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>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
103 #define NICE_0_LOAD SCHED_LOAD_SCALE
104 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
107 * These are the 'tuning knobs' of the scheduler:
109 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
110 * Timeslices get refilled after they expire.
112 #define DEF_TIMESLICE (100 * HZ / 1000)
115 * single value that denotes runtime == period, ie unlimited time.
117 #define RUNTIME_INF ((u64)~0ULL)
121 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
122 * Since cpu_power is a 'constant', we can use a reciprocal divide.
124 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
126 return reciprocal_divide(load, sg->reciprocal_cpu_power);
130 * Each time a sched group cpu_power is changed,
131 * we must compute its reciprocal value
133 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
135 sg->__cpu_power += val;
136 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
140 static inline int rt_policy(int policy)
142 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
147 static inline int task_has_rt_policy(struct task_struct *p)
149 return rt_policy(p->policy);
153 * This is the priority-queue data structure of the RT scheduling class:
155 struct rt_prio_array {
156 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
157 struct list_head queue[MAX_RT_PRIO];
160 struct rt_bandwidth {
161 /* nests inside the rq lock: */
162 spinlock_t rt_runtime_lock;
165 struct hrtimer rt_period_timer;
168 static struct rt_bandwidth def_rt_bandwidth;
170 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
172 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
174 struct rt_bandwidth *rt_b =
175 container_of(timer, struct rt_bandwidth, rt_period_timer);
181 now = hrtimer_cb_get_time(timer);
182 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
187 idle = do_sched_rt_period_timer(rt_b, overrun);
190 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
194 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
196 rt_b->rt_period = ns_to_ktime(period);
197 rt_b->rt_runtime = runtime;
199 spin_lock_init(&rt_b->rt_runtime_lock);
201 hrtimer_init(&rt_b->rt_period_timer,
202 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
203 rt_b->rt_period_timer.function = sched_rt_period_timer;
204 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
207 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
211 if (rt_b->rt_runtime == RUNTIME_INF)
214 if (hrtimer_active(&rt_b->rt_period_timer))
217 spin_lock(&rt_b->rt_runtime_lock);
219 if (hrtimer_active(&rt_b->rt_period_timer))
222 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
223 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
224 hrtimer_start(&rt_b->rt_period_timer,
225 rt_b->rt_period_timer.expires,
228 spin_unlock(&rt_b->rt_runtime_lock);
231 #ifdef CONFIG_RT_GROUP_SCHED
232 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
234 hrtimer_cancel(&rt_b->rt_period_timer);
239 * sched_domains_mutex serializes calls to arch_init_sched_domains,
240 * detach_destroy_domains and partition_sched_domains.
242 static DEFINE_MUTEX(sched_domains_mutex);
244 #ifdef CONFIG_GROUP_SCHED
246 #include <linux/cgroup.h>
250 static LIST_HEAD(task_groups);
252 /* task group related information */
254 #ifdef CONFIG_CGROUP_SCHED
255 struct cgroup_subsys_state css;
258 #ifdef CONFIG_FAIR_GROUP_SCHED
259 /* schedulable entities of this group on each cpu */
260 struct sched_entity **se;
261 /* runqueue "owned" by this group on each cpu */
262 struct cfs_rq **cfs_rq;
263 unsigned long shares;
266 #ifdef CONFIG_RT_GROUP_SCHED
267 struct sched_rt_entity **rt_se;
268 struct rt_rq **rt_rq;
270 struct rt_bandwidth rt_bandwidth;
274 struct list_head list;
276 struct task_group *parent;
277 struct list_head siblings;
278 struct list_head children;
281 #ifdef CONFIG_USER_SCHED
285 * Every UID task group (including init_task_group aka UID-0) will
286 * be a child to this group.
288 struct task_group root_task_group;
290 #ifdef CONFIG_FAIR_GROUP_SCHED
291 /* Default task group's sched entity on each cpu */
292 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
293 /* Default task group's cfs_rq on each cpu */
294 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
295 #endif /* CONFIG_FAIR_GROUP_SCHED */
297 #ifdef CONFIG_RT_GROUP_SCHED
298 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
299 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
300 #endif /* CONFIG_RT_GROUP_SCHED */
301 #else /* !CONFIG_FAIR_GROUP_SCHED */
302 #define root_task_group init_task_group
303 #endif /* CONFIG_FAIR_GROUP_SCHED */
305 /* task_group_lock serializes add/remove of task groups and also changes to
306 * a task group's cpu shares.
308 static DEFINE_SPINLOCK(task_group_lock);
310 #ifdef CONFIG_FAIR_GROUP_SCHED
311 #ifdef CONFIG_USER_SCHED
312 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
313 #else /* !CONFIG_USER_SCHED */
314 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
315 #endif /* CONFIG_USER_SCHED */
318 * A weight of 0 or 1 can cause arithmetics problems.
319 * A weight of a cfs_rq is the sum of weights of which entities
320 * are queued on this cfs_rq, so a weight of a entity should not be
321 * too large, so as the shares value of a task group.
322 * (The default weight is 1024 - so there's no practical
323 * limitation from this.)
326 #define MAX_SHARES (1UL << 18)
328 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
331 /* Default task group.
332 * Every task in system belong to this group at bootup.
334 struct task_group init_task_group;
336 /* return group to which a task belongs */
337 static inline struct task_group *task_group(struct task_struct *p)
339 struct task_group *tg;
341 #ifdef CONFIG_USER_SCHED
343 #elif defined(CONFIG_CGROUP_SCHED)
344 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
345 struct task_group, css);
347 tg = &init_task_group;
352 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
353 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
355 #ifdef CONFIG_FAIR_GROUP_SCHED
356 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
357 p->se.parent = task_group(p)->se[cpu];
360 #ifdef CONFIG_RT_GROUP_SCHED
361 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
362 p->rt.parent = task_group(p)->rt_se[cpu];
368 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
369 static inline struct task_group *task_group(struct task_struct *p)
374 #endif /* CONFIG_GROUP_SCHED */
376 /* CFS-related fields in a runqueue */
378 struct load_weight load;
379 unsigned long nr_running;
385 struct rb_root tasks_timeline;
386 struct rb_node *rb_leftmost;
388 struct list_head tasks;
389 struct list_head *balance_iterator;
392 * 'curr' points to currently running entity on this cfs_rq.
393 * It is set to NULL otherwise (i.e when none are currently running).
395 struct sched_entity *curr, *next;
397 unsigned long nr_spread_over;
399 #ifdef CONFIG_FAIR_GROUP_SCHED
400 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
403 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
404 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
405 * (like users, containers etc.)
407 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
408 * list is used during load balance.
410 struct list_head leaf_cfs_rq_list;
411 struct task_group *tg; /* group that "owns" this runqueue */
415 * the part of load.weight contributed by tasks
417 unsigned long task_weight;
420 * h_load = weight * f(tg)
422 * Where f(tg) is the recursive weight fraction assigned to
425 unsigned long h_load;
428 * this cpu's part of tg->shares
430 unsigned long shares;
433 * load.weight at the time we set shares
435 unsigned long rq_weight;
440 /* Real-Time classes' related field in a runqueue: */
442 struct rt_prio_array active;
443 unsigned long rt_nr_running;
444 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
445 int highest_prio; /* highest queued rt task prio */
448 unsigned long rt_nr_migratory;
454 /* Nests inside the rq lock: */
455 spinlock_t rt_runtime_lock;
457 #ifdef CONFIG_RT_GROUP_SCHED
458 unsigned long rt_nr_boosted;
461 struct list_head leaf_rt_rq_list;
462 struct task_group *tg;
463 struct sched_rt_entity *rt_se;
470 * We add the notion of a root-domain which will be used to define per-domain
471 * variables. Each exclusive cpuset essentially defines an island domain by
472 * fully partitioning the member cpus from any other cpuset. Whenever a new
473 * exclusive cpuset is created, we also create and attach a new root-domain
483 * The "RT overload" flag: it gets set if a CPU has more than
484 * one runnable RT task.
489 struct cpupri cpupri;
494 * By default the system creates a single root-domain with all cpus as
495 * members (mimicking the global state we have today).
497 static struct root_domain def_root_domain;
502 * This is the main, per-CPU runqueue data structure.
504 * Locking rule: those places that want to lock multiple runqueues
505 * (such as the load balancing or the thread migration code), lock
506 * acquire operations must be ordered by ascending &runqueue.
513 * nr_running and cpu_load should be in the same cacheline because
514 * remote CPUs use both these fields when doing load calculation.
516 unsigned long nr_running;
517 #define CPU_LOAD_IDX_MAX 5
518 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
519 unsigned char idle_at_tick;
521 unsigned long last_tick_seen;
522 unsigned char in_nohz_recently;
524 /* capture load from *all* tasks on this cpu: */
525 struct load_weight load;
526 unsigned long nr_load_updates;
532 #ifdef CONFIG_FAIR_GROUP_SCHED
533 /* list of leaf cfs_rq on this cpu: */
534 struct list_head leaf_cfs_rq_list;
536 #ifdef CONFIG_RT_GROUP_SCHED
537 struct list_head leaf_rt_rq_list;
541 * This is part of a global counter where only the total sum
542 * over all CPUs matters. A task can increase this counter on
543 * one CPU and if it got migrated afterwards it may decrease
544 * it on another CPU. Always updated under the runqueue lock:
546 unsigned long nr_uninterruptible;
548 struct task_struct *curr, *idle;
549 unsigned long next_balance;
550 struct mm_struct *prev_mm;
557 struct root_domain *rd;
558 struct sched_domain *sd;
560 /* For active balancing */
563 /* cpu of this runqueue: */
567 unsigned long avg_load_per_task;
569 struct task_struct *migration_thread;
570 struct list_head migration_queue;
573 #ifdef CONFIG_SCHED_HRTICK
575 int hrtick_csd_pending;
576 struct call_single_data hrtick_csd;
578 struct hrtimer hrtick_timer;
581 #ifdef CONFIG_SCHEDSTATS
583 struct sched_info rq_sched_info;
585 /* sys_sched_yield() stats */
586 unsigned int yld_exp_empty;
587 unsigned int yld_act_empty;
588 unsigned int yld_both_empty;
589 unsigned int yld_count;
591 /* schedule() stats */
592 unsigned int sched_switch;
593 unsigned int sched_count;
594 unsigned int sched_goidle;
596 /* try_to_wake_up() stats */
597 unsigned int ttwu_count;
598 unsigned int ttwu_local;
601 unsigned int bkl_count;
605 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
607 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
609 rq->curr->sched_class->check_preempt_curr(rq, p);
612 static inline int cpu_of(struct rq *rq)
622 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
623 * See detach_destroy_domains: synchronize_sched for details.
625 * The domain tree of any CPU may only be accessed from within
626 * preempt-disabled sections.
628 #define for_each_domain(cpu, __sd) \
629 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
631 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
632 #define this_rq() (&__get_cpu_var(runqueues))
633 #define task_rq(p) cpu_rq(task_cpu(p))
634 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
636 static inline void update_rq_clock(struct rq *rq)
638 rq->clock = sched_clock_cpu(cpu_of(rq));
642 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
644 #ifdef CONFIG_SCHED_DEBUG
645 # define const_debug __read_mostly
647 # define const_debug static const
653 * Returns true if the current cpu runqueue is locked.
654 * This interface allows printk to be called with the runqueue lock
655 * held and know whether or not it is OK to wake up the klogd.
657 int runqueue_is_locked(void)
660 struct rq *rq = cpu_rq(cpu);
663 ret = spin_is_locked(&rq->lock);
669 * Debugging: various feature bits
672 #define SCHED_FEAT(name, enabled) \
673 __SCHED_FEAT_##name ,
676 #include "sched_features.h"
681 #define SCHED_FEAT(name, enabled) \
682 (1UL << __SCHED_FEAT_##name) * enabled |
684 const_debug unsigned int sysctl_sched_features =
685 #include "sched_features.h"
690 #ifdef CONFIG_SCHED_DEBUG
691 #define SCHED_FEAT(name, enabled) \
694 static __read_mostly char *sched_feat_names[] = {
695 #include "sched_features.h"
701 static int sched_feat_open(struct inode *inode, struct file *filp)
703 filp->private_data = inode->i_private;
708 sched_feat_read(struct file *filp, char __user *ubuf,
709 size_t cnt, loff_t *ppos)
716 for (i = 0; sched_feat_names[i]; i++) {
717 len += strlen(sched_feat_names[i]);
721 buf = kmalloc(len + 2, GFP_KERNEL);
725 for (i = 0; sched_feat_names[i]; i++) {
726 if (sysctl_sched_features & (1UL << i))
727 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
729 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
732 r += sprintf(buf + r, "\n");
733 WARN_ON(r >= len + 2);
735 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
743 sched_feat_write(struct file *filp, const char __user *ubuf,
744 size_t cnt, loff_t *ppos)
754 if (copy_from_user(&buf, ubuf, cnt))
759 if (strncmp(buf, "NO_", 3) == 0) {
764 for (i = 0; sched_feat_names[i]; i++) {
765 int len = strlen(sched_feat_names[i]);
767 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
769 sysctl_sched_features &= ~(1UL << i);
771 sysctl_sched_features |= (1UL << i);
776 if (!sched_feat_names[i])
784 static struct file_operations sched_feat_fops = {
785 .open = sched_feat_open,
786 .read = sched_feat_read,
787 .write = sched_feat_write,
790 static __init int sched_init_debug(void)
792 debugfs_create_file("sched_features", 0644, NULL, NULL,
797 late_initcall(sched_init_debug);
801 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
804 * Number of tasks to iterate in a single balance run.
805 * Limited because this is done with IRQs disabled.
807 const_debug unsigned int sysctl_sched_nr_migrate = 32;
810 * ratelimit for updating the group shares.
813 unsigned int sysctl_sched_shares_ratelimit = 250000;
816 * period over which we measure -rt task cpu usage in us.
819 unsigned int sysctl_sched_rt_period = 1000000;
821 static __read_mostly int scheduler_running;
824 * part of the period that we allow rt tasks to run in us.
827 int sysctl_sched_rt_runtime = 950000;
829 static inline u64 global_rt_period(void)
831 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
834 static inline u64 global_rt_runtime(void)
836 if (sysctl_sched_rt_runtime < 0)
839 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
842 #ifndef prepare_arch_switch
843 # define prepare_arch_switch(next) do { } while (0)
845 #ifndef finish_arch_switch
846 # define finish_arch_switch(prev) do { } while (0)
849 static inline int task_current(struct rq *rq, struct task_struct *p)
851 return rq->curr == p;
854 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
855 static inline int task_running(struct rq *rq, struct task_struct *p)
857 return task_current(rq, p);
860 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
864 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
866 #ifdef CONFIG_DEBUG_SPINLOCK
867 /* this is a valid case when another task releases the spinlock */
868 rq->lock.owner = current;
871 * If we are tracking spinlock dependencies then we have to
872 * fix up the runqueue lock - which gets 'carried over' from
875 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
877 spin_unlock_irq(&rq->lock);
880 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
881 static inline int task_running(struct rq *rq, struct task_struct *p)
886 return task_current(rq, p);
890 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
894 * We can optimise this out completely for !SMP, because the
895 * SMP rebalancing from interrupt is the only thing that cares
900 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
901 spin_unlock_irq(&rq->lock);
903 spin_unlock(&rq->lock);
907 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
911 * After ->oncpu is cleared, the task can be moved to a different CPU.
912 * We must ensure this doesn't happen until the switch is completely
918 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
922 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
925 * __task_rq_lock - lock the runqueue a given task resides on.
926 * Must be called interrupts disabled.
928 static inline struct rq *__task_rq_lock(struct task_struct *p)
932 struct rq *rq = task_rq(p);
933 spin_lock(&rq->lock);
934 if (likely(rq == task_rq(p)))
936 spin_unlock(&rq->lock);
941 * task_rq_lock - lock the runqueue a given task resides on and disable
942 * interrupts. Note the ordering: we can safely lookup the task_rq without
943 * explicitly disabling preemption.
945 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
951 local_irq_save(*flags);
953 spin_lock(&rq->lock);
954 if (likely(rq == task_rq(p)))
956 spin_unlock_irqrestore(&rq->lock, *flags);
960 static void __task_rq_unlock(struct rq *rq)
963 spin_unlock(&rq->lock);
966 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
969 spin_unlock_irqrestore(&rq->lock, *flags);
973 * this_rq_lock - lock this runqueue and disable interrupts.
975 static struct rq *this_rq_lock(void)
982 spin_lock(&rq->lock);
987 #ifdef CONFIG_SCHED_HRTICK
989 * Use HR-timers to deliver accurate preemption points.
991 * Its all a bit involved since we cannot program an hrt while holding the
992 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
995 * When we get rescheduled we reprogram the hrtick_timer outside of the
1001 * - enabled by features
1002 * - hrtimer is actually high res
1004 static inline int hrtick_enabled(struct rq *rq)
1006 if (!sched_feat(HRTICK))
1008 if (!cpu_active(cpu_of(rq)))
1010 return hrtimer_is_hres_active(&rq->hrtick_timer);
1013 static void hrtick_clear(struct rq *rq)
1015 if (hrtimer_active(&rq->hrtick_timer))
1016 hrtimer_cancel(&rq->hrtick_timer);
1020 * High-resolution timer tick.
1021 * Runs from hardirq context with interrupts disabled.
1023 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1025 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1027 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1029 spin_lock(&rq->lock);
1030 update_rq_clock(rq);
1031 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1032 spin_unlock(&rq->lock);
1034 return HRTIMER_NORESTART;
1039 * called from hardirq (IPI) context
1041 static void __hrtick_start(void *arg)
1043 struct rq *rq = arg;
1045 spin_lock(&rq->lock);
1046 hrtimer_restart(&rq->hrtick_timer);
1047 rq->hrtick_csd_pending = 0;
1048 spin_unlock(&rq->lock);
1052 * Called to set the hrtick timer state.
1054 * called with rq->lock held and irqs disabled
1056 static void hrtick_start(struct rq *rq, u64 delay)
1058 struct hrtimer *timer = &rq->hrtick_timer;
1059 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1061 timer->expires = time;
1063 if (rq == this_rq()) {
1064 hrtimer_restart(timer);
1065 } else if (!rq->hrtick_csd_pending) {
1066 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1067 rq->hrtick_csd_pending = 1;
1072 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1074 int cpu = (int)(long)hcpu;
1077 case CPU_UP_CANCELED:
1078 case CPU_UP_CANCELED_FROZEN:
1079 case CPU_DOWN_PREPARE:
1080 case CPU_DOWN_PREPARE_FROZEN:
1082 case CPU_DEAD_FROZEN:
1083 hrtick_clear(cpu_rq(cpu));
1090 static void init_hrtick(void)
1092 hotcpu_notifier(hotplug_hrtick, 0);
1096 * Called to set the hrtick timer state.
1098 * called with rq->lock held and irqs disabled
1100 static void hrtick_start(struct rq *rq, u64 delay)
1102 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1105 static void init_hrtick(void)
1108 #endif /* CONFIG_SMP */
1110 static void init_rq_hrtick(struct rq *rq)
1113 rq->hrtick_csd_pending = 0;
1115 rq->hrtick_csd.flags = 0;
1116 rq->hrtick_csd.func = __hrtick_start;
1117 rq->hrtick_csd.info = rq;
1120 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1121 rq->hrtick_timer.function = hrtick;
1122 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1125 static inline void hrtick_clear(struct rq *rq)
1129 static inline void init_rq_hrtick(struct rq *rq)
1133 static inline void init_hrtick(void)
1139 * resched_task - mark a task 'to be rescheduled now'.
1141 * On UP this means the setting of the need_resched flag, on SMP it
1142 * might also involve a cross-CPU call to trigger the scheduler on
1147 #ifndef tsk_is_polling
1148 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1151 static void resched_task(struct task_struct *p)
1155 assert_spin_locked(&task_rq(p)->lock);
1157 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1160 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1163 if (cpu == smp_processor_id())
1166 /* NEED_RESCHED must be visible before we test polling */
1168 if (!tsk_is_polling(p))
1169 smp_send_reschedule(cpu);
1172 static void resched_cpu(int cpu)
1174 struct rq *rq = cpu_rq(cpu);
1175 unsigned long flags;
1177 if (!spin_trylock_irqsave(&rq->lock, flags))
1179 resched_task(cpu_curr(cpu));
1180 spin_unlock_irqrestore(&rq->lock, flags);
1185 * When add_timer_on() enqueues a timer into the timer wheel of an
1186 * idle CPU then this timer might expire before the next timer event
1187 * which is scheduled to wake up that CPU. In case of a completely
1188 * idle system the next event might even be infinite time into the
1189 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1190 * leaves the inner idle loop so the newly added timer is taken into
1191 * account when the CPU goes back to idle and evaluates the timer
1192 * wheel for the next timer event.
1194 void wake_up_idle_cpu(int cpu)
1196 struct rq *rq = cpu_rq(cpu);
1198 if (cpu == smp_processor_id())
1202 * This is safe, as this function is called with the timer
1203 * wheel base lock of (cpu) held. When the CPU is on the way
1204 * to idle and has not yet set rq->curr to idle then it will
1205 * be serialized on the timer wheel base lock and take the new
1206 * timer into account automatically.
1208 if (rq->curr != rq->idle)
1212 * We can set TIF_RESCHED on the idle task of the other CPU
1213 * lockless. The worst case is that the other CPU runs the
1214 * idle task through an additional NOOP schedule()
1216 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1218 /* NEED_RESCHED must be visible before we test polling */
1220 if (!tsk_is_polling(rq->idle))
1221 smp_send_reschedule(cpu);
1223 #endif /* CONFIG_NO_HZ */
1225 #else /* !CONFIG_SMP */
1226 static void resched_task(struct task_struct *p)
1228 assert_spin_locked(&task_rq(p)->lock);
1229 set_tsk_need_resched(p);
1231 #endif /* CONFIG_SMP */
1233 #if BITS_PER_LONG == 32
1234 # define WMULT_CONST (~0UL)
1236 # define WMULT_CONST (1UL << 32)
1239 #define WMULT_SHIFT 32
1242 * Shift right and round:
1244 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1247 * delta *= weight / lw
1249 static unsigned long
1250 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1251 struct load_weight *lw)
1255 if (!lw->inv_weight) {
1256 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1259 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1263 tmp = (u64)delta_exec * weight;
1265 * Check whether we'd overflow the 64-bit multiplication:
1267 if (unlikely(tmp > WMULT_CONST))
1268 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1271 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1273 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1276 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1282 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1289 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1290 * of tasks with abnormal "nice" values across CPUs the contribution that
1291 * each task makes to its run queue's load is weighted according to its
1292 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1293 * scaled version of the new time slice allocation that they receive on time
1297 #define WEIGHT_IDLEPRIO 2
1298 #define WMULT_IDLEPRIO (1 << 31)
1301 * Nice levels are multiplicative, with a gentle 10% change for every
1302 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1303 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1304 * that remained on nice 0.
1306 * The "10% effect" is relative and cumulative: from _any_ nice level,
1307 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1308 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1309 * If a task goes up by ~10% and another task goes down by ~10% then
1310 * the relative distance between them is ~25%.)
1312 static const int prio_to_weight[40] = {
1313 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1314 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1315 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1316 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1317 /* 0 */ 1024, 820, 655, 526, 423,
1318 /* 5 */ 335, 272, 215, 172, 137,
1319 /* 10 */ 110, 87, 70, 56, 45,
1320 /* 15 */ 36, 29, 23, 18, 15,
1324 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1326 * In cases where the weight does not change often, we can use the
1327 * precalculated inverse to speed up arithmetics by turning divisions
1328 * into multiplications:
1330 static const u32 prio_to_wmult[40] = {
1331 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1332 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1333 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1334 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1335 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1336 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1337 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1338 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1341 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1344 * runqueue iterator, to support SMP load-balancing between different
1345 * scheduling classes, without having to expose their internal data
1346 * structures to the load-balancing proper:
1348 struct rq_iterator {
1350 struct task_struct *(*start)(void *);
1351 struct task_struct *(*next)(void *);
1355 static unsigned long
1356 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1357 unsigned long max_load_move, struct sched_domain *sd,
1358 enum cpu_idle_type idle, int *all_pinned,
1359 int *this_best_prio, struct rq_iterator *iterator);
1362 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1363 struct sched_domain *sd, enum cpu_idle_type idle,
1364 struct rq_iterator *iterator);
1367 #ifdef CONFIG_CGROUP_CPUACCT
1368 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1370 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1373 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1375 update_load_add(&rq->load, load);
1378 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1380 update_load_sub(&rq->load, load);
1384 static unsigned long source_load(int cpu, int type);
1385 static unsigned long target_load(int cpu, int type);
1386 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1388 static unsigned long cpu_avg_load_per_task(int cpu)
1390 struct rq *rq = cpu_rq(cpu);
1393 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1395 return rq->avg_load_per_task;
1398 #ifdef CONFIG_FAIR_GROUP_SCHED
1400 typedef void (*tg_visitor)(struct task_group *, int, struct sched_domain *);
1403 * Iterate the full tree, calling @down when first entering a node and @up when
1404 * leaving it for the final time.
1407 walk_tg_tree(tg_visitor down, tg_visitor up, int cpu, struct sched_domain *sd)
1409 struct task_group *parent, *child;
1412 parent = &root_task_group;
1414 (*down)(parent, cpu, sd);
1415 list_for_each_entry_rcu(child, &parent->children, siblings) {
1422 (*up)(parent, cpu, sd);
1425 parent = parent->parent;
1431 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1434 * Calculate and set the cpu's group shares.
1437 __update_group_shares_cpu(struct task_group *tg, int cpu,
1438 unsigned long sd_shares, unsigned long sd_rq_weight)
1441 unsigned long shares;
1442 unsigned long rq_weight;
1447 rq_weight = tg->cfs_rq[cpu]->load.weight;
1450 * If there are currently no tasks on the cpu pretend there is one of
1451 * average load so that when a new task gets to run here it will not
1452 * get delayed by group starvation.
1456 rq_weight = NICE_0_LOAD;
1459 if (unlikely(rq_weight > sd_rq_weight))
1460 rq_weight = sd_rq_weight;
1463 * \Sum shares * rq_weight
1464 * shares = -----------------------
1468 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1471 * record the actual number of shares, not the boosted amount.
1473 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1474 tg->cfs_rq[cpu]->rq_weight = rq_weight;
1476 if (shares < MIN_SHARES)
1477 shares = MIN_SHARES;
1478 else if (shares > MAX_SHARES)
1479 shares = MAX_SHARES;
1481 __set_se_shares(tg->se[cpu], shares);
1485 * Re-compute the task group their per cpu shares over the given domain.
1486 * This needs to be done in a bottom-up fashion because the rq weight of a
1487 * parent group depends on the shares of its child groups.
1490 tg_shares_up(struct task_group *tg, int cpu, struct sched_domain *sd)
1492 unsigned long rq_weight = 0;
1493 unsigned long shares = 0;
1496 for_each_cpu_mask(i, sd->span) {
1497 rq_weight += tg->cfs_rq[i]->load.weight;
1498 shares += tg->cfs_rq[i]->shares;
1501 if ((!shares && rq_weight) || shares > tg->shares)
1502 shares = tg->shares;
1504 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1505 shares = tg->shares;
1508 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1510 for_each_cpu_mask(i, sd->span) {
1511 struct rq *rq = cpu_rq(i);
1512 unsigned long flags;
1514 spin_lock_irqsave(&rq->lock, flags);
1515 __update_group_shares_cpu(tg, i, shares, rq_weight);
1516 spin_unlock_irqrestore(&rq->lock, flags);
1521 * Compute the cpu's hierarchical load factor for each task group.
1522 * This needs to be done in a top-down fashion because the load of a child
1523 * group is a fraction of its parents load.
1526 tg_load_down(struct task_group *tg, int cpu, struct sched_domain *sd)
1531 load = cpu_rq(cpu)->load.weight;
1533 load = tg->parent->cfs_rq[cpu]->h_load;
1534 load *= tg->cfs_rq[cpu]->shares;
1535 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1538 tg->cfs_rq[cpu]->h_load = load;
1542 tg_nop(struct task_group *tg, int cpu, struct sched_domain *sd)
1546 static void update_shares(struct sched_domain *sd)
1548 u64 now = cpu_clock(raw_smp_processor_id());
1549 s64 elapsed = now - sd->last_update;
1551 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1552 sd->last_update = now;
1553 walk_tg_tree(tg_nop, tg_shares_up, 0, sd);
1557 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1559 spin_unlock(&rq->lock);
1561 spin_lock(&rq->lock);
1564 static void update_h_load(int cpu)
1566 walk_tg_tree(tg_load_down, tg_nop, cpu, NULL);
1571 static inline void update_shares(struct sched_domain *sd)
1575 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1583 #ifdef CONFIG_FAIR_GROUP_SCHED
1584 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1587 cfs_rq->shares = shares;
1592 #include "sched_stats.h"
1593 #include "sched_idletask.c"
1594 #include "sched_fair.c"
1595 #include "sched_rt.c"
1596 #ifdef CONFIG_SCHED_DEBUG
1597 # include "sched_debug.c"
1600 #define sched_class_highest (&rt_sched_class)
1601 #define for_each_class(class) \
1602 for (class = sched_class_highest; class; class = class->next)
1604 static void inc_nr_running(struct rq *rq)
1609 static void dec_nr_running(struct rq *rq)
1614 static void set_load_weight(struct task_struct *p)
1616 if (task_has_rt_policy(p)) {
1617 p->se.load.weight = prio_to_weight[0] * 2;
1618 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1623 * SCHED_IDLE tasks get minimal weight:
1625 if (p->policy == SCHED_IDLE) {
1626 p->se.load.weight = WEIGHT_IDLEPRIO;
1627 p->se.load.inv_weight = WMULT_IDLEPRIO;
1631 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1632 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1635 static void update_avg(u64 *avg, u64 sample)
1637 s64 diff = sample - *avg;
1641 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1643 sched_info_queued(p);
1644 p->sched_class->enqueue_task(rq, p, wakeup);
1648 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1650 if (sleep && p->se.last_wakeup) {
1651 update_avg(&p->se.avg_overlap,
1652 p->se.sum_exec_runtime - p->se.last_wakeup);
1653 p->se.last_wakeup = 0;
1656 sched_info_dequeued(p);
1657 p->sched_class->dequeue_task(rq, p, sleep);
1662 * __normal_prio - return the priority that is based on the static prio
1664 static inline int __normal_prio(struct task_struct *p)
1666 return p->static_prio;
1670 * Calculate the expected normal priority: i.e. priority
1671 * without taking RT-inheritance into account. Might be
1672 * boosted by interactivity modifiers. Changes upon fork,
1673 * setprio syscalls, and whenever the interactivity
1674 * estimator recalculates.
1676 static inline int normal_prio(struct task_struct *p)
1680 if (task_has_rt_policy(p))
1681 prio = MAX_RT_PRIO-1 - p->rt_priority;
1683 prio = __normal_prio(p);
1688 * Calculate the current priority, i.e. the priority
1689 * taken into account by the scheduler. This value might
1690 * be boosted by RT tasks, or might be boosted by
1691 * interactivity modifiers. Will be RT if the task got
1692 * RT-boosted. If not then it returns p->normal_prio.
1694 static int effective_prio(struct task_struct *p)
1696 p->normal_prio = normal_prio(p);
1698 * If we are RT tasks or we were boosted to RT priority,
1699 * keep the priority unchanged. Otherwise, update priority
1700 * to the normal priority:
1702 if (!rt_prio(p->prio))
1703 return p->normal_prio;
1708 * activate_task - move a task to the runqueue.
1710 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1712 if (task_contributes_to_load(p))
1713 rq->nr_uninterruptible--;
1715 enqueue_task(rq, p, wakeup);
1720 * deactivate_task - remove a task from the runqueue.
1722 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1724 if (task_contributes_to_load(p))
1725 rq->nr_uninterruptible++;
1727 dequeue_task(rq, p, sleep);
1732 * task_curr - is this task currently executing on a CPU?
1733 * @p: the task in question.
1735 inline int task_curr(const struct task_struct *p)
1737 return cpu_curr(task_cpu(p)) == p;
1740 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1742 set_task_rq(p, cpu);
1745 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1746 * successfuly executed on another CPU. We must ensure that updates of
1747 * per-task data have been completed by this moment.
1750 task_thread_info(p)->cpu = cpu;
1754 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1755 const struct sched_class *prev_class,
1756 int oldprio, int running)
1758 if (prev_class != p->sched_class) {
1759 if (prev_class->switched_from)
1760 prev_class->switched_from(rq, p, running);
1761 p->sched_class->switched_to(rq, p, running);
1763 p->sched_class->prio_changed(rq, p, oldprio, running);
1768 /* Used instead of source_load when we know the type == 0 */
1769 static unsigned long weighted_cpuload(const int cpu)
1771 return cpu_rq(cpu)->load.weight;
1775 * Is this task likely cache-hot:
1778 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1783 * Buddy candidates are cache hot:
1785 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1788 if (p->sched_class != &fair_sched_class)
1791 if (sysctl_sched_migration_cost == -1)
1793 if (sysctl_sched_migration_cost == 0)
1796 delta = now - p->se.exec_start;
1798 return delta < (s64)sysctl_sched_migration_cost;
1802 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1804 int old_cpu = task_cpu(p);
1805 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1806 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1807 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1810 clock_offset = old_rq->clock - new_rq->clock;
1812 #ifdef CONFIG_SCHEDSTATS
1813 if (p->se.wait_start)
1814 p->se.wait_start -= clock_offset;
1815 if (p->se.sleep_start)
1816 p->se.sleep_start -= clock_offset;
1817 if (p->se.block_start)
1818 p->se.block_start -= clock_offset;
1819 if (old_cpu != new_cpu) {
1820 schedstat_inc(p, se.nr_migrations);
1821 if (task_hot(p, old_rq->clock, NULL))
1822 schedstat_inc(p, se.nr_forced2_migrations);
1825 p->se.vruntime -= old_cfsrq->min_vruntime -
1826 new_cfsrq->min_vruntime;
1828 __set_task_cpu(p, new_cpu);
1831 struct migration_req {
1832 struct list_head list;
1834 struct task_struct *task;
1837 struct completion done;
1841 * The task's runqueue lock must be held.
1842 * Returns true if you have to wait for migration thread.
1845 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1847 struct rq *rq = task_rq(p);
1850 * If the task is not on a runqueue (and not running), then
1851 * it is sufficient to simply update the task's cpu field.
1853 if (!p->se.on_rq && !task_running(rq, p)) {
1854 set_task_cpu(p, dest_cpu);
1858 init_completion(&req->done);
1860 req->dest_cpu = dest_cpu;
1861 list_add(&req->list, &rq->migration_queue);
1867 * wait_task_inactive - wait for a thread to unschedule.
1869 * If @match_state is nonzero, it's the @p->state value just checked and
1870 * not expected to change. If it changes, i.e. @p might have woken up,
1871 * then return zero. When we succeed in waiting for @p to be off its CPU,
1872 * we return a positive number (its total switch count). If a second call
1873 * a short while later returns the same number, the caller can be sure that
1874 * @p has remained unscheduled the whole time.
1876 * The caller must ensure that the task *will* unschedule sometime soon,
1877 * else this function might spin for a *long* time. This function can't
1878 * be called with interrupts off, or it may introduce deadlock with
1879 * smp_call_function() if an IPI is sent by the same process we are
1880 * waiting to become inactive.
1882 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1884 unsigned long flags;
1891 * We do the initial early heuristics without holding
1892 * any task-queue locks at all. We'll only try to get
1893 * the runqueue lock when things look like they will
1899 * If the task is actively running on another CPU
1900 * still, just relax and busy-wait without holding
1903 * NOTE! Since we don't hold any locks, it's not
1904 * even sure that "rq" stays as the right runqueue!
1905 * But we don't care, since "task_running()" will
1906 * return false if the runqueue has changed and p
1907 * is actually now running somewhere else!
1909 while (task_running(rq, p)) {
1910 if (match_state && unlikely(p->state != match_state))
1916 * Ok, time to look more closely! We need the rq
1917 * lock now, to be *sure*. If we're wrong, we'll
1918 * just go back and repeat.
1920 rq = task_rq_lock(p, &flags);
1921 running = task_running(rq, p);
1922 on_rq = p->se.on_rq;
1924 if (!match_state || p->state == match_state) {
1925 ncsw = p->nivcsw + p->nvcsw;
1926 if (unlikely(!ncsw))
1929 task_rq_unlock(rq, &flags);
1932 * If it changed from the expected state, bail out now.
1934 if (unlikely(!ncsw))
1938 * Was it really running after all now that we
1939 * checked with the proper locks actually held?
1941 * Oops. Go back and try again..
1943 if (unlikely(running)) {
1949 * It's not enough that it's not actively running,
1950 * it must be off the runqueue _entirely_, and not
1953 * So if it wa still runnable (but just not actively
1954 * running right now), it's preempted, and we should
1955 * yield - it could be a while.
1957 if (unlikely(on_rq)) {
1958 schedule_timeout_uninterruptible(1);
1963 * Ahh, all good. It wasn't running, and it wasn't
1964 * runnable, which means that it will never become
1965 * running in the future either. We're all done!
1974 * kick_process - kick a running thread to enter/exit the kernel
1975 * @p: the to-be-kicked thread
1977 * Cause a process which is running on another CPU to enter
1978 * kernel-mode, without any delay. (to get signals handled.)
1980 * NOTE: this function doesnt have to take the runqueue lock,
1981 * because all it wants to ensure is that the remote task enters
1982 * the kernel. If the IPI races and the task has been migrated
1983 * to another CPU then no harm is done and the purpose has been
1986 void kick_process(struct task_struct *p)
1992 if ((cpu != smp_processor_id()) && task_curr(p))
1993 smp_send_reschedule(cpu);
1998 * Return a low guess at the load of a migration-source cpu weighted
1999 * according to the scheduling class and "nice" value.
2001 * We want to under-estimate the load of migration sources, to
2002 * balance conservatively.
2004 static unsigned long source_load(int cpu, int type)
2006 struct rq *rq = cpu_rq(cpu);
2007 unsigned long total = weighted_cpuload(cpu);
2009 if (type == 0 || !sched_feat(LB_BIAS))
2012 return min(rq->cpu_load[type-1], total);
2016 * Return a high guess at the load of a migration-target cpu weighted
2017 * according to the scheduling class and "nice" value.
2019 static unsigned long target_load(int cpu, int type)
2021 struct rq *rq = cpu_rq(cpu);
2022 unsigned long total = weighted_cpuload(cpu);
2024 if (type == 0 || !sched_feat(LB_BIAS))
2027 return max(rq->cpu_load[type-1], total);
2031 * find_idlest_group finds and returns the least busy CPU group within the
2034 static struct sched_group *
2035 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2037 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2038 unsigned long min_load = ULONG_MAX, this_load = 0;
2039 int load_idx = sd->forkexec_idx;
2040 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2043 unsigned long load, avg_load;
2047 /* Skip over this group if it has no CPUs allowed */
2048 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2051 local_group = cpu_isset(this_cpu, group->cpumask);
2053 /* Tally up the load of all CPUs in the group */
2056 for_each_cpu_mask_nr(i, group->cpumask) {
2057 /* Bias balancing toward cpus of our domain */
2059 load = source_load(i, load_idx);
2061 load = target_load(i, load_idx);
2066 /* Adjust by relative CPU power of the group */
2067 avg_load = sg_div_cpu_power(group,
2068 avg_load * SCHED_LOAD_SCALE);
2071 this_load = avg_load;
2073 } else if (avg_load < min_load) {
2074 min_load = avg_load;
2077 } while (group = group->next, group != sd->groups);
2079 if (!idlest || 100*this_load < imbalance*min_load)
2085 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2088 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2091 unsigned long load, min_load = ULONG_MAX;
2095 /* Traverse only the allowed CPUs */
2096 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2098 for_each_cpu_mask_nr(i, *tmp) {
2099 load = weighted_cpuload(i);
2101 if (load < min_load || (load == min_load && i == this_cpu)) {
2111 * sched_balance_self: balance the current task (running on cpu) in domains
2112 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2115 * Balance, ie. select the least loaded group.
2117 * Returns the target CPU number, or the same CPU if no balancing is needed.
2119 * preempt must be disabled.
2121 static int sched_balance_self(int cpu, int flag)
2123 struct task_struct *t = current;
2124 struct sched_domain *tmp, *sd = NULL;
2126 for_each_domain(cpu, tmp) {
2128 * If power savings logic is enabled for a domain, stop there.
2130 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2132 if (tmp->flags & flag)
2140 cpumask_t span, tmpmask;
2141 struct sched_group *group;
2142 int new_cpu, weight;
2144 if (!(sd->flags & flag)) {
2150 group = find_idlest_group(sd, t, cpu);
2156 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2157 if (new_cpu == -1 || new_cpu == cpu) {
2158 /* Now try balancing at a lower domain level of cpu */
2163 /* Now try balancing at a lower domain level of new_cpu */
2166 weight = cpus_weight(span);
2167 for_each_domain(cpu, tmp) {
2168 if (weight <= cpus_weight(tmp->span))
2170 if (tmp->flags & flag)
2173 /* while loop will break here if sd == NULL */
2179 #endif /* CONFIG_SMP */
2182 * try_to_wake_up - wake up a thread
2183 * @p: the to-be-woken-up thread
2184 * @state: the mask of task states that can be woken
2185 * @sync: do a synchronous wakeup?
2187 * Put it on the run-queue if it's not already there. The "current"
2188 * thread is always on the run-queue (except when the actual
2189 * re-schedule is in progress), and as such you're allowed to do
2190 * the simpler "current->state = TASK_RUNNING" to mark yourself
2191 * runnable without the overhead of this.
2193 * returns failure only if the task is already active.
2195 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2197 int cpu, orig_cpu, this_cpu, success = 0;
2198 unsigned long flags;
2202 if (!sched_feat(SYNC_WAKEUPS))
2206 if (sched_feat(LB_WAKEUP_UPDATE)) {
2207 struct sched_domain *sd;
2209 this_cpu = raw_smp_processor_id();
2212 for_each_domain(this_cpu, sd) {
2213 if (cpu_isset(cpu, sd->span)) {
2222 rq = task_rq_lock(p, &flags);
2223 old_state = p->state;
2224 if (!(old_state & state))
2232 this_cpu = smp_processor_id();
2235 if (unlikely(task_running(rq, p)))
2238 cpu = p->sched_class->select_task_rq(p, sync);
2239 if (cpu != orig_cpu) {
2240 set_task_cpu(p, cpu);
2241 task_rq_unlock(rq, &flags);
2242 /* might preempt at this point */
2243 rq = task_rq_lock(p, &flags);
2244 old_state = p->state;
2245 if (!(old_state & state))
2250 this_cpu = smp_processor_id();
2254 #ifdef CONFIG_SCHEDSTATS
2255 schedstat_inc(rq, ttwu_count);
2256 if (cpu == this_cpu)
2257 schedstat_inc(rq, ttwu_local);
2259 struct sched_domain *sd;
2260 for_each_domain(this_cpu, sd) {
2261 if (cpu_isset(cpu, sd->span)) {
2262 schedstat_inc(sd, ttwu_wake_remote);
2267 #endif /* CONFIG_SCHEDSTATS */
2270 #endif /* CONFIG_SMP */
2271 schedstat_inc(p, se.nr_wakeups);
2273 schedstat_inc(p, se.nr_wakeups_sync);
2274 if (orig_cpu != cpu)
2275 schedstat_inc(p, se.nr_wakeups_migrate);
2276 if (cpu == this_cpu)
2277 schedstat_inc(p, se.nr_wakeups_local);
2279 schedstat_inc(p, se.nr_wakeups_remote);
2280 update_rq_clock(rq);
2281 activate_task(rq, p, 1);
2285 trace_mark(kernel_sched_wakeup,
2286 "pid %d state %ld ## rq %p task %p rq->curr %p",
2287 p->pid, p->state, rq, p, rq->curr);
2288 check_preempt_curr(rq, p);
2290 p->state = TASK_RUNNING;
2292 if (p->sched_class->task_wake_up)
2293 p->sched_class->task_wake_up(rq, p);
2296 current->se.last_wakeup = current->se.sum_exec_runtime;
2298 task_rq_unlock(rq, &flags);
2303 int wake_up_process(struct task_struct *p)
2305 return try_to_wake_up(p, TASK_ALL, 0);
2307 EXPORT_SYMBOL(wake_up_process);
2309 int wake_up_state(struct task_struct *p, unsigned int state)
2311 return try_to_wake_up(p, state, 0);
2315 * Perform scheduler related setup for a newly forked process p.
2316 * p is forked by current.
2318 * __sched_fork() is basic setup used by init_idle() too:
2320 static void __sched_fork(struct task_struct *p)
2322 p->se.exec_start = 0;
2323 p->se.sum_exec_runtime = 0;
2324 p->se.prev_sum_exec_runtime = 0;
2325 p->se.last_wakeup = 0;
2326 p->se.avg_overlap = 0;
2328 #ifdef CONFIG_SCHEDSTATS
2329 p->se.wait_start = 0;
2330 p->se.sum_sleep_runtime = 0;
2331 p->se.sleep_start = 0;
2332 p->se.block_start = 0;
2333 p->se.sleep_max = 0;
2334 p->se.block_max = 0;
2336 p->se.slice_max = 0;
2340 INIT_LIST_HEAD(&p->rt.run_list);
2342 INIT_LIST_HEAD(&p->se.group_node);
2344 #ifdef CONFIG_PREEMPT_NOTIFIERS
2345 INIT_HLIST_HEAD(&p->preempt_notifiers);
2349 * We mark the process as running here, but have not actually
2350 * inserted it onto the runqueue yet. This guarantees that
2351 * nobody will actually run it, and a signal or other external
2352 * event cannot wake it up and insert it on the runqueue either.
2354 p->state = TASK_RUNNING;
2358 * fork()/clone()-time setup:
2360 void sched_fork(struct task_struct *p, int clone_flags)
2362 int cpu = get_cpu();
2367 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2369 set_task_cpu(p, cpu);
2372 * Make sure we do not leak PI boosting priority to the child:
2374 p->prio = current->normal_prio;
2375 if (!rt_prio(p->prio))
2376 p->sched_class = &fair_sched_class;
2378 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2379 if (likely(sched_info_on()))
2380 memset(&p->sched_info, 0, sizeof(p->sched_info));
2382 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2385 #ifdef CONFIG_PREEMPT
2386 /* Want to start with kernel preemption disabled. */
2387 task_thread_info(p)->preempt_count = 1;
2393 * wake_up_new_task - wake up a newly created task for the first time.
2395 * This function will do some initial scheduler statistics housekeeping
2396 * that must be done for every newly created context, then puts the task
2397 * on the runqueue and wakes it.
2399 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2401 unsigned long flags;
2404 rq = task_rq_lock(p, &flags);
2405 BUG_ON(p->state != TASK_RUNNING);
2406 update_rq_clock(rq);
2408 p->prio = effective_prio(p);
2410 if (!p->sched_class->task_new || !current->se.on_rq) {
2411 activate_task(rq, p, 0);
2414 * Let the scheduling class do new task startup
2415 * management (if any):
2417 p->sched_class->task_new(rq, p);
2420 trace_mark(kernel_sched_wakeup_new,
2421 "pid %d state %ld ## rq %p task %p rq->curr %p",
2422 p->pid, p->state, rq, p, rq->curr);
2423 check_preempt_curr(rq, p);
2425 if (p->sched_class->task_wake_up)
2426 p->sched_class->task_wake_up(rq, p);
2428 task_rq_unlock(rq, &flags);
2431 #ifdef CONFIG_PREEMPT_NOTIFIERS
2434 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2435 * @notifier: notifier struct to register
2437 void preempt_notifier_register(struct preempt_notifier *notifier)
2439 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2441 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2444 * preempt_notifier_unregister - no longer interested in preemption notifications
2445 * @notifier: notifier struct to unregister
2447 * This is safe to call from within a preemption notifier.
2449 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2451 hlist_del(¬ifier->link);
2453 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2455 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2457 struct preempt_notifier *notifier;
2458 struct hlist_node *node;
2460 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2461 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2465 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2466 struct task_struct *next)
2468 struct preempt_notifier *notifier;
2469 struct hlist_node *node;
2471 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2472 notifier->ops->sched_out(notifier, next);
2475 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2477 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2482 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2483 struct task_struct *next)
2487 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2490 * prepare_task_switch - prepare to switch tasks
2491 * @rq: the runqueue preparing to switch
2492 * @prev: the current task that is being switched out
2493 * @next: the task we are going to switch to.
2495 * This is called with the rq lock held and interrupts off. It must
2496 * be paired with a subsequent finish_task_switch after the context
2499 * prepare_task_switch sets up locking and calls architecture specific
2503 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2504 struct task_struct *next)
2506 fire_sched_out_preempt_notifiers(prev, next);
2507 prepare_lock_switch(rq, next);
2508 prepare_arch_switch(next);
2512 * finish_task_switch - clean up after a task-switch
2513 * @rq: runqueue associated with task-switch
2514 * @prev: the thread we just switched away from.
2516 * finish_task_switch must be called after the context switch, paired
2517 * with a prepare_task_switch call before the context switch.
2518 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2519 * and do any other architecture-specific cleanup actions.
2521 * Note that we may have delayed dropping an mm in context_switch(). If
2522 * so, we finish that here outside of the runqueue lock. (Doing it
2523 * with the lock held can cause deadlocks; see schedule() for
2526 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2527 __releases(rq->lock)
2529 struct mm_struct *mm = rq->prev_mm;
2535 * A task struct has one reference for the use as "current".
2536 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2537 * schedule one last time. The schedule call will never return, and
2538 * the scheduled task must drop that reference.
2539 * The test for TASK_DEAD must occur while the runqueue locks are
2540 * still held, otherwise prev could be scheduled on another cpu, die
2541 * there before we look at prev->state, and then the reference would
2543 * Manfred Spraul <manfred@colorfullife.com>
2545 prev_state = prev->state;
2546 finish_arch_switch(prev);
2547 finish_lock_switch(rq, prev);
2549 if (current->sched_class->post_schedule)
2550 current->sched_class->post_schedule(rq);
2553 fire_sched_in_preempt_notifiers(current);
2556 if (unlikely(prev_state == TASK_DEAD)) {
2558 * Remove function-return probe instances associated with this
2559 * task and put them back on the free list.
2561 kprobe_flush_task(prev);
2562 put_task_struct(prev);
2567 * schedule_tail - first thing a freshly forked thread must call.
2568 * @prev: the thread we just switched away from.
2570 asmlinkage void schedule_tail(struct task_struct *prev)
2571 __releases(rq->lock)
2573 struct rq *rq = this_rq();
2575 finish_task_switch(rq, prev);
2576 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2577 /* In this case, finish_task_switch does not reenable preemption */
2580 if (current->set_child_tid)
2581 put_user(task_pid_vnr(current), current->set_child_tid);
2585 * context_switch - switch to the new MM and the new
2586 * thread's register state.
2589 context_switch(struct rq *rq, struct task_struct *prev,
2590 struct task_struct *next)
2592 struct mm_struct *mm, *oldmm;
2594 prepare_task_switch(rq, prev, next);
2595 trace_mark(kernel_sched_schedule,
2596 "prev_pid %d next_pid %d prev_state %ld "
2597 "## rq %p prev %p next %p",
2598 prev->pid, next->pid, prev->state,
2601 oldmm = prev->active_mm;
2603 * For paravirt, this is coupled with an exit in switch_to to
2604 * combine the page table reload and the switch backend into
2607 arch_enter_lazy_cpu_mode();
2609 if (unlikely(!mm)) {
2610 next->active_mm = oldmm;
2611 atomic_inc(&oldmm->mm_count);
2612 enter_lazy_tlb(oldmm, next);
2614 switch_mm(oldmm, mm, next);
2616 if (unlikely(!prev->mm)) {
2617 prev->active_mm = NULL;
2618 rq->prev_mm = oldmm;
2621 * Since the runqueue lock will be released by the next
2622 * task (which is an invalid locking op but in the case
2623 * of the scheduler it's an obvious special-case), so we
2624 * do an early lockdep release here:
2626 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2627 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2630 /* Here we just switch the register state and the stack. */
2631 switch_to(prev, next, prev);
2635 * this_rq must be evaluated again because prev may have moved
2636 * CPUs since it called schedule(), thus the 'rq' on its stack
2637 * frame will be invalid.
2639 finish_task_switch(this_rq(), prev);
2643 * nr_running, nr_uninterruptible and nr_context_switches:
2645 * externally visible scheduler statistics: current number of runnable
2646 * threads, current number of uninterruptible-sleeping threads, total
2647 * number of context switches performed since bootup.
2649 unsigned long nr_running(void)
2651 unsigned long i, sum = 0;
2653 for_each_online_cpu(i)
2654 sum += cpu_rq(i)->nr_running;
2659 unsigned long nr_uninterruptible(void)
2661 unsigned long i, sum = 0;
2663 for_each_possible_cpu(i)
2664 sum += cpu_rq(i)->nr_uninterruptible;
2667 * Since we read the counters lockless, it might be slightly
2668 * inaccurate. Do not allow it to go below zero though:
2670 if (unlikely((long)sum < 0))
2676 unsigned long long nr_context_switches(void)
2679 unsigned long long sum = 0;
2681 for_each_possible_cpu(i)
2682 sum += cpu_rq(i)->nr_switches;
2687 unsigned long nr_iowait(void)
2689 unsigned long i, sum = 0;
2691 for_each_possible_cpu(i)
2692 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2697 unsigned long nr_active(void)
2699 unsigned long i, running = 0, uninterruptible = 0;
2701 for_each_online_cpu(i) {
2702 running += cpu_rq(i)->nr_running;
2703 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2706 if (unlikely((long)uninterruptible < 0))
2707 uninterruptible = 0;
2709 return running + uninterruptible;
2713 * Update rq->cpu_load[] statistics. This function is usually called every
2714 * scheduler tick (TICK_NSEC).
2716 static void update_cpu_load(struct rq *this_rq)
2718 unsigned long this_load = this_rq->load.weight;
2721 this_rq->nr_load_updates++;
2723 /* Update our load: */
2724 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2725 unsigned long old_load, new_load;
2727 /* scale is effectively 1 << i now, and >> i divides by scale */
2729 old_load = this_rq->cpu_load[i];
2730 new_load = this_load;
2732 * Round up the averaging division if load is increasing. This
2733 * prevents us from getting stuck on 9 if the load is 10, for
2736 if (new_load > old_load)
2737 new_load += scale-1;
2738 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2745 * double_rq_lock - safely lock two runqueues
2747 * Note this does not disable interrupts like task_rq_lock,
2748 * you need to do so manually before calling.
2750 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2751 __acquires(rq1->lock)
2752 __acquires(rq2->lock)
2754 BUG_ON(!irqs_disabled());
2756 spin_lock(&rq1->lock);
2757 __acquire(rq2->lock); /* Fake it out ;) */
2760 spin_lock(&rq1->lock);
2761 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2763 spin_lock(&rq2->lock);
2764 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2767 update_rq_clock(rq1);
2768 update_rq_clock(rq2);
2772 * double_rq_unlock - safely unlock two runqueues
2774 * Note this does not restore interrupts like task_rq_unlock,
2775 * you need to do so manually after calling.
2777 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2778 __releases(rq1->lock)
2779 __releases(rq2->lock)
2781 spin_unlock(&rq1->lock);
2783 spin_unlock(&rq2->lock);
2785 __release(rq2->lock);
2789 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2791 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2792 __releases(this_rq->lock)
2793 __acquires(busiest->lock)
2794 __acquires(this_rq->lock)
2798 if (unlikely(!irqs_disabled())) {
2799 /* printk() doesn't work good under rq->lock */
2800 spin_unlock(&this_rq->lock);
2803 if (unlikely(!spin_trylock(&busiest->lock))) {
2804 if (busiest < this_rq) {
2805 spin_unlock(&this_rq->lock);
2806 spin_lock(&busiest->lock);
2807 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2810 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2815 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2816 __releases(busiest->lock)
2818 spin_unlock(&busiest->lock);
2819 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2823 * If dest_cpu is allowed for this process, migrate the task to it.
2824 * This is accomplished by forcing the cpu_allowed mask to only
2825 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2826 * the cpu_allowed mask is restored.
2828 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2830 struct migration_req req;
2831 unsigned long flags;
2834 rq = task_rq_lock(p, &flags);
2835 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2836 || unlikely(!cpu_active(dest_cpu)))
2839 /* force the process onto the specified CPU */
2840 if (migrate_task(p, dest_cpu, &req)) {
2841 /* Need to wait for migration thread (might exit: take ref). */
2842 struct task_struct *mt = rq->migration_thread;
2844 get_task_struct(mt);
2845 task_rq_unlock(rq, &flags);
2846 wake_up_process(mt);
2847 put_task_struct(mt);
2848 wait_for_completion(&req.done);
2853 task_rq_unlock(rq, &flags);
2857 * sched_exec - execve() is a valuable balancing opportunity, because at
2858 * this point the task has the smallest effective memory and cache footprint.
2860 void sched_exec(void)
2862 int new_cpu, this_cpu = get_cpu();
2863 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2865 if (new_cpu != this_cpu)
2866 sched_migrate_task(current, new_cpu);
2870 * pull_task - move a task from a remote runqueue to the local runqueue.
2871 * Both runqueues must be locked.
2873 static void pull_task(struct rq *src_rq, struct task_struct *p,
2874 struct rq *this_rq, int this_cpu)
2876 deactivate_task(src_rq, p, 0);
2877 set_task_cpu(p, this_cpu);
2878 activate_task(this_rq, p, 0);
2880 * Note that idle threads have a prio of MAX_PRIO, for this test
2881 * to be always true for them.
2883 check_preempt_curr(this_rq, p);
2887 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2890 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2891 struct sched_domain *sd, enum cpu_idle_type idle,
2895 * We do not migrate tasks that are:
2896 * 1) running (obviously), or
2897 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2898 * 3) are cache-hot on their current CPU.
2900 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2901 schedstat_inc(p, se.nr_failed_migrations_affine);
2906 if (task_running(rq, p)) {
2907 schedstat_inc(p, se.nr_failed_migrations_running);
2912 * Aggressive migration if:
2913 * 1) task is cache cold, or
2914 * 2) too many balance attempts have failed.
2917 if (!task_hot(p, rq->clock, sd) ||
2918 sd->nr_balance_failed > sd->cache_nice_tries) {
2919 #ifdef CONFIG_SCHEDSTATS
2920 if (task_hot(p, rq->clock, sd)) {
2921 schedstat_inc(sd, lb_hot_gained[idle]);
2922 schedstat_inc(p, se.nr_forced_migrations);
2928 if (task_hot(p, rq->clock, sd)) {
2929 schedstat_inc(p, se.nr_failed_migrations_hot);
2935 static unsigned long
2936 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2937 unsigned long max_load_move, struct sched_domain *sd,
2938 enum cpu_idle_type idle, int *all_pinned,
2939 int *this_best_prio, struct rq_iterator *iterator)
2941 int loops = 0, pulled = 0, pinned = 0;
2942 struct task_struct *p;
2943 long rem_load_move = max_load_move;
2945 if (max_load_move == 0)
2951 * Start the load-balancing iterator:
2953 p = iterator->start(iterator->arg);
2955 if (!p || loops++ > sysctl_sched_nr_migrate)
2958 if ((p->se.load.weight >> 1) > rem_load_move ||
2959 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2960 p = iterator->next(iterator->arg);
2964 pull_task(busiest, p, this_rq, this_cpu);
2966 rem_load_move -= p->se.load.weight;
2969 * We only want to steal up to the prescribed amount of weighted load.
2971 if (rem_load_move > 0) {
2972 if (p->prio < *this_best_prio)
2973 *this_best_prio = p->prio;
2974 p = iterator->next(iterator->arg);
2979 * Right now, this is one of only two places pull_task() is called,
2980 * so we can safely collect pull_task() stats here rather than
2981 * inside pull_task().
2983 schedstat_add(sd, lb_gained[idle], pulled);
2986 *all_pinned = pinned;
2988 return max_load_move - rem_load_move;
2992 * move_tasks tries to move up to max_load_move weighted load from busiest to
2993 * this_rq, as part of a balancing operation within domain "sd".
2994 * Returns 1 if successful and 0 otherwise.
2996 * Called with both runqueues locked.
2998 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2999 unsigned long max_load_move,
3000 struct sched_domain *sd, enum cpu_idle_type idle,
3003 const struct sched_class *class = sched_class_highest;
3004 unsigned long total_load_moved = 0;
3005 int this_best_prio = this_rq->curr->prio;
3009 class->load_balance(this_rq, this_cpu, busiest,
3010 max_load_move - total_load_moved,
3011 sd, idle, all_pinned, &this_best_prio);
3012 class = class->next;
3014 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3017 } while (class && max_load_move > total_load_moved);
3019 return total_load_moved > 0;
3023 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3024 struct sched_domain *sd, enum cpu_idle_type idle,
3025 struct rq_iterator *iterator)
3027 struct task_struct *p = iterator->start(iterator->arg);
3031 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3032 pull_task(busiest, p, this_rq, this_cpu);
3034 * Right now, this is only the second place pull_task()
3035 * is called, so we can safely collect pull_task()
3036 * stats here rather than inside pull_task().
3038 schedstat_inc(sd, lb_gained[idle]);
3042 p = iterator->next(iterator->arg);
3049 * move_one_task tries to move exactly one task from busiest to this_rq, as
3050 * part of active balancing operations within "domain".
3051 * Returns 1 if successful and 0 otherwise.
3053 * Called with both runqueues locked.
3055 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3056 struct sched_domain *sd, enum cpu_idle_type idle)
3058 const struct sched_class *class;
3060 for (class = sched_class_highest; class; class = class->next)
3061 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3068 * find_busiest_group finds and returns the busiest CPU group within the
3069 * domain. It calculates and returns the amount of weighted load which
3070 * should be moved to restore balance via the imbalance parameter.
3072 static struct sched_group *
3073 find_busiest_group(struct sched_domain *sd, int this_cpu,
3074 unsigned long *imbalance, enum cpu_idle_type idle,
3075 int *sd_idle, const cpumask_t *cpus, int *balance)
3077 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3078 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3079 unsigned long max_pull;
3080 unsigned long busiest_load_per_task, busiest_nr_running;
3081 unsigned long this_load_per_task, this_nr_running;
3082 int load_idx, group_imb = 0;
3083 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3084 int power_savings_balance = 1;
3085 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3086 unsigned long min_nr_running = ULONG_MAX;
3087 struct sched_group *group_min = NULL, *group_leader = NULL;
3090 max_load = this_load = total_load = total_pwr = 0;
3091 busiest_load_per_task = busiest_nr_running = 0;
3092 this_load_per_task = this_nr_running = 0;
3094 if (idle == CPU_NOT_IDLE)
3095 load_idx = sd->busy_idx;
3096 else if (idle == CPU_NEWLY_IDLE)
3097 load_idx = sd->newidle_idx;
3099 load_idx = sd->idle_idx;
3102 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3105 int __group_imb = 0;
3106 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3107 unsigned long sum_nr_running, sum_weighted_load;
3108 unsigned long sum_avg_load_per_task;
3109 unsigned long avg_load_per_task;
3111 local_group = cpu_isset(this_cpu, group->cpumask);
3114 balance_cpu = first_cpu(group->cpumask);
3116 /* Tally up the load of all CPUs in the group */
3117 sum_weighted_load = sum_nr_running = avg_load = 0;
3118 sum_avg_load_per_task = avg_load_per_task = 0;
3121 min_cpu_load = ~0UL;
3123 for_each_cpu_mask_nr(i, group->cpumask) {
3126 if (!cpu_isset(i, *cpus))
3131 if (*sd_idle && rq->nr_running)
3134 /* Bias balancing toward cpus of our domain */
3136 if (idle_cpu(i) && !first_idle_cpu) {
3141 load = target_load(i, load_idx);
3143 load = source_load(i, load_idx);
3144 if (load > max_cpu_load)
3145 max_cpu_load = load;
3146 if (min_cpu_load > load)
3147 min_cpu_load = load;
3151 sum_nr_running += rq->nr_running;
3152 sum_weighted_load += weighted_cpuload(i);
3154 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3158 * First idle cpu or the first cpu(busiest) in this sched group
3159 * is eligible for doing load balancing at this and above
3160 * domains. In the newly idle case, we will allow all the cpu's
3161 * to do the newly idle load balance.
3163 if (idle != CPU_NEWLY_IDLE && local_group &&
3164 balance_cpu != this_cpu && balance) {
3169 total_load += avg_load;
3170 total_pwr += group->__cpu_power;
3172 /* Adjust by relative CPU power of the group */
3173 avg_load = sg_div_cpu_power(group,
3174 avg_load * SCHED_LOAD_SCALE);
3178 * Consider the group unbalanced when the imbalance is larger
3179 * than the average weight of two tasks.
3181 * APZ: with cgroup the avg task weight can vary wildly and
3182 * might not be a suitable number - should we keep a
3183 * normalized nr_running number somewhere that negates
3186 avg_load_per_task = sg_div_cpu_power(group,
3187 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3189 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3192 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3195 this_load = avg_load;
3197 this_nr_running = sum_nr_running;
3198 this_load_per_task = sum_weighted_load;
3199 } else if (avg_load > max_load &&
3200 (sum_nr_running > group_capacity || __group_imb)) {
3201 max_load = avg_load;
3203 busiest_nr_running = sum_nr_running;
3204 busiest_load_per_task = sum_weighted_load;
3205 group_imb = __group_imb;
3208 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3210 * Busy processors will not participate in power savings
3213 if (idle == CPU_NOT_IDLE ||
3214 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3218 * If the local group is idle or completely loaded
3219 * no need to do power savings balance at this domain
3221 if (local_group && (this_nr_running >= group_capacity ||
3223 power_savings_balance = 0;
3226 * If a group is already running at full capacity or idle,
3227 * don't include that group in power savings calculations
3229 if (!power_savings_balance || sum_nr_running >= group_capacity
3234 * Calculate the group which has the least non-idle load.
3235 * This is the group from where we need to pick up the load
3238 if ((sum_nr_running < min_nr_running) ||
3239 (sum_nr_running == min_nr_running &&
3240 first_cpu(group->cpumask) <
3241 first_cpu(group_min->cpumask))) {
3243 min_nr_running = sum_nr_running;
3244 min_load_per_task = sum_weighted_load /
3249 * Calculate the group which is almost near its
3250 * capacity but still has some space to pick up some load
3251 * from other group and save more power
3253 if (sum_nr_running <= group_capacity - 1) {
3254 if (sum_nr_running > leader_nr_running ||
3255 (sum_nr_running == leader_nr_running &&
3256 first_cpu(group->cpumask) >
3257 first_cpu(group_leader->cpumask))) {
3258 group_leader = group;
3259 leader_nr_running = sum_nr_running;
3264 group = group->next;
3265 } while (group != sd->groups);
3267 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3270 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3272 if (this_load >= avg_load ||
3273 100*max_load <= sd->imbalance_pct*this_load)
3276 busiest_load_per_task /= busiest_nr_running;
3278 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3281 * We're trying to get all the cpus to the average_load, so we don't
3282 * want to push ourselves above the average load, nor do we wish to
3283 * reduce the max loaded cpu below the average load, as either of these
3284 * actions would just result in more rebalancing later, and ping-pong
3285 * tasks around. Thus we look for the minimum possible imbalance.
3286 * Negative imbalances (*we* are more loaded than anyone else) will
3287 * be counted as no imbalance for these purposes -- we can't fix that
3288 * by pulling tasks to us. Be careful of negative numbers as they'll
3289 * appear as very large values with unsigned longs.
3291 if (max_load <= busiest_load_per_task)
3295 * In the presence of smp nice balancing, certain scenarios can have
3296 * max load less than avg load(as we skip the groups at or below
3297 * its cpu_power, while calculating max_load..)
3299 if (max_load < avg_load) {
3301 goto small_imbalance;
3304 /* Don't want to pull so many tasks that a group would go idle */
3305 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3307 /* How much load to actually move to equalise the imbalance */
3308 *imbalance = min(max_pull * busiest->__cpu_power,
3309 (avg_load - this_load) * this->__cpu_power)
3313 * if *imbalance is less than the average load per runnable task
3314 * there is no gaurantee that any tasks will be moved so we'll have
3315 * a think about bumping its value to force at least one task to be
3318 if (*imbalance < busiest_load_per_task) {
3319 unsigned long tmp, pwr_now, pwr_move;
3323 pwr_move = pwr_now = 0;
3325 if (this_nr_running) {
3326 this_load_per_task /= this_nr_running;
3327 if (busiest_load_per_task > this_load_per_task)
3330 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3332 if (max_load - this_load + 2*busiest_load_per_task >=
3333 busiest_load_per_task * imbn) {
3334 *imbalance = busiest_load_per_task;
3339 * OK, we don't have enough imbalance to justify moving tasks,
3340 * however we may be able to increase total CPU power used by
3344 pwr_now += busiest->__cpu_power *
3345 min(busiest_load_per_task, max_load);
3346 pwr_now += this->__cpu_power *
3347 min(this_load_per_task, this_load);
3348 pwr_now /= SCHED_LOAD_SCALE;
3350 /* Amount of load we'd subtract */
3351 tmp = sg_div_cpu_power(busiest,
3352 busiest_load_per_task * SCHED_LOAD_SCALE);
3354 pwr_move += busiest->__cpu_power *
3355 min(busiest_load_per_task, max_load - tmp);
3357 /* Amount of load we'd add */
3358 if (max_load * busiest->__cpu_power <
3359 busiest_load_per_task * SCHED_LOAD_SCALE)
3360 tmp = sg_div_cpu_power(this,
3361 max_load * busiest->__cpu_power);
3363 tmp = sg_div_cpu_power(this,
3364 busiest_load_per_task * SCHED_LOAD_SCALE);
3365 pwr_move += this->__cpu_power *
3366 min(this_load_per_task, this_load + tmp);
3367 pwr_move /= SCHED_LOAD_SCALE;
3369 /* Move if we gain throughput */
3370 if (pwr_move > pwr_now)
3371 *imbalance = busiest_load_per_task;
3377 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3378 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3381 if (this == group_leader && group_leader != group_min) {
3382 *imbalance = min_load_per_task;
3392 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3395 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3396 unsigned long imbalance, const cpumask_t *cpus)
3398 struct rq *busiest = NULL, *rq;
3399 unsigned long max_load = 0;
3402 for_each_cpu_mask_nr(i, group->cpumask) {
3405 if (!cpu_isset(i, *cpus))
3409 wl = weighted_cpuload(i);
3411 if (rq->nr_running == 1 && wl > imbalance)
3414 if (wl > max_load) {
3424 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3425 * so long as it is large enough.
3427 #define MAX_PINNED_INTERVAL 512
3430 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3431 * tasks if there is an imbalance.
3433 static int load_balance(int this_cpu, struct rq *this_rq,
3434 struct sched_domain *sd, enum cpu_idle_type idle,
3435 int *balance, cpumask_t *cpus)
3437 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3438 struct sched_group *group;
3439 unsigned long imbalance;
3441 unsigned long flags;
3446 * When power savings policy is enabled for the parent domain, idle
3447 * sibling can pick up load irrespective of busy siblings. In this case,
3448 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3449 * portraying it as CPU_NOT_IDLE.
3451 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3452 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3455 schedstat_inc(sd, lb_count[idle]);
3459 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3466 schedstat_inc(sd, lb_nobusyg[idle]);
3470 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3472 schedstat_inc(sd, lb_nobusyq[idle]);
3476 BUG_ON(busiest == this_rq);
3478 schedstat_add(sd, lb_imbalance[idle], imbalance);
3481 if (busiest->nr_running > 1) {
3483 * Attempt to move tasks. If find_busiest_group has found
3484 * an imbalance but busiest->nr_running <= 1, the group is
3485 * still unbalanced. ld_moved simply stays zero, so it is
3486 * correctly treated as an imbalance.
3488 local_irq_save(flags);
3489 double_rq_lock(this_rq, busiest);
3490 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3491 imbalance, sd, idle, &all_pinned);
3492 double_rq_unlock(this_rq, busiest);
3493 local_irq_restore(flags);
3496 * some other cpu did the load balance for us.
3498 if (ld_moved && this_cpu != smp_processor_id())
3499 resched_cpu(this_cpu);
3501 /* All tasks on this runqueue were pinned by CPU affinity */
3502 if (unlikely(all_pinned)) {
3503 cpu_clear(cpu_of(busiest), *cpus);
3504 if (!cpus_empty(*cpus))
3511 schedstat_inc(sd, lb_failed[idle]);
3512 sd->nr_balance_failed++;
3514 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3516 spin_lock_irqsave(&busiest->lock, flags);
3518 /* don't kick the migration_thread, if the curr
3519 * task on busiest cpu can't be moved to this_cpu
3521 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3522 spin_unlock_irqrestore(&busiest->lock, flags);
3524 goto out_one_pinned;
3527 if (!busiest->active_balance) {
3528 busiest->active_balance = 1;
3529 busiest->push_cpu = this_cpu;
3532 spin_unlock_irqrestore(&busiest->lock, flags);
3534 wake_up_process(busiest->migration_thread);
3537 * We've kicked active balancing, reset the failure
3540 sd->nr_balance_failed = sd->cache_nice_tries+1;
3543 sd->nr_balance_failed = 0;
3545 if (likely(!active_balance)) {
3546 /* We were unbalanced, so reset the balancing interval */
3547 sd->balance_interval = sd->min_interval;
3550 * If we've begun active balancing, start to back off. This
3551 * case may not be covered by the all_pinned logic if there
3552 * is only 1 task on the busy runqueue (because we don't call
3555 if (sd->balance_interval < sd->max_interval)
3556 sd->balance_interval *= 2;
3559 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3560 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3566 schedstat_inc(sd, lb_balanced[idle]);
3568 sd->nr_balance_failed = 0;
3571 /* tune up the balancing interval */
3572 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3573 (sd->balance_interval < sd->max_interval))
3574 sd->balance_interval *= 2;
3576 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3577 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3588 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3589 * tasks if there is an imbalance.
3591 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3592 * this_rq is locked.
3595 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3598 struct sched_group *group;
3599 struct rq *busiest = NULL;
3600 unsigned long imbalance;
3608 * When power savings policy is enabled for the parent domain, idle
3609 * sibling can pick up load irrespective of busy siblings. In this case,
3610 * let the state of idle sibling percolate up as IDLE, instead of
3611 * portraying it as CPU_NOT_IDLE.
3613 if (sd->flags & SD_SHARE_CPUPOWER &&
3614 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3617 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3619 update_shares_locked(this_rq, sd);
3620 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3621 &sd_idle, cpus, NULL);
3623 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3627 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3629 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3633 BUG_ON(busiest == this_rq);
3635 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3638 if (busiest->nr_running > 1) {
3639 /* Attempt to move tasks */
3640 double_lock_balance(this_rq, busiest);
3641 /* this_rq->clock is already updated */
3642 update_rq_clock(busiest);
3643 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3644 imbalance, sd, CPU_NEWLY_IDLE,
3646 double_unlock_balance(this_rq, busiest);
3648 if (unlikely(all_pinned)) {
3649 cpu_clear(cpu_of(busiest), *cpus);
3650 if (!cpus_empty(*cpus))
3656 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3657 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3658 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3661 sd->nr_balance_failed = 0;
3663 update_shares_locked(this_rq, sd);
3667 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3668 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3669 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3671 sd->nr_balance_failed = 0;
3677 * idle_balance is called by schedule() if this_cpu is about to become
3678 * idle. Attempts to pull tasks from other CPUs.
3680 static void idle_balance(int this_cpu, struct rq *this_rq)
3682 struct sched_domain *sd;
3683 int pulled_task = -1;
3684 unsigned long next_balance = jiffies + HZ;
3687 for_each_domain(this_cpu, sd) {
3688 unsigned long interval;
3690 if (!(sd->flags & SD_LOAD_BALANCE))
3693 if (sd->flags & SD_BALANCE_NEWIDLE)
3694 /* If we've pulled tasks over stop searching: */
3695 pulled_task = load_balance_newidle(this_cpu, this_rq,
3698 interval = msecs_to_jiffies(sd->balance_interval);
3699 if (time_after(next_balance, sd->last_balance + interval))
3700 next_balance = sd->last_balance + interval;
3704 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3706 * We are going idle. next_balance may be set based on
3707 * a busy processor. So reset next_balance.
3709 this_rq->next_balance = next_balance;
3714 * active_load_balance is run by migration threads. It pushes running tasks
3715 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3716 * running on each physical CPU where possible, and avoids physical /
3717 * logical imbalances.
3719 * Called with busiest_rq locked.
3721 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3723 int target_cpu = busiest_rq->push_cpu;
3724 struct sched_domain *sd;
3725 struct rq *target_rq;
3727 /* Is there any task to move? */
3728 if (busiest_rq->nr_running <= 1)
3731 target_rq = cpu_rq(target_cpu);
3734 * This condition is "impossible", if it occurs
3735 * we need to fix it. Originally reported by
3736 * Bjorn Helgaas on a 128-cpu setup.
3738 BUG_ON(busiest_rq == target_rq);
3740 /* move a task from busiest_rq to target_rq */
3741 double_lock_balance(busiest_rq, target_rq);
3742 update_rq_clock(busiest_rq);
3743 update_rq_clock(target_rq);
3745 /* Search for an sd spanning us and the target CPU. */
3746 for_each_domain(target_cpu, sd) {
3747 if ((sd->flags & SD_LOAD_BALANCE) &&
3748 cpu_isset(busiest_cpu, sd->span))
3753 schedstat_inc(sd, alb_count);
3755 if (move_one_task(target_rq, target_cpu, busiest_rq,
3757 schedstat_inc(sd, alb_pushed);
3759 schedstat_inc(sd, alb_failed);
3761 double_unlock_balance(busiest_rq, target_rq);
3766 atomic_t load_balancer;
3768 } nohz ____cacheline_aligned = {
3769 .load_balancer = ATOMIC_INIT(-1),
3770 .cpu_mask = CPU_MASK_NONE,
3774 * This routine will try to nominate the ilb (idle load balancing)
3775 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3776 * load balancing on behalf of all those cpus. If all the cpus in the system
3777 * go into this tickless mode, then there will be no ilb owner (as there is
3778 * no need for one) and all the cpus will sleep till the next wakeup event
3781 * For the ilb owner, tick is not stopped. And this tick will be used
3782 * for idle load balancing. ilb owner will still be part of
3785 * While stopping the tick, this cpu will become the ilb owner if there
3786 * is no other owner. And will be the owner till that cpu becomes busy
3787 * or if all cpus in the system stop their ticks at which point
3788 * there is no need for ilb owner.
3790 * When the ilb owner becomes busy, it nominates another owner, during the
3791 * next busy scheduler_tick()
3793 int select_nohz_load_balancer(int stop_tick)
3795 int cpu = smp_processor_id();
3798 cpu_set(cpu, nohz.cpu_mask);
3799 cpu_rq(cpu)->in_nohz_recently = 1;
3802 * If we are going offline and still the leader, give up!
3804 if (!cpu_active(cpu) &&
3805 atomic_read(&nohz.load_balancer) == cpu) {
3806 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3811 /* time for ilb owner also to sleep */
3812 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3813 if (atomic_read(&nohz.load_balancer) == cpu)
3814 atomic_set(&nohz.load_balancer, -1);
3818 if (atomic_read(&nohz.load_balancer) == -1) {
3819 /* make me the ilb owner */
3820 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3822 } else if (atomic_read(&nohz.load_balancer) == cpu)
3825 if (!cpu_isset(cpu, nohz.cpu_mask))
3828 cpu_clear(cpu, nohz.cpu_mask);
3830 if (atomic_read(&nohz.load_balancer) == cpu)
3831 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3838 static DEFINE_SPINLOCK(balancing);
3841 * It checks each scheduling domain to see if it is due to be balanced,
3842 * and initiates a balancing operation if so.
3844 * Balancing parameters are set up in arch_init_sched_domains.
3846 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3849 struct rq *rq = cpu_rq(cpu);
3850 unsigned long interval;
3851 struct sched_domain *sd;
3852 /* Earliest time when we have to do rebalance again */
3853 unsigned long next_balance = jiffies + 60*HZ;
3854 int update_next_balance = 0;
3858 for_each_domain(cpu, sd) {
3859 if (!(sd->flags & SD_LOAD_BALANCE))
3862 interval = sd->balance_interval;
3863 if (idle != CPU_IDLE)
3864 interval *= sd->busy_factor;
3866 /* scale ms to jiffies */
3867 interval = msecs_to_jiffies(interval);
3868 if (unlikely(!interval))
3870 if (interval > HZ*NR_CPUS/10)
3871 interval = HZ*NR_CPUS/10;
3873 need_serialize = sd->flags & SD_SERIALIZE;
3875 if (need_serialize) {
3876 if (!spin_trylock(&balancing))
3880 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3881 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3883 * We've pulled tasks over so either we're no
3884 * longer idle, or one of our SMT siblings is
3887 idle = CPU_NOT_IDLE;
3889 sd->last_balance = jiffies;
3892 spin_unlock(&balancing);
3894 if (time_after(next_balance, sd->last_balance + interval)) {
3895 next_balance = sd->last_balance + interval;
3896 update_next_balance = 1;
3900 * Stop the load balance at this level. There is another
3901 * CPU in our sched group which is doing load balancing more
3909 * next_balance will be updated only when there is a need.
3910 * When the cpu is attached to null domain for ex, it will not be
3913 if (likely(update_next_balance))
3914 rq->next_balance = next_balance;
3918 * run_rebalance_domains is triggered when needed from the scheduler tick.
3919 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3920 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3922 static void run_rebalance_domains(struct softirq_action *h)
3924 int this_cpu = smp_processor_id();
3925 struct rq *this_rq = cpu_rq(this_cpu);
3926 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3927 CPU_IDLE : CPU_NOT_IDLE;
3929 rebalance_domains(this_cpu, idle);
3933 * If this cpu is the owner for idle load balancing, then do the
3934 * balancing on behalf of the other idle cpus whose ticks are
3937 if (this_rq->idle_at_tick &&
3938 atomic_read(&nohz.load_balancer) == this_cpu) {
3939 cpumask_t cpus = nohz.cpu_mask;
3943 cpu_clear(this_cpu, cpus);
3944 for_each_cpu_mask_nr(balance_cpu, cpus) {
3946 * If this cpu gets work to do, stop the load balancing
3947 * work being done for other cpus. Next load
3948 * balancing owner will pick it up.
3953 rebalance_domains(balance_cpu, CPU_IDLE);
3955 rq = cpu_rq(balance_cpu);
3956 if (time_after(this_rq->next_balance, rq->next_balance))
3957 this_rq->next_balance = rq->next_balance;
3964 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3966 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3967 * idle load balancing owner or decide to stop the periodic load balancing,
3968 * if the whole system is idle.
3970 static inline void trigger_load_balance(struct rq *rq, int cpu)
3974 * If we were in the nohz mode recently and busy at the current
3975 * scheduler tick, then check if we need to nominate new idle
3978 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3979 rq->in_nohz_recently = 0;
3981 if (atomic_read(&nohz.load_balancer) == cpu) {
3982 cpu_clear(cpu, nohz.cpu_mask);
3983 atomic_set(&nohz.load_balancer, -1);
3986 if (atomic_read(&nohz.load_balancer) == -1) {
3988 * simple selection for now: Nominate the
3989 * first cpu in the nohz list to be the next
3992 * TBD: Traverse the sched domains and nominate
3993 * the nearest cpu in the nohz.cpu_mask.
3995 int ilb = first_cpu(nohz.cpu_mask);
3997 if (ilb < nr_cpu_ids)
4003 * If this cpu is idle and doing idle load balancing for all the
4004 * cpus with ticks stopped, is it time for that to stop?
4006 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4007 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4013 * If this cpu is idle and the idle load balancing is done by
4014 * someone else, then no need raise the SCHED_SOFTIRQ
4016 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4017 cpu_isset(cpu, nohz.cpu_mask))
4020 if (time_after_eq(jiffies, rq->next_balance))
4021 raise_softirq(SCHED_SOFTIRQ);
4024 #else /* CONFIG_SMP */
4027 * on UP we do not need to balance between CPUs:
4029 static inline void idle_balance(int cpu, struct rq *rq)
4035 DEFINE_PER_CPU(struct kernel_stat, kstat);
4037 EXPORT_PER_CPU_SYMBOL(kstat);
4040 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4041 * that have not yet been banked in case the task is currently running.
4043 unsigned long long task_sched_runtime(struct task_struct *p)
4045 unsigned long flags;
4049 rq = task_rq_lock(p, &flags);
4050 ns = p->se.sum_exec_runtime;
4051 if (task_current(rq, p)) {
4052 update_rq_clock(rq);
4053 delta_exec = rq->clock - p->se.exec_start;
4054 if ((s64)delta_exec > 0)
4057 task_rq_unlock(rq, &flags);
4063 * Account user cpu time to a process.
4064 * @p: the process that the cpu time gets accounted to
4065 * @cputime: the cpu time spent in user space since the last update
4067 void account_user_time(struct task_struct *p, cputime_t cputime)
4069 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4072 p->utime = cputime_add(p->utime, cputime);
4074 /* Add user time to cpustat. */
4075 tmp = cputime_to_cputime64(cputime);
4076 if (TASK_NICE(p) > 0)
4077 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4079 cpustat->user = cputime64_add(cpustat->user, tmp);
4080 /* Account for user time used */
4081 acct_update_integrals(p);
4085 * Account guest cpu time to a process.
4086 * @p: the process that the cpu time gets accounted to
4087 * @cputime: the cpu time spent in virtual machine since the last update
4089 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4092 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4094 tmp = cputime_to_cputime64(cputime);
4096 p->utime = cputime_add(p->utime, cputime);
4097 p->gtime = cputime_add(p->gtime, cputime);
4099 cpustat->user = cputime64_add(cpustat->user, tmp);
4100 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4104 * Account scaled user cpu time to a process.
4105 * @p: the process that the cpu time gets accounted to
4106 * @cputime: the cpu time spent in user space since the last update
4108 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4110 p->utimescaled = cputime_add(p->utimescaled, cputime);
4114 * Account system cpu time to a process.
4115 * @p: the process that the cpu time gets accounted to
4116 * @hardirq_offset: the offset to subtract from hardirq_count()
4117 * @cputime: the cpu time spent in kernel space since the last update
4119 void account_system_time(struct task_struct *p, int hardirq_offset,
4122 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4123 struct rq *rq = this_rq();
4126 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4127 account_guest_time(p, cputime);
4131 p->stime = cputime_add(p->stime, cputime);
4133 /* Add system time to cpustat. */
4134 tmp = cputime_to_cputime64(cputime);
4135 if (hardirq_count() - hardirq_offset)
4136 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4137 else if (softirq_count())
4138 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4139 else if (p != rq->idle)
4140 cpustat->system = cputime64_add(cpustat->system, tmp);
4141 else if (atomic_read(&rq->nr_iowait) > 0)
4142 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4144 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4145 /* Account for system time used */
4146 acct_update_integrals(p);
4150 * Account scaled system cpu time to a process.
4151 * @p: the process that the cpu time gets accounted to
4152 * @hardirq_offset: the offset to subtract from hardirq_count()
4153 * @cputime: the cpu time spent in kernel space since the last update
4155 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4157 p->stimescaled = cputime_add(p->stimescaled, cputime);
4161 * Account for involuntary wait time.
4162 * @p: the process from which the cpu time has been stolen
4163 * @steal: the cpu time spent in involuntary wait
4165 void account_steal_time(struct task_struct *p, cputime_t steal)
4167 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4168 cputime64_t tmp = cputime_to_cputime64(steal);
4169 struct rq *rq = this_rq();
4171 if (p == rq->idle) {
4172 p->stime = cputime_add(p->stime, steal);
4173 if (atomic_read(&rq->nr_iowait) > 0)
4174 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4176 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4178 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4182 * This function gets called by the timer code, with HZ frequency.
4183 * We call it with interrupts disabled.
4185 * It also gets called by the fork code, when changing the parent's
4188 void scheduler_tick(void)
4190 int cpu = smp_processor_id();
4191 struct rq *rq = cpu_rq(cpu);
4192 struct task_struct *curr = rq->curr;
4196 spin_lock(&rq->lock);
4197 update_rq_clock(rq);
4198 update_cpu_load(rq);
4199 curr->sched_class->task_tick(rq, curr, 0);
4200 spin_unlock(&rq->lock);
4203 rq->idle_at_tick = idle_cpu(cpu);
4204 trigger_load_balance(rq, cpu);
4208 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4209 defined(CONFIG_PREEMPT_TRACER))
4211 static inline unsigned long get_parent_ip(unsigned long addr)
4213 if (in_lock_functions(addr)) {
4214 addr = CALLER_ADDR2;
4215 if (in_lock_functions(addr))
4216 addr = CALLER_ADDR3;
4221 void __kprobes add_preempt_count(int val)
4223 #ifdef CONFIG_DEBUG_PREEMPT
4227 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4230 preempt_count() += val;
4231 #ifdef CONFIG_DEBUG_PREEMPT
4233 * Spinlock count overflowing soon?
4235 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4238 if (preempt_count() == val)
4239 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4241 EXPORT_SYMBOL(add_preempt_count);
4243 void __kprobes sub_preempt_count(int val)
4245 #ifdef CONFIG_DEBUG_PREEMPT
4249 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4252 * Is the spinlock portion underflowing?
4254 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4255 !(preempt_count() & PREEMPT_MASK)))
4259 if (preempt_count() == val)
4260 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4261 preempt_count() -= val;
4263 EXPORT_SYMBOL(sub_preempt_count);
4268 * Print scheduling while atomic bug:
4270 static noinline void __schedule_bug(struct task_struct *prev)
4272 struct pt_regs *regs = get_irq_regs();
4274 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4275 prev->comm, prev->pid, preempt_count());
4277 debug_show_held_locks(prev);
4279 if (irqs_disabled())
4280 print_irqtrace_events(prev);
4289 * Various schedule()-time debugging checks and statistics:
4291 static inline void schedule_debug(struct task_struct *prev)
4294 * Test if we are atomic. Since do_exit() needs to call into
4295 * schedule() atomically, we ignore that path for now.
4296 * Otherwise, whine if we are scheduling when we should not be.
4298 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4299 __schedule_bug(prev);
4301 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4303 schedstat_inc(this_rq(), sched_count);
4304 #ifdef CONFIG_SCHEDSTATS
4305 if (unlikely(prev->lock_depth >= 0)) {
4306 schedstat_inc(this_rq(), bkl_count);
4307 schedstat_inc(prev, sched_info.bkl_count);
4313 * Pick up the highest-prio task:
4315 static inline struct task_struct *
4316 pick_next_task(struct rq *rq, struct task_struct *prev)
4318 const struct sched_class *class;
4319 struct task_struct *p;
4322 * Optimization: we know that if all tasks are in
4323 * the fair class we can call that function directly:
4325 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4326 p = fair_sched_class.pick_next_task(rq);
4331 class = sched_class_highest;
4333 p = class->pick_next_task(rq);
4337 * Will never be NULL as the idle class always
4338 * returns a non-NULL p:
4340 class = class->next;
4345 * schedule() is the main scheduler function.
4347 asmlinkage void __sched schedule(void)
4349 struct task_struct *prev, *next;
4350 unsigned long *switch_count;
4356 cpu = smp_processor_id();
4360 switch_count = &prev->nivcsw;
4362 release_kernel_lock(prev);
4363 need_resched_nonpreemptible:
4365 schedule_debug(prev);
4367 if (sched_feat(HRTICK))
4371 * Do the rq-clock update outside the rq lock:
4373 local_irq_disable();
4374 update_rq_clock(rq);
4375 spin_lock(&rq->lock);
4376 clear_tsk_need_resched(prev);
4378 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4379 if (unlikely(signal_pending_state(prev->state, prev)))
4380 prev->state = TASK_RUNNING;
4382 deactivate_task(rq, prev, 1);
4383 switch_count = &prev->nvcsw;
4387 if (prev->sched_class->pre_schedule)
4388 prev->sched_class->pre_schedule(rq, prev);
4391 if (unlikely(!rq->nr_running))
4392 idle_balance(cpu, rq);
4394 prev->sched_class->put_prev_task(rq, prev);
4395 next = pick_next_task(rq, prev);
4397 if (likely(prev != next)) {
4398 sched_info_switch(prev, next);
4404 context_switch(rq, prev, next); /* unlocks the rq */
4406 * the context switch might have flipped the stack from under
4407 * us, hence refresh the local variables.
4409 cpu = smp_processor_id();
4412 spin_unlock_irq(&rq->lock);
4414 if (unlikely(reacquire_kernel_lock(current) < 0))
4415 goto need_resched_nonpreemptible;
4417 preempt_enable_no_resched();
4418 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4421 EXPORT_SYMBOL(schedule);
4423 #ifdef CONFIG_PREEMPT
4425 * this is the entry point to schedule() from in-kernel preemption
4426 * off of preempt_enable. Kernel preemptions off return from interrupt
4427 * occur there and call schedule directly.
4429 asmlinkage void __sched preempt_schedule(void)
4431 struct thread_info *ti = current_thread_info();
4434 * If there is a non-zero preempt_count or interrupts are disabled,
4435 * we do not want to preempt the current task. Just return..
4437 if (likely(ti->preempt_count || irqs_disabled()))
4441 add_preempt_count(PREEMPT_ACTIVE);
4443 sub_preempt_count(PREEMPT_ACTIVE);
4446 * Check again in case we missed a preemption opportunity
4447 * between schedule and now.
4450 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4452 EXPORT_SYMBOL(preempt_schedule);
4455 * this is the entry point to schedule() from kernel preemption
4456 * off of irq context.
4457 * Note, that this is called and return with irqs disabled. This will
4458 * protect us against recursive calling from irq.
4460 asmlinkage void __sched preempt_schedule_irq(void)
4462 struct thread_info *ti = current_thread_info();
4464 /* Catch callers which need to be fixed */
4465 BUG_ON(ti->preempt_count || !irqs_disabled());
4468 add_preempt_count(PREEMPT_ACTIVE);
4471 local_irq_disable();
4472 sub_preempt_count(PREEMPT_ACTIVE);
4475 * Check again in case we missed a preemption opportunity
4476 * between schedule and now.
4479 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4482 #endif /* CONFIG_PREEMPT */
4484 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4487 return try_to_wake_up(curr->private, mode, sync);
4489 EXPORT_SYMBOL(default_wake_function);
4492 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4493 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4494 * number) then we wake all the non-exclusive tasks and one exclusive task.
4496 * There are circumstances in which we can try to wake a task which has already
4497 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4498 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4500 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4501 int nr_exclusive, int sync, void *key)
4503 wait_queue_t *curr, *next;
4505 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4506 unsigned flags = curr->flags;
4508 if (curr->func(curr, mode, sync, key) &&
4509 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4515 * __wake_up - wake up threads blocked on a waitqueue.
4517 * @mode: which threads
4518 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4519 * @key: is directly passed to the wakeup function
4521 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4522 int nr_exclusive, void *key)
4524 unsigned long flags;
4526 spin_lock_irqsave(&q->lock, flags);
4527 __wake_up_common(q, mode, nr_exclusive, 0, key);
4528 spin_unlock_irqrestore(&q->lock, flags);
4530 EXPORT_SYMBOL(__wake_up);
4533 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4535 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4537 __wake_up_common(q, mode, 1, 0, NULL);
4541 * __wake_up_sync - wake up threads blocked on a waitqueue.
4543 * @mode: which threads
4544 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4546 * The sync wakeup differs that the waker knows that it will schedule
4547 * away soon, so while the target thread will be woken up, it will not
4548 * be migrated to another CPU - ie. the two threads are 'synchronized'
4549 * with each other. This can prevent needless bouncing between CPUs.
4551 * On UP it can prevent extra preemption.
4554 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4556 unsigned long flags;
4562 if (unlikely(!nr_exclusive))
4565 spin_lock_irqsave(&q->lock, flags);
4566 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4567 spin_unlock_irqrestore(&q->lock, flags);
4569 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4571 void complete(struct completion *x)
4573 unsigned long flags;
4575 spin_lock_irqsave(&x->wait.lock, flags);
4577 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4578 spin_unlock_irqrestore(&x->wait.lock, flags);
4580 EXPORT_SYMBOL(complete);
4582 void complete_all(struct completion *x)
4584 unsigned long flags;
4586 spin_lock_irqsave(&x->wait.lock, flags);
4587 x->done += UINT_MAX/2;
4588 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4589 spin_unlock_irqrestore(&x->wait.lock, flags);
4591 EXPORT_SYMBOL(complete_all);
4593 static inline long __sched
4594 do_wait_for_common(struct completion *x, long timeout, int state)
4597 DECLARE_WAITQUEUE(wait, current);
4599 wait.flags |= WQ_FLAG_EXCLUSIVE;
4600 __add_wait_queue_tail(&x->wait, &wait);
4602 if ((state == TASK_INTERRUPTIBLE &&
4603 signal_pending(current)) ||
4604 (state == TASK_KILLABLE &&
4605 fatal_signal_pending(current))) {
4606 timeout = -ERESTARTSYS;
4609 __set_current_state(state);
4610 spin_unlock_irq(&x->wait.lock);
4611 timeout = schedule_timeout(timeout);
4612 spin_lock_irq(&x->wait.lock);
4613 } while (!x->done && timeout);
4614 __remove_wait_queue(&x->wait, &wait);
4619 return timeout ?: 1;
4623 wait_for_common(struct completion *x, long timeout, int state)
4627 spin_lock_irq(&x->wait.lock);
4628 timeout = do_wait_for_common(x, timeout, state);
4629 spin_unlock_irq(&x->wait.lock);
4633 void __sched wait_for_completion(struct completion *x)
4635 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4637 EXPORT_SYMBOL(wait_for_completion);
4639 unsigned long __sched
4640 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4642 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4644 EXPORT_SYMBOL(wait_for_completion_timeout);
4646 int __sched wait_for_completion_interruptible(struct completion *x)
4648 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4649 if (t == -ERESTARTSYS)
4653 EXPORT_SYMBOL(wait_for_completion_interruptible);
4655 unsigned long __sched
4656 wait_for_completion_interruptible_timeout(struct completion *x,
4657 unsigned long timeout)
4659 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4661 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4663 int __sched wait_for_completion_killable(struct completion *x)
4665 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4666 if (t == -ERESTARTSYS)
4670 EXPORT_SYMBOL(wait_for_completion_killable);
4673 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4675 unsigned long flags;
4678 init_waitqueue_entry(&wait, current);
4680 __set_current_state(state);
4682 spin_lock_irqsave(&q->lock, flags);
4683 __add_wait_queue(q, &wait);
4684 spin_unlock(&q->lock);
4685 timeout = schedule_timeout(timeout);
4686 spin_lock_irq(&q->lock);
4687 __remove_wait_queue(q, &wait);
4688 spin_unlock_irqrestore(&q->lock, flags);
4693 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4695 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4697 EXPORT_SYMBOL(interruptible_sleep_on);
4700 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4702 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4704 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4706 void __sched sleep_on(wait_queue_head_t *q)
4708 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4710 EXPORT_SYMBOL(sleep_on);
4712 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4714 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4716 EXPORT_SYMBOL(sleep_on_timeout);
4718 #ifdef CONFIG_RT_MUTEXES
4721 * rt_mutex_setprio - set the current priority of a task
4723 * @prio: prio value (kernel-internal form)
4725 * This function changes the 'effective' priority of a task. It does
4726 * not touch ->normal_prio like __setscheduler().
4728 * Used by the rt_mutex code to implement priority inheritance logic.
4730 void rt_mutex_setprio(struct task_struct *p, int prio)
4732 unsigned long flags;
4733 int oldprio, on_rq, running;
4735 const struct sched_class *prev_class = p->sched_class;
4737 BUG_ON(prio < 0 || prio > MAX_PRIO);
4739 rq = task_rq_lock(p, &flags);
4740 update_rq_clock(rq);
4743 on_rq = p->se.on_rq;
4744 running = task_current(rq, p);
4746 dequeue_task(rq, p, 0);
4748 p->sched_class->put_prev_task(rq, p);
4751 p->sched_class = &rt_sched_class;
4753 p->sched_class = &fair_sched_class;
4758 p->sched_class->set_curr_task(rq);
4760 enqueue_task(rq, p, 0);
4762 check_class_changed(rq, p, prev_class, oldprio, running);
4764 task_rq_unlock(rq, &flags);
4769 void set_user_nice(struct task_struct *p, long nice)
4771 int old_prio, delta, on_rq;
4772 unsigned long flags;
4775 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4778 * We have to be careful, if called from sys_setpriority(),
4779 * the task might be in the middle of scheduling on another CPU.
4781 rq = task_rq_lock(p, &flags);
4782 update_rq_clock(rq);
4784 * The RT priorities are set via sched_setscheduler(), but we still
4785 * allow the 'normal' nice value to be set - but as expected
4786 * it wont have any effect on scheduling until the task is
4787 * SCHED_FIFO/SCHED_RR:
4789 if (task_has_rt_policy(p)) {
4790 p->static_prio = NICE_TO_PRIO(nice);
4793 on_rq = p->se.on_rq;
4795 dequeue_task(rq, p, 0);
4797 p->static_prio = NICE_TO_PRIO(nice);
4800 p->prio = effective_prio(p);
4801 delta = p->prio - old_prio;
4804 enqueue_task(rq, p, 0);
4806 * If the task increased its priority or is running and
4807 * lowered its priority, then reschedule its CPU:
4809 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4810 resched_task(rq->curr);
4813 task_rq_unlock(rq, &flags);
4815 EXPORT_SYMBOL(set_user_nice);
4818 * can_nice - check if a task can reduce its nice value
4822 int can_nice(const struct task_struct *p, const int nice)
4824 /* convert nice value [19,-20] to rlimit style value [1,40] */
4825 int nice_rlim = 20 - nice;
4827 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4828 capable(CAP_SYS_NICE));
4831 #ifdef __ARCH_WANT_SYS_NICE
4834 * sys_nice - change the priority of the current process.
4835 * @increment: priority increment
4837 * sys_setpriority is a more generic, but much slower function that
4838 * does similar things.
4840 asmlinkage long sys_nice(int increment)
4845 * Setpriority might change our priority at the same moment.
4846 * We don't have to worry. Conceptually one call occurs first
4847 * and we have a single winner.
4849 if (increment < -40)
4854 nice = PRIO_TO_NICE(current->static_prio) + increment;
4860 if (increment < 0 && !can_nice(current, nice))
4863 retval = security_task_setnice(current, nice);
4867 set_user_nice(current, nice);
4874 * task_prio - return the priority value of a given task.
4875 * @p: the task in question.
4877 * This is the priority value as seen by users in /proc.
4878 * RT tasks are offset by -200. Normal tasks are centered
4879 * around 0, value goes from -16 to +15.
4881 int task_prio(const struct task_struct *p)
4883 return p->prio - MAX_RT_PRIO;
4887 * task_nice - return the nice value of a given task.
4888 * @p: the task in question.
4890 int task_nice(const struct task_struct *p)
4892 return TASK_NICE(p);
4894 EXPORT_SYMBOL(task_nice);
4897 * idle_cpu - is a given cpu idle currently?
4898 * @cpu: the processor in question.
4900 int idle_cpu(int cpu)
4902 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4906 * idle_task - return the idle task for a given cpu.
4907 * @cpu: the processor in question.
4909 struct task_struct *idle_task(int cpu)
4911 return cpu_rq(cpu)->idle;
4915 * find_process_by_pid - find a process with a matching PID value.
4916 * @pid: the pid in question.
4918 static struct task_struct *find_process_by_pid(pid_t pid)
4920 return pid ? find_task_by_vpid(pid) : current;
4923 /* Actually do priority change: must hold rq lock. */
4925 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4927 BUG_ON(p->se.on_rq);
4930 switch (p->policy) {
4934 p->sched_class = &fair_sched_class;
4938 p->sched_class = &rt_sched_class;
4942 p->rt_priority = prio;
4943 p->normal_prio = normal_prio(p);
4944 /* we are holding p->pi_lock already */
4945 p->prio = rt_mutex_getprio(p);
4949 static int __sched_setscheduler(struct task_struct *p, int policy,
4950 struct sched_param *param, bool user)
4952 int retval, oldprio, oldpolicy = -1, on_rq, running;
4953 unsigned long flags;
4954 const struct sched_class *prev_class = p->sched_class;
4957 /* may grab non-irq protected spin_locks */
4958 BUG_ON(in_interrupt());
4960 /* double check policy once rq lock held */
4962 policy = oldpolicy = p->policy;
4963 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4964 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4965 policy != SCHED_IDLE)
4968 * Valid priorities for SCHED_FIFO and SCHED_RR are
4969 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4970 * SCHED_BATCH and SCHED_IDLE is 0.
4972 if (param->sched_priority < 0 ||
4973 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4974 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4976 if (rt_policy(policy) != (param->sched_priority != 0))
4980 * Allow unprivileged RT tasks to decrease priority:
4982 if (user && !capable(CAP_SYS_NICE)) {
4983 if (rt_policy(policy)) {
4984 unsigned long rlim_rtprio;
4986 if (!lock_task_sighand(p, &flags))
4988 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4989 unlock_task_sighand(p, &flags);
4991 /* can't set/change the rt policy */
4992 if (policy != p->policy && !rlim_rtprio)
4995 /* can't increase priority */
4996 if (param->sched_priority > p->rt_priority &&
4997 param->sched_priority > rlim_rtprio)
5001 * Like positive nice levels, dont allow tasks to
5002 * move out of SCHED_IDLE either:
5004 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5007 /* can't change other user's priorities */
5008 if ((current->euid != p->euid) &&
5009 (current->euid != p->uid))
5014 #ifdef CONFIG_RT_GROUP_SCHED
5016 * Do not allow realtime tasks into groups that have no runtime
5019 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5023 retval = security_task_setscheduler(p, policy, param);
5029 * make sure no PI-waiters arrive (or leave) while we are
5030 * changing the priority of the task:
5032 spin_lock_irqsave(&p->pi_lock, flags);
5034 * To be able to change p->policy safely, the apropriate
5035 * runqueue lock must be held.
5037 rq = __task_rq_lock(p);
5038 /* recheck policy now with rq lock held */
5039 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5040 policy = oldpolicy = -1;
5041 __task_rq_unlock(rq);
5042 spin_unlock_irqrestore(&p->pi_lock, flags);
5045 update_rq_clock(rq);
5046 on_rq = p->se.on_rq;
5047 running = task_current(rq, p);
5049 deactivate_task(rq, p, 0);
5051 p->sched_class->put_prev_task(rq, p);
5054 __setscheduler(rq, p, policy, param->sched_priority);
5057 p->sched_class->set_curr_task(rq);
5059 activate_task(rq, p, 0);
5061 check_class_changed(rq, p, prev_class, oldprio, running);
5063 __task_rq_unlock(rq);
5064 spin_unlock_irqrestore(&p->pi_lock, flags);
5066 rt_mutex_adjust_pi(p);
5072 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5073 * @p: the task in question.
5074 * @policy: new policy.
5075 * @param: structure containing the new RT priority.
5077 * NOTE that the task may be already dead.
5079 int sched_setscheduler(struct task_struct *p, int policy,
5080 struct sched_param *param)
5082 return __sched_setscheduler(p, policy, param, true);
5084 EXPORT_SYMBOL_GPL(sched_setscheduler);
5087 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5088 * @p: the task in question.
5089 * @policy: new policy.
5090 * @param: structure containing the new RT priority.
5092 * Just like sched_setscheduler, only don't bother checking if the
5093 * current context has permission. For example, this is needed in
5094 * stop_machine(): we create temporary high priority worker threads,
5095 * but our caller might not have that capability.
5097 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5098 struct sched_param *param)
5100 return __sched_setscheduler(p, policy, param, false);
5104 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5106 struct sched_param lparam;
5107 struct task_struct *p;
5110 if (!param || pid < 0)
5112 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5117 p = find_process_by_pid(pid);
5119 retval = sched_setscheduler(p, policy, &lparam);
5126 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5127 * @pid: the pid in question.
5128 * @policy: new policy.
5129 * @param: structure containing the new RT priority.
5132 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5134 /* negative values for policy are not valid */
5138 return do_sched_setscheduler(pid, policy, param);
5142 * sys_sched_setparam - set/change the RT priority of a thread
5143 * @pid: the pid in question.
5144 * @param: structure containing the new RT priority.
5146 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5148 return do_sched_setscheduler(pid, -1, param);
5152 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5153 * @pid: the pid in question.
5155 asmlinkage long sys_sched_getscheduler(pid_t pid)
5157 struct task_struct *p;
5164 read_lock(&tasklist_lock);
5165 p = find_process_by_pid(pid);
5167 retval = security_task_getscheduler(p);
5171 read_unlock(&tasklist_lock);
5176 * sys_sched_getscheduler - get the RT priority of a thread
5177 * @pid: the pid in question.
5178 * @param: structure containing the RT priority.
5180 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5182 struct sched_param lp;
5183 struct task_struct *p;
5186 if (!param || pid < 0)
5189 read_lock(&tasklist_lock);
5190 p = find_process_by_pid(pid);
5195 retval = security_task_getscheduler(p);
5199 lp.sched_priority = p->rt_priority;
5200 read_unlock(&tasklist_lock);
5203 * This one might sleep, we cannot do it with a spinlock held ...
5205 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5210 read_unlock(&tasklist_lock);
5214 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5216 cpumask_t cpus_allowed;
5217 cpumask_t new_mask = *in_mask;
5218 struct task_struct *p;
5222 read_lock(&tasklist_lock);
5224 p = find_process_by_pid(pid);
5226 read_unlock(&tasklist_lock);
5232 * It is not safe to call set_cpus_allowed with the
5233 * tasklist_lock held. We will bump the task_struct's
5234 * usage count and then drop tasklist_lock.
5237 read_unlock(&tasklist_lock);
5240 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5241 !capable(CAP_SYS_NICE))
5244 retval = security_task_setscheduler(p, 0, NULL);
5248 cpuset_cpus_allowed(p, &cpus_allowed);
5249 cpus_and(new_mask, new_mask, cpus_allowed);
5251 retval = set_cpus_allowed_ptr(p, &new_mask);
5254 cpuset_cpus_allowed(p, &cpus_allowed);
5255 if (!cpus_subset(new_mask, cpus_allowed)) {
5257 * We must have raced with a concurrent cpuset
5258 * update. Just reset the cpus_allowed to the
5259 * cpuset's cpus_allowed
5261 new_mask = cpus_allowed;
5271 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5272 cpumask_t *new_mask)
5274 if (len < sizeof(cpumask_t)) {
5275 memset(new_mask, 0, sizeof(cpumask_t));
5276 } else if (len > sizeof(cpumask_t)) {
5277 len = sizeof(cpumask_t);
5279 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5283 * sys_sched_setaffinity - set the cpu affinity of a process
5284 * @pid: pid of the process
5285 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5286 * @user_mask_ptr: user-space pointer to the new cpu mask
5288 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5289 unsigned long __user *user_mask_ptr)
5294 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5298 return sched_setaffinity(pid, &new_mask);
5301 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5303 struct task_struct *p;
5307 read_lock(&tasklist_lock);
5310 p = find_process_by_pid(pid);
5314 retval = security_task_getscheduler(p);
5318 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5321 read_unlock(&tasklist_lock);
5328 * sys_sched_getaffinity - get the cpu affinity of a process
5329 * @pid: pid of the process
5330 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5331 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5333 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5334 unsigned long __user *user_mask_ptr)
5339 if (len < sizeof(cpumask_t))
5342 ret = sched_getaffinity(pid, &mask);
5346 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5349 return sizeof(cpumask_t);
5353 * sys_sched_yield - yield the current processor to other threads.
5355 * This function yields the current CPU to other tasks. If there are no
5356 * other threads running on this CPU then this function will return.
5358 asmlinkage long sys_sched_yield(void)
5360 struct rq *rq = this_rq_lock();
5362 schedstat_inc(rq, yld_count);
5363 current->sched_class->yield_task(rq);
5366 * Since we are going to call schedule() anyway, there's
5367 * no need to preempt or enable interrupts:
5369 __release(rq->lock);
5370 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5371 _raw_spin_unlock(&rq->lock);
5372 preempt_enable_no_resched();
5379 static void __cond_resched(void)
5381 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5382 __might_sleep(__FILE__, __LINE__);
5385 * The BKS might be reacquired before we have dropped
5386 * PREEMPT_ACTIVE, which could trigger a second
5387 * cond_resched() call.
5390 add_preempt_count(PREEMPT_ACTIVE);
5392 sub_preempt_count(PREEMPT_ACTIVE);
5393 } while (need_resched());
5396 int __sched _cond_resched(void)
5398 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5399 system_state == SYSTEM_RUNNING) {
5405 EXPORT_SYMBOL(_cond_resched);
5408 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5409 * call schedule, and on return reacquire the lock.
5411 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5412 * operations here to prevent schedule() from being called twice (once via
5413 * spin_unlock(), once by hand).
5415 int cond_resched_lock(spinlock_t *lock)
5417 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5420 if (spin_needbreak(lock) || resched) {
5422 if (resched && need_resched())
5431 EXPORT_SYMBOL(cond_resched_lock);
5433 int __sched cond_resched_softirq(void)
5435 BUG_ON(!in_softirq());
5437 if (need_resched() && system_state == SYSTEM_RUNNING) {
5445 EXPORT_SYMBOL(cond_resched_softirq);
5448 * yield - yield the current processor to other threads.
5450 * This is a shortcut for kernel-space yielding - it marks the
5451 * thread runnable and calls sys_sched_yield().
5453 void __sched yield(void)
5455 set_current_state(TASK_RUNNING);
5458 EXPORT_SYMBOL(yield);
5461 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5462 * that process accounting knows that this is a task in IO wait state.
5464 * But don't do that if it is a deliberate, throttling IO wait (this task
5465 * has set its backing_dev_info: the queue against which it should throttle)
5467 void __sched io_schedule(void)
5469 struct rq *rq = &__raw_get_cpu_var(runqueues);
5471 delayacct_blkio_start();
5472 atomic_inc(&rq->nr_iowait);
5474 atomic_dec(&rq->nr_iowait);
5475 delayacct_blkio_end();
5477 EXPORT_SYMBOL(io_schedule);
5479 long __sched io_schedule_timeout(long timeout)
5481 struct rq *rq = &__raw_get_cpu_var(runqueues);
5484 delayacct_blkio_start();
5485 atomic_inc(&rq->nr_iowait);
5486 ret = schedule_timeout(timeout);
5487 atomic_dec(&rq->nr_iowait);
5488 delayacct_blkio_end();
5493 * sys_sched_get_priority_max - return maximum RT priority.
5494 * @policy: scheduling class.
5496 * this syscall returns the maximum rt_priority that can be used
5497 * by a given scheduling class.
5499 asmlinkage long sys_sched_get_priority_max(int policy)
5506 ret = MAX_USER_RT_PRIO-1;
5518 * sys_sched_get_priority_min - return minimum RT priority.
5519 * @policy: scheduling class.
5521 * this syscall returns the minimum rt_priority that can be used
5522 * by a given scheduling class.
5524 asmlinkage long sys_sched_get_priority_min(int policy)
5542 * sys_sched_rr_get_interval - return the default timeslice of a process.
5543 * @pid: pid of the process.
5544 * @interval: userspace pointer to the timeslice value.
5546 * this syscall writes the default timeslice value of a given process
5547 * into the user-space timespec buffer. A value of '0' means infinity.
5550 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5552 struct task_struct *p;
5553 unsigned int time_slice;
5561 read_lock(&tasklist_lock);
5562 p = find_process_by_pid(pid);
5566 retval = security_task_getscheduler(p);
5571 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5572 * tasks that are on an otherwise idle runqueue:
5575 if (p->policy == SCHED_RR) {
5576 time_slice = DEF_TIMESLICE;
5577 } else if (p->policy != SCHED_FIFO) {
5578 struct sched_entity *se = &p->se;
5579 unsigned long flags;
5582 rq = task_rq_lock(p, &flags);
5583 if (rq->cfs.load.weight)
5584 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5585 task_rq_unlock(rq, &flags);
5587 read_unlock(&tasklist_lock);
5588 jiffies_to_timespec(time_slice, &t);
5589 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5593 read_unlock(&tasklist_lock);
5597 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5599 void sched_show_task(struct task_struct *p)
5601 unsigned long free = 0;
5604 state = p->state ? __ffs(p->state) + 1 : 0;
5605 printk(KERN_INFO "%-13.13s %c", p->comm,
5606 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5607 #if BITS_PER_LONG == 32
5608 if (state == TASK_RUNNING)
5609 printk(KERN_CONT " running ");
5611 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5613 if (state == TASK_RUNNING)
5614 printk(KERN_CONT " running task ");
5616 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5618 #ifdef CONFIG_DEBUG_STACK_USAGE
5620 unsigned long *n = end_of_stack(p);
5623 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5626 printk(KERN_CONT "%5lu %5d %6d\n", free,
5627 task_pid_nr(p), task_pid_nr(p->real_parent));
5629 show_stack(p, NULL);
5632 void show_state_filter(unsigned long state_filter)
5634 struct task_struct *g, *p;
5636 #if BITS_PER_LONG == 32
5638 " task PC stack pid father\n");
5641 " task PC stack pid father\n");
5643 read_lock(&tasklist_lock);
5644 do_each_thread(g, p) {
5646 * reset the NMI-timeout, listing all files on a slow
5647 * console might take alot of time:
5649 touch_nmi_watchdog();
5650 if (!state_filter || (p->state & state_filter))
5652 } while_each_thread(g, p);
5654 touch_all_softlockup_watchdogs();
5656 #ifdef CONFIG_SCHED_DEBUG
5657 sysrq_sched_debug_show();
5659 read_unlock(&tasklist_lock);
5661 * Only show locks if all tasks are dumped:
5663 if (state_filter == -1)
5664 debug_show_all_locks();
5667 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5669 idle->sched_class = &idle_sched_class;
5673 * init_idle - set up an idle thread for a given CPU
5674 * @idle: task in question
5675 * @cpu: cpu the idle task belongs to
5677 * NOTE: this function does not set the idle thread's NEED_RESCHED
5678 * flag, to make booting more robust.
5680 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5682 struct rq *rq = cpu_rq(cpu);
5683 unsigned long flags;
5686 idle->se.exec_start = sched_clock();
5688 idle->prio = idle->normal_prio = MAX_PRIO;
5689 idle->cpus_allowed = cpumask_of_cpu(cpu);
5690 __set_task_cpu(idle, cpu);
5692 spin_lock_irqsave(&rq->lock, flags);
5693 rq->curr = rq->idle = idle;
5694 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5697 spin_unlock_irqrestore(&rq->lock, flags);
5699 /* Set the preempt count _outside_ the spinlocks! */
5700 #if defined(CONFIG_PREEMPT)
5701 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5703 task_thread_info(idle)->preempt_count = 0;
5706 * The idle tasks have their own, simple scheduling class:
5708 idle->sched_class = &idle_sched_class;
5712 * In a system that switches off the HZ timer nohz_cpu_mask
5713 * indicates which cpus entered this state. This is used
5714 * in the rcu update to wait only for active cpus. For system
5715 * which do not switch off the HZ timer nohz_cpu_mask should
5716 * always be CPU_MASK_NONE.
5718 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5721 * Increase the granularity value when there are more CPUs,
5722 * because with more CPUs the 'effective latency' as visible
5723 * to users decreases. But the relationship is not linear,
5724 * so pick a second-best guess by going with the log2 of the
5727 * This idea comes from the SD scheduler of Con Kolivas:
5729 static inline void sched_init_granularity(void)
5731 unsigned int factor = 1 + ilog2(num_online_cpus());
5732 const unsigned long limit = 200000000;
5734 sysctl_sched_min_granularity *= factor;
5735 if (sysctl_sched_min_granularity > limit)
5736 sysctl_sched_min_granularity = limit;
5738 sysctl_sched_latency *= factor;
5739 if (sysctl_sched_latency > limit)
5740 sysctl_sched_latency = limit;
5742 sysctl_sched_wakeup_granularity *= factor;
5744 sysctl_sched_shares_ratelimit *= factor;
5749 * This is how migration works:
5751 * 1) we queue a struct migration_req structure in the source CPU's
5752 * runqueue and wake up that CPU's migration thread.
5753 * 2) we down() the locked semaphore => thread blocks.
5754 * 3) migration thread wakes up (implicitly it forces the migrated
5755 * thread off the CPU)
5756 * 4) it gets the migration request and checks whether the migrated
5757 * task is still in the wrong runqueue.
5758 * 5) if it's in the wrong runqueue then the migration thread removes
5759 * it and puts it into the right queue.
5760 * 6) migration thread up()s the semaphore.
5761 * 7) we wake up and the migration is done.
5765 * Change a given task's CPU affinity. Migrate the thread to a
5766 * proper CPU and schedule it away if the CPU it's executing on
5767 * is removed from the allowed bitmask.
5769 * NOTE: the caller must have a valid reference to the task, the
5770 * task must not exit() & deallocate itself prematurely. The
5771 * call is not atomic; no spinlocks may be held.
5773 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5775 struct migration_req req;
5776 unsigned long flags;
5780 rq = task_rq_lock(p, &flags);
5781 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5786 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5787 !cpus_equal(p->cpus_allowed, *new_mask))) {
5792 if (p->sched_class->set_cpus_allowed)
5793 p->sched_class->set_cpus_allowed(p, new_mask);
5795 p->cpus_allowed = *new_mask;
5796 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5799 /* Can the task run on the task's current CPU? If so, we're done */
5800 if (cpu_isset(task_cpu(p), *new_mask))
5803 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5804 /* Need help from migration thread: drop lock and wait. */
5805 task_rq_unlock(rq, &flags);
5806 wake_up_process(rq->migration_thread);
5807 wait_for_completion(&req.done);
5808 tlb_migrate_finish(p->mm);
5812 task_rq_unlock(rq, &flags);
5816 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5819 * Move (not current) task off this cpu, onto dest cpu. We're doing
5820 * this because either it can't run here any more (set_cpus_allowed()
5821 * away from this CPU, or CPU going down), or because we're
5822 * attempting to rebalance this task on exec (sched_exec).
5824 * So we race with normal scheduler movements, but that's OK, as long
5825 * as the task is no longer on this CPU.
5827 * Returns non-zero if task was successfully migrated.
5829 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5831 struct rq *rq_dest, *rq_src;
5834 if (unlikely(!cpu_active(dest_cpu)))
5837 rq_src = cpu_rq(src_cpu);
5838 rq_dest = cpu_rq(dest_cpu);
5840 double_rq_lock(rq_src, rq_dest);
5841 /* Already moved. */
5842 if (task_cpu(p) != src_cpu)
5844 /* Affinity changed (again). */
5845 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5848 on_rq = p->se.on_rq;
5850 deactivate_task(rq_src, p, 0);
5852 set_task_cpu(p, dest_cpu);
5854 activate_task(rq_dest, p, 0);
5855 check_preempt_curr(rq_dest, p);
5860 double_rq_unlock(rq_src, rq_dest);
5865 * migration_thread - this is a highprio system thread that performs
5866 * thread migration by bumping thread off CPU then 'pushing' onto
5869 static int migration_thread(void *data)
5871 int cpu = (long)data;
5875 BUG_ON(rq->migration_thread != current);
5877 set_current_state(TASK_INTERRUPTIBLE);
5878 while (!kthread_should_stop()) {
5879 struct migration_req *req;
5880 struct list_head *head;
5882 spin_lock_irq(&rq->lock);
5884 if (cpu_is_offline(cpu)) {
5885 spin_unlock_irq(&rq->lock);
5889 if (rq->active_balance) {
5890 active_load_balance(rq, cpu);
5891 rq->active_balance = 0;
5894 head = &rq->migration_queue;
5896 if (list_empty(head)) {
5897 spin_unlock_irq(&rq->lock);
5899 set_current_state(TASK_INTERRUPTIBLE);
5902 req = list_entry(head->next, struct migration_req, list);
5903 list_del_init(head->next);
5905 spin_unlock(&rq->lock);
5906 __migrate_task(req->task, cpu, req->dest_cpu);
5909 complete(&req->done);
5911 __set_current_state(TASK_RUNNING);
5915 /* Wait for kthread_stop */
5916 set_current_state(TASK_INTERRUPTIBLE);
5917 while (!kthread_should_stop()) {
5919 set_current_state(TASK_INTERRUPTIBLE);
5921 __set_current_state(TASK_RUNNING);
5925 #ifdef CONFIG_HOTPLUG_CPU
5927 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5931 local_irq_disable();
5932 ret = __migrate_task(p, src_cpu, dest_cpu);
5938 * Figure out where task on dead CPU should go, use force if necessary.
5939 * NOTE: interrupts should be disabled by the caller
5941 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5943 unsigned long flags;
5950 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5951 cpus_and(mask, mask, p->cpus_allowed);
5952 dest_cpu = any_online_cpu(mask);
5954 /* On any allowed CPU? */
5955 if (dest_cpu >= nr_cpu_ids)
5956 dest_cpu = any_online_cpu(p->cpus_allowed);
5958 /* No more Mr. Nice Guy. */
5959 if (dest_cpu >= nr_cpu_ids) {
5960 cpumask_t cpus_allowed;
5962 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5964 * Try to stay on the same cpuset, where the
5965 * current cpuset may be a subset of all cpus.
5966 * The cpuset_cpus_allowed_locked() variant of
5967 * cpuset_cpus_allowed() will not block. It must be
5968 * called within calls to cpuset_lock/cpuset_unlock.
5970 rq = task_rq_lock(p, &flags);
5971 p->cpus_allowed = cpus_allowed;
5972 dest_cpu = any_online_cpu(p->cpus_allowed);
5973 task_rq_unlock(rq, &flags);
5976 * Don't tell them about moving exiting tasks or
5977 * kernel threads (both mm NULL), since they never
5980 if (p->mm && printk_ratelimit()) {
5981 printk(KERN_INFO "process %d (%s) no "
5982 "longer affine to cpu%d\n",
5983 task_pid_nr(p), p->comm, dead_cpu);
5986 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5990 * While a dead CPU has no uninterruptible tasks queued at this point,
5991 * it might still have a nonzero ->nr_uninterruptible counter, because
5992 * for performance reasons the counter is not stricly tracking tasks to
5993 * their home CPUs. So we just add the counter to another CPU's counter,
5994 * to keep the global sum constant after CPU-down:
5996 static void migrate_nr_uninterruptible(struct rq *rq_src)
5998 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5999 unsigned long flags;
6001 local_irq_save(flags);
6002 double_rq_lock(rq_src, rq_dest);
6003 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6004 rq_src->nr_uninterruptible = 0;
6005 double_rq_unlock(rq_src, rq_dest);
6006 local_irq_restore(flags);
6009 /* Run through task list and migrate tasks from the dead cpu. */
6010 static void migrate_live_tasks(int src_cpu)
6012 struct task_struct *p, *t;
6014 read_lock(&tasklist_lock);
6016 do_each_thread(t, p) {
6020 if (task_cpu(p) == src_cpu)
6021 move_task_off_dead_cpu(src_cpu, p);
6022 } while_each_thread(t, p);
6024 read_unlock(&tasklist_lock);
6028 * Schedules idle task to be the next runnable task on current CPU.
6029 * It does so by boosting its priority to highest possible.
6030 * Used by CPU offline code.
6032 void sched_idle_next(void)
6034 int this_cpu = smp_processor_id();
6035 struct rq *rq = cpu_rq(this_cpu);
6036 struct task_struct *p = rq->idle;
6037 unsigned long flags;
6039 /* cpu has to be offline */
6040 BUG_ON(cpu_online(this_cpu));
6043 * Strictly not necessary since rest of the CPUs are stopped by now
6044 * and interrupts disabled on the current cpu.
6046 spin_lock_irqsave(&rq->lock, flags);
6048 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6050 update_rq_clock(rq);
6051 activate_task(rq, p, 0);
6053 spin_unlock_irqrestore(&rq->lock, flags);
6057 * Ensures that the idle task is using init_mm right before its cpu goes
6060 void idle_task_exit(void)
6062 struct mm_struct *mm = current->active_mm;
6064 BUG_ON(cpu_online(smp_processor_id()));
6067 switch_mm(mm, &init_mm, current);
6071 /* called under rq->lock with disabled interrupts */
6072 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6074 struct rq *rq = cpu_rq(dead_cpu);
6076 /* Must be exiting, otherwise would be on tasklist. */
6077 BUG_ON(!p->exit_state);
6079 /* Cannot have done final schedule yet: would have vanished. */
6080 BUG_ON(p->state == TASK_DEAD);
6085 * Drop lock around migration; if someone else moves it,
6086 * that's OK. No task can be added to this CPU, so iteration is
6089 spin_unlock_irq(&rq->lock);
6090 move_task_off_dead_cpu(dead_cpu, p);
6091 spin_lock_irq(&rq->lock);
6096 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6097 static void migrate_dead_tasks(unsigned int dead_cpu)
6099 struct rq *rq = cpu_rq(dead_cpu);
6100 struct task_struct *next;
6103 if (!rq->nr_running)
6105 update_rq_clock(rq);
6106 next = pick_next_task(rq, rq->curr);
6109 next->sched_class->put_prev_task(rq, next);
6110 migrate_dead(dead_cpu, next);
6114 #endif /* CONFIG_HOTPLUG_CPU */
6116 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6118 static struct ctl_table sd_ctl_dir[] = {
6120 .procname = "sched_domain",
6126 static struct ctl_table sd_ctl_root[] = {
6128 .ctl_name = CTL_KERN,
6129 .procname = "kernel",
6131 .child = sd_ctl_dir,
6136 static struct ctl_table *sd_alloc_ctl_entry(int n)
6138 struct ctl_table *entry =
6139 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6144 static void sd_free_ctl_entry(struct ctl_table **tablep)
6146 struct ctl_table *entry;
6149 * In the intermediate directories, both the child directory and
6150 * procname are dynamically allocated and could fail but the mode
6151 * will always be set. In the lowest directory the names are
6152 * static strings and all have proc handlers.
6154 for (entry = *tablep; entry->mode; entry++) {
6156 sd_free_ctl_entry(&entry->child);
6157 if (entry->proc_handler == NULL)
6158 kfree(entry->procname);
6166 set_table_entry(struct ctl_table *entry,
6167 const char *procname, void *data, int maxlen,
6168 mode_t mode, proc_handler *proc_handler)
6170 entry->procname = procname;
6172 entry->maxlen = maxlen;
6174 entry->proc_handler = proc_handler;
6177 static struct ctl_table *
6178 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6180 struct ctl_table *table = sd_alloc_ctl_entry(12);
6185 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6186 sizeof(long), 0644, proc_doulongvec_minmax);
6187 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6188 sizeof(long), 0644, proc_doulongvec_minmax);
6189 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6190 sizeof(int), 0644, proc_dointvec_minmax);
6191 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6192 sizeof(int), 0644, proc_dointvec_minmax);
6193 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6194 sizeof(int), 0644, proc_dointvec_minmax);
6195 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6196 sizeof(int), 0644, proc_dointvec_minmax);
6197 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6198 sizeof(int), 0644, proc_dointvec_minmax);
6199 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6200 sizeof(int), 0644, proc_dointvec_minmax);
6201 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6202 sizeof(int), 0644, proc_dointvec_minmax);
6203 set_table_entry(&table[9], "cache_nice_tries",
6204 &sd->cache_nice_tries,
6205 sizeof(int), 0644, proc_dointvec_minmax);
6206 set_table_entry(&table[10], "flags", &sd->flags,
6207 sizeof(int), 0644, proc_dointvec_minmax);
6208 /* &table[11] is terminator */
6213 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6215 struct ctl_table *entry, *table;
6216 struct sched_domain *sd;
6217 int domain_num = 0, i;
6220 for_each_domain(cpu, sd)
6222 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6227 for_each_domain(cpu, sd) {
6228 snprintf(buf, 32, "domain%d", i);
6229 entry->procname = kstrdup(buf, GFP_KERNEL);
6231 entry->child = sd_alloc_ctl_domain_table(sd);
6238 static struct ctl_table_header *sd_sysctl_header;
6239 static void register_sched_domain_sysctl(void)
6241 int i, cpu_num = num_online_cpus();
6242 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6245 WARN_ON(sd_ctl_dir[0].child);
6246 sd_ctl_dir[0].child = entry;
6251 for_each_online_cpu(i) {
6252 snprintf(buf, 32, "cpu%d", i);
6253 entry->procname = kstrdup(buf, GFP_KERNEL);
6255 entry->child = sd_alloc_ctl_cpu_table(i);
6259 WARN_ON(sd_sysctl_header);
6260 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6263 /* may be called multiple times per register */
6264 static void unregister_sched_domain_sysctl(void)
6266 if (sd_sysctl_header)
6267 unregister_sysctl_table(sd_sysctl_header);
6268 sd_sysctl_header = NULL;
6269 if (sd_ctl_dir[0].child)
6270 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6273 static void register_sched_domain_sysctl(void)
6276 static void unregister_sched_domain_sysctl(void)
6281 static void set_rq_online(struct rq *rq)
6284 const struct sched_class *class;
6286 cpu_set(rq->cpu, rq->rd->online);
6289 for_each_class(class) {
6290 if (class->rq_online)
6291 class->rq_online(rq);
6296 static void set_rq_offline(struct rq *rq)
6299 const struct sched_class *class;
6301 for_each_class(class) {
6302 if (class->rq_offline)
6303 class->rq_offline(rq);
6306 cpu_clear(rq->cpu, rq->rd->online);
6312 * migration_call - callback that gets triggered when a CPU is added.
6313 * Here we can start up the necessary migration thread for the new CPU.
6315 static int __cpuinit
6316 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6318 struct task_struct *p;
6319 int cpu = (long)hcpu;
6320 unsigned long flags;
6325 case CPU_UP_PREPARE:
6326 case CPU_UP_PREPARE_FROZEN:
6327 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6330 kthread_bind(p, cpu);
6331 /* Must be high prio: stop_machine expects to yield to it. */
6332 rq = task_rq_lock(p, &flags);
6333 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6334 task_rq_unlock(rq, &flags);
6335 cpu_rq(cpu)->migration_thread = p;
6339 case CPU_ONLINE_FROZEN:
6340 /* Strictly unnecessary, as first user will wake it. */
6341 wake_up_process(cpu_rq(cpu)->migration_thread);
6343 /* Update our root-domain */
6345 spin_lock_irqsave(&rq->lock, flags);
6347 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6351 spin_unlock_irqrestore(&rq->lock, flags);
6354 #ifdef CONFIG_HOTPLUG_CPU
6355 case CPU_UP_CANCELED:
6356 case CPU_UP_CANCELED_FROZEN:
6357 if (!cpu_rq(cpu)->migration_thread)
6359 /* Unbind it from offline cpu so it can run. Fall thru. */
6360 kthread_bind(cpu_rq(cpu)->migration_thread,
6361 any_online_cpu(cpu_online_map));
6362 kthread_stop(cpu_rq(cpu)->migration_thread);
6363 cpu_rq(cpu)->migration_thread = NULL;
6367 case CPU_DEAD_FROZEN:
6368 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6369 migrate_live_tasks(cpu);
6371 kthread_stop(rq->migration_thread);
6372 rq->migration_thread = NULL;
6373 /* Idle task back to normal (off runqueue, low prio) */
6374 spin_lock_irq(&rq->lock);
6375 update_rq_clock(rq);
6376 deactivate_task(rq, rq->idle, 0);
6377 rq->idle->static_prio = MAX_PRIO;
6378 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6379 rq->idle->sched_class = &idle_sched_class;
6380 migrate_dead_tasks(cpu);
6381 spin_unlock_irq(&rq->lock);
6383 migrate_nr_uninterruptible(rq);
6384 BUG_ON(rq->nr_running != 0);
6387 * No need to migrate the tasks: it was best-effort if
6388 * they didn't take sched_hotcpu_mutex. Just wake up
6391 spin_lock_irq(&rq->lock);
6392 while (!list_empty(&rq->migration_queue)) {
6393 struct migration_req *req;
6395 req = list_entry(rq->migration_queue.next,
6396 struct migration_req, list);
6397 list_del_init(&req->list);
6398 complete(&req->done);
6400 spin_unlock_irq(&rq->lock);
6404 case CPU_DYING_FROZEN:
6405 /* Update our root-domain */
6407 spin_lock_irqsave(&rq->lock, flags);
6409 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6412 spin_unlock_irqrestore(&rq->lock, flags);
6419 /* Register at highest priority so that task migration (migrate_all_tasks)
6420 * happens before everything else.
6422 static struct notifier_block __cpuinitdata migration_notifier = {
6423 .notifier_call = migration_call,
6427 static int __init migration_init(void)
6429 void *cpu = (void *)(long)smp_processor_id();
6432 /* Start one for the boot CPU: */
6433 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6434 BUG_ON(err == NOTIFY_BAD);
6435 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6436 register_cpu_notifier(&migration_notifier);
6440 early_initcall(migration_init);
6445 #ifdef CONFIG_SCHED_DEBUG
6447 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6460 case SD_LV_ALLNODES:
6469 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6470 cpumask_t *groupmask)
6472 struct sched_group *group = sd->groups;
6475 cpulist_scnprintf(str, sizeof(str), sd->span);
6476 cpus_clear(*groupmask);
6478 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6480 if (!(sd->flags & SD_LOAD_BALANCE)) {
6481 printk("does not load-balance\n");
6483 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6488 printk(KERN_CONT "span %s level %s\n",
6489 str, sd_level_to_string(sd->level));
6491 if (!cpu_isset(cpu, sd->span)) {
6492 printk(KERN_ERR "ERROR: domain->span does not contain "
6495 if (!cpu_isset(cpu, group->cpumask)) {
6496 printk(KERN_ERR "ERROR: domain->groups does not contain"
6500 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6504 printk(KERN_ERR "ERROR: group is NULL\n");
6508 if (!group->__cpu_power) {
6509 printk(KERN_CONT "\n");
6510 printk(KERN_ERR "ERROR: domain->cpu_power not "
6515 if (!cpus_weight(group->cpumask)) {
6516 printk(KERN_CONT "\n");
6517 printk(KERN_ERR "ERROR: empty group\n");
6521 if (cpus_intersects(*groupmask, group->cpumask)) {
6522 printk(KERN_CONT "\n");
6523 printk(KERN_ERR "ERROR: repeated CPUs\n");
6527 cpus_or(*groupmask, *groupmask, group->cpumask);
6529 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6530 printk(KERN_CONT " %s", str);
6532 group = group->next;
6533 } while (group != sd->groups);
6534 printk(KERN_CONT "\n");
6536 if (!cpus_equal(sd->span, *groupmask))
6537 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6539 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6540 printk(KERN_ERR "ERROR: parent span is not a superset "
6541 "of domain->span\n");
6545 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6547 cpumask_t *groupmask;
6551 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6555 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6557 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6559 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6564 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6573 #else /* !CONFIG_SCHED_DEBUG */
6574 # define sched_domain_debug(sd, cpu) do { } while (0)
6575 #endif /* CONFIG_SCHED_DEBUG */
6577 static int sd_degenerate(struct sched_domain *sd)
6579 if (cpus_weight(sd->span) == 1)
6582 /* Following flags need at least 2 groups */
6583 if (sd->flags & (SD_LOAD_BALANCE |
6584 SD_BALANCE_NEWIDLE |
6588 SD_SHARE_PKG_RESOURCES)) {
6589 if (sd->groups != sd->groups->next)
6593 /* Following flags don't use groups */
6594 if (sd->flags & (SD_WAKE_IDLE |
6603 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6605 unsigned long cflags = sd->flags, pflags = parent->flags;
6607 if (sd_degenerate(parent))
6610 if (!cpus_equal(sd->span, parent->span))
6613 /* Does parent contain flags not in child? */
6614 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6615 if (cflags & SD_WAKE_AFFINE)
6616 pflags &= ~SD_WAKE_BALANCE;
6617 /* Flags needing groups don't count if only 1 group in parent */
6618 if (parent->groups == parent->groups->next) {
6619 pflags &= ~(SD_LOAD_BALANCE |
6620 SD_BALANCE_NEWIDLE |
6624 SD_SHARE_PKG_RESOURCES);
6626 if (~cflags & pflags)
6632 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6634 unsigned long flags;
6636 spin_lock_irqsave(&rq->lock, flags);
6639 struct root_domain *old_rd = rq->rd;
6641 if (cpu_isset(rq->cpu, old_rd->online))
6644 cpu_clear(rq->cpu, old_rd->span);
6646 if (atomic_dec_and_test(&old_rd->refcount))
6650 atomic_inc(&rd->refcount);
6653 cpu_set(rq->cpu, rd->span);
6654 if (cpu_isset(rq->cpu, cpu_online_map))
6657 spin_unlock_irqrestore(&rq->lock, flags);
6660 static void init_rootdomain(struct root_domain *rd)
6662 memset(rd, 0, sizeof(*rd));
6664 cpus_clear(rd->span);
6665 cpus_clear(rd->online);
6667 cpupri_init(&rd->cpupri);
6670 static void init_defrootdomain(void)
6672 init_rootdomain(&def_root_domain);
6673 atomic_set(&def_root_domain.refcount, 1);
6676 static struct root_domain *alloc_rootdomain(void)
6678 struct root_domain *rd;
6680 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6684 init_rootdomain(rd);
6690 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6691 * hold the hotplug lock.
6694 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6696 struct rq *rq = cpu_rq(cpu);
6697 struct sched_domain *tmp;
6699 /* Remove the sched domains which do not contribute to scheduling. */
6700 for (tmp = sd; tmp; tmp = tmp->parent) {
6701 struct sched_domain *parent = tmp->parent;
6704 if (sd_parent_degenerate(tmp, parent)) {
6705 tmp->parent = parent->parent;
6707 parent->parent->child = tmp;
6711 if (sd && sd_degenerate(sd)) {
6717 sched_domain_debug(sd, cpu);
6719 rq_attach_root(rq, rd);
6720 rcu_assign_pointer(rq->sd, sd);
6723 /* cpus with isolated domains */
6724 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6726 /* Setup the mask of cpus configured for isolated domains */
6727 static int __init isolated_cpu_setup(char *str)
6729 static int __initdata ints[NR_CPUS];
6732 str = get_options(str, ARRAY_SIZE(ints), ints);
6733 cpus_clear(cpu_isolated_map);
6734 for (i = 1; i <= ints[0]; i++)
6735 if (ints[i] < NR_CPUS)
6736 cpu_set(ints[i], cpu_isolated_map);
6740 __setup("isolcpus=", isolated_cpu_setup);
6743 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6744 * to a function which identifies what group(along with sched group) a CPU
6745 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6746 * (due to the fact that we keep track of groups covered with a cpumask_t).
6748 * init_sched_build_groups will build a circular linked list of the groups
6749 * covered by the given span, and will set each group's ->cpumask correctly,
6750 * and ->cpu_power to 0.
6753 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6754 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6755 struct sched_group **sg,
6756 cpumask_t *tmpmask),
6757 cpumask_t *covered, cpumask_t *tmpmask)
6759 struct sched_group *first = NULL, *last = NULL;
6762 cpus_clear(*covered);
6764 for_each_cpu_mask_nr(i, *span) {
6765 struct sched_group *sg;
6766 int group = group_fn(i, cpu_map, &sg, tmpmask);
6769 if (cpu_isset(i, *covered))
6772 cpus_clear(sg->cpumask);
6773 sg->__cpu_power = 0;
6775 for_each_cpu_mask_nr(j, *span) {
6776 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6779 cpu_set(j, *covered);
6780 cpu_set(j, sg->cpumask);
6791 #define SD_NODES_PER_DOMAIN 16
6796 * find_next_best_node - find the next node to include in a sched_domain
6797 * @node: node whose sched_domain we're building
6798 * @used_nodes: nodes already in the sched_domain
6800 * Find the next node to include in a given scheduling domain. Simply
6801 * finds the closest node not already in the @used_nodes map.
6803 * Should use nodemask_t.
6805 static int find_next_best_node(int node, nodemask_t *used_nodes)
6807 int i, n, val, min_val, best_node = 0;
6811 for (i = 0; i < nr_node_ids; i++) {
6812 /* Start at @node */
6813 n = (node + i) % nr_node_ids;
6815 if (!nr_cpus_node(n))
6818 /* Skip already used nodes */
6819 if (node_isset(n, *used_nodes))
6822 /* Simple min distance search */
6823 val = node_distance(node, n);
6825 if (val < min_val) {
6831 node_set(best_node, *used_nodes);
6836 * sched_domain_node_span - get a cpumask for a node's sched_domain
6837 * @node: node whose cpumask we're constructing
6838 * @span: resulting cpumask
6840 * Given a node, construct a good cpumask for its sched_domain to span. It
6841 * should be one that prevents unnecessary balancing, but also spreads tasks
6844 static void sched_domain_node_span(int node, cpumask_t *span)
6846 nodemask_t used_nodes;
6847 node_to_cpumask_ptr(nodemask, node);
6851 nodes_clear(used_nodes);
6853 cpus_or(*span, *span, *nodemask);
6854 node_set(node, used_nodes);
6856 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6857 int next_node = find_next_best_node(node, &used_nodes);
6859 node_to_cpumask_ptr_next(nodemask, next_node);
6860 cpus_or(*span, *span, *nodemask);
6863 #endif /* CONFIG_NUMA */
6865 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6868 * SMT sched-domains:
6870 #ifdef CONFIG_SCHED_SMT
6871 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6872 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6875 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6879 *sg = &per_cpu(sched_group_cpus, cpu);
6882 #endif /* CONFIG_SCHED_SMT */
6885 * multi-core sched-domains:
6887 #ifdef CONFIG_SCHED_MC
6888 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6889 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6890 #endif /* CONFIG_SCHED_MC */
6892 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6894 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6899 *mask = per_cpu(cpu_sibling_map, cpu);
6900 cpus_and(*mask, *mask, *cpu_map);
6901 group = first_cpu(*mask);
6903 *sg = &per_cpu(sched_group_core, group);
6906 #elif defined(CONFIG_SCHED_MC)
6908 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6912 *sg = &per_cpu(sched_group_core, cpu);
6917 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6918 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6921 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6925 #ifdef CONFIG_SCHED_MC
6926 *mask = cpu_coregroup_map(cpu);
6927 cpus_and(*mask, *mask, *cpu_map);
6928 group = first_cpu(*mask);
6929 #elif defined(CONFIG_SCHED_SMT)
6930 *mask = per_cpu(cpu_sibling_map, cpu);
6931 cpus_and(*mask, *mask, *cpu_map);
6932 group = first_cpu(*mask);
6937 *sg = &per_cpu(sched_group_phys, group);
6943 * The init_sched_build_groups can't handle what we want to do with node
6944 * groups, so roll our own. Now each node has its own list of groups which
6945 * gets dynamically allocated.
6947 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6948 static struct sched_group ***sched_group_nodes_bycpu;
6950 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6951 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6953 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6954 struct sched_group **sg, cpumask_t *nodemask)
6958 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6959 cpus_and(*nodemask, *nodemask, *cpu_map);
6960 group = first_cpu(*nodemask);
6963 *sg = &per_cpu(sched_group_allnodes, group);
6967 static void init_numa_sched_groups_power(struct sched_group *group_head)
6969 struct sched_group *sg = group_head;
6975 for_each_cpu_mask_nr(j, sg->cpumask) {
6976 struct sched_domain *sd;
6978 sd = &per_cpu(phys_domains, j);
6979 if (j != first_cpu(sd->groups->cpumask)) {
6981 * Only add "power" once for each
6987 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6990 } while (sg != group_head);
6992 #endif /* CONFIG_NUMA */
6995 /* Free memory allocated for various sched_group structures */
6996 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7000 for_each_cpu_mask_nr(cpu, *cpu_map) {
7001 struct sched_group **sched_group_nodes
7002 = sched_group_nodes_bycpu[cpu];
7004 if (!sched_group_nodes)
7007 for (i = 0; i < nr_node_ids; i++) {
7008 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7010 *nodemask = node_to_cpumask(i);
7011 cpus_and(*nodemask, *nodemask, *cpu_map);
7012 if (cpus_empty(*nodemask))
7022 if (oldsg != sched_group_nodes[i])
7025 kfree(sched_group_nodes);
7026 sched_group_nodes_bycpu[cpu] = NULL;
7029 #else /* !CONFIG_NUMA */
7030 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7033 #endif /* CONFIG_NUMA */
7036 * Initialize sched groups cpu_power.
7038 * cpu_power indicates the capacity of sched group, which is used while
7039 * distributing the load between different sched groups in a sched domain.
7040 * Typically cpu_power for all the groups in a sched domain will be same unless
7041 * there are asymmetries in the topology. If there are asymmetries, group
7042 * having more cpu_power will pickup more load compared to the group having
7045 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7046 * the maximum number of tasks a group can handle in the presence of other idle
7047 * or lightly loaded groups in the same sched domain.
7049 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7051 struct sched_domain *child;
7052 struct sched_group *group;
7054 WARN_ON(!sd || !sd->groups);
7056 if (cpu != first_cpu(sd->groups->cpumask))
7061 sd->groups->__cpu_power = 0;
7064 * For perf policy, if the groups in child domain share resources
7065 * (for example cores sharing some portions of the cache hierarchy
7066 * or SMT), then set this domain groups cpu_power such that each group
7067 * can handle only one task, when there are other idle groups in the
7068 * same sched domain.
7070 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7072 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7073 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7078 * add cpu_power of each child group to this groups cpu_power
7080 group = child->groups;
7082 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7083 group = group->next;
7084 } while (group != child->groups);
7088 * Initializers for schedule domains
7089 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7092 #define SD_INIT(sd, type) sd_init_##type(sd)
7093 #define SD_INIT_FUNC(type) \
7094 static noinline void sd_init_##type(struct sched_domain *sd) \
7096 memset(sd, 0, sizeof(*sd)); \
7097 *sd = SD_##type##_INIT; \
7098 sd->level = SD_LV_##type; \
7103 SD_INIT_FUNC(ALLNODES)
7106 #ifdef CONFIG_SCHED_SMT
7107 SD_INIT_FUNC(SIBLING)
7109 #ifdef CONFIG_SCHED_MC
7114 * To minimize stack usage kmalloc room for cpumasks and share the
7115 * space as the usage in build_sched_domains() dictates. Used only
7116 * if the amount of space is significant.
7119 cpumask_t tmpmask; /* make this one first */
7122 cpumask_t this_sibling_map;
7123 cpumask_t this_core_map;
7125 cpumask_t send_covered;
7128 cpumask_t domainspan;
7130 cpumask_t notcovered;
7135 #define SCHED_CPUMASK_ALLOC 1
7136 #define SCHED_CPUMASK_FREE(v) kfree(v)
7137 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7139 #define SCHED_CPUMASK_ALLOC 0
7140 #define SCHED_CPUMASK_FREE(v)
7141 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7144 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7145 ((unsigned long)(a) + offsetof(struct allmasks, v))
7147 static int default_relax_domain_level = -1;
7149 static int __init setup_relax_domain_level(char *str)
7153 val = simple_strtoul(str, NULL, 0);
7154 if (val < SD_LV_MAX)
7155 default_relax_domain_level = val;
7159 __setup("relax_domain_level=", setup_relax_domain_level);
7161 static void set_domain_attribute(struct sched_domain *sd,
7162 struct sched_domain_attr *attr)
7166 if (!attr || attr->relax_domain_level < 0) {
7167 if (default_relax_domain_level < 0)
7170 request = default_relax_domain_level;
7172 request = attr->relax_domain_level;
7173 if (request < sd->level) {
7174 /* turn off idle balance on this domain */
7175 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7177 /* turn on idle balance on this domain */
7178 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7183 * Build sched domains for a given set of cpus and attach the sched domains
7184 * to the individual cpus
7186 static int __build_sched_domains(const cpumask_t *cpu_map,
7187 struct sched_domain_attr *attr)
7190 struct root_domain *rd;
7191 SCHED_CPUMASK_DECLARE(allmasks);
7194 struct sched_group **sched_group_nodes = NULL;
7195 int sd_allnodes = 0;
7198 * Allocate the per-node list of sched groups
7200 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7202 if (!sched_group_nodes) {
7203 printk(KERN_WARNING "Can not alloc sched group node list\n");
7208 rd = alloc_rootdomain();
7210 printk(KERN_WARNING "Cannot alloc root domain\n");
7212 kfree(sched_group_nodes);
7217 #if SCHED_CPUMASK_ALLOC
7218 /* get space for all scratch cpumask variables */
7219 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7221 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7224 kfree(sched_group_nodes);
7229 tmpmask = (cpumask_t *)allmasks;
7233 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7237 * Set up domains for cpus specified by the cpu_map.
7239 for_each_cpu_mask_nr(i, *cpu_map) {
7240 struct sched_domain *sd = NULL, *p;
7241 SCHED_CPUMASK_VAR(nodemask, allmasks);
7243 *nodemask = node_to_cpumask(cpu_to_node(i));
7244 cpus_and(*nodemask, *nodemask, *cpu_map);
7247 if (cpus_weight(*cpu_map) >
7248 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7249 sd = &per_cpu(allnodes_domains, i);
7250 SD_INIT(sd, ALLNODES);
7251 set_domain_attribute(sd, attr);
7252 sd->span = *cpu_map;
7253 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7259 sd = &per_cpu(node_domains, i);
7261 set_domain_attribute(sd, attr);
7262 sched_domain_node_span(cpu_to_node(i), &sd->span);
7266 cpus_and(sd->span, sd->span, *cpu_map);
7270 sd = &per_cpu(phys_domains, i);
7272 set_domain_attribute(sd, attr);
7273 sd->span = *nodemask;
7277 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7279 #ifdef CONFIG_SCHED_MC
7281 sd = &per_cpu(core_domains, i);
7283 set_domain_attribute(sd, attr);
7284 sd->span = cpu_coregroup_map(i);
7285 cpus_and(sd->span, sd->span, *cpu_map);
7288 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7291 #ifdef CONFIG_SCHED_SMT
7293 sd = &per_cpu(cpu_domains, i);
7294 SD_INIT(sd, SIBLING);
7295 set_domain_attribute(sd, attr);
7296 sd->span = per_cpu(cpu_sibling_map, i);
7297 cpus_and(sd->span, sd->span, *cpu_map);
7300 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7304 #ifdef CONFIG_SCHED_SMT
7305 /* Set up CPU (sibling) groups */
7306 for_each_cpu_mask_nr(i, *cpu_map) {
7307 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7308 SCHED_CPUMASK_VAR(send_covered, allmasks);
7310 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7311 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7312 if (i != first_cpu(*this_sibling_map))
7315 init_sched_build_groups(this_sibling_map, cpu_map,
7317 send_covered, tmpmask);
7321 #ifdef CONFIG_SCHED_MC
7322 /* Set up multi-core groups */
7323 for_each_cpu_mask_nr(i, *cpu_map) {
7324 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7325 SCHED_CPUMASK_VAR(send_covered, allmasks);
7327 *this_core_map = cpu_coregroup_map(i);
7328 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7329 if (i != first_cpu(*this_core_map))
7332 init_sched_build_groups(this_core_map, cpu_map,
7334 send_covered, tmpmask);
7338 /* Set up physical groups */
7339 for (i = 0; i < nr_node_ids; i++) {
7340 SCHED_CPUMASK_VAR(nodemask, allmasks);
7341 SCHED_CPUMASK_VAR(send_covered, allmasks);
7343 *nodemask = node_to_cpumask(i);
7344 cpus_and(*nodemask, *nodemask, *cpu_map);
7345 if (cpus_empty(*nodemask))
7348 init_sched_build_groups(nodemask, cpu_map,
7350 send_covered, tmpmask);
7354 /* Set up node groups */
7356 SCHED_CPUMASK_VAR(send_covered, allmasks);
7358 init_sched_build_groups(cpu_map, cpu_map,
7359 &cpu_to_allnodes_group,
7360 send_covered, tmpmask);
7363 for (i = 0; i < nr_node_ids; i++) {
7364 /* Set up node groups */
7365 struct sched_group *sg, *prev;
7366 SCHED_CPUMASK_VAR(nodemask, allmasks);
7367 SCHED_CPUMASK_VAR(domainspan, allmasks);
7368 SCHED_CPUMASK_VAR(covered, allmasks);
7371 *nodemask = node_to_cpumask(i);
7372 cpus_clear(*covered);
7374 cpus_and(*nodemask, *nodemask, *cpu_map);
7375 if (cpus_empty(*nodemask)) {
7376 sched_group_nodes[i] = NULL;
7380 sched_domain_node_span(i, domainspan);
7381 cpus_and(*domainspan, *domainspan, *cpu_map);
7383 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7385 printk(KERN_WARNING "Can not alloc domain group for "
7389 sched_group_nodes[i] = sg;
7390 for_each_cpu_mask_nr(j, *nodemask) {
7391 struct sched_domain *sd;
7393 sd = &per_cpu(node_domains, j);
7396 sg->__cpu_power = 0;
7397 sg->cpumask = *nodemask;
7399 cpus_or(*covered, *covered, *nodemask);
7402 for (j = 0; j < nr_node_ids; j++) {
7403 SCHED_CPUMASK_VAR(notcovered, allmasks);
7404 int n = (i + j) % nr_node_ids;
7405 node_to_cpumask_ptr(pnodemask, n);
7407 cpus_complement(*notcovered, *covered);
7408 cpus_and(*tmpmask, *notcovered, *cpu_map);
7409 cpus_and(*tmpmask, *tmpmask, *domainspan);
7410 if (cpus_empty(*tmpmask))
7413 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7414 if (cpus_empty(*tmpmask))
7417 sg = kmalloc_node(sizeof(struct sched_group),
7421 "Can not alloc domain group for node %d\n", j);
7424 sg->__cpu_power = 0;
7425 sg->cpumask = *tmpmask;
7426 sg->next = prev->next;
7427 cpus_or(*covered, *covered, *tmpmask);
7434 /* Calculate CPU power for physical packages and nodes */
7435 #ifdef CONFIG_SCHED_SMT
7436 for_each_cpu_mask_nr(i, *cpu_map) {
7437 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7439 init_sched_groups_power(i, sd);
7442 #ifdef CONFIG_SCHED_MC
7443 for_each_cpu_mask_nr(i, *cpu_map) {
7444 struct sched_domain *sd = &per_cpu(core_domains, i);
7446 init_sched_groups_power(i, sd);
7450 for_each_cpu_mask_nr(i, *cpu_map) {
7451 struct sched_domain *sd = &per_cpu(phys_domains, i);
7453 init_sched_groups_power(i, sd);
7457 for (i = 0; i < nr_node_ids; i++)
7458 init_numa_sched_groups_power(sched_group_nodes[i]);
7461 struct sched_group *sg;
7463 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7465 init_numa_sched_groups_power(sg);
7469 /* Attach the domains */
7470 for_each_cpu_mask_nr(i, *cpu_map) {
7471 struct sched_domain *sd;
7472 #ifdef CONFIG_SCHED_SMT
7473 sd = &per_cpu(cpu_domains, i);
7474 #elif defined(CONFIG_SCHED_MC)
7475 sd = &per_cpu(core_domains, i);
7477 sd = &per_cpu(phys_domains, i);
7479 cpu_attach_domain(sd, rd, i);
7482 SCHED_CPUMASK_FREE((void *)allmasks);
7487 free_sched_groups(cpu_map, tmpmask);
7488 SCHED_CPUMASK_FREE((void *)allmasks);
7493 static int build_sched_domains(const cpumask_t *cpu_map)
7495 return __build_sched_domains(cpu_map, NULL);
7498 static cpumask_t *doms_cur; /* current sched domains */
7499 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7500 static struct sched_domain_attr *dattr_cur;
7501 /* attribues of custom domains in 'doms_cur' */
7504 * Special case: If a kmalloc of a doms_cur partition (array of
7505 * cpumask_t) fails, then fallback to a single sched domain,
7506 * as determined by the single cpumask_t fallback_doms.
7508 static cpumask_t fallback_doms;
7510 void __attribute__((weak)) arch_update_cpu_topology(void)
7515 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7516 * For now this just excludes isolated cpus, but could be used to
7517 * exclude other special cases in the future.
7519 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7523 arch_update_cpu_topology();
7525 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7527 doms_cur = &fallback_doms;
7528 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7530 err = build_sched_domains(doms_cur);
7531 register_sched_domain_sysctl();
7536 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7539 free_sched_groups(cpu_map, tmpmask);
7543 * Detach sched domains from a group of cpus specified in cpu_map
7544 * These cpus will now be attached to the NULL domain
7546 static void detach_destroy_domains(const cpumask_t *cpu_map)
7551 unregister_sched_domain_sysctl();
7553 for_each_cpu_mask_nr(i, *cpu_map)
7554 cpu_attach_domain(NULL, &def_root_domain, i);
7555 synchronize_sched();
7556 arch_destroy_sched_domains(cpu_map, &tmpmask);
7559 /* handle null as "default" */
7560 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7561 struct sched_domain_attr *new, int idx_new)
7563 struct sched_domain_attr tmp;
7570 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7571 new ? (new + idx_new) : &tmp,
7572 sizeof(struct sched_domain_attr));
7576 * Partition sched domains as specified by the 'ndoms_new'
7577 * cpumasks in the array doms_new[] of cpumasks. This compares
7578 * doms_new[] to the current sched domain partitioning, doms_cur[].
7579 * It destroys each deleted domain and builds each new domain.
7581 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7582 * The masks don't intersect (don't overlap.) We should setup one
7583 * sched domain for each mask. CPUs not in any of the cpumasks will
7584 * not be load balanced. If the same cpumask appears both in the
7585 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7588 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7589 * ownership of it and will kfree it when done with it. If the caller
7590 * failed the kmalloc call, then it can pass in doms_new == NULL,
7591 * and partition_sched_domains() will fallback to the single partition
7592 * 'fallback_doms', it also forces the domains to be rebuilt.
7594 * Call with hotplug lock held
7596 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7597 struct sched_domain_attr *dattr_new)
7601 mutex_lock(&sched_domains_mutex);
7603 /* always unregister in case we don't destroy any domains */
7604 unregister_sched_domain_sysctl();
7606 if (doms_new == NULL)
7609 /* Destroy deleted domains */
7610 for (i = 0; i < ndoms_cur; i++) {
7611 for (j = 0; j < ndoms_new; j++) {
7612 if (cpus_equal(doms_cur[i], doms_new[j])
7613 && dattrs_equal(dattr_cur, i, dattr_new, j))
7616 /* no match - a current sched domain not in new doms_new[] */
7617 detach_destroy_domains(doms_cur + i);
7622 if (doms_new == NULL) {
7625 doms_new = &fallback_doms;
7626 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7630 /* Build new domains */
7631 for (i = 0; i < ndoms_new; i++) {
7632 for (j = 0; j < ndoms_cur; j++) {
7633 if (cpus_equal(doms_new[i], doms_cur[j])
7634 && dattrs_equal(dattr_new, i, dattr_cur, j))
7637 /* no match - add a new doms_new */
7638 __build_sched_domains(doms_new + i,
7639 dattr_new ? dattr_new + i : NULL);
7644 /* Remember the new sched domains */
7645 if (doms_cur != &fallback_doms)
7647 kfree(dattr_cur); /* kfree(NULL) is safe */
7648 doms_cur = doms_new;
7649 dattr_cur = dattr_new;
7650 ndoms_cur = ndoms_new;
7652 register_sched_domain_sysctl();
7654 mutex_unlock(&sched_domains_mutex);
7657 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7658 int arch_reinit_sched_domains(void)
7661 rebuild_sched_domains();
7666 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7670 if (buf[0] != '0' && buf[0] != '1')
7674 sched_smt_power_savings = (buf[0] == '1');
7676 sched_mc_power_savings = (buf[0] == '1');
7678 ret = arch_reinit_sched_domains();
7680 return ret ? ret : count;
7683 #ifdef CONFIG_SCHED_MC
7684 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7687 return sprintf(page, "%u\n", sched_mc_power_savings);
7689 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7690 const char *buf, size_t count)
7692 return sched_power_savings_store(buf, count, 0);
7694 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7695 sched_mc_power_savings_show,
7696 sched_mc_power_savings_store);
7699 #ifdef CONFIG_SCHED_SMT
7700 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7703 return sprintf(page, "%u\n", sched_smt_power_savings);
7705 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7706 const char *buf, size_t count)
7708 return sched_power_savings_store(buf, count, 1);
7710 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7711 sched_smt_power_savings_show,
7712 sched_smt_power_savings_store);
7715 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7719 #ifdef CONFIG_SCHED_SMT
7721 err = sysfs_create_file(&cls->kset.kobj,
7722 &attr_sched_smt_power_savings.attr);
7724 #ifdef CONFIG_SCHED_MC
7725 if (!err && mc_capable())
7726 err = sysfs_create_file(&cls->kset.kobj,
7727 &attr_sched_mc_power_savings.attr);
7731 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7733 #ifndef CONFIG_CPUSETS
7735 * Add online and remove offline CPUs from the scheduler domains.
7736 * When cpusets are enabled they take over this function.
7738 static int update_sched_domains(struct notifier_block *nfb,
7739 unsigned long action, void *hcpu)
7743 case CPU_ONLINE_FROZEN:
7745 case CPU_DEAD_FROZEN:
7746 partition_sched_domains(0, NULL, NULL);
7755 static int update_runtime(struct notifier_block *nfb,
7756 unsigned long action, void *hcpu)
7758 int cpu = (int)(long)hcpu;
7761 case CPU_DOWN_PREPARE:
7762 case CPU_DOWN_PREPARE_FROZEN:
7763 disable_runtime(cpu_rq(cpu));
7766 case CPU_DOWN_FAILED:
7767 case CPU_DOWN_FAILED_FROZEN:
7769 case CPU_ONLINE_FROZEN:
7770 enable_runtime(cpu_rq(cpu));
7778 void __init sched_init_smp(void)
7780 cpumask_t non_isolated_cpus;
7782 #if defined(CONFIG_NUMA)
7783 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7785 BUG_ON(sched_group_nodes_bycpu == NULL);
7788 mutex_lock(&sched_domains_mutex);
7789 arch_init_sched_domains(&cpu_online_map);
7790 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7791 if (cpus_empty(non_isolated_cpus))
7792 cpu_set(smp_processor_id(), non_isolated_cpus);
7793 mutex_unlock(&sched_domains_mutex);
7796 #ifndef CONFIG_CPUSETS
7797 /* XXX: Theoretical race here - CPU may be hotplugged now */
7798 hotcpu_notifier(update_sched_domains, 0);
7801 /* RT runtime code needs to handle some hotplug events */
7802 hotcpu_notifier(update_runtime, 0);
7806 /* Move init over to a non-isolated CPU */
7807 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7809 sched_init_granularity();
7812 void __init sched_init_smp(void)
7814 sched_init_granularity();
7816 #endif /* CONFIG_SMP */
7818 int in_sched_functions(unsigned long addr)
7820 return in_lock_functions(addr) ||
7821 (addr >= (unsigned long)__sched_text_start
7822 && addr < (unsigned long)__sched_text_end);
7825 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7827 cfs_rq->tasks_timeline = RB_ROOT;
7828 INIT_LIST_HEAD(&cfs_rq->tasks);
7829 #ifdef CONFIG_FAIR_GROUP_SCHED
7832 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7835 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7837 struct rt_prio_array *array;
7840 array = &rt_rq->active;
7841 for (i = 0; i < MAX_RT_PRIO; i++) {
7842 INIT_LIST_HEAD(array->queue + i);
7843 __clear_bit(i, array->bitmap);
7845 /* delimiter for bitsearch: */
7846 __set_bit(MAX_RT_PRIO, array->bitmap);
7848 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7849 rt_rq->highest_prio = MAX_RT_PRIO;
7852 rt_rq->rt_nr_migratory = 0;
7853 rt_rq->overloaded = 0;
7857 rt_rq->rt_throttled = 0;
7858 rt_rq->rt_runtime = 0;
7859 spin_lock_init(&rt_rq->rt_runtime_lock);
7861 #ifdef CONFIG_RT_GROUP_SCHED
7862 rt_rq->rt_nr_boosted = 0;
7867 #ifdef CONFIG_FAIR_GROUP_SCHED
7868 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7869 struct sched_entity *se, int cpu, int add,
7870 struct sched_entity *parent)
7872 struct rq *rq = cpu_rq(cpu);
7873 tg->cfs_rq[cpu] = cfs_rq;
7874 init_cfs_rq(cfs_rq, rq);
7877 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7880 /* se could be NULL for init_task_group */
7885 se->cfs_rq = &rq->cfs;
7887 se->cfs_rq = parent->my_q;
7890 se->load.weight = tg->shares;
7891 se->load.inv_weight = 0;
7892 se->parent = parent;
7896 #ifdef CONFIG_RT_GROUP_SCHED
7897 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7898 struct sched_rt_entity *rt_se, int cpu, int add,
7899 struct sched_rt_entity *parent)
7901 struct rq *rq = cpu_rq(cpu);
7903 tg->rt_rq[cpu] = rt_rq;
7904 init_rt_rq(rt_rq, rq);
7906 rt_rq->rt_se = rt_se;
7907 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7909 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7911 tg->rt_se[cpu] = rt_se;
7916 rt_se->rt_rq = &rq->rt;
7918 rt_se->rt_rq = parent->my_q;
7920 rt_se->my_q = rt_rq;
7921 rt_se->parent = parent;
7922 INIT_LIST_HEAD(&rt_se->run_list);
7926 void __init sched_init(void)
7929 unsigned long alloc_size = 0, ptr;
7931 #ifdef CONFIG_FAIR_GROUP_SCHED
7932 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7934 #ifdef CONFIG_RT_GROUP_SCHED
7935 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7937 #ifdef CONFIG_USER_SCHED
7941 * As sched_init() is called before page_alloc is setup,
7942 * we use alloc_bootmem().
7945 ptr = (unsigned long)alloc_bootmem(alloc_size);
7947 #ifdef CONFIG_FAIR_GROUP_SCHED
7948 init_task_group.se = (struct sched_entity **)ptr;
7949 ptr += nr_cpu_ids * sizeof(void **);
7951 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7952 ptr += nr_cpu_ids * sizeof(void **);
7954 #ifdef CONFIG_USER_SCHED
7955 root_task_group.se = (struct sched_entity **)ptr;
7956 ptr += nr_cpu_ids * sizeof(void **);
7958 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7959 ptr += nr_cpu_ids * sizeof(void **);
7960 #endif /* CONFIG_USER_SCHED */
7961 #endif /* CONFIG_FAIR_GROUP_SCHED */
7962 #ifdef CONFIG_RT_GROUP_SCHED
7963 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7964 ptr += nr_cpu_ids * sizeof(void **);
7966 init_task_group.rt_rq = (struct rt_rq **)ptr;
7967 ptr += nr_cpu_ids * sizeof(void **);
7969 #ifdef CONFIG_USER_SCHED
7970 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7971 ptr += nr_cpu_ids * sizeof(void **);
7973 root_task_group.rt_rq = (struct rt_rq **)ptr;
7974 ptr += nr_cpu_ids * sizeof(void **);
7975 #endif /* CONFIG_USER_SCHED */
7976 #endif /* CONFIG_RT_GROUP_SCHED */
7980 init_defrootdomain();
7983 init_rt_bandwidth(&def_rt_bandwidth,
7984 global_rt_period(), global_rt_runtime());
7986 #ifdef CONFIG_RT_GROUP_SCHED
7987 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7988 global_rt_period(), global_rt_runtime());
7989 #ifdef CONFIG_USER_SCHED
7990 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7991 global_rt_period(), RUNTIME_INF);
7992 #endif /* CONFIG_USER_SCHED */
7993 #endif /* CONFIG_RT_GROUP_SCHED */
7995 #ifdef CONFIG_GROUP_SCHED
7996 list_add(&init_task_group.list, &task_groups);
7997 INIT_LIST_HEAD(&init_task_group.children);
7999 #ifdef CONFIG_USER_SCHED
8000 INIT_LIST_HEAD(&root_task_group.children);
8001 init_task_group.parent = &root_task_group;
8002 list_add(&init_task_group.siblings, &root_task_group.children);
8003 #endif /* CONFIG_USER_SCHED */
8004 #endif /* CONFIG_GROUP_SCHED */
8006 for_each_possible_cpu(i) {
8010 spin_lock_init(&rq->lock);
8012 init_cfs_rq(&rq->cfs, rq);
8013 init_rt_rq(&rq->rt, rq);
8014 #ifdef CONFIG_FAIR_GROUP_SCHED
8015 init_task_group.shares = init_task_group_load;
8016 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8017 #ifdef CONFIG_CGROUP_SCHED
8019 * How much cpu bandwidth does init_task_group get?
8021 * In case of task-groups formed thr' the cgroup filesystem, it
8022 * gets 100% of the cpu resources in the system. This overall
8023 * system cpu resource is divided among the tasks of
8024 * init_task_group and its child task-groups in a fair manner,
8025 * based on each entity's (task or task-group's) weight
8026 * (se->load.weight).
8028 * In other words, if init_task_group has 10 tasks of weight
8029 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8030 * then A0's share of the cpu resource is:
8032 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8034 * We achieve this by letting init_task_group's tasks sit
8035 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8037 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8038 #elif defined CONFIG_USER_SCHED
8039 root_task_group.shares = NICE_0_LOAD;
8040 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8042 * In case of task-groups formed thr' the user id of tasks,
8043 * init_task_group represents tasks belonging to root user.
8044 * Hence it forms a sibling of all subsequent groups formed.
8045 * In this case, init_task_group gets only a fraction of overall
8046 * system cpu resource, based on the weight assigned to root
8047 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8048 * by letting tasks of init_task_group sit in a separate cfs_rq
8049 * (init_cfs_rq) and having one entity represent this group of
8050 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8052 init_tg_cfs_entry(&init_task_group,
8053 &per_cpu(init_cfs_rq, i),
8054 &per_cpu(init_sched_entity, i), i, 1,
8055 root_task_group.se[i]);
8058 #endif /* CONFIG_FAIR_GROUP_SCHED */
8060 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8061 #ifdef CONFIG_RT_GROUP_SCHED
8062 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8063 #ifdef CONFIG_CGROUP_SCHED
8064 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8065 #elif defined CONFIG_USER_SCHED
8066 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8067 init_tg_rt_entry(&init_task_group,
8068 &per_cpu(init_rt_rq, i),
8069 &per_cpu(init_sched_rt_entity, i), i, 1,
8070 root_task_group.rt_se[i]);
8074 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8075 rq->cpu_load[j] = 0;
8079 rq->active_balance = 0;
8080 rq->next_balance = jiffies;
8084 rq->migration_thread = NULL;
8085 INIT_LIST_HEAD(&rq->migration_queue);
8086 rq_attach_root(rq, &def_root_domain);
8089 atomic_set(&rq->nr_iowait, 0);
8092 set_load_weight(&init_task);
8094 #ifdef CONFIG_PREEMPT_NOTIFIERS
8095 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8099 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8102 #ifdef CONFIG_RT_MUTEXES
8103 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8107 * The boot idle thread does lazy MMU switching as well:
8109 atomic_inc(&init_mm.mm_count);
8110 enter_lazy_tlb(&init_mm, current);
8113 * Make us the idle thread. Technically, schedule() should not be
8114 * called from this thread, however somewhere below it might be,
8115 * but because we are the idle thread, we just pick up running again
8116 * when this runqueue becomes "idle".
8118 init_idle(current, smp_processor_id());
8120 * During early bootup we pretend to be a normal task:
8122 current->sched_class = &fair_sched_class;
8124 scheduler_running = 1;
8127 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8128 void __might_sleep(char *file, int line)
8131 static unsigned long prev_jiffy; /* ratelimiting */
8133 if ((in_atomic() || irqs_disabled()) &&
8134 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8135 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8137 prev_jiffy = jiffies;
8138 printk(KERN_ERR "BUG: sleeping function called from invalid"
8139 " context at %s:%d\n", file, line);
8140 printk("in_atomic():%d, irqs_disabled():%d\n",
8141 in_atomic(), irqs_disabled());
8142 debug_show_held_locks(current);
8143 if (irqs_disabled())
8144 print_irqtrace_events(current);
8149 EXPORT_SYMBOL(__might_sleep);
8152 #ifdef CONFIG_MAGIC_SYSRQ
8153 static void normalize_task(struct rq *rq, struct task_struct *p)
8157 update_rq_clock(rq);
8158 on_rq = p->se.on_rq;
8160 deactivate_task(rq, p, 0);
8161 __setscheduler(rq, p, SCHED_NORMAL, 0);
8163 activate_task(rq, p, 0);
8164 resched_task(rq->curr);
8168 void normalize_rt_tasks(void)
8170 struct task_struct *g, *p;
8171 unsigned long flags;
8174 read_lock_irqsave(&tasklist_lock, flags);
8175 do_each_thread(g, p) {
8177 * Only normalize user tasks:
8182 p->se.exec_start = 0;
8183 #ifdef CONFIG_SCHEDSTATS
8184 p->se.wait_start = 0;
8185 p->se.sleep_start = 0;
8186 p->se.block_start = 0;
8191 * Renice negative nice level userspace
8194 if (TASK_NICE(p) < 0 && p->mm)
8195 set_user_nice(p, 0);
8199 spin_lock(&p->pi_lock);
8200 rq = __task_rq_lock(p);
8202 normalize_task(rq, p);
8204 __task_rq_unlock(rq);
8205 spin_unlock(&p->pi_lock);
8206 } while_each_thread(g, p);
8208 read_unlock_irqrestore(&tasklist_lock, flags);
8211 #endif /* CONFIG_MAGIC_SYSRQ */
8215 * These functions are only useful for the IA64 MCA handling.
8217 * They can only be called when the whole system has been
8218 * stopped - every CPU needs to be quiescent, and no scheduling
8219 * activity can take place. Using them for anything else would
8220 * be a serious bug, and as a result, they aren't even visible
8221 * under any other configuration.
8225 * curr_task - return the current task for a given cpu.
8226 * @cpu: the processor in question.
8228 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8230 struct task_struct *curr_task(int cpu)
8232 return cpu_curr(cpu);
8236 * set_curr_task - set the current task for a given cpu.
8237 * @cpu: the processor in question.
8238 * @p: the task pointer to set.
8240 * Description: This function must only be used when non-maskable interrupts
8241 * are serviced on a separate stack. It allows the architecture to switch the
8242 * notion of the current task on a cpu in a non-blocking manner. This function
8243 * must be called with all CPU's synchronized, and interrupts disabled, the
8244 * and caller must save the original value of the current task (see
8245 * curr_task() above) and restore that value before reenabling interrupts and
8246 * re-starting the system.
8248 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8250 void set_curr_task(int cpu, struct task_struct *p)
8257 #ifdef CONFIG_FAIR_GROUP_SCHED
8258 static void free_fair_sched_group(struct task_group *tg)
8262 for_each_possible_cpu(i) {
8264 kfree(tg->cfs_rq[i]);
8274 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8276 struct cfs_rq *cfs_rq;
8277 struct sched_entity *se, *parent_se;
8281 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8284 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8288 tg->shares = NICE_0_LOAD;
8290 for_each_possible_cpu(i) {
8293 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8294 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8298 se = kmalloc_node(sizeof(struct sched_entity),
8299 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8303 parent_se = parent ? parent->se[i] : NULL;
8304 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8313 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8315 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8316 &cpu_rq(cpu)->leaf_cfs_rq_list);
8319 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8321 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8323 #else /* !CONFG_FAIR_GROUP_SCHED */
8324 static inline void free_fair_sched_group(struct task_group *tg)
8329 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8334 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8338 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8341 #endif /* CONFIG_FAIR_GROUP_SCHED */
8343 #ifdef CONFIG_RT_GROUP_SCHED
8344 static void free_rt_sched_group(struct task_group *tg)
8348 destroy_rt_bandwidth(&tg->rt_bandwidth);
8350 for_each_possible_cpu(i) {
8352 kfree(tg->rt_rq[i]);
8354 kfree(tg->rt_se[i]);
8362 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8364 struct rt_rq *rt_rq;
8365 struct sched_rt_entity *rt_se, *parent_se;
8369 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8372 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8376 init_rt_bandwidth(&tg->rt_bandwidth,
8377 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8379 for_each_possible_cpu(i) {
8382 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8383 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8387 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8388 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8392 parent_se = parent ? parent->rt_se[i] : NULL;
8393 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8402 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8404 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8405 &cpu_rq(cpu)->leaf_rt_rq_list);
8408 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8410 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8412 #else /* !CONFIG_RT_GROUP_SCHED */
8413 static inline void free_rt_sched_group(struct task_group *tg)
8418 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8423 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8427 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8430 #endif /* CONFIG_RT_GROUP_SCHED */
8432 #ifdef CONFIG_GROUP_SCHED
8433 static void free_sched_group(struct task_group *tg)
8435 free_fair_sched_group(tg);
8436 free_rt_sched_group(tg);
8440 /* allocate runqueue etc for a new task group */
8441 struct task_group *sched_create_group(struct task_group *parent)
8443 struct task_group *tg;
8444 unsigned long flags;
8447 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8449 return ERR_PTR(-ENOMEM);
8451 if (!alloc_fair_sched_group(tg, parent))
8454 if (!alloc_rt_sched_group(tg, parent))
8457 spin_lock_irqsave(&task_group_lock, flags);
8458 for_each_possible_cpu(i) {
8459 register_fair_sched_group(tg, i);
8460 register_rt_sched_group(tg, i);
8462 list_add_rcu(&tg->list, &task_groups);
8464 WARN_ON(!parent); /* root should already exist */
8466 tg->parent = parent;
8467 INIT_LIST_HEAD(&tg->children);
8468 list_add_rcu(&tg->siblings, &parent->children);
8469 spin_unlock_irqrestore(&task_group_lock, flags);
8474 free_sched_group(tg);
8475 return ERR_PTR(-ENOMEM);
8478 /* rcu callback to free various structures associated with a task group */
8479 static void free_sched_group_rcu(struct rcu_head *rhp)
8481 /* now it should be safe to free those cfs_rqs */
8482 free_sched_group(container_of(rhp, struct task_group, rcu));
8485 /* Destroy runqueue etc associated with a task group */
8486 void sched_destroy_group(struct task_group *tg)
8488 unsigned long flags;
8491 spin_lock_irqsave(&task_group_lock, flags);
8492 for_each_possible_cpu(i) {
8493 unregister_fair_sched_group(tg, i);
8494 unregister_rt_sched_group(tg, i);
8496 list_del_rcu(&tg->list);
8497 list_del_rcu(&tg->siblings);
8498 spin_unlock_irqrestore(&task_group_lock, flags);
8500 /* wait for possible concurrent references to cfs_rqs complete */
8501 call_rcu(&tg->rcu, free_sched_group_rcu);
8504 /* change task's runqueue when it moves between groups.
8505 * The caller of this function should have put the task in its new group
8506 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8507 * reflect its new group.
8509 void sched_move_task(struct task_struct *tsk)
8512 unsigned long flags;
8515 rq = task_rq_lock(tsk, &flags);
8517 update_rq_clock(rq);
8519 running = task_current(rq, tsk);
8520 on_rq = tsk->se.on_rq;
8523 dequeue_task(rq, tsk, 0);
8524 if (unlikely(running))
8525 tsk->sched_class->put_prev_task(rq, tsk);
8527 set_task_rq(tsk, task_cpu(tsk));
8529 #ifdef CONFIG_FAIR_GROUP_SCHED
8530 if (tsk->sched_class->moved_group)
8531 tsk->sched_class->moved_group(tsk);
8534 if (unlikely(running))
8535 tsk->sched_class->set_curr_task(rq);
8537 enqueue_task(rq, tsk, 0);
8539 task_rq_unlock(rq, &flags);
8541 #endif /* CONFIG_GROUP_SCHED */
8543 #ifdef CONFIG_FAIR_GROUP_SCHED
8544 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8546 struct cfs_rq *cfs_rq = se->cfs_rq;
8551 dequeue_entity(cfs_rq, se, 0);
8553 se->load.weight = shares;
8554 se->load.inv_weight = 0;
8557 enqueue_entity(cfs_rq, se, 0);
8560 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8562 struct cfs_rq *cfs_rq = se->cfs_rq;
8563 struct rq *rq = cfs_rq->rq;
8564 unsigned long flags;
8566 spin_lock_irqsave(&rq->lock, flags);
8567 __set_se_shares(se, shares);
8568 spin_unlock_irqrestore(&rq->lock, flags);
8571 static DEFINE_MUTEX(shares_mutex);
8573 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8576 unsigned long flags;
8579 * We can't change the weight of the root cgroup.
8584 if (shares < MIN_SHARES)
8585 shares = MIN_SHARES;
8586 else if (shares > MAX_SHARES)
8587 shares = MAX_SHARES;
8589 mutex_lock(&shares_mutex);
8590 if (tg->shares == shares)
8593 spin_lock_irqsave(&task_group_lock, flags);
8594 for_each_possible_cpu(i)
8595 unregister_fair_sched_group(tg, i);
8596 list_del_rcu(&tg->siblings);
8597 spin_unlock_irqrestore(&task_group_lock, flags);
8599 /* wait for any ongoing reference to this group to finish */
8600 synchronize_sched();
8603 * Now we are free to modify the group's share on each cpu
8604 * w/o tripping rebalance_share or load_balance_fair.
8606 tg->shares = shares;
8607 for_each_possible_cpu(i) {
8611 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8612 set_se_shares(tg->se[i], shares);
8616 * Enable load balance activity on this group, by inserting it back on
8617 * each cpu's rq->leaf_cfs_rq_list.
8619 spin_lock_irqsave(&task_group_lock, flags);
8620 for_each_possible_cpu(i)
8621 register_fair_sched_group(tg, i);
8622 list_add_rcu(&tg->siblings, &tg->parent->children);
8623 spin_unlock_irqrestore(&task_group_lock, flags);
8625 mutex_unlock(&shares_mutex);
8629 unsigned long sched_group_shares(struct task_group *tg)
8635 #ifdef CONFIG_RT_GROUP_SCHED
8637 * Ensure that the real time constraints are schedulable.
8639 static DEFINE_MUTEX(rt_constraints_mutex);
8641 static unsigned long to_ratio(u64 period, u64 runtime)
8643 if (runtime == RUNTIME_INF)
8646 return div64_u64(runtime << 16, period);
8649 #ifdef CONFIG_CGROUP_SCHED
8650 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8652 struct task_group *tgi, *parent = tg->parent;
8653 unsigned long total = 0;
8656 if (global_rt_period() < period)
8659 return to_ratio(period, runtime) <
8660 to_ratio(global_rt_period(), global_rt_runtime());
8663 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8667 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8671 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8672 tgi->rt_bandwidth.rt_runtime);
8676 return total + to_ratio(period, runtime) <=
8677 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8678 parent->rt_bandwidth.rt_runtime);
8680 #elif defined CONFIG_USER_SCHED
8681 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8683 struct task_group *tgi;
8684 unsigned long total = 0;
8685 unsigned long global_ratio =
8686 to_ratio(global_rt_period(), global_rt_runtime());
8689 list_for_each_entry_rcu(tgi, &task_groups, list) {
8693 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8694 tgi->rt_bandwidth.rt_runtime);
8698 return total + to_ratio(period, runtime) < global_ratio;
8702 /* Must be called with tasklist_lock held */
8703 static inline int tg_has_rt_tasks(struct task_group *tg)
8705 struct task_struct *g, *p;
8706 do_each_thread(g, p) {
8707 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8709 } while_each_thread(g, p);
8713 static int tg_set_bandwidth(struct task_group *tg,
8714 u64 rt_period, u64 rt_runtime)
8718 mutex_lock(&rt_constraints_mutex);
8719 read_lock(&tasklist_lock);
8720 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8724 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8729 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8730 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8731 tg->rt_bandwidth.rt_runtime = rt_runtime;
8733 for_each_possible_cpu(i) {
8734 struct rt_rq *rt_rq = tg->rt_rq[i];
8736 spin_lock(&rt_rq->rt_runtime_lock);
8737 rt_rq->rt_runtime = rt_runtime;
8738 spin_unlock(&rt_rq->rt_runtime_lock);
8740 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8742 read_unlock(&tasklist_lock);
8743 mutex_unlock(&rt_constraints_mutex);
8748 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8750 u64 rt_runtime, rt_period;
8752 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8753 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8754 if (rt_runtime_us < 0)
8755 rt_runtime = RUNTIME_INF;
8757 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8760 long sched_group_rt_runtime(struct task_group *tg)
8764 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8767 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8768 do_div(rt_runtime_us, NSEC_PER_USEC);
8769 return rt_runtime_us;
8772 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8774 u64 rt_runtime, rt_period;
8776 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8777 rt_runtime = tg->rt_bandwidth.rt_runtime;
8782 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8785 long sched_group_rt_period(struct task_group *tg)
8789 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8790 do_div(rt_period_us, NSEC_PER_USEC);
8791 return rt_period_us;
8794 static int sched_rt_global_constraints(void)
8796 struct task_group *tg = &root_task_group;
8797 u64 rt_runtime, rt_period;
8800 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8801 rt_runtime = tg->rt_bandwidth.rt_runtime;
8803 mutex_lock(&rt_constraints_mutex);
8804 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8806 mutex_unlock(&rt_constraints_mutex);
8810 #else /* !CONFIG_RT_GROUP_SCHED */
8811 static int sched_rt_global_constraints(void)
8813 unsigned long flags;
8816 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8817 for_each_possible_cpu(i) {
8818 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8820 spin_lock(&rt_rq->rt_runtime_lock);
8821 rt_rq->rt_runtime = global_rt_runtime();
8822 spin_unlock(&rt_rq->rt_runtime_lock);
8824 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8828 #endif /* CONFIG_RT_GROUP_SCHED */
8830 int sched_rt_handler(struct ctl_table *table, int write,
8831 struct file *filp, void __user *buffer, size_t *lenp,
8835 int old_period, old_runtime;
8836 static DEFINE_MUTEX(mutex);
8839 old_period = sysctl_sched_rt_period;
8840 old_runtime = sysctl_sched_rt_runtime;
8842 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8844 if (!ret && write) {
8845 ret = sched_rt_global_constraints();
8847 sysctl_sched_rt_period = old_period;
8848 sysctl_sched_rt_runtime = old_runtime;
8850 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8851 def_rt_bandwidth.rt_period =
8852 ns_to_ktime(global_rt_period());
8855 mutex_unlock(&mutex);
8860 #ifdef CONFIG_CGROUP_SCHED
8862 /* return corresponding task_group object of a cgroup */
8863 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8865 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8866 struct task_group, css);
8869 static struct cgroup_subsys_state *
8870 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8872 struct task_group *tg, *parent;
8874 if (!cgrp->parent) {
8875 /* This is early initialization for the top cgroup */
8876 init_task_group.css.cgroup = cgrp;
8877 return &init_task_group.css;
8880 parent = cgroup_tg(cgrp->parent);
8881 tg = sched_create_group(parent);
8883 return ERR_PTR(-ENOMEM);
8885 /* Bind the cgroup to task_group object we just created */
8886 tg->css.cgroup = cgrp;
8892 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8894 struct task_group *tg = cgroup_tg(cgrp);
8896 sched_destroy_group(tg);
8900 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8901 struct task_struct *tsk)
8903 #ifdef CONFIG_RT_GROUP_SCHED
8904 /* Don't accept realtime tasks when there is no way for them to run */
8905 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8908 /* We don't support RT-tasks being in separate groups */
8909 if (tsk->sched_class != &fair_sched_class)
8917 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8918 struct cgroup *old_cont, struct task_struct *tsk)
8920 sched_move_task(tsk);
8923 #ifdef CONFIG_FAIR_GROUP_SCHED
8924 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8927 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8930 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8932 struct task_group *tg = cgroup_tg(cgrp);
8934 return (u64) tg->shares;
8936 #endif /* CONFIG_FAIR_GROUP_SCHED */
8938 #ifdef CONFIG_RT_GROUP_SCHED
8939 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8942 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8945 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8947 return sched_group_rt_runtime(cgroup_tg(cgrp));
8950 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8953 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8956 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8958 return sched_group_rt_period(cgroup_tg(cgrp));
8960 #endif /* CONFIG_RT_GROUP_SCHED */
8962 static struct cftype cpu_files[] = {
8963 #ifdef CONFIG_FAIR_GROUP_SCHED
8966 .read_u64 = cpu_shares_read_u64,
8967 .write_u64 = cpu_shares_write_u64,
8970 #ifdef CONFIG_RT_GROUP_SCHED
8972 .name = "rt_runtime_us",
8973 .read_s64 = cpu_rt_runtime_read,
8974 .write_s64 = cpu_rt_runtime_write,
8977 .name = "rt_period_us",
8978 .read_u64 = cpu_rt_period_read_uint,
8979 .write_u64 = cpu_rt_period_write_uint,
8984 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8986 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8989 struct cgroup_subsys cpu_cgroup_subsys = {
8991 .create = cpu_cgroup_create,
8992 .destroy = cpu_cgroup_destroy,
8993 .can_attach = cpu_cgroup_can_attach,
8994 .attach = cpu_cgroup_attach,
8995 .populate = cpu_cgroup_populate,
8996 .subsys_id = cpu_cgroup_subsys_id,
9000 #endif /* CONFIG_CGROUP_SCHED */
9002 #ifdef CONFIG_CGROUP_CPUACCT
9005 * CPU accounting code for task groups.
9007 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9008 * (balbir@in.ibm.com).
9011 /* track cpu usage of a group of tasks */
9013 struct cgroup_subsys_state css;
9014 /* cpuusage holds pointer to a u64-type object on every cpu */
9018 struct cgroup_subsys cpuacct_subsys;
9020 /* return cpu accounting group corresponding to this container */
9021 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9023 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9024 struct cpuacct, css);
9027 /* return cpu accounting group to which this task belongs */
9028 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9030 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9031 struct cpuacct, css);
9034 /* create a new cpu accounting group */
9035 static struct cgroup_subsys_state *cpuacct_create(
9036 struct cgroup_subsys *ss, struct cgroup *cgrp)
9038 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9041 return ERR_PTR(-ENOMEM);
9043 ca->cpuusage = alloc_percpu(u64);
9044 if (!ca->cpuusage) {
9046 return ERR_PTR(-ENOMEM);
9052 /* destroy an existing cpu accounting group */
9054 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9056 struct cpuacct *ca = cgroup_ca(cgrp);
9058 free_percpu(ca->cpuusage);
9062 /* return total cpu usage (in nanoseconds) of a group */
9063 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9065 struct cpuacct *ca = cgroup_ca(cgrp);
9066 u64 totalcpuusage = 0;
9069 for_each_possible_cpu(i) {
9070 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9073 * Take rq->lock to make 64-bit addition safe on 32-bit
9076 spin_lock_irq(&cpu_rq(i)->lock);
9077 totalcpuusage += *cpuusage;
9078 spin_unlock_irq(&cpu_rq(i)->lock);
9081 return totalcpuusage;
9084 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9087 struct cpuacct *ca = cgroup_ca(cgrp);
9096 for_each_possible_cpu(i) {
9097 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9099 spin_lock_irq(&cpu_rq(i)->lock);
9101 spin_unlock_irq(&cpu_rq(i)->lock);
9107 static struct cftype files[] = {
9110 .read_u64 = cpuusage_read,
9111 .write_u64 = cpuusage_write,
9115 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9117 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9121 * charge this task's execution time to its accounting group.
9123 * called with rq->lock held.
9125 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9129 if (!cpuacct_subsys.active)
9134 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9136 *cpuusage += cputime;
9140 struct cgroup_subsys cpuacct_subsys = {
9142 .create = cpuacct_create,
9143 .destroy = cpuacct_destroy,
9144 .populate = cpuacct_populate,
9145 .subsys_id = cpuacct_subsys_id,
9147 #endif /* CONFIG_CGROUP_CPUACCT */