4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
81 * Convert user-nice values [ -20 ... 0 ... 19 ]
82 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
86 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
87 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90 * 'User priority' is the nice value converted to something we
91 * can work with better when scaling various scheduler parameters,
92 * it's a [ 0 ... 39 ] range.
94 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
95 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
96 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99 * Helpers for converting nanosecond timing to jiffy resolution
101 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
103 #define NICE_0_LOAD SCHED_LOAD_SCALE
104 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
107 * These are the 'tuning knobs' of the scheduler:
109 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
110 * Timeslices get refilled after they expire.
112 #define DEF_TIMESLICE (100 * HZ / 1000)
115 * single value that denotes runtime == period, ie unlimited time.
117 #define RUNTIME_INF ((u64)~0ULL)
121 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
122 * Since cpu_power is a 'constant', we can use a reciprocal divide.
124 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
126 return reciprocal_divide(load, sg->reciprocal_cpu_power);
130 * Each time a sched group cpu_power is changed,
131 * we must compute its reciprocal value
133 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
135 sg->__cpu_power += val;
136 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
140 static inline int rt_policy(int policy)
142 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
147 static inline int task_has_rt_policy(struct task_struct *p)
149 return rt_policy(p->policy);
153 * This is the priority-queue data structure of the RT scheduling class:
155 struct rt_prio_array {
156 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
157 struct list_head queue[MAX_RT_PRIO];
160 struct rt_bandwidth {
161 /* nests inside the rq lock: */
162 spinlock_t rt_runtime_lock;
165 struct hrtimer rt_period_timer;
168 static struct rt_bandwidth def_rt_bandwidth;
170 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
172 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
174 struct rt_bandwidth *rt_b =
175 container_of(timer, struct rt_bandwidth, rt_period_timer);
181 now = hrtimer_cb_get_time(timer);
182 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
187 idle = do_sched_rt_period_timer(rt_b, overrun);
190 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
194 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
196 rt_b->rt_period = ns_to_ktime(period);
197 rt_b->rt_runtime = runtime;
199 spin_lock_init(&rt_b->rt_runtime_lock);
201 hrtimer_init(&rt_b->rt_period_timer,
202 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
203 rt_b->rt_period_timer.function = sched_rt_period_timer;
204 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
207 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
211 if (rt_b->rt_runtime == RUNTIME_INF)
214 if (hrtimer_active(&rt_b->rt_period_timer))
217 spin_lock(&rt_b->rt_runtime_lock);
219 if (hrtimer_active(&rt_b->rt_period_timer))
222 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
223 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
224 hrtimer_start(&rt_b->rt_period_timer,
225 rt_b->rt_period_timer.expires,
228 spin_unlock(&rt_b->rt_runtime_lock);
231 #ifdef CONFIG_RT_GROUP_SCHED
232 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
234 hrtimer_cancel(&rt_b->rt_period_timer);
239 * sched_domains_mutex serializes calls to arch_init_sched_domains,
240 * detach_destroy_domains and partition_sched_domains.
242 static DEFINE_MUTEX(sched_domains_mutex);
244 #ifdef CONFIG_GROUP_SCHED
246 #include <linux/cgroup.h>
250 static LIST_HEAD(task_groups);
252 /* task group related information */
254 #ifdef CONFIG_CGROUP_SCHED
255 struct cgroup_subsys_state css;
258 #ifdef CONFIG_FAIR_GROUP_SCHED
259 /* schedulable entities of this group on each cpu */
260 struct sched_entity **se;
261 /* runqueue "owned" by this group on each cpu */
262 struct cfs_rq **cfs_rq;
263 unsigned long shares;
266 #ifdef CONFIG_RT_GROUP_SCHED
267 struct sched_rt_entity **rt_se;
268 struct rt_rq **rt_rq;
270 struct rt_bandwidth rt_bandwidth;
274 struct list_head list;
276 struct task_group *parent;
277 struct list_head siblings;
278 struct list_head children;
281 #ifdef CONFIG_USER_SCHED
285 * Every UID task group (including init_task_group aka UID-0) will
286 * be a child to this group.
288 struct task_group root_task_group;
290 #ifdef CONFIG_FAIR_GROUP_SCHED
291 /* Default task group's sched entity on each cpu */
292 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
293 /* Default task group's cfs_rq on each cpu */
294 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
295 #endif /* CONFIG_FAIR_GROUP_SCHED */
297 #ifdef CONFIG_RT_GROUP_SCHED
298 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
299 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
300 #endif /* CONFIG_RT_GROUP_SCHED */
301 #else /* !CONFIG_FAIR_GROUP_SCHED */
302 #define root_task_group init_task_group
303 #endif /* CONFIG_FAIR_GROUP_SCHED */
305 /* task_group_lock serializes add/remove of task groups and also changes to
306 * a task group's cpu shares.
308 static DEFINE_SPINLOCK(task_group_lock);
310 #ifdef CONFIG_FAIR_GROUP_SCHED
311 #ifdef CONFIG_USER_SCHED
312 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
313 #else /* !CONFIG_USER_SCHED */
314 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
315 #endif /* CONFIG_USER_SCHED */
318 * A weight of 0 or 1 can cause arithmetics problems.
319 * A weight of a cfs_rq is the sum of weights of which entities
320 * are queued on this cfs_rq, so a weight of a entity should not be
321 * too large, so as the shares value of a task group.
322 * (The default weight is 1024 - so there's no practical
323 * limitation from this.)
326 #define MAX_SHARES (1UL << 18)
328 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
331 /* Default task group.
332 * Every task in system belong to this group at bootup.
334 struct task_group init_task_group;
336 /* return group to which a task belongs */
337 static inline struct task_group *task_group(struct task_struct *p)
339 struct task_group *tg;
341 #ifdef CONFIG_USER_SCHED
343 #elif defined(CONFIG_CGROUP_SCHED)
344 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
345 struct task_group, css);
347 tg = &init_task_group;
352 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
353 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
355 #ifdef CONFIG_FAIR_GROUP_SCHED
356 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
357 p->se.parent = task_group(p)->se[cpu];
360 #ifdef CONFIG_RT_GROUP_SCHED
361 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
362 p->rt.parent = task_group(p)->rt_se[cpu];
368 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
369 static inline struct task_group *task_group(struct task_struct *p)
374 #endif /* CONFIG_GROUP_SCHED */
376 /* CFS-related fields in a runqueue */
378 struct load_weight load;
379 unsigned long nr_running;
385 struct rb_root tasks_timeline;
386 struct rb_node *rb_leftmost;
388 struct list_head tasks;
389 struct list_head *balance_iterator;
392 * 'curr' points to currently running entity on this cfs_rq.
393 * It is set to NULL otherwise (i.e when none are currently running).
395 struct sched_entity *curr, *next;
397 unsigned long nr_spread_over;
399 #ifdef CONFIG_FAIR_GROUP_SCHED
400 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
403 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
404 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
405 * (like users, containers etc.)
407 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
408 * list is used during load balance.
410 struct list_head leaf_cfs_rq_list;
411 struct task_group *tg; /* group that "owns" this runqueue */
415 * the part of load.weight contributed by tasks
417 unsigned long task_weight;
420 * h_load = weight * f(tg)
422 * Where f(tg) is the recursive weight fraction assigned to
425 unsigned long h_load;
428 * this cpu's part of tg->shares
430 unsigned long shares;
433 * load.weight at the time we set shares
435 unsigned long rq_weight;
440 /* Real-Time classes' related field in a runqueue: */
442 struct rt_prio_array active;
443 unsigned long rt_nr_running;
444 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
445 int highest_prio; /* highest queued rt task prio */
448 unsigned long rt_nr_migratory;
454 /* Nests inside the rq lock: */
455 spinlock_t rt_runtime_lock;
457 #ifdef CONFIG_RT_GROUP_SCHED
458 unsigned long rt_nr_boosted;
461 struct list_head leaf_rt_rq_list;
462 struct task_group *tg;
463 struct sched_rt_entity *rt_se;
470 * We add the notion of a root-domain which will be used to define per-domain
471 * variables. Each exclusive cpuset essentially defines an island domain by
472 * fully partitioning the member cpus from any other cpuset. Whenever a new
473 * exclusive cpuset is created, we also create and attach a new root-domain
483 * The "RT overload" flag: it gets set if a CPU has more than
484 * one runnable RT task.
489 struct cpupri cpupri;
494 * By default the system creates a single root-domain with all cpus as
495 * members (mimicking the global state we have today).
497 static struct root_domain def_root_domain;
502 * This is the main, per-CPU runqueue data structure.
504 * Locking rule: those places that want to lock multiple runqueues
505 * (such as the load balancing or the thread migration code), lock
506 * acquire operations must be ordered by ascending &runqueue.
513 * nr_running and cpu_load should be in the same cacheline because
514 * remote CPUs use both these fields when doing load calculation.
516 unsigned long nr_running;
517 #define CPU_LOAD_IDX_MAX 5
518 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
519 unsigned char idle_at_tick;
521 unsigned long last_tick_seen;
522 unsigned char in_nohz_recently;
524 /* capture load from *all* tasks on this cpu: */
525 struct load_weight load;
526 unsigned long nr_load_updates;
532 #ifdef CONFIG_FAIR_GROUP_SCHED
533 /* list of leaf cfs_rq on this cpu: */
534 struct list_head leaf_cfs_rq_list;
536 #ifdef CONFIG_RT_GROUP_SCHED
537 struct list_head leaf_rt_rq_list;
541 * This is part of a global counter where only the total sum
542 * over all CPUs matters. A task can increase this counter on
543 * one CPU and if it got migrated afterwards it may decrease
544 * it on another CPU. Always updated under the runqueue lock:
546 unsigned long nr_uninterruptible;
548 struct task_struct *curr, *idle;
549 unsigned long next_balance;
550 struct mm_struct *prev_mm;
557 struct root_domain *rd;
558 struct sched_domain *sd;
560 /* For active balancing */
563 /* cpu of this runqueue: */
567 unsigned long avg_load_per_task;
569 struct task_struct *migration_thread;
570 struct list_head migration_queue;
573 #ifdef CONFIG_SCHED_HRTICK
575 int hrtick_csd_pending;
576 struct call_single_data hrtick_csd;
578 struct hrtimer hrtick_timer;
581 #ifdef CONFIG_SCHEDSTATS
583 struct sched_info rq_sched_info;
585 /* sys_sched_yield() stats */
586 unsigned int yld_exp_empty;
587 unsigned int yld_act_empty;
588 unsigned int yld_both_empty;
589 unsigned int yld_count;
591 /* schedule() stats */
592 unsigned int sched_switch;
593 unsigned int sched_count;
594 unsigned int sched_goidle;
596 /* try_to_wake_up() stats */
597 unsigned int ttwu_count;
598 unsigned int ttwu_local;
601 unsigned int bkl_count;
605 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
607 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
609 rq->curr->sched_class->check_preempt_curr(rq, p);
612 static inline int cpu_of(struct rq *rq)
622 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
623 * See detach_destroy_domains: synchronize_sched for details.
625 * The domain tree of any CPU may only be accessed from within
626 * preempt-disabled sections.
628 #define for_each_domain(cpu, __sd) \
629 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
631 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
632 #define this_rq() (&__get_cpu_var(runqueues))
633 #define task_rq(p) cpu_rq(task_cpu(p))
634 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
636 static inline void update_rq_clock(struct rq *rq)
638 rq->clock = sched_clock_cpu(cpu_of(rq));
642 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
644 #ifdef CONFIG_SCHED_DEBUG
645 # define const_debug __read_mostly
647 # define const_debug static const
653 * Returns true if the current cpu runqueue is locked.
654 * This interface allows printk to be called with the runqueue lock
655 * held and know whether or not it is OK to wake up the klogd.
657 int runqueue_is_locked(void)
660 struct rq *rq = cpu_rq(cpu);
663 ret = spin_is_locked(&rq->lock);
669 * Debugging: various feature bits
672 #define SCHED_FEAT(name, enabled) \
673 __SCHED_FEAT_##name ,
676 #include "sched_features.h"
681 #define SCHED_FEAT(name, enabled) \
682 (1UL << __SCHED_FEAT_##name) * enabled |
684 const_debug unsigned int sysctl_sched_features =
685 #include "sched_features.h"
690 #ifdef CONFIG_SCHED_DEBUG
691 #define SCHED_FEAT(name, enabled) \
694 static __read_mostly char *sched_feat_names[] = {
695 #include "sched_features.h"
701 static int sched_feat_open(struct inode *inode, struct file *filp)
703 filp->private_data = inode->i_private;
708 sched_feat_read(struct file *filp, char __user *ubuf,
709 size_t cnt, loff_t *ppos)
716 for (i = 0; sched_feat_names[i]; i++) {
717 len += strlen(sched_feat_names[i]);
721 buf = kmalloc(len + 2, GFP_KERNEL);
725 for (i = 0; sched_feat_names[i]; i++) {
726 if (sysctl_sched_features & (1UL << i))
727 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
729 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
732 r += sprintf(buf + r, "\n");
733 WARN_ON(r >= len + 2);
735 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
743 sched_feat_write(struct file *filp, const char __user *ubuf,
744 size_t cnt, loff_t *ppos)
754 if (copy_from_user(&buf, ubuf, cnt))
759 if (strncmp(buf, "NO_", 3) == 0) {
764 for (i = 0; sched_feat_names[i]; i++) {
765 int len = strlen(sched_feat_names[i]);
767 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
769 sysctl_sched_features &= ~(1UL << i);
771 sysctl_sched_features |= (1UL << i);
776 if (!sched_feat_names[i])
784 static struct file_operations sched_feat_fops = {
785 .open = sched_feat_open,
786 .read = sched_feat_read,
787 .write = sched_feat_write,
790 static __init int sched_init_debug(void)
792 debugfs_create_file("sched_features", 0644, NULL, NULL,
797 late_initcall(sched_init_debug);
801 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
804 * Number of tasks to iterate in a single balance run.
805 * Limited because this is done with IRQs disabled.
807 const_debug unsigned int sysctl_sched_nr_migrate = 32;
810 * ratelimit for updating the group shares.
813 unsigned int sysctl_sched_shares_ratelimit = 250000;
816 * period over which we measure -rt task cpu usage in us.
819 unsigned int sysctl_sched_rt_period = 1000000;
821 static __read_mostly int scheduler_running;
824 * part of the period that we allow rt tasks to run in us.
827 int sysctl_sched_rt_runtime = 950000;
829 static inline u64 global_rt_period(void)
831 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
834 static inline u64 global_rt_runtime(void)
836 if (sysctl_sched_rt_runtime < 0)
839 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
842 #ifndef prepare_arch_switch
843 # define prepare_arch_switch(next) do { } while (0)
845 #ifndef finish_arch_switch
846 # define finish_arch_switch(prev) do { } while (0)
849 static inline int task_current(struct rq *rq, struct task_struct *p)
851 return rq->curr == p;
854 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
855 static inline int task_running(struct rq *rq, struct task_struct *p)
857 return task_current(rq, p);
860 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
864 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
866 #ifdef CONFIG_DEBUG_SPINLOCK
867 /* this is a valid case when another task releases the spinlock */
868 rq->lock.owner = current;
871 * If we are tracking spinlock dependencies then we have to
872 * fix up the runqueue lock - which gets 'carried over' from
875 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
877 spin_unlock_irq(&rq->lock);
880 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
881 static inline int task_running(struct rq *rq, struct task_struct *p)
886 return task_current(rq, p);
890 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
894 * We can optimise this out completely for !SMP, because the
895 * SMP rebalancing from interrupt is the only thing that cares
900 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
901 spin_unlock_irq(&rq->lock);
903 spin_unlock(&rq->lock);
907 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
911 * After ->oncpu is cleared, the task can be moved to a different CPU.
912 * We must ensure this doesn't happen until the switch is completely
918 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
922 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
925 * __task_rq_lock - lock the runqueue a given task resides on.
926 * Must be called interrupts disabled.
928 static inline struct rq *__task_rq_lock(struct task_struct *p)
932 struct rq *rq = task_rq(p);
933 spin_lock(&rq->lock);
934 if (likely(rq == task_rq(p)))
936 spin_unlock(&rq->lock);
941 * task_rq_lock - lock the runqueue a given task resides on and disable
942 * interrupts. Note the ordering: we can safely lookup the task_rq without
943 * explicitly disabling preemption.
945 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
951 local_irq_save(*flags);
953 spin_lock(&rq->lock);
954 if (likely(rq == task_rq(p)))
956 spin_unlock_irqrestore(&rq->lock, *flags);
960 static void __task_rq_unlock(struct rq *rq)
963 spin_unlock(&rq->lock);
966 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
969 spin_unlock_irqrestore(&rq->lock, *flags);
973 * this_rq_lock - lock this runqueue and disable interrupts.
975 static struct rq *this_rq_lock(void)
982 spin_lock(&rq->lock);
987 #ifdef CONFIG_SCHED_HRTICK
989 * Use HR-timers to deliver accurate preemption points.
991 * Its all a bit involved since we cannot program an hrt while holding the
992 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
995 * When we get rescheduled we reprogram the hrtick_timer outside of the
1001 * - enabled by features
1002 * - hrtimer is actually high res
1004 static inline int hrtick_enabled(struct rq *rq)
1006 if (!sched_feat(HRTICK))
1008 if (!cpu_active(cpu_of(rq)))
1010 return hrtimer_is_hres_active(&rq->hrtick_timer);
1013 static void hrtick_clear(struct rq *rq)
1015 if (hrtimer_active(&rq->hrtick_timer))
1016 hrtimer_cancel(&rq->hrtick_timer);
1020 * High-resolution timer tick.
1021 * Runs from hardirq context with interrupts disabled.
1023 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1025 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1027 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1029 spin_lock(&rq->lock);
1030 update_rq_clock(rq);
1031 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1032 spin_unlock(&rq->lock);
1034 return HRTIMER_NORESTART;
1039 * called from hardirq (IPI) context
1041 static void __hrtick_start(void *arg)
1043 struct rq *rq = arg;
1045 spin_lock(&rq->lock);
1046 hrtimer_restart(&rq->hrtick_timer);
1047 rq->hrtick_csd_pending = 0;
1048 spin_unlock(&rq->lock);
1052 * Called to set the hrtick timer state.
1054 * called with rq->lock held and irqs disabled
1056 static void hrtick_start(struct rq *rq, u64 delay)
1058 struct hrtimer *timer = &rq->hrtick_timer;
1059 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1061 timer->expires = time;
1063 if (rq == this_rq()) {
1064 hrtimer_restart(timer);
1065 } else if (!rq->hrtick_csd_pending) {
1066 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1067 rq->hrtick_csd_pending = 1;
1072 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1074 int cpu = (int)(long)hcpu;
1077 case CPU_UP_CANCELED:
1078 case CPU_UP_CANCELED_FROZEN:
1079 case CPU_DOWN_PREPARE:
1080 case CPU_DOWN_PREPARE_FROZEN:
1082 case CPU_DEAD_FROZEN:
1083 hrtick_clear(cpu_rq(cpu));
1090 static void init_hrtick(void)
1092 hotcpu_notifier(hotplug_hrtick, 0);
1096 * Called to set the hrtick timer state.
1098 * called with rq->lock held and irqs disabled
1100 static void hrtick_start(struct rq *rq, u64 delay)
1102 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1105 static void init_hrtick(void)
1108 #endif /* CONFIG_SMP */
1110 static void init_rq_hrtick(struct rq *rq)
1113 rq->hrtick_csd_pending = 0;
1115 rq->hrtick_csd.flags = 0;
1116 rq->hrtick_csd.func = __hrtick_start;
1117 rq->hrtick_csd.info = rq;
1120 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1121 rq->hrtick_timer.function = hrtick;
1122 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1125 static inline void hrtick_clear(struct rq *rq)
1129 static inline void init_rq_hrtick(struct rq *rq)
1133 static inline void init_hrtick(void)
1139 * resched_task - mark a task 'to be rescheduled now'.
1141 * On UP this means the setting of the need_resched flag, on SMP it
1142 * might also involve a cross-CPU call to trigger the scheduler on
1147 #ifndef tsk_is_polling
1148 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1151 static void resched_task(struct task_struct *p)
1155 assert_spin_locked(&task_rq(p)->lock);
1157 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1160 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1163 if (cpu == smp_processor_id())
1166 /* NEED_RESCHED must be visible before we test polling */
1168 if (!tsk_is_polling(p))
1169 smp_send_reschedule(cpu);
1172 static void resched_cpu(int cpu)
1174 struct rq *rq = cpu_rq(cpu);
1175 unsigned long flags;
1177 if (!spin_trylock_irqsave(&rq->lock, flags))
1179 resched_task(cpu_curr(cpu));
1180 spin_unlock_irqrestore(&rq->lock, flags);
1185 * When add_timer_on() enqueues a timer into the timer wheel of an
1186 * idle CPU then this timer might expire before the next timer event
1187 * which is scheduled to wake up that CPU. In case of a completely
1188 * idle system the next event might even be infinite time into the
1189 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1190 * leaves the inner idle loop so the newly added timer is taken into
1191 * account when the CPU goes back to idle and evaluates the timer
1192 * wheel for the next timer event.
1194 void wake_up_idle_cpu(int cpu)
1196 struct rq *rq = cpu_rq(cpu);
1198 if (cpu == smp_processor_id())
1202 * This is safe, as this function is called with the timer
1203 * wheel base lock of (cpu) held. When the CPU is on the way
1204 * to idle and has not yet set rq->curr to idle then it will
1205 * be serialized on the timer wheel base lock and take the new
1206 * timer into account automatically.
1208 if (rq->curr != rq->idle)
1212 * We can set TIF_RESCHED on the idle task of the other CPU
1213 * lockless. The worst case is that the other CPU runs the
1214 * idle task through an additional NOOP schedule()
1216 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1218 /* NEED_RESCHED must be visible before we test polling */
1220 if (!tsk_is_polling(rq->idle))
1221 smp_send_reschedule(cpu);
1223 #endif /* CONFIG_NO_HZ */
1225 #else /* !CONFIG_SMP */
1226 static void resched_task(struct task_struct *p)
1228 assert_spin_locked(&task_rq(p)->lock);
1229 set_tsk_need_resched(p);
1231 #endif /* CONFIG_SMP */
1233 #if BITS_PER_LONG == 32
1234 # define WMULT_CONST (~0UL)
1236 # define WMULT_CONST (1UL << 32)
1239 #define WMULT_SHIFT 32
1242 * Shift right and round:
1244 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1247 * delta *= weight / lw
1249 static unsigned long
1250 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1251 struct load_weight *lw)
1255 if (!lw->inv_weight) {
1256 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1259 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1263 tmp = (u64)delta_exec * weight;
1265 * Check whether we'd overflow the 64-bit multiplication:
1267 if (unlikely(tmp > WMULT_CONST))
1268 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1271 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1273 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1276 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1282 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1289 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1290 * of tasks with abnormal "nice" values across CPUs the contribution that
1291 * each task makes to its run queue's load is weighted according to its
1292 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1293 * scaled version of the new time slice allocation that they receive on time
1297 #define WEIGHT_IDLEPRIO 2
1298 #define WMULT_IDLEPRIO (1 << 31)
1301 * Nice levels are multiplicative, with a gentle 10% change for every
1302 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1303 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1304 * that remained on nice 0.
1306 * The "10% effect" is relative and cumulative: from _any_ nice level,
1307 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1308 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1309 * If a task goes up by ~10% and another task goes down by ~10% then
1310 * the relative distance between them is ~25%.)
1312 static const int prio_to_weight[40] = {
1313 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1314 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1315 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1316 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1317 /* 0 */ 1024, 820, 655, 526, 423,
1318 /* 5 */ 335, 272, 215, 172, 137,
1319 /* 10 */ 110, 87, 70, 56, 45,
1320 /* 15 */ 36, 29, 23, 18, 15,
1324 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1326 * In cases where the weight does not change often, we can use the
1327 * precalculated inverse to speed up arithmetics by turning divisions
1328 * into multiplications:
1330 static const u32 prio_to_wmult[40] = {
1331 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1332 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1333 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1334 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1335 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1336 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1337 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1338 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1341 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1344 * runqueue iterator, to support SMP load-balancing between different
1345 * scheduling classes, without having to expose their internal data
1346 * structures to the load-balancing proper:
1348 struct rq_iterator {
1350 struct task_struct *(*start)(void *);
1351 struct task_struct *(*next)(void *);
1355 static unsigned long
1356 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1357 unsigned long max_load_move, struct sched_domain *sd,
1358 enum cpu_idle_type idle, int *all_pinned,
1359 int *this_best_prio, struct rq_iterator *iterator);
1362 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1363 struct sched_domain *sd, enum cpu_idle_type idle,
1364 struct rq_iterator *iterator);
1367 #ifdef CONFIG_CGROUP_CPUACCT
1368 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1370 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1373 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1375 update_load_add(&rq->load, load);
1378 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1380 update_load_sub(&rq->load, load);
1384 static unsigned long source_load(int cpu, int type);
1385 static unsigned long target_load(int cpu, int type);
1386 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1388 static unsigned long cpu_avg_load_per_task(int cpu)
1390 struct rq *rq = cpu_rq(cpu);
1393 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1395 return rq->avg_load_per_task;
1398 #ifdef CONFIG_FAIR_GROUP_SCHED
1400 typedef void (*tg_visitor)(struct task_group *, int, struct sched_domain *);
1403 * Iterate the full tree, calling @down when first entering a node and @up when
1404 * leaving it for the final time.
1407 walk_tg_tree(tg_visitor down, tg_visitor up, int cpu, struct sched_domain *sd)
1409 struct task_group *parent, *child;
1412 parent = &root_task_group;
1414 (*down)(parent, cpu, sd);
1415 list_for_each_entry_rcu(child, &parent->children, siblings) {
1422 (*up)(parent, cpu, sd);
1425 parent = parent->parent;
1431 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1434 * Calculate and set the cpu's group shares.
1437 __update_group_shares_cpu(struct task_group *tg, int cpu,
1438 unsigned long sd_shares, unsigned long sd_rq_weight)
1441 unsigned long shares;
1442 unsigned long rq_weight;
1447 rq_weight = tg->cfs_rq[cpu]->load.weight;
1450 * If there are currently no tasks on the cpu pretend there is one of
1451 * average load so that when a new task gets to run here it will not
1452 * get delayed by group starvation.
1456 rq_weight = NICE_0_LOAD;
1459 if (unlikely(rq_weight > sd_rq_weight))
1460 rq_weight = sd_rq_weight;
1463 * \Sum shares * rq_weight
1464 * shares = -----------------------
1468 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1471 * record the actual number of shares, not the boosted amount.
1473 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1474 tg->cfs_rq[cpu]->rq_weight = rq_weight;
1476 if (shares < MIN_SHARES)
1477 shares = MIN_SHARES;
1478 else if (shares > MAX_SHARES)
1479 shares = MAX_SHARES;
1481 __set_se_shares(tg->se[cpu], shares);
1485 * Re-compute the task group their per cpu shares over the given domain.
1486 * This needs to be done in a bottom-up fashion because the rq weight of a
1487 * parent group depends on the shares of its child groups.
1490 tg_shares_up(struct task_group *tg, int cpu, struct sched_domain *sd)
1492 unsigned long rq_weight = 0;
1493 unsigned long shares = 0;
1496 for_each_cpu_mask(i, sd->span) {
1497 rq_weight += tg->cfs_rq[i]->load.weight;
1498 shares += tg->cfs_rq[i]->shares;
1501 if ((!shares && rq_weight) || shares > tg->shares)
1502 shares = tg->shares;
1504 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1505 shares = tg->shares;
1508 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1510 for_each_cpu_mask(i, sd->span) {
1511 struct rq *rq = cpu_rq(i);
1512 unsigned long flags;
1514 spin_lock_irqsave(&rq->lock, flags);
1515 __update_group_shares_cpu(tg, i, shares, rq_weight);
1516 spin_unlock_irqrestore(&rq->lock, flags);
1521 * Compute the cpu's hierarchical load factor for each task group.
1522 * This needs to be done in a top-down fashion because the load of a child
1523 * group is a fraction of its parents load.
1526 tg_load_down(struct task_group *tg, int cpu, struct sched_domain *sd)
1531 load = cpu_rq(cpu)->load.weight;
1533 load = tg->parent->cfs_rq[cpu]->h_load;
1534 load *= tg->cfs_rq[cpu]->shares;
1535 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1538 tg->cfs_rq[cpu]->h_load = load;
1542 tg_nop(struct task_group *tg, int cpu, struct sched_domain *sd)
1546 static void update_shares(struct sched_domain *sd)
1548 u64 now = cpu_clock(raw_smp_processor_id());
1549 s64 elapsed = now - sd->last_update;
1551 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1552 sd->last_update = now;
1553 walk_tg_tree(tg_nop, tg_shares_up, 0, sd);
1557 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1559 spin_unlock(&rq->lock);
1561 spin_lock(&rq->lock);
1564 static void update_h_load(int cpu)
1566 walk_tg_tree(tg_load_down, tg_nop, cpu, NULL);
1571 static inline void update_shares(struct sched_domain *sd)
1575 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1583 #ifdef CONFIG_FAIR_GROUP_SCHED
1584 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1587 cfs_rq->shares = shares;
1592 #include "sched_stats.h"
1593 #include "sched_idletask.c"
1594 #include "sched_fair.c"
1595 #include "sched_rt.c"
1596 #ifdef CONFIG_SCHED_DEBUG
1597 # include "sched_debug.c"
1600 #define sched_class_highest (&rt_sched_class)
1601 #define for_each_class(class) \
1602 for (class = sched_class_highest; class; class = class->next)
1604 static void inc_nr_running(struct rq *rq)
1609 static void dec_nr_running(struct rq *rq)
1614 static void set_load_weight(struct task_struct *p)
1616 if (task_has_rt_policy(p)) {
1617 p->se.load.weight = prio_to_weight[0] * 2;
1618 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1623 * SCHED_IDLE tasks get minimal weight:
1625 if (p->policy == SCHED_IDLE) {
1626 p->se.load.weight = WEIGHT_IDLEPRIO;
1627 p->se.load.inv_weight = WMULT_IDLEPRIO;
1631 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1632 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1635 static void update_avg(u64 *avg, u64 sample)
1637 s64 diff = sample - *avg;
1641 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1643 sched_info_queued(p);
1644 p->sched_class->enqueue_task(rq, p, wakeup);
1648 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1650 if (sleep && p->se.last_wakeup) {
1651 update_avg(&p->se.avg_overlap,
1652 p->se.sum_exec_runtime - p->se.last_wakeup);
1653 p->se.last_wakeup = 0;
1656 sched_info_dequeued(p);
1657 p->sched_class->dequeue_task(rq, p, sleep);
1662 * __normal_prio - return the priority that is based on the static prio
1664 static inline int __normal_prio(struct task_struct *p)
1666 return p->static_prio;
1670 * Calculate the expected normal priority: i.e. priority
1671 * without taking RT-inheritance into account. Might be
1672 * boosted by interactivity modifiers. Changes upon fork,
1673 * setprio syscalls, and whenever the interactivity
1674 * estimator recalculates.
1676 static inline int normal_prio(struct task_struct *p)
1680 if (task_has_rt_policy(p))
1681 prio = MAX_RT_PRIO-1 - p->rt_priority;
1683 prio = __normal_prio(p);
1688 * Calculate the current priority, i.e. the priority
1689 * taken into account by the scheduler. This value might
1690 * be boosted by RT tasks, or might be boosted by
1691 * interactivity modifiers. Will be RT if the task got
1692 * RT-boosted. If not then it returns p->normal_prio.
1694 static int effective_prio(struct task_struct *p)
1696 p->normal_prio = normal_prio(p);
1698 * If we are RT tasks or we were boosted to RT priority,
1699 * keep the priority unchanged. Otherwise, update priority
1700 * to the normal priority:
1702 if (!rt_prio(p->prio))
1703 return p->normal_prio;
1708 * activate_task - move a task to the runqueue.
1710 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1712 if (task_contributes_to_load(p))
1713 rq->nr_uninterruptible--;
1715 enqueue_task(rq, p, wakeup);
1720 * deactivate_task - remove a task from the runqueue.
1722 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1724 if (task_contributes_to_load(p))
1725 rq->nr_uninterruptible++;
1727 dequeue_task(rq, p, sleep);
1732 * task_curr - is this task currently executing on a CPU?
1733 * @p: the task in question.
1735 inline int task_curr(const struct task_struct *p)
1737 return cpu_curr(task_cpu(p)) == p;
1740 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1742 set_task_rq(p, cpu);
1745 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1746 * successfuly executed on another CPU. We must ensure that updates of
1747 * per-task data have been completed by this moment.
1750 task_thread_info(p)->cpu = cpu;
1754 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1755 const struct sched_class *prev_class,
1756 int oldprio, int running)
1758 if (prev_class != p->sched_class) {
1759 if (prev_class->switched_from)
1760 prev_class->switched_from(rq, p, running);
1761 p->sched_class->switched_to(rq, p, running);
1763 p->sched_class->prio_changed(rq, p, oldprio, running);
1768 /* Used instead of source_load when we know the type == 0 */
1769 static unsigned long weighted_cpuload(const int cpu)
1771 return cpu_rq(cpu)->load.weight;
1775 * Is this task likely cache-hot:
1778 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1783 * Buddy candidates are cache hot:
1785 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1788 if (p->sched_class != &fair_sched_class)
1791 if (sysctl_sched_migration_cost == -1)
1793 if (sysctl_sched_migration_cost == 0)
1796 delta = now - p->se.exec_start;
1798 return delta < (s64)sysctl_sched_migration_cost;
1802 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1804 int old_cpu = task_cpu(p);
1805 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1806 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1807 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1810 clock_offset = old_rq->clock - new_rq->clock;
1812 #ifdef CONFIG_SCHEDSTATS
1813 if (p->se.wait_start)
1814 p->se.wait_start -= clock_offset;
1815 if (p->se.sleep_start)
1816 p->se.sleep_start -= clock_offset;
1817 if (p->se.block_start)
1818 p->se.block_start -= clock_offset;
1819 if (old_cpu != new_cpu) {
1820 schedstat_inc(p, se.nr_migrations);
1821 if (task_hot(p, old_rq->clock, NULL))
1822 schedstat_inc(p, se.nr_forced2_migrations);
1825 p->se.vruntime -= old_cfsrq->min_vruntime -
1826 new_cfsrq->min_vruntime;
1828 __set_task_cpu(p, new_cpu);
1831 struct migration_req {
1832 struct list_head list;
1834 struct task_struct *task;
1837 struct completion done;
1841 * The task's runqueue lock must be held.
1842 * Returns true if you have to wait for migration thread.
1845 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1847 struct rq *rq = task_rq(p);
1850 * If the task is not on a runqueue (and not running), then
1851 * it is sufficient to simply update the task's cpu field.
1853 if (!p->se.on_rq && !task_running(rq, p)) {
1854 set_task_cpu(p, dest_cpu);
1858 init_completion(&req->done);
1860 req->dest_cpu = dest_cpu;
1861 list_add(&req->list, &rq->migration_queue);
1867 * wait_task_inactive - wait for a thread to unschedule.
1869 * If @match_state is nonzero, it's the @p->state value just checked and
1870 * not expected to change. If it changes, i.e. @p might have woken up,
1871 * then return zero. When we succeed in waiting for @p to be off its CPU,
1872 * we return a positive number (its total switch count). If a second call
1873 * a short while later returns the same number, the caller can be sure that
1874 * @p has remained unscheduled the whole time.
1876 * The caller must ensure that the task *will* unschedule sometime soon,
1877 * else this function might spin for a *long* time. This function can't
1878 * be called with interrupts off, or it may introduce deadlock with
1879 * smp_call_function() if an IPI is sent by the same process we are
1880 * waiting to become inactive.
1882 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1884 unsigned long flags;
1891 * We do the initial early heuristics without holding
1892 * any task-queue locks at all. We'll only try to get
1893 * the runqueue lock when things look like they will
1899 * If the task is actively running on another CPU
1900 * still, just relax and busy-wait without holding
1903 * NOTE! Since we don't hold any locks, it's not
1904 * even sure that "rq" stays as the right runqueue!
1905 * But we don't care, since "task_running()" will
1906 * return false if the runqueue has changed and p
1907 * is actually now running somewhere else!
1909 while (task_running(rq, p)) {
1910 if (match_state && unlikely(p->state != match_state))
1916 * Ok, time to look more closely! We need the rq
1917 * lock now, to be *sure*. If we're wrong, we'll
1918 * just go back and repeat.
1920 rq = task_rq_lock(p, &flags);
1921 running = task_running(rq, p);
1922 on_rq = p->se.on_rq;
1924 if (!match_state || p->state == match_state) {
1925 ncsw = p->nivcsw + p->nvcsw;
1926 if (unlikely(!ncsw))
1929 task_rq_unlock(rq, &flags);
1932 * If it changed from the expected state, bail out now.
1934 if (unlikely(!ncsw))
1938 * Was it really running after all now that we
1939 * checked with the proper locks actually held?
1941 * Oops. Go back and try again..
1943 if (unlikely(running)) {
1949 * It's not enough that it's not actively running,
1950 * it must be off the runqueue _entirely_, and not
1953 * So if it wa still runnable (but just not actively
1954 * running right now), it's preempted, and we should
1955 * yield - it could be a while.
1957 if (unlikely(on_rq)) {
1958 schedule_timeout_uninterruptible(1);
1963 * Ahh, all good. It wasn't running, and it wasn't
1964 * runnable, which means that it will never become
1965 * running in the future either. We're all done!
1974 * kick_process - kick a running thread to enter/exit the kernel
1975 * @p: the to-be-kicked thread
1977 * Cause a process which is running on another CPU to enter
1978 * kernel-mode, without any delay. (to get signals handled.)
1980 * NOTE: this function doesnt have to take the runqueue lock,
1981 * because all it wants to ensure is that the remote task enters
1982 * the kernel. If the IPI races and the task has been migrated
1983 * to another CPU then no harm is done and the purpose has been
1986 void kick_process(struct task_struct *p)
1992 if ((cpu != smp_processor_id()) && task_curr(p))
1993 smp_send_reschedule(cpu);
1998 * Return a low guess at the load of a migration-source cpu weighted
1999 * according to the scheduling class and "nice" value.
2001 * We want to under-estimate the load of migration sources, to
2002 * balance conservatively.
2004 static unsigned long source_load(int cpu, int type)
2006 struct rq *rq = cpu_rq(cpu);
2007 unsigned long total = weighted_cpuload(cpu);
2009 if (type == 0 || !sched_feat(LB_BIAS))
2012 return min(rq->cpu_load[type-1], total);
2016 * Return a high guess at the load of a migration-target cpu weighted
2017 * according to the scheduling class and "nice" value.
2019 static unsigned long target_load(int cpu, int type)
2021 struct rq *rq = cpu_rq(cpu);
2022 unsigned long total = weighted_cpuload(cpu);
2024 if (type == 0 || !sched_feat(LB_BIAS))
2027 return max(rq->cpu_load[type-1], total);
2031 * find_idlest_group finds and returns the least busy CPU group within the
2034 static struct sched_group *
2035 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2037 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2038 unsigned long min_load = ULONG_MAX, this_load = 0;
2039 int load_idx = sd->forkexec_idx;
2040 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2043 unsigned long load, avg_load;
2047 /* Skip over this group if it has no CPUs allowed */
2048 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2051 local_group = cpu_isset(this_cpu, group->cpumask);
2053 /* Tally up the load of all CPUs in the group */
2056 for_each_cpu_mask_nr(i, group->cpumask) {
2057 /* Bias balancing toward cpus of our domain */
2059 load = source_load(i, load_idx);
2061 load = target_load(i, load_idx);
2066 /* Adjust by relative CPU power of the group */
2067 avg_load = sg_div_cpu_power(group,
2068 avg_load * SCHED_LOAD_SCALE);
2071 this_load = avg_load;
2073 } else if (avg_load < min_load) {
2074 min_load = avg_load;
2077 } while (group = group->next, group != sd->groups);
2079 if (!idlest || 100*this_load < imbalance*min_load)
2085 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2088 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2091 unsigned long load, min_load = ULONG_MAX;
2095 /* Traverse only the allowed CPUs */
2096 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2098 for_each_cpu_mask_nr(i, *tmp) {
2099 load = weighted_cpuload(i);
2101 if (load < min_load || (load == min_load && i == this_cpu)) {
2111 * sched_balance_self: balance the current task (running on cpu) in domains
2112 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2115 * Balance, ie. select the least loaded group.
2117 * Returns the target CPU number, or the same CPU if no balancing is needed.
2119 * preempt must be disabled.
2121 static int sched_balance_self(int cpu, int flag)
2123 struct task_struct *t = current;
2124 struct sched_domain *tmp, *sd = NULL;
2126 for_each_domain(cpu, tmp) {
2128 * If power savings logic is enabled for a domain, stop there.
2130 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2132 if (tmp->flags & flag)
2140 cpumask_t span, tmpmask;
2141 struct sched_group *group;
2142 int new_cpu, weight;
2144 if (!(sd->flags & flag)) {
2150 group = find_idlest_group(sd, t, cpu);
2156 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2157 if (new_cpu == -1 || new_cpu == cpu) {
2158 /* Now try balancing at a lower domain level of cpu */
2163 /* Now try balancing at a lower domain level of new_cpu */
2166 weight = cpus_weight(span);
2167 for_each_domain(cpu, tmp) {
2168 if (weight <= cpus_weight(tmp->span))
2170 if (tmp->flags & flag)
2173 /* while loop will break here if sd == NULL */
2179 #endif /* CONFIG_SMP */
2182 * try_to_wake_up - wake up a thread
2183 * @p: the to-be-woken-up thread
2184 * @state: the mask of task states that can be woken
2185 * @sync: do a synchronous wakeup?
2187 * Put it on the run-queue if it's not already there. The "current"
2188 * thread is always on the run-queue (except when the actual
2189 * re-schedule is in progress), and as such you're allowed to do
2190 * the simpler "current->state = TASK_RUNNING" to mark yourself
2191 * runnable without the overhead of this.
2193 * returns failure only if the task is already active.
2195 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2197 int cpu, orig_cpu, this_cpu, success = 0;
2198 unsigned long flags;
2202 if (!sched_feat(SYNC_WAKEUPS))
2206 if (sched_feat(LB_WAKEUP_UPDATE)) {
2207 struct sched_domain *sd;
2209 this_cpu = raw_smp_processor_id();
2212 for_each_domain(this_cpu, sd) {
2213 if (cpu_isset(cpu, sd->span)) {
2222 rq = task_rq_lock(p, &flags);
2223 old_state = p->state;
2224 if (!(old_state & state))
2232 this_cpu = smp_processor_id();
2235 if (unlikely(task_running(rq, p)))
2238 cpu = p->sched_class->select_task_rq(p, sync);
2239 if (cpu != orig_cpu) {
2240 set_task_cpu(p, cpu);
2241 task_rq_unlock(rq, &flags);
2242 /* might preempt at this point */
2243 rq = task_rq_lock(p, &flags);
2244 old_state = p->state;
2245 if (!(old_state & state))
2250 this_cpu = smp_processor_id();
2254 #ifdef CONFIG_SCHEDSTATS
2255 schedstat_inc(rq, ttwu_count);
2256 if (cpu == this_cpu)
2257 schedstat_inc(rq, ttwu_local);
2259 struct sched_domain *sd;
2260 for_each_domain(this_cpu, sd) {
2261 if (cpu_isset(cpu, sd->span)) {
2262 schedstat_inc(sd, ttwu_wake_remote);
2267 #endif /* CONFIG_SCHEDSTATS */
2270 #endif /* CONFIG_SMP */
2271 schedstat_inc(p, se.nr_wakeups);
2273 schedstat_inc(p, se.nr_wakeups_sync);
2274 if (orig_cpu != cpu)
2275 schedstat_inc(p, se.nr_wakeups_migrate);
2276 if (cpu == this_cpu)
2277 schedstat_inc(p, se.nr_wakeups_local);
2279 schedstat_inc(p, se.nr_wakeups_remote);
2280 update_rq_clock(rq);
2281 activate_task(rq, p, 1);
2285 trace_mark(kernel_sched_wakeup,
2286 "pid %d state %ld ## rq %p task %p rq->curr %p",
2287 p->pid, p->state, rq, p, rq->curr);
2288 check_preempt_curr(rq, p);
2290 p->state = TASK_RUNNING;
2292 if (p->sched_class->task_wake_up)
2293 p->sched_class->task_wake_up(rq, p);
2296 current->se.last_wakeup = current->se.sum_exec_runtime;
2298 task_rq_unlock(rq, &flags);
2303 int wake_up_process(struct task_struct *p)
2305 return try_to_wake_up(p, TASK_ALL, 0);
2307 EXPORT_SYMBOL(wake_up_process);
2309 int wake_up_state(struct task_struct *p, unsigned int state)
2311 return try_to_wake_up(p, state, 0);
2315 * Perform scheduler related setup for a newly forked process p.
2316 * p is forked by current.
2318 * __sched_fork() is basic setup used by init_idle() too:
2320 static void __sched_fork(struct task_struct *p)
2322 p->se.exec_start = 0;
2323 p->se.sum_exec_runtime = 0;
2324 p->se.prev_sum_exec_runtime = 0;
2325 p->se.last_wakeup = 0;
2326 p->se.avg_overlap = 0;
2328 #ifdef CONFIG_SCHEDSTATS
2329 p->se.wait_start = 0;
2330 p->se.sum_sleep_runtime = 0;
2331 p->se.sleep_start = 0;
2332 p->se.block_start = 0;
2333 p->se.sleep_max = 0;
2334 p->se.block_max = 0;
2336 p->se.slice_max = 0;
2340 INIT_LIST_HEAD(&p->rt.run_list);
2342 INIT_LIST_HEAD(&p->se.group_node);
2344 #ifdef CONFIG_PREEMPT_NOTIFIERS
2345 INIT_HLIST_HEAD(&p->preempt_notifiers);
2349 * We mark the process as running here, but have not actually
2350 * inserted it onto the runqueue yet. This guarantees that
2351 * nobody will actually run it, and a signal or other external
2352 * event cannot wake it up and insert it on the runqueue either.
2354 p->state = TASK_RUNNING;
2358 * fork()/clone()-time setup:
2360 void sched_fork(struct task_struct *p, int clone_flags)
2362 int cpu = get_cpu();
2367 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2369 set_task_cpu(p, cpu);
2372 * Make sure we do not leak PI boosting priority to the child:
2374 p->prio = current->normal_prio;
2375 if (!rt_prio(p->prio))
2376 p->sched_class = &fair_sched_class;
2378 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2379 if (likely(sched_info_on()))
2380 memset(&p->sched_info, 0, sizeof(p->sched_info));
2382 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2385 #ifdef CONFIG_PREEMPT
2386 /* Want to start with kernel preemption disabled. */
2387 task_thread_info(p)->preempt_count = 1;
2393 * wake_up_new_task - wake up a newly created task for the first time.
2395 * This function will do some initial scheduler statistics housekeeping
2396 * that must be done for every newly created context, then puts the task
2397 * on the runqueue and wakes it.
2399 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2401 unsigned long flags;
2404 rq = task_rq_lock(p, &flags);
2405 BUG_ON(p->state != TASK_RUNNING);
2406 update_rq_clock(rq);
2408 p->prio = effective_prio(p);
2410 if (!p->sched_class->task_new || !current->se.on_rq) {
2411 activate_task(rq, p, 0);
2414 * Let the scheduling class do new task startup
2415 * management (if any):
2417 p->sched_class->task_new(rq, p);
2420 trace_mark(kernel_sched_wakeup_new,
2421 "pid %d state %ld ## rq %p task %p rq->curr %p",
2422 p->pid, p->state, rq, p, rq->curr);
2423 check_preempt_curr(rq, p);
2425 if (p->sched_class->task_wake_up)
2426 p->sched_class->task_wake_up(rq, p);
2428 task_rq_unlock(rq, &flags);
2431 #ifdef CONFIG_PREEMPT_NOTIFIERS
2434 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2435 * @notifier: notifier struct to register
2437 void preempt_notifier_register(struct preempt_notifier *notifier)
2439 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2441 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2444 * preempt_notifier_unregister - no longer interested in preemption notifications
2445 * @notifier: notifier struct to unregister
2447 * This is safe to call from within a preemption notifier.
2449 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2451 hlist_del(¬ifier->link);
2453 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2455 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2457 struct preempt_notifier *notifier;
2458 struct hlist_node *node;
2460 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2461 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2465 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2466 struct task_struct *next)
2468 struct preempt_notifier *notifier;
2469 struct hlist_node *node;
2471 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2472 notifier->ops->sched_out(notifier, next);
2475 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2477 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2482 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2483 struct task_struct *next)
2487 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2490 * prepare_task_switch - prepare to switch tasks
2491 * @rq: the runqueue preparing to switch
2492 * @prev: the current task that is being switched out
2493 * @next: the task we are going to switch to.
2495 * This is called with the rq lock held and interrupts off. It must
2496 * be paired with a subsequent finish_task_switch after the context
2499 * prepare_task_switch sets up locking and calls architecture specific
2503 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2504 struct task_struct *next)
2506 fire_sched_out_preempt_notifiers(prev, next);
2507 prepare_lock_switch(rq, next);
2508 prepare_arch_switch(next);
2512 * finish_task_switch - clean up after a task-switch
2513 * @rq: runqueue associated with task-switch
2514 * @prev: the thread we just switched away from.
2516 * finish_task_switch must be called after the context switch, paired
2517 * with a prepare_task_switch call before the context switch.
2518 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2519 * and do any other architecture-specific cleanup actions.
2521 * Note that we may have delayed dropping an mm in context_switch(). If
2522 * so, we finish that here outside of the runqueue lock. (Doing it
2523 * with the lock held can cause deadlocks; see schedule() for
2526 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2527 __releases(rq->lock)
2529 struct mm_struct *mm = rq->prev_mm;
2535 * A task struct has one reference for the use as "current".
2536 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2537 * schedule one last time. The schedule call will never return, and
2538 * the scheduled task must drop that reference.
2539 * The test for TASK_DEAD must occur while the runqueue locks are
2540 * still held, otherwise prev could be scheduled on another cpu, die
2541 * there before we look at prev->state, and then the reference would
2543 * Manfred Spraul <manfred@colorfullife.com>
2545 prev_state = prev->state;
2546 finish_arch_switch(prev);
2547 finish_lock_switch(rq, prev);
2549 if (current->sched_class->post_schedule)
2550 current->sched_class->post_schedule(rq);
2553 fire_sched_in_preempt_notifiers(current);
2556 if (unlikely(prev_state == TASK_DEAD)) {
2558 * Remove function-return probe instances associated with this
2559 * task and put them back on the free list.
2561 kprobe_flush_task(prev);
2562 put_task_struct(prev);
2567 * schedule_tail - first thing a freshly forked thread must call.
2568 * @prev: the thread we just switched away from.
2570 asmlinkage void schedule_tail(struct task_struct *prev)
2571 __releases(rq->lock)
2573 struct rq *rq = this_rq();
2575 finish_task_switch(rq, prev);
2576 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2577 /* In this case, finish_task_switch does not reenable preemption */
2580 if (current->set_child_tid)
2581 put_user(task_pid_vnr(current), current->set_child_tid);
2585 * context_switch - switch to the new MM and the new
2586 * thread's register state.
2589 context_switch(struct rq *rq, struct task_struct *prev,
2590 struct task_struct *next)
2592 struct mm_struct *mm, *oldmm;
2594 prepare_task_switch(rq, prev, next);
2595 trace_mark(kernel_sched_schedule,
2596 "prev_pid %d next_pid %d prev_state %ld "
2597 "## rq %p prev %p next %p",
2598 prev->pid, next->pid, prev->state,
2601 oldmm = prev->active_mm;
2603 * For paravirt, this is coupled with an exit in switch_to to
2604 * combine the page table reload and the switch backend into
2607 arch_enter_lazy_cpu_mode();
2609 if (unlikely(!mm)) {
2610 next->active_mm = oldmm;
2611 atomic_inc(&oldmm->mm_count);
2612 enter_lazy_tlb(oldmm, next);
2614 switch_mm(oldmm, mm, next);
2616 if (unlikely(!prev->mm)) {
2617 prev->active_mm = NULL;
2618 rq->prev_mm = oldmm;
2621 * Since the runqueue lock will be released by the next
2622 * task (which is an invalid locking op but in the case
2623 * of the scheduler it's an obvious special-case), so we
2624 * do an early lockdep release here:
2626 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2627 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2630 /* Here we just switch the register state and the stack. */
2631 switch_to(prev, next, prev);
2635 * this_rq must be evaluated again because prev may have moved
2636 * CPUs since it called schedule(), thus the 'rq' on its stack
2637 * frame will be invalid.
2639 finish_task_switch(this_rq(), prev);
2643 * nr_running, nr_uninterruptible and nr_context_switches:
2645 * externally visible scheduler statistics: current number of runnable
2646 * threads, current number of uninterruptible-sleeping threads, total
2647 * number of context switches performed since bootup.
2649 unsigned long nr_running(void)
2651 unsigned long i, sum = 0;
2653 for_each_online_cpu(i)
2654 sum += cpu_rq(i)->nr_running;
2659 unsigned long nr_uninterruptible(void)
2661 unsigned long i, sum = 0;
2663 for_each_possible_cpu(i)
2664 sum += cpu_rq(i)->nr_uninterruptible;
2667 * Since we read the counters lockless, it might be slightly
2668 * inaccurate. Do not allow it to go below zero though:
2670 if (unlikely((long)sum < 0))
2676 unsigned long long nr_context_switches(void)
2679 unsigned long long sum = 0;
2681 for_each_possible_cpu(i)
2682 sum += cpu_rq(i)->nr_switches;
2687 unsigned long nr_iowait(void)
2689 unsigned long i, sum = 0;
2691 for_each_possible_cpu(i)
2692 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2697 unsigned long nr_active(void)
2699 unsigned long i, running = 0, uninterruptible = 0;
2701 for_each_online_cpu(i) {
2702 running += cpu_rq(i)->nr_running;
2703 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2706 if (unlikely((long)uninterruptible < 0))
2707 uninterruptible = 0;
2709 return running + uninterruptible;
2713 * Update rq->cpu_load[] statistics. This function is usually called every
2714 * scheduler tick (TICK_NSEC).
2716 static void update_cpu_load(struct rq *this_rq)
2718 unsigned long this_load = this_rq->load.weight;
2721 this_rq->nr_load_updates++;
2723 /* Update our load: */
2724 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2725 unsigned long old_load, new_load;
2727 /* scale is effectively 1 << i now, and >> i divides by scale */
2729 old_load = this_rq->cpu_load[i];
2730 new_load = this_load;
2732 * Round up the averaging division if load is increasing. This
2733 * prevents us from getting stuck on 9 if the load is 10, for
2736 if (new_load > old_load)
2737 new_load += scale-1;
2738 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2745 * double_rq_lock - safely lock two runqueues
2747 * Note this does not disable interrupts like task_rq_lock,
2748 * you need to do so manually before calling.
2750 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2751 __acquires(rq1->lock)
2752 __acquires(rq2->lock)
2754 BUG_ON(!irqs_disabled());
2756 spin_lock(&rq1->lock);
2757 __acquire(rq2->lock); /* Fake it out ;) */
2760 spin_lock(&rq1->lock);
2761 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2763 spin_lock(&rq2->lock);
2764 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2767 update_rq_clock(rq1);
2768 update_rq_clock(rq2);
2772 * double_rq_unlock - safely unlock two runqueues
2774 * Note this does not restore interrupts like task_rq_unlock,
2775 * you need to do so manually after calling.
2777 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2778 __releases(rq1->lock)
2779 __releases(rq2->lock)
2781 spin_unlock(&rq1->lock);
2783 spin_unlock(&rq2->lock);
2785 __release(rq2->lock);
2789 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2791 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2792 __releases(this_rq->lock)
2793 __acquires(busiest->lock)
2794 __acquires(this_rq->lock)
2798 if (unlikely(!irqs_disabled())) {
2799 /* printk() doesn't work good under rq->lock */
2800 spin_unlock(&this_rq->lock);
2803 if (unlikely(!spin_trylock(&busiest->lock))) {
2804 if (busiest < this_rq) {
2805 spin_unlock(&this_rq->lock);
2806 spin_lock(&busiest->lock);
2807 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2810 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2815 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2816 __releases(busiest->lock)
2818 spin_unlock(&busiest->lock);
2819 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2823 * If dest_cpu is allowed for this process, migrate the task to it.
2824 * This is accomplished by forcing the cpu_allowed mask to only
2825 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2826 * the cpu_allowed mask is restored.
2828 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2830 struct migration_req req;
2831 unsigned long flags;
2834 rq = task_rq_lock(p, &flags);
2835 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2836 || unlikely(!cpu_active(dest_cpu)))
2839 /* force the process onto the specified CPU */
2840 if (migrate_task(p, dest_cpu, &req)) {
2841 /* Need to wait for migration thread (might exit: take ref). */
2842 struct task_struct *mt = rq->migration_thread;
2844 get_task_struct(mt);
2845 task_rq_unlock(rq, &flags);
2846 wake_up_process(mt);
2847 put_task_struct(mt);
2848 wait_for_completion(&req.done);
2853 task_rq_unlock(rq, &flags);
2857 * sched_exec - execve() is a valuable balancing opportunity, because at
2858 * this point the task has the smallest effective memory and cache footprint.
2860 void sched_exec(void)
2862 int new_cpu, this_cpu = get_cpu();
2863 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2865 if (new_cpu != this_cpu)
2866 sched_migrate_task(current, new_cpu);
2870 * pull_task - move a task from a remote runqueue to the local runqueue.
2871 * Both runqueues must be locked.
2873 static void pull_task(struct rq *src_rq, struct task_struct *p,
2874 struct rq *this_rq, int this_cpu)
2876 deactivate_task(src_rq, p, 0);
2877 set_task_cpu(p, this_cpu);
2878 activate_task(this_rq, p, 0);
2880 * Note that idle threads have a prio of MAX_PRIO, for this test
2881 * to be always true for them.
2883 check_preempt_curr(this_rq, p);
2887 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2890 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2891 struct sched_domain *sd, enum cpu_idle_type idle,
2895 * We do not migrate tasks that are:
2896 * 1) running (obviously), or
2897 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2898 * 3) are cache-hot on their current CPU.
2900 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2901 schedstat_inc(p, se.nr_failed_migrations_affine);
2906 if (task_running(rq, p)) {
2907 schedstat_inc(p, se.nr_failed_migrations_running);
2912 * Aggressive migration if:
2913 * 1) task is cache cold, or
2914 * 2) too many balance attempts have failed.
2917 if (!task_hot(p, rq->clock, sd) ||
2918 sd->nr_balance_failed > sd->cache_nice_tries) {
2919 #ifdef CONFIG_SCHEDSTATS
2920 if (task_hot(p, rq->clock, sd)) {
2921 schedstat_inc(sd, lb_hot_gained[idle]);
2922 schedstat_inc(p, se.nr_forced_migrations);
2928 if (task_hot(p, rq->clock, sd)) {
2929 schedstat_inc(p, se.nr_failed_migrations_hot);
2935 static unsigned long
2936 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2937 unsigned long max_load_move, struct sched_domain *sd,
2938 enum cpu_idle_type idle, int *all_pinned,
2939 int *this_best_prio, struct rq_iterator *iterator)
2941 int loops = 0, pulled = 0, pinned = 0;
2942 struct task_struct *p;
2943 long rem_load_move = max_load_move;
2945 if (max_load_move == 0)
2951 * Start the load-balancing iterator:
2953 p = iterator->start(iterator->arg);
2955 if (!p || loops++ > sysctl_sched_nr_migrate)
2958 if ((p->se.load.weight >> 1) > rem_load_move ||
2959 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2960 p = iterator->next(iterator->arg);
2964 pull_task(busiest, p, this_rq, this_cpu);
2966 rem_load_move -= p->se.load.weight;
2969 * We only want to steal up to the prescribed amount of weighted load.
2971 if (rem_load_move > 0) {
2972 if (p->prio < *this_best_prio)
2973 *this_best_prio = p->prio;
2974 p = iterator->next(iterator->arg);
2979 * Right now, this is one of only two places pull_task() is called,
2980 * so we can safely collect pull_task() stats here rather than
2981 * inside pull_task().
2983 schedstat_add(sd, lb_gained[idle], pulled);
2986 *all_pinned = pinned;
2988 return max_load_move - rem_load_move;
2992 * move_tasks tries to move up to max_load_move weighted load from busiest to
2993 * this_rq, as part of a balancing operation within domain "sd".
2994 * Returns 1 if successful and 0 otherwise.
2996 * Called with both runqueues locked.
2998 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2999 unsigned long max_load_move,
3000 struct sched_domain *sd, enum cpu_idle_type idle,
3003 const struct sched_class *class = sched_class_highest;
3004 unsigned long total_load_moved = 0;
3005 int this_best_prio = this_rq->curr->prio;
3009 class->load_balance(this_rq, this_cpu, busiest,
3010 max_load_move - total_load_moved,
3011 sd, idle, all_pinned, &this_best_prio);
3012 class = class->next;
3014 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3017 } while (class && max_load_move > total_load_moved);
3019 return total_load_moved > 0;
3023 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3024 struct sched_domain *sd, enum cpu_idle_type idle,
3025 struct rq_iterator *iterator)
3027 struct task_struct *p = iterator->start(iterator->arg);
3031 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3032 pull_task(busiest, p, this_rq, this_cpu);
3034 * Right now, this is only the second place pull_task()
3035 * is called, so we can safely collect pull_task()
3036 * stats here rather than inside pull_task().
3038 schedstat_inc(sd, lb_gained[idle]);
3042 p = iterator->next(iterator->arg);
3049 * move_one_task tries to move exactly one task from busiest to this_rq, as
3050 * part of active balancing operations within "domain".
3051 * Returns 1 if successful and 0 otherwise.
3053 * Called with both runqueues locked.
3055 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3056 struct sched_domain *sd, enum cpu_idle_type idle)
3058 const struct sched_class *class;
3060 for (class = sched_class_highest; class; class = class->next)
3061 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3068 * find_busiest_group finds and returns the busiest CPU group within the
3069 * domain. It calculates and returns the amount of weighted load which
3070 * should be moved to restore balance via the imbalance parameter.
3072 static struct sched_group *
3073 find_busiest_group(struct sched_domain *sd, int this_cpu,
3074 unsigned long *imbalance, enum cpu_idle_type idle,
3075 int *sd_idle, const cpumask_t *cpus, int *balance)
3077 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3078 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3079 unsigned long max_pull;
3080 unsigned long busiest_load_per_task, busiest_nr_running;
3081 unsigned long this_load_per_task, this_nr_running;
3082 int load_idx, group_imb = 0;
3083 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3084 int power_savings_balance = 1;
3085 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3086 unsigned long min_nr_running = ULONG_MAX;
3087 struct sched_group *group_min = NULL, *group_leader = NULL;
3090 max_load = this_load = total_load = total_pwr = 0;
3091 busiest_load_per_task = busiest_nr_running = 0;
3092 this_load_per_task = this_nr_running = 0;
3094 if (idle == CPU_NOT_IDLE)
3095 load_idx = sd->busy_idx;
3096 else if (idle == CPU_NEWLY_IDLE)
3097 load_idx = sd->newidle_idx;
3099 load_idx = sd->idle_idx;
3102 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3105 int __group_imb = 0;
3106 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3107 unsigned long sum_nr_running, sum_weighted_load;
3108 unsigned long sum_avg_load_per_task;
3109 unsigned long avg_load_per_task;
3111 local_group = cpu_isset(this_cpu, group->cpumask);
3114 balance_cpu = first_cpu(group->cpumask);
3116 /* Tally up the load of all CPUs in the group */
3117 sum_weighted_load = sum_nr_running = avg_load = 0;
3118 sum_avg_load_per_task = avg_load_per_task = 0;
3121 min_cpu_load = ~0UL;
3123 for_each_cpu_mask_nr(i, group->cpumask) {
3126 if (!cpu_isset(i, *cpus))
3131 if (*sd_idle && rq->nr_running)
3134 /* Bias balancing toward cpus of our domain */
3136 if (idle_cpu(i) && !first_idle_cpu) {
3141 load = target_load(i, load_idx);
3143 load = source_load(i, load_idx);
3144 if (load > max_cpu_load)
3145 max_cpu_load = load;
3146 if (min_cpu_load > load)
3147 min_cpu_load = load;
3151 sum_nr_running += rq->nr_running;
3152 sum_weighted_load += weighted_cpuload(i);
3154 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3158 * First idle cpu or the first cpu(busiest) in this sched group
3159 * is eligible for doing load balancing at this and above
3160 * domains. In the newly idle case, we will allow all the cpu's
3161 * to do the newly idle load balance.
3163 if (idle != CPU_NEWLY_IDLE && local_group &&
3164 balance_cpu != this_cpu && balance) {
3169 total_load += avg_load;
3170 total_pwr += group->__cpu_power;
3172 /* Adjust by relative CPU power of the group */
3173 avg_load = sg_div_cpu_power(group,
3174 avg_load * SCHED_LOAD_SCALE);
3178 * Consider the group unbalanced when the imbalance is larger
3179 * than the average weight of two tasks.
3181 * APZ: with cgroup the avg task weight can vary wildly and
3182 * might not be a suitable number - should we keep a
3183 * normalized nr_running number somewhere that negates
3186 avg_load_per_task = sg_div_cpu_power(group,
3187 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3189 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3192 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3195 this_load = avg_load;
3197 this_nr_running = sum_nr_running;
3198 this_load_per_task = sum_weighted_load;
3199 } else if (avg_load > max_load &&
3200 (sum_nr_running > group_capacity || __group_imb)) {
3201 max_load = avg_load;
3203 busiest_nr_running = sum_nr_running;
3204 busiest_load_per_task = sum_weighted_load;
3205 group_imb = __group_imb;
3208 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3210 * Busy processors will not participate in power savings
3213 if (idle == CPU_NOT_IDLE ||
3214 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3218 * If the local group is idle or completely loaded
3219 * no need to do power savings balance at this domain
3221 if (local_group && (this_nr_running >= group_capacity ||
3223 power_savings_balance = 0;
3226 * If a group is already running at full capacity or idle,
3227 * don't include that group in power savings calculations
3229 if (!power_savings_balance || sum_nr_running >= group_capacity
3234 * Calculate the group which has the least non-idle load.
3235 * This is the group from where we need to pick up the load
3238 if ((sum_nr_running < min_nr_running) ||
3239 (sum_nr_running == min_nr_running &&
3240 first_cpu(group->cpumask) <
3241 first_cpu(group_min->cpumask))) {
3243 min_nr_running = sum_nr_running;
3244 min_load_per_task = sum_weighted_load /
3249 * Calculate the group which is almost near its
3250 * capacity but still has some space to pick up some load
3251 * from other group and save more power
3253 if (sum_nr_running <= group_capacity - 1) {
3254 if (sum_nr_running > leader_nr_running ||
3255 (sum_nr_running == leader_nr_running &&
3256 first_cpu(group->cpumask) >
3257 first_cpu(group_leader->cpumask))) {
3258 group_leader = group;
3259 leader_nr_running = sum_nr_running;
3264 group = group->next;
3265 } while (group != sd->groups);
3267 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3270 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3272 if (this_load >= avg_load ||
3273 100*max_load <= sd->imbalance_pct*this_load)
3276 busiest_load_per_task /= busiest_nr_running;
3278 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3281 * We're trying to get all the cpus to the average_load, so we don't
3282 * want to push ourselves above the average load, nor do we wish to
3283 * reduce the max loaded cpu below the average load, as either of these
3284 * actions would just result in more rebalancing later, and ping-pong
3285 * tasks around. Thus we look for the minimum possible imbalance.
3286 * Negative imbalances (*we* are more loaded than anyone else) will
3287 * be counted as no imbalance for these purposes -- we can't fix that
3288 * by pulling tasks to us. Be careful of negative numbers as they'll
3289 * appear as very large values with unsigned longs.
3291 if (max_load <= busiest_load_per_task)
3295 * In the presence of smp nice balancing, certain scenarios can have
3296 * max load less than avg load(as we skip the groups at or below
3297 * its cpu_power, while calculating max_load..)
3299 if (max_load < avg_load) {
3301 goto small_imbalance;
3304 /* Don't want to pull so many tasks that a group would go idle */
3305 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3307 /* How much load to actually move to equalise the imbalance */
3308 *imbalance = min(max_pull * busiest->__cpu_power,
3309 (avg_load - this_load) * this->__cpu_power)
3313 * if *imbalance is less than the average load per runnable task
3314 * there is no gaurantee that any tasks will be moved so we'll have
3315 * a think about bumping its value to force at least one task to be
3318 if (*imbalance < busiest_load_per_task) {
3319 unsigned long tmp, pwr_now, pwr_move;
3323 pwr_move = pwr_now = 0;
3325 if (this_nr_running) {
3326 this_load_per_task /= this_nr_running;
3327 if (busiest_load_per_task > this_load_per_task)
3330 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3332 if (max_load - this_load + 2*busiest_load_per_task >=
3333 busiest_load_per_task * imbn) {
3334 *imbalance = busiest_load_per_task;
3339 * OK, we don't have enough imbalance to justify moving tasks,
3340 * however we may be able to increase total CPU power used by
3344 pwr_now += busiest->__cpu_power *
3345 min(busiest_load_per_task, max_load);
3346 pwr_now += this->__cpu_power *
3347 min(this_load_per_task, this_load);
3348 pwr_now /= SCHED_LOAD_SCALE;
3350 /* Amount of load we'd subtract */
3351 tmp = sg_div_cpu_power(busiest,
3352 busiest_load_per_task * SCHED_LOAD_SCALE);
3354 pwr_move += busiest->__cpu_power *
3355 min(busiest_load_per_task, max_load - tmp);
3357 /* Amount of load we'd add */
3358 if (max_load * busiest->__cpu_power <
3359 busiest_load_per_task * SCHED_LOAD_SCALE)
3360 tmp = sg_div_cpu_power(this,
3361 max_load * busiest->__cpu_power);
3363 tmp = sg_div_cpu_power(this,
3364 busiest_load_per_task * SCHED_LOAD_SCALE);
3365 pwr_move += this->__cpu_power *
3366 min(this_load_per_task, this_load + tmp);
3367 pwr_move /= SCHED_LOAD_SCALE;
3369 /* Move if we gain throughput */
3370 if (pwr_move > pwr_now)
3371 *imbalance = busiest_load_per_task;
3377 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3378 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3381 if (this == group_leader && group_leader != group_min) {
3382 *imbalance = min_load_per_task;
3392 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3395 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3396 unsigned long imbalance, const cpumask_t *cpus)
3398 struct rq *busiest = NULL, *rq;
3399 unsigned long max_load = 0;
3402 for_each_cpu_mask_nr(i, group->cpumask) {
3405 if (!cpu_isset(i, *cpus))
3409 wl = weighted_cpuload(i);
3411 if (rq->nr_running == 1 && wl > imbalance)
3414 if (wl > max_load) {
3424 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3425 * so long as it is large enough.
3427 #define MAX_PINNED_INTERVAL 512
3430 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3431 * tasks if there is an imbalance.
3433 static int load_balance(int this_cpu, struct rq *this_rq,
3434 struct sched_domain *sd, enum cpu_idle_type idle,
3435 int *balance, cpumask_t *cpus)
3437 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3438 struct sched_group *group;
3439 unsigned long imbalance;
3441 unsigned long flags;
3446 * When power savings policy is enabled for the parent domain, idle
3447 * sibling can pick up load irrespective of busy siblings. In this case,
3448 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3449 * portraying it as CPU_NOT_IDLE.
3451 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3452 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3455 schedstat_inc(sd, lb_count[idle]);
3459 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3466 schedstat_inc(sd, lb_nobusyg[idle]);
3470 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3472 schedstat_inc(sd, lb_nobusyq[idle]);
3476 BUG_ON(busiest == this_rq);
3478 schedstat_add(sd, lb_imbalance[idle], imbalance);
3481 if (busiest->nr_running > 1) {
3483 * Attempt to move tasks. If find_busiest_group has found
3484 * an imbalance but busiest->nr_running <= 1, the group is
3485 * still unbalanced. ld_moved simply stays zero, so it is
3486 * correctly treated as an imbalance.
3488 local_irq_save(flags);
3489 double_rq_lock(this_rq, busiest);
3490 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3491 imbalance, sd, idle, &all_pinned);
3492 double_rq_unlock(this_rq, busiest);
3493 local_irq_restore(flags);
3496 * some other cpu did the load balance for us.
3498 if (ld_moved && this_cpu != smp_processor_id())
3499 resched_cpu(this_cpu);
3501 /* All tasks on this runqueue were pinned by CPU affinity */
3502 if (unlikely(all_pinned)) {
3503 cpu_clear(cpu_of(busiest), *cpus);
3504 if (!cpus_empty(*cpus))
3511 schedstat_inc(sd, lb_failed[idle]);
3512 sd->nr_balance_failed++;
3514 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3516 spin_lock_irqsave(&busiest->lock, flags);
3518 /* don't kick the migration_thread, if the curr
3519 * task on busiest cpu can't be moved to this_cpu
3521 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3522 spin_unlock_irqrestore(&busiest->lock, flags);
3524 goto out_one_pinned;
3527 if (!busiest->active_balance) {
3528 busiest->active_balance = 1;
3529 busiest->push_cpu = this_cpu;
3532 spin_unlock_irqrestore(&busiest->lock, flags);
3534 wake_up_process(busiest->migration_thread);
3537 * We've kicked active balancing, reset the failure
3540 sd->nr_balance_failed = sd->cache_nice_tries+1;
3543 sd->nr_balance_failed = 0;
3545 if (likely(!active_balance)) {
3546 /* We were unbalanced, so reset the balancing interval */
3547 sd->balance_interval = sd->min_interval;
3550 * If we've begun active balancing, start to back off. This
3551 * case may not be covered by the all_pinned logic if there
3552 * is only 1 task on the busy runqueue (because we don't call
3555 if (sd->balance_interval < sd->max_interval)
3556 sd->balance_interval *= 2;
3559 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3560 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3566 schedstat_inc(sd, lb_balanced[idle]);
3568 sd->nr_balance_failed = 0;
3571 /* tune up the balancing interval */
3572 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3573 (sd->balance_interval < sd->max_interval))
3574 sd->balance_interval *= 2;
3576 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3577 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3588 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3589 * tasks if there is an imbalance.
3591 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3592 * this_rq is locked.
3595 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3598 struct sched_group *group;
3599 struct rq *busiest = NULL;
3600 unsigned long imbalance;
3608 * When power savings policy is enabled for the parent domain, idle
3609 * sibling can pick up load irrespective of busy siblings. In this case,
3610 * let the state of idle sibling percolate up as IDLE, instead of
3611 * portraying it as CPU_NOT_IDLE.
3613 if (sd->flags & SD_SHARE_CPUPOWER &&
3614 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3617 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3619 update_shares_locked(this_rq, sd);
3620 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3621 &sd_idle, cpus, NULL);
3623 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3627 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3629 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3633 BUG_ON(busiest == this_rq);
3635 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3638 if (busiest->nr_running > 1) {
3639 /* Attempt to move tasks */
3640 double_lock_balance(this_rq, busiest);
3641 /* this_rq->clock is already updated */
3642 update_rq_clock(busiest);
3643 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3644 imbalance, sd, CPU_NEWLY_IDLE,
3646 double_unlock_balance(this_rq, busiest);
3648 if (unlikely(all_pinned)) {
3649 cpu_clear(cpu_of(busiest), *cpus);
3650 if (!cpus_empty(*cpus))
3656 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3657 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3658 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3661 sd->nr_balance_failed = 0;
3663 update_shares_locked(this_rq, sd);
3667 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3668 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3669 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3671 sd->nr_balance_failed = 0;
3677 * idle_balance is called by schedule() if this_cpu is about to become
3678 * idle. Attempts to pull tasks from other CPUs.
3680 static void idle_balance(int this_cpu, struct rq *this_rq)
3682 struct sched_domain *sd;
3683 int pulled_task = -1;
3684 unsigned long next_balance = jiffies + HZ;
3687 for_each_domain(this_cpu, sd) {
3688 unsigned long interval;
3690 if (!(sd->flags & SD_LOAD_BALANCE))
3693 if (sd->flags & SD_BALANCE_NEWIDLE)
3694 /* If we've pulled tasks over stop searching: */
3695 pulled_task = load_balance_newidle(this_cpu, this_rq,
3698 interval = msecs_to_jiffies(sd->balance_interval);
3699 if (time_after(next_balance, sd->last_balance + interval))
3700 next_balance = sd->last_balance + interval;
3704 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3706 * We are going idle. next_balance may be set based on
3707 * a busy processor. So reset next_balance.
3709 this_rq->next_balance = next_balance;
3714 * active_load_balance is run by migration threads. It pushes running tasks
3715 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3716 * running on each physical CPU where possible, and avoids physical /
3717 * logical imbalances.
3719 * Called with busiest_rq locked.
3721 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3723 int target_cpu = busiest_rq->push_cpu;
3724 struct sched_domain *sd;
3725 struct rq *target_rq;
3727 /* Is there any task to move? */
3728 if (busiest_rq->nr_running <= 1)
3731 target_rq = cpu_rq(target_cpu);
3734 * This condition is "impossible", if it occurs
3735 * we need to fix it. Originally reported by
3736 * Bjorn Helgaas on a 128-cpu setup.
3738 BUG_ON(busiest_rq == target_rq);
3740 /* move a task from busiest_rq to target_rq */
3741 double_lock_balance(busiest_rq, target_rq);
3742 update_rq_clock(busiest_rq);
3743 update_rq_clock(target_rq);
3745 /* Search for an sd spanning us and the target CPU. */
3746 for_each_domain(target_cpu, sd) {
3747 if ((sd->flags & SD_LOAD_BALANCE) &&
3748 cpu_isset(busiest_cpu, sd->span))
3753 schedstat_inc(sd, alb_count);
3755 if (move_one_task(target_rq, target_cpu, busiest_rq,
3757 schedstat_inc(sd, alb_pushed);
3759 schedstat_inc(sd, alb_failed);
3761 double_unlock_balance(busiest_rq, target_rq);
3766 atomic_t load_balancer;
3768 } nohz ____cacheline_aligned = {
3769 .load_balancer = ATOMIC_INIT(-1),
3770 .cpu_mask = CPU_MASK_NONE,
3774 * This routine will try to nominate the ilb (idle load balancing)
3775 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3776 * load balancing on behalf of all those cpus. If all the cpus in the system
3777 * go into this tickless mode, then there will be no ilb owner (as there is
3778 * no need for one) and all the cpus will sleep till the next wakeup event
3781 * For the ilb owner, tick is not stopped. And this tick will be used
3782 * for idle load balancing. ilb owner will still be part of
3785 * While stopping the tick, this cpu will become the ilb owner if there
3786 * is no other owner. And will be the owner till that cpu becomes busy
3787 * or if all cpus in the system stop their ticks at which point
3788 * there is no need for ilb owner.
3790 * When the ilb owner becomes busy, it nominates another owner, during the
3791 * next busy scheduler_tick()
3793 int select_nohz_load_balancer(int stop_tick)
3795 int cpu = smp_processor_id();
3798 cpu_set(cpu, nohz.cpu_mask);
3799 cpu_rq(cpu)->in_nohz_recently = 1;
3802 * If we are going offline and still the leader, give up!
3804 if (!cpu_active(cpu) &&
3805 atomic_read(&nohz.load_balancer) == cpu) {
3806 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3811 /* time for ilb owner also to sleep */
3812 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3813 if (atomic_read(&nohz.load_balancer) == cpu)
3814 atomic_set(&nohz.load_balancer, -1);
3818 if (atomic_read(&nohz.load_balancer) == -1) {
3819 /* make me the ilb owner */
3820 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3822 } else if (atomic_read(&nohz.load_balancer) == cpu)
3825 if (!cpu_isset(cpu, nohz.cpu_mask))
3828 cpu_clear(cpu, nohz.cpu_mask);
3830 if (atomic_read(&nohz.load_balancer) == cpu)
3831 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3838 static DEFINE_SPINLOCK(balancing);
3841 * It checks each scheduling domain to see if it is due to be balanced,
3842 * and initiates a balancing operation if so.
3844 * Balancing parameters are set up in arch_init_sched_domains.
3846 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3849 struct rq *rq = cpu_rq(cpu);
3850 unsigned long interval;
3851 struct sched_domain *sd;
3852 /* Earliest time when we have to do rebalance again */
3853 unsigned long next_balance = jiffies + 60*HZ;
3854 int update_next_balance = 0;
3858 for_each_domain(cpu, sd) {
3859 if (!(sd->flags & SD_LOAD_BALANCE))
3862 interval = sd->balance_interval;
3863 if (idle != CPU_IDLE)
3864 interval *= sd->busy_factor;
3866 /* scale ms to jiffies */
3867 interval = msecs_to_jiffies(interval);
3868 if (unlikely(!interval))
3870 if (interval > HZ*NR_CPUS/10)
3871 interval = HZ*NR_CPUS/10;
3873 need_serialize = sd->flags & SD_SERIALIZE;
3875 if (need_serialize) {
3876 if (!spin_trylock(&balancing))
3880 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3881 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3883 * We've pulled tasks over so either we're no
3884 * longer idle, or one of our SMT siblings is
3887 idle = CPU_NOT_IDLE;
3889 sd->last_balance = jiffies;
3892 spin_unlock(&balancing);
3894 if (time_after(next_balance, sd->last_balance + interval)) {
3895 next_balance = sd->last_balance + interval;
3896 update_next_balance = 1;
3900 * Stop the load balance at this level. There is another
3901 * CPU in our sched group which is doing load balancing more
3909 * next_balance will be updated only when there is a need.
3910 * When the cpu is attached to null domain for ex, it will not be
3913 if (likely(update_next_balance))
3914 rq->next_balance = next_balance;
3918 * run_rebalance_domains is triggered when needed from the scheduler tick.
3919 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3920 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3922 static void run_rebalance_domains(struct softirq_action *h)
3924 int this_cpu = smp_processor_id();
3925 struct rq *this_rq = cpu_rq(this_cpu);
3926 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3927 CPU_IDLE : CPU_NOT_IDLE;
3929 rebalance_domains(this_cpu, idle);
3933 * If this cpu is the owner for idle load balancing, then do the
3934 * balancing on behalf of the other idle cpus whose ticks are
3937 if (this_rq->idle_at_tick &&
3938 atomic_read(&nohz.load_balancer) == this_cpu) {
3939 cpumask_t cpus = nohz.cpu_mask;
3943 cpu_clear(this_cpu, cpus);
3944 for_each_cpu_mask_nr(balance_cpu, cpus) {
3946 * If this cpu gets work to do, stop the load balancing
3947 * work being done for other cpus. Next load
3948 * balancing owner will pick it up.
3953 rebalance_domains(balance_cpu, CPU_IDLE);
3955 rq = cpu_rq(balance_cpu);
3956 if (time_after(this_rq->next_balance, rq->next_balance))
3957 this_rq->next_balance = rq->next_balance;
3964 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3966 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3967 * idle load balancing owner or decide to stop the periodic load balancing,
3968 * if the whole system is idle.
3970 static inline void trigger_load_balance(struct rq *rq, int cpu)
3974 * If we were in the nohz mode recently and busy at the current
3975 * scheduler tick, then check if we need to nominate new idle
3978 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3979 rq->in_nohz_recently = 0;
3981 if (atomic_read(&nohz.load_balancer) == cpu) {
3982 cpu_clear(cpu, nohz.cpu_mask);
3983 atomic_set(&nohz.load_balancer, -1);
3986 if (atomic_read(&nohz.load_balancer) == -1) {
3988 * simple selection for now: Nominate the
3989 * first cpu in the nohz list to be the next
3992 * TBD: Traverse the sched domains and nominate
3993 * the nearest cpu in the nohz.cpu_mask.
3995 int ilb = first_cpu(nohz.cpu_mask);
3997 if (ilb < nr_cpu_ids)
4003 * If this cpu is idle and doing idle load balancing for all the
4004 * cpus with ticks stopped, is it time for that to stop?
4006 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4007 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4013 * If this cpu is idle and the idle load balancing is done by
4014 * someone else, then no need raise the SCHED_SOFTIRQ
4016 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4017 cpu_isset(cpu, nohz.cpu_mask))
4020 if (time_after_eq(jiffies, rq->next_balance))
4021 raise_softirq(SCHED_SOFTIRQ);
4024 #else /* CONFIG_SMP */
4027 * on UP we do not need to balance between CPUs:
4029 static inline void idle_balance(int cpu, struct rq *rq)
4035 DEFINE_PER_CPU(struct kernel_stat, kstat);
4037 EXPORT_PER_CPU_SYMBOL(kstat);
4040 * Return any ns on the sched_clock that have not yet been banked in
4041 * @p in case that task is currently running.
4043 unsigned long long task_delta_exec(struct task_struct *p)
4045 unsigned long flags;
4049 rq = task_rq_lock(p, &flags);
4051 if (task_current(rq, p)) {
4054 update_rq_clock(rq);
4055 delta_exec = rq->clock - p->se.exec_start;
4056 if ((s64)delta_exec > 0)
4060 task_rq_unlock(rq, &flags);
4066 * Account user cpu time to a process.
4067 * @p: the process that the cpu time gets accounted to
4068 * @cputime: the cpu time spent in user space since the last update
4070 void account_user_time(struct task_struct *p, cputime_t cputime)
4072 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4075 p->utime = cputime_add(p->utime, cputime);
4076 account_group_user_time(p, cputime);
4078 /* Add user time to cpustat. */
4079 tmp = cputime_to_cputime64(cputime);
4080 if (TASK_NICE(p) > 0)
4081 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4083 cpustat->user = cputime64_add(cpustat->user, tmp);
4084 /* Account for user time used */
4085 acct_update_integrals(p);
4089 * Account guest cpu time to a process.
4090 * @p: the process that the cpu time gets accounted to
4091 * @cputime: the cpu time spent in virtual machine since the last update
4093 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4096 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4098 tmp = cputime_to_cputime64(cputime);
4100 p->utime = cputime_add(p->utime, cputime);
4101 account_group_user_time(p, cputime);
4102 p->gtime = cputime_add(p->gtime, cputime);
4104 cpustat->user = cputime64_add(cpustat->user, tmp);
4105 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4109 * Account scaled user cpu time to a process.
4110 * @p: the process that the cpu time gets accounted to
4111 * @cputime: the cpu time spent in user space since the last update
4113 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4115 p->utimescaled = cputime_add(p->utimescaled, cputime);
4119 * Account system cpu time to a process.
4120 * @p: the process that the cpu time gets accounted to
4121 * @hardirq_offset: the offset to subtract from hardirq_count()
4122 * @cputime: the cpu time spent in kernel space since the last update
4124 void account_system_time(struct task_struct *p, int hardirq_offset,
4127 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4128 struct rq *rq = this_rq();
4131 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4132 account_guest_time(p, cputime);
4136 p->stime = cputime_add(p->stime, cputime);
4137 account_group_system_time(p, cputime);
4139 /* Add system time to cpustat. */
4140 tmp = cputime_to_cputime64(cputime);
4141 if (hardirq_count() - hardirq_offset)
4142 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4143 else if (softirq_count())
4144 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4145 else if (p != rq->idle)
4146 cpustat->system = cputime64_add(cpustat->system, tmp);
4147 else if (atomic_read(&rq->nr_iowait) > 0)
4148 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4150 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4151 /* Account for system time used */
4152 acct_update_integrals(p);
4156 * Account scaled system cpu time to a process.
4157 * @p: the process that the cpu time gets accounted to
4158 * @hardirq_offset: the offset to subtract from hardirq_count()
4159 * @cputime: the cpu time spent in kernel space since the last update
4161 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4163 p->stimescaled = cputime_add(p->stimescaled, cputime);
4167 * Account for involuntary wait time.
4168 * @p: the process from which the cpu time has been stolen
4169 * @steal: the cpu time spent in involuntary wait
4171 void account_steal_time(struct task_struct *p, cputime_t steal)
4173 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4174 cputime64_t tmp = cputime_to_cputime64(steal);
4175 struct rq *rq = this_rq();
4177 if (p == rq->idle) {
4178 p->stime = cputime_add(p->stime, steal);
4179 account_group_system_time(p, steal);
4180 if (atomic_read(&rq->nr_iowait) > 0)
4181 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4183 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4185 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4189 * Use precise platform statistics if available:
4191 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4192 cputime_t task_utime(struct task_struct *p)
4197 cputime_t task_stime(struct task_struct *p)
4202 cputime_t task_utime(struct task_struct *p)
4204 clock_t utime = cputime_to_clock_t(p->utime),
4205 total = utime + cputime_to_clock_t(p->stime);
4209 * Use CFS's precise accounting:
4211 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4215 do_div(temp, total);
4217 utime = (clock_t)temp;
4219 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4220 return p->prev_utime;
4223 cputime_t task_stime(struct task_struct *p)
4228 * Use CFS's precise accounting. (we subtract utime from
4229 * the total, to make sure the total observed by userspace
4230 * grows monotonically - apps rely on that):
4232 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4233 cputime_to_clock_t(task_utime(p));
4236 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4238 return p->prev_stime;
4242 inline cputime_t task_gtime(struct task_struct *p)
4248 * This function gets called by the timer code, with HZ frequency.
4249 * We call it with interrupts disabled.
4251 * It also gets called by the fork code, when changing the parent's
4254 void scheduler_tick(void)
4256 int cpu = smp_processor_id();
4257 struct rq *rq = cpu_rq(cpu);
4258 struct task_struct *curr = rq->curr;
4262 spin_lock(&rq->lock);
4263 update_rq_clock(rq);
4264 update_cpu_load(rq);
4265 curr->sched_class->task_tick(rq, curr, 0);
4266 spin_unlock(&rq->lock);
4269 rq->idle_at_tick = idle_cpu(cpu);
4270 trigger_load_balance(rq, cpu);
4274 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4275 defined(CONFIG_PREEMPT_TRACER))
4277 static inline unsigned long get_parent_ip(unsigned long addr)
4279 if (in_lock_functions(addr)) {
4280 addr = CALLER_ADDR2;
4281 if (in_lock_functions(addr))
4282 addr = CALLER_ADDR3;
4287 void __kprobes add_preempt_count(int val)
4289 #ifdef CONFIG_DEBUG_PREEMPT
4293 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4296 preempt_count() += val;
4297 #ifdef CONFIG_DEBUG_PREEMPT
4299 * Spinlock count overflowing soon?
4301 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4304 if (preempt_count() == val)
4305 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4307 EXPORT_SYMBOL(add_preempt_count);
4309 void __kprobes sub_preempt_count(int val)
4311 #ifdef CONFIG_DEBUG_PREEMPT
4315 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4318 * Is the spinlock portion underflowing?
4320 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4321 !(preempt_count() & PREEMPT_MASK)))
4325 if (preempt_count() == val)
4326 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4327 preempt_count() -= val;
4329 EXPORT_SYMBOL(sub_preempt_count);
4334 * Print scheduling while atomic bug:
4336 static noinline void __schedule_bug(struct task_struct *prev)
4338 struct pt_regs *regs = get_irq_regs();
4340 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4341 prev->comm, prev->pid, preempt_count());
4343 debug_show_held_locks(prev);
4345 if (irqs_disabled())
4346 print_irqtrace_events(prev);
4355 * Various schedule()-time debugging checks and statistics:
4357 static inline void schedule_debug(struct task_struct *prev)
4360 * Test if we are atomic. Since do_exit() needs to call into
4361 * schedule() atomically, we ignore that path for now.
4362 * Otherwise, whine if we are scheduling when we should not be.
4364 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4365 __schedule_bug(prev);
4367 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4369 schedstat_inc(this_rq(), sched_count);
4370 #ifdef CONFIG_SCHEDSTATS
4371 if (unlikely(prev->lock_depth >= 0)) {
4372 schedstat_inc(this_rq(), bkl_count);
4373 schedstat_inc(prev, sched_info.bkl_count);
4379 * Pick up the highest-prio task:
4381 static inline struct task_struct *
4382 pick_next_task(struct rq *rq, struct task_struct *prev)
4384 const struct sched_class *class;
4385 struct task_struct *p;
4388 * Optimization: we know that if all tasks are in
4389 * the fair class we can call that function directly:
4391 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4392 p = fair_sched_class.pick_next_task(rq);
4397 class = sched_class_highest;
4399 p = class->pick_next_task(rq);
4403 * Will never be NULL as the idle class always
4404 * returns a non-NULL p:
4406 class = class->next;
4411 * schedule() is the main scheduler function.
4413 asmlinkage void __sched schedule(void)
4415 struct task_struct *prev, *next;
4416 unsigned long *switch_count;
4422 cpu = smp_processor_id();
4426 switch_count = &prev->nivcsw;
4428 release_kernel_lock(prev);
4429 need_resched_nonpreemptible:
4431 schedule_debug(prev);
4433 if (sched_feat(HRTICK))
4437 * Do the rq-clock update outside the rq lock:
4439 local_irq_disable();
4440 update_rq_clock(rq);
4441 spin_lock(&rq->lock);
4442 clear_tsk_need_resched(prev);
4444 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4445 if (unlikely(signal_pending_state(prev->state, prev)))
4446 prev->state = TASK_RUNNING;
4448 deactivate_task(rq, prev, 1);
4449 switch_count = &prev->nvcsw;
4453 if (prev->sched_class->pre_schedule)
4454 prev->sched_class->pre_schedule(rq, prev);
4457 if (unlikely(!rq->nr_running))
4458 idle_balance(cpu, rq);
4460 prev->sched_class->put_prev_task(rq, prev);
4461 next = pick_next_task(rq, prev);
4463 if (likely(prev != next)) {
4464 sched_info_switch(prev, next);
4470 context_switch(rq, prev, next); /* unlocks the rq */
4472 * the context switch might have flipped the stack from under
4473 * us, hence refresh the local variables.
4475 cpu = smp_processor_id();
4478 spin_unlock_irq(&rq->lock);
4480 if (unlikely(reacquire_kernel_lock(current) < 0))
4481 goto need_resched_nonpreemptible;
4483 preempt_enable_no_resched();
4484 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4487 EXPORT_SYMBOL(schedule);
4489 #ifdef CONFIG_PREEMPT
4491 * this is the entry point to schedule() from in-kernel preemption
4492 * off of preempt_enable. Kernel preemptions off return from interrupt
4493 * occur there and call schedule directly.
4495 asmlinkage void __sched preempt_schedule(void)
4497 struct thread_info *ti = current_thread_info();
4500 * If there is a non-zero preempt_count or interrupts are disabled,
4501 * we do not want to preempt the current task. Just return..
4503 if (likely(ti->preempt_count || irqs_disabled()))
4507 add_preempt_count(PREEMPT_ACTIVE);
4509 sub_preempt_count(PREEMPT_ACTIVE);
4512 * Check again in case we missed a preemption opportunity
4513 * between schedule and now.
4516 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4518 EXPORT_SYMBOL(preempt_schedule);
4521 * this is the entry point to schedule() from kernel preemption
4522 * off of irq context.
4523 * Note, that this is called and return with irqs disabled. This will
4524 * protect us against recursive calling from irq.
4526 asmlinkage void __sched preempt_schedule_irq(void)
4528 struct thread_info *ti = current_thread_info();
4530 /* Catch callers which need to be fixed */
4531 BUG_ON(ti->preempt_count || !irqs_disabled());
4534 add_preempt_count(PREEMPT_ACTIVE);
4537 local_irq_disable();
4538 sub_preempt_count(PREEMPT_ACTIVE);
4541 * Check again in case we missed a preemption opportunity
4542 * between schedule and now.
4545 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4548 #endif /* CONFIG_PREEMPT */
4550 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4553 return try_to_wake_up(curr->private, mode, sync);
4555 EXPORT_SYMBOL(default_wake_function);
4558 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4559 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4560 * number) then we wake all the non-exclusive tasks and one exclusive task.
4562 * There are circumstances in which we can try to wake a task which has already
4563 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4564 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4566 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4567 int nr_exclusive, int sync, void *key)
4569 wait_queue_t *curr, *next;
4571 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4572 unsigned flags = curr->flags;
4574 if (curr->func(curr, mode, sync, key) &&
4575 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4581 * __wake_up - wake up threads blocked on a waitqueue.
4583 * @mode: which threads
4584 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4585 * @key: is directly passed to the wakeup function
4587 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4588 int nr_exclusive, void *key)
4590 unsigned long flags;
4592 spin_lock_irqsave(&q->lock, flags);
4593 __wake_up_common(q, mode, nr_exclusive, 0, key);
4594 spin_unlock_irqrestore(&q->lock, flags);
4596 EXPORT_SYMBOL(__wake_up);
4599 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4601 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4603 __wake_up_common(q, mode, 1, 0, NULL);
4607 * __wake_up_sync - wake up threads blocked on a waitqueue.
4609 * @mode: which threads
4610 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4612 * The sync wakeup differs that the waker knows that it will schedule
4613 * away soon, so while the target thread will be woken up, it will not
4614 * be migrated to another CPU - ie. the two threads are 'synchronized'
4615 * with each other. This can prevent needless bouncing between CPUs.
4617 * On UP it can prevent extra preemption.
4620 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4622 unsigned long flags;
4628 if (unlikely(!nr_exclusive))
4631 spin_lock_irqsave(&q->lock, flags);
4632 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4633 spin_unlock_irqrestore(&q->lock, flags);
4635 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4637 void complete(struct completion *x)
4639 unsigned long flags;
4641 spin_lock_irqsave(&x->wait.lock, flags);
4643 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4644 spin_unlock_irqrestore(&x->wait.lock, flags);
4646 EXPORT_SYMBOL(complete);
4648 void complete_all(struct completion *x)
4650 unsigned long flags;
4652 spin_lock_irqsave(&x->wait.lock, flags);
4653 x->done += UINT_MAX/2;
4654 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4655 spin_unlock_irqrestore(&x->wait.lock, flags);
4657 EXPORT_SYMBOL(complete_all);
4659 static inline long __sched
4660 do_wait_for_common(struct completion *x, long timeout, int state)
4663 DECLARE_WAITQUEUE(wait, current);
4665 wait.flags |= WQ_FLAG_EXCLUSIVE;
4666 __add_wait_queue_tail(&x->wait, &wait);
4668 if ((state == TASK_INTERRUPTIBLE &&
4669 signal_pending(current)) ||
4670 (state == TASK_KILLABLE &&
4671 fatal_signal_pending(current))) {
4672 timeout = -ERESTARTSYS;
4675 __set_current_state(state);
4676 spin_unlock_irq(&x->wait.lock);
4677 timeout = schedule_timeout(timeout);
4678 spin_lock_irq(&x->wait.lock);
4679 } while (!x->done && timeout);
4680 __remove_wait_queue(&x->wait, &wait);
4685 return timeout ?: 1;
4689 wait_for_common(struct completion *x, long timeout, int state)
4693 spin_lock_irq(&x->wait.lock);
4694 timeout = do_wait_for_common(x, timeout, state);
4695 spin_unlock_irq(&x->wait.lock);
4699 void __sched wait_for_completion(struct completion *x)
4701 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4703 EXPORT_SYMBOL(wait_for_completion);
4705 unsigned long __sched
4706 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4708 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4710 EXPORT_SYMBOL(wait_for_completion_timeout);
4712 int __sched wait_for_completion_interruptible(struct completion *x)
4714 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4715 if (t == -ERESTARTSYS)
4719 EXPORT_SYMBOL(wait_for_completion_interruptible);
4721 unsigned long __sched
4722 wait_for_completion_interruptible_timeout(struct completion *x,
4723 unsigned long timeout)
4725 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4727 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4729 int __sched wait_for_completion_killable(struct completion *x)
4731 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4732 if (t == -ERESTARTSYS)
4736 EXPORT_SYMBOL(wait_for_completion_killable);
4739 * try_wait_for_completion - try to decrement a completion without blocking
4740 * @x: completion structure
4742 * Returns: 0 if a decrement cannot be done without blocking
4743 * 1 if a decrement succeeded.
4745 * If a completion is being used as a counting completion,
4746 * attempt to decrement the counter without blocking. This
4747 * enables us to avoid waiting if the resource the completion
4748 * is protecting is not available.
4750 bool try_wait_for_completion(struct completion *x)
4754 spin_lock_irq(&x->wait.lock);
4759 spin_unlock_irq(&x->wait.lock);
4762 EXPORT_SYMBOL(try_wait_for_completion);
4765 * completion_done - Test to see if a completion has any waiters
4766 * @x: completion structure
4768 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4769 * 1 if there are no waiters.
4772 bool completion_done(struct completion *x)
4776 spin_lock_irq(&x->wait.lock);
4779 spin_unlock_irq(&x->wait.lock);
4782 EXPORT_SYMBOL(completion_done);
4785 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4787 unsigned long flags;
4790 init_waitqueue_entry(&wait, current);
4792 __set_current_state(state);
4794 spin_lock_irqsave(&q->lock, flags);
4795 __add_wait_queue(q, &wait);
4796 spin_unlock(&q->lock);
4797 timeout = schedule_timeout(timeout);
4798 spin_lock_irq(&q->lock);
4799 __remove_wait_queue(q, &wait);
4800 spin_unlock_irqrestore(&q->lock, flags);
4805 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4807 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4809 EXPORT_SYMBOL(interruptible_sleep_on);
4812 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4814 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4816 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4818 void __sched sleep_on(wait_queue_head_t *q)
4820 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4822 EXPORT_SYMBOL(sleep_on);
4824 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4826 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4828 EXPORT_SYMBOL(sleep_on_timeout);
4830 #ifdef CONFIG_RT_MUTEXES
4833 * rt_mutex_setprio - set the current priority of a task
4835 * @prio: prio value (kernel-internal form)
4837 * This function changes the 'effective' priority of a task. It does
4838 * not touch ->normal_prio like __setscheduler().
4840 * Used by the rt_mutex code to implement priority inheritance logic.
4842 void rt_mutex_setprio(struct task_struct *p, int prio)
4844 unsigned long flags;
4845 int oldprio, on_rq, running;
4847 const struct sched_class *prev_class = p->sched_class;
4849 BUG_ON(prio < 0 || prio > MAX_PRIO);
4851 rq = task_rq_lock(p, &flags);
4852 update_rq_clock(rq);
4855 on_rq = p->se.on_rq;
4856 running = task_current(rq, p);
4858 dequeue_task(rq, p, 0);
4860 p->sched_class->put_prev_task(rq, p);
4863 p->sched_class = &rt_sched_class;
4865 p->sched_class = &fair_sched_class;
4870 p->sched_class->set_curr_task(rq);
4872 enqueue_task(rq, p, 0);
4874 check_class_changed(rq, p, prev_class, oldprio, running);
4876 task_rq_unlock(rq, &flags);
4881 void set_user_nice(struct task_struct *p, long nice)
4883 int old_prio, delta, on_rq;
4884 unsigned long flags;
4887 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4890 * We have to be careful, if called from sys_setpriority(),
4891 * the task might be in the middle of scheduling on another CPU.
4893 rq = task_rq_lock(p, &flags);
4894 update_rq_clock(rq);
4896 * The RT priorities are set via sched_setscheduler(), but we still
4897 * allow the 'normal' nice value to be set - but as expected
4898 * it wont have any effect on scheduling until the task is
4899 * SCHED_FIFO/SCHED_RR:
4901 if (task_has_rt_policy(p)) {
4902 p->static_prio = NICE_TO_PRIO(nice);
4905 on_rq = p->se.on_rq;
4907 dequeue_task(rq, p, 0);
4909 p->static_prio = NICE_TO_PRIO(nice);
4912 p->prio = effective_prio(p);
4913 delta = p->prio - old_prio;
4916 enqueue_task(rq, p, 0);
4918 * If the task increased its priority or is running and
4919 * lowered its priority, then reschedule its CPU:
4921 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4922 resched_task(rq->curr);
4925 task_rq_unlock(rq, &flags);
4927 EXPORT_SYMBOL(set_user_nice);
4930 * can_nice - check if a task can reduce its nice value
4934 int can_nice(const struct task_struct *p, const int nice)
4936 /* convert nice value [19,-20] to rlimit style value [1,40] */
4937 int nice_rlim = 20 - nice;
4939 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4940 capable(CAP_SYS_NICE));
4943 #ifdef __ARCH_WANT_SYS_NICE
4946 * sys_nice - change the priority of the current process.
4947 * @increment: priority increment
4949 * sys_setpriority is a more generic, but much slower function that
4950 * does similar things.
4952 asmlinkage long sys_nice(int increment)
4957 * Setpriority might change our priority at the same moment.
4958 * We don't have to worry. Conceptually one call occurs first
4959 * and we have a single winner.
4961 if (increment < -40)
4966 nice = PRIO_TO_NICE(current->static_prio) + increment;
4972 if (increment < 0 && !can_nice(current, nice))
4975 retval = security_task_setnice(current, nice);
4979 set_user_nice(current, nice);
4986 * task_prio - return the priority value of a given task.
4987 * @p: the task in question.
4989 * This is the priority value as seen by users in /proc.
4990 * RT tasks are offset by -200. Normal tasks are centered
4991 * around 0, value goes from -16 to +15.
4993 int task_prio(const struct task_struct *p)
4995 return p->prio - MAX_RT_PRIO;
4999 * task_nice - return the nice value of a given task.
5000 * @p: the task in question.
5002 int task_nice(const struct task_struct *p)
5004 return TASK_NICE(p);
5006 EXPORT_SYMBOL(task_nice);
5009 * idle_cpu - is a given cpu idle currently?
5010 * @cpu: the processor in question.
5012 int idle_cpu(int cpu)
5014 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5018 * idle_task - return the idle task for a given cpu.
5019 * @cpu: the processor in question.
5021 struct task_struct *idle_task(int cpu)
5023 return cpu_rq(cpu)->idle;
5027 * find_process_by_pid - find a process with a matching PID value.
5028 * @pid: the pid in question.
5030 static struct task_struct *find_process_by_pid(pid_t pid)
5032 return pid ? find_task_by_vpid(pid) : current;
5035 /* Actually do priority change: must hold rq lock. */
5037 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5039 BUG_ON(p->se.on_rq);
5042 switch (p->policy) {
5046 p->sched_class = &fair_sched_class;
5050 p->sched_class = &rt_sched_class;
5054 p->rt_priority = prio;
5055 p->normal_prio = normal_prio(p);
5056 /* we are holding p->pi_lock already */
5057 p->prio = rt_mutex_getprio(p);
5061 static int __sched_setscheduler(struct task_struct *p, int policy,
5062 struct sched_param *param, bool user)
5064 int retval, oldprio, oldpolicy = -1, on_rq, running;
5065 unsigned long flags;
5066 const struct sched_class *prev_class = p->sched_class;
5069 /* may grab non-irq protected spin_locks */
5070 BUG_ON(in_interrupt());
5072 /* double check policy once rq lock held */
5074 policy = oldpolicy = p->policy;
5075 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5076 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5077 policy != SCHED_IDLE)
5080 * Valid priorities for SCHED_FIFO and SCHED_RR are
5081 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5082 * SCHED_BATCH and SCHED_IDLE is 0.
5084 if (param->sched_priority < 0 ||
5085 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5086 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5088 if (rt_policy(policy) != (param->sched_priority != 0))
5092 * Allow unprivileged RT tasks to decrease priority:
5094 if (user && !capable(CAP_SYS_NICE)) {
5095 if (rt_policy(policy)) {
5096 unsigned long rlim_rtprio;
5098 if (!lock_task_sighand(p, &flags))
5100 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5101 unlock_task_sighand(p, &flags);
5103 /* can't set/change the rt policy */
5104 if (policy != p->policy && !rlim_rtprio)
5107 /* can't increase priority */
5108 if (param->sched_priority > p->rt_priority &&
5109 param->sched_priority > rlim_rtprio)
5113 * Like positive nice levels, dont allow tasks to
5114 * move out of SCHED_IDLE either:
5116 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5119 /* can't change other user's priorities */
5120 if ((current->euid != p->euid) &&
5121 (current->euid != p->uid))
5126 #ifdef CONFIG_RT_GROUP_SCHED
5128 * Do not allow realtime tasks into groups that have no runtime
5131 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5135 retval = security_task_setscheduler(p, policy, param);
5141 * make sure no PI-waiters arrive (or leave) while we are
5142 * changing the priority of the task:
5144 spin_lock_irqsave(&p->pi_lock, flags);
5146 * To be able to change p->policy safely, the apropriate
5147 * runqueue lock must be held.
5149 rq = __task_rq_lock(p);
5150 /* recheck policy now with rq lock held */
5151 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5152 policy = oldpolicy = -1;
5153 __task_rq_unlock(rq);
5154 spin_unlock_irqrestore(&p->pi_lock, flags);
5157 update_rq_clock(rq);
5158 on_rq = p->se.on_rq;
5159 running = task_current(rq, p);
5161 deactivate_task(rq, p, 0);
5163 p->sched_class->put_prev_task(rq, p);
5166 __setscheduler(rq, p, policy, param->sched_priority);
5169 p->sched_class->set_curr_task(rq);
5171 activate_task(rq, p, 0);
5173 check_class_changed(rq, p, prev_class, oldprio, running);
5175 __task_rq_unlock(rq);
5176 spin_unlock_irqrestore(&p->pi_lock, flags);
5178 rt_mutex_adjust_pi(p);
5184 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5185 * @p: the task in question.
5186 * @policy: new policy.
5187 * @param: structure containing the new RT priority.
5189 * NOTE that the task may be already dead.
5191 int sched_setscheduler(struct task_struct *p, int policy,
5192 struct sched_param *param)
5194 return __sched_setscheduler(p, policy, param, true);
5196 EXPORT_SYMBOL_GPL(sched_setscheduler);
5199 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5200 * @p: the task in question.
5201 * @policy: new policy.
5202 * @param: structure containing the new RT priority.
5204 * Just like sched_setscheduler, only don't bother checking if the
5205 * current context has permission. For example, this is needed in
5206 * stop_machine(): we create temporary high priority worker threads,
5207 * but our caller might not have that capability.
5209 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5210 struct sched_param *param)
5212 return __sched_setscheduler(p, policy, param, false);
5216 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5218 struct sched_param lparam;
5219 struct task_struct *p;
5222 if (!param || pid < 0)
5224 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5229 p = find_process_by_pid(pid);
5231 retval = sched_setscheduler(p, policy, &lparam);
5238 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5239 * @pid: the pid in question.
5240 * @policy: new policy.
5241 * @param: structure containing the new RT priority.
5244 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5246 /* negative values for policy are not valid */
5250 return do_sched_setscheduler(pid, policy, param);
5254 * sys_sched_setparam - set/change the RT priority of a thread
5255 * @pid: the pid in question.
5256 * @param: structure containing the new RT priority.
5258 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5260 return do_sched_setscheduler(pid, -1, param);
5264 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5265 * @pid: the pid in question.
5267 asmlinkage long sys_sched_getscheduler(pid_t pid)
5269 struct task_struct *p;
5276 read_lock(&tasklist_lock);
5277 p = find_process_by_pid(pid);
5279 retval = security_task_getscheduler(p);
5283 read_unlock(&tasklist_lock);
5288 * sys_sched_getscheduler - get the RT priority of a thread
5289 * @pid: the pid in question.
5290 * @param: structure containing the RT priority.
5292 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5294 struct sched_param lp;
5295 struct task_struct *p;
5298 if (!param || pid < 0)
5301 read_lock(&tasklist_lock);
5302 p = find_process_by_pid(pid);
5307 retval = security_task_getscheduler(p);
5311 lp.sched_priority = p->rt_priority;
5312 read_unlock(&tasklist_lock);
5315 * This one might sleep, we cannot do it with a spinlock held ...
5317 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5322 read_unlock(&tasklist_lock);
5326 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5328 cpumask_t cpus_allowed;
5329 cpumask_t new_mask = *in_mask;
5330 struct task_struct *p;
5334 read_lock(&tasklist_lock);
5336 p = find_process_by_pid(pid);
5338 read_unlock(&tasklist_lock);
5344 * It is not safe to call set_cpus_allowed with the
5345 * tasklist_lock held. We will bump the task_struct's
5346 * usage count and then drop tasklist_lock.
5349 read_unlock(&tasklist_lock);
5352 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5353 !capable(CAP_SYS_NICE))
5356 retval = security_task_setscheduler(p, 0, NULL);
5360 cpuset_cpus_allowed(p, &cpus_allowed);
5361 cpus_and(new_mask, new_mask, cpus_allowed);
5363 retval = set_cpus_allowed_ptr(p, &new_mask);
5366 cpuset_cpus_allowed(p, &cpus_allowed);
5367 if (!cpus_subset(new_mask, cpus_allowed)) {
5369 * We must have raced with a concurrent cpuset
5370 * update. Just reset the cpus_allowed to the
5371 * cpuset's cpus_allowed
5373 new_mask = cpus_allowed;
5383 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5384 cpumask_t *new_mask)
5386 if (len < sizeof(cpumask_t)) {
5387 memset(new_mask, 0, sizeof(cpumask_t));
5388 } else if (len > sizeof(cpumask_t)) {
5389 len = sizeof(cpumask_t);
5391 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5395 * sys_sched_setaffinity - set the cpu affinity of a process
5396 * @pid: pid of the process
5397 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5398 * @user_mask_ptr: user-space pointer to the new cpu mask
5400 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5401 unsigned long __user *user_mask_ptr)
5406 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5410 return sched_setaffinity(pid, &new_mask);
5413 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5415 struct task_struct *p;
5419 read_lock(&tasklist_lock);
5422 p = find_process_by_pid(pid);
5426 retval = security_task_getscheduler(p);
5430 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5433 read_unlock(&tasklist_lock);
5440 * sys_sched_getaffinity - get the cpu affinity of a process
5441 * @pid: pid of the process
5442 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5443 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5445 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5446 unsigned long __user *user_mask_ptr)
5451 if (len < sizeof(cpumask_t))
5454 ret = sched_getaffinity(pid, &mask);
5458 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5461 return sizeof(cpumask_t);
5465 * sys_sched_yield - yield the current processor to other threads.
5467 * This function yields the current CPU to other tasks. If there are no
5468 * other threads running on this CPU then this function will return.
5470 asmlinkage long sys_sched_yield(void)
5472 struct rq *rq = this_rq_lock();
5474 schedstat_inc(rq, yld_count);
5475 current->sched_class->yield_task(rq);
5478 * Since we are going to call schedule() anyway, there's
5479 * no need to preempt or enable interrupts:
5481 __release(rq->lock);
5482 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5483 _raw_spin_unlock(&rq->lock);
5484 preempt_enable_no_resched();
5491 static void __cond_resched(void)
5493 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5494 __might_sleep(__FILE__, __LINE__);
5497 * The BKS might be reacquired before we have dropped
5498 * PREEMPT_ACTIVE, which could trigger a second
5499 * cond_resched() call.
5502 add_preempt_count(PREEMPT_ACTIVE);
5504 sub_preempt_count(PREEMPT_ACTIVE);
5505 } while (need_resched());
5508 int __sched _cond_resched(void)
5510 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5511 system_state == SYSTEM_RUNNING) {
5517 EXPORT_SYMBOL(_cond_resched);
5520 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5521 * call schedule, and on return reacquire the lock.
5523 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5524 * operations here to prevent schedule() from being called twice (once via
5525 * spin_unlock(), once by hand).
5527 int cond_resched_lock(spinlock_t *lock)
5529 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5532 if (spin_needbreak(lock) || resched) {
5534 if (resched && need_resched())
5543 EXPORT_SYMBOL(cond_resched_lock);
5545 int __sched cond_resched_softirq(void)
5547 BUG_ON(!in_softirq());
5549 if (need_resched() && system_state == SYSTEM_RUNNING) {
5557 EXPORT_SYMBOL(cond_resched_softirq);
5560 * yield - yield the current processor to other threads.
5562 * This is a shortcut for kernel-space yielding - it marks the
5563 * thread runnable and calls sys_sched_yield().
5565 void __sched yield(void)
5567 set_current_state(TASK_RUNNING);
5570 EXPORT_SYMBOL(yield);
5573 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5574 * that process accounting knows that this is a task in IO wait state.
5576 * But don't do that if it is a deliberate, throttling IO wait (this task
5577 * has set its backing_dev_info: the queue against which it should throttle)
5579 void __sched io_schedule(void)
5581 struct rq *rq = &__raw_get_cpu_var(runqueues);
5583 delayacct_blkio_start();
5584 atomic_inc(&rq->nr_iowait);
5586 atomic_dec(&rq->nr_iowait);
5587 delayacct_blkio_end();
5589 EXPORT_SYMBOL(io_schedule);
5591 long __sched io_schedule_timeout(long timeout)
5593 struct rq *rq = &__raw_get_cpu_var(runqueues);
5596 delayacct_blkio_start();
5597 atomic_inc(&rq->nr_iowait);
5598 ret = schedule_timeout(timeout);
5599 atomic_dec(&rq->nr_iowait);
5600 delayacct_blkio_end();
5605 * sys_sched_get_priority_max - return maximum RT priority.
5606 * @policy: scheduling class.
5608 * this syscall returns the maximum rt_priority that can be used
5609 * by a given scheduling class.
5611 asmlinkage long sys_sched_get_priority_max(int policy)
5618 ret = MAX_USER_RT_PRIO-1;
5630 * sys_sched_get_priority_min - return minimum RT priority.
5631 * @policy: scheduling class.
5633 * this syscall returns the minimum rt_priority that can be used
5634 * by a given scheduling class.
5636 asmlinkage long sys_sched_get_priority_min(int policy)
5654 * sys_sched_rr_get_interval - return the default timeslice of a process.
5655 * @pid: pid of the process.
5656 * @interval: userspace pointer to the timeslice value.
5658 * this syscall writes the default timeslice value of a given process
5659 * into the user-space timespec buffer. A value of '0' means infinity.
5662 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5664 struct task_struct *p;
5665 unsigned int time_slice;
5673 read_lock(&tasklist_lock);
5674 p = find_process_by_pid(pid);
5678 retval = security_task_getscheduler(p);
5683 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5684 * tasks that are on an otherwise idle runqueue:
5687 if (p->policy == SCHED_RR) {
5688 time_slice = DEF_TIMESLICE;
5689 } else if (p->policy != SCHED_FIFO) {
5690 struct sched_entity *se = &p->se;
5691 unsigned long flags;
5694 rq = task_rq_lock(p, &flags);
5695 if (rq->cfs.load.weight)
5696 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5697 task_rq_unlock(rq, &flags);
5699 read_unlock(&tasklist_lock);
5700 jiffies_to_timespec(time_slice, &t);
5701 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5705 read_unlock(&tasklist_lock);
5709 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5711 void sched_show_task(struct task_struct *p)
5713 unsigned long free = 0;
5716 state = p->state ? __ffs(p->state) + 1 : 0;
5717 printk(KERN_INFO "%-13.13s %c", p->comm,
5718 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5719 #if BITS_PER_LONG == 32
5720 if (state == TASK_RUNNING)
5721 printk(KERN_CONT " running ");
5723 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5725 if (state == TASK_RUNNING)
5726 printk(KERN_CONT " running task ");
5728 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5730 #ifdef CONFIG_DEBUG_STACK_USAGE
5732 unsigned long *n = end_of_stack(p);
5735 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5738 printk(KERN_CONT "%5lu %5d %6d\n", free,
5739 task_pid_nr(p), task_pid_nr(p->real_parent));
5741 show_stack(p, NULL);
5744 void show_state_filter(unsigned long state_filter)
5746 struct task_struct *g, *p;
5748 #if BITS_PER_LONG == 32
5750 " task PC stack pid father\n");
5753 " task PC stack pid father\n");
5755 read_lock(&tasklist_lock);
5756 do_each_thread(g, p) {
5758 * reset the NMI-timeout, listing all files on a slow
5759 * console might take alot of time:
5761 touch_nmi_watchdog();
5762 if (!state_filter || (p->state & state_filter))
5764 } while_each_thread(g, p);
5766 touch_all_softlockup_watchdogs();
5768 #ifdef CONFIG_SCHED_DEBUG
5769 sysrq_sched_debug_show();
5771 read_unlock(&tasklist_lock);
5773 * Only show locks if all tasks are dumped:
5775 if (state_filter == -1)
5776 debug_show_all_locks();
5779 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5781 idle->sched_class = &idle_sched_class;
5785 * init_idle - set up an idle thread for a given CPU
5786 * @idle: task in question
5787 * @cpu: cpu the idle task belongs to
5789 * NOTE: this function does not set the idle thread's NEED_RESCHED
5790 * flag, to make booting more robust.
5792 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5794 struct rq *rq = cpu_rq(cpu);
5795 unsigned long flags;
5798 idle->se.exec_start = sched_clock();
5800 idle->prio = idle->normal_prio = MAX_PRIO;
5801 idle->cpus_allowed = cpumask_of_cpu(cpu);
5802 __set_task_cpu(idle, cpu);
5804 spin_lock_irqsave(&rq->lock, flags);
5805 rq->curr = rq->idle = idle;
5806 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5809 spin_unlock_irqrestore(&rq->lock, flags);
5811 /* Set the preempt count _outside_ the spinlocks! */
5812 #if defined(CONFIG_PREEMPT)
5813 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5815 task_thread_info(idle)->preempt_count = 0;
5818 * The idle tasks have their own, simple scheduling class:
5820 idle->sched_class = &idle_sched_class;
5824 * In a system that switches off the HZ timer nohz_cpu_mask
5825 * indicates which cpus entered this state. This is used
5826 * in the rcu update to wait only for active cpus. For system
5827 * which do not switch off the HZ timer nohz_cpu_mask should
5828 * always be CPU_MASK_NONE.
5830 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5833 * Increase the granularity value when there are more CPUs,
5834 * because with more CPUs the 'effective latency' as visible
5835 * to users decreases. But the relationship is not linear,
5836 * so pick a second-best guess by going with the log2 of the
5839 * This idea comes from the SD scheduler of Con Kolivas:
5841 static inline void sched_init_granularity(void)
5843 unsigned int factor = 1 + ilog2(num_online_cpus());
5844 const unsigned long limit = 200000000;
5846 sysctl_sched_min_granularity *= factor;
5847 if (sysctl_sched_min_granularity > limit)
5848 sysctl_sched_min_granularity = limit;
5850 sysctl_sched_latency *= factor;
5851 if (sysctl_sched_latency > limit)
5852 sysctl_sched_latency = limit;
5854 sysctl_sched_wakeup_granularity *= factor;
5856 sysctl_sched_shares_ratelimit *= factor;
5861 * This is how migration works:
5863 * 1) we queue a struct migration_req structure in the source CPU's
5864 * runqueue and wake up that CPU's migration thread.
5865 * 2) we down() the locked semaphore => thread blocks.
5866 * 3) migration thread wakes up (implicitly it forces the migrated
5867 * thread off the CPU)
5868 * 4) it gets the migration request and checks whether the migrated
5869 * task is still in the wrong runqueue.
5870 * 5) if it's in the wrong runqueue then the migration thread removes
5871 * it and puts it into the right queue.
5872 * 6) migration thread up()s the semaphore.
5873 * 7) we wake up and the migration is done.
5877 * Change a given task's CPU affinity. Migrate the thread to a
5878 * proper CPU and schedule it away if the CPU it's executing on
5879 * is removed from the allowed bitmask.
5881 * NOTE: the caller must have a valid reference to the task, the
5882 * task must not exit() & deallocate itself prematurely. The
5883 * call is not atomic; no spinlocks may be held.
5885 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5887 struct migration_req req;
5888 unsigned long flags;
5892 rq = task_rq_lock(p, &flags);
5893 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5898 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5899 !cpus_equal(p->cpus_allowed, *new_mask))) {
5904 if (p->sched_class->set_cpus_allowed)
5905 p->sched_class->set_cpus_allowed(p, new_mask);
5907 p->cpus_allowed = *new_mask;
5908 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5911 /* Can the task run on the task's current CPU? If so, we're done */
5912 if (cpu_isset(task_cpu(p), *new_mask))
5915 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5916 /* Need help from migration thread: drop lock and wait. */
5917 task_rq_unlock(rq, &flags);
5918 wake_up_process(rq->migration_thread);
5919 wait_for_completion(&req.done);
5920 tlb_migrate_finish(p->mm);
5924 task_rq_unlock(rq, &flags);
5928 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5931 * Move (not current) task off this cpu, onto dest cpu. We're doing
5932 * this because either it can't run here any more (set_cpus_allowed()
5933 * away from this CPU, or CPU going down), or because we're
5934 * attempting to rebalance this task on exec (sched_exec).
5936 * So we race with normal scheduler movements, but that's OK, as long
5937 * as the task is no longer on this CPU.
5939 * Returns non-zero if task was successfully migrated.
5941 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5943 struct rq *rq_dest, *rq_src;
5946 if (unlikely(!cpu_active(dest_cpu)))
5949 rq_src = cpu_rq(src_cpu);
5950 rq_dest = cpu_rq(dest_cpu);
5952 double_rq_lock(rq_src, rq_dest);
5953 /* Already moved. */
5954 if (task_cpu(p) != src_cpu)
5956 /* Affinity changed (again). */
5957 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5960 on_rq = p->se.on_rq;
5962 deactivate_task(rq_src, p, 0);
5964 set_task_cpu(p, dest_cpu);
5966 activate_task(rq_dest, p, 0);
5967 check_preempt_curr(rq_dest, p);
5972 double_rq_unlock(rq_src, rq_dest);
5977 * migration_thread - this is a highprio system thread that performs
5978 * thread migration by bumping thread off CPU then 'pushing' onto
5981 static int migration_thread(void *data)
5983 int cpu = (long)data;
5987 BUG_ON(rq->migration_thread != current);
5989 set_current_state(TASK_INTERRUPTIBLE);
5990 while (!kthread_should_stop()) {
5991 struct migration_req *req;
5992 struct list_head *head;
5994 spin_lock_irq(&rq->lock);
5996 if (cpu_is_offline(cpu)) {
5997 spin_unlock_irq(&rq->lock);
6001 if (rq->active_balance) {
6002 active_load_balance(rq, cpu);
6003 rq->active_balance = 0;
6006 head = &rq->migration_queue;
6008 if (list_empty(head)) {
6009 spin_unlock_irq(&rq->lock);
6011 set_current_state(TASK_INTERRUPTIBLE);
6014 req = list_entry(head->next, struct migration_req, list);
6015 list_del_init(head->next);
6017 spin_unlock(&rq->lock);
6018 __migrate_task(req->task, cpu, req->dest_cpu);
6021 complete(&req->done);
6023 __set_current_state(TASK_RUNNING);
6027 /* Wait for kthread_stop */
6028 set_current_state(TASK_INTERRUPTIBLE);
6029 while (!kthread_should_stop()) {
6031 set_current_state(TASK_INTERRUPTIBLE);
6033 __set_current_state(TASK_RUNNING);
6037 #ifdef CONFIG_HOTPLUG_CPU
6039 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6043 local_irq_disable();
6044 ret = __migrate_task(p, src_cpu, dest_cpu);
6050 * Figure out where task on dead CPU should go, use force if necessary.
6051 * NOTE: interrupts should be disabled by the caller
6053 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6055 unsigned long flags;
6062 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6063 cpus_and(mask, mask, p->cpus_allowed);
6064 dest_cpu = any_online_cpu(mask);
6066 /* On any allowed CPU? */
6067 if (dest_cpu >= nr_cpu_ids)
6068 dest_cpu = any_online_cpu(p->cpus_allowed);
6070 /* No more Mr. Nice Guy. */
6071 if (dest_cpu >= nr_cpu_ids) {
6072 cpumask_t cpus_allowed;
6074 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6076 * Try to stay on the same cpuset, where the
6077 * current cpuset may be a subset of all cpus.
6078 * The cpuset_cpus_allowed_locked() variant of
6079 * cpuset_cpus_allowed() will not block. It must be
6080 * called within calls to cpuset_lock/cpuset_unlock.
6082 rq = task_rq_lock(p, &flags);
6083 p->cpus_allowed = cpus_allowed;
6084 dest_cpu = any_online_cpu(p->cpus_allowed);
6085 task_rq_unlock(rq, &flags);
6088 * Don't tell them about moving exiting tasks or
6089 * kernel threads (both mm NULL), since they never
6092 if (p->mm && printk_ratelimit()) {
6093 printk(KERN_INFO "process %d (%s) no "
6094 "longer affine to cpu%d\n",
6095 task_pid_nr(p), p->comm, dead_cpu);
6098 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6102 * While a dead CPU has no uninterruptible tasks queued at this point,
6103 * it might still have a nonzero ->nr_uninterruptible counter, because
6104 * for performance reasons the counter is not stricly tracking tasks to
6105 * their home CPUs. So we just add the counter to another CPU's counter,
6106 * to keep the global sum constant after CPU-down:
6108 static void migrate_nr_uninterruptible(struct rq *rq_src)
6110 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6111 unsigned long flags;
6113 local_irq_save(flags);
6114 double_rq_lock(rq_src, rq_dest);
6115 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6116 rq_src->nr_uninterruptible = 0;
6117 double_rq_unlock(rq_src, rq_dest);
6118 local_irq_restore(flags);
6121 /* Run through task list and migrate tasks from the dead cpu. */
6122 static void migrate_live_tasks(int src_cpu)
6124 struct task_struct *p, *t;
6126 read_lock(&tasklist_lock);
6128 do_each_thread(t, p) {
6132 if (task_cpu(p) == src_cpu)
6133 move_task_off_dead_cpu(src_cpu, p);
6134 } while_each_thread(t, p);
6136 read_unlock(&tasklist_lock);
6140 * Schedules idle task to be the next runnable task on current CPU.
6141 * It does so by boosting its priority to highest possible.
6142 * Used by CPU offline code.
6144 void sched_idle_next(void)
6146 int this_cpu = smp_processor_id();
6147 struct rq *rq = cpu_rq(this_cpu);
6148 struct task_struct *p = rq->idle;
6149 unsigned long flags;
6151 /* cpu has to be offline */
6152 BUG_ON(cpu_online(this_cpu));
6155 * Strictly not necessary since rest of the CPUs are stopped by now
6156 * and interrupts disabled on the current cpu.
6158 spin_lock_irqsave(&rq->lock, flags);
6160 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6162 update_rq_clock(rq);
6163 activate_task(rq, p, 0);
6165 spin_unlock_irqrestore(&rq->lock, flags);
6169 * Ensures that the idle task is using init_mm right before its cpu goes
6172 void idle_task_exit(void)
6174 struct mm_struct *mm = current->active_mm;
6176 BUG_ON(cpu_online(smp_processor_id()));
6179 switch_mm(mm, &init_mm, current);
6183 /* called under rq->lock with disabled interrupts */
6184 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6186 struct rq *rq = cpu_rq(dead_cpu);
6188 /* Must be exiting, otherwise would be on tasklist. */
6189 BUG_ON(!p->exit_state);
6191 /* Cannot have done final schedule yet: would have vanished. */
6192 BUG_ON(p->state == TASK_DEAD);
6197 * Drop lock around migration; if someone else moves it,
6198 * that's OK. No task can be added to this CPU, so iteration is
6201 spin_unlock_irq(&rq->lock);
6202 move_task_off_dead_cpu(dead_cpu, p);
6203 spin_lock_irq(&rq->lock);
6208 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6209 static void migrate_dead_tasks(unsigned int dead_cpu)
6211 struct rq *rq = cpu_rq(dead_cpu);
6212 struct task_struct *next;
6215 if (!rq->nr_running)
6217 update_rq_clock(rq);
6218 next = pick_next_task(rq, rq->curr);
6221 next->sched_class->put_prev_task(rq, next);
6222 migrate_dead(dead_cpu, next);
6226 #endif /* CONFIG_HOTPLUG_CPU */
6228 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6230 static struct ctl_table sd_ctl_dir[] = {
6232 .procname = "sched_domain",
6238 static struct ctl_table sd_ctl_root[] = {
6240 .ctl_name = CTL_KERN,
6241 .procname = "kernel",
6243 .child = sd_ctl_dir,
6248 static struct ctl_table *sd_alloc_ctl_entry(int n)
6250 struct ctl_table *entry =
6251 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6256 static void sd_free_ctl_entry(struct ctl_table **tablep)
6258 struct ctl_table *entry;
6261 * In the intermediate directories, both the child directory and
6262 * procname are dynamically allocated and could fail but the mode
6263 * will always be set. In the lowest directory the names are
6264 * static strings and all have proc handlers.
6266 for (entry = *tablep; entry->mode; entry++) {
6268 sd_free_ctl_entry(&entry->child);
6269 if (entry->proc_handler == NULL)
6270 kfree(entry->procname);
6278 set_table_entry(struct ctl_table *entry,
6279 const char *procname, void *data, int maxlen,
6280 mode_t mode, proc_handler *proc_handler)
6282 entry->procname = procname;
6284 entry->maxlen = maxlen;
6286 entry->proc_handler = proc_handler;
6289 static struct ctl_table *
6290 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6292 struct ctl_table *table = sd_alloc_ctl_entry(12);
6297 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6298 sizeof(long), 0644, proc_doulongvec_minmax);
6299 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6300 sizeof(long), 0644, proc_doulongvec_minmax);
6301 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6302 sizeof(int), 0644, proc_dointvec_minmax);
6303 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6304 sizeof(int), 0644, proc_dointvec_minmax);
6305 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6306 sizeof(int), 0644, proc_dointvec_minmax);
6307 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6308 sizeof(int), 0644, proc_dointvec_minmax);
6309 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6310 sizeof(int), 0644, proc_dointvec_minmax);
6311 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6312 sizeof(int), 0644, proc_dointvec_minmax);
6313 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6314 sizeof(int), 0644, proc_dointvec_minmax);
6315 set_table_entry(&table[9], "cache_nice_tries",
6316 &sd->cache_nice_tries,
6317 sizeof(int), 0644, proc_dointvec_minmax);
6318 set_table_entry(&table[10], "flags", &sd->flags,
6319 sizeof(int), 0644, proc_dointvec_minmax);
6320 /* &table[11] is terminator */
6325 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6327 struct ctl_table *entry, *table;
6328 struct sched_domain *sd;
6329 int domain_num = 0, i;
6332 for_each_domain(cpu, sd)
6334 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6339 for_each_domain(cpu, sd) {
6340 snprintf(buf, 32, "domain%d", i);
6341 entry->procname = kstrdup(buf, GFP_KERNEL);
6343 entry->child = sd_alloc_ctl_domain_table(sd);
6350 static struct ctl_table_header *sd_sysctl_header;
6351 static void register_sched_domain_sysctl(void)
6353 int i, cpu_num = num_online_cpus();
6354 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6357 WARN_ON(sd_ctl_dir[0].child);
6358 sd_ctl_dir[0].child = entry;
6363 for_each_online_cpu(i) {
6364 snprintf(buf, 32, "cpu%d", i);
6365 entry->procname = kstrdup(buf, GFP_KERNEL);
6367 entry->child = sd_alloc_ctl_cpu_table(i);
6371 WARN_ON(sd_sysctl_header);
6372 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6375 /* may be called multiple times per register */
6376 static void unregister_sched_domain_sysctl(void)
6378 if (sd_sysctl_header)
6379 unregister_sysctl_table(sd_sysctl_header);
6380 sd_sysctl_header = NULL;
6381 if (sd_ctl_dir[0].child)
6382 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6385 static void register_sched_domain_sysctl(void)
6388 static void unregister_sched_domain_sysctl(void)
6393 static void set_rq_online(struct rq *rq)
6396 const struct sched_class *class;
6398 cpu_set(rq->cpu, rq->rd->online);
6401 for_each_class(class) {
6402 if (class->rq_online)
6403 class->rq_online(rq);
6408 static void set_rq_offline(struct rq *rq)
6411 const struct sched_class *class;
6413 for_each_class(class) {
6414 if (class->rq_offline)
6415 class->rq_offline(rq);
6418 cpu_clear(rq->cpu, rq->rd->online);
6424 * migration_call - callback that gets triggered when a CPU is added.
6425 * Here we can start up the necessary migration thread for the new CPU.
6427 static int __cpuinit
6428 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6430 struct task_struct *p;
6431 int cpu = (long)hcpu;
6432 unsigned long flags;
6437 case CPU_UP_PREPARE:
6438 case CPU_UP_PREPARE_FROZEN:
6439 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6442 kthread_bind(p, cpu);
6443 /* Must be high prio: stop_machine expects to yield to it. */
6444 rq = task_rq_lock(p, &flags);
6445 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6446 task_rq_unlock(rq, &flags);
6447 cpu_rq(cpu)->migration_thread = p;
6451 case CPU_ONLINE_FROZEN:
6452 /* Strictly unnecessary, as first user will wake it. */
6453 wake_up_process(cpu_rq(cpu)->migration_thread);
6455 /* Update our root-domain */
6457 spin_lock_irqsave(&rq->lock, flags);
6459 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6463 spin_unlock_irqrestore(&rq->lock, flags);
6466 #ifdef CONFIG_HOTPLUG_CPU
6467 case CPU_UP_CANCELED:
6468 case CPU_UP_CANCELED_FROZEN:
6469 if (!cpu_rq(cpu)->migration_thread)
6471 /* Unbind it from offline cpu so it can run. Fall thru. */
6472 kthread_bind(cpu_rq(cpu)->migration_thread,
6473 any_online_cpu(cpu_online_map));
6474 kthread_stop(cpu_rq(cpu)->migration_thread);
6475 cpu_rq(cpu)->migration_thread = NULL;
6479 case CPU_DEAD_FROZEN:
6480 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6481 migrate_live_tasks(cpu);
6483 kthread_stop(rq->migration_thread);
6484 rq->migration_thread = NULL;
6485 /* Idle task back to normal (off runqueue, low prio) */
6486 spin_lock_irq(&rq->lock);
6487 update_rq_clock(rq);
6488 deactivate_task(rq, rq->idle, 0);
6489 rq->idle->static_prio = MAX_PRIO;
6490 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6491 rq->idle->sched_class = &idle_sched_class;
6492 migrate_dead_tasks(cpu);
6493 spin_unlock_irq(&rq->lock);
6495 migrate_nr_uninterruptible(rq);
6496 BUG_ON(rq->nr_running != 0);
6499 * No need to migrate the tasks: it was best-effort if
6500 * they didn't take sched_hotcpu_mutex. Just wake up
6503 spin_lock_irq(&rq->lock);
6504 while (!list_empty(&rq->migration_queue)) {
6505 struct migration_req *req;
6507 req = list_entry(rq->migration_queue.next,
6508 struct migration_req, list);
6509 list_del_init(&req->list);
6510 complete(&req->done);
6512 spin_unlock_irq(&rq->lock);
6516 case CPU_DYING_FROZEN:
6517 /* Update our root-domain */
6519 spin_lock_irqsave(&rq->lock, flags);
6521 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6524 spin_unlock_irqrestore(&rq->lock, flags);
6531 /* Register at highest priority so that task migration (migrate_all_tasks)
6532 * happens before everything else.
6534 static struct notifier_block __cpuinitdata migration_notifier = {
6535 .notifier_call = migration_call,
6539 static int __init migration_init(void)
6541 void *cpu = (void *)(long)smp_processor_id();
6544 /* Start one for the boot CPU: */
6545 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6546 BUG_ON(err == NOTIFY_BAD);
6547 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6548 register_cpu_notifier(&migration_notifier);
6552 early_initcall(migration_init);
6557 #ifdef CONFIG_SCHED_DEBUG
6559 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6572 case SD_LV_ALLNODES:
6581 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6582 cpumask_t *groupmask)
6584 struct sched_group *group = sd->groups;
6587 cpulist_scnprintf(str, sizeof(str), sd->span);
6588 cpus_clear(*groupmask);
6590 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6592 if (!(sd->flags & SD_LOAD_BALANCE)) {
6593 printk("does not load-balance\n");
6595 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6600 printk(KERN_CONT "span %s level %s\n",
6601 str, sd_level_to_string(sd->level));
6603 if (!cpu_isset(cpu, sd->span)) {
6604 printk(KERN_ERR "ERROR: domain->span does not contain "
6607 if (!cpu_isset(cpu, group->cpumask)) {
6608 printk(KERN_ERR "ERROR: domain->groups does not contain"
6612 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6616 printk(KERN_ERR "ERROR: group is NULL\n");
6620 if (!group->__cpu_power) {
6621 printk(KERN_CONT "\n");
6622 printk(KERN_ERR "ERROR: domain->cpu_power not "
6627 if (!cpus_weight(group->cpumask)) {
6628 printk(KERN_CONT "\n");
6629 printk(KERN_ERR "ERROR: empty group\n");
6633 if (cpus_intersects(*groupmask, group->cpumask)) {
6634 printk(KERN_CONT "\n");
6635 printk(KERN_ERR "ERROR: repeated CPUs\n");
6639 cpus_or(*groupmask, *groupmask, group->cpumask);
6641 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6642 printk(KERN_CONT " %s", str);
6644 group = group->next;
6645 } while (group != sd->groups);
6646 printk(KERN_CONT "\n");
6648 if (!cpus_equal(sd->span, *groupmask))
6649 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6651 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6652 printk(KERN_ERR "ERROR: parent span is not a superset "
6653 "of domain->span\n");
6657 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6659 cpumask_t *groupmask;
6663 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6667 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6669 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6671 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6676 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6685 #else /* !CONFIG_SCHED_DEBUG */
6686 # define sched_domain_debug(sd, cpu) do { } while (0)
6687 #endif /* CONFIG_SCHED_DEBUG */
6689 static int sd_degenerate(struct sched_domain *sd)
6691 if (cpus_weight(sd->span) == 1)
6694 /* Following flags need at least 2 groups */
6695 if (sd->flags & (SD_LOAD_BALANCE |
6696 SD_BALANCE_NEWIDLE |
6700 SD_SHARE_PKG_RESOURCES)) {
6701 if (sd->groups != sd->groups->next)
6705 /* Following flags don't use groups */
6706 if (sd->flags & (SD_WAKE_IDLE |
6715 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6717 unsigned long cflags = sd->flags, pflags = parent->flags;
6719 if (sd_degenerate(parent))
6722 if (!cpus_equal(sd->span, parent->span))
6725 /* Does parent contain flags not in child? */
6726 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6727 if (cflags & SD_WAKE_AFFINE)
6728 pflags &= ~SD_WAKE_BALANCE;
6729 /* Flags needing groups don't count if only 1 group in parent */
6730 if (parent->groups == parent->groups->next) {
6731 pflags &= ~(SD_LOAD_BALANCE |
6732 SD_BALANCE_NEWIDLE |
6736 SD_SHARE_PKG_RESOURCES);
6738 if (~cflags & pflags)
6744 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6746 unsigned long flags;
6748 spin_lock_irqsave(&rq->lock, flags);
6751 struct root_domain *old_rd = rq->rd;
6753 if (cpu_isset(rq->cpu, old_rd->online))
6756 cpu_clear(rq->cpu, old_rd->span);
6758 if (atomic_dec_and_test(&old_rd->refcount))
6762 atomic_inc(&rd->refcount);
6765 cpu_set(rq->cpu, rd->span);
6766 if (cpu_isset(rq->cpu, cpu_online_map))
6769 spin_unlock_irqrestore(&rq->lock, flags);
6772 static void init_rootdomain(struct root_domain *rd)
6774 memset(rd, 0, sizeof(*rd));
6776 cpus_clear(rd->span);
6777 cpus_clear(rd->online);
6779 cpupri_init(&rd->cpupri);
6782 static void init_defrootdomain(void)
6784 init_rootdomain(&def_root_domain);
6785 atomic_set(&def_root_domain.refcount, 1);
6788 static struct root_domain *alloc_rootdomain(void)
6790 struct root_domain *rd;
6792 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6796 init_rootdomain(rd);
6802 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6803 * hold the hotplug lock.
6806 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6808 struct rq *rq = cpu_rq(cpu);
6809 struct sched_domain *tmp;
6811 /* Remove the sched domains which do not contribute to scheduling. */
6812 for (tmp = sd; tmp; tmp = tmp->parent) {
6813 struct sched_domain *parent = tmp->parent;
6816 if (sd_parent_degenerate(tmp, parent)) {
6817 tmp->parent = parent->parent;
6819 parent->parent->child = tmp;
6823 if (sd && sd_degenerate(sd)) {
6829 sched_domain_debug(sd, cpu);
6831 rq_attach_root(rq, rd);
6832 rcu_assign_pointer(rq->sd, sd);
6835 /* cpus with isolated domains */
6836 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6838 /* Setup the mask of cpus configured for isolated domains */
6839 static int __init isolated_cpu_setup(char *str)
6841 static int __initdata ints[NR_CPUS];
6844 str = get_options(str, ARRAY_SIZE(ints), ints);
6845 cpus_clear(cpu_isolated_map);
6846 for (i = 1; i <= ints[0]; i++)
6847 if (ints[i] < NR_CPUS)
6848 cpu_set(ints[i], cpu_isolated_map);
6852 __setup("isolcpus=", isolated_cpu_setup);
6855 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6856 * to a function which identifies what group(along with sched group) a CPU
6857 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6858 * (due to the fact that we keep track of groups covered with a cpumask_t).
6860 * init_sched_build_groups will build a circular linked list of the groups
6861 * covered by the given span, and will set each group's ->cpumask correctly,
6862 * and ->cpu_power to 0.
6865 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6866 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6867 struct sched_group **sg,
6868 cpumask_t *tmpmask),
6869 cpumask_t *covered, cpumask_t *tmpmask)
6871 struct sched_group *first = NULL, *last = NULL;
6874 cpus_clear(*covered);
6876 for_each_cpu_mask_nr(i, *span) {
6877 struct sched_group *sg;
6878 int group = group_fn(i, cpu_map, &sg, tmpmask);
6881 if (cpu_isset(i, *covered))
6884 cpus_clear(sg->cpumask);
6885 sg->__cpu_power = 0;
6887 for_each_cpu_mask_nr(j, *span) {
6888 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6891 cpu_set(j, *covered);
6892 cpu_set(j, sg->cpumask);
6903 #define SD_NODES_PER_DOMAIN 16
6908 * find_next_best_node - find the next node to include in a sched_domain
6909 * @node: node whose sched_domain we're building
6910 * @used_nodes: nodes already in the sched_domain
6912 * Find the next node to include in a given scheduling domain. Simply
6913 * finds the closest node not already in the @used_nodes map.
6915 * Should use nodemask_t.
6917 static int find_next_best_node(int node, nodemask_t *used_nodes)
6919 int i, n, val, min_val, best_node = 0;
6923 for (i = 0; i < nr_node_ids; i++) {
6924 /* Start at @node */
6925 n = (node + i) % nr_node_ids;
6927 if (!nr_cpus_node(n))
6930 /* Skip already used nodes */
6931 if (node_isset(n, *used_nodes))
6934 /* Simple min distance search */
6935 val = node_distance(node, n);
6937 if (val < min_val) {
6943 node_set(best_node, *used_nodes);
6948 * sched_domain_node_span - get a cpumask for a node's sched_domain
6949 * @node: node whose cpumask we're constructing
6950 * @span: resulting cpumask
6952 * Given a node, construct a good cpumask for its sched_domain to span. It
6953 * should be one that prevents unnecessary balancing, but also spreads tasks
6956 static void sched_domain_node_span(int node, cpumask_t *span)
6958 nodemask_t used_nodes;
6959 node_to_cpumask_ptr(nodemask, node);
6963 nodes_clear(used_nodes);
6965 cpus_or(*span, *span, *nodemask);
6966 node_set(node, used_nodes);
6968 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6969 int next_node = find_next_best_node(node, &used_nodes);
6971 node_to_cpumask_ptr_next(nodemask, next_node);
6972 cpus_or(*span, *span, *nodemask);
6975 #endif /* CONFIG_NUMA */
6977 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6980 * SMT sched-domains:
6982 #ifdef CONFIG_SCHED_SMT
6983 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6984 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6987 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6991 *sg = &per_cpu(sched_group_cpus, cpu);
6994 #endif /* CONFIG_SCHED_SMT */
6997 * multi-core sched-domains:
6999 #ifdef CONFIG_SCHED_MC
7000 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7001 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7002 #endif /* CONFIG_SCHED_MC */
7004 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7006 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7011 *mask = per_cpu(cpu_sibling_map, cpu);
7012 cpus_and(*mask, *mask, *cpu_map);
7013 group = first_cpu(*mask);
7015 *sg = &per_cpu(sched_group_core, group);
7018 #elif defined(CONFIG_SCHED_MC)
7020 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7024 *sg = &per_cpu(sched_group_core, cpu);
7029 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7030 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7033 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7037 #ifdef CONFIG_SCHED_MC
7038 *mask = cpu_coregroup_map(cpu);
7039 cpus_and(*mask, *mask, *cpu_map);
7040 group = first_cpu(*mask);
7041 #elif defined(CONFIG_SCHED_SMT)
7042 *mask = per_cpu(cpu_sibling_map, cpu);
7043 cpus_and(*mask, *mask, *cpu_map);
7044 group = first_cpu(*mask);
7049 *sg = &per_cpu(sched_group_phys, group);
7055 * The init_sched_build_groups can't handle what we want to do with node
7056 * groups, so roll our own. Now each node has its own list of groups which
7057 * gets dynamically allocated.
7059 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7060 static struct sched_group ***sched_group_nodes_bycpu;
7062 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7063 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7065 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7066 struct sched_group **sg, cpumask_t *nodemask)
7070 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7071 cpus_and(*nodemask, *nodemask, *cpu_map);
7072 group = first_cpu(*nodemask);
7075 *sg = &per_cpu(sched_group_allnodes, group);
7079 static void init_numa_sched_groups_power(struct sched_group *group_head)
7081 struct sched_group *sg = group_head;
7087 for_each_cpu_mask_nr(j, sg->cpumask) {
7088 struct sched_domain *sd;
7090 sd = &per_cpu(phys_domains, j);
7091 if (j != first_cpu(sd->groups->cpumask)) {
7093 * Only add "power" once for each
7099 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7102 } while (sg != group_head);
7104 #endif /* CONFIG_NUMA */
7107 /* Free memory allocated for various sched_group structures */
7108 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7112 for_each_cpu_mask_nr(cpu, *cpu_map) {
7113 struct sched_group **sched_group_nodes
7114 = sched_group_nodes_bycpu[cpu];
7116 if (!sched_group_nodes)
7119 for (i = 0; i < nr_node_ids; i++) {
7120 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7122 *nodemask = node_to_cpumask(i);
7123 cpus_and(*nodemask, *nodemask, *cpu_map);
7124 if (cpus_empty(*nodemask))
7134 if (oldsg != sched_group_nodes[i])
7137 kfree(sched_group_nodes);
7138 sched_group_nodes_bycpu[cpu] = NULL;
7141 #else /* !CONFIG_NUMA */
7142 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7145 #endif /* CONFIG_NUMA */
7148 * Initialize sched groups cpu_power.
7150 * cpu_power indicates the capacity of sched group, which is used while
7151 * distributing the load between different sched groups in a sched domain.
7152 * Typically cpu_power for all the groups in a sched domain will be same unless
7153 * there are asymmetries in the topology. If there are asymmetries, group
7154 * having more cpu_power will pickup more load compared to the group having
7157 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7158 * the maximum number of tasks a group can handle in the presence of other idle
7159 * or lightly loaded groups in the same sched domain.
7161 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7163 struct sched_domain *child;
7164 struct sched_group *group;
7166 WARN_ON(!sd || !sd->groups);
7168 if (cpu != first_cpu(sd->groups->cpumask))
7173 sd->groups->__cpu_power = 0;
7176 * For perf policy, if the groups in child domain share resources
7177 * (for example cores sharing some portions of the cache hierarchy
7178 * or SMT), then set this domain groups cpu_power such that each group
7179 * can handle only one task, when there are other idle groups in the
7180 * same sched domain.
7182 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7184 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7185 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7190 * add cpu_power of each child group to this groups cpu_power
7192 group = child->groups;
7194 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7195 group = group->next;
7196 } while (group != child->groups);
7200 * Initializers for schedule domains
7201 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7204 #define SD_INIT(sd, type) sd_init_##type(sd)
7205 #define SD_INIT_FUNC(type) \
7206 static noinline void sd_init_##type(struct sched_domain *sd) \
7208 memset(sd, 0, sizeof(*sd)); \
7209 *sd = SD_##type##_INIT; \
7210 sd->level = SD_LV_##type; \
7215 SD_INIT_FUNC(ALLNODES)
7218 #ifdef CONFIG_SCHED_SMT
7219 SD_INIT_FUNC(SIBLING)
7221 #ifdef CONFIG_SCHED_MC
7226 * To minimize stack usage kmalloc room for cpumasks and share the
7227 * space as the usage in build_sched_domains() dictates. Used only
7228 * if the amount of space is significant.
7231 cpumask_t tmpmask; /* make this one first */
7234 cpumask_t this_sibling_map;
7235 cpumask_t this_core_map;
7237 cpumask_t send_covered;
7240 cpumask_t domainspan;
7242 cpumask_t notcovered;
7247 #define SCHED_CPUMASK_ALLOC 1
7248 #define SCHED_CPUMASK_FREE(v) kfree(v)
7249 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7251 #define SCHED_CPUMASK_ALLOC 0
7252 #define SCHED_CPUMASK_FREE(v)
7253 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7256 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7257 ((unsigned long)(a) + offsetof(struct allmasks, v))
7259 static int default_relax_domain_level = -1;
7261 static int __init setup_relax_domain_level(char *str)
7265 val = simple_strtoul(str, NULL, 0);
7266 if (val < SD_LV_MAX)
7267 default_relax_domain_level = val;
7271 __setup("relax_domain_level=", setup_relax_domain_level);
7273 static void set_domain_attribute(struct sched_domain *sd,
7274 struct sched_domain_attr *attr)
7278 if (!attr || attr->relax_domain_level < 0) {
7279 if (default_relax_domain_level < 0)
7282 request = default_relax_domain_level;
7284 request = attr->relax_domain_level;
7285 if (request < sd->level) {
7286 /* turn off idle balance on this domain */
7287 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7289 /* turn on idle balance on this domain */
7290 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7295 * Build sched domains for a given set of cpus and attach the sched domains
7296 * to the individual cpus
7298 static int __build_sched_domains(const cpumask_t *cpu_map,
7299 struct sched_domain_attr *attr)
7302 struct root_domain *rd;
7303 SCHED_CPUMASK_DECLARE(allmasks);
7306 struct sched_group **sched_group_nodes = NULL;
7307 int sd_allnodes = 0;
7310 * Allocate the per-node list of sched groups
7312 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7314 if (!sched_group_nodes) {
7315 printk(KERN_WARNING "Can not alloc sched group node list\n");
7320 rd = alloc_rootdomain();
7322 printk(KERN_WARNING "Cannot alloc root domain\n");
7324 kfree(sched_group_nodes);
7329 #if SCHED_CPUMASK_ALLOC
7330 /* get space for all scratch cpumask variables */
7331 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7333 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7336 kfree(sched_group_nodes);
7341 tmpmask = (cpumask_t *)allmasks;
7345 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7349 * Set up domains for cpus specified by the cpu_map.
7351 for_each_cpu_mask_nr(i, *cpu_map) {
7352 struct sched_domain *sd = NULL, *p;
7353 SCHED_CPUMASK_VAR(nodemask, allmasks);
7355 *nodemask = node_to_cpumask(cpu_to_node(i));
7356 cpus_and(*nodemask, *nodemask, *cpu_map);
7359 if (cpus_weight(*cpu_map) >
7360 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7361 sd = &per_cpu(allnodes_domains, i);
7362 SD_INIT(sd, ALLNODES);
7363 set_domain_attribute(sd, attr);
7364 sd->span = *cpu_map;
7365 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7371 sd = &per_cpu(node_domains, i);
7373 set_domain_attribute(sd, attr);
7374 sched_domain_node_span(cpu_to_node(i), &sd->span);
7378 cpus_and(sd->span, sd->span, *cpu_map);
7382 sd = &per_cpu(phys_domains, i);
7384 set_domain_attribute(sd, attr);
7385 sd->span = *nodemask;
7389 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7391 #ifdef CONFIG_SCHED_MC
7393 sd = &per_cpu(core_domains, i);
7395 set_domain_attribute(sd, attr);
7396 sd->span = cpu_coregroup_map(i);
7397 cpus_and(sd->span, sd->span, *cpu_map);
7400 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7403 #ifdef CONFIG_SCHED_SMT
7405 sd = &per_cpu(cpu_domains, i);
7406 SD_INIT(sd, SIBLING);
7407 set_domain_attribute(sd, attr);
7408 sd->span = per_cpu(cpu_sibling_map, i);
7409 cpus_and(sd->span, sd->span, *cpu_map);
7412 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7416 #ifdef CONFIG_SCHED_SMT
7417 /* Set up CPU (sibling) groups */
7418 for_each_cpu_mask_nr(i, *cpu_map) {
7419 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7420 SCHED_CPUMASK_VAR(send_covered, allmasks);
7422 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7423 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7424 if (i != first_cpu(*this_sibling_map))
7427 init_sched_build_groups(this_sibling_map, cpu_map,
7429 send_covered, tmpmask);
7433 #ifdef CONFIG_SCHED_MC
7434 /* Set up multi-core groups */
7435 for_each_cpu_mask_nr(i, *cpu_map) {
7436 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7437 SCHED_CPUMASK_VAR(send_covered, allmasks);
7439 *this_core_map = cpu_coregroup_map(i);
7440 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7441 if (i != first_cpu(*this_core_map))
7444 init_sched_build_groups(this_core_map, cpu_map,
7446 send_covered, tmpmask);
7450 /* Set up physical groups */
7451 for (i = 0; i < nr_node_ids; i++) {
7452 SCHED_CPUMASK_VAR(nodemask, allmasks);
7453 SCHED_CPUMASK_VAR(send_covered, allmasks);
7455 *nodemask = node_to_cpumask(i);
7456 cpus_and(*nodemask, *nodemask, *cpu_map);
7457 if (cpus_empty(*nodemask))
7460 init_sched_build_groups(nodemask, cpu_map,
7462 send_covered, tmpmask);
7466 /* Set up node groups */
7468 SCHED_CPUMASK_VAR(send_covered, allmasks);
7470 init_sched_build_groups(cpu_map, cpu_map,
7471 &cpu_to_allnodes_group,
7472 send_covered, tmpmask);
7475 for (i = 0; i < nr_node_ids; i++) {
7476 /* Set up node groups */
7477 struct sched_group *sg, *prev;
7478 SCHED_CPUMASK_VAR(nodemask, allmasks);
7479 SCHED_CPUMASK_VAR(domainspan, allmasks);
7480 SCHED_CPUMASK_VAR(covered, allmasks);
7483 *nodemask = node_to_cpumask(i);
7484 cpus_clear(*covered);
7486 cpus_and(*nodemask, *nodemask, *cpu_map);
7487 if (cpus_empty(*nodemask)) {
7488 sched_group_nodes[i] = NULL;
7492 sched_domain_node_span(i, domainspan);
7493 cpus_and(*domainspan, *domainspan, *cpu_map);
7495 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7497 printk(KERN_WARNING "Can not alloc domain group for "
7501 sched_group_nodes[i] = sg;
7502 for_each_cpu_mask_nr(j, *nodemask) {
7503 struct sched_domain *sd;
7505 sd = &per_cpu(node_domains, j);
7508 sg->__cpu_power = 0;
7509 sg->cpumask = *nodemask;
7511 cpus_or(*covered, *covered, *nodemask);
7514 for (j = 0; j < nr_node_ids; j++) {
7515 SCHED_CPUMASK_VAR(notcovered, allmasks);
7516 int n = (i + j) % nr_node_ids;
7517 node_to_cpumask_ptr(pnodemask, n);
7519 cpus_complement(*notcovered, *covered);
7520 cpus_and(*tmpmask, *notcovered, *cpu_map);
7521 cpus_and(*tmpmask, *tmpmask, *domainspan);
7522 if (cpus_empty(*tmpmask))
7525 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7526 if (cpus_empty(*tmpmask))
7529 sg = kmalloc_node(sizeof(struct sched_group),
7533 "Can not alloc domain group for node %d\n", j);
7536 sg->__cpu_power = 0;
7537 sg->cpumask = *tmpmask;
7538 sg->next = prev->next;
7539 cpus_or(*covered, *covered, *tmpmask);
7546 /* Calculate CPU power for physical packages and nodes */
7547 #ifdef CONFIG_SCHED_SMT
7548 for_each_cpu_mask_nr(i, *cpu_map) {
7549 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7551 init_sched_groups_power(i, sd);
7554 #ifdef CONFIG_SCHED_MC
7555 for_each_cpu_mask_nr(i, *cpu_map) {
7556 struct sched_domain *sd = &per_cpu(core_domains, i);
7558 init_sched_groups_power(i, sd);
7562 for_each_cpu_mask_nr(i, *cpu_map) {
7563 struct sched_domain *sd = &per_cpu(phys_domains, i);
7565 init_sched_groups_power(i, sd);
7569 for (i = 0; i < nr_node_ids; i++)
7570 init_numa_sched_groups_power(sched_group_nodes[i]);
7573 struct sched_group *sg;
7575 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7577 init_numa_sched_groups_power(sg);
7581 /* Attach the domains */
7582 for_each_cpu_mask_nr(i, *cpu_map) {
7583 struct sched_domain *sd;
7584 #ifdef CONFIG_SCHED_SMT
7585 sd = &per_cpu(cpu_domains, i);
7586 #elif defined(CONFIG_SCHED_MC)
7587 sd = &per_cpu(core_domains, i);
7589 sd = &per_cpu(phys_domains, i);
7591 cpu_attach_domain(sd, rd, i);
7594 SCHED_CPUMASK_FREE((void *)allmasks);
7599 free_sched_groups(cpu_map, tmpmask);
7600 SCHED_CPUMASK_FREE((void *)allmasks);
7605 static int build_sched_domains(const cpumask_t *cpu_map)
7607 return __build_sched_domains(cpu_map, NULL);
7610 static cpumask_t *doms_cur; /* current sched domains */
7611 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7612 static struct sched_domain_attr *dattr_cur;
7613 /* attribues of custom domains in 'doms_cur' */
7616 * Special case: If a kmalloc of a doms_cur partition (array of
7617 * cpumask_t) fails, then fallback to a single sched domain,
7618 * as determined by the single cpumask_t fallback_doms.
7620 static cpumask_t fallback_doms;
7622 void __attribute__((weak)) arch_update_cpu_topology(void)
7627 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7628 * For now this just excludes isolated cpus, but could be used to
7629 * exclude other special cases in the future.
7631 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7635 arch_update_cpu_topology();
7637 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7639 doms_cur = &fallback_doms;
7640 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7642 err = build_sched_domains(doms_cur);
7643 register_sched_domain_sysctl();
7648 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7651 free_sched_groups(cpu_map, tmpmask);
7655 * Detach sched domains from a group of cpus specified in cpu_map
7656 * These cpus will now be attached to the NULL domain
7658 static void detach_destroy_domains(const cpumask_t *cpu_map)
7663 unregister_sched_domain_sysctl();
7665 for_each_cpu_mask_nr(i, *cpu_map)
7666 cpu_attach_domain(NULL, &def_root_domain, i);
7667 synchronize_sched();
7668 arch_destroy_sched_domains(cpu_map, &tmpmask);
7671 /* handle null as "default" */
7672 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7673 struct sched_domain_attr *new, int idx_new)
7675 struct sched_domain_attr tmp;
7682 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7683 new ? (new + idx_new) : &tmp,
7684 sizeof(struct sched_domain_attr));
7688 * Partition sched domains as specified by the 'ndoms_new'
7689 * cpumasks in the array doms_new[] of cpumasks. This compares
7690 * doms_new[] to the current sched domain partitioning, doms_cur[].
7691 * It destroys each deleted domain and builds each new domain.
7693 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7694 * The masks don't intersect (don't overlap.) We should setup one
7695 * sched domain for each mask. CPUs not in any of the cpumasks will
7696 * not be load balanced. If the same cpumask appears both in the
7697 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7700 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7701 * ownership of it and will kfree it when done with it. If the caller
7702 * failed the kmalloc call, then it can pass in doms_new == NULL,
7703 * and partition_sched_domains() will fallback to the single partition
7704 * 'fallback_doms', it also forces the domains to be rebuilt.
7706 * If doms_new==NULL it will be replaced with cpu_online_map.
7707 * ndoms_new==0 is a special case for destroying existing domains.
7708 * It will not create the default domain.
7710 * Call with hotplug lock held
7712 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7713 struct sched_domain_attr *dattr_new)
7717 mutex_lock(&sched_domains_mutex);
7719 /* always unregister in case we don't destroy any domains */
7720 unregister_sched_domain_sysctl();
7722 n = doms_new ? ndoms_new : 0;
7724 /* Destroy deleted domains */
7725 for (i = 0; i < ndoms_cur; i++) {
7726 for (j = 0; j < n; j++) {
7727 if (cpus_equal(doms_cur[i], doms_new[j])
7728 && dattrs_equal(dattr_cur, i, dattr_new, j))
7731 /* no match - a current sched domain not in new doms_new[] */
7732 detach_destroy_domains(doms_cur + i);
7737 if (doms_new == NULL) {
7739 doms_new = &fallback_doms;
7740 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7744 /* Build new domains */
7745 for (i = 0; i < ndoms_new; i++) {
7746 for (j = 0; j < ndoms_cur; j++) {
7747 if (cpus_equal(doms_new[i], doms_cur[j])
7748 && dattrs_equal(dattr_new, i, dattr_cur, j))
7751 /* no match - add a new doms_new */
7752 __build_sched_domains(doms_new + i,
7753 dattr_new ? dattr_new + i : NULL);
7758 /* Remember the new sched domains */
7759 if (doms_cur != &fallback_doms)
7761 kfree(dattr_cur); /* kfree(NULL) is safe */
7762 doms_cur = doms_new;
7763 dattr_cur = dattr_new;
7764 ndoms_cur = ndoms_new;
7766 register_sched_domain_sysctl();
7768 mutex_unlock(&sched_domains_mutex);
7771 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7772 int arch_reinit_sched_domains(void)
7776 /* Destroy domains first to force the rebuild */
7777 partition_sched_domains(0, NULL, NULL);
7779 rebuild_sched_domains();
7785 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7789 if (buf[0] != '0' && buf[0] != '1')
7793 sched_smt_power_savings = (buf[0] == '1');
7795 sched_mc_power_savings = (buf[0] == '1');
7797 ret = arch_reinit_sched_domains();
7799 return ret ? ret : count;
7802 #ifdef CONFIG_SCHED_MC
7803 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7806 return sprintf(page, "%u\n", sched_mc_power_savings);
7808 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7809 const char *buf, size_t count)
7811 return sched_power_savings_store(buf, count, 0);
7813 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7814 sched_mc_power_savings_show,
7815 sched_mc_power_savings_store);
7818 #ifdef CONFIG_SCHED_SMT
7819 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7822 return sprintf(page, "%u\n", sched_smt_power_savings);
7824 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7825 const char *buf, size_t count)
7827 return sched_power_savings_store(buf, count, 1);
7829 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7830 sched_smt_power_savings_show,
7831 sched_smt_power_savings_store);
7834 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7838 #ifdef CONFIG_SCHED_SMT
7840 err = sysfs_create_file(&cls->kset.kobj,
7841 &attr_sched_smt_power_savings.attr);
7843 #ifdef CONFIG_SCHED_MC
7844 if (!err && mc_capable())
7845 err = sysfs_create_file(&cls->kset.kobj,
7846 &attr_sched_mc_power_savings.attr);
7850 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7852 #ifndef CONFIG_CPUSETS
7854 * Add online and remove offline CPUs from the scheduler domains.
7855 * When cpusets are enabled they take over this function.
7857 static int update_sched_domains(struct notifier_block *nfb,
7858 unsigned long action, void *hcpu)
7862 case CPU_ONLINE_FROZEN:
7864 case CPU_DEAD_FROZEN:
7865 partition_sched_domains(1, NULL, NULL);
7874 static int update_runtime(struct notifier_block *nfb,
7875 unsigned long action, void *hcpu)
7877 int cpu = (int)(long)hcpu;
7880 case CPU_DOWN_PREPARE:
7881 case CPU_DOWN_PREPARE_FROZEN:
7882 disable_runtime(cpu_rq(cpu));
7885 case CPU_DOWN_FAILED:
7886 case CPU_DOWN_FAILED_FROZEN:
7888 case CPU_ONLINE_FROZEN:
7889 enable_runtime(cpu_rq(cpu));
7897 void __init sched_init_smp(void)
7899 cpumask_t non_isolated_cpus;
7901 #if defined(CONFIG_NUMA)
7902 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7904 BUG_ON(sched_group_nodes_bycpu == NULL);
7907 mutex_lock(&sched_domains_mutex);
7908 arch_init_sched_domains(&cpu_online_map);
7909 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7910 if (cpus_empty(non_isolated_cpus))
7911 cpu_set(smp_processor_id(), non_isolated_cpus);
7912 mutex_unlock(&sched_domains_mutex);
7915 #ifndef CONFIG_CPUSETS
7916 /* XXX: Theoretical race here - CPU may be hotplugged now */
7917 hotcpu_notifier(update_sched_domains, 0);
7920 /* RT runtime code needs to handle some hotplug events */
7921 hotcpu_notifier(update_runtime, 0);
7925 /* Move init over to a non-isolated CPU */
7926 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7928 sched_init_granularity();
7931 void __init sched_init_smp(void)
7933 sched_init_granularity();
7935 #endif /* CONFIG_SMP */
7937 int in_sched_functions(unsigned long addr)
7939 return in_lock_functions(addr) ||
7940 (addr >= (unsigned long)__sched_text_start
7941 && addr < (unsigned long)__sched_text_end);
7944 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7946 cfs_rq->tasks_timeline = RB_ROOT;
7947 INIT_LIST_HEAD(&cfs_rq->tasks);
7948 #ifdef CONFIG_FAIR_GROUP_SCHED
7951 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7954 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7956 struct rt_prio_array *array;
7959 array = &rt_rq->active;
7960 for (i = 0; i < MAX_RT_PRIO; i++) {
7961 INIT_LIST_HEAD(array->queue + i);
7962 __clear_bit(i, array->bitmap);
7964 /* delimiter for bitsearch: */
7965 __set_bit(MAX_RT_PRIO, array->bitmap);
7967 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7968 rt_rq->highest_prio = MAX_RT_PRIO;
7971 rt_rq->rt_nr_migratory = 0;
7972 rt_rq->overloaded = 0;
7976 rt_rq->rt_throttled = 0;
7977 rt_rq->rt_runtime = 0;
7978 spin_lock_init(&rt_rq->rt_runtime_lock);
7980 #ifdef CONFIG_RT_GROUP_SCHED
7981 rt_rq->rt_nr_boosted = 0;
7986 #ifdef CONFIG_FAIR_GROUP_SCHED
7987 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7988 struct sched_entity *se, int cpu, int add,
7989 struct sched_entity *parent)
7991 struct rq *rq = cpu_rq(cpu);
7992 tg->cfs_rq[cpu] = cfs_rq;
7993 init_cfs_rq(cfs_rq, rq);
7996 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7999 /* se could be NULL for init_task_group */
8004 se->cfs_rq = &rq->cfs;
8006 se->cfs_rq = parent->my_q;
8009 se->load.weight = tg->shares;
8010 se->load.inv_weight = 0;
8011 se->parent = parent;
8015 #ifdef CONFIG_RT_GROUP_SCHED
8016 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8017 struct sched_rt_entity *rt_se, int cpu, int add,
8018 struct sched_rt_entity *parent)
8020 struct rq *rq = cpu_rq(cpu);
8022 tg->rt_rq[cpu] = rt_rq;
8023 init_rt_rq(rt_rq, rq);
8025 rt_rq->rt_se = rt_se;
8026 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8028 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8030 tg->rt_se[cpu] = rt_se;
8035 rt_se->rt_rq = &rq->rt;
8037 rt_se->rt_rq = parent->my_q;
8039 rt_se->my_q = rt_rq;
8040 rt_se->parent = parent;
8041 INIT_LIST_HEAD(&rt_se->run_list);
8045 void __init sched_init(void)
8048 unsigned long alloc_size = 0, ptr;
8050 #ifdef CONFIG_FAIR_GROUP_SCHED
8051 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8053 #ifdef CONFIG_RT_GROUP_SCHED
8054 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8056 #ifdef CONFIG_USER_SCHED
8060 * As sched_init() is called before page_alloc is setup,
8061 * we use alloc_bootmem().
8064 ptr = (unsigned long)alloc_bootmem(alloc_size);
8066 #ifdef CONFIG_FAIR_GROUP_SCHED
8067 init_task_group.se = (struct sched_entity **)ptr;
8068 ptr += nr_cpu_ids * sizeof(void **);
8070 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8071 ptr += nr_cpu_ids * sizeof(void **);
8073 #ifdef CONFIG_USER_SCHED
8074 root_task_group.se = (struct sched_entity **)ptr;
8075 ptr += nr_cpu_ids * sizeof(void **);
8077 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8078 ptr += nr_cpu_ids * sizeof(void **);
8079 #endif /* CONFIG_USER_SCHED */
8080 #endif /* CONFIG_FAIR_GROUP_SCHED */
8081 #ifdef CONFIG_RT_GROUP_SCHED
8082 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8083 ptr += nr_cpu_ids * sizeof(void **);
8085 init_task_group.rt_rq = (struct rt_rq **)ptr;
8086 ptr += nr_cpu_ids * sizeof(void **);
8088 #ifdef CONFIG_USER_SCHED
8089 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8090 ptr += nr_cpu_ids * sizeof(void **);
8092 root_task_group.rt_rq = (struct rt_rq **)ptr;
8093 ptr += nr_cpu_ids * sizeof(void **);
8094 #endif /* CONFIG_USER_SCHED */
8095 #endif /* CONFIG_RT_GROUP_SCHED */
8099 init_defrootdomain();
8102 init_rt_bandwidth(&def_rt_bandwidth,
8103 global_rt_period(), global_rt_runtime());
8105 #ifdef CONFIG_RT_GROUP_SCHED
8106 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8107 global_rt_period(), global_rt_runtime());
8108 #ifdef CONFIG_USER_SCHED
8109 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8110 global_rt_period(), RUNTIME_INF);
8111 #endif /* CONFIG_USER_SCHED */
8112 #endif /* CONFIG_RT_GROUP_SCHED */
8114 #ifdef CONFIG_GROUP_SCHED
8115 list_add(&init_task_group.list, &task_groups);
8116 INIT_LIST_HEAD(&init_task_group.children);
8118 #ifdef CONFIG_USER_SCHED
8119 INIT_LIST_HEAD(&root_task_group.children);
8120 init_task_group.parent = &root_task_group;
8121 list_add(&init_task_group.siblings, &root_task_group.children);
8122 #endif /* CONFIG_USER_SCHED */
8123 #endif /* CONFIG_GROUP_SCHED */
8125 for_each_possible_cpu(i) {
8129 spin_lock_init(&rq->lock);
8131 init_cfs_rq(&rq->cfs, rq);
8132 init_rt_rq(&rq->rt, rq);
8133 #ifdef CONFIG_FAIR_GROUP_SCHED
8134 init_task_group.shares = init_task_group_load;
8135 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8136 #ifdef CONFIG_CGROUP_SCHED
8138 * How much cpu bandwidth does init_task_group get?
8140 * In case of task-groups formed thr' the cgroup filesystem, it
8141 * gets 100% of the cpu resources in the system. This overall
8142 * system cpu resource is divided among the tasks of
8143 * init_task_group and its child task-groups in a fair manner,
8144 * based on each entity's (task or task-group's) weight
8145 * (se->load.weight).
8147 * In other words, if init_task_group has 10 tasks of weight
8148 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8149 * then A0's share of the cpu resource is:
8151 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8153 * We achieve this by letting init_task_group's tasks sit
8154 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8156 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8157 #elif defined CONFIG_USER_SCHED
8158 root_task_group.shares = NICE_0_LOAD;
8159 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8161 * In case of task-groups formed thr' the user id of tasks,
8162 * init_task_group represents tasks belonging to root user.
8163 * Hence it forms a sibling of all subsequent groups formed.
8164 * In this case, init_task_group gets only a fraction of overall
8165 * system cpu resource, based on the weight assigned to root
8166 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8167 * by letting tasks of init_task_group sit in a separate cfs_rq
8168 * (init_cfs_rq) and having one entity represent this group of
8169 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8171 init_tg_cfs_entry(&init_task_group,
8172 &per_cpu(init_cfs_rq, i),
8173 &per_cpu(init_sched_entity, i), i, 1,
8174 root_task_group.se[i]);
8177 #endif /* CONFIG_FAIR_GROUP_SCHED */
8179 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8180 #ifdef CONFIG_RT_GROUP_SCHED
8181 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8182 #ifdef CONFIG_CGROUP_SCHED
8183 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8184 #elif defined CONFIG_USER_SCHED
8185 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8186 init_tg_rt_entry(&init_task_group,
8187 &per_cpu(init_rt_rq, i),
8188 &per_cpu(init_sched_rt_entity, i), i, 1,
8189 root_task_group.rt_se[i]);
8193 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8194 rq->cpu_load[j] = 0;
8198 rq->active_balance = 0;
8199 rq->next_balance = jiffies;
8203 rq->migration_thread = NULL;
8204 INIT_LIST_HEAD(&rq->migration_queue);
8205 rq_attach_root(rq, &def_root_domain);
8208 atomic_set(&rq->nr_iowait, 0);
8211 set_load_weight(&init_task);
8213 #ifdef CONFIG_PREEMPT_NOTIFIERS
8214 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8218 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8221 #ifdef CONFIG_RT_MUTEXES
8222 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8226 * The boot idle thread does lazy MMU switching as well:
8228 atomic_inc(&init_mm.mm_count);
8229 enter_lazy_tlb(&init_mm, current);
8232 * Make us the idle thread. Technically, schedule() should not be
8233 * called from this thread, however somewhere below it might be,
8234 * but because we are the idle thread, we just pick up running again
8235 * when this runqueue becomes "idle".
8237 init_idle(current, smp_processor_id());
8239 * During early bootup we pretend to be a normal task:
8241 current->sched_class = &fair_sched_class;
8243 scheduler_running = 1;
8246 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8247 void __might_sleep(char *file, int line)
8250 static unsigned long prev_jiffy; /* ratelimiting */
8252 if ((in_atomic() || irqs_disabled()) &&
8253 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8254 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8256 prev_jiffy = jiffies;
8257 printk(KERN_ERR "BUG: sleeping function called from invalid"
8258 " context at %s:%d\n", file, line);
8259 printk("in_atomic():%d, irqs_disabled():%d\n",
8260 in_atomic(), irqs_disabled());
8261 debug_show_held_locks(current);
8262 if (irqs_disabled())
8263 print_irqtrace_events(current);
8268 EXPORT_SYMBOL(__might_sleep);
8271 #ifdef CONFIG_MAGIC_SYSRQ
8272 static void normalize_task(struct rq *rq, struct task_struct *p)
8276 update_rq_clock(rq);
8277 on_rq = p->se.on_rq;
8279 deactivate_task(rq, p, 0);
8280 __setscheduler(rq, p, SCHED_NORMAL, 0);
8282 activate_task(rq, p, 0);
8283 resched_task(rq->curr);
8287 void normalize_rt_tasks(void)
8289 struct task_struct *g, *p;
8290 unsigned long flags;
8293 read_lock_irqsave(&tasklist_lock, flags);
8294 do_each_thread(g, p) {
8296 * Only normalize user tasks:
8301 p->se.exec_start = 0;
8302 #ifdef CONFIG_SCHEDSTATS
8303 p->se.wait_start = 0;
8304 p->se.sleep_start = 0;
8305 p->se.block_start = 0;
8310 * Renice negative nice level userspace
8313 if (TASK_NICE(p) < 0 && p->mm)
8314 set_user_nice(p, 0);
8318 spin_lock(&p->pi_lock);
8319 rq = __task_rq_lock(p);
8321 normalize_task(rq, p);
8323 __task_rq_unlock(rq);
8324 spin_unlock(&p->pi_lock);
8325 } while_each_thread(g, p);
8327 read_unlock_irqrestore(&tasklist_lock, flags);
8330 #endif /* CONFIG_MAGIC_SYSRQ */
8334 * These functions are only useful for the IA64 MCA handling.
8336 * They can only be called when the whole system has been
8337 * stopped - every CPU needs to be quiescent, and no scheduling
8338 * activity can take place. Using them for anything else would
8339 * be a serious bug, and as a result, they aren't even visible
8340 * under any other configuration.
8344 * curr_task - return the current task for a given cpu.
8345 * @cpu: the processor in question.
8347 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8349 struct task_struct *curr_task(int cpu)
8351 return cpu_curr(cpu);
8355 * set_curr_task - set the current task for a given cpu.
8356 * @cpu: the processor in question.
8357 * @p: the task pointer to set.
8359 * Description: This function must only be used when non-maskable interrupts
8360 * are serviced on a separate stack. It allows the architecture to switch the
8361 * notion of the current task on a cpu in a non-blocking manner. This function
8362 * must be called with all CPU's synchronized, and interrupts disabled, the
8363 * and caller must save the original value of the current task (see
8364 * curr_task() above) and restore that value before reenabling interrupts and
8365 * re-starting the system.
8367 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8369 void set_curr_task(int cpu, struct task_struct *p)
8376 #ifdef CONFIG_FAIR_GROUP_SCHED
8377 static void free_fair_sched_group(struct task_group *tg)
8381 for_each_possible_cpu(i) {
8383 kfree(tg->cfs_rq[i]);
8393 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8395 struct cfs_rq *cfs_rq;
8396 struct sched_entity *se, *parent_se;
8400 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8403 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8407 tg->shares = NICE_0_LOAD;
8409 for_each_possible_cpu(i) {
8412 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8413 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8417 se = kmalloc_node(sizeof(struct sched_entity),
8418 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8422 parent_se = parent ? parent->se[i] : NULL;
8423 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8432 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8434 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8435 &cpu_rq(cpu)->leaf_cfs_rq_list);
8438 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8440 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8442 #else /* !CONFG_FAIR_GROUP_SCHED */
8443 static inline void free_fair_sched_group(struct task_group *tg)
8448 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8453 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8457 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8460 #endif /* CONFIG_FAIR_GROUP_SCHED */
8462 #ifdef CONFIG_RT_GROUP_SCHED
8463 static void free_rt_sched_group(struct task_group *tg)
8467 destroy_rt_bandwidth(&tg->rt_bandwidth);
8469 for_each_possible_cpu(i) {
8471 kfree(tg->rt_rq[i]);
8473 kfree(tg->rt_se[i]);
8481 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8483 struct rt_rq *rt_rq;
8484 struct sched_rt_entity *rt_se, *parent_se;
8488 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8491 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8495 init_rt_bandwidth(&tg->rt_bandwidth,
8496 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8498 for_each_possible_cpu(i) {
8501 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8502 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8506 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8507 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8511 parent_se = parent ? parent->rt_se[i] : NULL;
8512 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8521 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8523 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8524 &cpu_rq(cpu)->leaf_rt_rq_list);
8527 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8529 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8531 #else /* !CONFIG_RT_GROUP_SCHED */
8532 static inline void free_rt_sched_group(struct task_group *tg)
8537 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8542 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8546 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8549 #endif /* CONFIG_RT_GROUP_SCHED */
8551 #ifdef CONFIG_GROUP_SCHED
8552 static void free_sched_group(struct task_group *tg)
8554 free_fair_sched_group(tg);
8555 free_rt_sched_group(tg);
8559 /* allocate runqueue etc for a new task group */
8560 struct task_group *sched_create_group(struct task_group *parent)
8562 struct task_group *tg;
8563 unsigned long flags;
8566 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8568 return ERR_PTR(-ENOMEM);
8570 if (!alloc_fair_sched_group(tg, parent))
8573 if (!alloc_rt_sched_group(tg, parent))
8576 spin_lock_irqsave(&task_group_lock, flags);
8577 for_each_possible_cpu(i) {
8578 register_fair_sched_group(tg, i);
8579 register_rt_sched_group(tg, i);
8581 list_add_rcu(&tg->list, &task_groups);
8583 WARN_ON(!parent); /* root should already exist */
8585 tg->parent = parent;
8586 INIT_LIST_HEAD(&tg->children);
8587 list_add_rcu(&tg->siblings, &parent->children);
8588 spin_unlock_irqrestore(&task_group_lock, flags);
8593 free_sched_group(tg);
8594 return ERR_PTR(-ENOMEM);
8597 /* rcu callback to free various structures associated with a task group */
8598 static void free_sched_group_rcu(struct rcu_head *rhp)
8600 /* now it should be safe to free those cfs_rqs */
8601 free_sched_group(container_of(rhp, struct task_group, rcu));
8604 /* Destroy runqueue etc associated with a task group */
8605 void sched_destroy_group(struct task_group *tg)
8607 unsigned long flags;
8610 spin_lock_irqsave(&task_group_lock, flags);
8611 for_each_possible_cpu(i) {
8612 unregister_fair_sched_group(tg, i);
8613 unregister_rt_sched_group(tg, i);
8615 list_del_rcu(&tg->list);
8616 list_del_rcu(&tg->siblings);
8617 spin_unlock_irqrestore(&task_group_lock, flags);
8619 /* wait for possible concurrent references to cfs_rqs complete */
8620 call_rcu(&tg->rcu, free_sched_group_rcu);
8623 /* change task's runqueue when it moves between groups.
8624 * The caller of this function should have put the task in its new group
8625 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8626 * reflect its new group.
8628 void sched_move_task(struct task_struct *tsk)
8631 unsigned long flags;
8634 rq = task_rq_lock(tsk, &flags);
8636 update_rq_clock(rq);
8638 running = task_current(rq, tsk);
8639 on_rq = tsk->se.on_rq;
8642 dequeue_task(rq, tsk, 0);
8643 if (unlikely(running))
8644 tsk->sched_class->put_prev_task(rq, tsk);
8646 set_task_rq(tsk, task_cpu(tsk));
8648 #ifdef CONFIG_FAIR_GROUP_SCHED
8649 if (tsk->sched_class->moved_group)
8650 tsk->sched_class->moved_group(tsk);
8653 if (unlikely(running))
8654 tsk->sched_class->set_curr_task(rq);
8656 enqueue_task(rq, tsk, 0);
8658 task_rq_unlock(rq, &flags);
8660 #endif /* CONFIG_GROUP_SCHED */
8662 #ifdef CONFIG_FAIR_GROUP_SCHED
8663 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8665 struct cfs_rq *cfs_rq = se->cfs_rq;
8670 dequeue_entity(cfs_rq, se, 0);
8672 se->load.weight = shares;
8673 se->load.inv_weight = 0;
8676 enqueue_entity(cfs_rq, se, 0);
8679 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8681 struct cfs_rq *cfs_rq = se->cfs_rq;
8682 struct rq *rq = cfs_rq->rq;
8683 unsigned long flags;
8685 spin_lock_irqsave(&rq->lock, flags);
8686 __set_se_shares(se, shares);
8687 spin_unlock_irqrestore(&rq->lock, flags);
8690 static DEFINE_MUTEX(shares_mutex);
8692 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8695 unsigned long flags;
8698 * We can't change the weight of the root cgroup.
8703 if (shares < MIN_SHARES)
8704 shares = MIN_SHARES;
8705 else if (shares > MAX_SHARES)
8706 shares = MAX_SHARES;
8708 mutex_lock(&shares_mutex);
8709 if (tg->shares == shares)
8712 spin_lock_irqsave(&task_group_lock, flags);
8713 for_each_possible_cpu(i)
8714 unregister_fair_sched_group(tg, i);
8715 list_del_rcu(&tg->siblings);
8716 spin_unlock_irqrestore(&task_group_lock, flags);
8718 /* wait for any ongoing reference to this group to finish */
8719 synchronize_sched();
8722 * Now we are free to modify the group's share on each cpu
8723 * w/o tripping rebalance_share or load_balance_fair.
8725 tg->shares = shares;
8726 for_each_possible_cpu(i) {
8730 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8731 set_se_shares(tg->se[i], shares);
8735 * Enable load balance activity on this group, by inserting it back on
8736 * each cpu's rq->leaf_cfs_rq_list.
8738 spin_lock_irqsave(&task_group_lock, flags);
8739 for_each_possible_cpu(i)
8740 register_fair_sched_group(tg, i);
8741 list_add_rcu(&tg->siblings, &tg->parent->children);
8742 spin_unlock_irqrestore(&task_group_lock, flags);
8744 mutex_unlock(&shares_mutex);
8748 unsigned long sched_group_shares(struct task_group *tg)
8754 #ifdef CONFIG_RT_GROUP_SCHED
8756 * Ensure that the real time constraints are schedulable.
8758 static DEFINE_MUTEX(rt_constraints_mutex);
8760 static unsigned long to_ratio(u64 period, u64 runtime)
8762 if (runtime == RUNTIME_INF)
8765 return div64_u64(runtime << 16, period);
8768 #ifdef CONFIG_CGROUP_SCHED
8769 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8771 struct task_group *tgi, *parent = tg->parent;
8772 unsigned long total = 0;
8775 if (global_rt_period() < period)
8778 return to_ratio(period, runtime) <
8779 to_ratio(global_rt_period(), global_rt_runtime());
8782 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8786 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8790 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8791 tgi->rt_bandwidth.rt_runtime);
8795 return total + to_ratio(period, runtime) <=
8796 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8797 parent->rt_bandwidth.rt_runtime);
8799 #elif defined CONFIG_USER_SCHED
8800 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8802 struct task_group *tgi;
8803 unsigned long total = 0;
8804 unsigned long global_ratio =
8805 to_ratio(global_rt_period(), global_rt_runtime());
8808 list_for_each_entry_rcu(tgi, &task_groups, list) {
8812 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8813 tgi->rt_bandwidth.rt_runtime);
8817 return total + to_ratio(period, runtime) < global_ratio;
8821 /* Must be called with tasklist_lock held */
8822 static inline int tg_has_rt_tasks(struct task_group *tg)
8824 struct task_struct *g, *p;
8825 do_each_thread(g, p) {
8826 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8828 } while_each_thread(g, p);
8832 static int tg_set_bandwidth(struct task_group *tg,
8833 u64 rt_period, u64 rt_runtime)
8837 mutex_lock(&rt_constraints_mutex);
8838 read_lock(&tasklist_lock);
8839 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8843 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8848 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8849 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8850 tg->rt_bandwidth.rt_runtime = rt_runtime;
8852 for_each_possible_cpu(i) {
8853 struct rt_rq *rt_rq = tg->rt_rq[i];
8855 spin_lock(&rt_rq->rt_runtime_lock);
8856 rt_rq->rt_runtime = rt_runtime;
8857 spin_unlock(&rt_rq->rt_runtime_lock);
8859 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8861 read_unlock(&tasklist_lock);
8862 mutex_unlock(&rt_constraints_mutex);
8867 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8869 u64 rt_runtime, rt_period;
8871 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8872 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8873 if (rt_runtime_us < 0)
8874 rt_runtime = RUNTIME_INF;
8876 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8879 long sched_group_rt_runtime(struct task_group *tg)
8883 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8886 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8887 do_div(rt_runtime_us, NSEC_PER_USEC);
8888 return rt_runtime_us;
8891 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8893 u64 rt_runtime, rt_period;
8895 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8896 rt_runtime = tg->rt_bandwidth.rt_runtime;
8901 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8904 long sched_group_rt_period(struct task_group *tg)
8908 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8909 do_div(rt_period_us, NSEC_PER_USEC);
8910 return rt_period_us;
8913 static int sched_rt_global_constraints(void)
8915 struct task_group *tg = &root_task_group;
8916 u64 rt_runtime, rt_period;
8919 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8920 rt_runtime = tg->rt_bandwidth.rt_runtime;
8922 mutex_lock(&rt_constraints_mutex);
8923 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8925 mutex_unlock(&rt_constraints_mutex);
8929 #else /* !CONFIG_RT_GROUP_SCHED */
8930 static int sched_rt_global_constraints(void)
8932 unsigned long flags;
8935 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8936 for_each_possible_cpu(i) {
8937 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8939 spin_lock(&rt_rq->rt_runtime_lock);
8940 rt_rq->rt_runtime = global_rt_runtime();
8941 spin_unlock(&rt_rq->rt_runtime_lock);
8943 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8947 #endif /* CONFIG_RT_GROUP_SCHED */
8949 int sched_rt_handler(struct ctl_table *table, int write,
8950 struct file *filp, void __user *buffer, size_t *lenp,
8954 int old_period, old_runtime;
8955 static DEFINE_MUTEX(mutex);
8958 old_period = sysctl_sched_rt_period;
8959 old_runtime = sysctl_sched_rt_runtime;
8961 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8963 if (!ret && write) {
8964 ret = sched_rt_global_constraints();
8966 sysctl_sched_rt_period = old_period;
8967 sysctl_sched_rt_runtime = old_runtime;
8969 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8970 def_rt_bandwidth.rt_period =
8971 ns_to_ktime(global_rt_period());
8974 mutex_unlock(&mutex);
8979 #ifdef CONFIG_CGROUP_SCHED
8981 /* return corresponding task_group object of a cgroup */
8982 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8984 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8985 struct task_group, css);
8988 static struct cgroup_subsys_state *
8989 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8991 struct task_group *tg, *parent;
8993 if (!cgrp->parent) {
8994 /* This is early initialization for the top cgroup */
8995 init_task_group.css.cgroup = cgrp;
8996 return &init_task_group.css;
8999 parent = cgroup_tg(cgrp->parent);
9000 tg = sched_create_group(parent);
9002 return ERR_PTR(-ENOMEM);
9004 /* Bind the cgroup to task_group object we just created */
9005 tg->css.cgroup = cgrp;
9011 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9013 struct task_group *tg = cgroup_tg(cgrp);
9015 sched_destroy_group(tg);
9019 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9020 struct task_struct *tsk)
9022 #ifdef CONFIG_RT_GROUP_SCHED
9023 /* Don't accept realtime tasks when there is no way for them to run */
9024 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9027 /* We don't support RT-tasks being in separate groups */
9028 if (tsk->sched_class != &fair_sched_class)
9036 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9037 struct cgroup *old_cont, struct task_struct *tsk)
9039 sched_move_task(tsk);
9042 #ifdef CONFIG_FAIR_GROUP_SCHED
9043 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9046 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9049 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9051 struct task_group *tg = cgroup_tg(cgrp);
9053 return (u64) tg->shares;
9055 #endif /* CONFIG_FAIR_GROUP_SCHED */
9057 #ifdef CONFIG_RT_GROUP_SCHED
9058 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9061 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9064 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9066 return sched_group_rt_runtime(cgroup_tg(cgrp));
9069 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9072 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9075 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9077 return sched_group_rt_period(cgroup_tg(cgrp));
9079 #endif /* CONFIG_RT_GROUP_SCHED */
9081 static struct cftype cpu_files[] = {
9082 #ifdef CONFIG_FAIR_GROUP_SCHED
9085 .read_u64 = cpu_shares_read_u64,
9086 .write_u64 = cpu_shares_write_u64,
9089 #ifdef CONFIG_RT_GROUP_SCHED
9091 .name = "rt_runtime_us",
9092 .read_s64 = cpu_rt_runtime_read,
9093 .write_s64 = cpu_rt_runtime_write,
9096 .name = "rt_period_us",
9097 .read_u64 = cpu_rt_period_read_uint,
9098 .write_u64 = cpu_rt_period_write_uint,
9103 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9105 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9108 struct cgroup_subsys cpu_cgroup_subsys = {
9110 .create = cpu_cgroup_create,
9111 .destroy = cpu_cgroup_destroy,
9112 .can_attach = cpu_cgroup_can_attach,
9113 .attach = cpu_cgroup_attach,
9114 .populate = cpu_cgroup_populate,
9115 .subsys_id = cpu_cgroup_subsys_id,
9119 #endif /* CONFIG_CGROUP_SCHED */
9121 #ifdef CONFIG_CGROUP_CPUACCT
9124 * CPU accounting code for task groups.
9126 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9127 * (balbir@in.ibm.com).
9130 /* track cpu usage of a group of tasks */
9132 struct cgroup_subsys_state css;
9133 /* cpuusage holds pointer to a u64-type object on every cpu */
9137 struct cgroup_subsys cpuacct_subsys;
9139 /* return cpu accounting group corresponding to this container */
9140 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9142 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9143 struct cpuacct, css);
9146 /* return cpu accounting group to which this task belongs */
9147 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9149 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9150 struct cpuacct, css);
9153 /* create a new cpu accounting group */
9154 static struct cgroup_subsys_state *cpuacct_create(
9155 struct cgroup_subsys *ss, struct cgroup *cgrp)
9157 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9160 return ERR_PTR(-ENOMEM);
9162 ca->cpuusage = alloc_percpu(u64);
9163 if (!ca->cpuusage) {
9165 return ERR_PTR(-ENOMEM);
9171 /* destroy an existing cpu accounting group */
9173 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9175 struct cpuacct *ca = cgroup_ca(cgrp);
9177 free_percpu(ca->cpuusage);
9181 /* return total cpu usage (in nanoseconds) of a group */
9182 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9184 struct cpuacct *ca = cgroup_ca(cgrp);
9185 u64 totalcpuusage = 0;
9188 for_each_possible_cpu(i) {
9189 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9192 * Take rq->lock to make 64-bit addition safe on 32-bit
9195 spin_lock_irq(&cpu_rq(i)->lock);
9196 totalcpuusage += *cpuusage;
9197 spin_unlock_irq(&cpu_rq(i)->lock);
9200 return totalcpuusage;
9203 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9206 struct cpuacct *ca = cgroup_ca(cgrp);
9215 for_each_possible_cpu(i) {
9216 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9218 spin_lock_irq(&cpu_rq(i)->lock);
9220 spin_unlock_irq(&cpu_rq(i)->lock);
9226 static struct cftype files[] = {
9229 .read_u64 = cpuusage_read,
9230 .write_u64 = cpuusage_write,
9234 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9236 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9240 * charge this task's execution time to its accounting group.
9242 * called with rq->lock held.
9244 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9248 if (!cpuacct_subsys.active)
9253 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9255 *cpuusage += cputime;
9259 struct cgroup_subsys cpuacct_subsys = {
9261 .create = cpuacct_create,
9262 .destroy = cpuacct_destroy,
9263 .populate = cpuacct_populate,
9264 .subsys_id = cpuacct_subsys_id,
9266 #endif /* CONFIG_CGROUP_CPUACCT */