4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
75 #include <asm/irq_regs.h>
77 #include "sched_cpupri.h"
80 * Convert user-nice values [ -20 ... 0 ... 19 ]
81 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
84 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
85 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
86 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
89 * 'User priority' is the nice value converted to something we
90 * can work with better when scaling various scheduler parameters,
91 * it's a [ 0 ... 39 ] range.
93 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
94 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
95 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
98 * Helpers for converting nanosecond timing to jiffy resolution
100 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
109 * Timeslices get refilled after they expire.
111 #define DEF_TIMESLICE (100 * HZ / 1000)
114 * single value that denotes runtime == period, ie unlimited time.
116 #define RUNTIME_INF ((u64)~0ULL)
120 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
121 * Since cpu_power is a 'constant', we can use a reciprocal divide.
123 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
125 return reciprocal_divide(load, sg->reciprocal_cpu_power);
129 * Each time a sched group cpu_power is changed,
130 * we must compute its reciprocal value
132 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
134 sg->__cpu_power += val;
135 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
139 static inline int rt_policy(int policy)
141 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
146 static inline int task_has_rt_policy(struct task_struct *p)
148 return rt_policy(p->policy);
152 * This is the priority-queue data structure of the RT scheduling class:
154 struct rt_prio_array {
155 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
156 struct list_head queue[MAX_RT_PRIO];
159 struct rt_bandwidth {
160 /* nests inside the rq lock: */
161 spinlock_t rt_runtime_lock;
164 struct hrtimer rt_period_timer;
167 static struct rt_bandwidth def_rt_bandwidth;
169 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
171 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
173 struct rt_bandwidth *rt_b =
174 container_of(timer, struct rt_bandwidth, rt_period_timer);
180 now = hrtimer_cb_get_time(timer);
181 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
186 idle = do_sched_rt_period_timer(rt_b, overrun);
189 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
193 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
195 rt_b->rt_period = ns_to_ktime(period);
196 rt_b->rt_runtime = runtime;
198 spin_lock_init(&rt_b->rt_runtime_lock);
200 hrtimer_init(&rt_b->rt_period_timer,
201 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
202 rt_b->rt_period_timer.function = sched_rt_period_timer;
203 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
206 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
210 if (rt_b->rt_runtime == RUNTIME_INF)
213 if (hrtimer_active(&rt_b->rt_period_timer))
216 spin_lock(&rt_b->rt_runtime_lock);
218 if (hrtimer_active(&rt_b->rt_period_timer))
221 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
222 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
223 hrtimer_start(&rt_b->rt_period_timer,
224 rt_b->rt_period_timer.expires,
227 spin_unlock(&rt_b->rt_runtime_lock);
230 #ifdef CONFIG_RT_GROUP_SCHED
231 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
233 hrtimer_cancel(&rt_b->rt_period_timer);
238 * sched_domains_mutex serializes calls to arch_init_sched_domains,
239 * detach_destroy_domains and partition_sched_domains.
241 static DEFINE_MUTEX(sched_domains_mutex);
243 #ifdef CONFIG_GROUP_SCHED
245 #include <linux/cgroup.h>
249 static LIST_HEAD(task_groups);
251 /* task group related information */
253 #ifdef CONFIG_CGROUP_SCHED
254 struct cgroup_subsys_state css;
257 #ifdef CONFIG_FAIR_GROUP_SCHED
258 /* schedulable entities of this group on each cpu */
259 struct sched_entity **se;
260 /* runqueue "owned" by this group on each cpu */
261 struct cfs_rq **cfs_rq;
262 unsigned long shares;
265 #ifdef CONFIG_RT_GROUP_SCHED
266 struct sched_rt_entity **rt_se;
267 struct rt_rq **rt_rq;
269 struct rt_bandwidth rt_bandwidth;
273 struct list_head list;
275 struct task_group *parent;
276 struct list_head siblings;
277 struct list_head children;
280 #ifdef CONFIG_USER_SCHED
284 * Every UID task group (including init_task_group aka UID-0) will
285 * be a child to this group.
287 struct task_group root_task_group;
289 #ifdef CONFIG_FAIR_GROUP_SCHED
290 /* Default task group's sched entity on each cpu */
291 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
292 /* Default task group's cfs_rq on each cpu */
293 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
294 #endif /* CONFIG_FAIR_GROUP_SCHED */
296 #ifdef CONFIG_RT_GROUP_SCHED
297 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
298 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
299 #endif /* CONFIG_RT_GROUP_SCHED */
300 #else /* !CONFIG_FAIR_GROUP_SCHED */
301 #define root_task_group init_task_group
302 #endif /* CONFIG_FAIR_GROUP_SCHED */
304 /* task_group_lock serializes add/remove of task groups and also changes to
305 * a task group's cpu shares.
307 static DEFINE_SPINLOCK(task_group_lock);
309 #ifdef CONFIG_FAIR_GROUP_SCHED
310 #ifdef CONFIG_USER_SCHED
311 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
312 #else /* !CONFIG_USER_SCHED */
313 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
314 #endif /* CONFIG_USER_SCHED */
317 * A weight of 0 or 1 can cause arithmetics problems.
318 * A weight of a cfs_rq is the sum of weights of which entities
319 * are queued on this cfs_rq, so a weight of a entity should not be
320 * too large, so as the shares value of a task group.
321 * (The default weight is 1024 - so there's no practical
322 * limitation from this.)
325 #define MAX_SHARES (1UL << 18)
327 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
330 /* Default task group.
331 * Every task in system belong to this group at bootup.
333 struct task_group init_task_group;
335 /* return group to which a task belongs */
336 static inline struct task_group *task_group(struct task_struct *p)
338 struct task_group *tg;
340 #ifdef CONFIG_USER_SCHED
342 #elif defined(CONFIG_CGROUP_SCHED)
343 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
344 struct task_group, css);
346 tg = &init_task_group;
351 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
352 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
354 #ifdef CONFIG_FAIR_GROUP_SCHED
355 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
356 p->se.parent = task_group(p)->se[cpu];
359 #ifdef CONFIG_RT_GROUP_SCHED
360 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
361 p->rt.parent = task_group(p)->rt_se[cpu];
367 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
369 #endif /* CONFIG_GROUP_SCHED */
371 /* CFS-related fields in a runqueue */
373 struct load_weight load;
374 unsigned long nr_running;
380 struct rb_root tasks_timeline;
381 struct rb_node *rb_leftmost;
383 struct list_head tasks;
384 struct list_head *balance_iterator;
387 * 'curr' points to currently running entity on this cfs_rq.
388 * It is set to NULL otherwise (i.e when none are currently running).
390 struct sched_entity *curr, *next;
392 unsigned long nr_spread_over;
394 #ifdef CONFIG_FAIR_GROUP_SCHED
395 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
398 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
399 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
400 * (like users, containers etc.)
402 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
403 * list is used during load balance.
405 struct list_head leaf_cfs_rq_list;
406 struct task_group *tg; /* group that "owns" this runqueue */
410 * the part of load.weight contributed by tasks
412 unsigned long task_weight;
415 * h_load = weight * f(tg)
417 * Where f(tg) is the recursive weight fraction assigned to
420 unsigned long h_load;
423 * this cpu's part of tg->shares
425 unsigned long shares;
430 /* Real-Time classes' related field in a runqueue: */
432 struct rt_prio_array active;
433 unsigned long rt_nr_running;
434 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
435 int highest_prio; /* highest queued rt task prio */
438 unsigned long rt_nr_migratory;
444 /* Nests inside the rq lock: */
445 spinlock_t rt_runtime_lock;
447 #ifdef CONFIG_RT_GROUP_SCHED
448 unsigned long rt_nr_boosted;
451 struct list_head leaf_rt_rq_list;
452 struct task_group *tg;
453 struct sched_rt_entity *rt_se;
460 * We add the notion of a root-domain which will be used to define per-domain
461 * variables. Each exclusive cpuset essentially defines an island domain by
462 * fully partitioning the member cpus from any other cpuset. Whenever a new
463 * exclusive cpuset is created, we also create and attach a new root-domain
473 * The "RT overload" flag: it gets set if a CPU has more than
474 * one runnable RT task.
479 struct cpupri cpupri;
484 * By default the system creates a single root-domain with all cpus as
485 * members (mimicking the global state we have today).
487 static struct root_domain def_root_domain;
492 * This is the main, per-CPU runqueue data structure.
494 * Locking rule: those places that want to lock multiple runqueues
495 * (such as the load balancing or the thread migration code), lock
496 * acquire operations must be ordered by ascending &runqueue.
503 * nr_running and cpu_load should be in the same cacheline because
504 * remote CPUs use both these fields when doing load calculation.
506 unsigned long nr_running;
507 #define CPU_LOAD_IDX_MAX 5
508 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
509 unsigned char idle_at_tick;
511 unsigned long last_tick_seen;
512 unsigned char in_nohz_recently;
514 /* capture load from *all* tasks on this cpu: */
515 struct load_weight load;
516 unsigned long nr_load_updates;
522 #ifdef CONFIG_FAIR_GROUP_SCHED
523 /* list of leaf cfs_rq on this cpu: */
524 struct list_head leaf_cfs_rq_list;
526 #ifdef CONFIG_RT_GROUP_SCHED
527 struct list_head leaf_rt_rq_list;
531 * This is part of a global counter where only the total sum
532 * over all CPUs matters. A task can increase this counter on
533 * one CPU and if it got migrated afterwards it may decrease
534 * it on another CPU. Always updated under the runqueue lock:
536 unsigned long nr_uninterruptible;
538 struct task_struct *curr, *idle;
539 unsigned long next_balance;
540 struct mm_struct *prev_mm;
547 struct root_domain *rd;
548 struct sched_domain *sd;
550 /* For active balancing */
553 /* cpu of this runqueue: */
557 struct task_struct *migration_thread;
558 struct list_head migration_queue;
561 #ifdef CONFIG_SCHED_HRTICK
562 unsigned long hrtick_flags;
563 ktime_t hrtick_expire;
564 struct hrtimer hrtick_timer;
567 #ifdef CONFIG_SCHEDSTATS
569 struct sched_info rq_sched_info;
571 /* sys_sched_yield() stats */
572 unsigned int yld_exp_empty;
573 unsigned int yld_act_empty;
574 unsigned int yld_both_empty;
575 unsigned int yld_count;
577 /* schedule() stats */
578 unsigned int sched_switch;
579 unsigned int sched_count;
580 unsigned int sched_goidle;
582 /* try_to_wake_up() stats */
583 unsigned int ttwu_count;
584 unsigned int ttwu_local;
587 unsigned int bkl_count;
589 struct lock_class_key rq_lock_key;
592 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
594 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
596 rq->curr->sched_class->check_preempt_curr(rq, p);
599 static inline int cpu_of(struct rq *rq)
609 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
610 * See detach_destroy_domains: synchronize_sched for details.
612 * The domain tree of any CPU may only be accessed from within
613 * preempt-disabled sections.
615 #define for_each_domain(cpu, __sd) \
616 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
618 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
619 #define this_rq() (&__get_cpu_var(runqueues))
620 #define task_rq(p) cpu_rq(task_cpu(p))
621 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
623 static inline void update_rq_clock(struct rq *rq)
625 rq->clock = sched_clock_cpu(cpu_of(rq));
629 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
631 #ifdef CONFIG_SCHED_DEBUG
632 # define const_debug __read_mostly
634 # define const_debug static const
638 * Debugging: various feature bits
641 #define SCHED_FEAT(name, enabled) \
642 __SCHED_FEAT_##name ,
645 #include "sched_features.h"
650 #define SCHED_FEAT(name, enabled) \
651 (1UL << __SCHED_FEAT_##name) * enabled |
653 const_debug unsigned int sysctl_sched_features =
654 #include "sched_features.h"
659 #ifdef CONFIG_SCHED_DEBUG
660 #define SCHED_FEAT(name, enabled) \
663 static __read_mostly char *sched_feat_names[] = {
664 #include "sched_features.h"
670 static int sched_feat_open(struct inode *inode, struct file *filp)
672 filp->private_data = inode->i_private;
677 sched_feat_read(struct file *filp, char __user *ubuf,
678 size_t cnt, loff_t *ppos)
685 for (i = 0; sched_feat_names[i]; i++) {
686 len += strlen(sched_feat_names[i]);
690 buf = kmalloc(len + 2, GFP_KERNEL);
694 for (i = 0; sched_feat_names[i]; i++) {
695 if (sysctl_sched_features & (1UL << i))
696 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
698 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
701 r += sprintf(buf + r, "\n");
702 WARN_ON(r >= len + 2);
704 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
712 sched_feat_write(struct file *filp, const char __user *ubuf,
713 size_t cnt, loff_t *ppos)
723 if (copy_from_user(&buf, ubuf, cnt))
728 if (strncmp(buf, "NO_", 3) == 0) {
733 for (i = 0; sched_feat_names[i]; i++) {
734 int len = strlen(sched_feat_names[i]);
736 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
738 sysctl_sched_features &= ~(1UL << i);
740 sysctl_sched_features |= (1UL << i);
745 if (!sched_feat_names[i])
753 static struct file_operations sched_feat_fops = {
754 .open = sched_feat_open,
755 .read = sched_feat_read,
756 .write = sched_feat_write,
759 static __init int sched_init_debug(void)
761 debugfs_create_file("sched_features", 0644, NULL, NULL,
766 late_initcall(sched_init_debug);
770 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
773 * Number of tasks to iterate in a single balance run.
774 * Limited because this is done with IRQs disabled.
776 const_debug unsigned int sysctl_sched_nr_migrate = 32;
779 * period over which we measure -rt task cpu usage in us.
782 unsigned int sysctl_sched_rt_period = 1000000;
784 static __read_mostly int scheduler_running;
787 * part of the period that we allow rt tasks to run in us.
790 int sysctl_sched_rt_runtime = 950000;
792 static inline u64 global_rt_period(void)
794 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
797 static inline u64 global_rt_runtime(void)
799 if (sysctl_sched_rt_period < 0)
802 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
805 #ifndef prepare_arch_switch
806 # define prepare_arch_switch(next) do { } while (0)
808 #ifndef finish_arch_switch
809 # define finish_arch_switch(prev) do { } while (0)
812 static inline int task_current(struct rq *rq, struct task_struct *p)
814 return rq->curr == p;
817 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
818 static inline int task_running(struct rq *rq, struct task_struct *p)
820 return task_current(rq, p);
823 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
827 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
829 #ifdef CONFIG_DEBUG_SPINLOCK
830 /* this is a valid case when another task releases the spinlock */
831 rq->lock.owner = current;
834 * If we are tracking spinlock dependencies then we have to
835 * fix up the runqueue lock - which gets 'carried over' from
838 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
840 spin_unlock_irq(&rq->lock);
843 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
844 static inline int task_running(struct rq *rq, struct task_struct *p)
849 return task_current(rq, p);
853 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
857 * We can optimise this out completely for !SMP, because the
858 * SMP rebalancing from interrupt is the only thing that cares
863 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
864 spin_unlock_irq(&rq->lock);
866 spin_unlock(&rq->lock);
870 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
874 * After ->oncpu is cleared, the task can be moved to a different CPU.
875 * We must ensure this doesn't happen until the switch is completely
881 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
885 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
888 * __task_rq_lock - lock the runqueue a given task resides on.
889 * Must be called interrupts disabled.
891 static inline struct rq *__task_rq_lock(struct task_struct *p)
895 struct rq *rq = task_rq(p);
896 spin_lock(&rq->lock);
897 if (likely(rq == task_rq(p)))
899 spin_unlock(&rq->lock);
904 * task_rq_lock - lock the runqueue a given task resides on and disable
905 * interrupts. Note the ordering: we can safely lookup the task_rq without
906 * explicitly disabling preemption.
908 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
914 local_irq_save(*flags);
916 spin_lock(&rq->lock);
917 if (likely(rq == task_rq(p)))
919 spin_unlock_irqrestore(&rq->lock, *flags);
923 static void __task_rq_unlock(struct rq *rq)
926 spin_unlock(&rq->lock);
929 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
932 spin_unlock_irqrestore(&rq->lock, *flags);
936 * this_rq_lock - lock this runqueue and disable interrupts.
938 static struct rq *this_rq_lock(void)
945 spin_lock(&rq->lock);
950 static void __resched_task(struct task_struct *p, int tif_bit);
952 static inline void resched_task(struct task_struct *p)
954 __resched_task(p, TIF_NEED_RESCHED);
957 #ifdef CONFIG_SCHED_HRTICK
959 * Use HR-timers to deliver accurate preemption points.
961 * Its all a bit involved since we cannot program an hrt while holding the
962 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
965 * When we get rescheduled we reprogram the hrtick_timer outside of the
968 static inline void resched_hrt(struct task_struct *p)
970 __resched_task(p, TIF_HRTICK_RESCHED);
973 static inline void resched_rq(struct rq *rq)
977 spin_lock_irqsave(&rq->lock, flags);
978 resched_task(rq->curr);
979 spin_unlock_irqrestore(&rq->lock, flags);
983 HRTICK_SET, /* re-programm hrtick_timer */
984 HRTICK_RESET, /* not a new slice */
985 HRTICK_BLOCK, /* stop hrtick operations */
990 * - enabled by features
991 * - hrtimer is actually high res
993 static inline int hrtick_enabled(struct rq *rq)
995 if (!sched_feat(HRTICK))
997 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
999 return hrtimer_is_hres_active(&rq->hrtick_timer);
1003 * Called to set the hrtick timer state.
1005 * called with rq->lock held and irqs disabled
1007 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1009 assert_spin_locked(&rq->lock);
1012 * preempt at: now + delay
1015 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1017 * indicate we need to program the timer
1019 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1021 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1024 * New slices are called from the schedule path and don't need a
1025 * forced reschedule.
1028 resched_hrt(rq->curr);
1031 static void hrtick_clear(struct rq *rq)
1033 if (hrtimer_active(&rq->hrtick_timer))
1034 hrtimer_cancel(&rq->hrtick_timer);
1038 * Update the timer from the possible pending state.
1040 static void hrtick_set(struct rq *rq)
1044 unsigned long flags;
1046 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1048 spin_lock_irqsave(&rq->lock, flags);
1049 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1050 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1051 time = rq->hrtick_expire;
1052 clear_thread_flag(TIF_HRTICK_RESCHED);
1053 spin_unlock_irqrestore(&rq->lock, flags);
1056 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1057 if (reset && !hrtimer_active(&rq->hrtick_timer))
1064 * High-resolution timer tick.
1065 * Runs from hardirq context with interrupts disabled.
1067 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1069 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1071 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1073 spin_lock(&rq->lock);
1074 update_rq_clock(rq);
1075 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1076 spin_unlock(&rq->lock);
1078 return HRTIMER_NORESTART;
1082 static void hotplug_hrtick_disable(int cpu)
1084 struct rq *rq = cpu_rq(cpu);
1085 unsigned long flags;
1087 spin_lock_irqsave(&rq->lock, flags);
1088 rq->hrtick_flags = 0;
1089 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1090 spin_unlock_irqrestore(&rq->lock, flags);
1095 static void hotplug_hrtick_enable(int cpu)
1097 struct rq *rq = cpu_rq(cpu);
1098 unsigned long flags;
1100 spin_lock_irqsave(&rq->lock, flags);
1101 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1102 spin_unlock_irqrestore(&rq->lock, flags);
1106 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1108 int cpu = (int)(long)hcpu;
1111 case CPU_UP_CANCELED:
1112 case CPU_UP_CANCELED_FROZEN:
1113 case CPU_DOWN_PREPARE:
1114 case CPU_DOWN_PREPARE_FROZEN:
1116 case CPU_DEAD_FROZEN:
1117 hotplug_hrtick_disable(cpu);
1120 case CPU_UP_PREPARE:
1121 case CPU_UP_PREPARE_FROZEN:
1122 case CPU_DOWN_FAILED:
1123 case CPU_DOWN_FAILED_FROZEN:
1125 case CPU_ONLINE_FROZEN:
1126 hotplug_hrtick_enable(cpu);
1133 static void init_hrtick(void)
1135 hotcpu_notifier(hotplug_hrtick, 0);
1137 #endif /* CONFIG_SMP */
1139 static void init_rq_hrtick(struct rq *rq)
1141 rq->hrtick_flags = 0;
1142 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1143 rq->hrtick_timer.function = hrtick;
1144 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1147 void hrtick_resched(void)
1150 unsigned long flags;
1152 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1155 local_irq_save(flags);
1156 rq = cpu_rq(smp_processor_id());
1158 local_irq_restore(flags);
1161 static inline void hrtick_clear(struct rq *rq)
1165 static inline void hrtick_set(struct rq *rq)
1169 static inline void init_rq_hrtick(struct rq *rq)
1173 void hrtick_resched(void)
1177 static inline void init_hrtick(void)
1183 * resched_task - mark a task 'to be rescheduled now'.
1185 * On UP this means the setting of the need_resched flag, on SMP it
1186 * might also involve a cross-CPU call to trigger the scheduler on
1191 #ifndef tsk_is_polling
1192 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1195 static void __resched_task(struct task_struct *p, int tif_bit)
1199 assert_spin_locked(&task_rq(p)->lock);
1201 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1204 set_tsk_thread_flag(p, tif_bit);
1207 if (cpu == smp_processor_id())
1210 /* NEED_RESCHED must be visible before we test polling */
1212 if (!tsk_is_polling(p))
1213 smp_send_reschedule(cpu);
1216 static void resched_cpu(int cpu)
1218 struct rq *rq = cpu_rq(cpu);
1219 unsigned long flags;
1221 if (!spin_trylock_irqsave(&rq->lock, flags))
1223 resched_task(cpu_curr(cpu));
1224 spin_unlock_irqrestore(&rq->lock, flags);
1229 * When add_timer_on() enqueues a timer into the timer wheel of an
1230 * idle CPU then this timer might expire before the next timer event
1231 * which is scheduled to wake up that CPU. In case of a completely
1232 * idle system the next event might even be infinite time into the
1233 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1234 * leaves the inner idle loop so the newly added timer is taken into
1235 * account when the CPU goes back to idle and evaluates the timer
1236 * wheel for the next timer event.
1238 void wake_up_idle_cpu(int cpu)
1240 struct rq *rq = cpu_rq(cpu);
1242 if (cpu == smp_processor_id())
1246 * This is safe, as this function is called with the timer
1247 * wheel base lock of (cpu) held. When the CPU is on the way
1248 * to idle and has not yet set rq->curr to idle then it will
1249 * be serialized on the timer wheel base lock and take the new
1250 * timer into account automatically.
1252 if (rq->curr != rq->idle)
1256 * We can set TIF_RESCHED on the idle task of the other CPU
1257 * lockless. The worst case is that the other CPU runs the
1258 * idle task through an additional NOOP schedule()
1260 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1262 /* NEED_RESCHED must be visible before we test polling */
1264 if (!tsk_is_polling(rq->idle))
1265 smp_send_reschedule(cpu);
1267 #endif /* CONFIG_NO_HZ */
1269 #else /* !CONFIG_SMP */
1270 static void __resched_task(struct task_struct *p, int tif_bit)
1272 assert_spin_locked(&task_rq(p)->lock);
1273 set_tsk_thread_flag(p, tif_bit);
1275 #endif /* CONFIG_SMP */
1277 #if BITS_PER_LONG == 32
1278 # define WMULT_CONST (~0UL)
1280 # define WMULT_CONST (1UL << 32)
1283 #define WMULT_SHIFT 32
1286 * Shift right and round:
1288 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1291 * delta *= weight / lw
1293 static unsigned long
1294 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1295 struct load_weight *lw)
1299 if (!lw->inv_weight) {
1300 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1303 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1307 tmp = (u64)delta_exec * weight;
1309 * Check whether we'd overflow the 64-bit multiplication:
1311 if (unlikely(tmp > WMULT_CONST))
1312 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1315 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1317 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1320 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1326 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1333 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1334 * of tasks with abnormal "nice" values across CPUs the contribution that
1335 * each task makes to its run queue's load is weighted according to its
1336 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1337 * scaled version of the new time slice allocation that they receive on time
1341 #define WEIGHT_IDLEPRIO 2
1342 #define WMULT_IDLEPRIO (1 << 31)
1345 * Nice levels are multiplicative, with a gentle 10% change for every
1346 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1347 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1348 * that remained on nice 0.
1350 * The "10% effect" is relative and cumulative: from _any_ nice level,
1351 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1352 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1353 * If a task goes up by ~10% and another task goes down by ~10% then
1354 * the relative distance between them is ~25%.)
1356 static const int prio_to_weight[40] = {
1357 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1358 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1359 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1360 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1361 /* 0 */ 1024, 820, 655, 526, 423,
1362 /* 5 */ 335, 272, 215, 172, 137,
1363 /* 10 */ 110, 87, 70, 56, 45,
1364 /* 15 */ 36, 29, 23, 18, 15,
1368 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1370 * In cases where the weight does not change often, we can use the
1371 * precalculated inverse to speed up arithmetics by turning divisions
1372 * into multiplications:
1374 static const u32 prio_to_wmult[40] = {
1375 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1376 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1377 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1378 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1379 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1380 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1381 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1382 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1385 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1388 * runqueue iterator, to support SMP load-balancing between different
1389 * scheduling classes, without having to expose their internal data
1390 * structures to the load-balancing proper:
1392 struct rq_iterator {
1394 struct task_struct *(*start)(void *);
1395 struct task_struct *(*next)(void *);
1399 static unsigned long
1400 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1401 unsigned long max_load_move, struct sched_domain *sd,
1402 enum cpu_idle_type idle, int *all_pinned,
1403 int *this_best_prio, struct rq_iterator *iterator);
1406 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1407 struct sched_domain *sd, enum cpu_idle_type idle,
1408 struct rq_iterator *iterator);
1411 #ifdef CONFIG_CGROUP_CPUACCT
1412 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1414 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1417 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1419 update_load_add(&rq->load, load);
1422 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1424 update_load_sub(&rq->load, load);
1428 static unsigned long source_load(int cpu, int type);
1429 static unsigned long target_load(int cpu, int type);
1430 static unsigned long cpu_avg_load_per_task(int cpu);
1431 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1433 #ifdef CONFIG_FAIR_GROUP_SCHED
1435 typedef void (*tg_visitor)(struct task_group *, int, struct sched_domain *);
1438 * Iterate the full tree, calling @down when first entering a node and @up when
1439 * leaving it for the final time.
1442 walk_tg_tree(tg_visitor down, tg_visitor up, int cpu, struct sched_domain *sd)
1444 struct task_group *parent, *child;
1447 parent = &root_task_group;
1449 (*down)(parent, cpu, sd);
1450 list_for_each_entry_rcu(child, &parent->children, siblings) {
1457 (*up)(parent, cpu, sd);
1460 parent = parent->parent;
1466 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1469 * Calculate and set the cpu's group shares.
1472 __update_group_shares_cpu(struct task_group *tg, int cpu,
1473 unsigned long sd_shares, unsigned long sd_rq_weight)
1476 unsigned long shares;
1477 unsigned long rq_weight;
1482 rq_weight = tg->cfs_rq[cpu]->load.weight;
1485 * If there are currently no tasks on the cpu pretend there is one of
1486 * average load so that when a new task gets to run here it will not
1487 * get delayed by group starvation.
1491 rq_weight = NICE_0_LOAD;
1494 if (unlikely(rq_weight > sd_rq_weight))
1495 rq_weight = sd_rq_weight;
1498 * \Sum shares * rq_weight
1499 * shares = -----------------------
1503 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1506 * record the actual number of shares, not the boosted amount.
1508 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1510 if (shares < MIN_SHARES)
1511 shares = MIN_SHARES;
1512 else if (shares > MAX_SHARES)
1513 shares = MAX_SHARES;
1515 __set_se_shares(tg->se[cpu], shares);
1519 * Re-compute the task group their per cpu shares over the given domain.
1520 * This needs to be done in a bottom-up fashion because the rq weight of a
1521 * parent group depends on the shares of its child groups.
1524 tg_shares_up(struct task_group *tg, int cpu, struct sched_domain *sd)
1526 unsigned long rq_weight = 0;
1527 unsigned long shares = 0;
1530 for_each_cpu_mask(i, sd->span) {
1531 rq_weight += tg->cfs_rq[i]->load.weight;
1532 shares += tg->cfs_rq[i]->shares;
1535 if ((!shares && rq_weight) || shares > tg->shares)
1536 shares = tg->shares;
1538 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1539 shares = tg->shares;
1541 for_each_cpu_mask(i, sd->span) {
1542 struct rq *rq = cpu_rq(i);
1543 unsigned long flags;
1545 spin_lock_irqsave(&rq->lock, flags);
1546 __update_group_shares_cpu(tg, i, shares, rq_weight);
1547 spin_unlock_irqrestore(&rq->lock, flags);
1552 * Compute the cpu's hierarchical load factor for each task group.
1553 * This needs to be done in a top-down fashion because the load of a child
1554 * group is a fraction of its parents load.
1557 tg_load_down(struct task_group *tg, int cpu, struct sched_domain *sd)
1562 load = cpu_rq(cpu)->load.weight;
1564 load = tg->parent->cfs_rq[cpu]->h_load;
1565 load *= tg->cfs_rq[cpu]->shares;
1566 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1569 tg->cfs_rq[cpu]->h_load = load;
1573 tg_nop(struct task_group *tg, int cpu, struct sched_domain *sd)
1577 static void update_shares(struct sched_domain *sd)
1579 walk_tg_tree(tg_nop, tg_shares_up, 0, sd);
1582 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1584 spin_unlock(&rq->lock);
1586 spin_lock(&rq->lock);
1589 static void update_h_load(int cpu)
1591 walk_tg_tree(tg_load_down, tg_nop, cpu, NULL);
1594 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1596 cfs_rq->shares = shares;
1601 static inline void update_shares(struct sched_domain *sd)
1605 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1613 #include "sched_stats.h"
1614 #include "sched_idletask.c"
1615 #include "sched_fair.c"
1616 #include "sched_rt.c"
1617 #ifdef CONFIG_SCHED_DEBUG
1618 # include "sched_debug.c"
1621 #define sched_class_highest (&rt_sched_class)
1622 #define for_each_class(class) \
1623 for (class = sched_class_highest; class; class = class->next)
1625 static void inc_nr_running(struct rq *rq)
1630 static void dec_nr_running(struct rq *rq)
1635 static void set_load_weight(struct task_struct *p)
1637 if (task_has_rt_policy(p)) {
1638 p->se.load.weight = prio_to_weight[0] * 2;
1639 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1644 * SCHED_IDLE tasks get minimal weight:
1646 if (p->policy == SCHED_IDLE) {
1647 p->se.load.weight = WEIGHT_IDLEPRIO;
1648 p->se.load.inv_weight = WMULT_IDLEPRIO;
1652 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1653 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1656 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1658 sched_info_queued(p);
1659 p->sched_class->enqueue_task(rq, p, wakeup);
1663 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1665 p->sched_class->dequeue_task(rq, p, sleep);
1670 * __normal_prio - return the priority that is based on the static prio
1672 static inline int __normal_prio(struct task_struct *p)
1674 return p->static_prio;
1678 * Calculate the expected normal priority: i.e. priority
1679 * without taking RT-inheritance into account. Might be
1680 * boosted by interactivity modifiers. Changes upon fork,
1681 * setprio syscalls, and whenever the interactivity
1682 * estimator recalculates.
1684 static inline int normal_prio(struct task_struct *p)
1688 if (task_has_rt_policy(p))
1689 prio = MAX_RT_PRIO-1 - p->rt_priority;
1691 prio = __normal_prio(p);
1696 * Calculate the current priority, i.e. the priority
1697 * taken into account by the scheduler. This value might
1698 * be boosted by RT tasks, or might be boosted by
1699 * interactivity modifiers. Will be RT if the task got
1700 * RT-boosted. If not then it returns p->normal_prio.
1702 static int effective_prio(struct task_struct *p)
1704 p->normal_prio = normal_prio(p);
1706 * If we are RT tasks or we were boosted to RT priority,
1707 * keep the priority unchanged. Otherwise, update priority
1708 * to the normal priority:
1710 if (!rt_prio(p->prio))
1711 return p->normal_prio;
1716 * activate_task - move a task to the runqueue.
1718 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1720 if (task_contributes_to_load(p))
1721 rq->nr_uninterruptible--;
1723 enqueue_task(rq, p, wakeup);
1728 * deactivate_task - remove a task from the runqueue.
1730 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1732 if (task_contributes_to_load(p))
1733 rq->nr_uninterruptible++;
1735 dequeue_task(rq, p, sleep);
1740 * task_curr - is this task currently executing on a CPU?
1741 * @p: the task in question.
1743 inline int task_curr(const struct task_struct *p)
1745 return cpu_curr(task_cpu(p)) == p;
1748 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1750 set_task_rq(p, cpu);
1753 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1754 * successfuly executed on another CPU. We must ensure that updates of
1755 * per-task data have been completed by this moment.
1758 task_thread_info(p)->cpu = cpu;
1762 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1763 const struct sched_class *prev_class,
1764 int oldprio, int running)
1766 if (prev_class != p->sched_class) {
1767 if (prev_class->switched_from)
1768 prev_class->switched_from(rq, p, running);
1769 p->sched_class->switched_to(rq, p, running);
1771 p->sched_class->prio_changed(rq, p, oldprio, running);
1776 /* Used instead of source_load when we know the type == 0 */
1777 static unsigned long weighted_cpuload(const int cpu)
1779 return cpu_rq(cpu)->load.weight;
1783 * Is this task likely cache-hot:
1786 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1791 * Buddy candidates are cache hot:
1793 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1796 if (p->sched_class != &fair_sched_class)
1799 if (sysctl_sched_migration_cost == -1)
1801 if (sysctl_sched_migration_cost == 0)
1804 delta = now - p->se.exec_start;
1806 return delta < (s64)sysctl_sched_migration_cost;
1810 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1812 int old_cpu = task_cpu(p);
1813 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1814 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1815 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1818 clock_offset = old_rq->clock - new_rq->clock;
1820 #ifdef CONFIG_SCHEDSTATS
1821 if (p->se.wait_start)
1822 p->se.wait_start -= clock_offset;
1823 if (p->se.sleep_start)
1824 p->se.sleep_start -= clock_offset;
1825 if (p->se.block_start)
1826 p->se.block_start -= clock_offset;
1827 if (old_cpu != new_cpu) {
1828 schedstat_inc(p, se.nr_migrations);
1829 if (task_hot(p, old_rq->clock, NULL))
1830 schedstat_inc(p, se.nr_forced2_migrations);
1833 p->se.vruntime -= old_cfsrq->min_vruntime -
1834 new_cfsrq->min_vruntime;
1836 __set_task_cpu(p, new_cpu);
1839 struct migration_req {
1840 struct list_head list;
1842 struct task_struct *task;
1845 struct completion done;
1849 * The task's runqueue lock must be held.
1850 * Returns true if you have to wait for migration thread.
1853 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1855 struct rq *rq = task_rq(p);
1858 * If the task is not on a runqueue (and not running), then
1859 * it is sufficient to simply update the task's cpu field.
1861 if (!p->se.on_rq && !task_running(rq, p)) {
1862 set_task_cpu(p, dest_cpu);
1866 init_completion(&req->done);
1868 req->dest_cpu = dest_cpu;
1869 list_add(&req->list, &rq->migration_queue);
1875 * wait_task_inactive - wait for a thread to unschedule.
1877 * The caller must ensure that the task *will* unschedule sometime soon,
1878 * else this function might spin for a *long* time. This function can't
1879 * be called with interrupts off, or it may introduce deadlock with
1880 * smp_call_function() if an IPI is sent by the same process we are
1881 * waiting to become inactive.
1883 void wait_task_inactive(struct task_struct *p)
1885 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))
1913 * Ok, time to look more closely! We need the rq
1914 * lock now, to be *sure*. If we're wrong, we'll
1915 * just go back and repeat.
1917 rq = task_rq_lock(p, &flags);
1918 running = task_running(rq, p);
1919 on_rq = p->se.on_rq;
1920 task_rq_unlock(rq, &flags);
1923 * Was it really running after all now that we
1924 * checked with the proper locks actually held?
1926 * Oops. Go back and try again..
1928 if (unlikely(running)) {
1934 * It's not enough that it's not actively running,
1935 * it must be off the runqueue _entirely_, and not
1938 * So if it wa still runnable (but just not actively
1939 * running right now), it's preempted, and we should
1940 * yield - it could be a while.
1942 if (unlikely(on_rq)) {
1943 schedule_timeout_uninterruptible(1);
1948 * Ahh, all good. It wasn't running, and it wasn't
1949 * runnable, which means that it will never become
1950 * running in the future either. We're all done!
1957 * kick_process - kick a running thread to enter/exit the kernel
1958 * @p: the to-be-kicked thread
1960 * Cause a process which is running on another CPU to enter
1961 * kernel-mode, without any delay. (to get signals handled.)
1963 * NOTE: this function doesnt have to take the runqueue lock,
1964 * because all it wants to ensure is that the remote task enters
1965 * the kernel. If the IPI races and the task has been migrated
1966 * to another CPU then no harm is done and the purpose has been
1969 void kick_process(struct task_struct *p)
1975 if ((cpu != smp_processor_id()) && task_curr(p))
1976 smp_send_reschedule(cpu);
1981 * Return a low guess at the load of a migration-source cpu weighted
1982 * according to the scheduling class and "nice" value.
1984 * We want to under-estimate the load of migration sources, to
1985 * balance conservatively.
1987 static unsigned long source_load(int cpu, int type)
1989 struct rq *rq = cpu_rq(cpu);
1990 unsigned long total = weighted_cpuload(cpu);
1995 return min(rq->cpu_load[type-1], total);
1999 * Return a high guess at the load of a migration-target cpu weighted
2000 * according to the scheduling class and "nice" value.
2002 static unsigned long target_load(int cpu, int type)
2004 struct rq *rq = cpu_rq(cpu);
2005 unsigned long total = weighted_cpuload(cpu);
2010 return max(rq->cpu_load[type-1], total);
2014 * Return the average load per task on the cpu's run queue
2016 static unsigned long cpu_avg_load_per_task(int cpu)
2018 struct rq *rq = cpu_rq(cpu);
2019 unsigned long total = weighted_cpuload(cpu);
2020 unsigned long n = rq->nr_running;
2022 return n ? total / n : SCHED_LOAD_SCALE;
2026 * find_idlest_group finds and returns the least busy CPU group within the
2029 static struct sched_group *
2030 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2032 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2033 unsigned long min_load = ULONG_MAX, this_load = 0;
2034 int load_idx = sd->forkexec_idx;
2035 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2038 unsigned long load, avg_load;
2042 /* Skip over this group if it has no CPUs allowed */
2043 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2046 local_group = cpu_isset(this_cpu, group->cpumask);
2048 /* Tally up the load of all CPUs in the group */
2051 for_each_cpu_mask(i, group->cpumask) {
2052 /* Bias balancing toward cpus of our domain */
2054 load = source_load(i, load_idx);
2056 load = target_load(i, load_idx);
2061 /* Adjust by relative CPU power of the group */
2062 avg_load = sg_div_cpu_power(group,
2063 avg_load * SCHED_LOAD_SCALE);
2066 this_load = avg_load;
2068 } else if (avg_load < min_load) {
2069 min_load = avg_load;
2072 } while (group = group->next, group != sd->groups);
2074 if (!idlest || 100*this_load < imbalance*min_load)
2080 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2083 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2086 unsigned long load, min_load = ULONG_MAX;
2090 /* Traverse only the allowed CPUs */
2091 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2093 for_each_cpu_mask(i, *tmp) {
2094 load = weighted_cpuload(i);
2096 if (load < min_load || (load == min_load && i == this_cpu)) {
2106 * sched_balance_self: balance the current task (running on cpu) in domains
2107 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2110 * Balance, ie. select the least loaded group.
2112 * Returns the target CPU number, or the same CPU if no balancing is needed.
2114 * preempt must be disabled.
2116 static int sched_balance_self(int cpu, int flag)
2118 struct task_struct *t = current;
2119 struct sched_domain *tmp, *sd = NULL;
2121 for_each_domain(cpu, tmp) {
2123 * If power savings logic is enabled for a domain, stop there.
2125 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2127 if (tmp->flags & flag)
2135 cpumask_t span, tmpmask;
2136 struct sched_group *group;
2137 int new_cpu, weight;
2139 if (!(sd->flags & flag)) {
2145 group = find_idlest_group(sd, t, cpu);
2151 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2152 if (new_cpu == -1 || new_cpu == cpu) {
2153 /* Now try balancing at a lower domain level of cpu */
2158 /* Now try balancing at a lower domain level of new_cpu */
2161 weight = cpus_weight(span);
2162 for_each_domain(cpu, tmp) {
2163 if (weight <= cpus_weight(tmp->span))
2165 if (tmp->flags & flag)
2168 /* while loop will break here if sd == NULL */
2174 #endif /* CONFIG_SMP */
2177 * try_to_wake_up - wake up a thread
2178 * @p: the to-be-woken-up thread
2179 * @state: the mask of task states that can be woken
2180 * @sync: do a synchronous wakeup?
2182 * Put it on the run-queue if it's not already there. The "current"
2183 * thread is always on the run-queue (except when the actual
2184 * re-schedule is in progress), and as such you're allowed to do
2185 * the simpler "current->state = TASK_RUNNING" to mark yourself
2186 * runnable without the overhead of this.
2188 * returns failure only if the task is already active.
2190 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2192 int cpu, orig_cpu, this_cpu, success = 0;
2193 unsigned long flags;
2197 if (!sched_feat(SYNC_WAKEUPS))
2201 rq = task_rq_lock(p, &flags);
2202 old_state = p->state;
2203 if (!(old_state & state))
2211 this_cpu = smp_processor_id();
2214 if (unlikely(task_running(rq, p)))
2217 cpu = p->sched_class->select_task_rq(p, sync);
2218 if (cpu != orig_cpu) {
2219 set_task_cpu(p, cpu);
2220 task_rq_unlock(rq, &flags);
2221 /* might preempt at this point */
2222 rq = task_rq_lock(p, &flags);
2223 old_state = p->state;
2224 if (!(old_state & state))
2229 this_cpu = smp_processor_id();
2233 #ifdef CONFIG_SCHEDSTATS
2234 schedstat_inc(rq, ttwu_count);
2235 if (cpu == this_cpu)
2236 schedstat_inc(rq, ttwu_local);
2238 struct sched_domain *sd;
2239 for_each_domain(this_cpu, sd) {
2240 if (cpu_isset(cpu, sd->span)) {
2241 schedstat_inc(sd, ttwu_wake_remote);
2246 #endif /* CONFIG_SCHEDSTATS */
2249 #endif /* CONFIG_SMP */
2250 schedstat_inc(p, se.nr_wakeups);
2252 schedstat_inc(p, se.nr_wakeups_sync);
2253 if (orig_cpu != cpu)
2254 schedstat_inc(p, se.nr_wakeups_migrate);
2255 if (cpu == this_cpu)
2256 schedstat_inc(p, se.nr_wakeups_local);
2258 schedstat_inc(p, se.nr_wakeups_remote);
2259 update_rq_clock(rq);
2260 activate_task(rq, p, 1);
2264 check_preempt_curr(rq, p);
2266 p->state = TASK_RUNNING;
2268 if (p->sched_class->task_wake_up)
2269 p->sched_class->task_wake_up(rq, p);
2272 task_rq_unlock(rq, &flags);
2277 int wake_up_process(struct task_struct *p)
2279 return try_to_wake_up(p, TASK_ALL, 0);
2281 EXPORT_SYMBOL(wake_up_process);
2283 int wake_up_state(struct task_struct *p, unsigned int state)
2285 return try_to_wake_up(p, state, 0);
2289 * Perform scheduler related setup for a newly forked process p.
2290 * p is forked by current.
2292 * __sched_fork() is basic setup used by init_idle() too:
2294 static void __sched_fork(struct task_struct *p)
2296 p->se.exec_start = 0;
2297 p->se.sum_exec_runtime = 0;
2298 p->se.prev_sum_exec_runtime = 0;
2299 p->se.last_wakeup = 0;
2300 p->se.avg_overlap = 0;
2302 #ifdef CONFIG_SCHEDSTATS
2303 p->se.wait_start = 0;
2304 p->se.sum_sleep_runtime = 0;
2305 p->se.sleep_start = 0;
2306 p->se.block_start = 0;
2307 p->se.sleep_max = 0;
2308 p->se.block_max = 0;
2310 p->se.slice_max = 0;
2314 INIT_LIST_HEAD(&p->rt.run_list);
2316 INIT_LIST_HEAD(&p->se.group_node);
2318 #ifdef CONFIG_PREEMPT_NOTIFIERS
2319 INIT_HLIST_HEAD(&p->preempt_notifiers);
2323 * We mark the process as running here, but have not actually
2324 * inserted it onto the runqueue yet. This guarantees that
2325 * nobody will actually run it, and a signal or other external
2326 * event cannot wake it up and insert it on the runqueue either.
2328 p->state = TASK_RUNNING;
2332 * fork()/clone()-time setup:
2334 void sched_fork(struct task_struct *p, int clone_flags)
2336 int cpu = get_cpu();
2341 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2343 set_task_cpu(p, cpu);
2346 * Make sure we do not leak PI boosting priority to the child:
2348 p->prio = current->normal_prio;
2349 if (!rt_prio(p->prio))
2350 p->sched_class = &fair_sched_class;
2352 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2353 if (likely(sched_info_on()))
2354 memset(&p->sched_info, 0, sizeof(p->sched_info));
2356 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2359 #ifdef CONFIG_PREEMPT
2360 /* Want to start with kernel preemption disabled. */
2361 task_thread_info(p)->preempt_count = 1;
2367 * wake_up_new_task - wake up a newly created task for the first time.
2369 * This function will do some initial scheduler statistics housekeeping
2370 * that must be done for every newly created context, then puts the task
2371 * on the runqueue and wakes it.
2373 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2375 unsigned long flags;
2378 rq = task_rq_lock(p, &flags);
2379 BUG_ON(p->state != TASK_RUNNING);
2380 update_rq_clock(rq);
2382 p->prio = effective_prio(p);
2384 if (!p->sched_class->task_new || !current->se.on_rq) {
2385 activate_task(rq, p, 0);
2388 * Let the scheduling class do new task startup
2389 * management (if any):
2391 p->sched_class->task_new(rq, p);
2394 check_preempt_curr(rq, p);
2396 if (p->sched_class->task_wake_up)
2397 p->sched_class->task_wake_up(rq, p);
2399 task_rq_unlock(rq, &flags);
2402 #ifdef CONFIG_PREEMPT_NOTIFIERS
2405 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2406 * @notifier: notifier struct to register
2408 void preempt_notifier_register(struct preempt_notifier *notifier)
2410 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2412 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2415 * preempt_notifier_unregister - no longer interested in preemption notifications
2416 * @notifier: notifier struct to unregister
2418 * This is safe to call from within a preemption notifier.
2420 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2422 hlist_del(¬ifier->link);
2424 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2426 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2428 struct preempt_notifier *notifier;
2429 struct hlist_node *node;
2431 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2432 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2436 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2437 struct task_struct *next)
2439 struct preempt_notifier *notifier;
2440 struct hlist_node *node;
2442 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2443 notifier->ops->sched_out(notifier, next);
2446 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2448 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2453 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2454 struct task_struct *next)
2458 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2461 * prepare_task_switch - prepare to switch tasks
2462 * @rq: the runqueue preparing to switch
2463 * @prev: the current task that is being switched out
2464 * @next: the task we are going to switch to.
2466 * This is called with the rq lock held and interrupts off. It must
2467 * be paired with a subsequent finish_task_switch after the context
2470 * prepare_task_switch sets up locking and calls architecture specific
2474 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2475 struct task_struct *next)
2477 fire_sched_out_preempt_notifiers(prev, next);
2478 prepare_lock_switch(rq, next);
2479 prepare_arch_switch(next);
2483 * finish_task_switch - clean up after a task-switch
2484 * @rq: runqueue associated with task-switch
2485 * @prev: the thread we just switched away from.
2487 * finish_task_switch must be called after the context switch, paired
2488 * with a prepare_task_switch call before the context switch.
2489 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2490 * and do any other architecture-specific cleanup actions.
2492 * Note that we may have delayed dropping an mm in context_switch(). If
2493 * so, we finish that here outside of the runqueue lock. (Doing it
2494 * with the lock held can cause deadlocks; see schedule() for
2497 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2498 __releases(rq->lock)
2500 struct mm_struct *mm = rq->prev_mm;
2506 * A task struct has one reference for the use as "current".
2507 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2508 * schedule one last time. The schedule call will never return, and
2509 * the scheduled task must drop that reference.
2510 * The test for TASK_DEAD must occur while the runqueue locks are
2511 * still held, otherwise prev could be scheduled on another cpu, die
2512 * there before we look at prev->state, and then the reference would
2514 * Manfred Spraul <manfred@colorfullife.com>
2516 prev_state = prev->state;
2517 finish_arch_switch(prev);
2518 finish_lock_switch(rq, prev);
2520 if (current->sched_class->post_schedule)
2521 current->sched_class->post_schedule(rq);
2524 fire_sched_in_preempt_notifiers(current);
2527 if (unlikely(prev_state == TASK_DEAD)) {
2529 * Remove function-return probe instances associated with this
2530 * task and put them back on the free list.
2532 kprobe_flush_task(prev);
2533 put_task_struct(prev);
2538 * schedule_tail - first thing a freshly forked thread must call.
2539 * @prev: the thread we just switched away from.
2541 asmlinkage void schedule_tail(struct task_struct *prev)
2542 __releases(rq->lock)
2544 struct rq *rq = this_rq();
2546 finish_task_switch(rq, prev);
2547 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2548 /* In this case, finish_task_switch does not reenable preemption */
2551 if (current->set_child_tid)
2552 put_user(task_pid_vnr(current), current->set_child_tid);
2556 * context_switch - switch to the new MM and the new
2557 * thread's register state.
2560 context_switch(struct rq *rq, struct task_struct *prev,
2561 struct task_struct *next)
2563 struct mm_struct *mm, *oldmm;
2565 prepare_task_switch(rq, prev, next);
2567 oldmm = prev->active_mm;
2569 * For paravirt, this is coupled with an exit in switch_to to
2570 * combine the page table reload and the switch backend into
2573 arch_enter_lazy_cpu_mode();
2575 if (unlikely(!mm)) {
2576 next->active_mm = oldmm;
2577 atomic_inc(&oldmm->mm_count);
2578 enter_lazy_tlb(oldmm, next);
2580 switch_mm(oldmm, mm, next);
2582 if (unlikely(!prev->mm)) {
2583 prev->active_mm = NULL;
2584 rq->prev_mm = oldmm;
2587 * Since the runqueue lock will be released by the next
2588 * task (which is an invalid locking op but in the case
2589 * of the scheduler it's an obvious special-case), so we
2590 * do an early lockdep release here:
2592 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2593 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2596 /* Here we just switch the register state and the stack. */
2597 switch_to(prev, next, prev);
2601 * this_rq must be evaluated again because prev may have moved
2602 * CPUs since it called schedule(), thus the 'rq' on its stack
2603 * frame will be invalid.
2605 finish_task_switch(this_rq(), prev);
2609 * nr_running, nr_uninterruptible and nr_context_switches:
2611 * externally visible scheduler statistics: current number of runnable
2612 * threads, current number of uninterruptible-sleeping threads, total
2613 * number of context switches performed since bootup.
2615 unsigned long nr_running(void)
2617 unsigned long i, sum = 0;
2619 for_each_online_cpu(i)
2620 sum += cpu_rq(i)->nr_running;
2625 unsigned long nr_uninterruptible(void)
2627 unsigned long i, sum = 0;
2629 for_each_possible_cpu(i)
2630 sum += cpu_rq(i)->nr_uninterruptible;
2633 * Since we read the counters lockless, it might be slightly
2634 * inaccurate. Do not allow it to go below zero though:
2636 if (unlikely((long)sum < 0))
2642 unsigned long long nr_context_switches(void)
2645 unsigned long long sum = 0;
2647 for_each_possible_cpu(i)
2648 sum += cpu_rq(i)->nr_switches;
2653 unsigned long nr_iowait(void)
2655 unsigned long i, sum = 0;
2657 for_each_possible_cpu(i)
2658 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2663 unsigned long nr_active(void)
2665 unsigned long i, running = 0, uninterruptible = 0;
2667 for_each_online_cpu(i) {
2668 running += cpu_rq(i)->nr_running;
2669 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2672 if (unlikely((long)uninterruptible < 0))
2673 uninterruptible = 0;
2675 return running + uninterruptible;
2679 * Update rq->cpu_load[] statistics. This function is usually called every
2680 * scheduler tick (TICK_NSEC).
2682 static void update_cpu_load(struct rq *this_rq)
2684 unsigned long this_load = this_rq->load.weight;
2687 this_rq->nr_load_updates++;
2689 /* Update our load: */
2690 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2691 unsigned long old_load, new_load;
2693 /* scale is effectively 1 << i now, and >> i divides by scale */
2695 old_load = this_rq->cpu_load[i];
2696 new_load = this_load;
2698 * Round up the averaging division if load is increasing. This
2699 * prevents us from getting stuck on 9 if the load is 10, for
2702 if (new_load > old_load)
2703 new_load += scale-1;
2704 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2711 * double_rq_lock - safely lock two runqueues
2713 * Note this does not disable interrupts like task_rq_lock,
2714 * you need to do so manually before calling.
2716 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2717 __acquires(rq1->lock)
2718 __acquires(rq2->lock)
2720 BUG_ON(!irqs_disabled());
2722 spin_lock(&rq1->lock);
2723 __acquire(rq2->lock); /* Fake it out ;) */
2726 spin_lock(&rq1->lock);
2727 spin_lock(&rq2->lock);
2729 spin_lock(&rq2->lock);
2730 spin_lock(&rq1->lock);
2733 update_rq_clock(rq1);
2734 update_rq_clock(rq2);
2738 * double_rq_unlock - safely unlock two runqueues
2740 * Note this does not restore interrupts like task_rq_unlock,
2741 * you need to do so manually after calling.
2743 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2744 __releases(rq1->lock)
2745 __releases(rq2->lock)
2747 spin_unlock(&rq1->lock);
2749 spin_unlock(&rq2->lock);
2751 __release(rq2->lock);
2755 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2757 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2758 __releases(this_rq->lock)
2759 __acquires(busiest->lock)
2760 __acquires(this_rq->lock)
2764 if (unlikely(!irqs_disabled())) {
2765 /* printk() doesn't work good under rq->lock */
2766 spin_unlock(&this_rq->lock);
2769 if (unlikely(!spin_trylock(&busiest->lock))) {
2770 if (busiest < this_rq) {
2771 spin_unlock(&this_rq->lock);
2772 spin_lock(&busiest->lock);
2773 spin_lock(&this_rq->lock);
2776 spin_lock(&busiest->lock);
2782 * If dest_cpu is allowed for this process, migrate the task to it.
2783 * This is accomplished by forcing the cpu_allowed mask to only
2784 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2785 * the cpu_allowed mask is restored.
2787 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2789 struct migration_req req;
2790 unsigned long flags;
2793 rq = task_rq_lock(p, &flags);
2794 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2795 || unlikely(cpu_is_offline(dest_cpu)))
2798 /* force the process onto the specified CPU */
2799 if (migrate_task(p, dest_cpu, &req)) {
2800 /* Need to wait for migration thread (might exit: take ref). */
2801 struct task_struct *mt = rq->migration_thread;
2803 get_task_struct(mt);
2804 task_rq_unlock(rq, &flags);
2805 wake_up_process(mt);
2806 put_task_struct(mt);
2807 wait_for_completion(&req.done);
2812 task_rq_unlock(rq, &flags);
2816 * sched_exec - execve() is a valuable balancing opportunity, because at
2817 * this point the task has the smallest effective memory and cache footprint.
2819 void sched_exec(void)
2821 int new_cpu, this_cpu = get_cpu();
2822 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2824 if (new_cpu != this_cpu)
2825 sched_migrate_task(current, new_cpu);
2829 * pull_task - move a task from a remote runqueue to the local runqueue.
2830 * Both runqueues must be locked.
2832 static void pull_task(struct rq *src_rq, struct task_struct *p,
2833 struct rq *this_rq, int this_cpu)
2835 deactivate_task(src_rq, p, 0);
2836 set_task_cpu(p, this_cpu);
2837 activate_task(this_rq, p, 0);
2839 * Note that idle threads have a prio of MAX_PRIO, for this test
2840 * to be always true for them.
2842 check_preempt_curr(this_rq, p);
2846 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2849 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2850 struct sched_domain *sd, enum cpu_idle_type idle,
2854 * We do not migrate tasks that are:
2855 * 1) running (obviously), or
2856 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2857 * 3) are cache-hot on their current CPU.
2859 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2860 schedstat_inc(p, se.nr_failed_migrations_affine);
2865 if (task_running(rq, p)) {
2866 schedstat_inc(p, se.nr_failed_migrations_running);
2871 * Aggressive migration if:
2872 * 1) task is cache cold, or
2873 * 2) too many balance attempts have failed.
2876 if (!task_hot(p, rq->clock, sd) ||
2877 sd->nr_balance_failed > sd->cache_nice_tries) {
2878 #ifdef CONFIG_SCHEDSTATS
2879 if (task_hot(p, rq->clock, sd)) {
2880 schedstat_inc(sd, lb_hot_gained[idle]);
2881 schedstat_inc(p, se.nr_forced_migrations);
2887 if (task_hot(p, rq->clock, sd)) {
2888 schedstat_inc(p, se.nr_failed_migrations_hot);
2894 static unsigned long
2895 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2896 unsigned long max_load_move, struct sched_domain *sd,
2897 enum cpu_idle_type idle, int *all_pinned,
2898 int *this_best_prio, struct rq_iterator *iterator)
2900 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
2901 struct task_struct *p;
2902 long rem_load_move = max_load_move;
2904 if (max_load_move == 0)
2910 * Start the load-balancing iterator:
2912 p = iterator->start(iterator->arg);
2914 if (!p || loops++ > sysctl_sched_nr_migrate)
2917 * To help distribute high priority tasks across CPUs we don't
2918 * skip a task if it will be the highest priority task (i.e. smallest
2919 * prio value) on its new queue regardless of its load weight
2921 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
2922 SCHED_LOAD_SCALE_FUZZ;
2923 if ((skip_for_load && p->prio >= *this_best_prio) ||
2924 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2925 p = iterator->next(iterator->arg);
2929 pull_task(busiest, p, this_rq, this_cpu);
2931 rem_load_move -= p->se.load.weight;
2934 * We only want to steal up to the prescribed amount of weighted load.
2936 if (rem_load_move > 0) {
2937 if (p->prio < *this_best_prio)
2938 *this_best_prio = p->prio;
2939 p = iterator->next(iterator->arg);
2944 * Right now, this is one of only two places pull_task() is called,
2945 * so we can safely collect pull_task() stats here rather than
2946 * inside pull_task().
2948 schedstat_add(sd, lb_gained[idle], pulled);
2951 *all_pinned = pinned;
2953 return max_load_move - rem_load_move;
2957 * move_tasks tries to move up to max_load_move weighted load from busiest to
2958 * this_rq, as part of a balancing operation within domain "sd".
2959 * Returns 1 if successful and 0 otherwise.
2961 * Called with both runqueues locked.
2963 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2964 unsigned long max_load_move,
2965 struct sched_domain *sd, enum cpu_idle_type idle,
2968 const struct sched_class *class = sched_class_highest;
2969 unsigned long total_load_moved = 0;
2970 int this_best_prio = this_rq->curr->prio;
2974 class->load_balance(this_rq, this_cpu, busiest,
2975 max_load_move - total_load_moved,
2976 sd, idle, all_pinned, &this_best_prio);
2977 class = class->next;
2978 } while (class && max_load_move > total_load_moved);
2980 return total_load_moved > 0;
2984 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2985 struct sched_domain *sd, enum cpu_idle_type idle,
2986 struct rq_iterator *iterator)
2988 struct task_struct *p = iterator->start(iterator->arg);
2992 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2993 pull_task(busiest, p, this_rq, this_cpu);
2995 * Right now, this is only the second place pull_task()
2996 * is called, so we can safely collect pull_task()
2997 * stats here rather than inside pull_task().
2999 schedstat_inc(sd, lb_gained[idle]);
3003 p = iterator->next(iterator->arg);
3010 * move_one_task tries to move exactly one task from busiest to this_rq, as
3011 * part of active balancing operations within "domain".
3012 * Returns 1 if successful and 0 otherwise.
3014 * Called with both runqueues locked.
3016 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3017 struct sched_domain *sd, enum cpu_idle_type idle)
3019 const struct sched_class *class;
3021 for (class = sched_class_highest; class; class = class->next)
3022 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3029 * find_busiest_group finds and returns the busiest CPU group within the
3030 * domain. It calculates and returns the amount of weighted load which
3031 * should be moved to restore balance via the imbalance parameter.
3033 static struct sched_group *
3034 find_busiest_group(struct sched_domain *sd, int this_cpu,
3035 unsigned long *imbalance, enum cpu_idle_type idle,
3036 int *sd_idle, const cpumask_t *cpus, int *balance)
3038 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3039 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3040 unsigned long max_pull;
3041 unsigned long busiest_load_per_task, busiest_nr_running;
3042 unsigned long this_load_per_task, this_nr_running;
3043 int load_idx, group_imb = 0;
3044 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3045 int power_savings_balance = 1;
3046 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3047 unsigned long min_nr_running = ULONG_MAX;
3048 struct sched_group *group_min = NULL, *group_leader = NULL;
3051 max_load = this_load = total_load = total_pwr = 0;
3052 busiest_load_per_task = busiest_nr_running = 0;
3053 this_load_per_task = this_nr_running = 0;
3054 if (idle == CPU_NOT_IDLE)
3055 load_idx = sd->busy_idx;
3056 else if (idle == CPU_NEWLY_IDLE)
3057 load_idx = sd->newidle_idx;
3059 load_idx = sd->idle_idx;
3062 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3065 int __group_imb = 0;
3066 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3067 unsigned long sum_nr_running, sum_weighted_load;
3069 local_group = cpu_isset(this_cpu, group->cpumask);
3072 balance_cpu = first_cpu(group->cpumask);
3074 /* Tally up the load of all CPUs in the group */
3075 sum_weighted_load = sum_nr_running = avg_load = 0;
3077 min_cpu_load = ~0UL;
3079 for_each_cpu_mask(i, group->cpumask) {
3082 if (!cpu_isset(i, *cpus))
3087 if (*sd_idle && rq->nr_running)
3090 /* Bias balancing toward cpus of our domain */
3092 if (idle_cpu(i) && !first_idle_cpu) {
3097 load = target_load(i, load_idx);
3099 load = source_load(i, load_idx);
3100 if (load > max_cpu_load)
3101 max_cpu_load = load;
3102 if (min_cpu_load > load)
3103 min_cpu_load = load;
3107 sum_nr_running += rq->nr_running;
3108 sum_weighted_load += weighted_cpuload(i);
3112 * First idle cpu or the first cpu(busiest) in this sched group
3113 * is eligible for doing load balancing at this and above
3114 * domains. In the newly idle case, we will allow all the cpu's
3115 * to do the newly idle load balance.
3117 if (idle != CPU_NEWLY_IDLE && local_group &&
3118 balance_cpu != this_cpu && balance) {
3123 total_load += avg_load;
3124 total_pwr += group->__cpu_power;
3126 /* Adjust by relative CPU power of the group */
3127 avg_load = sg_div_cpu_power(group,
3128 avg_load * SCHED_LOAD_SCALE);
3130 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3133 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3136 this_load = avg_load;
3138 this_nr_running = sum_nr_running;
3139 this_load_per_task = sum_weighted_load;
3140 } else if (avg_load > max_load &&
3141 (sum_nr_running > group_capacity || __group_imb)) {
3142 max_load = avg_load;
3144 busiest_nr_running = sum_nr_running;
3145 busiest_load_per_task = sum_weighted_load;
3146 group_imb = __group_imb;
3149 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3151 * Busy processors will not participate in power savings
3154 if (idle == CPU_NOT_IDLE ||
3155 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3159 * If the local group is idle or completely loaded
3160 * no need to do power savings balance at this domain
3162 if (local_group && (this_nr_running >= group_capacity ||
3164 power_savings_balance = 0;
3167 * If a group is already running at full capacity or idle,
3168 * don't include that group in power savings calculations
3170 if (!power_savings_balance || sum_nr_running >= group_capacity
3175 * Calculate the group which has the least non-idle load.
3176 * This is the group from where we need to pick up the load
3179 if ((sum_nr_running < min_nr_running) ||
3180 (sum_nr_running == min_nr_running &&
3181 first_cpu(group->cpumask) <
3182 first_cpu(group_min->cpumask))) {
3184 min_nr_running = sum_nr_running;
3185 min_load_per_task = sum_weighted_load /
3190 * Calculate the group which is almost near its
3191 * capacity but still has some space to pick up some load
3192 * from other group and save more power
3194 if (sum_nr_running <= group_capacity - 1) {
3195 if (sum_nr_running > leader_nr_running ||
3196 (sum_nr_running == leader_nr_running &&
3197 first_cpu(group->cpumask) >
3198 first_cpu(group_leader->cpumask))) {
3199 group_leader = group;
3200 leader_nr_running = sum_nr_running;
3205 group = group->next;
3206 } while (group != sd->groups);
3208 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3211 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3213 if (this_load >= avg_load ||
3214 100*max_load <= sd->imbalance_pct*this_load)
3217 busiest_load_per_task /= busiest_nr_running;
3219 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3222 * We're trying to get all the cpus to the average_load, so we don't
3223 * want to push ourselves above the average load, nor do we wish to
3224 * reduce the max loaded cpu below the average load, as either of these
3225 * actions would just result in more rebalancing later, and ping-pong
3226 * tasks around. Thus we look for the minimum possible imbalance.
3227 * Negative imbalances (*we* are more loaded than anyone else) will
3228 * be counted as no imbalance for these purposes -- we can't fix that
3229 * by pulling tasks to us. Be careful of negative numbers as they'll
3230 * appear as very large values with unsigned longs.
3232 if (max_load <= busiest_load_per_task)
3236 * In the presence of smp nice balancing, certain scenarios can have
3237 * max load less than avg load(as we skip the groups at or below
3238 * its cpu_power, while calculating max_load..)
3240 if (max_load < avg_load) {
3242 goto small_imbalance;
3245 /* Don't want to pull so many tasks that a group would go idle */
3246 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3248 /* How much load to actually move to equalise the imbalance */
3249 *imbalance = min(max_pull * busiest->__cpu_power,
3250 (avg_load - this_load) * this->__cpu_power)
3254 * if *imbalance is less than the average load per runnable task
3255 * there is no gaurantee that any tasks will be moved so we'll have
3256 * a think about bumping its value to force at least one task to be
3259 if (*imbalance < busiest_load_per_task) {
3260 unsigned long tmp, pwr_now, pwr_move;
3264 pwr_move = pwr_now = 0;
3266 if (this_nr_running) {
3267 this_load_per_task /= this_nr_running;
3268 if (busiest_load_per_task > this_load_per_task)
3271 this_load_per_task = SCHED_LOAD_SCALE;
3273 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3274 busiest_load_per_task * imbn) {
3275 *imbalance = busiest_load_per_task;
3280 * OK, we don't have enough imbalance to justify moving tasks,
3281 * however we may be able to increase total CPU power used by
3285 pwr_now += busiest->__cpu_power *
3286 min(busiest_load_per_task, max_load);
3287 pwr_now += this->__cpu_power *
3288 min(this_load_per_task, this_load);
3289 pwr_now /= SCHED_LOAD_SCALE;
3291 /* Amount of load we'd subtract */
3292 tmp = sg_div_cpu_power(busiest,
3293 busiest_load_per_task * SCHED_LOAD_SCALE);
3295 pwr_move += busiest->__cpu_power *
3296 min(busiest_load_per_task, max_load - tmp);
3298 /* Amount of load we'd add */
3299 if (max_load * busiest->__cpu_power <
3300 busiest_load_per_task * SCHED_LOAD_SCALE)
3301 tmp = sg_div_cpu_power(this,
3302 max_load * busiest->__cpu_power);
3304 tmp = sg_div_cpu_power(this,
3305 busiest_load_per_task * SCHED_LOAD_SCALE);
3306 pwr_move += this->__cpu_power *
3307 min(this_load_per_task, this_load + tmp);
3308 pwr_move /= SCHED_LOAD_SCALE;
3310 /* Move if we gain throughput */
3311 if (pwr_move > pwr_now)
3312 *imbalance = busiest_load_per_task;
3318 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3319 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3322 if (this == group_leader && group_leader != group_min) {
3323 *imbalance = min_load_per_task;
3333 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3336 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3337 unsigned long imbalance, const cpumask_t *cpus)
3339 struct rq *busiest = NULL, *rq;
3340 unsigned long max_load = 0;
3343 for_each_cpu_mask(i, group->cpumask) {
3346 if (!cpu_isset(i, *cpus))
3350 wl = weighted_cpuload(i);
3352 if (rq->nr_running == 1 && wl > imbalance)
3355 if (wl > max_load) {
3365 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3366 * so long as it is large enough.
3368 #define MAX_PINNED_INTERVAL 512
3371 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3372 * tasks if there is an imbalance.
3374 static int load_balance(int this_cpu, struct rq *this_rq,
3375 struct sched_domain *sd, enum cpu_idle_type idle,
3376 int *balance, cpumask_t *cpus)
3378 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3379 struct sched_group *group;
3380 unsigned long imbalance;
3382 unsigned long flags;
3387 * When power savings policy is enabled for the parent domain, idle
3388 * sibling can pick up load irrespective of busy siblings. In this case,
3389 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3390 * portraying it as CPU_NOT_IDLE.
3392 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3393 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3396 schedstat_inc(sd, lb_count[idle]);
3400 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3407 schedstat_inc(sd, lb_nobusyg[idle]);
3411 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3413 schedstat_inc(sd, lb_nobusyq[idle]);
3417 BUG_ON(busiest == this_rq);
3419 schedstat_add(sd, lb_imbalance[idle], imbalance);
3422 if (busiest->nr_running > 1) {
3424 * Attempt to move tasks. If find_busiest_group has found
3425 * an imbalance but busiest->nr_running <= 1, the group is
3426 * still unbalanced. ld_moved simply stays zero, so it is
3427 * correctly treated as an imbalance.
3429 local_irq_save(flags);
3430 double_rq_lock(this_rq, busiest);
3431 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3432 imbalance, sd, idle, &all_pinned);
3433 double_rq_unlock(this_rq, busiest);
3434 local_irq_restore(flags);
3437 * some other cpu did the load balance for us.
3439 if (ld_moved && this_cpu != smp_processor_id())
3440 resched_cpu(this_cpu);
3442 /* All tasks on this runqueue were pinned by CPU affinity */
3443 if (unlikely(all_pinned)) {
3444 cpu_clear(cpu_of(busiest), *cpus);
3445 if (!cpus_empty(*cpus))
3452 schedstat_inc(sd, lb_failed[idle]);
3453 sd->nr_balance_failed++;
3455 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3457 spin_lock_irqsave(&busiest->lock, flags);
3459 /* don't kick the migration_thread, if the curr
3460 * task on busiest cpu can't be moved to this_cpu
3462 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3463 spin_unlock_irqrestore(&busiest->lock, flags);
3465 goto out_one_pinned;
3468 if (!busiest->active_balance) {
3469 busiest->active_balance = 1;
3470 busiest->push_cpu = this_cpu;
3473 spin_unlock_irqrestore(&busiest->lock, flags);
3475 wake_up_process(busiest->migration_thread);
3478 * We've kicked active balancing, reset the failure
3481 sd->nr_balance_failed = sd->cache_nice_tries+1;
3484 sd->nr_balance_failed = 0;
3486 if (likely(!active_balance)) {
3487 /* We were unbalanced, so reset the balancing interval */
3488 sd->balance_interval = sd->min_interval;
3491 * If we've begun active balancing, start to back off. This
3492 * case may not be covered by the all_pinned logic if there
3493 * is only 1 task on the busy runqueue (because we don't call
3496 if (sd->balance_interval < sd->max_interval)
3497 sd->balance_interval *= 2;
3500 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3501 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3507 schedstat_inc(sd, lb_balanced[idle]);
3509 sd->nr_balance_failed = 0;
3512 /* tune up the balancing interval */
3513 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3514 (sd->balance_interval < sd->max_interval))
3515 sd->balance_interval *= 2;
3517 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3518 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3529 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3530 * tasks if there is an imbalance.
3532 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3533 * this_rq is locked.
3536 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3539 struct sched_group *group;
3540 struct rq *busiest = NULL;
3541 unsigned long imbalance;
3549 * When power savings policy is enabled for the parent domain, idle
3550 * sibling can pick up load irrespective of busy siblings. In this case,
3551 * let the state of idle sibling percolate up as IDLE, instead of
3552 * portraying it as CPU_NOT_IDLE.
3554 if (sd->flags & SD_SHARE_CPUPOWER &&
3555 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3558 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3560 update_shares_locked(this_rq, sd);
3561 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3562 &sd_idle, cpus, NULL);
3564 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3568 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3570 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3574 BUG_ON(busiest == this_rq);
3576 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3579 if (busiest->nr_running > 1) {
3580 /* Attempt to move tasks */
3581 double_lock_balance(this_rq, busiest);
3582 /* this_rq->clock is already updated */
3583 update_rq_clock(busiest);
3584 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3585 imbalance, sd, CPU_NEWLY_IDLE,
3587 spin_unlock(&busiest->lock);
3589 if (unlikely(all_pinned)) {
3590 cpu_clear(cpu_of(busiest), *cpus);
3591 if (!cpus_empty(*cpus))
3597 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3598 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3599 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3602 sd->nr_balance_failed = 0;
3604 update_shares_locked(this_rq, sd);
3608 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3609 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3610 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3612 sd->nr_balance_failed = 0;
3618 * idle_balance is called by schedule() if this_cpu is about to become
3619 * idle. Attempts to pull tasks from other CPUs.
3621 static void idle_balance(int this_cpu, struct rq *this_rq)
3623 struct sched_domain *sd;
3624 int pulled_task = -1;
3625 unsigned long next_balance = jiffies + HZ;
3628 for_each_domain(this_cpu, sd) {
3629 unsigned long interval;
3631 if (!(sd->flags & SD_LOAD_BALANCE))
3634 if (sd->flags & SD_BALANCE_NEWIDLE)
3635 /* If we've pulled tasks over stop searching: */
3636 pulled_task = load_balance_newidle(this_cpu, this_rq,
3639 interval = msecs_to_jiffies(sd->balance_interval);
3640 if (time_after(next_balance, sd->last_balance + interval))
3641 next_balance = sd->last_balance + interval;
3645 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3647 * We are going idle. next_balance may be set based on
3648 * a busy processor. So reset next_balance.
3650 this_rq->next_balance = next_balance;
3655 * active_load_balance is run by migration threads. It pushes running tasks
3656 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3657 * running on each physical CPU where possible, and avoids physical /
3658 * logical imbalances.
3660 * Called with busiest_rq locked.
3662 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3664 int target_cpu = busiest_rq->push_cpu;
3665 struct sched_domain *sd;
3666 struct rq *target_rq;
3668 /* Is there any task to move? */
3669 if (busiest_rq->nr_running <= 1)
3672 target_rq = cpu_rq(target_cpu);
3675 * This condition is "impossible", if it occurs
3676 * we need to fix it. Originally reported by
3677 * Bjorn Helgaas on a 128-cpu setup.
3679 BUG_ON(busiest_rq == target_rq);
3681 /* move a task from busiest_rq to target_rq */
3682 double_lock_balance(busiest_rq, target_rq);
3683 update_rq_clock(busiest_rq);
3684 update_rq_clock(target_rq);
3686 /* Search for an sd spanning us and the target CPU. */
3687 for_each_domain(target_cpu, sd) {
3688 if ((sd->flags & SD_LOAD_BALANCE) &&
3689 cpu_isset(busiest_cpu, sd->span))
3694 schedstat_inc(sd, alb_count);
3696 if (move_one_task(target_rq, target_cpu, busiest_rq,
3698 schedstat_inc(sd, alb_pushed);
3700 schedstat_inc(sd, alb_failed);
3702 spin_unlock(&target_rq->lock);
3707 atomic_t load_balancer;
3709 } nohz ____cacheline_aligned = {
3710 .load_balancer = ATOMIC_INIT(-1),
3711 .cpu_mask = CPU_MASK_NONE,
3715 * This routine will try to nominate the ilb (idle load balancing)
3716 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3717 * load balancing on behalf of all those cpus. If all the cpus in the system
3718 * go into this tickless mode, then there will be no ilb owner (as there is
3719 * no need for one) and all the cpus will sleep till the next wakeup event
3722 * For the ilb owner, tick is not stopped. And this tick will be used
3723 * for idle load balancing. ilb owner will still be part of
3726 * While stopping the tick, this cpu will become the ilb owner if there
3727 * is no other owner. And will be the owner till that cpu becomes busy
3728 * or if all cpus in the system stop their ticks at which point
3729 * there is no need for ilb owner.
3731 * When the ilb owner becomes busy, it nominates another owner, during the
3732 * next busy scheduler_tick()
3734 int select_nohz_load_balancer(int stop_tick)
3736 int cpu = smp_processor_id();
3739 cpu_set(cpu, nohz.cpu_mask);
3740 cpu_rq(cpu)->in_nohz_recently = 1;
3743 * If we are going offline and still the leader, give up!
3745 if (cpu_is_offline(cpu) &&
3746 atomic_read(&nohz.load_balancer) == cpu) {
3747 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3752 /* time for ilb owner also to sleep */
3753 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3754 if (atomic_read(&nohz.load_balancer) == cpu)
3755 atomic_set(&nohz.load_balancer, -1);
3759 if (atomic_read(&nohz.load_balancer) == -1) {
3760 /* make me the ilb owner */
3761 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3763 } else if (atomic_read(&nohz.load_balancer) == cpu)
3766 if (!cpu_isset(cpu, nohz.cpu_mask))
3769 cpu_clear(cpu, nohz.cpu_mask);
3771 if (atomic_read(&nohz.load_balancer) == cpu)
3772 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3779 static DEFINE_SPINLOCK(balancing);
3782 * It checks each scheduling domain to see if it is due to be balanced,
3783 * and initiates a balancing operation if so.
3785 * Balancing parameters are set up in arch_init_sched_domains.
3787 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3790 struct rq *rq = cpu_rq(cpu);
3791 unsigned long interval;
3792 struct sched_domain *sd;
3793 /* Earliest time when we have to do rebalance again */
3794 unsigned long next_balance = jiffies + 60*HZ;
3795 int update_next_balance = 0;
3799 for_each_domain(cpu, sd) {
3800 if (!(sd->flags & SD_LOAD_BALANCE))
3803 interval = sd->balance_interval;
3804 if (idle != CPU_IDLE)
3805 interval *= sd->busy_factor;
3807 /* scale ms to jiffies */
3808 interval = msecs_to_jiffies(interval);
3809 if (unlikely(!interval))
3811 if (interval > HZ*NR_CPUS/10)
3812 interval = HZ*NR_CPUS/10;
3814 need_serialize = sd->flags & SD_SERIALIZE;
3816 if (need_serialize) {
3817 if (!spin_trylock(&balancing))
3821 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3822 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3824 * We've pulled tasks over so either we're no
3825 * longer idle, or one of our SMT siblings is
3828 idle = CPU_NOT_IDLE;
3830 sd->last_balance = jiffies;
3833 spin_unlock(&balancing);
3835 if (time_after(next_balance, sd->last_balance + interval)) {
3836 next_balance = sd->last_balance + interval;
3837 update_next_balance = 1;
3841 * Stop the load balance at this level. There is another
3842 * CPU in our sched group which is doing load balancing more
3850 * next_balance will be updated only when there is a need.
3851 * When the cpu is attached to null domain for ex, it will not be
3854 if (likely(update_next_balance))
3855 rq->next_balance = next_balance;
3859 * run_rebalance_domains is triggered when needed from the scheduler tick.
3860 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3861 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3863 static void run_rebalance_domains(struct softirq_action *h)
3865 int this_cpu = smp_processor_id();
3866 struct rq *this_rq = cpu_rq(this_cpu);
3867 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3868 CPU_IDLE : CPU_NOT_IDLE;
3870 rebalance_domains(this_cpu, idle);
3874 * If this cpu is the owner for idle load balancing, then do the
3875 * balancing on behalf of the other idle cpus whose ticks are
3878 if (this_rq->idle_at_tick &&
3879 atomic_read(&nohz.load_balancer) == this_cpu) {
3880 cpumask_t cpus = nohz.cpu_mask;
3884 cpu_clear(this_cpu, cpus);
3885 for_each_cpu_mask(balance_cpu, cpus) {
3887 * If this cpu gets work to do, stop the load balancing
3888 * work being done for other cpus. Next load
3889 * balancing owner will pick it up.
3894 rebalance_domains(balance_cpu, CPU_IDLE);
3896 rq = cpu_rq(balance_cpu);
3897 if (time_after(this_rq->next_balance, rq->next_balance))
3898 this_rq->next_balance = rq->next_balance;
3905 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3907 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3908 * idle load balancing owner or decide to stop the periodic load balancing,
3909 * if the whole system is idle.
3911 static inline void trigger_load_balance(struct rq *rq, int cpu)
3915 * If we were in the nohz mode recently and busy at the current
3916 * scheduler tick, then check if we need to nominate new idle
3919 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3920 rq->in_nohz_recently = 0;
3922 if (atomic_read(&nohz.load_balancer) == cpu) {
3923 cpu_clear(cpu, nohz.cpu_mask);
3924 atomic_set(&nohz.load_balancer, -1);
3927 if (atomic_read(&nohz.load_balancer) == -1) {
3929 * simple selection for now: Nominate the
3930 * first cpu in the nohz list to be the next
3933 * TBD: Traverse the sched domains and nominate
3934 * the nearest cpu in the nohz.cpu_mask.
3936 int ilb = first_cpu(nohz.cpu_mask);
3938 if (ilb < nr_cpu_ids)
3944 * If this cpu is idle and doing idle load balancing for all the
3945 * cpus with ticks stopped, is it time for that to stop?
3947 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3948 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3954 * If this cpu is idle and the idle load balancing is done by
3955 * someone else, then no need raise the SCHED_SOFTIRQ
3957 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3958 cpu_isset(cpu, nohz.cpu_mask))
3961 if (time_after_eq(jiffies, rq->next_balance))
3962 raise_softirq(SCHED_SOFTIRQ);
3965 #else /* CONFIG_SMP */
3968 * on UP we do not need to balance between CPUs:
3970 static inline void idle_balance(int cpu, struct rq *rq)
3976 DEFINE_PER_CPU(struct kernel_stat, kstat);
3978 EXPORT_PER_CPU_SYMBOL(kstat);
3981 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3982 * that have not yet been banked in case the task is currently running.
3984 unsigned long long task_sched_runtime(struct task_struct *p)
3986 unsigned long flags;
3990 rq = task_rq_lock(p, &flags);
3991 ns = p->se.sum_exec_runtime;
3992 if (task_current(rq, p)) {
3993 update_rq_clock(rq);
3994 delta_exec = rq->clock - p->se.exec_start;
3995 if ((s64)delta_exec > 0)
3998 task_rq_unlock(rq, &flags);
4004 * Account user cpu time to a process.
4005 * @p: the process that the cpu time gets accounted to
4006 * @cputime: the cpu time spent in user space since the last update
4008 void account_user_time(struct task_struct *p, cputime_t cputime)
4010 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4013 p->utime = cputime_add(p->utime, cputime);
4015 /* Add user time to cpustat. */
4016 tmp = cputime_to_cputime64(cputime);
4017 if (TASK_NICE(p) > 0)
4018 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4020 cpustat->user = cputime64_add(cpustat->user, tmp);
4024 * Account guest cpu time to a process.
4025 * @p: the process that the cpu time gets accounted to
4026 * @cputime: the cpu time spent in virtual machine since the last update
4028 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4031 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4033 tmp = cputime_to_cputime64(cputime);
4035 p->utime = cputime_add(p->utime, cputime);
4036 p->gtime = cputime_add(p->gtime, cputime);
4038 cpustat->user = cputime64_add(cpustat->user, tmp);
4039 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4043 * Account scaled user cpu time to a process.
4044 * @p: the process that the cpu time gets accounted to
4045 * @cputime: the cpu time spent in user space since the last update
4047 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4049 p->utimescaled = cputime_add(p->utimescaled, cputime);
4053 * Account system cpu time to a process.
4054 * @p: the process that the cpu time gets accounted to
4055 * @hardirq_offset: the offset to subtract from hardirq_count()
4056 * @cputime: the cpu time spent in kernel space since the last update
4058 void account_system_time(struct task_struct *p, int hardirq_offset,
4061 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4062 struct rq *rq = this_rq();
4065 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4066 account_guest_time(p, cputime);
4070 p->stime = cputime_add(p->stime, cputime);
4072 /* Add system time to cpustat. */
4073 tmp = cputime_to_cputime64(cputime);
4074 if (hardirq_count() - hardirq_offset)
4075 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4076 else if (softirq_count())
4077 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4078 else if (p != rq->idle)
4079 cpustat->system = cputime64_add(cpustat->system, tmp);
4080 else if (atomic_read(&rq->nr_iowait) > 0)
4081 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4083 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4084 /* Account for system time used */
4085 acct_update_integrals(p);
4089 * Account scaled system cpu time to a process.
4090 * @p: the process that the cpu time gets accounted to
4091 * @hardirq_offset: the offset to subtract from hardirq_count()
4092 * @cputime: the cpu time spent in kernel space since the last update
4094 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4096 p->stimescaled = cputime_add(p->stimescaled, cputime);
4100 * Account for involuntary wait time.
4101 * @p: the process from which the cpu time has been stolen
4102 * @steal: the cpu time spent in involuntary wait
4104 void account_steal_time(struct task_struct *p, cputime_t steal)
4106 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4107 cputime64_t tmp = cputime_to_cputime64(steal);
4108 struct rq *rq = this_rq();
4110 if (p == rq->idle) {
4111 p->stime = cputime_add(p->stime, steal);
4112 if (atomic_read(&rq->nr_iowait) > 0)
4113 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4115 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4117 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4121 * This function gets called by the timer code, with HZ frequency.
4122 * We call it with interrupts disabled.
4124 * It also gets called by the fork code, when changing the parent's
4127 void scheduler_tick(void)
4129 int cpu = smp_processor_id();
4130 struct rq *rq = cpu_rq(cpu);
4131 struct task_struct *curr = rq->curr;
4135 spin_lock(&rq->lock);
4136 update_rq_clock(rq);
4137 update_cpu_load(rq);
4138 curr->sched_class->task_tick(rq, curr, 0);
4139 spin_unlock(&rq->lock);
4142 rq->idle_at_tick = idle_cpu(cpu);
4143 trigger_load_balance(rq, cpu);
4147 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4149 void __kprobes add_preempt_count(int val)
4154 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4156 preempt_count() += val;
4158 * Spinlock count overflowing soon?
4160 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4163 EXPORT_SYMBOL(add_preempt_count);
4165 void __kprobes sub_preempt_count(int val)
4170 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4173 * Is the spinlock portion underflowing?
4175 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4176 !(preempt_count() & PREEMPT_MASK)))
4179 preempt_count() -= val;
4181 EXPORT_SYMBOL(sub_preempt_count);
4186 * Print scheduling while atomic bug:
4188 static noinline void __schedule_bug(struct task_struct *prev)
4190 struct pt_regs *regs = get_irq_regs();
4192 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4193 prev->comm, prev->pid, preempt_count());
4195 debug_show_held_locks(prev);
4197 if (irqs_disabled())
4198 print_irqtrace_events(prev);
4207 * Various schedule()-time debugging checks and statistics:
4209 static inline void schedule_debug(struct task_struct *prev)
4212 * Test if we are atomic. Since do_exit() needs to call into
4213 * schedule() atomically, we ignore that path for now.
4214 * Otherwise, whine if we are scheduling when we should not be.
4216 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4217 __schedule_bug(prev);
4219 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4221 schedstat_inc(this_rq(), sched_count);
4222 #ifdef CONFIG_SCHEDSTATS
4223 if (unlikely(prev->lock_depth >= 0)) {
4224 schedstat_inc(this_rq(), bkl_count);
4225 schedstat_inc(prev, sched_info.bkl_count);
4231 * Pick up the highest-prio task:
4233 static inline struct task_struct *
4234 pick_next_task(struct rq *rq, struct task_struct *prev)
4236 const struct sched_class *class;
4237 struct task_struct *p;
4240 * Optimization: we know that if all tasks are in
4241 * the fair class we can call that function directly:
4243 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4244 p = fair_sched_class.pick_next_task(rq);
4249 class = sched_class_highest;
4251 p = class->pick_next_task(rq);
4255 * Will never be NULL as the idle class always
4256 * returns a non-NULL p:
4258 class = class->next;
4263 * schedule() is the main scheduler function.
4265 asmlinkage void __sched schedule(void)
4267 struct task_struct *prev, *next;
4268 unsigned long *switch_count;
4270 int cpu, hrtick = sched_feat(HRTICK);
4274 cpu = smp_processor_id();
4278 switch_count = &prev->nivcsw;
4280 release_kernel_lock(prev);
4281 need_resched_nonpreemptible:
4283 schedule_debug(prev);
4289 * Do the rq-clock update outside the rq lock:
4291 local_irq_disable();
4292 update_rq_clock(rq);
4293 spin_lock(&rq->lock);
4294 clear_tsk_need_resched(prev);
4296 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4297 if (unlikely(signal_pending_state(prev->state, prev)))
4298 prev->state = TASK_RUNNING;
4300 deactivate_task(rq, prev, 1);
4301 switch_count = &prev->nvcsw;
4305 if (prev->sched_class->pre_schedule)
4306 prev->sched_class->pre_schedule(rq, prev);
4309 if (unlikely(!rq->nr_running))
4310 idle_balance(cpu, rq);
4312 prev->sched_class->put_prev_task(rq, prev);
4313 next = pick_next_task(rq, prev);
4315 if (likely(prev != next)) {
4316 sched_info_switch(prev, next);
4322 context_switch(rq, prev, next); /* unlocks the rq */
4324 * the context switch might have flipped the stack from under
4325 * us, hence refresh the local variables.
4327 cpu = smp_processor_id();
4330 spin_unlock_irq(&rq->lock);
4335 if (unlikely(reacquire_kernel_lock(current) < 0))
4336 goto need_resched_nonpreemptible;
4338 preempt_enable_no_resched();
4339 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4342 EXPORT_SYMBOL(schedule);
4344 #ifdef CONFIG_PREEMPT
4346 * this is the entry point to schedule() from in-kernel preemption
4347 * off of preempt_enable. Kernel preemptions off return from interrupt
4348 * occur there and call schedule directly.
4350 asmlinkage void __sched preempt_schedule(void)
4352 struct thread_info *ti = current_thread_info();
4355 * If there is a non-zero preempt_count or interrupts are disabled,
4356 * we do not want to preempt the current task. Just return..
4358 if (likely(ti->preempt_count || irqs_disabled()))
4362 add_preempt_count(PREEMPT_ACTIVE);
4364 sub_preempt_count(PREEMPT_ACTIVE);
4367 * Check again in case we missed a preemption opportunity
4368 * between schedule and now.
4371 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4373 EXPORT_SYMBOL(preempt_schedule);
4376 * this is the entry point to schedule() from kernel preemption
4377 * off of irq context.
4378 * Note, that this is called and return with irqs disabled. This will
4379 * protect us against recursive calling from irq.
4381 asmlinkage void __sched preempt_schedule_irq(void)
4383 struct thread_info *ti = current_thread_info();
4385 /* Catch callers which need to be fixed */
4386 BUG_ON(ti->preempt_count || !irqs_disabled());
4389 add_preempt_count(PREEMPT_ACTIVE);
4392 local_irq_disable();
4393 sub_preempt_count(PREEMPT_ACTIVE);
4396 * Check again in case we missed a preemption opportunity
4397 * between schedule and now.
4400 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4403 #endif /* CONFIG_PREEMPT */
4405 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4408 return try_to_wake_up(curr->private, mode, sync);
4410 EXPORT_SYMBOL(default_wake_function);
4413 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4414 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4415 * number) then we wake all the non-exclusive tasks and one exclusive task.
4417 * There are circumstances in which we can try to wake a task which has already
4418 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4419 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4421 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4422 int nr_exclusive, int sync, void *key)
4424 wait_queue_t *curr, *next;
4426 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4427 unsigned flags = curr->flags;
4429 if (curr->func(curr, mode, sync, key) &&
4430 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4436 * __wake_up - wake up threads blocked on a waitqueue.
4438 * @mode: which threads
4439 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4440 * @key: is directly passed to the wakeup function
4442 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4443 int nr_exclusive, void *key)
4445 unsigned long flags;
4447 spin_lock_irqsave(&q->lock, flags);
4448 __wake_up_common(q, mode, nr_exclusive, 0, key);
4449 spin_unlock_irqrestore(&q->lock, flags);
4451 EXPORT_SYMBOL(__wake_up);
4454 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4456 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4458 __wake_up_common(q, mode, 1, 0, NULL);
4462 * __wake_up_sync - wake up threads blocked on a waitqueue.
4464 * @mode: which threads
4465 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4467 * The sync wakeup differs that the waker knows that it will schedule
4468 * away soon, so while the target thread will be woken up, it will not
4469 * be migrated to another CPU - ie. the two threads are 'synchronized'
4470 * with each other. This can prevent needless bouncing between CPUs.
4472 * On UP it can prevent extra preemption.
4475 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4477 unsigned long flags;
4483 if (unlikely(!nr_exclusive))
4486 spin_lock_irqsave(&q->lock, flags);
4487 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4488 spin_unlock_irqrestore(&q->lock, flags);
4490 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4492 void complete(struct completion *x)
4494 unsigned long flags;
4496 spin_lock_irqsave(&x->wait.lock, flags);
4498 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4499 spin_unlock_irqrestore(&x->wait.lock, flags);
4501 EXPORT_SYMBOL(complete);
4503 void complete_all(struct completion *x)
4505 unsigned long flags;
4507 spin_lock_irqsave(&x->wait.lock, flags);
4508 x->done += UINT_MAX/2;
4509 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4510 spin_unlock_irqrestore(&x->wait.lock, flags);
4512 EXPORT_SYMBOL(complete_all);
4514 static inline long __sched
4515 do_wait_for_common(struct completion *x, long timeout, int state)
4518 DECLARE_WAITQUEUE(wait, current);
4520 wait.flags |= WQ_FLAG_EXCLUSIVE;
4521 __add_wait_queue_tail(&x->wait, &wait);
4523 if ((state == TASK_INTERRUPTIBLE &&
4524 signal_pending(current)) ||
4525 (state == TASK_KILLABLE &&
4526 fatal_signal_pending(current))) {
4527 timeout = -ERESTARTSYS;
4530 __set_current_state(state);
4531 spin_unlock_irq(&x->wait.lock);
4532 timeout = schedule_timeout(timeout);
4533 spin_lock_irq(&x->wait.lock);
4534 } while (!x->done && timeout);
4535 __remove_wait_queue(&x->wait, &wait);
4540 return timeout ?: 1;
4544 wait_for_common(struct completion *x, long timeout, int state)
4548 spin_lock_irq(&x->wait.lock);
4549 timeout = do_wait_for_common(x, timeout, state);
4550 spin_unlock_irq(&x->wait.lock);
4554 void __sched wait_for_completion(struct completion *x)
4556 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4558 EXPORT_SYMBOL(wait_for_completion);
4560 unsigned long __sched
4561 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4563 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4565 EXPORT_SYMBOL(wait_for_completion_timeout);
4567 int __sched wait_for_completion_interruptible(struct completion *x)
4569 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4570 if (t == -ERESTARTSYS)
4574 EXPORT_SYMBOL(wait_for_completion_interruptible);
4576 unsigned long __sched
4577 wait_for_completion_interruptible_timeout(struct completion *x,
4578 unsigned long timeout)
4580 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4582 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4584 int __sched wait_for_completion_killable(struct completion *x)
4586 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4587 if (t == -ERESTARTSYS)
4591 EXPORT_SYMBOL(wait_for_completion_killable);
4594 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4596 unsigned long flags;
4599 init_waitqueue_entry(&wait, current);
4601 __set_current_state(state);
4603 spin_lock_irqsave(&q->lock, flags);
4604 __add_wait_queue(q, &wait);
4605 spin_unlock(&q->lock);
4606 timeout = schedule_timeout(timeout);
4607 spin_lock_irq(&q->lock);
4608 __remove_wait_queue(q, &wait);
4609 spin_unlock_irqrestore(&q->lock, flags);
4614 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4616 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4618 EXPORT_SYMBOL(interruptible_sleep_on);
4621 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4623 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4625 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4627 void __sched sleep_on(wait_queue_head_t *q)
4629 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4631 EXPORT_SYMBOL(sleep_on);
4633 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4635 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4637 EXPORT_SYMBOL(sleep_on_timeout);
4639 #ifdef CONFIG_RT_MUTEXES
4642 * rt_mutex_setprio - set the current priority of a task
4644 * @prio: prio value (kernel-internal form)
4646 * This function changes the 'effective' priority of a task. It does
4647 * not touch ->normal_prio like __setscheduler().
4649 * Used by the rt_mutex code to implement priority inheritance logic.
4651 void rt_mutex_setprio(struct task_struct *p, int prio)
4653 unsigned long flags;
4654 int oldprio, on_rq, running;
4656 const struct sched_class *prev_class = p->sched_class;
4658 BUG_ON(prio < 0 || prio > MAX_PRIO);
4660 rq = task_rq_lock(p, &flags);
4661 update_rq_clock(rq);
4664 on_rq = p->se.on_rq;
4665 running = task_current(rq, p);
4667 dequeue_task(rq, p, 0);
4669 p->sched_class->put_prev_task(rq, p);
4672 p->sched_class = &rt_sched_class;
4674 p->sched_class = &fair_sched_class;
4679 p->sched_class->set_curr_task(rq);
4681 enqueue_task(rq, p, 0);
4683 check_class_changed(rq, p, prev_class, oldprio, running);
4685 task_rq_unlock(rq, &flags);
4690 void set_user_nice(struct task_struct *p, long nice)
4692 int old_prio, delta, on_rq;
4693 unsigned long flags;
4696 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4699 * We have to be careful, if called from sys_setpriority(),
4700 * the task might be in the middle of scheduling on another CPU.
4702 rq = task_rq_lock(p, &flags);
4703 update_rq_clock(rq);
4705 * The RT priorities are set via sched_setscheduler(), but we still
4706 * allow the 'normal' nice value to be set - but as expected
4707 * it wont have any effect on scheduling until the task is
4708 * SCHED_FIFO/SCHED_RR:
4710 if (task_has_rt_policy(p)) {
4711 p->static_prio = NICE_TO_PRIO(nice);
4714 on_rq = p->se.on_rq;
4716 dequeue_task(rq, p, 0);
4718 p->static_prio = NICE_TO_PRIO(nice);
4721 p->prio = effective_prio(p);
4722 delta = p->prio - old_prio;
4725 enqueue_task(rq, p, 0);
4727 * If the task increased its priority or is running and
4728 * lowered its priority, then reschedule its CPU:
4730 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4731 resched_task(rq->curr);
4734 task_rq_unlock(rq, &flags);
4736 EXPORT_SYMBOL(set_user_nice);
4739 * can_nice - check if a task can reduce its nice value
4743 int can_nice(const struct task_struct *p, const int nice)
4745 /* convert nice value [19,-20] to rlimit style value [1,40] */
4746 int nice_rlim = 20 - nice;
4748 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4749 capable(CAP_SYS_NICE));
4752 #ifdef __ARCH_WANT_SYS_NICE
4755 * sys_nice - change the priority of the current process.
4756 * @increment: priority increment
4758 * sys_setpriority is a more generic, but much slower function that
4759 * does similar things.
4761 asmlinkage long sys_nice(int increment)
4766 * Setpriority might change our priority at the same moment.
4767 * We don't have to worry. Conceptually one call occurs first
4768 * and we have a single winner.
4770 if (increment < -40)
4775 nice = PRIO_TO_NICE(current->static_prio) + increment;
4781 if (increment < 0 && !can_nice(current, nice))
4784 retval = security_task_setnice(current, nice);
4788 set_user_nice(current, nice);
4795 * task_prio - return the priority value of a given task.
4796 * @p: the task in question.
4798 * This is the priority value as seen by users in /proc.
4799 * RT tasks are offset by -200. Normal tasks are centered
4800 * around 0, value goes from -16 to +15.
4802 int task_prio(const struct task_struct *p)
4804 return p->prio - MAX_RT_PRIO;
4808 * task_nice - return the nice value of a given task.
4809 * @p: the task in question.
4811 int task_nice(const struct task_struct *p)
4813 return TASK_NICE(p);
4815 EXPORT_SYMBOL(task_nice);
4818 * idle_cpu - is a given cpu idle currently?
4819 * @cpu: the processor in question.
4821 int idle_cpu(int cpu)
4823 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4827 * idle_task - return the idle task for a given cpu.
4828 * @cpu: the processor in question.
4830 struct task_struct *idle_task(int cpu)
4832 return cpu_rq(cpu)->idle;
4836 * find_process_by_pid - find a process with a matching PID value.
4837 * @pid: the pid in question.
4839 static struct task_struct *find_process_by_pid(pid_t pid)
4841 return pid ? find_task_by_vpid(pid) : current;
4844 /* Actually do priority change: must hold rq lock. */
4846 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4848 BUG_ON(p->se.on_rq);
4851 switch (p->policy) {
4855 p->sched_class = &fair_sched_class;
4859 p->sched_class = &rt_sched_class;
4863 p->rt_priority = prio;
4864 p->normal_prio = normal_prio(p);
4865 /* we are holding p->pi_lock already */
4866 p->prio = rt_mutex_getprio(p);
4871 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4872 * @p: the task in question.
4873 * @policy: new policy.
4874 * @param: structure containing the new RT priority.
4876 * NOTE that the task may be already dead.
4878 int sched_setscheduler(struct task_struct *p, int policy,
4879 struct sched_param *param)
4881 int retval, oldprio, oldpolicy = -1, on_rq, running;
4882 unsigned long flags;
4883 const struct sched_class *prev_class = p->sched_class;
4886 /* may grab non-irq protected spin_locks */
4887 BUG_ON(in_interrupt());
4889 /* double check policy once rq lock held */
4891 policy = oldpolicy = p->policy;
4892 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4893 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4894 policy != SCHED_IDLE)
4897 * Valid priorities for SCHED_FIFO and SCHED_RR are
4898 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4899 * SCHED_BATCH and SCHED_IDLE is 0.
4901 if (param->sched_priority < 0 ||
4902 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4903 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4905 if (rt_policy(policy) != (param->sched_priority != 0))
4909 * Allow unprivileged RT tasks to decrease priority:
4911 if (!capable(CAP_SYS_NICE)) {
4912 if (rt_policy(policy)) {
4913 unsigned long rlim_rtprio;
4915 if (!lock_task_sighand(p, &flags))
4917 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4918 unlock_task_sighand(p, &flags);
4920 /* can't set/change the rt policy */
4921 if (policy != p->policy && !rlim_rtprio)
4924 /* can't increase priority */
4925 if (param->sched_priority > p->rt_priority &&
4926 param->sched_priority > rlim_rtprio)
4930 * Like positive nice levels, dont allow tasks to
4931 * move out of SCHED_IDLE either:
4933 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4936 /* can't change other user's priorities */
4937 if ((current->euid != p->euid) &&
4938 (current->euid != p->uid))
4942 #ifdef CONFIG_RT_GROUP_SCHED
4944 * Do not allow realtime tasks into groups that have no runtime
4947 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4951 retval = security_task_setscheduler(p, policy, param);
4955 * make sure no PI-waiters arrive (or leave) while we are
4956 * changing the priority of the task:
4958 spin_lock_irqsave(&p->pi_lock, flags);
4960 * To be able to change p->policy safely, the apropriate
4961 * runqueue lock must be held.
4963 rq = __task_rq_lock(p);
4964 /* recheck policy now with rq lock held */
4965 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4966 policy = oldpolicy = -1;
4967 __task_rq_unlock(rq);
4968 spin_unlock_irqrestore(&p->pi_lock, flags);
4971 update_rq_clock(rq);
4972 on_rq = p->se.on_rq;
4973 running = task_current(rq, p);
4975 deactivate_task(rq, p, 0);
4977 p->sched_class->put_prev_task(rq, p);
4980 __setscheduler(rq, p, policy, param->sched_priority);
4983 p->sched_class->set_curr_task(rq);
4985 activate_task(rq, p, 0);
4987 check_class_changed(rq, p, prev_class, oldprio, running);
4989 __task_rq_unlock(rq);
4990 spin_unlock_irqrestore(&p->pi_lock, flags);
4992 rt_mutex_adjust_pi(p);
4996 EXPORT_SYMBOL_GPL(sched_setscheduler);
4999 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5001 struct sched_param lparam;
5002 struct task_struct *p;
5005 if (!param || pid < 0)
5007 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5012 p = find_process_by_pid(pid);
5014 retval = sched_setscheduler(p, policy, &lparam);
5021 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5022 * @pid: the pid in question.
5023 * @policy: new policy.
5024 * @param: structure containing the new RT priority.
5027 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5029 /* negative values for policy are not valid */
5033 return do_sched_setscheduler(pid, policy, param);
5037 * sys_sched_setparam - set/change the RT priority of a thread
5038 * @pid: the pid in question.
5039 * @param: structure containing the new RT priority.
5041 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5043 return do_sched_setscheduler(pid, -1, param);
5047 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5048 * @pid: the pid in question.
5050 asmlinkage long sys_sched_getscheduler(pid_t pid)
5052 struct task_struct *p;
5059 read_lock(&tasklist_lock);
5060 p = find_process_by_pid(pid);
5062 retval = security_task_getscheduler(p);
5066 read_unlock(&tasklist_lock);
5071 * sys_sched_getscheduler - get the RT priority of a thread
5072 * @pid: the pid in question.
5073 * @param: structure containing the RT priority.
5075 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5077 struct sched_param lp;
5078 struct task_struct *p;
5081 if (!param || pid < 0)
5084 read_lock(&tasklist_lock);
5085 p = find_process_by_pid(pid);
5090 retval = security_task_getscheduler(p);
5094 lp.sched_priority = p->rt_priority;
5095 read_unlock(&tasklist_lock);
5098 * This one might sleep, we cannot do it with a spinlock held ...
5100 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5105 read_unlock(&tasklist_lock);
5109 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5111 cpumask_t cpus_allowed;
5112 cpumask_t new_mask = *in_mask;
5113 struct task_struct *p;
5117 read_lock(&tasklist_lock);
5119 p = find_process_by_pid(pid);
5121 read_unlock(&tasklist_lock);
5127 * It is not safe to call set_cpus_allowed with the
5128 * tasklist_lock held. We will bump the task_struct's
5129 * usage count and then drop tasklist_lock.
5132 read_unlock(&tasklist_lock);
5135 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5136 !capable(CAP_SYS_NICE))
5139 retval = security_task_setscheduler(p, 0, NULL);
5143 cpuset_cpus_allowed(p, &cpus_allowed);
5144 cpus_and(new_mask, new_mask, cpus_allowed);
5146 retval = set_cpus_allowed_ptr(p, &new_mask);
5149 cpuset_cpus_allowed(p, &cpus_allowed);
5150 if (!cpus_subset(new_mask, cpus_allowed)) {
5152 * We must have raced with a concurrent cpuset
5153 * update. Just reset the cpus_allowed to the
5154 * cpuset's cpus_allowed
5156 new_mask = cpus_allowed;
5166 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5167 cpumask_t *new_mask)
5169 if (len < sizeof(cpumask_t)) {
5170 memset(new_mask, 0, sizeof(cpumask_t));
5171 } else if (len > sizeof(cpumask_t)) {
5172 len = sizeof(cpumask_t);
5174 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5178 * sys_sched_setaffinity - set the cpu affinity of a process
5179 * @pid: pid of the process
5180 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5181 * @user_mask_ptr: user-space pointer to the new cpu mask
5183 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5184 unsigned long __user *user_mask_ptr)
5189 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5193 return sched_setaffinity(pid, &new_mask);
5196 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5198 struct task_struct *p;
5202 read_lock(&tasklist_lock);
5205 p = find_process_by_pid(pid);
5209 retval = security_task_getscheduler(p);
5213 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5216 read_unlock(&tasklist_lock);
5223 * sys_sched_getaffinity - get the cpu affinity of a process
5224 * @pid: pid of the process
5225 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5226 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5228 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5229 unsigned long __user *user_mask_ptr)
5234 if (len < sizeof(cpumask_t))
5237 ret = sched_getaffinity(pid, &mask);
5241 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5244 return sizeof(cpumask_t);
5248 * sys_sched_yield - yield the current processor to other threads.
5250 * This function yields the current CPU to other tasks. If there are no
5251 * other threads running on this CPU then this function will return.
5253 asmlinkage long sys_sched_yield(void)
5255 struct rq *rq = this_rq_lock();
5257 schedstat_inc(rq, yld_count);
5258 current->sched_class->yield_task(rq);
5261 * Since we are going to call schedule() anyway, there's
5262 * no need to preempt or enable interrupts:
5264 __release(rq->lock);
5265 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5266 _raw_spin_unlock(&rq->lock);
5267 preempt_enable_no_resched();
5274 static void __cond_resched(void)
5276 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5277 __might_sleep(__FILE__, __LINE__);
5280 * The BKS might be reacquired before we have dropped
5281 * PREEMPT_ACTIVE, which could trigger a second
5282 * cond_resched() call.
5285 add_preempt_count(PREEMPT_ACTIVE);
5287 sub_preempt_count(PREEMPT_ACTIVE);
5288 } while (need_resched());
5291 int __sched _cond_resched(void)
5293 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5294 system_state == SYSTEM_RUNNING) {
5300 EXPORT_SYMBOL(_cond_resched);
5303 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5304 * call schedule, and on return reacquire the lock.
5306 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5307 * operations here to prevent schedule() from being called twice (once via
5308 * spin_unlock(), once by hand).
5310 int cond_resched_lock(spinlock_t *lock)
5312 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5315 if (spin_needbreak(lock) || resched) {
5317 if (resched && need_resched())
5326 EXPORT_SYMBOL(cond_resched_lock);
5328 int __sched cond_resched_softirq(void)
5330 BUG_ON(!in_softirq());
5332 if (need_resched() && system_state == SYSTEM_RUNNING) {
5340 EXPORT_SYMBOL(cond_resched_softirq);
5343 * yield - yield the current processor to other threads.
5345 * This is a shortcut for kernel-space yielding - it marks the
5346 * thread runnable and calls sys_sched_yield().
5348 void __sched yield(void)
5350 set_current_state(TASK_RUNNING);
5353 EXPORT_SYMBOL(yield);
5356 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5357 * that process accounting knows that this is a task in IO wait state.
5359 * But don't do that if it is a deliberate, throttling IO wait (this task
5360 * has set its backing_dev_info: the queue against which it should throttle)
5362 void __sched io_schedule(void)
5364 struct rq *rq = &__raw_get_cpu_var(runqueues);
5366 delayacct_blkio_start();
5367 atomic_inc(&rq->nr_iowait);
5369 atomic_dec(&rq->nr_iowait);
5370 delayacct_blkio_end();
5372 EXPORT_SYMBOL(io_schedule);
5374 long __sched io_schedule_timeout(long timeout)
5376 struct rq *rq = &__raw_get_cpu_var(runqueues);
5379 delayacct_blkio_start();
5380 atomic_inc(&rq->nr_iowait);
5381 ret = schedule_timeout(timeout);
5382 atomic_dec(&rq->nr_iowait);
5383 delayacct_blkio_end();
5388 * sys_sched_get_priority_max - return maximum RT priority.
5389 * @policy: scheduling class.
5391 * this syscall returns the maximum rt_priority that can be used
5392 * by a given scheduling class.
5394 asmlinkage long sys_sched_get_priority_max(int policy)
5401 ret = MAX_USER_RT_PRIO-1;
5413 * sys_sched_get_priority_min - return minimum RT priority.
5414 * @policy: scheduling class.
5416 * this syscall returns the minimum rt_priority that can be used
5417 * by a given scheduling class.
5419 asmlinkage long sys_sched_get_priority_min(int policy)
5437 * sys_sched_rr_get_interval - return the default timeslice of a process.
5438 * @pid: pid of the process.
5439 * @interval: userspace pointer to the timeslice value.
5441 * this syscall writes the default timeslice value of a given process
5442 * into the user-space timespec buffer. A value of '0' means infinity.
5445 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5447 struct task_struct *p;
5448 unsigned int time_slice;
5456 read_lock(&tasklist_lock);
5457 p = find_process_by_pid(pid);
5461 retval = security_task_getscheduler(p);
5466 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5467 * tasks that are on an otherwise idle runqueue:
5470 if (p->policy == SCHED_RR) {
5471 time_slice = DEF_TIMESLICE;
5472 } else if (p->policy != SCHED_FIFO) {
5473 struct sched_entity *se = &p->se;
5474 unsigned long flags;
5477 rq = task_rq_lock(p, &flags);
5478 if (rq->cfs.load.weight)
5479 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5480 task_rq_unlock(rq, &flags);
5482 read_unlock(&tasklist_lock);
5483 jiffies_to_timespec(time_slice, &t);
5484 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5488 read_unlock(&tasklist_lock);
5492 static const char stat_nam[] = "RSDTtZX";
5494 void sched_show_task(struct task_struct *p)
5496 unsigned long free = 0;
5499 state = p->state ? __ffs(p->state) + 1 : 0;
5500 printk(KERN_INFO "%-13.13s %c", p->comm,
5501 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5502 #if BITS_PER_LONG == 32
5503 if (state == TASK_RUNNING)
5504 printk(KERN_CONT " running ");
5506 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5508 if (state == TASK_RUNNING)
5509 printk(KERN_CONT " running task ");
5511 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5513 #ifdef CONFIG_DEBUG_STACK_USAGE
5515 unsigned long *n = end_of_stack(p);
5518 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5521 printk(KERN_CONT "%5lu %5d %6d\n", free,
5522 task_pid_nr(p), task_pid_nr(p->real_parent));
5524 show_stack(p, NULL);
5527 void show_state_filter(unsigned long state_filter)
5529 struct task_struct *g, *p;
5531 #if BITS_PER_LONG == 32
5533 " task PC stack pid father\n");
5536 " task PC stack pid father\n");
5538 read_lock(&tasklist_lock);
5539 do_each_thread(g, p) {
5541 * reset the NMI-timeout, listing all files on a slow
5542 * console might take alot of time:
5544 touch_nmi_watchdog();
5545 if (!state_filter || (p->state & state_filter))
5547 } while_each_thread(g, p);
5549 touch_all_softlockup_watchdogs();
5551 #ifdef CONFIG_SCHED_DEBUG
5552 sysrq_sched_debug_show();
5554 read_unlock(&tasklist_lock);
5556 * Only show locks if all tasks are dumped:
5558 if (state_filter == -1)
5559 debug_show_all_locks();
5562 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5564 idle->sched_class = &idle_sched_class;
5568 * init_idle - set up an idle thread for a given CPU
5569 * @idle: task in question
5570 * @cpu: cpu the idle task belongs to
5572 * NOTE: this function does not set the idle thread's NEED_RESCHED
5573 * flag, to make booting more robust.
5575 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5577 struct rq *rq = cpu_rq(cpu);
5578 unsigned long flags;
5581 idle->se.exec_start = sched_clock();
5583 idle->prio = idle->normal_prio = MAX_PRIO;
5584 idle->cpus_allowed = cpumask_of_cpu(cpu);
5585 __set_task_cpu(idle, cpu);
5587 spin_lock_irqsave(&rq->lock, flags);
5588 rq->curr = rq->idle = idle;
5589 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5592 spin_unlock_irqrestore(&rq->lock, flags);
5594 /* Set the preempt count _outside_ the spinlocks! */
5595 #if defined(CONFIG_PREEMPT)
5596 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5598 task_thread_info(idle)->preempt_count = 0;
5601 * The idle tasks have their own, simple scheduling class:
5603 idle->sched_class = &idle_sched_class;
5607 * In a system that switches off the HZ timer nohz_cpu_mask
5608 * indicates which cpus entered this state. This is used
5609 * in the rcu update to wait only for active cpus. For system
5610 * which do not switch off the HZ timer nohz_cpu_mask should
5611 * always be CPU_MASK_NONE.
5613 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5616 * Increase the granularity value when there are more CPUs,
5617 * because with more CPUs the 'effective latency' as visible
5618 * to users decreases. But the relationship is not linear,
5619 * so pick a second-best guess by going with the log2 of the
5622 * This idea comes from the SD scheduler of Con Kolivas:
5624 static inline void sched_init_granularity(void)
5626 unsigned int factor = 1 + ilog2(num_online_cpus());
5627 const unsigned long limit = 200000000;
5629 sysctl_sched_min_granularity *= factor;
5630 if (sysctl_sched_min_granularity > limit)
5631 sysctl_sched_min_granularity = limit;
5633 sysctl_sched_latency *= factor;
5634 if (sysctl_sched_latency > limit)
5635 sysctl_sched_latency = limit;
5637 sysctl_sched_wakeup_granularity *= factor;
5642 * This is how migration works:
5644 * 1) we queue a struct migration_req structure in the source CPU's
5645 * runqueue and wake up that CPU's migration thread.
5646 * 2) we down() the locked semaphore => thread blocks.
5647 * 3) migration thread wakes up (implicitly it forces the migrated
5648 * thread off the CPU)
5649 * 4) it gets the migration request and checks whether the migrated
5650 * task is still in the wrong runqueue.
5651 * 5) if it's in the wrong runqueue then the migration thread removes
5652 * it and puts it into the right queue.
5653 * 6) migration thread up()s the semaphore.
5654 * 7) we wake up and the migration is done.
5658 * Change a given task's CPU affinity. Migrate the thread to a
5659 * proper CPU and schedule it away if the CPU it's executing on
5660 * is removed from the allowed bitmask.
5662 * NOTE: the caller must have a valid reference to the task, the
5663 * task must not exit() & deallocate itself prematurely. The
5664 * call is not atomic; no spinlocks may be held.
5666 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5668 struct migration_req req;
5669 unsigned long flags;
5673 rq = task_rq_lock(p, &flags);
5674 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5679 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5680 !cpus_equal(p->cpus_allowed, *new_mask))) {
5685 if (p->sched_class->set_cpus_allowed)
5686 p->sched_class->set_cpus_allowed(p, new_mask);
5688 p->cpus_allowed = *new_mask;
5689 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5692 /* Can the task run on the task's current CPU? If so, we're done */
5693 if (cpu_isset(task_cpu(p), *new_mask))
5696 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5697 /* Need help from migration thread: drop lock and wait. */
5698 task_rq_unlock(rq, &flags);
5699 wake_up_process(rq->migration_thread);
5700 wait_for_completion(&req.done);
5701 tlb_migrate_finish(p->mm);
5705 task_rq_unlock(rq, &flags);
5709 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5712 * Move (not current) task off this cpu, onto dest cpu. We're doing
5713 * this because either it can't run here any more (set_cpus_allowed()
5714 * away from this CPU, or CPU going down), or because we're
5715 * attempting to rebalance this task on exec (sched_exec).
5717 * So we race with normal scheduler movements, but that's OK, as long
5718 * as the task is no longer on this CPU.
5720 * Returns non-zero if task was successfully migrated.
5722 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5724 struct rq *rq_dest, *rq_src;
5727 if (unlikely(cpu_is_offline(dest_cpu)))
5730 rq_src = cpu_rq(src_cpu);
5731 rq_dest = cpu_rq(dest_cpu);
5733 double_rq_lock(rq_src, rq_dest);
5734 /* Already moved. */
5735 if (task_cpu(p) != src_cpu)
5737 /* Affinity changed (again). */
5738 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5741 on_rq = p->se.on_rq;
5743 deactivate_task(rq_src, p, 0);
5745 set_task_cpu(p, dest_cpu);
5747 activate_task(rq_dest, p, 0);
5748 check_preempt_curr(rq_dest, p);
5752 double_rq_unlock(rq_src, rq_dest);
5757 * migration_thread - this is a highprio system thread that performs
5758 * thread migration by bumping thread off CPU then 'pushing' onto
5761 static int migration_thread(void *data)
5763 int cpu = (long)data;
5767 BUG_ON(rq->migration_thread != current);
5769 set_current_state(TASK_INTERRUPTIBLE);
5770 while (!kthread_should_stop()) {
5771 struct migration_req *req;
5772 struct list_head *head;
5774 spin_lock_irq(&rq->lock);
5776 if (cpu_is_offline(cpu)) {
5777 spin_unlock_irq(&rq->lock);
5781 if (rq->active_balance) {
5782 active_load_balance(rq, cpu);
5783 rq->active_balance = 0;
5786 head = &rq->migration_queue;
5788 if (list_empty(head)) {
5789 spin_unlock_irq(&rq->lock);
5791 set_current_state(TASK_INTERRUPTIBLE);
5794 req = list_entry(head->next, struct migration_req, list);
5795 list_del_init(head->next);
5797 spin_unlock(&rq->lock);
5798 __migrate_task(req->task, cpu, req->dest_cpu);
5801 complete(&req->done);
5803 __set_current_state(TASK_RUNNING);
5807 /* Wait for kthread_stop */
5808 set_current_state(TASK_INTERRUPTIBLE);
5809 while (!kthread_should_stop()) {
5811 set_current_state(TASK_INTERRUPTIBLE);
5813 __set_current_state(TASK_RUNNING);
5817 #ifdef CONFIG_HOTPLUG_CPU
5819 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5823 local_irq_disable();
5824 ret = __migrate_task(p, src_cpu, dest_cpu);
5830 * Figure out where task on dead CPU should go, use force if necessary.
5831 * NOTE: interrupts should be disabled by the caller
5833 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5835 unsigned long flags;
5842 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5843 cpus_and(mask, mask, p->cpus_allowed);
5844 dest_cpu = any_online_cpu(mask);
5846 /* On any allowed CPU? */
5847 if (dest_cpu >= nr_cpu_ids)
5848 dest_cpu = any_online_cpu(p->cpus_allowed);
5850 /* No more Mr. Nice Guy. */
5851 if (dest_cpu >= nr_cpu_ids) {
5852 cpumask_t cpus_allowed;
5854 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5856 * Try to stay on the same cpuset, where the
5857 * current cpuset may be a subset of all cpus.
5858 * The cpuset_cpus_allowed_locked() variant of
5859 * cpuset_cpus_allowed() will not block. It must be
5860 * called within calls to cpuset_lock/cpuset_unlock.
5862 rq = task_rq_lock(p, &flags);
5863 p->cpus_allowed = cpus_allowed;
5864 dest_cpu = any_online_cpu(p->cpus_allowed);
5865 task_rq_unlock(rq, &flags);
5868 * Don't tell them about moving exiting tasks or
5869 * kernel threads (both mm NULL), since they never
5872 if (p->mm && printk_ratelimit()) {
5873 printk(KERN_INFO "process %d (%s) no "
5874 "longer affine to cpu%d\n",
5875 task_pid_nr(p), p->comm, dead_cpu);
5878 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5882 * While a dead CPU has no uninterruptible tasks queued at this point,
5883 * it might still have a nonzero ->nr_uninterruptible counter, because
5884 * for performance reasons the counter is not stricly tracking tasks to
5885 * their home CPUs. So we just add the counter to another CPU's counter,
5886 * to keep the global sum constant after CPU-down:
5888 static void migrate_nr_uninterruptible(struct rq *rq_src)
5890 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5891 unsigned long flags;
5893 local_irq_save(flags);
5894 double_rq_lock(rq_src, rq_dest);
5895 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5896 rq_src->nr_uninterruptible = 0;
5897 double_rq_unlock(rq_src, rq_dest);
5898 local_irq_restore(flags);
5901 /* Run through task list and migrate tasks from the dead cpu. */
5902 static void migrate_live_tasks(int src_cpu)
5904 struct task_struct *p, *t;
5906 read_lock(&tasklist_lock);
5908 do_each_thread(t, p) {
5912 if (task_cpu(p) == src_cpu)
5913 move_task_off_dead_cpu(src_cpu, p);
5914 } while_each_thread(t, p);
5916 read_unlock(&tasklist_lock);
5920 * Schedules idle task to be the next runnable task on current CPU.
5921 * It does so by boosting its priority to highest possible.
5922 * Used by CPU offline code.
5924 void sched_idle_next(void)
5926 int this_cpu = smp_processor_id();
5927 struct rq *rq = cpu_rq(this_cpu);
5928 struct task_struct *p = rq->idle;
5929 unsigned long flags;
5931 /* cpu has to be offline */
5932 BUG_ON(cpu_online(this_cpu));
5935 * Strictly not necessary since rest of the CPUs are stopped by now
5936 * and interrupts disabled on the current cpu.
5938 spin_lock_irqsave(&rq->lock, flags);
5940 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5942 update_rq_clock(rq);
5943 activate_task(rq, p, 0);
5945 spin_unlock_irqrestore(&rq->lock, flags);
5949 * Ensures that the idle task is using init_mm right before its cpu goes
5952 void idle_task_exit(void)
5954 struct mm_struct *mm = current->active_mm;
5956 BUG_ON(cpu_online(smp_processor_id()));
5959 switch_mm(mm, &init_mm, current);
5963 /* called under rq->lock with disabled interrupts */
5964 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5966 struct rq *rq = cpu_rq(dead_cpu);
5968 /* Must be exiting, otherwise would be on tasklist. */
5969 BUG_ON(!p->exit_state);
5971 /* Cannot have done final schedule yet: would have vanished. */
5972 BUG_ON(p->state == TASK_DEAD);
5977 * Drop lock around migration; if someone else moves it,
5978 * that's OK. No task can be added to this CPU, so iteration is
5981 spin_unlock_irq(&rq->lock);
5982 move_task_off_dead_cpu(dead_cpu, p);
5983 spin_lock_irq(&rq->lock);
5988 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5989 static void migrate_dead_tasks(unsigned int dead_cpu)
5991 struct rq *rq = cpu_rq(dead_cpu);
5992 struct task_struct *next;
5995 if (!rq->nr_running)
5997 update_rq_clock(rq);
5998 next = pick_next_task(rq, rq->curr);
6001 migrate_dead(dead_cpu, next);
6005 #endif /* CONFIG_HOTPLUG_CPU */
6007 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6009 static struct ctl_table sd_ctl_dir[] = {
6011 .procname = "sched_domain",
6017 static struct ctl_table sd_ctl_root[] = {
6019 .ctl_name = CTL_KERN,
6020 .procname = "kernel",
6022 .child = sd_ctl_dir,
6027 static struct ctl_table *sd_alloc_ctl_entry(int n)
6029 struct ctl_table *entry =
6030 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6035 static void sd_free_ctl_entry(struct ctl_table **tablep)
6037 struct ctl_table *entry;
6040 * In the intermediate directories, both the child directory and
6041 * procname are dynamically allocated and could fail but the mode
6042 * will always be set. In the lowest directory the names are
6043 * static strings and all have proc handlers.
6045 for (entry = *tablep; entry->mode; entry++) {
6047 sd_free_ctl_entry(&entry->child);
6048 if (entry->proc_handler == NULL)
6049 kfree(entry->procname);
6057 set_table_entry(struct ctl_table *entry,
6058 const char *procname, void *data, int maxlen,
6059 mode_t mode, proc_handler *proc_handler)
6061 entry->procname = procname;
6063 entry->maxlen = maxlen;
6065 entry->proc_handler = proc_handler;
6068 static struct ctl_table *
6069 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6071 struct ctl_table *table = sd_alloc_ctl_entry(12);
6076 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6077 sizeof(long), 0644, proc_doulongvec_minmax);
6078 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6079 sizeof(long), 0644, proc_doulongvec_minmax);
6080 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6081 sizeof(int), 0644, proc_dointvec_minmax);
6082 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6083 sizeof(int), 0644, proc_dointvec_minmax);
6084 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6085 sizeof(int), 0644, proc_dointvec_minmax);
6086 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6087 sizeof(int), 0644, proc_dointvec_minmax);
6088 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6089 sizeof(int), 0644, proc_dointvec_minmax);
6090 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6091 sizeof(int), 0644, proc_dointvec_minmax);
6092 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6093 sizeof(int), 0644, proc_dointvec_minmax);
6094 set_table_entry(&table[9], "cache_nice_tries",
6095 &sd->cache_nice_tries,
6096 sizeof(int), 0644, proc_dointvec_minmax);
6097 set_table_entry(&table[10], "flags", &sd->flags,
6098 sizeof(int), 0644, proc_dointvec_minmax);
6099 /* &table[11] is terminator */
6104 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6106 struct ctl_table *entry, *table;
6107 struct sched_domain *sd;
6108 int domain_num = 0, i;
6111 for_each_domain(cpu, sd)
6113 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6118 for_each_domain(cpu, sd) {
6119 snprintf(buf, 32, "domain%d", i);
6120 entry->procname = kstrdup(buf, GFP_KERNEL);
6122 entry->child = sd_alloc_ctl_domain_table(sd);
6129 static struct ctl_table_header *sd_sysctl_header;
6130 static void register_sched_domain_sysctl(void)
6132 int i, cpu_num = num_online_cpus();
6133 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6136 WARN_ON(sd_ctl_dir[0].child);
6137 sd_ctl_dir[0].child = entry;
6142 for_each_online_cpu(i) {
6143 snprintf(buf, 32, "cpu%d", i);
6144 entry->procname = kstrdup(buf, GFP_KERNEL);
6146 entry->child = sd_alloc_ctl_cpu_table(i);
6150 WARN_ON(sd_sysctl_header);
6151 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6154 /* may be called multiple times per register */
6155 static void unregister_sched_domain_sysctl(void)
6157 if (sd_sysctl_header)
6158 unregister_sysctl_table(sd_sysctl_header);
6159 sd_sysctl_header = NULL;
6160 if (sd_ctl_dir[0].child)
6161 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6164 static void register_sched_domain_sysctl(void)
6167 static void unregister_sched_domain_sysctl(void)
6172 static void set_rq_online(struct rq *rq)
6175 const struct sched_class *class;
6177 cpu_set(rq->cpu, rq->rd->online);
6180 for_each_class(class) {
6181 if (class->rq_online)
6182 class->rq_online(rq);
6187 static void set_rq_offline(struct rq *rq)
6190 const struct sched_class *class;
6192 for_each_class(class) {
6193 if (class->rq_offline)
6194 class->rq_offline(rq);
6197 cpu_clear(rq->cpu, rq->rd->online);
6203 * migration_call - callback that gets triggered when a CPU is added.
6204 * Here we can start up the necessary migration thread for the new CPU.
6206 static int __cpuinit
6207 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6209 struct task_struct *p;
6210 int cpu = (long)hcpu;
6211 unsigned long flags;
6216 case CPU_UP_PREPARE:
6217 case CPU_UP_PREPARE_FROZEN:
6218 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6221 kthread_bind(p, cpu);
6222 /* Must be high prio: stop_machine expects to yield to it. */
6223 rq = task_rq_lock(p, &flags);
6224 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6225 task_rq_unlock(rq, &flags);
6226 cpu_rq(cpu)->migration_thread = p;
6230 case CPU_ONLINE_FROZEN:
6231 /* Strictly unnecessary, as first user will wake it. */
6232 wake_up_process(cpu_rq(cpu)->migration_thread);
6234 /* Update our root-domain */
6236 spin_lock_irqsave(&rq->lock, flags);
6238 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6242 spin_unlock_irqrestore(&rq->lock, flags);
6245 #ifdef CONFIG_HOTPLUG_CPU
6246 case CPU_UP_CANCELED:
6247 case CPU_UP_CANCELED_FROZEN:
6248 if (!cpu_rq(cpu)->migration_thread)
6250 /* Unbind it from offline cpu so it can run. Fall thru. */
6251 kthread_bind(cpu_rq(cpu)->migration_thread,
6252 any_online_cpu(cpu_online_map));
6253 kthread_stop(cpu_rq(cpu)->migration_thread);
6254 cpu_rq(cpu)->migration_thread = NULL;
6258 case CPU_DEAD_FROZEN:
6259 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6260 migrate_live_tasks(cpu);
6262 kthread_stop(rq->migration_thread);
6263 rq->migration_thread = NULL;
6264 /* Idle task back to normal (off runqueue, low prio) */
6265 spin_lock_irq(&rq->lock);
6266 update_rq_clock(rq);
6267 deactivate_task(rq, rq->idle, 0);
6268 rq->idle->static_prio = MAX_PRIO;
6269 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6270 rq->idle->sched_class = &idle_sched_class;
6271 migrate_dead_tasks(cpu);
6272 spin_unlock_irq(&rq->lock);
6274 migrate_nr_uninterruptible(rq);
6275 BUG_ON(rq->nr_running != 0);
6278 * No need to migrate the tasks: it was best-effort if
6279 * they didn't take sched_hotcpu_mutex. Just wake up
6282 spin_lock_irq(&rq->lock);
6283 while (!list_empty(&rq->migration_queue)) {
6284 struct migration_req *req;
6286 req = list_entry(rq->migration_queue.next,
6287 struct migration_req, list);
6288 list_del_init(&req->list);
6289 complete(&req->done);
6291 spin_unlock_irq(&rq->lock);
6295 case CPU_DYING_FROZEN:
6296 /* Update our root-domain */
6298 spin_lock_irqsave(&rq->lock, flags);
6300 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6303 spin_unlock_irqrestore(&rq->lock, flags);
6310 /* Register at highest priority so that task migration (migrate_all_tasks)
6311 * happens before everything else.
6313 static struct notifier_block __cpuinitdata migration_notifier = {
6314 .notifier_call = migration_call,
6318 void __init migration_init(void)
6320 void *cpu = (void *)(long)smp_processor_id();
6323 /* Start one for the boot CPU: */
6324 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6325 BUG_ON(err == NOTIFY_BAD);
6326 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6327 register_cpu_notifier(&migration_notifier);
6333 #ifdef CONFIG_SCHED_DEBUG
6335 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6348 case SD_LV_ALLNODES:
6357 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6358 cpumask_t *groupmask)
6360 struct sched_group *group = sd->groups;
6363 cpulist_scnprintf(str, sizeof(str), sd->span);
6364 cpus_clear(*groupmask);
6366 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6368 if (!(sd->flags & SD_LOAD_BALANCE)) {
6369 printk("does not load-balance\n");
6371 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6376 printk(KERN_CONT "span %s level %s\n",
6377 str, sd_level_to_string(sd->level));
6379 if (!cpu_isset(cpu, sd->span)) {
6380 printk(KERN_ERR "ERROR: domain->span does not contain "
6383 if (!cpu_isset(cpu, group->cpumask)) {
6384 printk(KERN_ERR "ERROR: domain->groups does not contain"
6388 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6392 printk(KERN_ERR "ERROR: group is NULL\n");
6396 if (!group->__cpu_power) {
6397 printk(KERN_CONT "\n");
6398 printk(KERN_ERR "ERROR: domain->cpu_power not "
6403 if (!cpus_weight(group->cpumask)) {
6404 printk(KERN_CONT "\n");
6405 printk(KERN_ERR "ERROR: empty group\n");
6409 if (cpus_intersects(*groupmask, group->cpumask)) {
6410 printk(KERN_CONT "\n");
6411 printk(KERN_ERR "ERROR: repeated CPUs\n");
6415 cpus_or(*groupmask, *groupmask, group->cpumask);
6417 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6418 printk(KERN_CONT " %s", str);
6420 group = group->next;
6421 } while (group != sd->groups);
6422 printk(KERN_CONT "\n");
6424 if (!cpus_equal(sd->span, *groupmask))
6425 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6427 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6428 printk(KERN_ERR "ERROR: parent span is not a superset "
6429 "of domain->span\n");
6433 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6435 cpumask_t *groupmask;
6439 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6443 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6445 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6447 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6452 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6461 #else /* !CONFIG_SCHED_DEBUG */
6462 # define sched_domain_debug(sd, cpu) do { } while (0)
6463 #endif /* CONFIG_SCHED_DEBUG */
6465 static int sd_degenerate(struct sched_domain *sd)
6467 if (cpus_weight(sd->span) == 1)
6470 /* Following flags need at least 2 groups */
6471 if (sd->flags & (SD_LOAD_BALANCE |
6472 SD_BALANCE_NEWIDLE |
6476 SD_SHARE_PKG_RESOURCES)) {
6477 if (sd->groups != sd->groups->next)
6481 /* Following flags don't use groups */
6482 if (sd->flags & (SD_WAKE_IDLE |
6491 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6493 unsigned long cflags = sd->flags, pflags = parent->flags;
6495 if (sd_degenerate(parent))
6498 if (!cpus_equal(sd->span, parent->span))
6501 /* Does parent contain flags not in child? */
6502 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6503 if (cflags & SD_WAKE_AFFINE)
6504 pflags &= ~SD_WAKE_BALANCE;
6505 /* Flags needing groups don't count if only 1 group in parent */
6506 if (parent->groups == parent->groups->next) {
6507 pflags &= ~(SD_LOAD_BALANCE |
6508 SD_BALANCE_NEWIDLE |
6512 SD_SHARE_PKG_RESOURCES);
6514 if (~cflags & pflags)
6520 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6522 unsigned long flags;
6524 spin_lock_irqsave(&rq->lock, flags);
6527 struct root_domain *old_rd = rq->rd;
6529 if (cpu_isset(rq->cpu, old_rd->online))
6532 cpu_clear(rq->cpu, old_rd->span);
6534 if (atomic_dec_and_test(&old_rd->refcount))
6538 atomic_inc(&rd->refcount);
6541 cpu_set(rq->cpu, rd->span);
6542 if (cpu_isset(rq->cpu, cpu_online_map))
6545 spin_unlock_irqrestore(&rq->lock, flags);
6548 static void init_rootdomain(struct root_domain *rd)
6550 memset(rd, 0, sizeof(*rd));
6552 cpus_clear(rd->span);
6553 cpus_clear(rd->online);
6555 cpupri_init(&rd->cpupri);
6558 static void init_defrootdomain(void)
6560 init_rootdomain(&def_root_domain);
6561 atomic_set(&def_root_domain.refcount, 1);
6564 static struct root_domain *alloc_rootdomain(void)
6566 struct root_domain *rd;
6568 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6572 init_rootdomain(rd);
6578 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6579 * hold the hotplug lock.
6582 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6584 struct rq *rq = cpu_rq(cpu);
6585 struct sched_domain *tmp;
6587 /* Remove the sched domains which do not contribute to scheduling. */
6588 for (tmp = sd; tmp; tmp = tmp->parent) {
6589 struct sched_domain *parent = tmp->parent;
6592 if (sd_parent_degenerate(tmp, parent)) {
6593 tmp->parent = parent->parent;
6595 parent->parent->child = tmp;
6599 if (sd && sd_degenerate(sd)) {
6605 sched_domain_debug(sd, cpu);
6607 rq_attach_root(rq, rd);
6608 rcu_assign_pointer(rq->sd, sd);
6611 /* cpus with isolated domains */
6612 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6614 /* Setup the mask of cpus configured for isolated domains */
6615 static int __init isolated_cpu_setup(char *str)
6617 int ints[NR_CPUS], i;
6619 str = get_options(str, ARRAY_SIZE(ints), ints);
6620 cpus_clear(cpu_isolated_map);
6621 for (i = 1; i <= ints[0]; i++)
6622 if (ints[i] < NR_CPUS)
6623 cpu_set(ints[i], cpu_isolated_map);
6627 __setup("isolcpus=", isolated_cpu_setup);
6630 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6631 * to a function which identifies what group(along with sched group) a CPU
6632 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6633 * (due to the fact that we keep track of groups covered with a cpumask_t).
6635 * init_sched_build_groups will build a circular linked list of the groups
6636 * covered by the given span, and will set each group's ->cpumask correctly,
6637 * and ->cpu_power to 0.
6640 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6641 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6642 struct sched_group **sg,
6643 cpumask_t *tmpmask),
6644 cpumask_t *covered, cpumask_t *tmpmask)
6646 struct sched_group *first = NULL, *last = NULL;
6649 cpus_clear(*covered);
6651 for_each_cpu_mask(i, *span) {
6652 struct sched_group *sg;
6653 int group = group_fn(i, cpu_map, &sg, tmpmask);
6656 if (cpu_isset(i, *covered))
6659 cpus_clear(sg->cpumask);
6660 sg->__cpu_power = 0;
6662 for_each_cpu_mask(j, *span) {
6663 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6666 cpu_set(j, *covered);
6667 cpu_set(j, sg->cpumask);
6678 #define SD_NODES_PER_DOMAIN 16
6683 * find_next_best_node - find the next node to include in a sched_domain
6684 * @node: node whose sched_domain we're building
6685 * @used_nodes: nodes already in the sched_domain
6687 * Find the next node to include in a given scheduling domain. Simply
6688 * finds the closest node not already in the @used_nodes map.
6690 * Should use nodemask_t.
6692 static int find_next_best_node(int node, nodemask_t *used_nodes)
6694 int i, n, val, min_val, best_node = 0;
6698 for (i = 0; i < MAX_NUMNODES; i++) {
6699 /* Start at @node */
6700 n = (node + i) % MAX_NUMNODES;
6702 if (!nr_cpus_node(n))
6705 /* Skip already used nodes */
6706 if (node_isset(n, *used_nodes))
6709 /* Simple min distance search */
6710 val = node_distance(node, n);
6712 if (val < min_val) {
6718 node_set(best_node, *used_nodes);
6723 * sched_domain_node_span - get a cpumask for a node's sched_domain
6724 * @node: node whose cpumask we're constructing
6725 * @span: resulting cpumask
6727 * Given a node, construct a good cpumask for its sched_domain to span. It
6728 * should be one that prevents unnecessary balancing, but also spreads tasks
6731 static void sched_domain_node_span(int node, cpumask_t *span)
6733 nodemask_t used_nodes;
6734 node_to_cpumask_ptr(nodemask, node);
6738 nodes_clear(used_nodes);
6740 cpus_or(*span, *span, *nodemask);
6741 node_set(node, used_nodes);
6743 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6744 int next_node = find_next_best_node(node, &used_nodes);
6746 node_to_cpumask_ptr_next(nodemask, next_node);
6747 cpus_or(*span, *span, *nodemask);
6750 #endif /* CONFIG_NUMA */
6752 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6755 * SMT sched-domains:
6757 #ifdef CONFIG_SCHED_SMT
6758 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6759 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6762 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6766 *sg = &per_cpu(sched_group_cpus, cpu);
6769 #endif /* CONFIG_SCHED_SMT */
6772 * multi-core sched-domains:
6774 #ifdef CONFIG_SCHED_MC
6775 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6776 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6777 #endif /* CONFIG_SCHED_MC */
6779 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6781 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6786 *mask = per_cpu(cpu_sibling_map, cpu);
6787 cpus_and(*mask, *mask, *cpu_map);
6788 group = first_cpu(*mask);
6790 *sg = &per_cpu(sched_group_core, group);
6793 #elif defined(CONFIG_SCHED_MC)
6795 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6799 *sg = &per_cpu(sched_group_core, cpu);
6804 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6805 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6808 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6812 #ifdef CONFIG_SCHED_MC
6813 *mask = cpu_coregroup_map(cpu);
6814 cpus_and(*mask, *mask, *cpu_map);
6815 group = first_cpu(*mask);
6816 #elif defined(CONFIG_SCHED_SMT)
6817 *mask = per_cpu(cpu_sibling_map, cpu);
6818 cpus_and(*mask, *mask, *cpu_map);
6819 group = first_cpu(*mask);
6824 *sg = &per_cpu(sched_group_phys, group);
6830 * The init_sched_build_groups can't handle what we want to do with node
6831 * groups, so roll our own. Now each node has its own list of groups which
6832 * gets dynamically allocated.
6834 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6835 static struct sched_group ***sched_group_nodes_bycpu;
6837 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6838 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6840 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6841 struct sched_group **sg, cpumask_t *nodemask)
6845 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6846 cpus_and(*nodemask, *nodemask, *cpu_map);
6847 group = first_cpu(*nodemask);
6850 *sg = &per_cpu(sched_group_allnodes, group);
6854 static void init_numa_sched_groups_power(struct sched_group *group_head)
6856 struct sched_group *sg = group_head;
6862 for_each_cpu_mask(j, sg->cpumask) {
6863 struct sched_domain *sd;
6865 sd = &per_cpu(phys_domains, j);
6866 if (j != first_cpu(sd->groups->cpumask)) {
6868 * Only add "power" once for each
6874 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6877 } while (sg != group_head);
6879 #endif /* CONFIG_NUMA */
6882 /* Free memory allocated for various sched_group structures */
6883 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6887 for_each_cpu_mask(cpu, *cpu_map) {
6888 struct sched_group **sched_group_nodes
6889 = sched_group_nodes_bycpu[cpu];
6891 if (!sched_group_nodes)
6894 for (i = 0; i < MAX_NUMNODES; i++) {
6895 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6897 *nodemask = node_to_cpumask(i);
6898 cpus_and(*nodemask, *nodemask, *cpu_map);
6899 if (cpus_empty(*nodemask))
6909 if (oldsg != sched_group_nodes[i])
6912 kfree(sched_group_nodes);
6913 sched_group_nodes_bycpu[cpu] = NULL;
6916 #else /* !CONFIG_NUMA */
6917 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6920 #endif /* CONFIG_NUMA */
6923 * Initialize sched groups cpu_power.
6925 * cpu_power indicates the capacity of sched group, which is used while
6926 * distributing the load between different sched groups in a sched domain.
6927 * Typically cpu_power for all the groups in a sched domain will be same unless
6928 * there are asymmetries in the topology. If there are asymmetries, group
6929 * having more cpu_power will pickup more load compared to the group having
6932 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6933 * the maximum number of tasks a group can handle in the presence of other idle
6934 * or lightly loaded groups in the same sched domain.
6936 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6938 struct sched_domain *child;
6939 struct sched_group *group;
6941 WARN_ON(!sd || !sd->groups);
6943 if (cpu != first_cpu(sd->groups->cpumask))
6948 sd->groups->__cpu_power = 0;
6951 * For perf policy, if the groups in child domain share resources
6952 * (for example cores sharing some portions of the cache hierarchy
6953 * or SMT), then set this domain groups cpu_power such that each group
6954 * can handle only one task, when there are other idle groups in the
6955 * same sched domain.
6957 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6959 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6960 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6965 * add cpu_power of each child group to this groups cpu_power
6967 group = child->groups;
6969 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6970 group = group->next;
6971 } while (group != child->groups);
6975 * Initializers for schedule domains
6976 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6979 #define SD_INIT(sd, type) sd_init_##type(sd)
6980 #define SD_INIT_FUNC(type) \
6981 static noinline void sd_init_##type(struct sched_domain *sd) \
6983 memset(sd, 0, sizeof(*sd)); \
6984 *sd = SD_##type##_INIT; \
6985 sd->level = SD_LV_##type; \
6990 SD_INIT_FUNC(ALLNODES)
6993 #ifdef CONFIG_SCHED_SMT
6994 SD_INIT_FUNC(SIBLING)
6996 #ifdef CONFIG_SCHED_MC
7001 * To minimize stack usage kmalloc room for cpumasks and share the
7002 * space as the usage in build_sched_domains() dictates. Used only
7003 * if the amount of space is significant.
7006 cpumask_t tmpmask; /* make this one first */
7009 cpumask_t this_sibling_map;
7010 cpumask_t this_core_map;
7012 cpumask_t send_covered;
7015 cpumask_t domainspan;
7017 cpumask_t notcovered;
7022 #define SCHED_CPUMASK_ALLOC 1
7023 #define SCHED_CPUMASK_FREE(v) kfree(v)
7024 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7026 #define SCHED_CPUMASK_ALLOC 0
7027 #define SCHED_CPUMASK_FREE(v)
7028 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7031 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7032 ((unsigned long)(a) + offsetof(struct allmasks, v))
7034 static int default_relax_domain_level = -1;
7036 static int __init setup_relax_domain_level(char *str)
7040 val = simple_strtoul(str, NULL, 0);
7041 if (val < SD_LV_MAX)
7042 default_relax_domain_level = val;
7046 __setup("relax_domain_level=", setup_relax_domain_level);
7048 static void set_domain_attribute(struct sched_domain *sd,
7049 struct sched_domain_attr *attr)
7053 if (!attr || attr->relax_domain_level < 0) {
7054 if (default_relax_domain_level < 0)
7057 request = default_relax_domain_level;
7059 request = attr->relax_domain_level;
7060 if (request < sd->level) {
7061 /* turn off idle balance on this domain */
7062 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7064 /* turn on idle balance on this domain */
7065 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7070 * Build sched domains for a given set of cpus and attach the sched domains
7071 * to the individual cpus
7073 static int __build_sched_domains(const cpumask_t *cpu_map,
7074 struct sched_domain_attr *attr)
7077 struct root_domain *rd;
7078 SCHED_CPUMASK_DECLARE(allmasks);
7081 struct sched_group **sched_group_nodes = NULL;
7082 int sd_allnodes = 0;
7085 * Allocate the per-node list of sched groups
7087 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7089 if (!sched_group_nodes) {
7090 printk(KERN_WARNING "Can not alloc sched group node list\n");
7095 rd = alloc_rootdomain();
7097 printk(KERN_WARNING "Cannot alloc root domain\n");
7099 kfree(sched_group_nodes);
7104 #if SCHED_CPUMASK_ALLOC
7105 /* get space for all scratch cpumask variables */
7106 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7108 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7111 kfree(sched_group_nodes);
7116 tmpmask = (cpumask_t *)allmasks;
7120 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7124 * Set up domains for cpus specified by the cpu_map.
7126 for_each_cpu_mask(i, *cpu_map) {
7127 struct sched_domain *sd = NULL, *p;
7128 SCHED_CPUMASK_VAR(nodemask, allmasks);
7130 *nodemask = node_to_cpumask(cpu_to_node(i));
7131 cpus_and(*nodemask, *nodemask, *cpu_map);
7134 if (cpus_weight(*cpu_map) >
7135 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7136 sd = &per_cpu(allnodes_domains, i);
7137 SD_INIT(sd, ALLNODES);
7138 set_domain_attribute(sd, attr);
7139 sd->span = *cpu_map;
7140 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7146 sd = &per_cpu(node_domains, i);
7148 set_domain_attribute(sd, attr);
7149 sched_domain_node_span(cpu_to_node(i), &sd->span);
7153 cpus_and(sd->span, sd->span, *cpu_map);
7157 sd = &per_cpu(phys_domains, i);
7159 set_domain_attribute(sd, attr);
7160 sd->span = *nodemask;
7164 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7166 #ifdef CONFIG_SCHED_MC
7168 sd = &per_cpu(core_domains, i);
7170 set_domain_attribute(sd, attr);
7171 sd->span = cpu_coregroup_map(i);
7172 cpus_and(sd->span, sd->span, *cpu_map);
7175 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7178 #ifdef CONFIG_SCHED_SMT
7180 sd = &per_cpu(cpu_domains, i);
7181 SD_INIT(sd, SIBLING);
7182 set_domain_attribute(sd, attr);
7183 sd->span = per_cpu(cpu_sibling_map, i);
7184 cpus_and(sd->span, sd->span, *cpu_map);
7187 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7191 #ifdef CONFIG_SCHED_SMT
7192 /* Set up CPU (sibling) groups */
7193 for_each_cpu_mask(i, *cpu_map) {
7194 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7195 SCHED_CPUMASK_VAR(send_covered, allmasks);
7197 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7198 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7199 if (i != first_cpu(*this_sibling_map))
7202 init_sched_build_groups(this_sibling_map, cpu_map,
7204 send_covered, tmpmask);
7208 #ifdef CONFIG_SCHED_MC
7209 /* Set up multi-core groups */
7210 for_each_cpu_mask(i, *cpu_map) {
7211 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7212 SCHED_CPUMASK_VAR(send_covered, allmasks);
7214 *this_core_map = cpu_coregroup_map(i);
7215 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7216 if (i != first_cpu(*this_core_map))
7219 init_sched_build_groups(this_core_map, cpu_map,
7221 send_covered, tmpmask);
7225 /* Set up physical groups */
7226 for (i = 0; i < MAX_NUMNODES; i++) {
7227 SCHED_CPUMASK_VAR(nodemask, allmasks);
7228 SCHED_CPUMASK_VAR(send_covered, allmasks);
7230 *nodemask = node_to_cpumask(i);
7231 cpus_and(*nodemask, *nodemask, *cpu_map);
7232 if (cpus_empty(*nodemask))
7235 init_sched_build_groups(nodemask, cpu_map,
7237 send_covered, tmpmask);
7241 /* Set up node groups */
7243 SCHED_CPUMASK_VAR(send_covered, allmasks);
7245 init_sched_build_groups(cpu_map, cpu_map,
7246 &cpu_to_allnodes_group,
7247 send_covered, tmpmask);
7250 for (i = 0; i < MAX_NUMNODES; i++) {
7251 /* Set up node groups */
7252 struct sched_group *sg, *prev;
7253 SCHED_CPUMASK_VAR(nodemask, allmasks);
7254 SCHED_CPUMASK_VAR(domainspan, allmasks);
7255 SCHED_CPUMASK_VAR(covered, allmasks);
7258 *nodemask = node_to_cpumask(i);
7259 cpus_clear(*covered);
7261 cpus_and(*nodemask, *nodemask, *cpu_map);
7262 if (cpus_empty(*nodemask)) {
7263 sched_group_nodes[i] = NULL;
7267 sched_domain_node_span(i, domainspan);
7268 cpus_and(*domainspan, *domainspan, *cpu_map);
7270 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7272 printk(KERN_WARNING "Can not alloc domain group for "
7276 sched_group_nodes[i] = sg;
7277 for_each_cpu_mask(j, *nodemask) {
7278 struct sched_domain *sd;
7280 sd = &per_cpu(node_domains, j);
7283 sg->__cpu_power = 0;
7284 sg->cpumask = *nodemask;
7286 cpus_or(*covered, *covered, *nodemask);
7289 for (j = 0; j < MAX_NUMNODES; j++) {
7290 SCHED_CPUMASK_VAR(notcovered, allmasks);
7291 int n = (i + j) % MAX_NUMNODES;
7292 node_to_cpumask_ptr(pnodemask, n);
7294 cpus_complement(*notcovered, *covered);
7295 cpus_and(*tmpmask, *notcovered, *cpu_map);
7296 cpus_and(*tmpmask, *tmpmask, *domainspan);
7297 if (cpus_empty(*tmpmask))
7300 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7301 if (cpus_empty(*tmpmask))
7304 sg = kmalloc_node(sizeof(struct sched_group),
7308 "Can not alloc domain group for node %d\n", j);
7311 sg->__cpu_power = 0;
7312 sg->cpumask = *tmpmask;
7313 sg->next = prev->next;
7314 cpus_or(*covered, *covered, *tmpmask);
7321 /* Calculate CPU power for physical packages and nodes */
7322 #ifdef CONFIG_SCHED_SMT
7323 for_each_cpu_mask(i, *cpu_map) {
7324 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7326 init_sched_groups_power(i, sd);
7329 #ifdef CONFIG_SCHED_MC
7330 for_each_cpu_mask(i, *cpu_map) {
7331 struct sched_domain *sd = &per_cpu(core_domains, i);
7333 init_sched_groups_power(i, sd);
7337 for_each_cpu_mask(i, *cpu_map) {
7338 struct sched_domain *sd = &per_cpu(phys_domains, i);
7340 init_sched_groups_power(i, sd);
7344 for (i = 0; i < MAX_NUMNODES; i++)
7345 init_numa_sched_groups_power(sched_group_nodes[i]);
7348 struct sched_group *sg;
7350 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7352 init_numa_sched_groups_power(sg);
7356 /* Attach the domains */
7357 for_each_cpu_mask(i, *cpu_map) {
7358 struct sched_domain *sd;
7359 #ifdef CONFIG_SCHED_SMT
7360 sd = &per_cpu(cpu_domains, i);
7361 #elif defined(CONFIG_SCHED_MC)
7362 sd = &per_cpu(core_domains, i);
7364 sd = &per_cpu(phys_domains, i);
7366 cpu_attach_domain(sd, rd, i);
7369 SCHED_CPUMASK_FREE((void *)allmasks);
7374 free_sched_groups(cpu_map, tmpmask);
7375 SCHED_CPUMASK_FREE((void *)allmasks);
7380 static int build_sched_domains(const cpumask_t *cpu_map)
7382 return __build_sched_domains(cpu_map, NULL);
7385 static cpumask_t *doms_cur; /* current sched domains */
7386 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7387 static struct sched_domain_attr *dattr_cur;
7388 /* attribues of custom domains in 'doms_cur' */
7391 * Special case: If a kmalloc of a doms_cur partition (array of
7392 * cpumask_t) fails, then fallback to a single sched domain,
7393 * as determined by the single cpumask_t fallback_doms.
7395 static cpumask_t fallback_doms;
7397 void __attribute__((weak)) arch_update_cpu_topology(void)
7402 * Free current domain masks.
7403 * Called after all cpus are attached to NULL domain.
7405 static void free_sched_domains(void)
7408 if (doms_cur != &fallback_doms)
7410 doms_cur = &fallback_doms;
7414 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7415 * For now this just excludes isolated cpus, but could be used to
7416 * exclude other special cases in the future.
7418 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7422 arch_update_cpu_topology();
7424 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7426 doms_cur = &fallback_doms;
7427 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7429 err = build_sched_domains(doms_cur);
7430 register_sched_domain_sysctl();
7435 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7438 free_sched_groups(cpu_map, tmpmask);
7442 * Detach sched domains from a group of cpus specified in cpu_map
7443 * These cpus will now be attached to the NULL domain
7445 static void detach_destroy_domains(const cpumask_t *cpu_map)
7450 unregister_sched_domain_sysctl();
7452 for_each_cpu_mask(i, *cpu_map)
7453 cpu_attach_domain(NULL, &def_root_domain, i);
7454 synchronize_sched();
7455 arch_destroy_sched_domains(cpu_map, &tmpmask);
7458 /* handle null as "default" */
7459 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7460 struct sched_domain_attr *new, int idx_new)
7462 struct sched_domain_attr tmp;
7469 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7470 new ? (new + idx_new) : &tmp,
7471 sizeof(struct sched_domain_attr));
7475 * Partition sched domains as specified by the 'ndoms_new'
7476 * cpumasks in the array doms_new[] of cpumasks. This compares
7477 * doms_new[] to the current sched domain partitioning, doms_cur[].
7478 * It destroys each deleted domain and builds each new domain.
7480 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7481 * The masks don't intersect (don't overlap.) We should setup one
7482 * sched domain for each mask. CPUs not in any of the cpumasks will
7483 * not be load balanced. If the same cpumask appears both in the
7484 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7487 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7488 * ownership of it and will kfree it when done with it. If the caller
7489 * failed the kmalloc call, then it can pass in doms_new == NULL,
7490 * and partition_sched_domains() will fallback to the single partition
7493 * Call with hotplug lock held
7495 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7496 struct sched_domain_attr *dattr_new)
7500 mutex_lock(&sched_domains_mutex);
7502 /* always unregister in case we don't destroy any domains */
7503 unregister_sched_domain_sysctl();
7505 if (doms_new == NULL) {
7507 doms_new = &fallback_doms;
7508 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7512 /* Destroy deleted domains */
7513 for (i = 0; i < ndoms_cur; i++) {
7514 for (j = 0; j < ndoms_new; j++) {
7515 if (cpus_equal(doms_cur[i], doms_new[j])
7516 && dattrs_equal(dattr_cur, i, dattr_new, j))
7519 /* no match - a current sched domain not in new doms_new[] */
7520 detach_destroy_domains(doms_cur + i);
7525 /* Build new domains */
7526 for (i = 0; i < ndoms_new; i++) {
7527 for (j = 0; j < ndoms_cur; j++) {
7528 if (cpus_equal(doms_new[i], doms_cur[j])
7529 && dattrs_equal(dattr_new, i, dattr_cur, j))
7532 /* no match - add a new doms_new */
7533 __build_sched_domains(doms_new + i,
7534 dattr_new ? dattr_new + i : NULL);
7539 /* Remember the new sched domains */
7540 if (doms_cur != &fallback_doms)
7542 kfree(dattr_cur); /* kfree(NULL) is safe */
7543 doms_cur = doms_new;
7544 dattr_cur = dattr_new;
7545 ndoms_cur = ndoms_new;
7547 register_sched_domain_sysctl();
7549 mutex_unlock(&sched_domains_mutex);
7552 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7553 int arch_reinit_sched_domains(void)
7558 mutex_lock(&sched_domains_mutex);
7559 detach_destroy_domains(&cpu_online_map);
7560 free_sched_domains();
7561 err = arch_init_sched_domains(&cpu_online_map);
7562 mutex_unlock(&sched_domains_mutex);
7568 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7572 if (buf[0] != '0' && buf[0] != '1')
7576 sched_smt_power_savings = (buf[0] == '1');
7578 sched_mc_power_savings = (buf[0] == '1');
7580 ret = arch_reinit_sched_domains();
7582 return ret ? ret : count;
7585 #ifdef CONFIG_SCHED_MC
7586 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7588 return sprintf(page, "%u\n", sched_mc_power_savings);
7590 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7591 const char *buf, size_t count)
7593 return sched_power_savings_store(buf, count, 0);
7595 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7596 sched_mc_power_savings_store);
7599 #ifdef CONFIG_SCHED_SMT
7600 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7602 return sprintf(page, "%u\n", sched_smt_power_savings);
7604 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7605 const char *buf, size_t count)
7607 return sched_power_savings_store(buf, count, 1);
7609 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7610 sched_smt_power_savings_store);
7613 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7617 #ifdef CONFIG_SCHED_SMT
7619 err = sysfs_create_file(&cls->kset.kobj,
7620 &attr_sched_smt_power_savings.attr);
7622 #ifdef CONFIG_SCHED_MC
7623 if (!err && mc_capable())
7624 err = sysfs_create_file(&cls->kset.kobj,
7625 &attr_sched_mc_power_savings.attr);
7629 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7632 * Force a reinitialization of the sched domains hierarchy. The domains
7633 * and groups cannot be updated in place without racing with the balancing
7634 * code, so we temporarily attach all running cpus to the NULL domain
7635 * which will prevent rebalancing while the sched domains are recalculated.
7637 static int update_sched_domains(struct notifier_block *nfb,
7638 unsigned long action, void *hcpu)
7640 int cpu = (int)(long)hcpu;
7643 case CPU_DOWN_PREPARE:
7644 case CPU_DOWN_PREPARE_FROZEN:
7645 disable_runtime(cpu_rq(cpu));
7647 case CPU_UP_PREPARE:
7648 case CPU_UP_PREPARE_FROZEN:
7649 detach_destroy_domains(&cpu_online_map);
7650 free_sched_domains();
7654 case CPU_DOWN_FAILED:
7655 case CPU_DOWN_FAILED_FROZEN:
7657 case CPU_ONLINE_FROZEN:
7658 enable_runtime(cpu_rq(cpu));
7660 case CPU_UP_CANCELED:
7661 case CPU_UP_CANCELED_FROZEN:
7663 case CPU_DEAD_FROZEN:
7665 * Fall through and re-initialise the domains.
7672 #ifndef CONFIG_CPUSETS
7674 * Create default domain partitioning if cpusets are disabled.
7675 * Otherwise we let cpusets rebuild the domains based on the
7679 /* The hotplug lock is already held by cpu_up/cpu_down */
7680 arch_init_sched_domains(&cpu_online_map);
7686 void __init sched_init_smp(void)
7688 cpumask_t non_isolated_cpus;
7690 #if defined(CONFIG_NUMA)
7691 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7693 BUG_ON(sched_group_nodes_bycpu == NULL);
7696 mutex_lock(&sched_domains_mutex);
7697 arch_init_sched_domains(&cpu_online_map);
7698 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7699 if (cpus_empty(non_isolated_cpus))
7700 cpu_set(smp_processor_id(), non_isolated_cpus);
7701 mutex_unlock(&sched_domains_mutex);
7703 /* XXX: Theoretical race here - CPU may be hotplugged now */
7704 hotcpu_notifier(update_sched_domains, 0);
7707 /* Move init over to a non-isolated CPU */
7708 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7710 sched_init_granularity();
7713 void __init sched_init_smp(void)
7715 sched_init_granularity();
7717 #endif /* CONFIG_SMP */
7719 int in_sched_functions(unsigned long addr)
7721 return in_lock_functions(addr) ||
7722 (addr >= (unsigned long)__sched_text_start
7723 && addr < (unsigned long)__sched_text_end);
7726 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7728 cfs_rq->tasks_timeline = RB_ROOT;
7729 INIT_LIST_HEAD(&cfs_rq->tasks);
7730 #ifdef CONFIG_FAIR_GROUP_SCHED
7733 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7736 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7738 struct rt_prio_array *array;
7741 array = &rt_rq->active;
7742 for (i = 0; i < MAX_RT_PRIO; i++) {
7743 INIT_LIST_HEAD(array->queue + i);
7744 __clear_bit(i, array->bitmap);
7746 /* delimiter for bitsearch: */
7747 __set_bit(MAX_RT_PRIO, array->bitmap);
7749 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7750 rt_rq->highest_prio = MAX_RT_PRIO;
7753 rt_rq->rt_nr_migratory = 0;
7754 rt_rq->overloaded = 0;
7758 rt_rq->rt_throttled = 0;
7759 rt_rq->rt_runtime = 0;
7760 spin_lock_init(&rt_rq->rt_runtime_lock);
7762 #ifdef CONFIG_RT_GROUP_SCHED
7763 rt_rq->rt_nr_boosted = 0;
7768 #ifdef CONFIG_FAIR_GROUP_SCHED
7769 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7770 struct sched_entity *se, int cpu, int add,
7771 struct sched_entity *parent)
7773 struct rq *rq = cpu_rq(cpu);
7774 tg->cfs_rq[cpu] = cfs_rq;
7775 init_cfs_rq(cfs_rq, rq);
7778 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7781 /* se could be NULL for init_task_group */
7786 se->cfs_rq = &rq->cfs;
7788 se->cfs_rq = parent->my_q;
7791 se->load.weight = tg->shares;
7792 se->load.inv_weight = 0;
7793 se->parent = parent;
7797 #ifdef CONFIG_RT_GROUP_SCHED
7798 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7799 struct sched_rt_entity *rt_se, int cpu, int add,
7800 struct sched_rt_entity *parent)
7802 struct rq *rq = cpu_rq(cpu);
7804 tg->rt_rq[cpu] = rt_rq;
7805 init_rt_rq(rt_rq, rq);
7807 rt_rq->rt_se = rt_se;
7808 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7810 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7812 tg->rt_se[cpu] = rt_se;
7817 rt_se->rt_rq = &rq->rt;
7819 rt_se->rt_rq = parent->my_q;
7821 rt_se->my_q = rt_rq;
7822 rt_se->parent = parent;
7823 INIT_LIST_HEAD(&rt_se->run_list);
7827 void __init sched_init(void)
7830 unsigned long alloc_size = 0, ptr;
7832 #ifdef CONFIG_FAIR_GROUP_SCHED
7833 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7835 #ifdef CONFIG_RT_GROUP_SCHED
7836 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7838 #ifdef CONFIG_USER_SCHED
7842 * As sched_init() is called before page_alloc is setup,
7843 * we use alloc_bootmem().
7846 ptr = (unsigned long)alloc_bootmem(alloc_size);
7848 #ifdef CONFIG_FAIR_GROUP_SCHED
7849 init_task_group.se = (struct sched_entity **)ptr;
7850 ptr += nr_cpu_ids * sizeof(void **);
7852 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7853 ptr += nr_cpu_ids * sizeof(void **);
7855 #ifdef CONFIG_USER_SCHED
7856 root_task_group.se = (struct sched_entity **)ptr;
7857 ptr += nr_cpu_ids * sizeof(void **);
7859 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7860 ptr += nr_cpu_ids * sizeof(void **);
7861 #endif /* CONFIG_USER_SCHED */
7862 #endif /* CONFIG_FAIR_GROUP_SCHED */
7863 #ifdef CONFIG_RT_GROUP_SCHED
7864 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7865 ptr += nr_cpu_ids * sizeof(void **);
7867 init_task_group.rt_rq = (struct rt_rq **)ptr;
7868 ptr += nr_cpu_ids * sizeof(void **);
7870 #ifdef CONFIG_USER_SCHED
7871 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7872 ptr += nr_cpu_ids * sizeof(void **);
7874 root_task_group.rt_rq = (struct rt_rq **)ptr;
7875 ptr += nr_cpu_ids * sizeof(void **);
7876 #endif /* CONFIG_USER_SCHED */
7877 #endif /* CONFIG_RT_GROUP_SCHED */
7881 init_defrootdomain();
7884 init_rt_bandwidth(&def_rt_bandwidth,
7885 global_rt_period(), global_rt_runtime());
7887 #ifdef CONFIG_RT_GROUP_SCHED
7888 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7889 global_rt_period(), global_rt_runtime());
7890 #ifdef CONFIG_USER_SCHED
7891 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7892 global_rt_period(), RUNTIME_INF);
7893 #endif /* CONFIG_USER_SCHED */
7894 #endif /* CONFIG_RT_GROUP_SCHED */
7896 #ifdef CONFIG_GROUP_SCHED
7897 list_add(&init_task_group.list, &task_groups);
7898 INIT_LIST_HEAD(&init_task_group.children);
7900 #ifdef CONFIG_USER_SCHED
7901 INIT_LIST_HEAD(&root_task_group.children);
7902 init_task_group.parent = &root_task_group;
7903 list_add(&init_task_group.siblings, &root_task_group.children);
7904 #endif /* CONFIG_USER_SCHED */
7905 #endif /* CONFIG_GROUP_SCHED */
7907 for_each_possible_cpu(i) {
7911 spin_lock_init(&rq->lock);
7912 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7914 init_cfs_rq(&rq->cfs, rq);
7915 init_rt_rq(&rq->rt, rq);
7916 #ifdef CONFIG_FAIR_GROUP_SCHED
7917 init_task_group.shares = init_task_group_load;
7918 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7919 #ifdef CONFIG_CGROUP_SCHED
7921 * How much cpu bandwidth does init_task_group get?
7923 * In case of task-groups formed thr' the cgroup filesystem, it
7924 * gets 100% of the cpu resources in the system. This overall
7925 * system cpu resource is divided among the tasks of
7926 * init_task_group and its child task-groups in a fair manner,
7927 * based on each entity's (task or task-group's) weight
7928 * (se->load.weight).
7930 * In other words, if init_task_group has 10 tasks of weight
7931 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7932 * then A0's share of the cpu resource is:
7934 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7936 * We achieve this by letting init_task_group's tasks sit
7937 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7939 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7940 #elif defined CONFIG_USER_SCHED
7941 root_task_group.shares = NICE_0_LOAD;
7942 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
7944 * In case of task-groups formed thr' the user id of tasks,
7945 * init_task_group represents tasks belonging to root user.
7946 * Hence it forms a sibling of all subsequent groups formed.
7947 * In this case, init_task_group gets only a fraction of overall
7948 * system cpu resource, based on the weight assigned to root
7949 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7950 * by letting tasks of init_task_group sit in a separate cfs_rq
7951 * (init_cfs_rq) and having one entity represent this group of
7952 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7954 init_tg_cfs_entry(&init_task_group,
7955 &per_cpu(init_cfs_rq, i),
7956 &per_cpu(init_sched_entity, i), i, 1,
7957 root_task_group.se[i]);
7960 #endif /* CONFIG_FAIR_GROUP_SCHED */
7962 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7963 #ifdef CONFIG_RT_GROUP_SCHED
7964 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7965 #ifdef CONFIG_CGROUP_SCHED
7966 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7967 #elif defined CONFIG_USER_SCHED
7968 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
7969 init_tg_rt_entry(&init_task_group,
7970 &per_cpu(init_rt_rq, i),
7971 &per_cpu(init_sched_rt_entity, i), i, 1,
7972 root_task_group.rt_se[i]);
7976 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7977 rq->cpu_load[j] = 0;
7981 rq->active_balance = 0;
7982 rq->next_balance = jiffies;
7986 rq->migration_thread = NULL;
7987 INIT_LIST_HEAD(&rq->migration_queue);
7988 rq_attach_root(rq, &def_root_domain);
7991 atomic_set(&rq->nr_iowait, 0);
7994 set_load_weight(&init_task);
7996 #ifdef CONFIG_PREEMPT_NOTIFIERS
7997 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8001 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8004 #ifdef CONFIG_RT_MUTEXES
8005 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8009 * The boot idle thread does lazy MMU switching as well:
8011 atomic_inc(&init_mm.mm_count);
8012 enter_lazy_tlb(&init_mm, current);
8015 * Make us the idle thread. Technically, schedule() should not be
8016 * called from this thread, however somewhere below it might be,
8017 * but because we are the idle thread, we just pick up running again
8018 * when this runqueue becomes "idle".
8020 init_idle(current, smp_processor_id());
8022 * During early bootup we pretend to be a normal task:
8024 current->sched_class = &fair_sched_class;
8026 scheduler_running = 1;
8029 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8030 void __might_sleep(char *file, int line)
8033 static unsigned long prev_jiffy; /* ratelimiting */
8035 if ((in_atomic() || irqs_disabled()) &&
8036 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8037 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8039 prev_jiffy = jiffies;
8040 printk(KERN_ERR "BUG: sleeping function called from invalid"
8041 " context at %s:%d\n", file, line);
8042 printk("in_atomic():%d, irqs_disabled():%d\n",
8043 in_atomic(), irqs_disabled());
8044 debug_show_held_locks(current);
8045 if (irqs_disabled())
8046 print_irqtrace_events(current);
8051 EXPORT_SYMBOL(__might_sleep);
8054 #ifdef CONFIG_MAGIC_SYSRQ
8055 static void normalize_task(struct rq *rq, struct task_struct *p)
8059 update_rq_clock(rq);
8060 on_rq = p->se.on_rq;
8062 deactivate_task(rq, p, 0);
8063 __setscheduler(rq, p, SCHED_NORMAL, 0);
8065 activate_task(rq, p, 0);
8066 resched_task(rq->curr);
8070 void normalize_rt_tasks(void)
8072 struct task_struct *g, *p;
8073 unsigned long flags;
8076 read_lock_irqsave(&tasklist_lock, flags);
8077 do_each_thread(g, p) {
8079 * Only normalize user tasks:
8084 p->se.exec_start = 0;
8085 #ifdef CONFIG_SCHEDSTATS
8086 p->se.wait_start = 0;
8087 p->se.sleep_start = 0;
8088 p->se.block_start = 0;
8093 * Renice negative nice level userspace
8096 if (TASK_NICE(p) < 0 && p->mm)
8097 set_user_nice(p, 0);
8101 spin_lock(&p->pi_lock);
8102 rq = __task_rq_lock(p);
8104 normalize_task(rq, p);
8106 __task_rq_unlock(rq);
8107 spin_unlock(&p->pi_lock);
8108 } while_each_thread(g, p);
8110 read_unlock_irqrestore(&tasklist_lock, flags);
8113 #endif /* CONFIG_MAGIC_SYSRQ */
8117 * These functions are only useful for the IA64 MCA handling.
8119 * They can only be called when the whole system has been
8120 * stopped - every CPU needs to be quiescent, and no scheduling
8121 * activity can take place. Using them for anything else would
8122 * be a serious bug, and as a result, they aren't even visible
8123 * under any other configuration.
8127 * curr_task - return the current task for a given cpu.
8128 * @cpu: the processor in question.
8130 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8132 struct task_struct *curr_task(int cpu)
8134 return cpu_curr(cpu);
8138 * set_curr_task - set the current task for a given cpu.
8139 * @cpu: the processor in question.
8140 * @p: the task pointer to set.
8142 * Description: This function must only be used when non-maskable interrupts
8143 * are serviced on a separate stack. It allows the architecture to switch the
8144 * notion of the current task on a cpu in a non-blocking manner. This function
8145 * must be called with all CPU's synchronized, and interrupts disabled, the
8146 * and caller must save the original value of the current task (see
8147 * curr_task() above) and restore that value before reenabling interrupts and
8148 * re-starting the system.
8150 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8152 void set_curr_task(int cpu, struct task_struct *p)
8159 #ifdef CONFIG_FAIR_GROUP_SCHED
8160 static void free_fair_sched_group(struct task_group *tg)
8164 for_each_possible_cpu(i) {
8166 kfree(tg->cfs_rq[i]);
8176 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8178 struct cfs_rq *cfs_rq;
8179 struct sched_entity *se, *parent_se;
8183 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8186 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8190 tg->shares = NICE_0_LOAD;
8192 for_each_possible_cpu(i) {
8195 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8196 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8200 se = kmalloc_node(sizeof(struct sched_entity),
8201 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8205 parent_se = parent ? parent->se[i] : NULL;
8206 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8215 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8217 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8218 &cpu_rq(cpu)->leaf_cfs_rq_list);
8221 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8223 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8225 #else /* !CONFG_FAIR_GROUP_SCHED */
8226 static inline void free_fair_sched_group(struct task_group *tg)
8231 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8236 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8240 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8243 #endif /* CONFIG_FAIR_GROUP_SCHED */
8245 #ifdef CONFIG_RT_GROUP_SCHED
8246 static void free_rt_sched_group(struct task_group *tg)
8250 destroy_rt_bandwidth(&tg->rt_bandwidth);
8252 for_each_possible_cpu(i) {
8254 kfree(tg->rt_rq[i]);
8256 kfree(tg->rt_se[i]);
8264 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8266 struct rt_rq *rt_rq;
8267 struct sched_rt_entity *rt_se, *parent_se;
8271 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8274 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8278 init_rt_bandwidth(&tg->rt_bandwidth,
8279 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8281 for_each_possible_cpu(i) {
8284 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8285 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8289 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8290 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8294 parent_se = parent ? parent->rt_se[i] : NULL;
8295 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8304 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8306 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8307 &cpu_rq(cpu)->leaf_rt_rq_list);
8310 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8312 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8314 #else /* !CONFIG_RT_GROUP_SCHED */
8315 static inline void free_rt_sched_group(struct task_group *tg)
8320 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8325 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8329 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8332 #endif /* CONFIG_RT_GROUP_SCHED */
8334 #ifdef CONFIG_GROUP_SCHED
8335 static void free_sched_group(struct task_group *tg)
8337 free_fair_sched_group(tg);
8338 free_rt_sched_group(tg);
8342 /* allocate runqueue etc for a new task group */
8343 struct task_group *sched_create_group(struct task_group *parent)
8345 struct task_group *tg;
8346 unsigned long flags;
8349 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8351 return ERR_PTR(-ENOMEM);
8353 if (!alloc_fair_sched_group(tg, parent))
8356 if (!alloc_rt_sched_group(tg, parent))
8359 spin_lock_irqsave(&task_group_lock, flags);
8360 for_each_possible_cpu(i) {
8361 register_fair_sched_group(tg, i);
8362 register_rt_sched_group(tg, i);
8364 list_add_rcu(&tg->list, &task_groups);
8366 WARN_ON(!parent); /* root should already exist */
8368 tg->parent = parent;
8369 list_add_rcu(&tg->siblings, &parent->children);
8370 INIT_LIST_HEAD(&tg->children);
8371 spin_unlock_irqrestore(&task_group_lock, flags);
8376 free_sched_group(tg);
8377 return ERR_PTR(-ENOMEM);
8380 /* rcu callback to free various structures associated with a task group */
8381 static void free_sched_group_rcu(struct rcu_head *rhp)
8383 /* now it should be safe to free those cfs_rqs */
8384 free_sched_group(container_of(rhp, struct task_group, rcu));
8387 /* Destroy runqueue etc associated with a task group */
8388 void sched_destroy_group(struct task_group *tg)
8390 unsigned long flags;
8393 spin_lock_irqsave(&task_group_lock, flags);
8394 for_each_possible_cpu(i) {
8395 unregister_fair_sched_group(tg, i);
8396 unregister_rt_sched_group(tg, i);
8398 list_del_rcu(&tg->list);
8399 list_del_rcu(&tg->siblings);
8400 spin_unlock_irqrestore(&task_group_lock, flags);
8402 /* wait for possible concurrent references to cfs_rqs complete */
8403 call_rcu(&tg->rcu, free_sched_group_rcu);
8406 /* change task's runqueue when it moves between groups.
8407 * The caller of this function should have put the task in its new group
8408 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8409 * reflect its new group.
8411 void sched_move_task(struct task_struct *tsk)
8414 unsigned long flags;
8417 rq = task_rq_lock(tsk, &flags);
8419 update_rq_clock(rq);
8421 running = task_current(rq, tsk);
8422 on_rq = tsk->se.on_rq;
8425 dequeue_task(rq, tsk, 0);
8426 if (unlikely(running))
8427 tsk->sched_class->put_prev_task(rq, tsk);
8429 set_task_rq(tsk, task_cpu(tsk));
8431 #ifdef CONFIG_FAIR_GROUP_SCHED
8432 if (tsk->sched_class->moved_group)
8433 tsk->sched_class->moved_group(tsk);
8436 if (unlikely(running))
8437 tsk->sched_class->set_curr_task(rq);
8439 enqueue_task(rq, tsk, 0);
8441 task_rq_unlock(rq, &flags);
8443 #endif /* CONFIG_GROUP_SCHED */
8445 #ifdef CONFIG_FAIR_GROUP_SCHED
8446 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8448 struct cfs_rq *cfs_rq = se->cfs_rq;
8453 dequeue_entity(cfs_rq, se, 0);
8455 se->load.weight = shares;
8456 se->load.inv_weight = 0;
8459 enqueue_entity(cfs_rq, se, 0);
8462 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8464 struct cfs_rq *cfs_rq = se->cfs_rq;
8465 struct rq *rq = cfs_rq->rq;
8466 unsigned long flags;
8468 spin_lock_irqsave(&rq->lock, flags);
8469 __set_se_shares(se, shares);
8470 spin_unlock_irqrestore(&rq->lock, flags);
8473 static DEFINE_MUTEX(shares_mutex);
8475 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8478 unsigned long flags;
8481 * We can't change the weight of the root cgroup.
8486 if (shares < MIN_SHARES)
8487 shares = MIN_SHARES;
8488 else if (shares > MAX_SHARES)
8489 shares = MAX_SHARES;
8491 mutex_lock(&shares_mutex);
8492 if (tg->shares == shares)
8495 spin_lock_irqsave(&task_group_lock, flags);
8496 for_each_possible_cpu(i)
8497 unregister_fair_sched_group(tg, i);
8498 list_del_rcu(&tg->siblings);
8499 spin_unlock_irqrestore(&task_group_lock, flags);
8501 /* wait for any ongoing reference to this group to finish */
8502 synchronize_sched();
8505 * Now we are free to modify the group's share on each cpu
8506 * w/o tripping rebalance_share or load_balance_fair.
8508 tg->shares = shares;
8509 for_each_possible_cpu(i) {
8513 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8514 set_se_shares(tg->se[i], shares);
8518 * Enable load balance activity on this group, by inserting it back on
8519 * each cpu's rq->leaf_cfs_rq_list.
8521 spin_lock_irqsave(&task_group_lock, flags);
8522 for_each_possible_cpu(i)
8523 register_fair_sched_group(tg, i);
8524 list_add_rcu(&tg->siblings, &tg->parent->children);
8525 spin_unlock_irqrestore(&task_group_lock, flags);
8527 mutex_unlock(&shares_mutex);
8531 unsigned long sched_group_shares(struct task_group *tg)
8537 #ifdef CONFIG_RT_GROUP_SCHED
8539 * Ensure that the real time constraints are schedulable.
8541 static DEFINE_MUTEX(rt_constraints_mutex);
8543 static unsigned long to_ratio(u64 period, u64 runtime)
8545 if (runtime == RUNTIME_INF)
8548 return div64_u64(runtime << 16, period);
8551 #ifdef CONFIG_CGROUP_SCHED
8552 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8554 struct task_group *tgi, *parent = tg->parent;
8555 unsigned long total = 0;
8558 if (global_rt_period() < period)
8561 return to_ratio(period, runtime) <
8562 to_ratio(global_rt_period(), global_rt_runtime());
8565 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8569 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8573 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8574 tgi->rt_bandwidth.rt_runtime);
8578 return total + to_ratio(period, runtime) <=
8579 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8580 parent->rt_bandwidth.rt_runtime);
8582 #elif defined CONFIG_USER_SCHED
8583 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8585 struct task_group *tgi;
8586 unsigned long total = 0;
8587 unsigned long global_ratio =
8588 to_ratio(global_rt_period(), global_rt_runtime());
8591 list_for_each_entry_rcu(tgi, &task_groups, list) {
8595 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8596 tgi->rt_bandwidth.rt_runtime);
8600 return total + to_ratio(period, runtime) < global_ratio;
8604 /* Must be called with tasklist_lock held */
8605 static inline int tg_has_rt_tasks(struct task_group *tg)
8607 struct task_struct *g, *p;
8608 do_each_thread(g, p) {
8609 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8611 } while_each_thread(g, p);
8615 static int tg_set_bandwidth(struct task_group *tg,
8616 u64 rt_period, u64 rt_runtime)
8620 mutex_lock(&rt_constraints_mutex);
8621 read_lock(&tasklist_lock);
8622 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8626 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8631 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8632 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8633 tg->rt_bandwidth.rt_runtime = rt_runtime;
8635 for_each_possible_cpu(i) {
8636 struct rt_rq *rt_rq = tg->rt_rq[i];
8638 spin_lock(&rt_rq->rt_runtime_lock);
8639 rt_rq->rt_runtime = rt_runtime;
8640 spin_unlock(&rt_rq->rt_runtime_lock);
8642 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8644 read_unlock(&tasklist_lock);
8645 mutex_unlock(&rt_constraints_mutex);
8650 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8652 u64 rt_runtime, rt_period;
8654 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8655 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8656 if (rt_runtime_us < 0)
8657 rt_runtime = RUNTIME_INF;
8659 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8662 long sched_group_rt_runtime(struct task_group *tg)
8666 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8669 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8670 do_div(rt_runtime_us, NSEC_PER_USEC);
8671 return rt_runtime_us;
8674 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8676 u64 rt_runtime, rt_period;
8678 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8679 rt_runtime = tg->rt_bandwidth.rt_runtime;
8681 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8684 long sched_group_rt_period(struct task_group *tg)
8688 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8689 do_div(rt_period_us, NSEC_PER_USEC);
8690 return rt_period_us;
8693 static int sched_rt_global_constraints(void)
8695 struct task_group *tg = &root_task_group;
8696 u64 rt_runtime, rt_period;
8699 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8700 rt_runtime = tg->rt_bandwidth.rt_runtime;
8702 mutex_lock(&rt_constraints_mutex);
8703 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8705 mutex_unlock(&rt_constraints_mutex);
8709 #else /* !CONFIG_RT_GROUP_SCHED */
8710 static int sched_rt_global_constraints(void)
8712 unsigned long flags;
8715 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8716 for_each_possible_cpu(i) {
8717 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8719 spin_lock(&rt_rq->rt_runtime_lock);
8720 rt_rq->rt_runtime = global_rt_runtime();
8721 spin_unlock(&rt_rq->rt_runtime_lock);
8723 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8727 #endif /* CONFIG_RT_GROUP_SCHED */
8729 int sched_rt_handler(struct ctl_table *table, int write,
8730 struct file *filp, void __user *buffer, size_t *lenp,
8734 int old_period, old_runtime;
8735 static DEFINE_MUTEX(mutex);
8738 old_period = sysctl_sched_rt_period;
8739 old_runtime = sysctl_sched_rt_runtime;
8741 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8743 if (!ret && write) {
8744 ret = sched_rt_global_constraints();
8746 sysctl_sched_rt_period = old_period;
8747 sysctl_sched_rt_runtime = old_runtime;
8749 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8750 def_rt_bandwidth.rt_period =
8751 ns_to_ktime(global_rt_period());
8754 mutex_unlock(&mutex);
8759 #ifdef CONFIG_CGROUP_SCHED
8761 /* return corresponding task_group object of a cgroup */
8762 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8764 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8765 struct task_group, css);
8768 static struct cgroup_subsys_state *
8769 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8771 struct task_group *tg, *parent;
8773 if (!cgrp->parent) {
8774 /* This is early initialization for the top cgroup */
8775 init_task_group.css.cgroup = cgrp;
8776 return &init_task_group.css;
8779 parent = cgroup_tg(cgrp->parent);
8780 tg = sched_create_group(parent);
8782 return ERR_PTR(-ENOMEM);
8784 /* Bind the cgroup to task_group object we just created */
8785 tg->css.cgroup = cgrp;
8791 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8793 struct task_group *tg = cgroup_tg(cgrp);
8795 sched_destroy_group(tg);
8799 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8800 struct task_struct *tsk)
8802 #ifdef CONFIG_RT_GROUP_SCHED
8803 /* Don't accept realtime tasks when there is no way for them to run */
8804 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8807 /* We don't support RT-tasks being in separate groups */
8808 if (tsk->sched_class != &fair_sched_class)
8816 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8817 struct cgroup *old_cont, struct task_struct *tsk)
8819 sched_move_task(tsk);
8822 #ifdef CONFIG_FAIR_GROUP_SCHED
8823 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8826 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8829 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8831 struct task_group *tg = cgroup_tg(cgrp);
8833 return (u64) tg->shares;
8835 #endif /* CONFIG_FAIR_GROUP_SCHED */
8837 #ifdef CONFIG_RT_GROUP_SCHED
8838 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8841 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8844 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8846 return sched_group_rt_runtime(cgroup_tg(cgrp));
8849 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8852 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8855 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8857 return sched_group_rt_period(cgroup_tg(cgrp));
8859 #endif /* CONFIG_RT_GROUP_SCHED */
8861 static struct cftype cpu_files[] = {
8862 #ifdef CONFIG_FAIR_GROUP_SCHED
8865 .read_u64 = cpu_shares_read_u64,
8866 .write_u64 = cpu_shares_write_u64,
8869 #ifdef CONFIG_RT_GROUP_SCHED
8871 .name = "rt_runtime_us",
8872 .read_s64 = cpu_rt_runtime_read,
8873 .write_s64 = cpu_rt_runtime_write,
8876 .name = "rt_period_us",
8877 .read_u64 = cpu_rt_period_read_uint,
8878 .write_u64 = cpu_rt_period_write_uint,
8883 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8885 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8888 struct cgroup_subsys cpu_cgroup_subsys = {
8890 .create = cpu_cgroup_create,
8891 .destroy = cpu_cgroup_destroy,
8892 .can_attach = cpu_cgroup_can_attach,
8893 .attach = cpu_cgroup_attach,
8894 .populate = cpu_cgroup_populate,
8895 .subsys_id = cpu_cgroup_subsys_id,
8899 #endif /* CONFIG_CGROUP_SCHED */
8901 #ifdef CONFIG_CGROUP_CPUACCT
8904 * CPU accounting code for task groups.
8906 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8907 * (balbir@in.ibm.com).
8910 /* track cpu usage of a group of tasks */
8912 struct cgroup_subsys_state css;
8913 /* cpuusage holds pointer to a u64-type object on every cpu */
8917 struct cgroup_subsys cpuacct_subsys;
8919 /* return cpu accounting group corresponding to this container */
8920 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8922 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8923 struct cpuacct, css);
8926 /* return cpu accounting group to which this task belongs */
8927 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8929 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8930 struct cpuacct, css);
8933 /* create a new cpu accounting group */
8934 static struct cgroup_subsys_state *cpuacct_create(
8935 struct cgroup_subsys *ss, struct cgroup *cgrp)
8937 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8940 return ERR_PTR(-ENOMEM);
8942 ca->cpuusage = alloc_percpu(u64);
8943 if (!ca->cpuusage) {
8945 return ERR_PTR(-ENOMEM);
8951 /* destroy an existing cpu accounting group */
8953 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8955 struct cpuacct *ca = cgroup_ca(cgrp);
8957 free_percpu(ca->cpuusage);
8961 /* return total cpu usage (in nanoseconds) of a group */
8962 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8964 struct cpuacct *ca = cgroup_ca(cgrp);
8965 u64 totalcpuusage = 0;
8968 for_each_possible_cpu(i) {
8969 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8972 * Take rq->lock to make 64-bit addition safe on 32-bit
8975 spin_lock_irq(&cpu_rq(i)->lock);
8976 totalcpuusage += *cpuusage;
8977 spin_unlock_irq(&cpu_rq(i)->lock);
8980 return totalcpuusage;
8983 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8986 struct cpuacct *ca = cgroup_ca(cgrp);
8995 for_each_possible_cpu(i) {
8996 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8998 spin_lock_irq(&cpu_rq(i)->lock);
9000 spin_unlock_irq(&cpu_rq(i)->lock);
9006 static struct cftype files[] = {
9009 .read_u64 = cpuusage_read,
9010 .write_u64 = cpuusage_write,
9014 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9016 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9020 * charge this task's execution time to its accounting group.
9022 * called with rq->lock held.
9024 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9028 if (!cpuacct_subsys.active)
9033 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9035 *cpuusage += cputime;
9039 struct cgroup_subsys cpuacct_subsys = {
9041 .create = cpuacct_create,
9042 .destroy = cpuacct_destroy,
9043 .populate = cpuacct_populate,
9044 .subsys_id = cpuacct_subsys_id,
9046 #endif /* CONFIG_CGROUP_CPUACCT */