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 const_debug unsigned int sysctl_sched_shares_ratelimit = 500000;
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_period < 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);
2816 * If dest_cpu is allowed for this process, migrate the task to it.
2817 * This is accomplished by forcing the cpu_allowed mask to only
2818 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2819 * the cpu_allowed mask is restored.
2821 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2823 struct migration_req req;
2824 unsigned long flags;
2827 rq = task_rq_lock(p, &flags);
2828 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2829 || unlikely(!cpu_active(dest_cpu)))
2832 /* force the process onto the specified CPU */
2833 if (migrate_task(p, dest_cpu, &req)) {
2834 /* Need to wait for migration thread (might exit: take ref). */
2835 struct task_struct *mt = rq->migration_thread;
2837 get_task_struct(mt);
2838 task_rq_unlock(rq, &flags);
2839 wake_up_process(mt);
2840 put_task_struct(mt);
2841 wait_for_completion(&req.done);
2846 task_rq_unlock(rq, &flags);
2850 * sched_exec - execve() is a valuable balancing opportunity, because at
2851 * this point the task has the smallest effective memory and cache footprint.
2853 void sched_exec(void)
2855 int new_cpu, this_cpu = get_cpu();
2856 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2858 if (new_cpu != this_cpu)
2859 sched_migrate_task(current, new_cpu);
2863 * pull_task - move a task from a remote runqueue to the local runqueue.
2864 * Both runqueues must be locked.
2866 static void pull_task(struct rq *src_rq, struct task_struct *p,
2867 struct rq *this_rq, int this_cpu)
2869 deactivate_task(src_rq, p, 0);
2870 set_task_cpu(p, this_cpu);
2871 activate_task(this_rq, p, 0);
2873 * Note that idle threads have a prio of MAX_PRIO, for this test
2874 * to be always true for them.
2876 check_preempt_curr(this_rq, p);
2880 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2883 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2884 struct sched_domain *sd, enum cpu_idle_type idle,
2888 * We do not migrate tasks that are:
2889 * 1) running (obviously), or
2890 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2891 * 3) are cache-hot on their current CPU.
2893 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2894 schedstat_inc(p, se.nr_failed_migrations_affine);
2899 if (task_running(rq, p)) {
2900 schedstat_inc(p, se.nr_failed_migrations_running);
2905 * Aggressive migration if:
2906 * 1) task is cache cold, or
2907 * 2) too many balance attempts have failed.
2910 if (!task_hot(p, rq->clock, sd) ||
2911 sd->nr_balance_failed > sd->cache_nice_tries) {
2912 #ifdef CONFIG_SCHEDSTATS
2913 if (task_hot(p, rq->clock, sd)) {
2914 schedstat_inc(sd, lb_hot_gained[idle]);
2915 schedstat_inc(p, se.nr_forced_migrations);
2921 if (task_hot(p, rq->clock, sd)) {
2922 schedstat_inc(p, se.nr_failed_migrations_hot);
2928 static unsigned long
2929 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2930 unsigned long max_load_move, struct sched_domain *sd,
2931 enum cpu_idle_type idle, int *all_pinned,
2932 int *this_best_prio, struct rq_iterator *iterator)
2934 int loops = 0, pulled = 0, pinned = 0;
2935 struct task_struct *p;
2936 long rem_load_move = max_load_move;
2938 if (max_load_move == 0)
2944 * Start the load-balancing iterator:
2946 p = iterator->start(iterator->arg);
2948 if (!p || loops++ > sysctl_sched_nr_migrate)
2951 if ((p->se.load.weight >> 1) > rem_load_move ||
2952 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2953 p = iterator->next(iterator->arg);
2957 pull_task(busiest, p, this_rq, this_cpu);
2959 rem_load_move -= p->se.load.weight;
2962 * We only want to steal up to the prescribed amount of weighted load.
2964 if (rem_load_move > 0) {
2965 if (p->prio < *this_best_prio)
2966 *this_best_prio = p->prio;
2967 p = iterator->next(iterator->arg);
2972 * Right now, this is one of only two places pull_task() is called,
2973 * so we can safely collect pull_task() stats here rather than
2974 * inside pull_task().
2976 schedstat_add(sd, lb_gained[idle], pulled);
2979 *all_pinned = pinned;
2981 return max_load_move - rem_load_move;
2985 * move_tasks tries to move up to max_load_move weighted load from busiest to
2986 * this_rq, as part of a balancing operation within domain "sd".
2987 * Returns 1 if successful and 0 otherwise.
2989 * Called with both runqueues locked.
2991 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2992 unsigned long max_load_move,
2993 struct sched_domain *sd, enum cpu_idle_type idle,
2996 const struct sched_class *class = sched_class_highest;
2997 unsigned long total_load_moved = 0;
2998 int this_best_prio = this_rq->curr->prio;
3002 class->load_balance(this_rq, this_cpu, busiest,
3003 max_load_move - total_load_moved,
3004 sd, idle, all_pinned, &this_best_prio);
3005 class = class->next;
3007 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3010 } while (class && max_load_move > total_load_moved);
3012 return total_load_moved > 0;
3016 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3017 struct sched_domain *sd, enum cpu_idle_type idle,
3018 struct rq_iterator *iterator)
3020 struct task_struct *p = iterator->start(iterator->arg);
3024 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3025 pull_task(busiest, p, this_rq, this_cpu);
3027 * Right now, this is only the second place pull_task()
3028 * is called, so we can safely collect pull_task()
3029 * stats here rather than inside pull_task().
3031 schedstat_inc(sd, lb_gained[idle]);
3035 p = iterator->next(iterator->arg);
3042 * move_one_task tries to move exactly one task from busiest to this_rq, as
3043 * part of active balancing operations within "domain".
3044 * Returns 1 if successful and 0 otherwise.
3046 * Called with both runqueues locked.
3048 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3049 struct sched_domain *sd, enum cpu_idle_type idle)
3051 const struct sched_class *class;
3053 for (class = sched_class_highest; class; class = class->next)
3054 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3061 * find_busiest_group finds and returns the busiest CPU group within the
3062 * domain. It calculates and returns the amount of weighted load which
3063 * should be moved to restore balance via the imbalance parameter.
3065 static struct sched_group *
3066 find_busiest_group(struct sched_domain *sd, int this_cpu,
3067 unsigned long *imbalance, enum cpu_idle_type idle,
3068 int *sd_idle, const cpumask_t *cpus, int *balance)
3070 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3071 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3072 unsigned long max_pull;
3073 unsigned long busiest_load_per_task, busiest_nr_running;
3074 unsigned long this_load_per_task, this_nr_running;
3075 int load_idx, group_imb = 0;
3076 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3077 int power_savings_balance = 1;
3078 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3079 unsigned long min_nr_running = ULONG_MAX;
3080 struct sched_group *group_min = NULL, *group_leader = NULL;
3083 max_load = this_load = total_load = total_pwr = 0;
3084 busiest_load_per_task = busiest_nr_running = 0;
3085 this_load_per_task = this_nr_running = 0;
3087 if (idle == CPU_NOT_IDLE)
3088 load_idx = sd->busy_idx;
3089 else if (idle == CPU_NEWLY_IDLE)
3090 load_idx = sd->newidle_idx;
3092 load_idx = sd->idle_idx;
3095 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3098 int __group_imb = 0;
3099 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3100 unsigned long sum_nr_running, sum_weighted_load;
3101 unsigned long sum_avg_load_per_task;
3102 unsigned long avg_load_per_task;
3104 local_group = cpu_isset(this_cpu, group->cpumask);
3107 balance_cpu = first_cpu(group->cpumask);
3109 /* Tally up the load of all CPUs in the group */
3110 sum_weighted_load = sum_nr_running = avg_load = 0;
3111 sum_avg_load_per_task = avg_load_per_task = 0;
3114 min_cpu_load = ~0UL;
3116 for_each_cpu_mask_nr(i, group->cpumask) {
3119 if (!cpu_isset(i, *cpus))
3124 if (*sd_idle && rq->nr_running)
3127 /* Bias balancing toward cpus of our domain */
3129 if (idle_cpu(i) && !first_idle_cpu) {
3134 load = target_load(i, load_idx);
3136 load = source_load(i, load_idx);
3137 if (load > max_cpu_load)
3138 max_cpu_load = load;
3139 if (min_cpu_load > load)
3140 min_cpu_load = load;
3144 sum_nr_running += rq->nr_running;
3145 sum_weighted_load += weighted_cpuload(i);
3147 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3151 * First idle cpu or the first cpu(busiest) in this sched group
3152 * is eligible for doing load balancing at this and above
3153 * domains. In the newly idle case, we will allow all the cpu's
3154 * to do the newly idle load balance.
3156 if (idle != CPU_NEWLY_IDLE && local_group &&
3157 balance_cpu != this_cpu && balance) {
3162 total_load += avg_load;
3163 total_pwr += group->__cpu_power;
3165 /* Adjust by relative CPU power of the group */
3166 avg_load = sg_div_cpu_power(group,
3167 avg_load * SCHED_LOAD_SCALE);
3171 * Consider the group unbalanced when the imbalance is larger
3172 * than the average weight of two tasks.
3174 * APZ: with cgroup the avg task weight can vary wildly and
3175 * might not be a suitable number - should we keep a
3176 * normalized nr_running number somewhere that negates
3179 avg_load_per_task = sg_div_cpu_power(group,
3180 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3182 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3185 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3188 this_load = avg_load;
3190 this_nr_running = sum_nr_running;
3191 this_load_per_task = sum_weighted_load;
3192 } else if (avg_load > max_load &&
3193 (sum_nr_running > group_capacity || __group_imb)) {
3194 max_load = avg_load;
3196 busiest_nr_running = sum_nr_running;
3197 busiest_load_per_task = sum_weighted_load;
3198 group_imb = __group_imb;
3201 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3203 * Busy processors will not participate in power savings
3206 if (idle == CPU_NOT_IDLE ||
3207 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3211 * If the local group is idle or completely loaded
3212 * no need to do power savings balance at this domain
3214 if (local_group && (this_nr_running >= group_capacity ||
3216 power_savings_balance = 0;
3219 * If a group is already running at full capacity or idle,
3220 * don't include that group in power savings calculations
3222 if (!power_savings_balance || sum_nr_running >= group_capacity
3227 * Calculate the group which has the least non-idle load.
3228 * This is the group from where we need to pick up the load
3231 if ((sum_nr_running < min_nr_running) ||
3232 (sum_nr_running == min_nr_running &&
3233 first_cpu(group->cpumask) <
3234 first_cpu(group_min->cpumask))) {
3236 min_nr_running = sum_nr_running;
3237 min_load_per_task = sum_weighted_load /
3242 * Calculate the group which is almost near its
3243 * capacity but still has some space to pick up some load
3244 * from other group and save more power
3246 if (sum_nr_running <= group_capacity - 1) {
3247 if (sum_nr_running > leader_nr_running ||
3248 (sum_nr_running == leader_nr_running &&
3249 first_cpu(group->cpumask) >
3250 first_cpu(group_leader->cpumask))) {
3251 group_leader = group;
3252 leader_nr_running = sum_nr_running;
3257 group = group->next;
3258 } while (group != sd->groups);
3260 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3263 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3265 if (this_load >= avg_load ||
3266 100*max_load <= sd->imbalance_pct*this_load)
3269 busiest_load_per_task /= busiest_nr_running;
3271 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3274 * We're trying to get all the cpus to the average_load, so we don't
3275 * want to push ourselves above the average load, nor do we wish to
3276 * reduce the max loaded cpu below the average load, as either of these
3277 * actions would just result in more rebalancing later, and ping-pong
3278 * tasks around. Thus we look for the minimum possible imbalance.
3279 * Negative imbalances (*we* are more loaded than anyone else) will
3280 * be counted as no imbalance for these purposes -- we can't fix that
3281 * by pulling tasks to us. Be careful of negative numbers as they'll
3282 * appear as very large values with unsigned longs.
3284 if (max_load <= busiest_load_per_task)
3288 * In the presence of smp nice balancing, certain scenarios can have
3289 * max load less than avg load(as we skip the groups at or below
3290 * its cpu_power, while calculating max_load..)
3292 if (max_load < avg_load) {
3294 goto small_imbalance;
3297 /* Don't want to pull so many tasks that a group would go idle */
3298 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3300 /* How much load to actually move to equalise the imbalance */
3301 *imbalance = min(max_pull * busiest->__cpu_power,
3302 (avg_load - this_load) * this->__cpu_power)
3306 * if *imbalance is less than the average load per runnable task
3307 * there is no gaurantee that any tasks will be moved so we'll have
3308 * a think about bumping its value to force at least one task to be
3311 if (*imbalance < busiest_load_per_task) {
3312 unsigned long tmp, pwr_now, pwr_move;
3316 pwr_move = pwr_now = 0;
3318 if (this_nr_running) {
3319 this_load_per_task /= this_nr_running;
3320 if (busiest_load_per_task > this_load_per_task)
3323 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3325 if (max_load - this_load + 2*busiest_load_per_task >=
3326 busiest_load_per_task * imbn) {
3327 *imbalance = busiest_load_per_task;
3332 * OK, we don't have enough imbalance to justify moving tasks,
3333 * however we may be able to increase total CPU power used by
3337 pwr_now += busiest->__cpu_power *
3338 min(busiest_load_per_task, max_load);
3339 pwr_now += this->__cpu_power *
3340 min(this_load_per_task, this_load);
3341 pwr_now /= SCHED_LOAD_SCALE;
3343 /* Amount of load we'd subtract */
3344 tmp = sg_div_cpu_power(busiest,
3345 busiest_load_per_task * SCHED_LOAD_SCALE);
3347 pwr_move += busiest->__cpu_power *
3348 min(busiest_load_per_task, max_load - tmp);
3350 /* Amount of load we'd add */
3351 if (max_load * busiest->__cpu_power <
3352 busiest_load_per_task * SCHED_LOAD_SCALE)
3353 tmp = sg_div_cpu_power(this,
3354 max_load * busiest->__cpu_power);
3356 tmp = sg_div_cpu_power(this,
3357 busiest_load_per_task * SCHED_LOAD_SCALE);
3358 pwr_move += this->__cpu_power *
3359 min(this_load_per_task, this_load + tmp);
3360 pwr_move /= SCHED_LOAD_SCALE;
3362 /* Move if we gain throughput */
3363 if (pwr_move > pwr_now)
3364 *imbalance = busiest_load_per_task;
3370 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3371 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3374 if (this == group_leader && group_leader != group_min) {
3375 *imbalance = min_load_per_task;
3385 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3388 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3389 unsigned long imbalance, const cpumask_t *cpus)
3391 struct rq *busiest = NULL, *rq;
3392 unsigned long max_load = 0;
3395 for_each_cpu_mask_nr(i, group->cpumask) {
3398 if (!cpu_isset(i, *cpus))
3402 wl = weighted_cpuload(i);
3404 if (rq->nr_running == 1 && wl > imbalance)
3407 if (wl > max_load) {
3417 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3418 * so long as it is large enough.
3420 #define MAX_PINNED_INTERVAL 512
3423 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3424 * tasks if there is an imbalance.
3426 static int load_balance(int this_cpu, struct rq *this_rq,
3427 struct sched_domain *sd, enum cpu_idle_type idle,
3428 int *balance, cpumask_t *cpus)
3430 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3431 struct sched_group *group;
3432 unsigned long imbalance;
3434 unsigned long flags;
3439 * When power savings policy is enabled for the parent domain, idle
3440 * sibling can pick up load irrespective of busy siblings. In this case,
3441 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3442 * portraying it as CPU_NOT_IDLE.
3444 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3445 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3448 schedstat_inc(sd, lb_count[idle]);
3452 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3459 schedstat_inc(sd, lb_nobusyg[idle]);
3463 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3465 schedstat_inc(sd, lb_nobusyq[idle]);
3469 BUG_ON(busiest == this_rq);
3471 schedstat_add(sd, lb_imbalance[idle], imbalance);
3474 if (busiest->nr_running > 1) {
3476 * Attempt to move tasks. If find_busiest_group has found
3477 * an imbalance but busiest->nr_running <= 1, the group is
3478 * still unbalanced. ld_moved simply stays zero, so it is
3479 * correctly treated as an imbalance.
3481 local_irq_save(flags);
3482 double_rq_lock(this_rq, busiest);
3483 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3484 imbalance, sd, idle, &all_pinned);
3485 double_rq_unlock(this_rq, busiest);
3486 local_irq_restore(flags);
3489 * some other cpu did the load balance for us.
3491 if (ld_moved && this_cpu != smp_processor_id())
3492 resched_cpu(this_cpu);
3494 /* All tasks on this runqueue were pinned by CPU affinity */
3495 if (unlikely(all_pinned)) {
3496 cpu_clear(cpu_of(busiest), *cpus);
3497 if (!cpus_empty(*cpus))
3504 schedstat_inc(sd, lb_failed[idle]);
3505 sd->nr_balance_failed++;
3507 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3509 spin_lock_irqsave(&busiest->lock, flags);
3511 /* don't kick the migration_thread, if the curr
3512 * task on busiest cpu can't be moved to this_cpu
3514 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3515 spin_unlock_irqrestore(&busiest->lock, flags);
3517 goto out_one_pinned;
3520 if (!busiest->active_balance) {
3521 busiest->active_balance = 1;
3522 busiest->push_cpu = this_cpu;
3525 spin_unlock_irqrestore(&busiest->lock, flags);
3527 wake_up_process(busiest->migration_thread);
3530 * We've kicked active balancing, reset the failure
3533 sd->nr_balance_failed = sd->cache_nice_tries+1;
3536 sd->nr_balance_failed = 0;
3538 if (likely(!active_balance)) {
3539 /* We were unbalanced, so reset the balancing interval */
3540 sd->balance_interval = sd->min_interval;
3543 * If we've begun active balancing, start to back off. This
3544 * case may not be covered by the all_pinned logic if there
3545 * is only 1 task on the busy runqueue (because we don't call
3548 if (sd->balance_interval < sd->max_interval)
3549 sd->balance_interval *= 2;
3552 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3553 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3559 schedstat_inc(sd, lb_balanced[idle]);
3561 sd->nr_balance_failed = 0;
3564 /* tune up the balancing interval */
3565 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3566 (sd->balance_interval < sd->max_interval))
3567 sd->balance_interval *= 2;
3569 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3570 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3581 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3582 * tasks if there is an imbalance.
3584 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3585 * this_rq is locked.
3588 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3591 struct sched_group *group;
3592 struct rq *busiest = NULL;
3593 unsigned long imbalance;
3601 * When power savings policy is enabled for the parent domain, idle
3602 * sibling can pick up load irrespective of busy siblings. In this case,
3603 * let the state of idle sibling percolate up as IDLE, instead of
3604 * portraying it as CPU_NOT_IDLE.
3606 if (sd->flags & SD_SHARE_CPUPOWER &&
3607 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3610 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3612 update_shares_locked(this_rq, sd);
3613 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3614 &sd_idle, cpus, NULL);
3616 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3620 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3622 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3626 BUG_ON(busiest == this_rq);
3628 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3631 if (busiest->nr_running > 1) {
3632 /* Attempt to move tasks */
3633 double_lock_balance(this_rq, busiest);
3634 /* this_rq->clock is already updated */
3635 update_rq_clock(busiest);
3636 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3637 imbalance, sd, CPU_NEWLY_IDLE,
3639 spin_unlock(&busiest->lock);
3641 if (unlikely(all_pinned)) {
3642 cpu_clear(cpu_of(busiest), *cpus);
3643 if (!cpus_empty(*cpus))
3649 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3650 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3651 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3654 sd->nr_balance_failed = 0;
3656 update_shares_locked(this_rq, sd);
3660 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3661 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3662 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3664 sd->nr_balance_failed = 0;
3670 * idle_balance is called by schedule() if this_cpu is about to become
3671 * idle. Attempts to pull tasks from other CPUs.
3673 static void idle_balance(int this_cpu, struct rq *this_rq)
3675 struct sched_domain *sd;
3676 int pulled_task = -1;
3677 unsigned long next_balance = jiffies + HZ;
3680 for_each_domain(this_cpu, sd) {
3681 unsigned long interval;
3683 if (!(sd->flags & SD_LOAD_BALANCE))
3686 if (sd->flags & SD_BALANCE_NEWIDLE)
3687 /* If we've pulled tasks over stop searching: */
3688 pulled_task = load_balance_newidle(this_cpu, this_rq,
3691 interval = msecs_to_jiffies(sd->balance_interval);
3692 if (time_after(next_balance, sd->last_balance + interval))
3693 next_balance = sd->last_balance + interval;
3697 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3699 * We are going idle. next_balance may be set based on
3700 * a busy processor. So reset next_balance.
3702 this_rq->next_balance = next_balance;
3707 * active_load_balance is run by migration threads. It pushes running tasks
3708 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3709 * running on each physical CPU where possible, and avoids physical /
3710 * logical imbalances.
3712 * Called with busiest_rq locked.
3714 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3716 int target_cpu = busiest_rq->push_cpu;
3717 struct sched_domain *sd;
3718 struct rq *target_rq;
3720 /* Is there any task to move? */
3721 if (busiest_rq->nr_running <= 1)
3724 target_rq = cpu_rq(target_cpu);
3727 * This condition is "impossible", if it occurs
3728 * we need to fix it. Originally reported by
3729 * Bjorn Helgaas on a 128-cpu setup.
3731 BUG_ON(busiest_rq == target_rq);
3733 /* move a task from busiest_rq to target_rq */
3734 double_lock_balance(busiest_rq, target_rq);
3735 update_rq_clock(busiest_rq);
3736 update_rq_clock(target_rq);
3738 /* Search for an sd spanning us and the target CPU. */
3739 for_each_domain(target_cpu, sd) {
3740 if ((sd->flags & SD_LOAD_BALANCE) &&
3741 cpu_isset(busiest_cpu, sd->span))
3746 schedstat_inc(sd, alb_count);
3748 if (move_one_task(target_rq, target_cpu, busiest_rq,
3750 schedstat_inc(sd, alb_pushed);
3752 schedstat_inc(sd, alb_failed);
3754 spin_unlock(&target_rq->lock);
3759 atomic_t load_balancer;
3761 } nohz ____cacheline_aligned = {
3762 .load_balancer = ATOMIC_INIT(-1),
3763 .cpu_mask = CPU_MASK_NONE,
3767 * This routine will try to nominate the ilb (idle load balancing)
3768 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3769 * load balancing on behalf of all those cpus. If all the cpus in the system
3770 * go into this tickless mode, then there will be no ilb owner (as there is
3771 * no need for one) and all the cpus will sleep till the next wakeup event
3774 * For the ilb owner, tick is not stopped. And this tick will be used
3775 * for idle load balancing. ilb owner will still be part of
3778 * While stopping the tick, this cpu will become the ilb owner if there
3779 * is no other owner. And will be the owner till that cpu becomes busy
3780 * or if all cpus in the system stop their ticks at which point
3781 * there is no need for ilb owner.
3783 * When the ilb owner becomes busy, it nominates another owner, during the
3784 * next busy scheduler_tick()
3786 int select_nohz_load_balancer(int stop_tick)
3788 int cpu = smp_processor_id();
3791 cpu_set(cpu, nohz.cpu_mask);
3792 cpu_rq(cpu)->in_nohz_recently = 1;
3795 * If we are going offline and still the leader, give up!
3797 if (!cpu_active(cpu) &&
3798 atomic_read(&nohz.load_balancer) == cpu) {
3799 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3804 /* time for ilb owner also to sleep */
3805 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3806 if (atomic_read(&nohz.load_balancer) == cpu)
3807 atomic_set(&nohz.load_balancer, -1);
3811 if (atomic_read(&nohz.load_balancer) == -1) {
3812 /* make me the ilb owner */
3813 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3815 } else if (atomic_read(&nohz.load_balancer) == cpu)
3818 if (!cpu_isset(cpu, nohz.cpu_mask))
3821 cpu_clear(cpu, nohz.cpu_mask);
3823 if (atomic_read(&nohz.load_balancer) == cpu)
3824 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3831 static DEFINE_SPINLOCK(balancing);
3834 * It checks each scheduling domain to see if it is due to be balanced,
3835 * and initiates a balancing operation if so.
3837 * Balancing parameters are set up in arch_init_sched_domains.
3839 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3842 struct rq *rq = cpu_rq(cpu);
3843 unsigned long interval;
3844 struct sched_domain *sd;
3845 /* Earliest time when we have to do rebalance again */
3846 unsigned long next_balance = jiffies + 60*HZ;
3847 int update_next_balance = 0;
3851 for_each_domain(cpu, sd) {
3852 if (!(sd->flags & SD_LOAD_BALANCE))
3855 interval = sd->balance_interval;
3856 if (idle != CPU_IDLE)
3857 interval *= sd->busy_factor;
3859 /* scale ms to jiffies */
3860 interval = msecs_to_jiffies(interval);
3861 if (unlikely(!interval))
3863 if (interval > HZ*NR_CPUS/10)
3864 interval = HZ*NR_CPUS/10;
3866 need_serialize = sd->flags & SD_SERIALIZE;
3868 if (need_serialize) {
3869 if (!spin_trylock(&balancing))
3873 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3874 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3876 * We've pulled tasks over so either we're no
3877 * longer idle, or one of our SMT siblings is
3880 idle = CPU_NOT_IDLE;
3882 sd->last_balance = jiffies;
3885 spin_unlock(&balancing);
3887 if (time_after(next_balance, sd->last_balance + interval)) {
3888 next_balance = sd->last_balance + interval;
3889 update_next_balance = 1;
3893 * Stop the load balance at this level. There is another
3894 * CPU in our sched group which is doing load balancing more
3902 * next_balance will be updated only when there is a need.
3903 * When the cpu is attached to null domain for ex, it will not be
3906 if (likely(update_next_balance))
3907 rq->next_balance = next_balance;
3911 * run_rebalance_domains is triggered when needed from the scheduler tick.
3912 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3913 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3915 static void run_rebalance_domains(struct softirq_action *h)
3917 int this_cpu = smp_processor_id();
3918 struct rq *this_rq = cpu_rq(this_cpu);
3919 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3920 CPU_IDLE : CPU_NOT_IDLE;
3922 rebalance_domains(this_cpu, idle);
3926 * If this cpu is the owner for idle load balancing, then do the
3927 * balancing on behalf of the other idle cpus whose ticks are
3930 if (this_rq->idle_at_tick &&
3931 atomic_read(&nohz.load_balancer) == this_cpu) {
3932 cpumask_t cpus = nohz.cpu_mask;
3936 cpu_clear(this_cpu, cpus);
3937 for_each_cpu_mask_nr(balance_cpu, cpus) {
3939 * If this cpu gets work to do, stop the load balancing
3940 * work being done for other cpus. Next load
3941 * balancing owner will pick it up.
3946 rebalance_domains(balance_cpu, CPU_IDLE);
3948 rq = cpu_rq(balance_cpu);
3949 if (time_after(this_rq->next_balance, rq->next_balance))
3950 this_rq->next_balance = rq->next_balance;
3957 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3959 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3960 * idle load balancing owner or decide to stop the periodic load balancing,
3961 * if the whole system is idle.
3963 static inline void trigger_load_balance(struct rq *rq, int cpu)
3967 * If we were in the nohz mode recently and busy at the current
3968 * scheduler tick, then check if we need to nominate new idle
3971 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3972 rq->in_nohz_recently = 0;
3974 if (atomic_read(&nohz.load_balancer) == cpu) {
3975 cpu_clear(cpu, nohz.cpu_mask);
3976 atomic_set(&nohz.load_balancer, -1);
3979 if (atomic_read(&nohz.load_balancer) == -1) {
3981 * simple selection for now: Nominate the
3982 * first cpu in the nohz list to be the next
3985 * TBD: Traverse the sched domains and nominate
3986 * the nearest cpu in the nohz.cpu_mask.
3988 int ilb = first_cpu(nohz.cpu_mask);
3990 if (ilb < nr_cpu_ids)
3996 * If this cpu is idle and doing idle load balancing for all the
3997 * cpus with ticks stopped, is it time for that to stop?
3999 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4000 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4006 * If this cpu is idle and the idle load balancing is done by
4007 * someone else, then no need raise the SCHED_SOFTIRQ
4009 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4010 cpu_isset(cpu, nohz.cpu_mask))
4013 if (time_after_eq(jiffies, rq->next_balance))
4014 raise_softirq(SCHED_SOFTIRQ);
4017 #else /* CONFIG_SMP */
4020 * on UP we do not need to balance between CPUs:
4022 static inline void idle_balance(int cpu, struct rq *rq)
4028 DEFINE_PER_CPU(struct kernel_stat, kstat);
4030 EXPORT_PER_CPU_SYMBOL(kstat);
4033 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4034 * that have not yet been banked in case the task is currently running.
4036 unsigned long long task_sched_runtime(struct task_struct *p)
4038 unsigned long flags;
4042 rq = task_rq_lock(p, &flags);
4043 ns = p->se.sum_exec_runtime;
4044 if (task_current(rq, p)) {
4045 update_rq_clock(rq);
4046 delta_exec = rq->clock - p->se.exec_start;
4047 if ((s64)delta_exec > 0)
4050 task_rq_unlock(rq, &flags);
4056 * Account user cpu time to a process.
4057 * @p: the process that the cpu time gets accounted to
4058 * @cputime: the cpu time spent in user space since the last update
4060 void account_user_time(struct task_struct *p, cputime_t cputime)
4062 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4065 p->utime = cputime_add(p->utime, cputime);
4067 /* Add user time to cpustat. */
4068 tmp = cputime_to_cputime64(cputime);
4069 if (TASK_NICE(p) > 0)
4070 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4072 cpustat->user = cputime64_add(cpustat->user, tmp);
4073 /* Account for user time used */
4074 acct_update_integrals(p);
4078 * Account guest cpu time to a process.
4079 * @p: the process that the cpu time gets accounted to
4080 * @cputime: the cpu time spent in virtual machine since the last update
4082 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4085 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4087 tmp = cputime_to_cputime64(cputime);
4089 p->utime = cputime_add(p->utime, cputime);
4090 p->gtime = cputime_add(p->gtime, cputime);
4092 cpustat->user = cputime64_add(cpustat->user, tmp);
4093 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4097 * Account scaled user cpu time to a process.
4098 * @p: the process that the cpu time gets accounted to
4099 * @cputime: the cpu time spent in user space since the last update
4101 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4103 p->utimescaled = cputime_add(p->utimescaled, cputime);
4107 * Account system cpu time to a process.
4108 * @p: the process that the cpu time gets accounted to
4109 * @hardirq_offset: the offset to subtract from hardirq_count()
4110 * @cputime: the cpu time spent in kernel space since the last update
4112 void account_system_time(struct task_struct *p, int hardirq_offset,
4115 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4116 struct rq *rq = this_rq();
4119 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4120 account_guest_time(p, cputime);
4124 p->stime = cputime_add(p->stime, cputime);
4126 /* Add system time to cpustat. */
4127 tmp = cputime_to_cputime64(cputime);
4128 if (hardirq_count() - hardirq_offset)
4129 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4130 else if (softirq_count())
4131 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4132 else if (p != rq->idle)
4133 cpustat->system = cputime64_add(cpustat->system, tmp);
4134 else if (atomic_read(&rq->nr_iowait) > 0)
4135 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4137 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4138 /* Account for system time used */
4139 acct_update_integrals(p);
4143 * Account scaled system cpu time to a process.
4144 * @p: the process that the cpu time gets accounted to
4145 * @hardirq_offset: the offset to subtract from hardirq_count()
4146 * @cputime: the cpu time spent in kernel space since the last update
4148 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4150 p->stimescaled = cputime_add(p->stimescaled, cputime);
4154 * Account for involuntary wait time.
4155 * @p: the process from which the cpu time has been stolen
4156 * @steal: the cpu time spent in involuntary wait
4158 void account_steal_time(struct task_struct *p, cputime_t steal)
4160 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4161 cputime64_t tmp = cputime_to_cputime64(steal);
4162 struct rq *rq = this_rq();
4164 if (p == rq->idle) {
4165 p->stime = cputime_add(p->stime, steal);
4166 if (atomic_read(&rq->nr_iowait) > 0)
4167 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4169 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4171 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4175 * This function gets called by the timer code, with HZ frequency.
4176 * We call it with interrupts disabled.
4178 * It also gets called by the fork code, when changing the parent's
4181 void scheduler_tick(void)
4183 int cpu = smp_processor_id();
4184 struct rq *rq = cpu_rq(cpu);
4185 struct task_struct *curr = rq->curr;
4189 spin_lock(&rq->lock);
4190 update_rq_clock(rq);
4191 update_cpu_load(rq);
4192 curr->sched_class->task_tick(rq, curr, 0);
4193 spin_unlock(&rq->lock);
4196 rq->idle_at_tick = idle_cpu(cpu);
4197 trigger_load_balance(rq, cpu);
4201 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4202 defined(CONFIG_PREEMPT_TRACER))
4204 static inline unsigned long get_parent_ip(unsigned long addr)
4206 if (in_lock_functions(addr)) {
4207 addr = CALLER_ADDR2;
4208 if (in_lock_functions(addr))
4209 addr = CALLER_ADDR3;
4214 void __kprobes add_preempt_count(int val)
4216 #ifdef CONFIG_DEBUG_PREEMPT
4220 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4223 preempt_count() += val;
4224 #ifdef CONFIG_DEBUG_PREEMPT
4226 * Spinlock count overflowing soon?
4228 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4231 if (preempt_count() == val)
4232 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4234 EXPORT_SYMBOL(add_preempt_count);
4236 void __kprobes sub_preempt_count(int val)
4238 #ifdef CONFIG_DEBUG_PREEMPT
4242 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4245 * Is the spinlock portion underflowing?
4247 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4248 !(preempt_count() & PREEMPT_MASK)))
4252 if (preempt_count() == val)
4253 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4254 preempt_count() -= val;
4256 EXPORT_SYMBOL(sub_preempt_count);
4261 * Print scheduling while atomic bug:
4263 static noinline void __schedule_bug(struct task_struct *prev)
4265 struct pt_regs *regs = get_irq_regs();
4267 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4268 prev->comm, prev->pid, preempt_count());
4270 debug_show_held_locks(prev);
4272 if (irqs_disabled())
4273 print_irqtrace_events(prev);
4282 * Various schedule()-time debugging checks and statistics:
4284 static inline void schedule_debug(struct task_struct *prev)
4287 * Test if we are atomic. Since do_exit() needs to call into
4288 * schedule() atomically, we ignore that path for now.
4289 * Otherwise, whine if we are scheduling when we should not be.
4291 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4292 __schedule_bug(prev);
4294 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4296 schedstat_inc(this_rq(), sched_count);
4297 #ifdef CONFIG_SCHEDSTATS
4298 if (unlikely(prev->lock_depth >= 0)) {
4299 schedstat_inc(this_rq(), bkl_count);
4300 schedstat_inc(prev, sched_info.bkl_count);
4306 * Pick up the highest-prio task:
4308 static inline struct task_struct *
4309 pick_next_task(struct rq *rq, struct task_struct *prev)
4311 const struct sched_class *class;
4312 struct task_struct *p;
4315 * Optimization: we know that if all tasks are in
4316 * the fair class we can call that function directly:
4318 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4319 p = fair_sched_class.pick_next_task(rq);
4324 class = sched_class_highest;
4326 p = class->pick_next_task(rq);
4330 * Will never be NULL as the idle class always
4331 * returns a non-NULL p:
4333 class = class->next;
4338 * schedule() is the main scheduler function.
4340 asmlinkage void __sched schedule(void)
4342 struct task_struct *prev, *next;
4343 unsigned long *switch_count;
4349 cpu = smp_processor_id();
4353 switch_count = &prev->nivcsw;
4355 release_kernel_lock(prev);
4356 need_resched_nonpreemptible:
4358 schedule_debug(prev);
4360 if (sched_feat(HRTICK))
4364 * Do the rq-clock update outside the rq lock:
4366 local_irq_disable();
4367 update_rq_clock(rq);
4368 spin_lock(&rq->lock);
4369 clear_tsk_need_resched(prev);
4371 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4372 if (unlikely(signal_pending_state(prev->state, prev)))
4373 prev->state = TASK_RUNNING;
4375 deactivate_task(rq, prev, 1);
4376 switch_count = &prev->nvcsw;
4380 if (prev->sched_class->pre_schedule)
4381 prev->sched_class->pre_schedule(rq, prev);
4384 if (unlikely(!rq->nr_running))
4385 idle_balance(cpu, rq);
4387 prev->sched_class->put_prev_task(rq, prev);
4388 next = pick_next_task(rq, prev);
4390 if (likely(prev != next)) {
4391 sched_info_switch(prev, next);
4397 context_switch(rq, prev, next); /* unlocks the rq */
4399 * the context switch might have flipped the stack from under
4400 * us, hence refresh the local variables.
4402 cpu = smp_processor_id();
4405 spin_unlock_irq(&rq->lock);
4407 if (unlikely(reacquire_kernel_lock(current) < 0))
4408 goto need_resched_nonpreemptible;
4410 preempt_enable_no_resched();
4411 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4414 EXPORT_SYMBOL(schedule);
4416 #ifdef CONFIG_PREEMPT
4418 * this is the entry point to schedule() from in-kernel preemption
4419 * off of preempt_enable. Kernel preemptions off return from interrupt
4420 * occur there and call schedule directly.
4422 asmlinkage void __sched preempt_schedule(void)
4424 struct thread_info *ti = current_thread_info();
4427 * If there is a non-zero preempt_count or interrupts are disabled,
4428 * we do not want to preempt the current task. Just return..
4430 if (likely(ti->preempt_count || irqs_disabled()))
4434 add_preempt_count(PREEMPT_ACTIVE);
4436 sub_preempt_count(PREEMPT_ACTIVE);
4439 * Check again in case we missed a preemption opportunity
4440 * between schedule and now.
4443 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4445 EXPORT_SYMBOL(preempt_schedule);
4448 * this is the entry point to schedule() from kernel preemption
4449 * off of irq context.
4450 * Note, that this is called and return with irqs disabled. This will
4451 * protect us against recursive calling from irq.
4453 asmlinkage void __sched preempt_schedule_irq(void)
4455 struct thread_info *ti = current_thread_info();
4457 /* Catch callers which need to be fixed */
4458 BUG_ON(ti->preempt_count || !irqs_disabled());
4461 add_preempt_count(PREEMPT_ACTIVE);
4464 local_irq_disable();
4465 sub_preempt_count(PREEMPT_ACTIVE);
4468 * Check again in case we missed a preemption opportunity
4469 * between schedule and now.
4472 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4475 #endif /* CONFIG_PREEMPT */
4477 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4480 return try_to_wake_up(curr->private, mode, sync);
4482 EXPORT_SYMBOL(default_wake_function);
4485 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4486 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4487 * number) then we wake all the non-exclusive tasks and one exclusive task.
4489 * There are circumstances in which we can try to wake a task which has already
4490 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4491 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4493 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4494 int nr_exclusive, int sync, void *key)
4496 wait_queue_t *curr, *next;
4498 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4499 unsigned flags = curr->flags;
4501 if (curr->func(curr, mode, sync, key) &&
4502 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4508 * __wake_up - wake up threads blocked on a waitqueue.
4510 * @mode: which threads
4511 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4512 * @key: is directly passed to the wakeup function
4514 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4515 int nr_exclusive, void *key)
4517 unsigned long flags;
4519 spin_lock_irqsave(&q->lock, flags);
4520 __wake_up_common(q, mode, nr_exclusive, 0, key);
4521 spin_unlock_irqrestore(&q->lock, flags);
4523 EXPORT_SYMBOL(__wake_up);
4526 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4528 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4530 __wake_up_common(q, mode, 1, 0, NULL);
4534 * __wake_up_sync - wake up threads blocked on a waitqueue.
4536 * @mode: which threads
4537 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4539 * The sync wakeup differs that the waker knows that it will schedule
4540 * away soon, so while the target thread will be woken up, it will not
4541 * be migrated to another CPU - ie. the two threads are 'synchronized'
4542 * with each other. This can prevent needless bouncing between CPUs.
4544 * On UP it can prevent extra preemption.
4547 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4549 unsigned long flags;
4555 if (unlikely(!nr_exclusive))
4558 spin_lock_irqsave(&q->lock, flags);
4559 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4560 spin_unlock_irqrestore(&q->lock, flags);
4562 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4564 void complete(struct completion *x)
4566 unsigned long flags;
4568 spin_lock_irqsave(&x->wait.lock, flags);
4570 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4571 spin_unlock_irqrestore(&x->wait.lock, flags);
4573 EXPORT_SYMBOL(complete);
4575 void complete_all(struct completion *x)
4577 unsigned long flags;
4579 spin_lock_irqsave(&x->wait.lock, flags);
4580 x->done += UINT_MAX/2;
4581 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4582 spin_unlock_irqrestore(&x->wait.lock, flags);
4584 EXPORT_SYMBOL(complete_all);
4586 static inline long __sched
4587 do_wait_for_common(struct completion *x, long timeout, int state)
4590 DECLARE_WAITQUEUE(wait, current);
4592 wait.flags |= WQ_FLAG_EXCLUSIVE;
4593 __add_wait_queue_tail(&x->wait, &wait);
4595 if ((state == TASK_INTERRUPTIBLE &&
4596 signal_pending(current)) ||
4597 (state == TASK_KILLABLE &&
4598 fatal_signal_pending(current))) {
4599 timeout = -ERESTARTSYS;
4602 __set_current_state(state);
4603 spin_unlock_irq(&x->wait.lock);
4604 timeout = schedule_timeout(timeout);
4605 spin_lock_irq(&x->wait.lock);
4606 } while (!x->done && timeout);
4607 __remove_wait_queue(&x->wait, &wait);
4612 return timeout ?: 1;
4616 wait_for_common(struct completion *x, long timeout, int state)
4620 spin_lock_irq(&x->wait.lock);
4621 timeout = do_wait_for_common(x, timeout, state);
4622 spin_unlock_irq(&x->wait.lock);
4626 void __sched wait_for_completion(struct completion *x)
4628 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4630 EXPORT_SYMBOL(wait_for_completion);
4632 unsigned long __sched
4633 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4635 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4637 EXPORT_SYMBOL(wait_for_completion_timeout);
4639 int __sched wait_for_completion_interruptible(struct completion *x)
4641 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4642 if (t == -ERESTARTSYS)
4646 EXPORT_SYMBOL(wait_for_completion_interruptible);
4648 unsigned long __sched
4649 wait_for_completion_interruptible_timeout(struct completion *x,
4650 unsigned long timeout)
4652 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4654 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4656 int __sched wait_for_completion_killable(struct completion *x)
4658 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4659 if (t == -ERESTARTSYS)
4663 EXPORT_SYMBOL(wait_for_completion_killable);
4666 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4668 unsigned long flags;
4671 init_waitqueue_entry(&wait, current);
4673 __set_current_state(state);
4675 spin_lock_irqsave(&q->lock, flags);
4676 __add_wait_queue(q, &wait);
4677 spin_unlock(&q->lock);
4678 timeout = schedule_timeout(timeout);
4679 spin_lock_irq(&q->lock);
4680 __remove_wait_queue(q, &wait);
4681 spin_unlock_irqrestore(&q->lock, flags);
4686 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4688 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4690 EXPORT_SYMBOL(interruptible_sleep_on);
4693 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4695 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4697 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4699 void __sched sleep_on(wait_queue_head_t *q)
4701 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4703 EXPORT_SYMBOL(sleep_on);
4705 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4707 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4709 EXPORT_SYMBOL(sleep_on_timeout);
4711 #ifdef CONFIG_RT_MUTEXES
4714 * rt_mutex_setprio - set the current priority of a task
4716 * @prio: prio value (kernel-internal form)
4718 * This function changes the 'effective' priority of a task. It does
4719 * not touch ->normal_prio like __setscheduler().
4721 * Used by the rt_mutex code to implement priority inheritance logic.
4723 void rt_mutex_setprio(struct task_struct *p, int prio)
4725 unsigned long flags;
4726 int oldprio, on_rq, running;
4728 const struct sched_class *prev_class = p->sched_class;
4730 BUG_ON(prio < 0 || prio > MAX_PRIO);
4732 rq = task_rq_lock(p, &flags);
4733 update_rq_clock(rq);
4736 on_rq = p->se.on_rq;
4737 running = task_current(rq, p);
4739 dequeue_task(rq, p, 0);
4741 p->sched_class->put_prev_task(rq, p);
4744 p->sched_class = &rt_sched_class;
4746 p->sched_class = &fair_sched_class;
4751 p->sched_class->set_curr_task(rq);
4753 enqueue_task(rq, p, 0);
4755 check_class_changed(rq, p, prev_class, oldprio, running);
4757 task_rq_unlock(rq, &flags);
4762 void set_user_nice(struct task_struct *p, long nice)
4764 int old_prio, delta, on_rq;
4765 unsigned long flags;
4768 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4771 * We have to be careful, if called from sys_setpriority(),
4772 * the task might be in the middle of scheduling on another CPU.
4774 rq = task_rq_lock(p, &flags);
4775 update_rq_clock(rq);
4777 * The RT priorities are set via sched_setscheduler(), but we still
4778 * allow the 'normal' nice value to be set - but as expected
4779 * it wont have any effect on scheduling until the task is
4780 * SCHED_FIFO/SCHED_RR:
4782 if (task_has_rt_policy(p)) {
4783 p->static_prio = NICE_TO_PRIO(nice);
4786 on_rq = p->se.on_rq;
4788 dequeue_task(rq, p, 0);
4790 p->static_prio = NICE_TO_PRIO(nice);
4793 p->prio = effective_prio(p);
4794 delta = p->prio - old_prio;
4797 enqueue_task(rq, p, 0);
4799 * If the task increased its priority or is running and
4800 * lowered its priority, then reschedule its CPU:
4802 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4803 resched_task(rq->curr);
4806 task_rq_unlock(rq, &flags);
4808 EXPORT_SYMBOL(set_user_nice);
4811 * can_nice - check if a task can reduce its nice value
4815 int can_nice(const struct task_struct *p, const int nice)
4817 /* convert nice value [19,-20] to rlimit style value [1,40] */
4818 int nice_rlim = 20 - nice;
4820 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4821 capable(CAP_SYS_NICE));
4824 #ifdef __ARCH_WANT_SYS_NICE
4827 * sys_nice - change the priority of the current process.
4828 * @increment: priority increment
4830 * sys_setpriority is a more generic, but much slower function that
4831 * does similar things.
4833 asmlinkage long sys_nice(int increment)
4838 * Setpriority might change our priority at the same moment.
4839 * We don't have to worry. Conceptually one call occurs first
4840 * and we have a single winner.
4842 if (increment < -40)
4847 nice = PRIO_TO_NICE(current->static_prio) + increment;
4853 if (increment < 0 && !can_nice(current, nice))
4856 retval = security_task_setnice(current, nice);
4860 set_user_nice(current, nice);
4867 * task_prio - return the priority value of a given task.
4868 * @p: the task in question.
4870 * This is the priority value as seen by users in /proc.
4871 * RT tasks are offset by -200. Normal tasks are centered
4872 * around 0, value goes from -16 to +15.
4874 int task_prio(const struct task_struct *p)
4876 return p->prio - MAX_RT_PRIO;
4880 * task_nice - return the nice value of a given task.
4881 * @p: the task in question.
4883 int task_nice(const struct task_struct *p)
4885 return TASK_NICE(p);
4887 EXPORT_SYMBOL(task_nice);
4890 * idle_cpu - is a given cpu idle currently?
4891 * @cpu: the processor in question.
4893 int idle_cpu(int cpu)
4895 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4899 * idle_task - return the idle task for a given cpu.
4900 * @cpu: the processor in question.
4902 struct task_struct *idle_task(int cpu)
4904 return cpu_rq(cpu)->idle;
4908 * find_process_by_pid - find a process with a matching PID value.
4909 * @pid: the pid in question.
4911 static struct task_struct *find_process_by_pid(pid_t pid)
4913 return pid ? find_task_by_vpid(pid) : current;
4916 /* Actually do priority change: must hold rq lock. */
4918 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4920 BUG_ON(p->se.on_rq);
4923 switch (p->policy) {
4927 p->sched_class = &fair_sched_class;
4931 p->sched_class = &rt_sched_class;
4935 p->rt_priority = prio;
4936 p->normal_prio = normal_prio(p);
4937 /* we are holding p->pi_lock already */
4938 p->prio = rt_mutex_getprio(p);
4942 static int __sched_setscheduler(struct task_struct *p, int policy,
4943 struct sched_param *param, bool user)
4945 int retval, oldprio, oldpolicy = -1, on_rq, running;
4946 unsigned long flags;
4947 const struct sched_class *prev_class = p->sched_class;
4950 /* may grab non-irq protected spin_locks */
4951 BUG_ON(in_interrupt());
4953 /* double check policy once rq lock held */
4955 policy = oldpolicy = p->policy;
4956 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4957 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4958 policy != SCHED_IDLE)
4961 * Valid priorities for SCHED_FIFO and SCHED_RR are
4962 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4963 * SCHED_BATCH and SCHED_IDLE is 0.
4965 if (param->sched_priority < 0 ||
4966 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4967 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4969 if (rt_policy(policy) != (param->sched_priority != 0))
4973 * Allow unprivileged RT tasks to decrease priority:
4975 if (user && !capable(CAP_SYS_NICE)) {
4976 if (rt_policy(policy)) {
4977 unsigned long rlim_rtprio;
4979 if (!lock_task_sighand(p, &flags))
4981 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4982 unlock_task_sighand(p, &flags);
4984 /* can't set/change the rt policy */
4985 if (policy != p->policy && !rlim_rtprio)
4988 /* can't increase priority */
4989 if (param->sched_priority > p->rt_priority &&
4990 param->sched_priority > rlim_rtprio)
4994 * Like positive nice levels, dont allow tasks to
4995 * move out of SCHED_IDLE either:
4997 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5000 /* can't change other user's priorities */
5001 if ((current->euid != p->euid) &&
5002 (current->euid != p->uid))
5006 #ifdef CONFIG_RT_GROUP_SCHED
5008 * Do not allow realtime tasks into groups that have no runtime
5012 && rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5016 retval = security_task_setscheduler(p, policy, param);
5020 * make sure no PI-waiters arrive (or leave) while we are
5021 * changing the priority of the task:
5023 spin_lock_irqsave(&p->pi_lock, flags);
5025 * To be able to change p->policy safely, the apropriate
5026 * runqueue lock must be held.
5028 rq = __task_rq_lock(p);
5029 /* recheck policy now with rq lock held */
5030 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5031 policy = oldpolicy = -1;
5032 __task_rq_unlock(rq);
5033 spin_unlock_irqrestore(&p->pi_lock, flags);
5036 update_rq_clock(rq);
5037 on_rq = p->se.on_rq;
5038 running = task_current(rq, p);
5040 deactivate_task(rq, p, 0);
5042 p->sched_class->put_prev_task(rq, p);
5045 __setscheduler(rq, p, policy, param->sched_priority);
5048 p->sched_class->set_curr_task(rq);
5050 activate_task(rq, p, 0);
5052 check_class_changed(rq, p, prev_class, oldprio, running);
5054 __task_rq_unlock(rq);
5055 spin_unlock_irqrestore(&p->pi_lock, flags);
5057 rt_mutex_adjust_pi(p);
5063 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5064 * @p: the task in question.
5065 * @policy: new policy.
5066 * @param: structure containing the new RT priority.
5068 * NOTE that the task may be already dead.
5070 int sched_setscheduler(struct task_struct *p, int policy,
5071 struct sched_param *param)
5073 return __sched_setscheduler(p, policy, param, true);
5075 EXPORT_SYMBOL_GPL(sched_setscheduler);
5078 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5079 * @p: the task in question.
5080 * @policy: new policy.
5081 * @param: structure containing the new RT priority.
5083 * Just like sched_setscheduler, only don't bother checking if the
5084 * current context has permission. For example, this is needed in
5085 * stop_machine(): we create temporary high priority worker threads,
5086 * but our caller might not have that capability.
5088 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5089 struct sched_param *param)
5091 return __sched_setscheduler(p, policy, param, false);
5095 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5097 struct sched_param lparam;
5098 struct task_struct *p;
5101 if (!param || pid < 0)
5103 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5108 p = find_process_by_pid(pid);
5110 retval = sched_setscheduler(p, policy, &lparam);
5117 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5118 * @pid: the pid in question.
5119 * @policy: new policy.
5120 * @param: structure containing the new RT priority.
5123 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5125 /* negative values for policy are not valid */
5129 return do_sched_setscheduler(pid, policy, param);
5133 * sys_sched_setparam - set/change the RT priority of a thread
5134 * @pid: the pid in question.
5135 * @param: structure containing the new RT priority.
5137 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5139 return do_sched_setscheduler(pid, -1, param);
5143 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5144 * @pid: the pid in question.
5146 asmlinkage long sys_sched_getscheduler(pid_t pid)
5148 struct task_struct *p;
5155 read_lock(&tasklist_lock);
5156 p = find_process_by_pid(pid);
5158 retval = security_task_getscheduler(p);
5162 read_unlock(&tasklist_lock);
5167 * sys_sched_getscheduler - get the RT priority of a thread
5168 * @pid: the pid in question.
5169 * @param: structure containing the RT priority.
5171 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5173 struct sched_param lp;
5174 struct task_struct *p;
5177 if (!param || pid < 0)
5180 read_lock(&tasklist_lock);
5181 p = find_process_by_pid(pid);
5186 retval = security_task_getscheduler(p);
5190 lp.sched_priority = p->rt_priority;
5191 read_unlock(&tasklist_lock);
5194 * This one might sleep, we cannot do it with a spinlock held ...
5196 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5201 read_unlock(&tasklist_lock);
5205 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5207 cpumask_t cpus_allowed;
5208 cpumask_t new_mask = *in_mask;
5209 struct task_struct *p;
5213 read_lock(&tasklist_lock);
5215 p = find_process_by_pid(pid);
5217 read_unlock(&tasklist_lock);
5223 * It is not safe to call set_cpus_allowed with the
5224 * tasklist_lock held. We will bump the task_struct's
5225 * usage count and then drop tasklist_lock.
5228 read_unlock(&tasklist_lock);
5231 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5232 !capable(CAP_SYS_NICE))
5235 retval = security_task_setscheduler(p, 0, NULL);
5239 cpuset_cpus_allowed(p, &cpus_allowed);
5240 cpus_and(new_mask, new_mask, cpus_allowed);
5242 retval = set_cpus_allowed_ptr(p, &new_mask);
5245 cpuset_cpus_allowed(p, &cpus_allowed);
5246 if (!cpus_subset(new_mask, cpus_allowed)) {
5248 * We must have raced with a concurrent cpuset
5249 * update. Just reset the cpus_allowed to the
5250 * cpuset's cpus_allowed
5252 new_mask = cpus_allowed;
5262 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5263 cpumask_t *new_mask)
5265 if (len < sizeof(cpumask_t)) {
5266 memset(new_mask, 0, sizeof(cpumask_t));
5267 } else if (len > sizeof(cpumask_t)) {
5268 len = sizeof(cpumask_t);
5270 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5274 * sys_sched_setaffinity - set the cpu affinity of a process
5275 * @pid: pid of the process
5276 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5277 * @user_mask_ptr: user-space pointer to the new cpu mask
5279 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5280 unsigned long __user *user_mask_ptr)
5285 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5289 return sched_setaffinity(pid, &new_mask);
5292 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5294 struct task_struct *p;
5298 read_lock(&tasklist_lock);
5301 p = find_process_by_pid(pid);
5305 retval = security_task_getscheduler(p);
5309 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5312 read_unlock(&tasklist_lock);
5319 * sys_sched_getaffinity - get the cpu affinity of a process
5320 * @pid: pid of the process
5321 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5322 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5324 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5325 unsigned long __user *user_mask_ptr)
5330 if (len < sizeof(cpumask_t))
5333 ret = sched_getaffinity(pid, &mask);
5337 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5340 return sizeof(cpumask_t);
5344 * sys_sched_yield - yield the current processor to other threads.
5346 * This function yields the current CPU to other tasks. If there are no
5347 * other threads running on this CPU then this function will return.
5349 asmlinkage long sys_sched_yield(void)
5351 struct rq *rq = this_rq_lock();
5353 schedstat_inc(rq, yld_count);
5354 current->sched_class->yield_task(rq);
5357 * Since we are going to call schedule() anyway, there's
5358 * no need to preempt or enable interrupts:
5360 __release(rq->lock);
5361 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5362 _raw_spin_unlock(&rq->lock);
5363 preempt_enable_no_resched();
5370 static void __cond_resched(void)
5372 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5373 __might_sleep(__FILE__, __LINE__);
5376 * The BKS might be reacquired before we have dropped
5377 * PREEMPT_ACTIVE, which could trigger a second
5378 * cond_resched() call.
5381 add_preempt_count(PREEMPT_ACTIVE);
5383 sub_preempt_count(PREEMPT_ACTIVE);
5384 } while (need_resched());
5387 int __sched _cond_resched(void)
5389 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5390 system_state == SYSTEM_RUNNING) {
5396 EXPORT_SYMBOL(_cond_resched);
5399 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5400 * call schedule, and on return reacquire the lock.
5402 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5403 * operations here to prevent schedule() from being called twice (once via
5404 * spin_unlock(), once by hand).
5406 int cond_resched_lock(spinlock_t *lock)
5408 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5411 if (spin_needbreak(lock) || resched) {
5413 if (resched && need_resched())
5422 EXPORT_SYMBOL(cond_resched_lock);
5424 int __sched cond_resched_softirq(void)
5426 BUG_ON(!in_softirq());
5428 if (need_resched() && system_state == SYSTEM_RUNNING) {
5436 EXPORT_SYMBOL(cond_resched_softirq);
5439 * yield - yield the current processor to other threads.
5441 * This is a shortcut for kernel-space yielding - it marks the
5442 * thread runnable and calls sys_sched_yield().
5444 void __sched yield(void)
5446 set_current_state(TASK_RUNNING);
5449 EXPORT_SYMBOL(yield);
5452 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5453 * that process accounting knows that this is a task in IO wait state.
5455 * But don't do that if it is a deliberate, throttling IO wait (this task
5456 * has set its backing_dev_info: the queue against which it should throttle)
5458 void __sched io_schedule(void)
5460 struct rq *rq = &__raw_get_cpu_var(runqueues);
5462 delayacct_blkio_start();
5463 atomic_inc(&rq->nr_iowait);
5465 atomic_dec(&rq->nr_iowait);
5466 delayacct_blkio_end();
5468 EXPORT_SYMBOL(io_schedule);
5470 long __sched io_schedule_timeout(long timeout)
5472 struct rq *rq = &__raw_get_cpu_var(runqueues);
5475 delayacct_blkio_start();
5476 atomic_inc(&rq->nr_iowait);
5477 ret = schedule_timeout(timeout);
5478 atomic_dec(&rq->nr_iowait);
5479 delayacct_blkio_end();
5484 * sys_sched_get_priority_max - return maximum RT priority.
5485 * @policy: scheduling class.
5487 * this syscall returns the maximum rt_priority that can be used
5488 * by a given scheduling class.
5490 asmlinkage long sys_sched_get_priority_max(int policy)
5497 ret = MAX_USER_RT_PRIO-1;
5509 * sys_sched_get_priority_min - return minimum RT priority.
5510 * @policy: scheduling class.
5512 * this syscall returns the minimum rt_priority that can be used
5513 * by a given scheduling class.
5515 asmlinkage long sys_sched_get_priority_min(int policy)
5533 * sys_sched_rr_get_interval - return the default timeslice of a process.
5534 * @pid: pid of the process.
5535 * @interval: userspace pointer to the timeslice value.
5537 * this syscall writes the default timeslice value of a given process
5538 * into the user-space timespec buffer. A value of '0' means infinity.
5541 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5543 struct task_struct *p;
5544 unsigned int time_slice;
5552 read_lock(&tasklist_lock);
5553 p = find_process_by_pid(pid);
5557 retval = security_task_getscheduler(p);
5562 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5563 * tasks that are on an otherwise idle runqueue:
5566 if (p->policy == SCHED_RR) {
5567 time_slice = DEF_TIMESLICE;
5568 } else if (p->policy != SCHED_FIFO) {
5569 struct sched_entity *se = &p->se;
5570 unsigned long flags;
5573 rq = task_rq_lock(p, &flags);
5574 if (rq->cfs.load.weight)
5575 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5576 task_rq_unlock(rq, &flags);
5578 read_unlock(&tasklist_lock);
5579 jiffies_to_timespec(time_slice, &t);
5580 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5584 read_unlock(&tasklist_lock);
5588 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5590 void sched_show_task(struct task_struct *p)
5592 unsigned long free = 0;
5595 state = p->state ? __ffs(p->state) + 1 : 0;
5596 printk(KERN_INFO "%-13.13s %c", p->comm,
5597 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5598 #if BITS_PER_LONG == 32
5599 if (state == TASK_RUNNING)
5600 printk(KERN_CONT " running ");
5602 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5604 if (state == TASK_RUNNING)
5605 printk(KERN_CONT " running task ");
5607 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5609 #ifdef CONFIG_DEBUG_STACK_USAGE
5611 unsigned long *n = end_of_stack(p);
5614 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5617 printk(KERN_CONT "%5lu %5d %6d\n", free,
5618 task_pid_nr(p), task_pid_nr(p->real_parent));
5620 show_stack(p, NULL);
5623 void show_state_filter(unsigned long state_filter)
5625 struct task_struct *g, *p;
5627 #if BITS_PER_LONG == 32
5629 " task PC stack pid father\n");
5632 " task PC stack pid father\n");
5634 read_lock(&tasklist_lock);
5635 do_each_thread(g, p) {
5637 * reset the NMI-timeout, listing all files on a slow
5638 * console might take alot of time:
5640 touch_nmi_watchdog();
5641 if (!state_filter || (p->state & state_filter))
5643 } while_each_thread(g, p);
5645 touch_all_softlockup_watchdogs();
5647 #ifdef CONFIG_SCHED_DEBUG
5648 sysrq_sched_debug_show();
5650 read_unlock(&tasklist_lock);
5652 * Only show locks if all tasks are dumped:
5654 if (state_filter == -1)
5655 debug_show_all_locks();
5658 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5660 idle->sched_class = &idle_sched_class;
5664 * init_idle - set up an idle thread for a given CPU
5665 * @idle: task in question
5666 * @cpu: cpu the idle task belongs to
5668 * NOTE: this function does not set the idle thread's NEED_RESCHED
5669 * flag, to make booting more robust.
5671 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5673 struct rq *rq = cpu_rq(cpu);
5674 unsigned long flags;
5677 idle->se.exec_start = sched_clock();
5679 idle->prio = idle->normal_prio = MAX_PRIO;
5680 idle->cpus_allowed = cpumask_of_cpu(cpu);
5681 __set_task_cpu(idle, cpu);
5683 spin_lock_irqsave(&rq->lock, flags);
5684 rq->curr = rq->idle = idle;
5685 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5688 spin_unlock_irqrestore(&rq->lock, flags);
5690 /* Set the preempt count _outside_ the spinlocks! */
5691 #if defined(CONFIG_PREEMPT)
5692 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5694 task_thread_info(idle)->preempt_count = 0;
5697 * The idle tasks have their own, simple scheduling class:
5699 idle->sched_class = &idle_sched_class;
5703 * In a system that switches off the HZ timer nohz_cpu_mask
5704 * indicates which cpus entered this state. This is used
5705 * in the rcu update to wait only for active cpus. For system
5706 * which do not switch off the HZ timer nohz_cpu_mask should
5707 * always be CPU_MASK_NONE.
5709 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5712 * Increase the granularity value when there are more CPUs,
5713 * because with more CPUs the 'effective latency' as visible
5714 * to users decreases. But the relationship is not linear,
5715 * so pick a second-best guess by going with the log2 of the
5718 * This idea comes from the SD scheduler of Con Kolivas:
5720 static inline void sched_init_granularity(void)
5722 unsigned int factor = 1 + ilog2(num_online_cpus());
5723 const unsigned long limit = 200000000;
5725 sysctl_sched_min_granularity *= factor;
5726 if (sysctl_sched_min_granularity > limit)
5727 sysctl_sched_min_granularity = limit;
5729 sysctl_sched_latency *= factor;
5730 if (sysctl_sched_latency > limit)
5731 sysctl_sched_latency = limit;
5733 sysctl_sched_wakeup_granularity *= factor;
5738 * This is how migration works:
5740 * 1) we queue a struct migration_req structure in the source CPU's
5741 * runqueue and wake up that CPU's migration thread.
5742 * 2) we down() the locked semaphore => thread blocks.
5743 * 3) migration thread wakes up (implicitly it forces the migrated
5744 * thread off the CPU)
5745 * 4) it gets the migration request and checks whether the migrated
5746 * task is still in the wrong runqueue.
5747 * 5) if it's in the wrong runqueue then the migration thread removes
5748 * it and puts it into the right queue.
5749 * 6) migration thread up()s the semaphore.
5750 * 7) we wake up and the migration is done.
5754 * Change a given task's CPU affinity. Migrate the thread to a
5755 * proper CPU and schedule it away if the CPU it's executing on
5756 * is removed from the allowed bitmask.
5758 * NOTE: the caller must have a valid reference to the task, the
5759 * task must not exit() & deallocate itself prematurely. The
5760 * call is not atomic; no spinlocks may be held.
5762 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5764 struct migration_req req;
5765 unsigned long flags;
5769 rq = task_rq_lock(p, &flags);
5770 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5775 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5776 !cpus_equal(p->cpus_allowed, *new_mask))) {
5781 if (p->sched_class->set_cpus_allowed)
5782 p->sched_class->set_cpus_allowed(p, new_mask);
5784 p->cpus_allowed = *new_mask;
5785 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5788 /* Can the task run on the task's current CPU? If so, we're done */
5789 if (cpu_isset(task_cpu(p), *new_mask))
5792 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5793 /* Need help from migration thread: drop lock and wait. */
5794 task_rq_unlock(rq, &flags);
5795 wake_up_process(rq->migration_thread);
5796 wait_for_completion(&req.done);
5797 tlb_migrate_finish(p->mm);
5801 task_rq_unlock(rq, &flags);
5805 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5808 * Move (not current) task off this cpu, onto dest cpu. We're doing
5809 * this because either it can't run here any more (set_cpus_allowed()
5810 * away from this CPU, or CPU going down), or because we're
5811 * attempting to rebalance this task on exec (sched_exec).
5813 * So we race with normal scheduler movements, but that's OK, as long
5814 * as the task is no longer on this CPU.
5816 * Returns non-zero if task was successfully migrated.
5818 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5820 struct rq *rq_dest, *rq_src;
5823 if (unlikely(!cpu_active(dest_cpu)))
5826 rq_src = cpu_rq(src_cpu);
5827 rq_dest = cpu_rq(dest_cpu);
5829 double_rq_lock(rq_src, rq_dest);
5830 /* Already moved. */
5831 if (task_cpu(p) != src_cpu)
5833 /* Affinity changed (again). */
5834 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5837 on_rq = p->se.on_rq;
5839 deactivate_task(rq_src, p, 0);
5841 set_task_cpu(p, dest_cpu);
5843 activate_task(rq_dest, p, 0);
5844 check_preempt_curr(rq_dest, p);
5849 double_rq_unlock(rq_src, rq_dest);
5854 * migration_thread - this is a highprio system thread that performs
5855 * thread migration by bumping thread off CPU then 'pushing' onto
5858 static int migration_thread(void *data)
5860 int cpu = (long)data;
5864 BUG_ON(rq->migration_thread != current);
5866 set_current_state(TASK_INTERRUPTIBLE);
5867 while (!kthread_should_stop()) {
5868 struct migration_req *req;
5869 struct list_head *head;
5871 spin_lock_irq(&rq->lock);
5873 if (cpu_is_offline(cpu)) {
5874 spin_unlock_irq(&rq->lock);
5878 if (rq->active_balance) {
5879 active_load_balance(rq, cpu);
5880 rq->active_balance = 0;
5883 head = &rq->migration_queue;
5885 if (list_empty(head)) {
5886 spin_unlock_irq(&rq->lock);
5888 set_current_state(TASK_INTERRUPTIBLE);
5891 req = list_entry(head->next, struct migration_req, list);
5892 list_del_init(head->next);
5894 spin_unlock(&rq->lock);
5895 __migrate_task(req->task, cpu, req->dest_cpu);
5898 complete(&req->done);
5900 __set_current_state(TASK_RUNNING);
5904 /* Wait for kthread_stop */
5905 set_current_state(TASK_INTERRUPTIBLE);
5906 while (!kthread_should_stop()) {
5908 set_current_state(TASK_INTERRUPTIBLE);
5910 __set_current_state(TASK_RUNNING);
5914 #ifdef CONFIG_HOTPLUG_CPU
5916 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5920 local_irq_disable();
5921 ret = __migrate_task(p, src_cpu, dest_cpu);
5927 * Figure out where task on dead CPU should go, use force if necessary.
5928 * NOTE: interrupts should be disabled by the caller
5930 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5932 unsigned long flags;
5939 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5940 cpus_and(mask, mask, p->cpus_allowed);
5941 dest_cpu = any_online_cpu(mask);
5943 /* On any allowed CPU? */
5944 if (dest_cpu >= nr_cpu_ids)
5945 dest_cpu = any_online_cpu(p->cpus_allowed);
5947 /* No more Mr. Nice Guy. */
5948 if (dest_cpu >= nr_cpu_ids) {
5949 cpumask_t cpus_allowed;
5951 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5953 * Try to stay on the same cpuset, where the
5954 * current cpuset may be a subset of all cpus.
5955 * The cpuset_cpus_allowed_locked() variant of
5956 * cpuset_cpus_allowed() will not block. It must be
5957 * called within calls to cpuset_lock/cpuset_unlock.
5959 rq = task_rq_lock(p, &flags);
5960 p->cpus_allowed = cpus_allowed;
5961 dest_cpu = any_online_cpu(p->cpus_allowed);
5962 task_rq_unlock(rq, &flags);
5965 * Don't tell them about moving exiting tasks or
5966 * kernel threads (both mm NULL), since they never
5969 if (p->mm && printk_ratelimit()) {
5970 printk(KERN_INFO "process %d (%s) no "
5971 "longer affine to cpu%d\n",
5972 task_pid_nr(p), p->comm, dead_cpu);
5975 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5979 * While a dead CPU has no uninterruptible tasks queued at this point,
5980 * it might still have a nonzero ->nr_uninterruptible counter, because
5981 * for performance reasons the counter is not stricly tracking tasks to
5982 * their home CPUs. So we just add the counter to another CPU's counter,
5983 * to keep the global sum constant after CPU-down:
5985 static void migrate_nr_uninterruptible(struct rq *rq_src)
5987 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5988 unsigned long flags;
5990 local_irq_save(flags);
5991 double_rq_lock(rq_src, rq_dest);
5992 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5993 rq_src->nr_uninterruptible = 0;
5994 double_rq_unlock(rq_src, rq_dest);
5995 local_irq_restore(flags);
5998 /* Run through task list and migrate tasks from the dead cpu. */
5999 static void migrate_live_tasks(int src_cpu)
6001 struct task_struct *p, *t;
6003 read_lock(&tasklist_lock);
6005 do_each_thread(t, p) {
6009 if (task_cpu(p) == src_cpu)
6010 move_task_off_dead_cpu(src_cpu, p);
6011 } while_each_thread(t, p);
6013 read_unlock(&tasklist_lock);
6017 * Schedules idle task to be the next runnable task on current CPU.
6018 * It does so by boosting its priority to highest possible.
6019 * Used by CPU offline code.
6021 void sched_idle_next(void)
6023 int this_cpu = smp_processor_id();
6024 struct rq *rq = cpu_rq(this_cpu);
6025 struct task_struct *p = rq->idle;
6026 unsigned long flags;
6028 /* cpu has to be offline */
6029 BUG_ON(cpu_online(this_cpu));
6032 * Strictly not necessary since rest of the CPUs are stopped by now
6033 * and interrupts disabled on the current cpu.
6035 spin_lock_irqsave(&rq->lock, flags);
6037 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6039 update_rq_clock(rq);
6040 activate_task(rq, p, 0);
6042 spin_unlock_irqrestore(&rq->lock, flags);
6046 * Ensures that the idle task is using init_mm right before its cpu goes
6049 void idle_task_exit(void)
6051 struct mm_struct *mm = current->active_mm;
6053 BUG_ON(cpu_online(smp_processor_id()));
6056 switch_mm(mm, &init_mm, current);
6060 /* called under rq->lock with disabled interrupts */
6061 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6063 struct rq *rq = cpu_rq(dead_cpu);
6065 /* Must be exiting, otherwise would be on tasklist. */
6066 BUG_ON(!p->exit_state);
6068 /* Cannot have done final schedule yet: would have vanished. */
6069 BUG_ON(p->state == TASK_DEAD);
6074 * Drop lock around migration; if someone else moves it,
6075 * that's OK. No task can be added to this CPU, so iteration is
6078 spin_unlock_irq(&rq->lock);
6079 move_task_off_dead_cpu(dead_cpu, p);
6080 spin_lock_irq(&rq->lock);
6085 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6086 static void migrate_dead_tasks(unsigned int dead_cpu)
6088 struct rq *rq = cpu_rq(dead_cpu);
6089 struct task_struct *next;
6092 if (!rq->nr_running)
6094 update_rq_clock(rq);
6095 next = pick_next_task(rq, rq->curr);
6098 next->sched_class->put_prev_task(rq, next);
6099 migrate_dead(dead_cpu, next);
6103 #endif /* CONFIG_HOTPLUG_CPU */
6105 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6107 static struct ctl_table sd_ctl_dir[] = {
6109 .procname = "sched_domain",
6115 static struct ctl_table sd_ctl_root[] = {
6117 .ctl_name = CTL_KERN,
6118 .procname = "kernel",
6120 .child = sd_ctl_dir,
6125 static struct ctl_table *sd_alloc_ctl_entry(int n)
6127 struct ctl_table *entry =
6128 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6133 static void sd_free_ctl_entry(struct ctl_table **tablep)
6135 struct ctl_table *entry;
6138 * In the intermediate directories, both the child directory and
6139 * procname are dynamically allocated and could fail but the mode
6140 * will always be set. In the lowest directory the names are
6141 * static strings and all have proc handlers.
6143 for (entry = *tablep; entry->mode; entry++) {
6145 sd_free_ctl_entry(&entry->child);
6146 if (entry->proc_handler == NULL)
6147 kfree(entry->procname);
6155 set_table_entry(struct ctl_table *entry,
6156 const char *procname, void *data, int maxlen,
6157 mode_t mode, proc_handler *proc_handler)
6159 entry->procname = procname;
6161 entry->maxlen = maxlen;
6163 entry->proc_handler = proc_handler;
6166 static struct ctl_table *
6167 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6169 struct ctl_table *table = sd_alloc_ctl_entry(12);
6174 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6175 sizeof(long), 0644, proc_doulongvec_minmax);
6176 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6177 sizeof(long), 0644, proc_doulongvec_minmax);
6178 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6179 sizeof(int), 0644, proc_dointvec_minmax);
6180 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6181 sizeof(int), 0644, proc_dointvec_minmax);
6182 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6183 sizeof(int), 0644, proc_dointvec_minmax);
6184 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6185 sizeof(int), 0644, proc_dointvec_minmax);
6186 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6187 sizeof(int), 0644, proc_dointvec_minmax);
6188 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6189 sizeof(int), 0644, proc_dointvec_minmax);
6190 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6191 sizeof(int), 0644, proc_dointvec_minmax);
6192 set_table_entry(&table[9], "cache_nice_tries",
6193 &sd->cache_nice_tries,
6194 sizeof(int), 0644, proc_dointvec_minmax);
6195 set_table_entry(&table[10], "flags", &sd->flags,
6196 sizeof(int), 0644, proc_dointvec_minmax);
6197 /* &table[11] is terminator */
6202 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6204 struct ctl_table *entry, *table;
6205 struct sched_domain *sd;
6206 int domain_num = 0, i;
6209 for_each_domain(cpu, sd)
6211 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6216 for_each_domain(cpu, sd) {
6217 snprintf(buf, 32, "domain%d", i);
6218 entry->procname = kstrdup(buf, GFP_KERNEL);
6220 entry->child = sd_alloc_ctl_domain_table(sd);
6227 static struct ctl_table_header *sd_sysctl_header;
6228 static void register_sched_domain_sysctl(void)
6230 int i, cpu_num = num_online_cpus();
6231 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6234 WARN_ON(sd_ctl_dir[0].child);
6235 sd_ctl_dir[0].child = entry;
6240 for_each_online_cpu(i) {
6241 snprintf(buf, 32, "cpu%d", i);
6242 entry->procname = kstrdup(buf, GFP_KERNEL);
6244 entry->child = sd_alloc_ctl_cpu_table(i);
6248 WARN_ON(sd_sysctl_header);
6249 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6252 /* may be called multiple times per register */
6253 static void unregister_sched_domain_sysctl(void)
6255 if (sd_sysctl_header)
6256 unregister_sysctl_table(sd_sysctl_header);
6257 sd_sysctl_header = NULL;
6258 if (sd_ctl_dir[0].child)
6259 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6262 static void register_sched_domain_sysctl(void)
6265 static void unregister_sched_domain_sysctl(void)
6270 static void set_rq_online(struct rq *rq)
6273 const struct sched_class *class;
6275 cpu_set(rq->cpu, rq->rd->online);
6278 for_each_class(class) {
6279 if (class->rq_online)
6280 class->rq_online(rq);
6285 static void set_rq_offline(struct rq *rq)
6288 const struct sched_class *class;
6290 for_each_class(class) {
6291 if (class->rq_offline)
6292 class->rq_offline(rq);
6295 cpu_clear(rq->cpu, rq->rd->online);
6301 * migration_call - callback that gets triggered when a CPU is added.
6302 * Here we can start up the necessary migration thread for the new CPU.
6304 static int __cpuinit
6305 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6307 struct task_struct *p;
6308 int cpu = (long)hcpu;
6309 unsigned long flags;
6314 case CPU_UP_PREPARE:
6315 case CPU_UP_PREPARE_FROZEN:
6316 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6319 kthread_bind(p, cpu);
6320 /* Must be high prio: stop_machine expects to yield to it. */
6321 rq = task_rq_lock(p, &flags);
6322 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6323 task_rq_unlock(rq, &flags);
6324 cpu_rq(cpu)->migration_thread = p;
6328 case CPU_ONLINE_FROZEN:
6329 /* Strictly unnecessary, as first user will wake it. */
6330 wake_up_process(cpu_rq(cpu)->migration_thread);
6332 /* Update our root-domain */
6334 spin_lock_irqsave(&rq->lock, flags);
6336 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6340 spin_unlock_irqrestore(&rq->lock, flags);
6343 #ifdef CONFIG_HOTPLUG_CPU
6344 case CPU_UP_CANCELED:
6345 case CPU_UP_CANCELED_FROZEN:
6346 if (!cpu_rq(cpu)->migration_thread)
6348 /* Unbind it from offline cpu so it can run. Fall thru. */
6349 kthread_bind(cpu_rq(cpu)->migration_thread,
6350 any_online_cpu(cpu_online_map));
6351 kthread_stop(cpu_rq(cpu)->migration_thread);
6352 cpu_rq(cpu)->migration_thread = NULL;
6356 case CPU_DEAD_FROZEN:
6357 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6358 migrate_live_tasks(cpu);
6360 kthread_stop(rq->migration_thread);
6361 rq->migration_thread = NULL;
6362 /* Idle task back to normal (off runqueue, low prio) */
6363 spin_lock_irq(&rq->lock);
6364 update_rq_clock(rq);
6365 deactivate_task(rq, rq->idle, 0);
6366 rq->idle->static_prio = MAX_PRIO;
6367 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6368 rq->idle->sched_class = &idle_sched_class;
6369 migrate_dead_tasks(cpu);
6370 spin_unlock_irq(&rq->lock);
6372 migrate_nr_uninterruptible(rq);
6373 BUG_ON(rq->nr_running != 0);
6376 * No need to migrate the tasks: it was best-effort if
6377 * they didn't take sched_hotcpu_mutex. Just wake up
6380 spin_lock_irq(&rq->lock);
6381 while (!list_empty(&rq->migration_queue)) {
6382 struct migration_req *req;
6384 req = list_entry(rq->migration_queue.next,
6385 struct migration_req, list);
6386 list_del_init(&req->list);
6387 complete(&req->done);
6389 spin_unlock_irq(&rq->lock);
6393 case CPU_DYING_FROZEN:
6394 /* Update our root-domain */
6396 spin_lock_irqsave(&rq->lock, flags);
6398 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6401 spin_unlock_irqrestore(&rq->lock, flags);
6408 /* Register at highest priority so that task migration (migrate_all_tasks)
6409 * happens before everything else.
6411 static struct notifier_block __cpuinitdata migration_notifier = {
6412 .notifier_call = migration_call,
6416 static int __init migration_init(void)
6418 void *cpu = (void *)(long)smp_processor_id();
6421 /* Start one for the boot CPU: */
6422 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6423 BUG_ON(err == NOTIFY_BAD);
6424 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6425 register_cpu_notifier(&migration_notifier);
6429 early_initcall(migration_init);
6434 #ifdef CONFIG_SCHED_DEBUG
6436 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6449 case SD_LV_ALLNODES:
6458 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6459 cpumask_t *groupmask)
6461 struct sched_group *group = sd->groups;
6464 cpulist_scnprintf(str, sizeof(str), sd->span);
6465 cpus_clear(*groupmask);
6467 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6469 if (!(sd->flags & SD_LOAD_BALANCE)) {
6470 printk("does not load-balance\n");
6472 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6477 printk(KERN_CONT "span %s level %s\n",
6478 str, sd_level_to_string(sd->level));
6480 if (!cpu_isset(cpu, sd->span)) {
6481 printk(KERN_ERR "ERROR: domain->span does not contain "
6484 if (!cpu_isset(cpu, group->cpumask)) {
6485 printk(KERN_ERR "ERROR: domain->groups does not contain"
6489 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6493 printk(KERN_ERR "ERROR: group is NULL\n");
6497 if (!group->__cpu_power) {
6498 printk(KERN_CONT "\n");
6499 printk(KERN_ERR "ERROR: domain->cpu_power not "
6504 if (!cpus_weight(group->cpumask)) {
6505 printk(KERN_CONT "\n");
6506 printk(KERN_ERR "ERROR: empty group\n");
6510 if (cpus_intersects(*groupmask, group->cpumask)) {
6511 printk(KERN_CONT "\n");
6512 printk(KERN_ERR "ERROR: repeated CPUs\n");
6516 cpus_or(*groupmask, *groupmask, group->cpumask);
6518 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6519 printk(KERN_CONT " %s", str);
6521 group = group->next;
6522 } while (group != sd->groups);
6523 printk(KERN_CONT "\n");
6525 if (!cpus_equal(sd->span, *groupmask))
6526 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6528 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6529 printk(KERN_ERR "ERROR: parent span is not a superset "
6530 "of domain->span\n");
6534 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6536 cpumask_t *groupmask;
6540 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6544 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6546 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6548 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6553 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6562 #else /* !CONFIG_SCHED_DEBUG */
6563 # define sched_domain_debug(sd, cpu) do { } while (0)
6564 #endif /* CONFIG_SCHED_DEBUG */
6566 static int sd_degenerate(struct sched_domain *sd)
6568 if (cpus_weight(sd->span) == 1)
6571 /* Following flags need at least 2 groups */
6572 if (sd->flags & (SD_LOAD_BALANCE |
6573 SD_BALANCE_NEWIDLE |
6577 SD_SHARE_PKG_RESOURCES)) {
6578 if (sd->groups != sd->groups->next)
6582 /* Following flags don't use groups */
6583 if (sd->flags & (SD_WAKE_IDLE |
6592 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6594 unsigned long cflags = sd->flags, pflags = parent->flags;
6596 if (sd_degenerate(parent))
6599 if (!cpus_equal(sd->span, parent->span))
6602 /* Does parent contain flags not in child? */
6603 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6604 if (cflags & SD_WAKE_AFFINE)
6605 pflags &= ~SD_WAKE_BALANCE;
6606 /* Flags needing groups don't count if only 1 group in parent */
6607 if (parent->groups == parent->groups->next) {
6608 pflags &= ~(SD_LOAD_BALANCE |
6609 SD_BALANCE_NEWIDLE |
6613 SD_SHARE_PKG_RESOURCES);
6615 if (~cflags & pflags)
6621 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6623 unsigned long flags;
6625 spin_lock_irqsave(&rq->lock, flags);
6628 struct root_domain *old_rd = rq->rd;
6630 if (cpu_isset(rq->cpu, old_rd->online))
6633 cpu_clear(rq->cpu, old_rd->span);
6635 if (atomic_dec_and_test(&old_rd->refcount))
6639 atomic_inc(&rd->refcount);
6642 cpu_set(rq->cpu, rd->span);
6643 if (cpu_isset(rq->cpu, cpu_online_map))
6646 spin_unlock_irqrestore(&rq->lock, flags);
6649 static void init_rootdomain(struct root_domain *rd)
6651 memset(rd, 0, sizeof(*rd));
6653 cpus_clear(rd->span);
6654 cpus_clear(rd->online);
6656 cpupri_init(&rd->cpupri);
6659 static void init_defrootdomain(void)
6661 init_rootdomain(&def_root_domain);
6662 atomic_set(&def_root_domain.refcount, 1);
6665 static struct root_domain *alloc_rootdomain(void)
6667 struct root_domain *rd;
6669 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6673 init_rootdomain(rd);
6679 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6680 * hold the hotplug lock.
6683 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6685 struct rq *rq = cpu_rq(cpu);
6686 struct sched_domain *tmp;
6688 /* Remove the sched domains which do not contribute to scheduling. */
6689 for (tmp = sd; tmp; tmp = tmp->parent) {
6690 struct sched_domain *parent = tmp->parent;
6693 if (sd_parent_degenerate(tmp, parent)) {
6694 tmp->parent = parent->parent;
6696 parent->parent->child = tmp;
6700 if (sd && sd_degenerate(sd)) {
6706 sched_domain_debug(sd, cpu);
6708 rq_attach_root(rq, rd);
6709 rcu_assign_pointer(rq->sd, sd);
6712 /* cpus with isolated domains */
6713 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6715 /* Setup the mask of cpus configured for isolated domains */
6716 static int __init isolated_cpu_setup(char *str)
6718 static int __initdata ints[NR_CPUS];
6721 str = get_options(str, ARRAY_SIZE(ints), ints);
6722 cpus_clear(cpu_isolated_map);
6723 for (i = 1; i <= ints[0]; i++)
6724 if (ints[i] < NR_CPUS)
6725 cpu_set(ints[i], cpu_isolated_map);
6729 __setup("isolcpus=", isolated_cpu_setup);
6732 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6733 * to a function which identifies what group(along with sched group) a CPU
6734 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6735 * (due to the fact that we keep track of groups covered with a cpumask_t).
6737 * init_sched_build_groups will build a circular linked list of the groups
6738 * covered by the given span, and will set each group's ->cpumask correctly,
6739 * and ->cpu_power to 0.
6742 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6743 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6744 struct sched_group **sg,
6745 cpumask_t *tmpmask),
6746 cpumask_t *covered, cpumask_t *tmpmask)
6748 struct sched_group *first = NULL, *last = NULL;
6751 cpus_clear(*covered);
6753 for_each_cpu_mask_nr(i, *span) {
6754 struct sched_group *sg;
6755 int group = group_fn(i, cpu_map, &sg, tmpmask);
6758 if (cpu_isset(i, *covered))
6761 cpus_clear(sg->cpumask);
6762 sg->__cpu_power = 0;
6764 for_each_cpu_mask_nr(j, *span) {
6765 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6768 cpu_set(j, *covered);
6769 cpu_set(j, sg->cpumask);
6780 #define SD_NODES_PER_DOMAIN 16
6785 * find_next_best_node - find the next node to include in a sched_domain
6786 * @node: node whose sched_domain we're building
6787 * @used_nodes: nodes already in the sched_domain
6789 * Find the next node to include in a given scheduling domain. Simply
6790 * finds the closest node not already in the @used_nodes map.
6792 * Should use nodemask_t.
6794 static int find_next_best_node(int node, nodemask_t *used_nodes)
6796 int i, n, val, min_val, best_node = 0;
6800 for (i = 0; i < nr_node_ids; i++) {
6801 /* Start at @node */
6802 n = (node + i) % nr_node_ids;
6804 if (!nr_cpus_node(n))
6807 /* Skip already used nodes */
6808 if (node_isset(n, *used_nodes))
6811 /* Simple min distance search */
6812 val = node_distance(node, n);
6814 if (val < min_val) {
6820 node_set(best_node, *used_nodes);
6825 * sched_domain_node_span - get a cpumask for a node's sched_domain
6826 * @node: node whose cpumask we're constructing
6827 * @span: resulting cpumask
6829 * Given a node, construct a good cpumask for its sched_domain to span. It
6830 * should be one that prevents unnecessary balancing, but also spreads tasks
6833 static void sched_domain_node_span(int node, cpumask_t *span)
6835 nodemask_t used_nodes;
6836 node_to_cpumask_ptr(nodemask, node);
6840 nodes_clear(used_nodes);
6842 cpus_or(*span, *span, *nodemask);
6843 node_set(node, used_nodes);
6845 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6846 int next_node = find_next_best_node(node, &used_nodes);
6848 node_to_cpumask_ptr_next(nodemask, next_node);
6849 cpus_or(*span, *span, *nodemask);
6852 #endif /* CONFIG_NUMA */
6854 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6857 * SMT sched-domains:
6859 #ifdef CONFIG_SCHED_SMT
6860 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6861 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6864 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6868 *sg = &per_cpu(sched_group_cpus, cpu);
6871 #endif /* CONFIG_SCHED_SMT */
6874 * multi-core sched-domains:
6876 #ifdef CONFIG_SCHED_MC
6877 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6878 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6879 #endif /* CONFIG_SCHED_MC */
6881 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6883 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6888 *mask = per_cpu(cpu_sibling_map, cpu);
6889 cpus_and(*mask, *mask, *cpu_map);
6890 group = first_cpu(*mask);
6892 *sg = &per_cpu(sched_group_core, group);
6895 #elif defined(CONFIG_SCHED_MC)
6897 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6901 *sg = &per_cpu(sched_group_core, cpu);
6906 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6907 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6910 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6914 #ifdef CONFIG_SCHED_MC
6915 *mask = cpu_coregroup_map(cpu);
6916 cpus_and(*mask, *mask, *cpu_map);
6917 group = first_cpu(*mask);
6918 #elif defined(CONFIG_SCHED_SMT)
6919 *mask = per_cpu(cpu_sibling_map, cpu);
6920 cpus_and(*mask, *mask, *cpu_map);
6921 group = first_cpu(*mask);
6926 *sg = &per_cpu(sched_group_phys, group);
6932 * The init_sched_build_groups can't handle what we want to do with node
6933 * groups, so roll our own. Now each node has its own list of groups which
6934 * gets dynamically allocated.
6936 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6937 static struct sched_group ***sched_group_nodes_bycpu;
6939 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6940 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6942 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6943 struct sched_group **sg, cpumask_t *nodemask)
6947 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6948 cpus_and(*nodemask, *nodemask, *cpu_map);
6949 group = first_cpu(*nodemask);
6952 *sg = &per_cpu(sched_group_allnodes, group);
6956 static void init_numa_sched_groups_power(struct sched_group *group_head)
6958 struct sched_group *sg = group_head;
6964 for_each_cpu_mask_nr(j, sg->cpumask) {
6965 struct sched_domain *sd;
6967 sd = &per_cpu(phys_domains, j);
6968 if (j != first_cpu(sd->groups->cpumask)) {
6970 * Only add "power" once for each
6976 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6979 } while (sg != group_head);
6981 #endif /* CONFIG_NUMA */
6984 /* Free memory allocated for various sched_group structures */
6985 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6989 for_each_cpu_mask_nr(cpu, *cpu_map) {
6990 struct sched_group **sched_group_nodes
6991 = sched_group_nodes_bycpu[cpu];
6993 if (!sched_group_nodes)
6996 for (i = 0; i < nr_node_ids; i++) {
6997 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6999 *nodemask = node_to_cpumask(i);
7000 cpus_and(*nodemask, *nodemask, *cpu_map);
7001 if (cpus_empty(*nodemask))
7011 if (oldsg != sched_group_nodes[i])
7014 kfree(sched_group_nodes);
7015 sched_group_nodes_bycpu[cpu] = NULL;
7018 #else /* !CONFIG_NUMA */
7019 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7022 #endif /* CONFIG_NUMA */
7025 * Initialize sched groups cpu_power.
7027 * cpu_power indicates the capacity of sched group, which is used while
7028 * distributing the load between different sched groups in a sched domain.
7029 * Typically cpu_power for all the groups in a sched domain will be same unless
7030 * there are asymmetries in the topology. If there are asymmetries, group
7031 * having more cpu_power will pickup more load compared to the group having
7034 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7035 * the maximum number of tasks a group can handle in the presence of other idle
7036 * or lightly loaded groups in the same sched domain.
7038 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7040 struct sched_domain *child;
7041 struct sched_group *group;
7043 WARN_ON(!sd || !sd->groups);
7045 if (cpu != first_cpu(sd->groups->cpumask))
7050 sd->groups->__cpu_power = 0;
7053 * For perf policy, if the groups in child domain share resources
7054 * (for example cores sharing some portions of the cache hierarchy
7055 * or SMT), then set this domain groups cpu_power such that each group
7056 * can handle only one task, when there are other idle groups in the
7057 * same sched domain.
7059 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7061 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7062 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7067 * add cpu_power of each child group to this groups cpu_power
7069 group = child->groups;
7071 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7072 group = group->next;
7073 } while (group != child->groups);
7077 * Initializers for schedule domains
7078 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7081 #define SD_INIT(sd, type) sd_init_##type(sd)
7082 #define SD_INIT_FUNC(type) \
7083 static noinline void sd_init_##type(struct sched_domain *sd) \
7085 memset(sd, 0, sizeof(*sd)); \
7086 *sd = SD_##type##_INIT; \
7087 sd->level = SD_LV_##type; \
7092 SD_INIT_FUNC(ALLNODES)
7095 #ifdef CONFIG_SCHED_SMT
7096 SD_INIT_FUNC(SIBLING)
7098 #ifdef CONFIG_SCHED_MC
7103 * To minimize stack usage kmalloc room for cpumasks and share the
7104 * space as the usage in build_sched_domains() dictates. Used only
7105 * if the amount of space is significant.
7108 cpumask_t tmpmask; /* make this one first */
7111 cpumask_t this_sibling_map;
7112 cpumask_t this_core_map;
7114 cpumask_t send_covered;
7117 cpumask_t domainspan;
7119 cpumask_t notcovered;
7124 #define SCHED_CPUMASK_ALLOC 1
7125 #define SCHED_CPUMASK_FREE(v) kfree(v)
7126 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7128 #define SCHED_CPUMASK_ALLOC 0
7129 #define SCHED_CPUMASK_FREE(v)
7130 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7133 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7134 ((unsigned long)(a) + offsetof(struct allmasks, v))
7136 static int default_relax_domain_level = -1;
7138 static int __init setup_relax_domain_level(char *str)
7142 val = simple_strtoul(str, NULL, 0);
7143 if (val < SD_LV_MAX)
7144 default_relax_domain_level = val;
7148 __setup("relax_domain_level=", setup_relax_domain_level);
7150 static void set_domain_attribute(struct sched_domain *sd,
7151 struct sched_domain_attr *attr)
7155 if (!attr || attr->relax_domain_level < 0) {
7156 if (default_relax_domain_level < 0)
7159 request = default_relax_domain_level;
7161 request = attr->relax_domain_level;
7162 if (request < sd->level) {
7163 /* turn off idle balance on this domain */
7164 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7166 /* turn on idle balance on this domain */
7167 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7172 * Build sched domains for a given set of cpus and attach the sched domains
7173 * to the individual cpus
7175 static int __build_sched_domains(const cpumask_t *cpu_map,
7176 struct sched_domain_attr *attr)
7179 struct root_domain *rd;
7180 SCHED_CPUMASK_DECLARE(allmasks);
7183 struct sched_group **sched_group_nodes = NULL;
7184 int sd_allnodes = 0;
7187 * Allocate the per-node list of sched groups
7189 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7191 if (!sched_group_nodes) {
7192 printk(KERN_WARNING "Can not alloc sched group node list\n");
7197 rd = alloc_rootdomain();
7199 printk(KERN_WARNING "Cannot alloc root domain\n");
7201 kfree(sched_group_nodes);
7206 #if SCHED_CPUMASK_ALLOC
7207 /* get space for all scratch cpumask variables */
7208 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7210 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7213 kfree(sched_group_nodes);
7218 tmpmask = (cpumask_t *)allmasks;
7222 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7226 * Set up domains for cpus specified by the cpu_map.
7228 for_each_cpu_mask_nr(i, *cpu_map) {
7229 struct sched_domain *sd = NULL, *p;
7230 SCHED_CPUMASK_VAR(nodemask, allmasks);
7232 *nodemask = node_to_cpumask(cpu_to_node(i));
7233 cpus_and(*nodemask, *nodemask, *cpu_map);
7236 if (cpus_weight(*cpu_map) >
7237 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7238 sd = &per_cpu(allnodes_domains, i);
7239 SD_INIT(sd, ALLNODES);
7240 set_domain_attribute(sd, attr);
7241 sd->span = *cpu_map;
7242 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7248 sd = &per_cpu(node_domains, i);
7250 set_domain_attribute(sd, attr);
7251 sched_domain_node_span(cpu_to_node(i), &sd->span);
7255 cpus_and(sd->span, sd->span, *cpu_map);
7259 sd = &per_cpu(phys_domains, i);
7261 set_domain_attribute(sd, attr);
7262 sd->span = *nodemask;
7266 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7268 #ifdef CONFIG_SCHED_MC
7270 sd = &per_cpu(core_domains, i);
7272 set_domain_attribute(sd, attr);
7273 sd->span = cpu_coregroup_map(i);
7274 cpus_and(sd->span, sd->span, *cpu_map);
7277 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7280 #ifdef CONFIG_SCHED_SMT
7282 sd = &per_cpu(cpu_domains, i);
7283 SD_INIT(sd, SIBLING);
7284 set_domain_attribute(sd, attr);
7285 sd->span = per_cpu(cpu_sibling_map, i);
7286 cpus_and(sd->span, sd->span, *cpu_map);
7289 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7293 #ifdef CONFIG_SCHED_SMT
7294 /* Set up CPU (sibling) groups */
7295 for_each_cpu_mask_nr(i, *cpu_map) {
7296 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7297 SCHED_CPUMASK_VAR(send_covered, allmasks);
7299 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7300 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7301 if (i != first_cpu(*this_sibling_map))
7304 init_sched_build_groups(this_sibling_map, cpu_map,
7306 send_covered, tmpmask);
7310 #ifdef CONFIG_SCHED_MC
7311 /* Set up multi-core groups */
7312 for_each_cpu_mask_nr(i, *cpu_map) {
7313 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7314 SCHED_CPUMASK_VAR(send_covered, allmasks);
7316 *this_core_map = cpu_coregroup_map(i);
7317 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7318 if (i != first_cpu(*this_core_map))
7321 init_sched_build_groups(this_core_map, cpu_map,
7323 send_covered, tmpmask);
7327 /* Set up physical groups */
7328 for (i = 0; i < nr_node_ids; i++) {
7329 SCHED_CPUMASK_VAR(nodemask, allmasks);
7330 SCHED_CPUMASK_VAR(send_covered, allmasks);
7332 *nodemask = node_to_cpumask(i);
7333 cpus_and(*nodemask, *nodemask, *cpu_map);
7334 if (cpus_empty(*nodemask))
7337 init_sched_build_groups(nodemask, cpu_map,
7339 send_covered, tmpmask);
7343 /* Set up node groups */
7345 SCHED_CPUMASK_VAR(send_covered, allmasks);
7347 init_sched_build_groups(cpu_map, cpu_map,
7348 &cpu_to_allnodes_group,
7349 send_covered, tmpmask);
7352 for (i = 0; i < nr_node_ids; i++) {
7353 /* Set up node groups */
7354 struct sched_group *sg, *prev;
7355 SCHED_CPUMASK_VAR(nodemask, allmasks);
7356 SCHED_CPUMASK_VAR(domainspan, allmasks);
7357 SCHED_CPUMASK_VAR(covered, allmasks);
7360 *nodemask = node_to_cpumask(i);
7361 cpus_clear(*covered);
7363 cpus_and(*nodemask, *nodemask, *cpu_map);
7364 if (cpus_empty(*nodemask)) {
7365 sched_group_nodes[i] = NULL;
7369 sched_domain_node_span(i, domainspan);
7370 cpus_and(*domainspan, *domainspan, *cpu_map);
7372 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7374 printk(KERN_WARNING "Can not alloc domain group for "
7378 sched_group_nodes[i] = sg;
7379 for_each_cpu_mask_nr(j, *nodemask) {
7380 struct sched_domain *sd;
7382 sd = &per_cpu(node_domains, j);
7385 sg->__cpu_power = 0;
7386 sg->cpumask = *nodemask;
7388 cpus_or(*covered, *covered, *nodemask);
7391 for (j = 0; j < nr_node_ids; j++) {
7392 SCHED_CPUMASK_VAR(notcovered, allmasks);
7393 int n = (i + j) % nr_node_ids;
7394 node_to_cpumask_ptr(pnodemask, n);
7396 cpus_complement(*notcovered, *covered);
7397 cpus_and(*tmpmask, *notcovered, *cpu_map);
7398 cpus_and(*tmpmask, *tmpmask, *domainspan);
7399 if (cpus_empty(*tmpmask))
7402 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7403 if (cpus_empty(*tmpmask))
7406 sg = kmalloc_node(sizeof(struct sched_group),
7410 "Can not alloc domain group for node %d\n", j);
7413 sg->__cpu_power = 0;
7414 sg->cpumask = *tmpmask;
7415 sg->next = prev->next;
7416 cpus_or(*covered, *covered, *tmpmask);
7423 /* Calculate CPU power for physical packages and nodes */
7424 #ifdef CONFIG_SCHED_SMT
7425 for_each_cpu_mask_nr(i, *cpu_map) {
7426 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7428 init_sched_groups_power(i, sd);
7431 #ifdef CONFIG_SCHED_MC
7432 for_each_cpu_mask_nr(i, *cpu_map) {
7433 struct sched_domain *sd = &per_cpu(core_domains, i);
7435 init_sched_groups_power(i, sd);
7439 for_each_cpu_mask_nr(i, *cpu_map) {
7440 struct sched_domain *sd = &per_cpu(phys_domains, i);
7442 init_sched_groups_power(i, sd);
7446 for (i = 0; i < nr_node_ids; i++)
7447 init_numa_sched_groups_power(sched_group_nodes[i]);
7450 struct sched_group *sg;
7452 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7454 init_numa_sched_groups_power(sg);
7458 /* Attach the domains */
7459 for_each_cpu_mask_nr(i, *cpu_map) {
7460 struct sched_domain *sd;
7461 #ifdef CONFIG_SCHED_SMT
7462 sd = &per_cpu(cpu_domains, i);
7463 #elif defined(CONFIG_SCHED_MC)
7464 sd = &per_cpu(core_domains, i);
7466 sd = &per_cpu(phys_domains, i);
7468 cpu_attach_domain(sd, rd, i);
7471 SCHED_CPUMASK_FREE((void *)allmasks);
7476 free_sched_groups(cpu_map, tmpmask);
7477 SCHED_CPUMASK_FREE((void *)allmasks);
7482 static int build_sched_domains(const cpumask_t *cpu_map)
7484 return __build_sched_domains(cpu_map, NULL);
7487 static cpumask_t *doms_cur; /* current sched domains */
7488 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7489 static struct sched_domain_attr *dattr_cur;
7490 /* attribues of custom domains in 'doms_cur' */
7493 * Special case: If a kmalloc of a doms_cur partition (array of
7494 * cpumask_t) fails, then fallback to a single sched domain,
7495 * as determined by the single cpumask_t fallback_doms.
7497 static cpumask_t fallback_doms;
7499 void __attribute__((weak)) arch_update_cpu_topology(void)
7504 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7505 * For now this just excludes isolated cpus, but could be used to
7506 * exclude other special cases in the future.
7508 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7512 arch_update_cpu_topology();
7514 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7516 doms_cur = &fallback_doms;
7517 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7519 err = build_sched_domains(doms_cur);
7520 register_sched_domain_sysctl();
7525 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7528 free_sched_groups(cpu_map, tmpmask);
7532 * Detach sched domains from a group of cpus specified in cpu_map
7533 * These cpus will now be attached to the NULL domain
7535 static void detach_destroy_domains(const cpumask_t *cpu_map)
7540 unregister_sched_domain_sysctl();
7542 for_each_cpu_mask_nr(i, *cpu_map)
7543 cpu_attach_domain(NULL, &def_root_domain, i);
7544 synchronize_sched();
7545 arch_destroy_sched_domains(cpu_map, &tmpmask);
7548 /* handle null as "default" */
7549 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7550 struct sched_domain_attr *new, int idx_new)
7552 struct sched_domain_attr tmp;
7559 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7560 new ? (new + idx_new) : &tmp,
7561 sizeof(struct sched_domain_attr));
7565 * Partition sched domains as specified by the 'ndoms_new'
7566 * cpumasks in the array doms_new[] of cpumasks. This compares
7567 * doms_new[] to the current sched domain partitioning, doms_cur[].
7568 * It destroys each deleted domain and builds each new domain.
7570 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7571 * The masks don't intersect (don't overlap.) We should setup one
7572 * sched domain for each mask. CPUs not in any of the cpumasks will
7573 * not be load balanced. If the same cpumask appears both in the
7574 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7577 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7578 * ownership of it and will kfree it when done with it. If the caller
7579 * failed the kmalloc call, then it can pass in doms_new == NULL,
7580 * and partition_sched_domains() will fallback to the single partition
7581 * 'fallback_doms', it also forces the domains to be rebuilt.
7583 * Call with hotplug lock held
7585 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7586 struct sched_domain_attr *dattr_new)
7590 mutex_lock(&sched_domains_mutex);
7592 /* always unregister in case we don't destroy any domains */
7593 unregister_sched_domain_sysctl();
7595 if (doms_new == NULL)
7598 /* Destroy deleted domains */
7599 for (i = 0; i < ndoms_cur; i++) {
7600 for (j = 0; j < ndoms_new; j++) {
7601 if (cpus_equal(doms_cur[i], doms_new[j])
7602 && dattrs_equal(dattr_cur, i, dattr_new, j))
7605 /* no match - a current sched domain not in new doms_new[] */
7606 detach_destroy_domains(doms_cur + i);
7611 if (doms_new == NULL) {
7614 doms_new = &fallback_doms;
7615 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7619 /* Build new domains */
7620 for (i = 0; i < ndoms_new; i++) {
7621 for (j = 0; j < ndoms_cur; j++) {
7622 if (cpus_equal(doms_new[i], doms_cur[j])
7623 && dattrs_equal(dattr_new, i, dattr_cur, j))
7626 /* no match - add a new doms_new */
7627 __build_sched_domains(doms_new + i,
7628 dattr_new ? dattr_new + i : NULL);
7633 /* Remember the new sched domains */
7634 if (doms_cur != &fallback_doms)
7636 kfree(dattr_cur); /* kfree(NULL) is safe */
7637 doms_cur = doms_new;
7638 dattr_cur = dattr_new;
7639 ndoms_cur = ndoms_new;
7641 register_sched_domain_sysctl();
7643 mutex_unlock(&sched_domains_mutex);
7646 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7647 int arch_reinit_sched_domains(void)
7650 rebuild_sched_domains();
7655 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7659 if (buf[0] != '0' && buf[0] != '1')
7663 sched_smt_power_savings = (buf[0] == '1');
7665 sched_mc_power_savings = (buf[0] == '1');
7667 ret = arch_reinit_sched_domains();
7669 return ret ? ret : count;
7672 #ifdef CONFIG_SCHED_MC
7673 static ssize_t sched_mc_power_savings_show(struct sys_device *dev,
7674 struct sysdev_attribute *attr, char *page)
7676 return sprintf(page, "%u\n", sched_mc_power_savings);
7678 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7679 struct sysdev_attribute *attr,
7680 const char *buf, size_t count)
7682 return sched_power_savings_store(buf, count, 0);
7684 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7685 sched_mc_power_savings_store);
7688 #ifdef CONFIG_SCHED_SMT
7689 static ssize_t sched_smt_power_savings_show(struct sys_device *dev,
7690 struct sysdev_attribute *attr, char *page)
7692 return sprintf(page, "%u\n", sched_smt_power_savings);
7694 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7695 struct sysdev_attribute *attr,
7696 const char *buf, size_t count)
7698 return sched_power_savings_store(buf, count, 1);
7700 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7701 sched_smt_power_savings_store);
7704 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7708 #ifdef CONFIG_SCHED_SMT
7710 err = sysfs_create_file(&cls->kset.kobj,
7711 &attr_sched_smt_power_savings.attr);
7713 #ifdef CONFIG_SCHED_MC
7714 if (!err && mc_capable())
7715 err = sysfs_create_file(&cls->kset.kobj,
7716 &attr_sched_mc_power_savings.attr);
7720 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7722 #ifndef CONFIG_CPUSETS
7724 * Add online and remove offline CPUs from the scheduler domains.
7725 * When cpusets are enabled they take over this function.
7727 static int update_sched_domains(struct notifier_block *nfb,
7728 unsigned long action, void *hcpu)
7732 case CPU_ONLINE_FROZEN:
7734 case CPU_DEAD_FROZEN:
7735 partition_sched_domains(0, NULL, NULL);
7744 static int update_runtime(struct notifier_block *nfb,
7745 unsigned long action, void *hcpu)
7747 int cpu = (int)(long)hcpu;
7750 case CPU_DOWN_PREPARE:
7751 case CPU_DOWN_PREPARE_FROZEN:
7752 disable_runtime(cpu_rq(cpu));
7755 case CPU_DOWN_FAILED:
7756 case CPU_DOWN_FAILED_FROZEN:
7758 case CPU_ONLINE_FROZEN:
7759 enable_runtime(cpu_rq(cpu));
7767 void __init sched_init_smp(void)
7769 cpumask_t non_isolated_cpus;
7771 #if defined(CONFIG_NUMA)
7772 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7774 BUG_ON(sched_group_nodes_bycpu == NULL);
7777 mutex_lock(&sched_domains_mutex);
7778 arch_init_sched_domains(&cpu_online_map);
7779 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7780 if (cpus_empty(non_isolated_cpus))
7781 cpu_set(smp_processor_id(), non_isolated_cpus);
7782 mutex_unlock(&sched_domains_mutex);
7785 #ifndef CONFIG_CPUSETS
7786 /* XXX: Theoretical race here - CPU may be hotplugged now */
7787 hotcpu_notifier(update_sched_domains, 0);
7790 /* RT runtime code needs to handle some hotplug events */
7791 hotcpu_notifier(update_runtime, 0);
7795 /* Move init over to a non-isolated CPU */
7796 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7798 sched_init_granularity();
7801 void __init sched_init_smp(void)
7803 sched_init_granularity();
7805 #endif /* CONFIG_SMP */
7807 int in_sched_functions(unsigned long addr)
7809 return in_lock_functions(addr) ||
7810 (addr >= (unsigned long)__sched_text_start
7811 && addr < (unsigned long)__sched_text_end);
7814 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7816 cfs_rq->tasks_timeline = RB_ROOT;
7817 INIT_LIST_HEAD(&cfs_rq->tasks);
7818 #ifdef CONFIG_FAIR_GROUP_SCHED
7821 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7824 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7826 struct rt_prio_array *array;
7829 array = &rt_rq->active;
7830 for (i = 0; i < MAX_RT_PRIO; i++) {
7831 INIT_LIST_HEAD(array->queue + i);
7832 __clear_bit(i, array->bitmap);
7834 /* delimiter for bitsearch: */
7835 __set_bit(MAX_RT_PRIO, array->bitmap);
7837 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7838 rt_rq->highest_prio = MAX_RT_PRIO;
7841 rt_rq->rt_nr_migratory = 0;
7842 rt_rq->overloaded = 0;
7846 rt_rq->rt_throttled = 0;
7847 rt_rq->rt_runtime = 0;
7848 spin_lock_init(&rt_rq->rt_runtime_lock);
7850 #ifdef CONFIG_RT_GROUP_SCHED
7851 rt_rq->rt_nr_boosted = 0;
7856 #ifdef CONFIG_FAIR_GROUP_SCHED
7857 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7858 struct sched_entity *se, int cpu, int add,
7859 struct sched_entity *parent)
7861 struct rq *rq = cpu_rq(cpu);
7862 tg->cfs_rq[cpu] = cfs_rq;
7863 init_cfs_rq(cfs_rq, rq);
7866 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7869 /* se could be NULL for init_task_group */
7874 se->cfs_rq = &rq->cfs;
7876 se->cfs_rq = parent->my_q;
7879 se->load.weight = tg->shares;
7880 se->load.inv_weight = 0;
7881 se->parent = parent;
7885 #ifdef CONFIG_RT_GROUP_SCHED
7886 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7887 struct sched_rt_entity *rt_se, int cpu, int add,
7888 struct sched_rt_entity *parent)
7890 struct rq *rq = cpu_rq(cpu);
7892 tg->rt_rq[cpu] = rt_rq;
7893 init_rt_rq(rt_rq, rq);
7895 rt_rq->rt_se = rt_se;
7896 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7898 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7900 tg->rt_se[cpu] = rt_se;
7905 rt_se->rt_rq = &rq->rt;
7907 rt_se->rt_rq = parent->my_q;
7909 rt_se->my_q = rt_rq;
7910 rt_se->parent = parent;
7911 INIT_LIST_HEAD(&rt_se->run_list);
7915 void __init sched_init(void)
7918 unsigned long alloc_size = 0, ptr;
7920 #ifdef CONFIG_FAIR_GROUP_SCHED
7921 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7923 #ifdef CONFIG_RT_GROUP_SCHED
7924 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7926 #ifdef CONFIG_USER_SCHED
7930 * As sched_init() is called before page_alloc is setup,
7931 * we use alloc_bootmem().
7934 ptr = (unsigned long)alloc_bootmem(alloc_size);
7936 #ifdef CONFIG_FAIR_GROUP_SCHED
7937 init_task_group.se = (struct sched_entity **)ptr;
7938 ptr += nr_cpu_ids * sizeof(void **);
7940 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7941 ptr += nr_cpu_ids * sizeof(void **);
7943 #ifdef CONFIG_USER_SCHED
7944 root_task_group.se = (struct sched_entity **)ptr;
7945 ptr += nr_cpu_ids * sizeof(void **);
7947 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7948 ptr += nr_cpu_ids * sizeof(void **);
7949 #endif /* CONFIG_USER_SCHED */
7950 #endif /* CONFIG_FAIR_GROUP_SCHED */
7951 #ifdef CONFIG_RT_GROUP_SCHED
7952 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7953 ptr += nr_cpu_ids * sizeof(void **);
7955 init_task_group.rt_rq = (struct rt_rq **)ptr;
7956 ptr += nr_cpu_ids * sizeof(void **);
7958 #ifdef CONFIG_USER_SCHED
7959 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7960 ptr += nr_cpu_ids * sizeof(void **);
7962 root_task_group.rt_rq = (struct rt_rq **)ptr;
7963 ptr += nr_cpu_ids * sizeof(void **);
7964 #endif /* CONFIG_USER_SCHED */
7965 #endif /* CONFIG_RT_GROUP_SCHED */
7969 init_defrootdomain();
7972 init_rt_bandwidth(&def_rt_bandwidth,
7973 global_rt_period(), global_rt_runtime());
7975 #ifdef CONFIG_RT_GROUP_SCHED
7976 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7977 global_rt_period(), global_rt_runtime());
7978 #ifdef CONFIG_USER_SCHED
7979 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7980 global_rt_period(), RUNTIME_INF);
7981 #endif /* CONFIG_USER_SCHED */
7982 #endif /* CONFIG_RT_GROUP_SCHED */
7984 #ifdef CONFIG_GROUP_SCHED
7985 list_add(&init_task_group.list, &task_groups);
7986 INIT_LIST_HEAD(&init_task_group.children);
7988 #ifdef CONFIG_USER_SCHED
7989 INIT_LIST_HEAD(&root_task_group.children);
7990 init_task_group.parent = &root_task_group;
7991 list_add(&init_task_group.siblings, &root_task_group.children);
7992 #endif /* CONFIG_USER_SCHED */
7993 #endif /* CONFIG_GROUP_SCHED */
7995 for_each_possible_cpu(i) {
7999 spin_lock_init(&rq->lock);
8001 init_cfs_rq(&rq->cfs, rq);
8002 init_rt_rq(&rq->rt, rq);
8003 #ifdef CONFIG_FAIR_GROUP_SCHED
8004 init_task_group.shares = init_task_group_load;
8005 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8006 #ifdef CONFIG_CGROUP_SCHED
8008 * How much cpu bandwidth does init_task_group get?
8010 * In case of task-groups formed thr' the cgroup filesystem, it
8011 * gets 100% of the cpu resources in the system. This overall
8012 * system cpu resource is divided among the tasks of
8013 * init_task_group and its child task-groups in a fair manner,
8014 * based on each entity's (task or task-group's) weight
8015 * (se->load.weight).
8017 * In other words, if init_task_group has 10 tasks of weight
8018 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8019 * then A0's share of the cpu resource is:
8021 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8023 * We achieve this by letting init_task_group's tasks sit
8024 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8026 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8027 #elif defined CONFIG_USER_SCHED
8028 root_task_group.shares = NICE_0_LOAD;
8029 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8031 * In case of task-groups formed thr' the user id of tasks,
8032 * init_task_group represents tasks belonging to root user.
8033 * Hence it forms a sibling of all subsequent groups formed.
8034 * In this case, init_task_group gets only a fraction of overall
8035 * system cpu resource, based on the weight assigned to root
8036 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8037 * by letting tasks of init_task_group sit in a separate cfs_rq
8038 * (init_cfs_rq) and having one entity represent this group of
8039 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8041 init_tg_cfs_entry(&init_task_group,
8042 &per_cpu(init_cfs_rq, i),
8043 &per_cpu(init_sched_entity, i), i, 1,
8044 root_task_group.se[i]);
8047 #endif /* CONFIG_FAIR_GROUP_SCHED */
8049 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8050 #ifdef CONFIG_RT_GROUP_SCHED
8051 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8052 #ifdef CONFIG_CGROUP_SCHED
8053 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8054 #elif defined CONFIG_USER_SCHED
8055 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8056 init_tg_rt_entry(&init_task_group,
8057 &per_cpu(init_rt_rq, i),
8058 &per_cpu(init_sched_rt_entity, i), i, 1,
8059 root_task_group.rt_se[i]);
8063 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8064 rq->cpu_load[j] = 0;
8068 rq->active_balance = 0;
8069 rq->next_balance = jiffies;
8073 rq->migration_thread = NULL;
8074 INIT_LIST_HEAD(&rq->migration_queue);
8075 rq_attach_root(rq, &def_root_domain);
8078 atomic_set(&rq->nr_iowait, 0);
8081 set_load_weight(&init_task);
8083 #ifdef CONFIG_PREEMPT_NOTIFIERS
8084 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8088 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8091 #ifdef CONFIG_RT_MUTEXES
8092 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8096 * The boot idle thread does lazy MMU switching as well:
8098 atomic_inc(&init_mm.mm_count);
8099 enter_lazy_tlb(&init_mm, current);
8102 * Make us the idle thread. Technically, schedule() should not be
8103 * called from this thread, however somewhere below it might be,
8104 * but because we are the idle thread, we just pick up running again
8105 * when this runqueue becomes "idle".
8107 init_idle(current, smp_processor_id());
8109 * During early bootup we pretend to be a normal task:
8111 current->sched_class = &fair_sched_class;
8113 scheduler_running = 1;
8116 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8117 void __might_sleep(char *file, int line)
8120 static unsigned long prev_jiffy; /* ratelimiting */
8122 if ((in_atomic() || irqs_disabled()) &&
8123 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8124 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8126 prev_jiffy = jiffies;
8127 printk(KERN_ERR "BUG: sleeping function called from invalid"
8128 " context at %s:%d\n", file, line);
8129 printk("in_atomic():%d, irqs_disabled():%d\n",
8130 in_atomic(), irqs_disabled());
8131 debug_show_held_locks(current);
8132 if (irqs_disabled())
8133 print_irqtrace_events(current);
8138 EXPORT_SYMBOL(__might_sleep);
8141 #ifdef CONFIG_MAGIC_SYSRQ
8142 static void normalize_task(struct rq *rq, struct task_struct *p)
8146 update_rq_clock(rq);
8147 on_rq = p->se.on_rq;
8149 deactivate_task(rq, p, 0);
8150 __setscheduler(rq, p, SCHED_NORMAL, 0);
8152 activate_task(rq, p, 0);
8153 resched_task(rq->curr);
8157 void normalize_rt_tasks(void)
8159 struct task_struct *g, *p;
8160 unsigned long flags;
8163 read_lock_irqsave(&tasklist_lock, flags);
8164 do_each_thread(g, p) {
8166 * Only normalize user tasks:
8171 p->se.exec_start = 0;
8172 #ifdef CONFIG_SCHEDSTATS
8173 p->se.wait_start = 0;
8174 p->se.sleep_start = 0;
8175 p->se.block_start = 0;
8180 * Renice negative nice level userspace
8183 if (TASK_NICE(p) < 0 && p->mm)
8184 set_user_nice(p, 0);
8188 spin_lock(&p->pi_lock);
8189 rq = __task_rq_lock(p);
8191 normalize_task(rq, p);
8193 __task_rq_unlock(rq);
8194 spin_unlock(&p->pi_lock);
8195 } while_each_thread(g, p);
8197 read_unlock_irqrestore(&tasklist_lock, flags);
8200 #endif /* CONFIG_MAGIC_SYSRQ */
8204 * These functions are only useful for the IA64 MCA handling.
8206 * They can only be called when the whole system has been
8207 * stopped - every CPU needs to be quiescent, and no scheduling
8208 * activity can take place. Using them for anything else would
8209 * be a serious bug, and as a result, they aren't even visible
8210 * under any other configuration.
8214 * curr_task - return the current task for a given cpu.
8215 * @cpu: the processor in question.
8217 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8219 struct task_struct *curr_task(int cpu)
8221 return cpu_curr(cpu);
8225 * set_curr_task - set the current task for a given cpu.
8226 * @cpu: the processor in question.
8227 * @p: the task pointer to set.
8229 * Description: This function must only be used when non-maskable interrupts
8230 * are serviced on a separate stack. It allows the architecture to switch the
8231 * notion of the current task on a cpu in a non-blocking manner. This function
8232 * must be called with all CPU's synchronized, and interrupts disabled, the
8233 * and caller must save the original value of the current task (see
8234 * curr_task() above) and restore that value before reenabling interrupts and
8235 * re-starting the system.
8237 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8239 void set_curr_task(int cpu, struct task_struct *p)
8246 #ifdef CONFIG_FAIR_GROUP_SCHED
8247 static void free_fair_sched_group(struct task_group *tg)
8251 for_each_possible_cpu(i) {
8253 kfree(tg->cfs_rq[i]);
8263 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8265 struct cfs_rq *cfs_rq;
8266 struct sched_entity *se, *parent_se;
8270 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8273 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8277 tg->shares = NICE_0_LOAD;
8279 for_each_possible_cpu(i) {
8282 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8283 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8287 se = kmalloc_node(sizeof(struct sched_entity),
8288 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8292 parent_se = parent ? parent->se[i] : NULL;
8293 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8302 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8304 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8305 &cpu_rq(cpu)->leaf_cfs_rq_list);
8308 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8310 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8312 #else /* !CONFG_FAIR_GROUP_SCHED */
8313 static inline void free_fair_sched_group(struct task_group *tg)
8318 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8323 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8327 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8330 #endif /* CONFIG_FAIR_GROUP_SCHED */
8332 #ifdef CONFIG_RT_GROUP_SCHED
8333 static void free_rt_sched_group(struct task_group *tg)
8337 destroy_rt_bandwidth(&tg->rt_bandwidth);
8339 for_each_possible_cpu(i) {
8341 kfree(tg->rt_rq[i]);
8343 kfree(tg->rt_se[i]);
8351 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8353 struct rt_rq *rt_rq;
8354 struct sched_rt_entity *rt_se, *parent_se;
8358 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8361 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8365 init_rt_bandwidth(&tg->rt_bandwidth,
8366 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8368 for_each_possible_cpu(i) {
8371 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8372 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8376 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8377 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8381 parent_se = parent ? parent->rt_se[i] : NULL;
8382 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8391 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8393 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8394 &cpu_rq(cpu)->leaf_rt_rq_list);
8397 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8399 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8401 #else /* !CONFIG_RT_GROUP_SCHED */
8402 static inline void free_rt_sched_group(struct task_group *tg)
8407 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8412 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8416 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8419 #endif /* CONFIG_RT_GROUP_SCHED */
8421 #ifdef CONFIG_GROUP_SCHED
8422 static void free_sched_group(struct task_group *tg)
8424 free_fair_sched_group(tg);
8425 free_rt_sched_group(tg);
8429 /* allocate runqueue etc for a new task group */
8430 struct task_group *sched_create_group(struct task_group *parent)
8432 struct task_group *tg;
8433 unsigned long flags;
8436 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8438 return ERR_PTR(-ENOMEM);
8440 if (!alloc_fair_sched_group(tg, parent))
8443 if (!alloc_rt_sched_group(tg, parent))
8446 spin_lock_irqsave(&task_group_lock, flags);
8447 for_each_possible_cpu(i) {
8448 register_fair_sched_group(tg, i);
8449 register_rt_sched_group(tg, i);
8451 list_add_rcu(&tg->list, &task_groups);
8453 WARN_ON(!parent); /* root should already exist */
8455 tg->parent = parent;
8456 list_add_rcu(&tg->siblings, &parent->children);
8457 INIT_LIST_HEAD(&tg->children);
8458 spin_unlock_irqrestore(&task_group_lock, flags);
8463 free_sched_group(tg);
8464 return ERR_PTR(-ENOMEM);
8467 /* rcu callback to free various structures associated with a task group */
8468 static void free_sched_group_rcu(struct rcu_head *rhp)
8470 /* now it should be safe to free those cfs_rqs */
8471 free_sched_group(container_of(rhp, struct task_group, rcu));
8474 /* Destroy runqueue etc associated with a task group */
8475 void sched_destroy_group(struct task_group *tg)
8477 unsigned long flags;
8480 spin_lock_irqsave(&task_group_lock, flags);
8481 for_each_possible_cpu(i) {
8482 unregister_fair_sched_group(tg, i);
8483 unregister_rt_sched_group(tg, i);
8485 list_del_rcu(&tg->list);
8486 list_del_rcu(&tg->siblings);
8487 spin_unlock_irqrestore(&task_group_lock, flags);
8489 /* wait for possible concurrent references to cfs_rqs complete */
8490 call_rcu(&tg->rcu, free_sched_group_rcu);
8493 /* change task's runqueue when it moves between groups.
8494 * The caller of this function should have put the task in its new group
8495 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8496 * reflect its new group.
8498 void sched_move_task(struct task_struct *tsk)
8501 unsigned long flags;
8504 rq = task_rq_lock(tsk, &flags);
8506 update_rq_clock(rq);
8508 running = task_current(rq, tsk);
8509 on_rq = tsk->se.on_rq;
8512 dequeue_task(rq, tsk, 0);
8513 if (unlikely(running))
8514 tsk->sched_class->put_prev_task(rq, tsk);
8516 set_task_rq(tsk, task_cpu(tsk));
8518 #ifdef CONFIG_FAIR_GROUP_SCHED
8519 if (tsk->sched_class->moved_group)
8520 tsk->sched_class->moved_group(tsk);
8523 if (unlikely(running))
8524 tsk->sched_class->set_curr_task(rq);
8526 enqueue_task(rq, tsk, 0);
8528 task_rq_unlock(rq, &flags);
8530 #endif /* CONFIG_GROUP_SCHED */
8532 #ifdef CONFIG_FAIR_GROUP_SCHED
8533 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8535 struct cfs_rq *cfs_rq = se->cfs_rq;
8540 dequeue_entity(cfs_rq, se, 0);
8542 se->load.weight = shares;
8543 se->load.inv_weight = 0;
8546 enqueue_entity(cfs_rq, se, 0);
8549 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8551 struct cfs_rq *cfs_rq = se->cfs_rq;
8552 struct rq *rq = cfs_rq->rq;
8553 unsigned long flags;
8555 spin_lock_irqsave(&rq->lock, flags);
8556 __set_se_shares(se, shares);
8557 spin_unlock_irqrestore(&rq->lock, flags);
8560 static DEFINE_MUTEX(shares_mutex);
8562 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8565 unsigned long flags;
8568 * We can't change the weight of the root cgroup.
8573 if (shares < MIN_SHARES)
8574 shares = MIN_SHARES;
8575 else if (shares > MAX_SHARES)
8576 shares = MAX_SHARES;
8578 mutex_lock(&shares_mutex);
8579 if (tg->shares == shares)
8582 spin_lock_irqsave(&task_group_lock, flags);
8583 for_each_possible_cpu(i)
8584 unregister_fair_sched_group(tg, i);
8585 list_del_rcu(&tg->siblings);
8586 spin_unlock_irqrestore(&task_group_lock, flags);
8588 /* wait for any ongoing reference to this group to finish */
8589 synchronize_sched();
8592 * Now we are free to modify the group's share on each cpu
8593 * w/o tripping rebalance_share or load_balance_fair.
8595 tg->shares = shares;
8596 for_each_possible_cpu(i) {
8600 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8601 set_se_shares(tg->se[i], shares);
8605 * Enable load balance activity on this group, by inserting it back on
8606 * each cpu's rq->leaf_cfs_rq_list.
8608 spin_lock_irqsave(&task_group_lock, flags);
8609 for_each_possible_cpu(i)
8610 register_fair_sched_group(tg, i);
8611 list_add_rcu(&tg->siblings, &tg->parent->children);
8612 spin_unlock_irqrestore(&task_group_lock, flags);
8614 mutex_unlock(&shares_mutex);
8618 unsigned long sched_group_shares(struct task_group *tg)
8624 #ifdef CONFIG_RT_GROUP_SCHED
8626 * Ensure that the real time constraints are schedulable.
8628 static DEFINE_MUTEX(rt_constraints_mutex);
8630 static unsigned long to_ratio(u64 period, u64 runtime)
8632 if (runtime == RUNTIME_INF)
8635 return div64_u64(runtime << 16, period);
8638 #ifdef CONFIG_CGROUP_SCHED
8639 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8641 struct task_group *tgi, *parent = tg->parent;
8642 unsigned long total = 0;
8645 if (global_rt_period() < period)
8648 return to_ratio(period, runtime) <
8649 to_ratio(global_rt_period(), global_rt_runtime());
8652 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8656 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8660 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8661 tgi->rt_bandwidth.rt_runtime);
8665 return total + to_ratio(period, runtime) <=
8666 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8667 parent->rt_bandwidth.rt_runtime);
8669 #elif defined CONFIG_USER_SCHED
8670 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8672 struct task_group *tgi;
8673 unsigned long total = 0;
8674 unsigned long global_ratio =
8675 to_ratio(global_rt_period(), global_rt_runtime());
8678 list_for_each_entry_rcu(tgi, &task_groups, list) {
8682 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8683 tgi->rt_bandwidth.rt_runtime);
8687 return total + to_ratio(period, runtime) < global_ratio;
8691 /* Must be called with tasklist_lock held */
8692 static inline int tg_has_rt_tasks(struct task_group *tg)
8694 struct task_struct *g, *p;
8695 do_each_thread(g, p) {
8696 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8698 } while_each_thread(g, p);
8702 static int tg_set_bandwidth(struct task_group *tg,
8703 u64 rt_period, u64 rt_runtime)
8707 mutex_lock(&rt_constraints_mutex);
8708 read_lock(&tasklist_lock);
8709 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8713 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8718 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8719 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8720 tg->rt_bandwidth.rt_runtime = rt_runtime;
8722 for_each_possible_cpu(i) {
8723 struct rt_rq *rt_rq = tg->rt_rq[i];
8725 spin_lock(&rt_rq->rt_runtime_lock);
8726 rt_rq->rt_runtime = rt_runtime;
8727 spin_unlock(&rt_rq->rt_runtime_lock);
8729 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8731 read_unlock(&tasklist_lock);
8732 mutex_unlock(&rt_constraints_mutex);
8737 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8739 u64 rt_runtime, rt_period;
8741 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8742 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8743 if (rt_runtime_us < 0)
8744 rt_runtime = RUNTIME_INF;
8746 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8749 long sched_group_rt_runtime(struct task_group *tg)
8753 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8756 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8757 do_div(rt_runtime_us, NSEC_PER_USEC);
8758 return rt_runtime_us;
8761 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8763 u64 rt_runtime, rt_period;
8765 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8766 rt_runtime = tg->rt_bandwidth.rt_runtime;
8771 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8774 long sched_group_rt_period(struct task_group *tg)
8778 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8779 do_div(rt_period_us, NSEC_PER_USEC);
8780 return rt_period_us;
8783 static int sched_rt_global_constraints(void)
8785 struct task_group *tg = &root_task_group;
8786 u64 rt_runtime, rt_period;
8789 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8790 rt_runtime = tg->rt_bandwidth.rt_runtime;
8792 mutex_lock(&rt_constraints_mutex);
8793 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8795 mutex_unlock(&rt_constraints_mutex);
8799 #else /* !CONFIG_RT_GROUP_SCHED */
8800 static int sched_rt_global_constraints(void)
8802 unsigned long flags;
8805 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8806 for_each_possible_cpu(i) {
8807 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8809 spin_lock(&rt_rq->rt_runtime_lock);
8810 rt_rq->rt_runtime = global_rt_runtime();
8811 spin_unlock(&rt_rq->rt_runtime_lock);
8813 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8817 #endif /* CONFIG_RT_GROUP_SCHED */
8819 int sched_rt_handler(struct ctl_table *table, int write,
8820 struct file *filp, void __user *buffer, size_t *lenp,
8824 int old_period, old_runtime;
8825 static DEFINE_MUTEX(mutex);
8828 old_period = sysctl_sched_rt_period;
8829 old_runtime = sysctl_sched_rt_runtime;
8831 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8833 if (!ret && write) {
8834 ret = sched_rt_global_constraints();
8836 sysctl_sched_rt_period = old_period;
8837 sysctl_sched_rt_runtime = old_runtime;
8839 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8840 def_rt_bandwidth.rt_period =
8841 ns_to_ktime(global_rt_period());
8844 mutex_unlock(&mutex);
8849 #ifdef CONFIG_CGROUP_SCHED
8851 /* return corresponding task_group object of a cgroup */
8852 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8854 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8855 struct task_group, css);
8858 static struct cgroup_subsys_state *
8859 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8861 struct task_group *tg, *parent;
8863 if (!cgrp->parent) {
8864 /* This is early initialization for the top cgroup */
8865 init_task_group.css.cgroup = cgrp;
8866 return &init_task_group.css;
8869 parent = cgroup_tg(cgrp->parent);
8870 tg = sched_create_group(parent);
8872 return ERR_PTR(-ENOMEM);
8874 /* Bind the cgroup to task_group object we just created */
8875 tg->css.cgroup = cgrp;
8881 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8883 struct task_group *tg = cgroup_tg(cgrp);
8885 sched_destroy_group(tg);
8889 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8890 struct task_struct *tsk)
8892 #ifdef CONFIG_RT_GROUP_SCHED
8893 /* Don't accept realtime tasks when there is no way for them to run */
8894 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8897 /* We don't support RT-tasks being in separate groups */
8898 if (tsk->sched_class != &fair_sched_class)
8906 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8907 struct cgroup *old_cont, struct task_struct *tsk)
8909 sched_move_task(tsk);
8912 #ifdef CONFIG_FAIR_GROUP_SCHED
8913 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8916 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8919 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8921 struct task_group *tg = cgroup_tg(cgrp);
8923 return (u64) tg->shares;
8925 #endif /* CONFIG_FAIR_GROUP_SCHED */
8927 #ifdef CONFIG_RT_GROUP_SCHED
8928 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8931 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8934 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8936 return sched_group_rt_runtime(cgroup_tg(cgrp));
8939 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8942 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8945 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8947 return sched_group_rt_period(cgroup_tg(cgrp));
8949 #endif /* CONFIG_RT_GROUP_SCHED */
8951 static struct cftype cpu_files[] = {
8952 #ifdef CONFIG_FAIR_GROUP_SCHED
8955 .read_u64 = cpu_shares_read_u64,
8956 .write_u64 = cpu_shares_write_u64,
8959 #ifdef CONFIG_RT_GROUP_SCHED
8961 .name = "rt_runtime_us",
8962 .read_s64 = cpu_rt_runtime_read,
8963 .write_s64 = cpu_rt_runtime_write,
8966 .name = "rt_period_us",
8967 .read_u64 = cpu_rt_period_read_uint,
8968 .write_u64 = cpu_rt_period_write_uint,
8973 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8975 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8978 struct cgroup_subsys cpu_cgroup_subsys = {
8980 .create = cpu_cgroup_create,
8981 .destroy = cpu_cgroup_destroy,
8982 .can_attach = cpu_cgroup_can_attach,
8983 .attach = cpu_cgroup_attach,
8984 .populate = cpu_cgroup_populate,
8985 .subsys_id = cpu_cgroup_subsys_id,
8989 #endif /* CONFIG_CGROUP_SCHED */
8991 #ifdef CONFIG_CGROUP_CPUACCT
8994 * CPU accounting code for task groups.
8996 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8997 * (balbir@in.ibm.com).
9000 /* track cpu usage of a group of tasks */
9002 struct cgroup_subsys_state css;
9003 /* cpuusage holds pointer to a u64-type object on every cpu */
9007 struct cgroup_subsys cpuacct_subsys;
9009 /* return cpu accounting group corresponding to this container */
9010 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9012 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9013 struct cpuacct, css);
9016 /* return cpu accounting group to which this task belongs */
9017 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9019 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9020 struct cpuacct, css);
9023 /* create a new cpu accounting group */
9024 static struct cgroup_subsys_state *cpuacct_create(
9025 struct cgroup_subsys *ss, struct cgroup *cgrp)
9027 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9030 return ERR_PTR(-ENOMEM);
9032 ca->cpuusage = alloc_percpu(u64);
9033 if (!ca->cpuusage) {
9035 return ERR_PTR(-ENOMEM);
9041 /* destroy an existing cpu accounting group */
9043 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9045 struct cpuacct *ca = cgroup_ca(cgrp);
9047 free_percpu(ca->cpuusage);
9051 /* return total cpu usage (in nanoseconds) of a group */
9052 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9054 struct cpuacct *ca = cgroup_ca(cgrp);
9055 u64 totalcpuusage = 0;
9058 for_each_possible_cpu(i) {
9059 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9062 * Take rq->lock to make 64-bit addition safe on 32-bit
9065 spin_lock_irq(&cpu_rq(i)->lock);
9066 totalcpuusage += *cpuusage;
9067 spin_unlock_irq(&cpu_rq(i)->lock);
9070 return totalcpuusage;
9073 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9076 struct cpuacct *ca = cgroup_ca(cgrp);
9085 for_each_possible_cpu(i) {
9086 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9088 spin_lock_irq(&cpu_rq(i)->lock);
9090 spin_unlock_irq(&cpu_rq(i)->lock);
9096 static struct cftype files[] = {
9099 .read_u64 = cpuusage_read,
9100 .write_u64 = cpuusage_write,
9104 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9106 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9110 * charge this task's execution time to its accounting group.
9112 * called with rq->lock held.
9114 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9118 if (!cpuacct_subsys.active)
9123 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9125 *cpuusage += cputime;
9129 struct cgroup_subsys cpuacct_subsys = {
9131 .create = cpuacct_create,
9132 .destroy = cpuacct_destroy,
9133 .populate = cpuacct_populate,
9134 .subsys_id = cpuacct_subsys_id,
9136 #endif /* CONFIG_CGROUP_CPUACCT */