_up_process(rq->migration_thread);
5579 wait_for_completion(&req.done);
5580 tlb_migrate_finish(p->mm);
5584 task_rq_unlock(rq, &flags);
5588 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5591 * Move (not current) task off this cpu, onto dest cpu. We're doing
5592 * this because either it can't run here any more (set_cpus_allowed()
5593 * away from this CPU, or CPU going down), or because we're
5594 * attempting to rebalance this task on exec (sched_exec).
5596 * So we race with normal scheduler movements, but that's OK, as long
5597 * as the task is no longer on this CPU.
5599 * Returns non-zero if task was successfully migrated.
5601 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5603 struct rq *rq_dest, *rq_src;
5606 if (unlikely(cpu_is_offline(dest_cpu)))
5609 rq_src = cpu_rq(src_cpu);
5610 rq_dest = cpu_rq(dest_cpu);
5612 double_rq_lock(rq_src, rq_dest);
5613 /* Already moved. */
5614 if (task_cpu(p) != src_cpu)
5616 /* Affinity changed (again). */
5617 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5620 on_rq = p->se.on_rq;
5622 deactivate_task(rq_src, p, 0);
5624 set_task_cpu(p, dest_cpu);
5626 activate_task(rq_dest, p, 0);
5627 check_preempt_curr(rq_dest, p);
5631 double_rq_unlock(rq_src, rq_dest);
5636 * migration_thread - this is a highprio system thread that performs
5637 * thread migration by bumping thread off CPU then 'pushing' onto
5640 static int migration_thread(void *data)
5642 int cpu = (long)data;
5646 BUG_ON(rq->migration_thread != current);
5648 set_current_state(TASK_INTERRUPTIBLE);
5649 while (!kthread_should_stop()) {
5650 struct migration_req *req;
5651 struct list_head *head;
5653 spin_lock_irq(&rq->lock);
5655 if (cpu_is_offline(cpu)) {
5656 spin_unlock_irq(&rq->lock);
5660 if (rq->active_balance) {
5661 active_load_balance(rq, cpu);
5662 rq->active_balance = 0;
5665 head = &rq->migration_queue;
5667 if (list_empty(head)) {
5668 spin_unlock_irq(&rq->lock);
5670 set_current_state(TASK_INTERRUPTIBLE);
5673 req = list_entry(head->next, struct migration_req, list);
5674 list_del_init(head->next);
5676 spin_unlock(&rq->lock);
5677 __migrate_task(req->task, cpu, req->dest_cpu);
5680 complete(&req->done);
5682 __set_current_state(TASK_RUNNING);
5686 /* Wait for kthread_stop */
5687 set_current_state(TASK_INTERRUPTIBLE);
5688 while (!kthread_should_stop()) {
5690 set_current_state(TASK_INTERRUPTIBLE);
5692 __set_current_state(TASK_RUNNING);
5696 #ifdef CONFIG_HOTPLUG_CPU
5698 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5702 local_irq_disable();
5703 ret = __migrate_task(p, src_cpu, dest_cpu);
5709 * Figure out where task on dead CPU should go, use force if necessary.
5710 * NOTE: interrupts should be disabled by the caller
5712 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5714 unsigned long flags;
5721 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5722 cpus_and(mask, mask, p->cpus_allowed);
5723 dest_cpu = any_online_cpu(mask);
5725 /* On any allowed CPU? */
5726 if (dest_cpu >= nr_cpu_ids)
5727 dest_cpu = any_online_cpu(p->cpus_allowed);
5729 /* No more Mr. Nice Guy. */
5730 if (dest_cpu >= nr_cpu_ids) {
5731 cpumask_t cpus_allowed;
5733 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5735 * Try to stay on the same cpuset, where the
5736 * current cpuset may be a subset of all cpus.
5737 * The cpuset_cpus_allowed_locked() variant of
5738 * cpuset_cpus_allowed() will not block. It must be
5739 * called within calls to cpuset_lock/cpuset_unlock.
5741 rq = task_rq_lock(p, &flags);
5742 p->cpus_allowed = cpus_allowed;
5743 dest_cpu = any_online_cpu(p->cpus_allowed);
5744 task_rq_unlock(rq, &flags);
5747 * Don't tell them about moving exiting tasks or
5748 * kernel threads (both mm NULL), since they never
5751 if (p->mm && printk_ratelimit()) {
5752 printk(KERN_INFO "process %d (%s) no "
5753 "longer affine to cpu%d\n",
5754 task_pid_nr(p), p->comm, dead_cpu);
5757 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5761 * While a dead CPU has no uninterruptible tasks queued at this point,
5762 * it might still have a nonzero ->nr_uninterruptible counter, because
5763 * for performance reasons the counter is not stricly tracking tasks to
5764 * their home CPUs. So we just add the counter to another CPU's counter,
5765 * to keep the global sum constant after CPU-down:
5767 static void migrate_nr_uninterruptible(struct rq *rq_src)
5769 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5770 unsigned long flags;
5772 local_irq_save(flags);
5773 double_rq_lock(rq_src, rq_dest);
5774 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5775 rq_src->nr_uninterruptible = 0;
5776 double_rq_unlock(rq_src, rq_dest);
5777 local_irq_restore(flags);
5780 /* Run through task list and migrate tasks from the dead cpu. */
5781 static void migrate_live_tasks(int src_cpu)
5783 struct task_struct *p, *t;
5785 read_lock(&tasklist_lock);
5787 do_each_thread(t, p) {
5791 if (task_cpu(p) == src_cpu)
5792 move_task_off_dead_cpu(src_cpu, p);
5793 } while_each_thread(t, p);
5795 read_unlock(&tasklist_lock);
5799 * Schedules idle task to be the next runnable task on current CPU.
5800 * It does so by boosting its priority to highest possible.
5801 * Used by CPU offline code.
5803 void sched_idle_next(void)
5805 int this_cpu = smp_processor_id();
5806 struct rq *rq = cpu_rq(this_cpu);
5807 struct task_struct *p = rq->idle;
5808 unsigned long flags;
5810 /* cpu has to be offline */
5811 BUG_ON(cpu_online(this_cpu));
5814 * Strictly not necessary since rest of the CPUs are stopped by now
5815 * and interrupts disabled on the current cpu.
5817 spin_lock_irqsave(&rq->lock, flags);
5819 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5821 update_rq_clock(rq);
5822 activate_task(rq, p, 0);
5824 spin_unlock_irqrestore(&rq->lock, flags);
5828 * Ensures that the idle task is using init_mm right before its cpu goes
5831 void idle_task_exit(void)
5833 struct mm_struct *mm = current->active_mm;
5835 BUG_ON(cpu_online(smp_processor_id()));
5838 switch_mm(mm, &init_mm, current);
5842 /* called under rq->lock with disabled interrupts */
5843 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5845 struct rq *rq = cpu_rq(dead_cpu);
5847 /* Must be exiting, otherwise would be on tasklist. */
5848 BUG_ON(!p->exit_state);
5850 /* Cannot have done final schedule yet: would have vanished. */
5851 BUG_ON(p->state == TASK_DEAD);
5856 * Drop lock around migration; if someone else moves it,
5857 * that's OK. No task can be added to this CPU, so iteration is
5860 spin_unlock_irq(&rq->lock);
5861 move_task_off_dead_cpu(dead_cpu, p);
5862 spin_lock_irq(&rq->lock);
5867 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5868 static void migrate_dead_tasks(unsigned int dead_cpu)
5870 struct rq *rq = cpu_rq(dead_cpu);
5871 struct task_struct *next;
5874 if (!rq->nr_running)
5876 update_rq_clock(rq);
5877 next = pick_next_task(rq, rq->curr);
5880 migrate_dead(dead_cpu, next);
5884 #endif /* CONFIG_HOTPLUG_CPU */
5886 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5888 static struct ctl_table sd_ctl_dir[] = {
5890 .procname = "sched_domain",
5896 static struct ctl_table sd_ctl_root[] = {
5898 .ctl_name = CTL_KERN,
5899 .procname = "kernel",
5901 .child = sd_ctl_dir,
5906 static struct ctl_table *sd_alloc_ctl_entry(int n)
5908 struct ctl_table *entry =
5909 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5914 static void sd_free_ctl_entry(struct ctl_table **tablep)
5916 struct ctl_table *entry;
5919 * In the intermediate directories, both the child directory and
5920 * procname are dynamically allocated and could fail but the mode
5921 * will always be set. In the lowest directory the names are
5922 * static strings and all have proc handlers.
5924 for (entry = *tablep; entry->mode; entry++) {
5926 sd_free_ctl_entry(&entry->child);
5927 if (entry->proc_handler == NULL)
5928 kfree(entry->procname);
5936 set_table_entry(struct ctl_table *entry,
5937 const char *procname, void *data, int maxlen,
5938 mode_t mode, proc_handler *proc_handler)
5940 entry->procname = procname;
5942 entry->maxlen = maxlen;
5944 entry->proc_handler = proc_handler;
5947 static struct ctl_table *
5948 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5950 struct ctl_table *table = sd_alloc_ctl_entry(12);
5955 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5956 sizeof(long), 0644, proc_doulongvec_minmax);
5957 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5958 sizeof(long), 0644, proc_doulongvec_minmax);
5959 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5960 sizeof(int), 0644, proc_dointvec_minmax);
5961 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5962 sizeof(int), 0644, proc_dointvec_minmax);
5963 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5964 sizeof(int), 0644, proc_dointvec_minmax);
5965 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5966 sizeof(int), 0644, proc_dointvec_minmax);
5967 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5968 sizeof(int), 0644, proc_dointvec_minmax);
5969 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5970 sizeof(int), 0644, proc_dointvec_minmax);
5971 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5972 sizeof(int), 0644, proc_dointvec_minmax);
5973 set_table_entry(&table[9], "cache_nice_tries",
5974 &sd->cache_nice_tries,
5975 sizeof(int), 0644, proc_dointvec_minmax);
5976 set_table_entry(&table[10], "flags", &sd->flags,
5977 sizeof(int), 0644, proc_dointvec_minmax);
5978 /* &table[11] is terminator */
5983 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5985 struct ctl_table *entry, *table;
5986 struct sched_domain *sd;
5987 int domain_num = 0, i;
5990 for_each_domain(cpu, sd)
5992 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5997 for_each_domain(cpu, sd) {
5998 snprintf(buf, 32, "domain%d", i);
5999 entry->procname = kstrdup(buf, GFP_KERNEL);
6001 entry->child = sd_alloc_ctl_domain_table(sd);
6008 static struct ctl_table_header *sd_sysctl_header;
6009 static void register_sched_domain_sysctl(void)
6011 int i, cpu_num = num_online_cpus();
6012 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6015 WARN_ON(sd_ctl_dir[0].child);
6016 sd_ctl_dir[0].child = entry;
6021 for_each_online_cpu(i) {
6022 snprintf(buf, 32, "cpu%d", i);
6023 entry->procname = kstrdup(buf, GFP_KERNEL);
6025 entry->child = sd_alloc_ctl_cpu_table(i);
6029 WARN_ON(sd_sysctl_header);
6030 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6033 /* may be called multiple times per register */
6034 static void unregister_sched_domain_sysctl(void)
6036 if (sd_sysctl_header)
6037 unregister_sysctl_table(sd_sysctl_header);
6038 sd_sysctl_header = NULL;
6039 if (sd_ctl_dir[0].child)
6040 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6043 static void register_sched_domain_sysctl(void)
6046 static void unregister_sched_domain_sysctl(void)
6051 static void set_rq_online(struct rq *rq)
6054 const struct sched_class *class;
6056 cpu_set(rq->cpu, rq->rd->online);
6059 for_each_class(class) {
6060 if (class->rq_online)
6061 class->rq_online(rq);
6066 static void set_rq_offline(struct rq *rq)
6069 const struct sched_class *class;
6071 for_each_class(class) {
6072 if (class->rq_offline)
6073 class->rq_offline(rq);
6076 cpu_clear(rq->cpu, rq->rd->online);
6082 * migration_call - callback that gets triggered when a CPU is added.
6083 * Here we can start up the necessary migration thread for the new CPU.
6085 static int __cpuinit
6086 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6088 struct task_struct *p;
6089 int cpu = (long)hcpu;
6090 unsigned long flags;
6095 case CPU_UP_PREPARE:
6096 case CPU_UP_PREPARE_FROZEN:
6097 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6100 kthread_bind(p, cpu);
6101 /* Must be high prio: stop_machine expects to yield to it. */
6102 rq = task_rq_lock(p, &flags);
6103 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6104 task_rq_unlock(rq, &flags);
6105 cpu_rq(cpu)->migration_thread = p;
6109 case CPU_ONLINE_FROZEN:
6110 /* Strictly unnecessary, as first user will wake it. */
6111 wake_up_process(cpu_rq(cpu)->migration_thread);
6113 /* Update our root-domain */
6115 spin_lock_irqsave(&rq->lock, flags);
6117 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6121 spin_unlock_irqrestore(&rq->lock, flags);
6124 #ifdef CONFIG_HOTPLUG_CPU
6125 case CPU_UP_CANCELED:
6126 case CPU_UP_CANCELED_FROZEN:
6127 if (!cpu_rq(cpu)->migration_thread)
6129 /* Unbind it from offline cpu so it can run. Fall thru. */
6130 kthread_bind(cpu_rq(cpu)->migration_thread,
6131 any_online_cpu(cpu_online_map));
6132 kthread_stop(cpu_rq(cpu)->migration_thread);
6133 cpu_rq(cpu)->migration_thread = NULL;
6137 case CPU_DEAD_FROZEN:
6138 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6139 migrate_live_tasks(cpu);
6141 kthread_stop(rq->migration_thread);
6142 rq->migration_thread = NULL;
6143 /* Idle task back to normal (off runqueue, low prio) */
6144 spin_lock_irq(&rq->lock);
6145 update_rq_clock(rq);
6146 deactivate_task(rq, rq->idle, 0);
6147 rq->idle->static_prio = MAX_PRIO;
6148 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6149 rq->idle->sched_class = &idle_sched_class;
6150 migrate_dead_tasks(cpu);
6151 spin_unlock_irq(&rq->lock);
6153 migrate_nr_uninterruptible(rq);
6154 BUG_ON(rq->nr_running != 0);
6157 * No need to migrate the tasks: it was best-effort if
6158 * they didn't take sched_hotcpu_mutex. Just wake up
6161 spin_lock_irq(&rq->lock);
6162 while (!list_empty(&rq->migration_queue)) {
6163 struct migration_req *req;
6165 req = list_entry(rq->migration_queue.next,
6166 struct migration_req, list);
6167 list_del_init(&req->list);
6168 complete(&req->done);
6170 spin_unlock_irq(&rq->lock);
6174 case CPU_DYING_FROZEN:
6175 /* Update our root-domain */
6177 spin_lock_irqsave(&rq->lock, flags);
6179 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6182 spin_unlock_irqrestore(&rq->lock, flags);
6189 /* Register at highest priority so that task migration (migrate_all_tasks)
6190 * happens before everything else.
6192 static struct notifier_block __cpuinitdata migration_notifier = {
6193 .notifier_call = migration_call,
6197 void __init migration_init(void)
6199 void *cpu = (void *)(long)smp_processor_id();
6202 /* Start one for the boot CPU: */
6203 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6204 BUG_ON(err == NOTIFY_BAD);
6205 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6206 register_cpu_notifier(&migration_notifier);
6212 #ifdef CONFIG_SCHED_DEBUG
6214 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6227 case SD_LV_ALLNODES:
6236 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6237 cpumask_t *groupmask)
6239 struct sched_group *group = sd->groups;
6242 cpulist_scnprintf(str, sizeof(str), sd->span);
6243 cpus_clear(*groupmask);
6245 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6247 if (!(sd->flags & SD_LOAD_BALANCE)) {
6248 printk("does not load-balance\n");
6250 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6255 printk(KERN_CONT "span %s level %s\n",
6256 str, sd_level_to_string(sd->level));
6258 if (!cpu_isset(cpu, sd->span)) {
6259 printk(KERN_ERR "ERROR: domain->span does not contain "
6262 if (!cpu_isset(cpu, group->cpumask)) {
6263 printk(KERN_ERR "ERROR: domain->groups does not contain"
6267 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6271 printk(KERN_ERR "ERROR: group is NULL\n");
6275 if (!group->__cpu_power) {
6276 printk(KERN_CONT "\n");
6277 printk(KERN_ERR "ERROR: domain->cpu_power not "
6282 if (!cpus_weight(group->cpumask)) {
6283 printk(KERN_CONT "\n");
6284 printk(KERN_ERR "ERROR: empty group\n");
6288 if (cpus_intersects(*groupmask, group->cpumask)) {
6289 printk(KERN_CONT "\n");
6290 printk(KERN_ERR "ERROR: repeated CPUs\n");
6294 cpus_or(*groupmask, *groupmask, group->cpumask);
6296 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6297 printk(KERN_CONT " %s", str);
6299 group = group->next;
6300 } while (group != sd->groups);
6301 printk(KERN_CONT "\n");
6303 if (!cpus_equal(sd->span, *groupmask))
6304 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6306 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6307 printk(KERN_ERR "ERROR: parent span is not a superset "
6308 "of domain->span\n");
6312 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6314 cpumask_t *groupmask;
6318 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6322 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6324 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6326 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6331 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6340 #else /* !CONFIG_SCHED_DEBUG */
6341 # define sched_domain_debug(sd, cpu) do { } while (0)
6342 #endif /* CONFIG_SCHED_DEBUG */
6344 static int sd_degenerate(struct sched_domain *sd)
6346 if (cpus_weight(sd->span) == 1)
6349 /* Following flags need at least 2 groups */
6350 if (sd->flags & (SD_LOAD_BALANCE |
6351 SD_BALANCE_NEWIDLE |
6355 SD_SHARE_PKG_RESOURCES)) {
6356 if (sd->groups != sd->groups->next)
6360 /* Following flags don't use groups */
6361 if (sd->flags & (SD_WAKE_IDLE |
6370 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6372 unsigned long cflags = sd->flags, pflags = parent->flags;
6374 if (sd_degenerate(parent))
6377 if (!cpus_equal(sd->span, parent->span))
6380 /* Does parent contain flags not in child? */
6381 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6382 if (cflags & SD_WAKE_AFFINE)
6383 pflags &= ~SD_WAKE_BALANCE;
6384 /* Flags needing groups don't count if only 1 group in parent */
6385 if (parent->groups == parent->groups->next) {
6386 pflags &= ~(SD_LOAD_BALANCE |
6387 SD_BALANCE_NEWIDLE |
6391 SD_SHARE_PKG_RESOURCES);
6393 if (~cflags & pflags)
6399 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6401 unsigned long flags;
6403 spin_lock_irqsave(&rq->lock, flags);
6406 struct root_domain *old_rd = rq->rd;
6408 if (cpu_isset(rq->cpu, old_rd->online))
6411 cpu_clear(rq->cpu, old_rd->span);
6413 if (atomic_dec_and_test(&old_rd->refcount))
6417 atomic_inc(&rd->refcount);
6420 cpu_set(rq->cpu, rd->span);
6421 if (cpu_isset(rq->cpu, cpu_online_map))
6424 spin_unlock_irqrestore(&rq->lock, flags);
6427 static void init_rootdomain(struct root_domain *rd)
6429 memset(rd, 0, sizeof(*rd));
6431 cpus_clear(rd->span);
6432 cpus_clear(rd->online);
6434 cpupri_init(&rd->cpupri);
6437 static void init_defrootdomain(void)
6439 init_rootdomain(&def_root_domain);
6440 atomic_set(&def_root_domain.refcount, 1);
6443 static struct root_domain *alloc_rootdomain(void)
6445 struct root_domain *rd;
6447 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6451 init_rootdomain(rd);
6457 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6458 * hold the hotplug lock.
6461 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6463 struct rq *rq = cpu_rq(cpu);
6464 struct sched_domain *tmp;
6466 /* Remove the sched domains which do not contribute to scheduling. */
6467 for (tmp = sd; tmp; tmp = tmp->parent) {
6468 struct sched_domain *parent = tmp->parent;
6471 if (sd_parent_degenerate(tmp, parent)) {
6472 tmp->parent = parent->parent;
6474 parent->parent->child = tmp;
6478 if (sd && sd_degenerate(sd)) {
6484 sched_domain_debug(sd, cpu);
6486 rq_attach_root(rq, rd);
6487 rcu_assign_pointer(rq->sd, sd);
6490 /* cpus with isolated domains */
6491 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6493 /* Setup the mask of cpus configured for isolated domains */
6494 static int __init isolated_cpu_setup(char *str)
6496 int ints[NR_CPUS], i;
6498 str = get_options(str, ARRAY_SIZE(ints), ints);
6499 cpus_clear(cpu_isolated_map);
6500 for (i = 1; i <= ints[0]; i++)
6501 if (ints[i] < NR_CPUS)
6502 cpu_set(ints[i], cpu_isolated_map);
6506 __setup("isolcpus=", isolated_cpu_setup);
6509 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6510 * to a function which identifies what group(along with sched group) a CPU
6511 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6512 * (due to the fact that we keep track of groups covered with a cpumask_t).
6514 * init_sched_build_groups will build a circular linked list of the groups
6515 * covered by the given span, and will set each group's ->cpumask correctly,
6516 * and ->cpu_power to 0.
6519 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6520 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6521 struct sched_group **sg,
6522 cpumask_t *tmpmask),
6523 cpumask_t *covered, cpumask_t *tmpmask)
6525 struct sched_group *first = NULL, *last = NULL;
6528 cpus_clear(*covered);
6530 for_each_cpu_mask(i, *span) {
6531 struct sched_group *sg;
6532 int group = group_fn(i, cpu_map, &sg, tmpmask);
6535 if (cpu_isset(i, *covered))
6538 cpus_clear(sg->cpumask);
6539 sg->__cpu_power = 0;
6541 for_each_cpu_mask(j, *span) {
6542 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6545 cpu_set(j, *covered);
6546 cpu_set(j, sg->cpumask);
6557 #define SD_NODES_PER_DOMAIN 16
6562 * find_next_best_node - find the next node to include in a sched_domain
6563 * @node: node whose sched_domain we're building
6564 * @used_nodes: nodes already in the sched_domain
6566 * Find the next node to include in a given scheduling domain. Simply
6567 * finds the closest node not already in the @used_nodes map.
6569 * Should use nodemask_t.
6571 static int find_next_best_node(int node, nodemask_t *used_nodes)
6573 int i, n, val, min_val, best_node = 0;
6577 for (i = 0; i < MAX_NUMNODES; i++) {
6578 /* Start at @node */
6579 n = (node + i) % MAX_NUMNODES;
6581 if (!nr_cpus_node(n))
6584 /* Skip already used nodes */
6585 if (node_isset(n, *used_nodes))
6588 /* Simple min distance search */
6589 val = node_distance(node, n);
6591 if (val < min_val) {
6597 node_set(best_node, *used_nodes);
6602 * sched_domain_node_span - get a cpumask for a node's sched_domain
6603 * @node: node whose cpumask we're constructing
6604 * @span: resulting cpumask
6606 * Given a node, construct a good cpumask for its sched_domain to span. It
6607 * should be one that prevents unnecessary balancing, but also spreads tasks
6610 static void sched_domain_node_span(int node, cpumask_t *span)
6612 nodemask_t used_nodes;
6613 node_to_cpumask_ptr(nodemask, node);
6617 nodes_clear(used_nodes);
6619 cpus_or(*span, *span, *nodemask);
6620 node_set(node, used_nodes);
6622 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6623 int next_node = find_next_best_node(node, &used_nodes);
6625 node_to_cpumask_ptr_next(nodemask, next_node);
6626 cpus_or(*span, *span, *nodemask);
6629 #endif /* CONFIG_NUMA */
6631 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6634 * SMT sched-domains:
6636 #ifdef CONFIG_SCHED_SMT
6637 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6638 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6641 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6645 *sg = &per_cpu(sched_group_cpus, cpu);
6648 #endif /* CONFIG_SCHED_SMT */
6651 * multi-core sched-domains:
6653 #ifdef CONFIG_SCHED_MC
6654 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6655 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6656 #endif /* CONFIG_SCHED_MC */
6658 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6660 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6665 *mask = per_cpu(cpu_sibling_map, cpu);
6666 cpus_and(*mask, *mask, *cpu_map);
6667 group = first_cpu(*mask);
6669 *sg = &per_cpu(sched_group_core, group);
6672 #elif defined(CONFIG_SCHED_MC)
6674 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6678 *sg = &per_cpu(sched_group_core, cpu);
6683 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6684 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6687 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6691 #ifdef CONFIG_SCHED_MC
6692 *mask = cpu_coregroup_map(cpu);
6693 cpus_and(*mask, *mask, *cpu_map);
6694 group = first_cpu(*mask);
6695 #elif defined(CONFIG_SCHED_SMT)
6696 *mask = per_cpu(cpu_sibling_map, cpu);
6697 cpus_and(*mask, *mask, *cpu_map);
6698 group = first_cpu(*mask);
6703 *sg = &per_cpu(sched_group_phys, group);
6709 * The init_sched_build_groups can't handle what we want to do with node
6710 * groups, so roll our own. Now each node has its own list of groups which
6711 * gets dynamically allocated.
6713 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6714 static struct sched_group ***sched_group_nodes_bycpu;
6716 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6717 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6719 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6720 struct sched_group **sg, cpumask_t *nodemask)
6724 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6725 cpus_and(*nodemask, *nodemask, *cpu_map);
6726 group = first_cpu(*nodemask);
6729 *sg = &per_cpu(sched_group_allnodes, group);
6733 static void init_numa_sched_groups_power(struct sched_group *group_head)
6735 struct sched_group *sg = group_head;
6741 for_each_cpu_mask(j, sg->cpumask) {
6742 struct sched_domain *sd;
6744 sd = &per_cpu(phys_domains, j);
6745 if (j != first_cpu(sd->groups->cpumask)) {
6747 * Only add "power" once for each
6753 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6756 } while (sg != group_head);
6758 #endif /* CONFIG_NUMA */
6761 /* Free memory allocated for various sched_group structures */
6762 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6766 for_each_cpu_mask(cpu, *cpu_map) {
6767 struct sched_group **sched_group_nodes
6768 = sched_group_nodes_bycpu[cpu];
6770 if (!sched_group_nodes)
6773 for (i = 0; i < MAX_NUMNODES; i++) {
6774 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6776 *nodemask = node_to_cpumask(i);
6777 cpus_and(*nodemask, *nodemask, *cpu_map);
6778 if (cpus_empty(*nodemask))
6788 if (oldsg != sched_group_nodes[i])
6791 kfree(sched_group_nodes);
6792 sched_group_nodes_bycpu[cpu] = NULL;
6795 #else /* !CONFIG_NUMA */
6796 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6799 #endif /* CONFIG_NUMA */
6802 * Initialize sched groups cpu_power.
6804 * cpu_power indicates the capacity of sched group, which is used while
6805 * distributing the load between different sched groups in a sched domain.
6806 * Typically cpu_power for all the groups in a sched domain will be same unless
6807 * there are asymmetries in the topology. If there are asymmetries, group
6808 * having more cpu_power will pickup more load compared to the group having
6811 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6812 * the maximum number of tasks a group can handle in the presence of other idle
6813 * or lightly loaded groups in the same sched domain.
6815 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6817 struct sched_domain *child;
6818 struct sched_group *group;
6820 WARN_ON(!sd || !sd->groups);
6822 if (cpu != first_cpu(sd->groups->cpumask))
6827 sd->groups->__cpu_power = 0;
6830 * For perf policy, if the groups in child domain share resources
6831 * (for example cores sharing some portions of the cache hierarchy
6832 * or SMT), then set this domain groups cpu_power such that each group
6833 * can handle only one task, when there are other idle groups in the
6834 * same sched domain.
6836 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6838 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6839 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6844 * add cpu_power of each child group to this groups cpu_power
6846 group = child->groups;
6848 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6849 group = group->next;
6850 } while (group != child->groups);
6854 * Initializers for schedule domains
6855 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6858 #define SD_INIT(sd, type) sd_init_##type(sd)
6859 #define SD_INIT_FUNC(type) \
6860 static noinline void sd_init_##type(struct sched_domain *sd) \
6862 memset(sd, 0, sizeof(*sd)); \
6863 *sd = SD_##type##_INIT; \
6864 sd->level = SD_LV_##type; \
6869 SD_INIT_FUNC(ALLNODES)
6872 #ifdef CONFIG_SCHED_SMT
6873 SD_INIT_FUNC(SIBLING)
6875 #ifdef CONFIG_SCHED_MC
6880 * To minimize stack usage kmalloc room for cpumasks and share the
6881 * space as the usage in build_sched_domains() dictates. Used only
6882 * if the amount of space is significant.
6885 cpumask_t tmpmask; /* make this one first */
6888 cpumask_t this_sibling_map;
6889 cpumask_t this_core_map;
6891 cpumask_t send_covered;
6894 cpumask_t domainspan;
6896 cpumask_t notcovered;
6901 #define SCHED_CPUMASK_ALLOC 1
6902 #define SCHED_CPUMASK_FREE(v) kfree(v)
6903 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6905 #define SCHED_CPUMASK_ALLOC 0
6906 #define SCHED_CPUMASK_FREE(v)
6907 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6910 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6911 ((unsigned long)(a) + offsetof(struct allmasks, v))
6913 static int default_relax_domain_level = -1;
6915 static int __init setup_relax_domain_level(char *str)
6919 val = simple_strtoul(str, NULL, 0);
6920 if (val < SD_LV_MAX)
6921 default_relax_domain_level = val;
6925 __setup("relax_domain_level=", setup_relax_domain_level);
6927 static void set_domain_attribute(struct sched_domain *sd,
6928 struct sched_domain_attr *attr)
6932 if (!attr || attr->relax_domain_level < 0) {
6933 if (default_relax_domain_level < 0)
6936 request = default_relax_domain_level;
6938 request = attr->relax_domain_level;
6939 if (request < sd->level) {
6940 /* turn off idle balance on this domain */
6941 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
6943 /* turn on idle balance on this domain */
6944 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
6949 * Build sched domains for a given set of cpus and attach the sched domains
6950 * to the individual cpus
6952 static int __build_sched_domains(const cpumask_t *cpu_map,
6953 struct sched_domain_attr *attr)
6956 struct root_domain *rd;
6957 SCHED_CPUMASK_DECLARE(allmasks);
6960 struct sched_group **sched_group_nodes = NULL;
6961 int sd_allnodes = 0;
6964 * Allocate the per-node list of sched groups
6966 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
6968 if (!sched_group_nodes) {
6969 printk(KERN_WARNING "Can not alloc sched group node list\n");
6974 rd = alloc_rootdomain();
6976 printk(KERN_WARNING "Cannot alloc root domain\n");
6978 kfree(sched_group_nodes);
6983 #if SCHED_CPUMASK_ALLOC
6984 /* get space for all scratch cpumask variables */
6985 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
6987 printk(KERN_WARNING "Cannot alloc cpumask array\n");
6990 kfree(sched_group_nodes);
6995 tmpmask = (cpumask_t *)allmasks;
6999 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7003 * Set up domains for cpus specified by the cpu_map.
7005 for_each_cpu_mask(i, *cpu_map) {
7006 struct sched_domain *sd = NULL, *p;
7007 SCHED_CPUMASK_VAR(nodemask, allmasks);
7009 *nodemask = node_to_cpumask(cpu_to_node(i));
7010 cpus_and(*nodemask, *nodemask, *cpu_map);
7013 if (cpus_weight(*cpu_map) >
7014 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7015 sd = &per_cpu(allnodes_domains, i);
7016 SD_INIT(sd, ALLNODES);
7017 set_domain_attribute(sd, attr);
7018 sd->span = *cpu_map;
7019 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7025 sd = &per_cpu(node_domains, i);
7027 set_domain_attribute(sd, attr);
7028 sched_domain_node_span(cpu_to_node(i), &sd->span);
7032 cpus_and(sd->span, sd->span, *cpu_map);
7036 sd = &per_cpu(phys_domains, i);
7038 set_domain_attribute(sd, attr);
7039 sd->span = *nodemask;
7043 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7045 #ifdef CONFIG_SCHED_MC
7047 sd = &per_cpu(core_domains, i);
7049 set_domain_attribute(sd, attr);
7050 sd->span = cpu_coregroup_map(i);
7051 cpus_and(sd->span, sd->span, *cpu_map);
7054 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7057 #ifdef CONFIG_SCHED_SMT
7059 sd = &per_cpu(cpu_domains, i);
7060 SD_INIT(sd, SIBLING);
7061 set_domain_attribute(sd, attr);
7062 sd->span = per_cpu(cpu_sibling_map, i);
7063 cpus_and(sd->span, sd->span, *cpu_map);
7066 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7070 #ifdef CONFIG_SCHED_SMT
7071 /* Set up CPU (sibling) groups */
7072 for_each_cpu_mask(i, *cpu_map) {
7073 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7074 SCHED_CPUMASK_VAR(send_covered, allmasks);
7076 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7077 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7078 if (i != first_cpu(*this_sibling_map))
7081 init_sched_build_groups(this_sibling_map, cpu_map,
7083 send_covered, tmpmask);
7087 #ifdef CONFIG_SCHED_MC
7088 /* Set up multi-core groups */
7089 for_each_cpu_mask(i, *cpu_map) {
7090 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7091 SCHED_CPUMASK_VAR(send_covered, allmasks);
7093 *this_core_map = cpu_coregroup_map(i);
7094 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7095 if (i != first_cpu(*this_core_map))
7098 init_sched_build_groups(this_core_map, cpu_map,
7100 send_covered, tmpmask);
7104 /* Set up physical groups */
7105 for (i = 0; i < MAX_NUMNODES; i++) {
7106 SCHED_CPUMASK_VAR(nodemask, allmasks);
7107 SCHED_CPUMASK_VAR(send_covered, allmasks);
7109 *nodemask = node_to_cpumask(i);
7110 cpus_and(*nodemask, *nodemask, *cpu_map);
7111 if (cpus_empty(*nodemask))
7114 init_sched_build_groups(nodemask, cpu_map,
7116 send_covered, tmpmask);
7120 /* Set up node groups */
7122 SCHED_CPUMASK_VAR(send_covered, allmasks);
7124 init_sched_build_groups(cpu_map, cpu_map,
7125 &cpu_to_allnodes_group,
7126 send_covered, tmpmask);
7129 for (i = 0; i < MAX_NUMNODES; i++) {
7130 /* Set up node groups */
7131 struct sched_group *sg, *prev;
7132 SCHED_CPUMASK_VAR(nodemask, allmasks);
7133 SCHED_CPUMASK_VAR(domainspan, allmasks);
7134 SCHED_CPUMASK_VAR(covered, allmasks);
7137 *nodemask = node_to_cpumask(i);
7138 cpus_clear(*covered);
7140 cpus_and(*nodemask, *nodemask, *cpu_map);
7141 if (cpus_empty(*nodemask)) {
7142 sched_group_nodes[i] = NULL;
7146 sched_domain_node_span(i, domainspan);
7147 cpus_and(*domainspan, *domainspan, *cpu_map);
7149 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7151 printk(KERN_WARNING "Can not alloc domain group for "
7155 sched_group_nodes[i] = sg;
7156 for_each_cpu_mask(j, *nodemask) {
7157 struct sched_domain *sd;
7159 sd = &per_cpu(node_domains, j);
7162 sg->__cpu_power = 0;
7163 sg->cpumask = *nodemask;
7165 cpus_or(*covered, *covered, *nodemask);
7168 for (j = 0; j < MAX_NUMNODES; j++) {
7169 SCHED_CPUMASK_VAR(notcovered, allmasks);
7170 int n = (i + j) % MAX_NUMNODES;
7171 node_to_cpumask_ptr(pnodemask, n);
7173 cpus_complement(*notcovered, *covered);
7174 cpus_and(*tmpmask, *notcovered, *cpu_map);
7175 cpus_and(*tmpmask, *tmpmask, *domainspan);
7176 if (cpus_empty(*tmpmask))
7179 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7180 if (cpus_empty(*tmpmask))
7183 sg = kmalloc_node(sizeof(struct sched_group),
7187 "Can not alloc domain group for node %d\n", j);
7190 sg->__cpu_power = 0;
7191 sg->cpumask = *tmpmask;
7192 sg->next = prev->next;
7193 cpus_or(*covered, *covered, *tmpmask);
7200 /* Calculate CPU power for physical packages and nodes */
7201 #ifdef CONFIG_SCHED_SMT
7202 for_each_cpu_mask(i, *cpu_map) {
7203 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7205 init_sched_groups_power(i, sd);
7208 #ifdef CONFIG_SCHED_MC
7209 for_each_cpu_mask(i, *cpu_map) {
7210 struct sched_domain *sd = &per_cpu(core_domains, i);
7212 init_sched_groups_power(i, sd);
7216 for_each_cpu_mask(i, *cpu_map) {
7217 struct sched_domain *sd = &per_cpu(phys_domains, i);
7219 init_sched_groups_power(i, sd);
7223 for (i = 0; i < MAX_NUMNODES; i++)
7224 init_numa_sched_groups_power(sched_group_nodes[i]);
7227 struct sched_group *sg;
7229 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7231 init_numa_sched_groups_power(sg);
7235 /* Attach the domains */
7236 for_each_cpu_mask(i, *cpu_map) {
7237 struct sched_domain *sd;
7238 #ifdef CONFIG_SCHED_SMT
7239 sd = &per_cpu(cpu_domains, i);
7240 #elif defined(CONFIG_SCHED_MC)
7241 sd = &per_cpu(core_domains, i);
7243 sd = &per_cpu(phys_domains, i);
7245 cpu_attach_domain(sd, rd, i);
7248 SCHED_CPUMASK_FREE((void *)allmasks);
7253 free_sched_groups(cpu_map, tmpmask);
7254 SCHED_CPUMASK_FREE((void *)allmasks);
7259 static int build_sched_domains(const cpumask_t *cpu_map)
7261 return __build_sched_domains(cpu_map, NULL);
7264 static cpumask_t *doms_cur; /* current sched domains */
7265 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7266 static struct sched_domain_attr *dattr_cur;
7267 /* attribues of custom domains in 'doms_cur' */
7270 * Special case: If a kmalloc of a doms_cur partition (array of
7271 * cpumask_t) fails, then fallback to a single sched domain,
7272 * as determined by the single cpumask_t fallback_doms.
7274 static cpumask_t fallback_doms;
7276 void __attribute__((weak)) arch_update_cpu_topology(void)
7281 * Free current domain masks.
7282 * Called after all cpus are attached to NULL domain.
7284 static void free_sched_domains(void)
7287 if (doms_cur != &fallback_doms)
7289 doms_cur = &fallback_doms;
7293 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7294 * For now this just excludes isolated cpus, but could be used to
7295 * exclude other special cases in the future.
7297 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7301 arch_update_cpu_topology();
7303 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7305 doms_cur = &fallback_doms;
7306 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7308 err = build_sched_domains(doms_cur);
7309 register_sched_domain_sysctl();
7314 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7317 free_sched_groups(cpu_map, tmpmask);
7321 * Detach sched domains from a group of cpus specified in cpu_map
7322 * These cpus will now be attached to the NULL domain
7324 static void detach_destroy_domains(const cpumask_t *cpu_map)
7329 unregister_sched_domain_sysctl();
7331 for_each_cpu_mask(i, *cpu_map)
7332 cpu_attach_domain(NULL, &def_root_domain, i);
7333 synchronize_sched();
7334 arch_destroy_sched_domains(cpu_map, &tmpmask);
7337 /* handle null as "default" */
7338 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7339 struct sched_domain_attr *new, int idx_new)
7341 struct sched_domain_attr tmp;
7348 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7349 new ? (new + idx_new) : &tmp,
7350 sizeof(struct sched_domain_attr));
7354 * Partition sched domains as specified by the 'ndoms_new'
7355 * cpumasks in the array doms_new[] of cpumasks. This compares
7356 * doms_new[] to the current sched domain partitioning, doms_cur[].
7357 * It destroys each deleted domain and builds each new domain.
7359 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7360 * The masks don't intersect (don't overlap.) We should setup one
7361 * sched domain for each mask. CPUs not in any of the cpumasks will
7362 * not be load balanced. If the same cpumask appears both in the
7363 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7366 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7367 * ownership of it and will kfree it when done with it. If the caller
7368 * failed the kmalloc call, then it can pass in doms_new == NULL,
7369 * and partition_sched_domains() will fallback to the single partition
7372 * Call with hotplug lock held
7374 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7375 struct sched_domain_attr *dattr_new)
7379 mutex_lock(&sched_domains_mutex);
7381 /* always unregister in case we don't destroy any domains */
7382 unregister_sched_domain_sysctl();
7384 if (doms_new == NULL) {
7386 doms_new = &fallback_doms;
7387 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7391 /* Destroy deleted domains */
7392 for (i = 0; i < ndoms_cur; i++) {
7393 for (j = 0; j < ndoms_new; j++) {
7394 if (cpus_equal(doms_cur[i], doms_new[j])
7395 && dattrs_equal(dattr_cur, i, dattr_new, j))
7398 /* no match - a current sched domain not in new doms_new[] */
7399 detach_destroy_domains(doms_cur + i);
7404 /* Build new domains */
7405 for (i = 0; i < ndoms_new; i++) {
7406 for (j = 0; j < ndoms_cur; j++) {
7407 if (cpus_equal(doms_new[i], doms_cur[j])
7408 && dattrs_equal(dattr_new, i, dattr_cur, j))
7411 /* no match - add a new doms_new */
7412 __build_sched_domains(doms_new + i,
7413 dattr_new ? dattr_new + i : NULL);
7418 /* Remember the new sched domains */
7419 if (doms_cur != &fallback_doms)
7421 kfree(dattr_cur); /* kfree(NULL) is safe */
7422 doms_cur = doms_new;
7423 dattr_cur = dattr_new;
7424 ndoms_cur = ndoms_new;
7426 register_sched_domain_sysctl();
7428 mutex_unlock(&sched_domains_mutex);
7431 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7432 int arch_reinit_sched_domains(void)
7437 mutex_lock(&sched_domains_mutex);
7438 detach_destroy_domains(&cpu_online_map);
7439 free_sched_domains();
7440 err = arch_init_sched_domains(&cpu_online_map);
7441 mutex_unlock(&sched_domains_mutex);
7447 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7451 if (buf[0] != '0' && buf[0] != '1')
7455 sched_smt_power_savings = (buf[0] == '1');
7457 sched_mc_power_savings = (buf[0] == '1');
7459 ret = arch_reinit_sched_domains();
7461 return ret ? ret : count;
7464 #ifdef CONFIG_SCHED_MC
7465 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7467 return sprintf(page, "%u\n", sched_mc_power_savings);
7469 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7470 const char *buf, size_t count)
7472 return sched_power_savings_store(buf, count, 0);
7474 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7475 sched_mc_power_savings_store);
7478 #ifdef CONFIG_SCHED_SMT
7479 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7481 return sprintf(page, "%u\n", sched_smt_power_savings);
7483 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7484 const char *buf, size_t count)
7486 return sched_power_savings_store(buf, count, 1);
7488 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7489 sched_smt_power_savings_store);
7492 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7496 #ifdef CONFIG_SCHED_SMT
7498 err = sysfs_create_file(&cls->kset.kobj,
7499 &attr_sched_smt_power_savings.attr);
7501 #ifdef CONFIG_SCHED_MC
7502 if (!err && mc_capable())
7503 err = sysfs_create_file(&cls->kset.kobj,
7504 &attr_sched_mc_power_savings.attr);
7508 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7511 * Force a reinitialization of the sched domains hierarchy. The domains
7512 * and groups cannot be updated in place without racing with the balancing
7513 * code, so we temporarily attach all running cpus to the NULL domain
7514 * which will prevent rebalancing while the sched domains are recalculated.
7516 static int update_sched_domains(struct notifier_block *nfb,
7517 unsigned long action, void *hcpu)
7519 int cpu = (int)(long)hcpu;
7522 case CPU_DOWN_PREPARE:
7523 case CPU_DOWN_PREPARE_FROZEN:
7524 disable_runtime(cpu_rq(cpu));
7526 case CPU_UP_PREPARE:
7527 case CPU_UP_PREPARE_FROZEN:
7528 detach_destroy_domains(&cpu_online_map);
7529 free_sched_domains();
7533 case CPU_DOWN_FAILED:
7534 case CPU_DOWN_FAILED_FROZEN:
7536 case CPU_ONLINE_FROZEN:
7537 enable_runtime(cpu_rq(cpu));
7539 case CPU_UP_CANCELED:
7540 case CPU_UP_CANCELED_FROZEN:
7542 case CPU_DEAD_FROZEN:
7544 * Fall through and re-initialise the domains.
7551 #ifndef CONFIG_CPUSETS
7553 * Create default domain partitioning if cpusets are disabled.
7554 * Otherwise we let cpusets rebuild the domains based on the
7558 /* The hotplug lock is already held by cpu_up/cpu_down */
7559 arch_init_sched_domains(&cpu_online_map);
7565 void __init sched_init_smp(void)
7567 cpumask_t non_isolated_cpus;
7569 #if defined(CONFIG_NUMA)
7570 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7572 BUG_ON(sched_group_nodes_bycpu == NULL);
7575 mutex_lock(&sched_domains_mutex);
7576 arch_init_sched_domains(&cpu_online_map);
7577 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7578 if (cpus_empty(non_isolated_cpus))
7579 cpu_set(smp_processor_id(), non_isolated_cpus);
7580 mutex_unlock(&sched_domains_mutex);
7582 /* XXX: Theoretical race here - CPU may be hotplugged now */
7583 hotcpu_notifier(update_sched_domains, 0);
7586 /* Move init over to a non-isolated CPU */
7587 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7589 sched_init_granularity();
7592 void __init sched_init_smp(void)
7594 sched_init_granularity();
7596 #endif /* CONFIG_SMP */
7598 int in_sched_functions(unsigned long addr)
7600 return in_lock_functions(addr) ||
7601 (addr >= (unsigned long)__sched_text_start
7602 && addr < (unsigned long)__sched_text_end);
7605 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7607 cfs_rq->tasks_timeline = RB_ROOT;
7608 INIT_LIST_HEAD(&cfs_rq->tasks);
7609 #ifdef CONFIG_FAIR_GROUP_SCHED
7612 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7615 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7617 struct rt_prio_array *array;
7620 array = &rt_rq->active;
7621 for (i = 0; i < MAX_RT_PRIO; i++) {
7622 INIT_LIST_HEAD(array->queue + i);
7623 __clear_bit(i, array->bitmap);
7625 /* delimiter for bitsearch: */
7626 __set_bit(MAX_RT_PRIO, array->bitmap);
7628 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7629 rt_rq->highest_prio = MAX_RT_PRIO;
7632 rt_rq->rt_nr_migratory = 0;
7633 rt_rq->overloaded = 0;
7637 rt_rq->rt_throttled = 0;
7638 rt_rq->rt_runtime = 0;
7639 spin_lock_init(&rt_rq->rt_runtime_lock);
7641 #ifdef CONFIG_RT_GROUP_SCHED
7642 rt_rq->rt_nr_boosted = 0;
7647 #ifdef CONFIG_FAIR_GROUP_SCHED
7648 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7649 struct sched_entity *se, int cpu, int add,
7650 struct sched_entity *parent)
7652 struct rq *rq = cpu_rq(cpu);
7653 tg->cfs_rq[cpu] = cfs_rq;
7654 init_cfs_rq(cfs_rq, rq);
7657 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7660 /* se could be NULL for init_task_group */
7665 se->cfs_rq = &rq->cfs;
7667 se->cfs_rq = parent->my_q;
7670 se->load.weight = tg->shares;
7671 se->load.inv_weight = 0;
7672 se->parent = parent;
7676 #ifdef CONFIG_RT_GROUP_SCHED
7677 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7678 struct sched_rt_entity *rt_se, int cpu, int add,
7679 struct sched_rt_entity *parent)
7681 struct rq *rq = cpu_rq(cpu);
7683 tg->rt_rq[cpu] = rt_rq;
7684 init_rt_rq(rt_rq, rq);
7686 rt_rq->rt_se = rt_se;
7687 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7689 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7691 tg->rt_se[cpu] = rt_se;
7696 rt_se->rt_rq = &rq->rt;
7698 rt_se->rt_rq = parent->my_q;
7700 rt_se->my_q = rt_rq;
7701 rt_se->parent = parent;
7702 INIT_LIST_HEAD(&rt_se->run_list);
7706 void __init sched_init(void)
7709 unsigned long alloc_size = 0, ptr;
7711 #ifdef CONFIG_FAIR_GROUP_SCHED
7712 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7714 #ifdef CONFIG_RT_GROUP_SCHED
7715 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7717 #ifdef CONFIG_USER_SCHED
7721 * As sched_init() is called before page_alloc is setup,
7722 * we use alloc_bootmem().
7725 ptr = (unsigned long)alloc_bootmem(alloc_size);
7727 #ifdef CONFIG_FAIR_GROUP_SCHED
7728 init_task_group.se = (struct sched_entity **)ptr;
7729 ptr += nr_cpu_ids * sizeof(void **);
7731 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7732 ptr += nr_cpu_ids * sizeof(void **);
7734 #ifdef CONFIG_USER_SCHED
7735 root_task_group.se = (struct sched_entity **)ptr;
7736 ptr += nr_cpu_ids * sizeof(void **);
7738 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7739 ptr += nr_cpu_ids * sizeof(void **);
7740 #endif /* CONFIG_USER_SCHED */
7741 #endif /* CONFIG_FAIR_GROUP_SCHED */
7742 #ifdef CONFIG_RT_GROUP_SCHED
7743 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7744 ptr += nr_cpu_ids * sizeof(void **);
7746 init_task_group.rt_rq = (struct rt_rq **)ptr;
7747 ptr += nr_cpu_ids * sizeof(void **);
7749 #ifdef CONFIG_USER_SCHED
7750 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7751 ptr += nr_cpu_ids * sizeof(void **);
7753 root_task_group.rt_rq = (struct rt_rq **)ptr;
7754 ptr += nr_cpu_ids * sizeof(void **);
7755 #endif /* CONFIG_USER_SCHED */
7756 #endif /* CONFIG_RT_GROUP_SCHED */
7760 init_defrootdomain();
7763 init_rt_bandwidth(&def_rt_bandwidth,
7764 global_rt_period(), global_rt_runtime());
7766 #ifdef CONFIG_RT_GROUP_SCHED
7767 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7768 global_rt_period(), global_rt_runtime());
7769 #ifdef CONFIG_USER_SCHED
7770 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7771 global_rt_period(), RUNTIME_INF);
7772 #endif /* CONFIG_USER_SCHED */
7773 #endif /* CONFIG_RT_GROUP_SCHED */
7775 #ifdef CONFIG_GROUP_SCHED
7776 list_add(&init_task_group.list, &task_groups);
7777 INIT_LIST_HEAD(&init_task_group.children);
7779 #ifdef CONFIG_USER_SCHED
7780 INIT_LIST_HEAD(&root_task_group.children);
7781 init_task_group.parent = &root_task_group;
7782 list_add(&init_task_group.siblings, &root_task_group.children);
7783 #endif /* CONFIG_USER_SCHED */
7784 #endif /* CONFIG_GROUP_SCHED */
7786 for_each_possible_cpu(i) {
7790 spin_lock_init(&rq->lock);
7791 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7793 init_cfs_rq(&rq->cfs, rq);
7794 init_rt_rq(&rq->rt, rq);
7795 #ifdef CONFIG_FAIR_GROUP_SCHED
7796 init_task_group.shares = init_task_group_load;
7797 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7798 #ifdef CONFIG_CGROUP_SCHED
7800 * How much cpu bandwidth does init_task_group get?
7802 * In case of task-groups formed thr' the cgroup filesystem, it
7803 * gets 100% of the cpu resources in the system. This overall
7804 * system cpu resource is divided among the tasks of
7805 * init_task_group and its child task-groups in a fair manner,
7806 * based on each entity's (task or task-group's) weight
7807 * (se->load.weight).
7809 * In other words, if init_task_group has 10 tasks of weight
7810 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7811 * then A0's share of the cpu resource is:
7813 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7815 * We achieve this by letting init_task_group's tasks sit
7816 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7818 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7819 #elif defined CONFIG_USER_SCHED
7820 root_task_group.shares = NICE_0_LOAD;
7821 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
7823 * In case of task-groups formed thr' the user id of tasks,
7824 * init_task_group represents tasks belonging to root user.
7825 * Hence it forms a sibling of all subsequent groups formed.
7826 * In this case, init_task_group gets only a fraction of overall
7827 * system cpu resource, based on the weight assigned to root
7828 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7829 * by letting tasks of init_task_group sit in a separate cfs_rq
7830 * (init_cfs_rq) and having one entity represent this group of
7831 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7833 init_tg_cfs_entry(&init_task_group,
7834 &per_cpu(init_cfs_rq, i),
7835 &per_cpu(init_sched_entity, i), i, 1,
7836 root_task_group.se[i]);
7839 #endif /* CONFIG_FAIR_GROUP_SCHED */
7841 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7842 #ifdef CONFIG_RT_GROUP_SCHED
7843 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7844 #ifdef CONFIG_CGROUP_SCHED
7845 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7846 #elif defined CONFIG_USER_SCHED
7847 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
7848 init_tg_rt_entry(&init_task_group,
7849 &per_cpu(init_rt_rq, i),
7850 &per_cpu(init_sched_rt_entity, i), i, 1,
7851 root_task_group.rt_se[i]);
7855 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7856 rq->cpu_load[j] = 0;
7860 rq->active_balance = 0;
7861 rq->next_balance = jiffies;
7865 rq->migration_thread = NULL;
7866 INIT_LIST_HEAD(&rq->migration_queue);
7867 rq_attach_root(rq, &def_root_domain);
7870 atomic_set(&rq->nr_iowait, 0);
7873 set_load_weight(&init_task);
7875 #ifdef CONFIG_PREEMPT_NOTIFIERS
7876 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7880 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
7883 #ifdef CONFIG_RT_MUTEXES
7884 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
7888 * The boot idle thread does lazy MMU switching as well:
7890 atomic_inc(&init_mm.mm_count);
7891 enter_lazy_tlb(&init_mm, current);
7894 * Make us the idle thread. Technically, schedule() should not be
7895 * called from this thread, however somewhere below it might be,
7896 * but because we are the idle thread, we just pick up running again
7897 * when this runqueue becomes "idle".
7899 init_idle(current, smp_processor_id());
7901 * During early bootup we pretend to be a normal task:
7903 current->sched_class = &fair_sched_class;
7905 scheduler_running = 1;
7908 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7909 void __might_sleep(char *file, int line)
7912 static unsigned long prev_jiffy; /* ratelimiting */
7914 if ((in_atomic() || irqs_disabled()) &&
7915 system_state == SYSTEM_RUNNING && !oops_in_progress) {
7916 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7918 prev_jiffy = jiffies;
7919 printk(KERN_ERR "BUG: sleeping function called from invalid"
7920 " context at %s:%d\n", file, line);
7921 printk("in_atomic():%d, irqs_disabled():%d\n",
7922 in_atomic(), irqs_disabled());
7923 debug_show_held_locks(current);
7924 if (irqs_disabled())
7925 print_irqtrace_events(current);
7930 EXPORT_SYMBOL(__might_sleep);
7933 #ifdef CONFIG_MAGIC_SYSRQ
7934 static void normalize_task(struct rq *rq, struct task_struct *p)
7938 update_rq_clock(rq);
7939 on_rq = p->se.on_rq;
7941 deactivate_task(rq, p, 0);
7942 __setscheduler(rq, p, SCHED_NORMAL, 0);
7944 activate_task(rq, p, 0);
7945 resched_task(rq->curr);
7949 void normalize_rt_tasks(void)
7951 struct task_struct *g, *p;
7952 unsigned long flags;
7955 read_lock_irqsave(&tasklist_lock, flags);
7956 do_each_thread(g, p) {
7958 * Only normalize user tasks:
7963 p->se.exec_start = 0;
7964 #ifdef CONFIG_SCHEDSTATS
7965 p->se.wait_start = 0;
7966 p->se.sleep_start = 0;
7967 p->se.block_start = 0;
7972 * Renice negative nice level userspace
7975 if (TASK_NICE(p) < 0 && p->mm)
7976 set_user_nice(p, 0);
7980 spin_lock(&p->pi_lock);
7981 rq = __task_rq_lock(p);
7983 normalize_task(rq, p);
7985 __task_rq_unlock(rq);
7986 spin_unlock(&p->pi_lock);
7987 } while_each_thread(g, p);
7989 read_unlock_irqrestore(&tasklist_lock, flags);
7992 #endif /* CONFIG_MAGIC_SYSRQ */
7996 * These functions are only useful for the IA64 MCA handling.
7998 * They can only be called when the whole system has been
7999 * stopped - every CPU needs to be quiescent, and no scheduling
8000 * activity can take place. Using them for anything else would
8001 * be a serious bug, and as a result, they aren't even visible
8002 * under any other configuration.
8006 * curr_task - return the current task for a given cpu.
8007 * @cpu: the processor in question.
8009 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8011 struct task_struct *curr_task(int cpu)
8013 return cpu_curr(cpu);
8017 * set_curr_task - set the current task for a given cpu.
8018 * @cpu: the processor in question.
8019 * @p: the task pointer to set.
8021 * Description: This function must only be used when non-maskable interrupts
8022 * are serviced on a separate stack. It allows the architecture to switch the
8023 * notion of the current task on a cpu in a non-blocking manner. This function
8024 * must be called with all CPU's synchronized, and interrupts disabled, the
8025 * and caller must save the original value of the current task (see
8026 * curr_task() above) and restore that value before reenabling interrupts and
8027 * re-starting the system.
8029 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8031 void set_curr_task(int cpu, struct task_struct *p)
8038 #ifdef CONFIG_FAIR_GROUP_SCHED
8039 static void free_fair_sched_group(struct task_group *tg)
8043 for_each_possible_cpu(i) {
8045 kfree(tg->cfs_rq[i]);
8055 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8057 struct cfs_rq *cfs_rq;
8058 struct sched_entity *se, *parent_se;
8062 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8065 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8069 tg->shares = NICE_0_LOAD;
8071 for_each_possible_cpu(i) {
8074 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8075 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8079 se = kmalloc_node(sizeof(struct sched_entity),
8080 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8084 parent_se = parent ? parent->se[i] : NULL;
8085 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8094 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8096 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8097 &cpu_rq(cpu)->leaf_cfs_rq_list);
8100 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8102 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8104 #else /* !CONFG_FAIR_GROUP_SCHED */
8105 static inline void free_fair_sched_group(struct task_group *tg)
8110 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8115 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8119 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8122 #endif /* CONFIG_FAIR_GROUP_SCHED */
8124 #ifdef CONFIG_RT_GROUP_SCHED
8125 static void free_rt_sched_group(struct task_group *tg)
8129 destroy_rt_bandwidth(&tg->rt_bandwidth);
8131 for_each_possible_cpu(i) {
8133 kfree(tg->rt_rq[i]);
8135 kfree(tg->rt_se[i]);
8143 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8145 struct rt_rq *rt_rq;
8146 struct sched_rt_entity *rt_se, *parent_se;
8150 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8153 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8157 init_rt_bandwidth(&tg->rt_bandwidth,
8158 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8160 for_each_possible_cpu(i) {
8163 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8164 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8168 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8169 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8173 parent_se = parent ? parent->rt_se[i] : NULL;
8174 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8183 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8185 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8186 &cpu_rq(cpu)->leaf_rt_rq_list);
8189 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8191 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8193 #else /* !CONFIG_RT_GROUP_SCHED */
8194 static inline void free_rt_sched_group(struct task_group *tg)
8199 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8204 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8208 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8211 #endif /* CONFIG_RT_GROUP_SCHED */
8213 #ifdef CONFIG_GROUP_SCHED
8214 static void free_sched_group(struct task_group *tg)
8216 free_fair_sched_group(tg);
8217 free_rt_sched_group(tg);
8221 /* allocate runqueue etc for a new task group */
8222 struct task_group *sched_create_group(struct task_group *parent)
8224 struct task_group *tg;
8225 unsigned long flags;
8228 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8230 return ERR_PTR(-ENOMEM);
8232 if (!alloc_fair_sched_group(tg, parent))
8235 if (!alloc_rt_sched_group(tg, parent))
8238 spin_lock_irqsave(&task_group_lock, flags);
8239 for_each_possible_cpu(i) {
8240 register_fair_sched_group(tg, i);
8241 register_rt_sched_group(tg, i);
8243 list_add_rcu(&tg->list, &task_groups);
8245 WARN_ON(!parent); /* root should already exist */
8247 tg->parent = parent;
8248 list_add_rcu(&tg->siblings, &parent->children);
8249 INIT_LIST_HEAD(&tg->children);
8250 spin_unlock_irqrestore(&task_group_lock, flags);
8255 free_sched_group(tg);
8256 return ERR_PTR(-ENOMEM);
8259 /* rcu callback to free various structures associated with a task group */
8260 static void free_sched_group_rcu(struct rcu_head *rhp)
8262 /* now it should be safe to free those cfs_rqs */
8263 free_sched_group(container_of(rhp, struct task_group, rcu));
8266 /* Destroy runqueue etc associated with a task group */
8267 void sched_destroy_group(struct task_group *tg)
8269 unsigned long flags;
8272 spin_lock_irqsave(&task_group_lock, flags);
8273 for_each_possible_cpu(i) {
8274 unregister_fair_sched_group(tg, i);
8275 unregister_rt_sched_group(tg, i);
8277 list_del_rcu(&tg->list);
8278 list_del_rcu(&tg->siblings);
8279 spin_unlock_irqrestore(&task_group_lock, flags);
8281 /* wait for possible concurrent references to cfs_rqs complete */
8282 call_rcu(&tg->rcu, free_sched_group_rcu);
8285 /* change task's runqueue when it moves between groups.
8286 * The caller of this function should have put the task in its new group
8287 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8288 * reflect its new group.
8290 void sched_move_task(struct task_struct *tsk)
8293 unsigned long flags;
8296 rq = task_rq_lock(tsk, &flags);
8298 update_rq_clock(rq);
8300 running = task_current(rq, tsk);
8301 on_rq = tsk->se.on_rq;
8304 dequeue_task(rq, tsk, 0);
8305 if (unlikely(running))
8306 tsk->sched_class->put_prev_task(rq, tsk);
8308 set_task_rq(tsk, task_cpu(tsk));
8310 #ifdef CONFIG_FAIR_GROUP_SCHED
8311 if (tsk->sched_class->moved_group)
8312 tsk->sched_class->moved_group(tsk);
8315 if (unlikely(running))
8316 tsk->sched_class->set_curr_task(rq);
8318 enqueue_task(rq, tsk, 0);
8320 task_rq_unlock(rq, &flags);
8322 #endif /* CONFIG_GROUP_SCHED */
8324 #ifdef CONFIG_FAIR_GROUP_SCHED
8325 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8327 struct cfs_rq *cfs_rq = se->cfs_rq;
8328 struct rq *rq = cfs_rq->rq;
8331 spin_lock_irq(&rq->lock);
8335 dequeue_entity(cfs_rq, se, 0);
8337 se->load.weight = shares;
8338 se->load.inv_weight = 0;
8341 enqueue_entity(cfs_rq, se, 0);
8343 spin_unlock_irq(&rq->lock);
8346 static DEFINE_MUTEX(shares_mutex);
8348 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8351 unsigned long flags;
8354 * We can't change the weight of the root cgroup.
8359 if (shares < MIN_SHARES)
8360 shares = MIN_SHARES;
8361 else if (shares > MAX_SHARES)
8362 shares = MAX_SHARES;
8364 mutex_lock(&shares_mutex);
8365 if (tg->shares == shares)
8368 spin_lock_irqsave(&task_group_lock, flags);
8369 for_each_possible_cpu(i)
8370 unregister_fair_sched_group(tg, i);
8371 list_del_rcu(&tg->siblings);
8372 spin_unlock_irqrestore(&task_group_lock, flags);
8374 /* wait for any ongoing reference to this group to finish */
8375 synchronize_sched();
8378 * Now we are free to modify the group's share on each cpu
8379 * w/o tripping rebalance_share or load_balance_fair.
8381 tg->shares = shares;
8382 for_each_possible_cpu(i)
8383 set_se_shares(tg->se[i], shares);
8386 * Enable load balance activity on this group, by inserting it back on
8387 * each cpu's rq->leaf_cfs_rq_list.
8389 spin_lock_irqsave(&task_group_lock, flags);
8390 for_each_possible_cpu(i)
8391 register_fair_sched_group(tg, i);
8392 list_add_rcu(&tg->siblings, &tg->parent->children);
8393 spin_unlock_irqrestore(&task_group_lock, flags);
8395 mutex_unlock(&shares_mutex);
8399 unsigned long sched_group_shares(struct task_group *tg)
8405 #ifdef CONFIG_RT_GROUP_SCHED
8407 * Ensure that the real time constraints are schedulable.
8409 static DEFINE_MUTEX(rt_constraints_mutex);
8411 static unsigned long to_ratio(u64 period, u64 runtime)
8413 if (runtime == RUNTIME_INF)
8416 return div64_u64(runtime << 16, period);
8419 #ifdef CONFIG_CGROUP_SCHED
8420 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8422 struct task_group *tgi, *parent = tg->parent;
8423 unsigned long total = 0;
8426 if (global_rt_period() < period)
8429 return to_ratio(period, runtime) <
8430 to_ratio(global_rt_period(), global_rt_runtime());
8433 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8437 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8441 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8442 tgi->rt_bandwidth.rt_runtime);
8446 return total + to_ratio(period, runtime) <=
8447 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8448 parent->rt_bandwidth.rt_runtime);
8450 #elif defined CONFIG_USER_SCHED
8451 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8453 struct task_group *tgi;
8454 unsigned long total = 0;
8455 unsigned long global_ratio =
8456 to_ratio(global_rt_period(), global_rt_runtime());
8459 list_for_each_entry_rcu(tgi, &task_groups, list) {
8463 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8464 tgi->rt_bandwidth.rt_runtime);
8468 return total + to_ratio(period, runtime) < global_ratio;
8472 /* Must be called with tasklist_lock held */
8473 static inline int tg_has_rt_tasks(struct task_group *tg)
8475 struct task_struct *g, *p;
8476 do_each_thread(g, p) {
8477 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8479 } while_each_thread(g, p);
8483 static int tg_set_bandwidth(struct task_group *tg,
8484 u64 rt_period, u64 rt_runtime)
8488 mutex_lock(&rt_constraints_mutex);
8489 read_lock(&tasklist_lock);
8490 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8494 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8499 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8500 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8501 tg->rt_bandwidth.rt_runtime = rt_runtime;
8503 for_each_possible_cpu(i) {
8504 struct rt_rq *rt_rq = tg->rt_rq[i];
8506 spin_lock(&rt_rq->rt_runtime_lock);
8507 rt_rq->rt_runtime = rt_runtime;
8508 spin_unlock(&rt_rq->rt_runtime_lock);
8510 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8512 read_unlock(&tasklist_lock);
8513 mutex_unlock(&rt_constraints_mutex);
8518 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8520 u64 rt_runtime, rt_period;
8522 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8523 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8524 if (rt_runtime_us < 0)
8525 rt_runtime = RUNTIME_INF;
8527 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8530 long sched_group_rt_runtime(struct task_group *tg)
8534 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8537 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8538 do_div(rt_runtime_us, NSEC_PER_USEC);
8539 return rt_runtime_us;
8542 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8544 u64 rt_runtime, rt_period;
8546 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8547 rt_runtime = tg->rt_bandwidth.rt_runtime;
8549 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8552 long sched_group_rt_period(struct task_group *tg)
8556 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8557 do_div(rt_period_us, NSEC_PER_USEC);
8558 return rt_period_us;
8561 static int sched_rt_global_constraints(void)
8563 struct task_group *tg = &root_task_group;
8564 u64 rt_runtime, rt_period;
8567 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8568 rt_runtime = tg->rt_bandwidth.rt_runtime;
8570 mutex_lock(&rt_constraints_mutex);
8571 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8573 mutex_unlock(&rt_constraints_mutex);
8577 #else /* !CONFIG_RT_GROUP_SCHED */
8578 static int sched_rt_global_constraints(void)
8580 unsigned long flags;
8583 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8584 for_each_possible_cpu(i) {
8585 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8587 spin_lock(&rt_rq->rt_runtime_lock);
8588 rt_rq->rt_runtime = global_rt_runtime();
8589 spin_unlock(&rt_rq->rt_runtime_lock);
8591 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8595 #endif /* CONFIG_RT_GROUP_SCHED */
8597 int sched_rt_handler(struct ctl_table *table, int write,
8598 struct file *filp, void __user *buffer, size_t *lenp,
8602 int old_period, old_runtime;
8603 static DEFINE_MUTEX(mutex);
8606 old_period = sysctl_sched_rt_period;
8607 old_runtime = sysctl_sched_rt_runtime;
8609 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8611 if (!ret && write) {
8612 ret = sched_rt_global_constraints();
8614 sysctl_sched_rt_period = old_period;
8615 sysctl_sched_rt_runtime = old_runtime;
8617 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8618 def_rt_bandwidth.rt_period =
8619 ns_to_ktime(global_rt_period());
8622 mutex_unlock(&mutex);
8627 #ifdef CONFIG_CGROUP_SCHED
8629 /* return corresponding task_group object of a cgroup */
8630 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8632 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8633 struct task_group, css);
8636 static struct cgroup_subsys_state *
8637 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8639 struct task_group *tg, *parent;
8641 if (!cgrp->parent) {
8642 /* This is early initialization for the top cgroup */
8643 init_task_group.css.cgroup = cgrp;
8644 return &init_task_group.css;
8647 parent = cgroup_tg(cgrp->parent);
8648 tg = sched_create_group(parent);
8650 return ERR_PTR(-ENOMEM);
8652 /* Bind the cgroup to task_group object we just created */
8653 tg->css.cgroup = cgrp;
8659 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8661 struct task_group *tg = cgroup_tg(cgrp);
8663 sched_destroy_group(tg);
8667 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8668 struct task_struct *tsk)
8670 #ifdef CONFIG_RT_GROUP_SCHED
8671 /* Don't accept realtime tasks when there is no way for them to run */
8672 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8675 /* We don't support RT-tasks being in separate groups */
8676 if (tsk->sched_class != &fair_sched_class)
8684 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8685 struct cgroup *old_cont, struct task_struct *tsk)
8687 sched_move_task(tsk);
8690 #ifdef CONFIG_FAIR_GROUP_SCHED
8691 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8694 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8697 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8699 struct task_group *tg = cgroup_tg(cgrp);
8701 return (u64) tg->shares;
8703 #endif /* CONFIG_FAIR_GROUP_SCHED */
8705 #ifdef CONFIG_RT_GROUP_SCHED
8706 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8709 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8712 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8714 return sched_group_rt_runtime(cgroup_tg(cgrp));
8717 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8720 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8723 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8725 return sched_group_rt_period(cgroup_tg(cgrp));
8727 #endif /* CONFIG_RT_GROUP_SCHED */
8729 static struct cftype cpu_files[] = {
8730 #ifdef CONFIG_FAIR_GROUP_SCHED
8733 .read_u64 = cpu_shares_read_u64,
8734 .write_u64 = cpu_shares_write_u64,
8737 #ifdef CONFIG_RT_GROUP_SCHED
8739 .name = "rt_runtime_us",
8740 .read_s64 = cpu_rt_runtime_read,
8741 .write_s64 = cpu_rt_runtime_write,
8744 .name = "rt_period_us",
8745 .read_u64 = cpu_rt_period_read_uint,
8746 .write_u64 = cpu_rt_period_write_uint,
8751 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8753 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8756 struct cgroup_subsys cpu_cgroup_subsys = {
8758 .create = cpu_cgroup_create,
8759 .destroy = cpu_cgroup_destroy,
8760 .can_attach = cpu_cgroup_can_attach,
8761 .attach = cpu_cgroup_attach,
8762 .populate = cpu_cgroup_populate,
8763 .subsys_id = cpu_cgroup_subsys_id,
8767 #endif /* CONFIG_CGROUP_SCHED */
8769 #ifdef CONFIG_CGROUP_CPUACCT
8772 * CPU accounting code for task groups.
8774 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8775 * (balbir@in.ibm.com).
8778 /* track cpu usage of a group of tasks */
8780 struct cgroup_subsys_state css;
8781 /* cpuusage holds pointer to a u64-type object on every cpu */
8785 struct cgroup_subsys cpuacct_subsys;
8787 /* return cpu accounting group corresponding to this container */
8788 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8790 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8791 struct cpuacct, css);
8794 /* return cpu accounting group to which this task belongs */
8795 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8797 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8798 struct cpuacct, css);
8801 /* create a new cpu accounting group */
8802 static struct cgroup_subsys_state *cpuacct_create(
8803 struct cgroup_subsys *ss, struct cgroup *cgrp)
8805 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8808 return ERR_PTR(-ENOMEM);
8810 ca->cpuusage = alloc_percpu(u64);
8811 if (!ca->cpuusage) {
8813 return ERR_PTR(-ENOMEM);
8819 /* destroy an existing cpu accounting group */
8821 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8823 struct cpuacct *ca = cgroup_ca(cgrp);
8825 free_percpu(ca->cpuusage);
8829 /* return total cpu usage (in nanoseconds) of a group */
8830 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8832 struct cpuacct *ca = cgroup_ca(cgrp);
8833 u64 totalcpuusage = 0;
8836 for_each_possible_cpu(i) {
8837 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8840 * Take rq->lock to make 64-bit addition safe on 32-bit
8843 spin_lock_irq(&cpu_rq(i)->lock);
8844 totalcpuusage += *cpuusage;
8845 spin_unlock_irq(&cpu_rq(i)->lock);
8848 return totalcpuusage;
8851 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8854 struct cpuacct *ca = cgroup_ca(cgrp);
8863 for_each_possible_cpu(i) {
8864 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8866 spin_lock_irq(&cpu_rq(i)->lock);
8868 spin_unlock_irq(&cpu_rq(i)->lock);
8874 static struct cftype files[] = {
8877 .read_u64 = cpuusage_read,
8878 .write_u64 = cpuusage_write,
8882 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8884 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8888 * charge this task's execution time to its accounting group.
8890 * called with rq->lock held.
8892 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8896 if (!cpuacct_subsys.active)
8901 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
8903 *cpuusage += cputime;
8907 struct cgroup_subsys cpuacct_subsys = {
8909 .create = cpuacct_create,
8910 .destroy = cpuacct_destroy,
8911 .populate = cpuacct_populate,
8912 .subsys_id = cpuacct_subsys_id,
8914 #endif /* CONFIG_CGROUP_CPUACCT */