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 unsigned long avg_load_per_task;
559 struct task_struct *migration_thread;
560 struct list_head migration_queue;
563 #ifdef CONFIG_SCHED_HRTICK
564 unsigned long hrtick_flags;
565 ktime_t hrtick_expire;
566 struct hrtimer hrtick_timer;
569 #ifdef CONFIG_SCHEDSTATS
571 struct sched_info rq_sched_info;
573 /* sys_sched_yield() stats */
574 unsigned int yld_exp_empty;
575 unsigned int yld_act_empty;
576 unsigned int yld_both_empty;
577 unsigned int yld_count;
579 /* schedule() stats */
580 unsigned int sched_switch;
581 unsigned int sched_count;
582 unsigned int sched_goidle;
584 /* try_to_wake_up() stats */
585 unsigned int ttwu_count;
586 unsigned int ttwu_local;
589 unsigned int bkl_count;
591 struct lock_class_key rq_lock_key;
594 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
596 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
598 rq->curr->sched_class->check_preempt_curr(rq, p);
601 static inline int cpu_of(struct rq *rq)
611 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
612 * See detach_destroy_domains: synchronize_sched for details.
614 * The domain tree of any CPU may only be accessed from within
615 * preempt-disabled sections.
617 #define for_each_domain(cpu, __sd) \
618 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
620 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
621 #define this_rq() (&__get_cpu_var(runqueues))
622 #define task_rq(p) cpu_rq(task_cpu(p))
623 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
625 static inline void update_rq_clock(struct rq *rq)
627 rq->clock = sched_clock_cpu(cpu_of(rq));
631 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
633 #ifdef CONFIG_SCHED_DEBUG
634 # define const_debug __read_mostly
636 # define const_debug static const
640 * Debugging: various feature bits
643 #define SCHED_FEAT(name, enabled) \
644 __SCHED_FEAT_##name ,
647 #include "sched_features.h"
652 #define SCHED_FEAT(name, enabled) \
653 (1UL << __SCHED_FEAT_##name) * enabled |
655 const_debug unsigned int sysctl_sched_features =
656 #include "sched_features.h"
661 #ifdef CONFIG_SCHED_DEBUG
662 #define SCHED_FEAT(name, enabled) \
665 static __read_mostly char *sched_feat_names[] = {
666 #include "sched_features.h"
672 static int sched_feat_open(struct inode *inode, struct file *filp)
674 filp->private_data = inode->i_private;
679 sched_feat_read(struct file *filp, char __user *ubuf,
680 size_t cnt, loff_t *ppos)
687 for (i = 0; sched_feat_names[i]; i++) {
688 len += strlen(sched_feat_names[i]);
692 buf = kmalloc(len + 2, GFP_KERNEL);
696 for (i = 0; sched_feat_names[i]; i++) {
697 if (sysctl_sched_features & (1UL << i))
698 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
700 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
703 r += sprintf(buf + r, "\n");
704 WARN_ON(r >= len + 2);
706 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
714 sched_feat_write(struct file *filp, const char __user *ubuf,
715 size_t cnt, loff_t *ppos)
725 if (copy_from_user(&buf, ubuf, cnt))
730 if (strncmp(buf, "NO_", 3) == 0) {
735 for (i = 0; sched_feat_names[i]; i++) {
736 int len = strlen(sched_feat_names[i]);
738 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
740 sysctl_sched_features &= ~(1UL << i);
742 sysctl_sched_features |= (1UL << i);
747 if (!sched_feat_names[i])
755 static struct file_operations sched_feat_fops = {
756 .open = sched_feat_open,
757 .read = sched_feat_read,
758 .write = sched_feat_write,
761 static __init int sched_init_debug(void)
763 debugfs_create_file("sched_features", 0644, NULL, NULL,
768 late_initcall(sched_init_debug);
772 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
775 * Number of tasks to iterate in a single balance run.
776 * Limited because this is done with IRQs disabled.
778 const_debug unsigned int sysctl_sched_nr_migrate = 32;
781 * ratelimit for updating the group shares.
784 const_debug unsigned int sysctl_sched_shares_ratelimit = 500000;
787 * period over which we measure -rt task cpu usage in us.
790 unsigned int sysctl_sched_rt_period = 1000000;
792 static __read_mostly int scheduler_running;
795 * part of the period that we allow rt tasks to run in us.
798 int sysctl_sched_rt_runtime = 950000;
800 static inline u64 global_rt_period(void)
802 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
805 static inline u64 global_rt_runtime(void)
807 if (sysctl_sched_rt_period < 0)
810 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
813 #ifndef prepare_arch_switch
814 # define prepare_arch_switch(next) do { } while (0)
816 #ifndef finish_arch_switch
817 # define finish_arch_switch(prev) do { } while (0)
820 static inline int task_current(struct rq *rq, struct task_struct *p)
822 return rq->curr == p;
825 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
826 static inline int task_running(struct rq *rq, struct task_struct *p)
828 return task_current(rq, p);
831 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
835 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
837 #ifdef CONFIG_DEBUG_SPINLOCK
838 /* this is a valid case when another task releases the spinlock */
839 rq->lock.owner = current;
842 * If we are tracking spinlock dependencies then we have to
843 * fix up the runqueue lock - which gets 'carried over' from
846 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
848 spin_unlock_irq(&rq->lock);
851 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
852 static inline int task_running(struct rq *rq, struct task_struct *p)
857 return task_current(rq, p);
861 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
865 * We can optimise this out completely for !SMP, because the
866 * SMP rebalancing from interrupt is the only thing that cares
871 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
872 spin_unlock_irq(&rq->lock);
874 spin_unlock(&rq->lock);
878 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
882 * After ->oncpu is cleared, the task can be moved to a different CPU.
883 * We must ensure this doesn't happen until the switch is completely
889 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
893 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
896 * __task_rq_lock - lock the runqueue a given task resides on.
897 * Must be called interrupts disabled.
899 static inline struct rq *__task_rq_lock(struct task_struct *p)
903 struct rq *rq = task_rq(p);
904 spin_lock(&rq->lock);
905 if (likely(rq == task_rq(p)))
907 spin_unlock(&rq->lock);
912 * task_rq_lock - lock the runqueue a given task resides on and disable
913 * interrupts. Note the ordering: we can safely lookup the task_rq without
914 * explicitly disabling preemption.
916 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
922 local_irq_save(*flags);
924 spin_lock(&rq->lock);
925 if (likely(rq == task_rq(p)))
927 spin_unlock_irqrestore(&rq->lock, *flags);
931 static void __task_rq_unlock(struct rq *rq)
934 spin_unlock(&rq->lock);
937 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
940 spin_unlock_irqrestore(&rq->lock, *flags);
944 * this_rq_lock - lock this runqueue and disable interrupts.
946 static struct rq *this_rq_lock(void)
953 spin_lock(&rq->lock);
958 static void __resched_task(struct task_struct *p, int tif_bit);
960 static inline void resched_task(struct task_struct *p)
962 __resched_task(p, TIF_NEED_RESCHED);
965 #ifdef CONFIG_SCHED_HRTICK
967 * Use HR-timers to deliver accurate preemption points.
969 * Its all a bit involved since we cannot program an hrt while holding the
970 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
973 * When we get rescheduled we reprogram the hrtick_timer outside of the
976 static inline void resched_hrt(struct task_struct *p)
978 __resched_task(p, TIF_HRTICK_RESCHED);
981 static inline void resched_rq(struct rq *rq)
985 spin_lock_irqsave(&rq->lock, flags);
986 resched_task(rq->curr);
987 spin_unlock_irqrestore(&rq->lock, flags);
991 HRTICK_SET, /* re-programm hrtick_timer */
992 HRTICK_RESET, /* not a new slice */
993 HRTICK_BLOCK, /* stop hrtick operations */
998 * - enabled by features
999 * - hrtimer is actually high res
1001 static inline int hrtick_enabled(struct rq *rq)
1003 if (!sched_feat(HRTICK))
1005 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1007 return hrtimer_is_hres_active(&rq->hrtick_timer);
1011 * Called to set the hrtick timer state.
1013 * called with rq->lock held and irqs disabled
1015 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1017 assert_spin_locked(&rq->lock);
1020 * preempt at: now + delay
1023 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1025 * indicate we need to program the timer
1027 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1029 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1032 * New slices are called from the schedule path and don't need a
1033 * forced reschedule.
1036 resched_hrt(rq->curr);
1039 static void hrtick_clear(struct rq *rq)
1041 if (hrtimer_active(&rq->hrtick_timer))
1042 hrtimer_cancel(&rq->hrtick_timer);
1046 * Update the timer from the possible pending state.
1048 static void hrtick_set(struct rq *rq)
1052 unsigned long flags;
1054 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1056 spin_lock_irqsave(&rq->lock, flags);
1057 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1058 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1059 time = rq->hrtick_expire;
1060 clear_thread_flag(TIF_HRTICK_RESCHED);
1061 spin_unlock_irqrestore(&rq->lock, flags);
1064 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1065 if (reset && !hrtimer_active(&rq->hrtick_timer))
1072 * High-resolution timer tick.
1073 * Runs from hardirq context with interrupts disabled.
1075 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1077 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1079 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1081 spin_lock(&rq->lock);
1082 update_rq_clock(rq);
1083 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1084 spin_unlock(&rq->lock);
1086 return HRTIMER_NORESTART;
1090 static void hotplug_hrtick_disable(int cpu)
1092 struct rq *rq = cpu_rq(cpu);
1093 unsigned long flags;
1095 spin_lock_irqsave(&rq->lock, flags);
1096 rq->hrtick_flags = 0;
1097 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1098 spin_unlock_irqrestore(&rq->lock, flags);
1103 static void hotplug_hrtick_enable(int cpu)
1105 struct rq *rq = cpu_rq(cpu);
1106 unsigned long flags;
1108 spin_lock_irqsave(&rq->lock, flags);
1109 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1110 spin_unlock_irqrestore(&rq->lock, flags);
1114 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1116 int cpu = (int)(long)hcpu;
1119 case CPU_UP_CANCELED:
1120 case CPU_UP_CANCELED_FROZEN:
1121 case CPU_DOWN_PREPARE:
1122 case CPU_DOWN_PREPARE_FROZEN:
1124 case CPU_DEAD_FROZEN:
1125 hotplug_hrtick_disable(cpu);
1128 case CPU_UP_PREPARE:
1129 case CPU_UP_PREPARE_FROZEN:
1130 case CPU_DOWN_FAILED:
1131 case CPU_DOWN_FAILED_FROZEN:
1133 case CPU_ONLINE_FROZEN:
1134 hotplug_hrtick_enable(cpu);
1141 static void init_hrtick(void)
1143 hotcpu_notifier(hotplug_hrtick, 0);
1145 #endif /* CONFIG_SMP */
1147 static void init_rq_hrtick(struct rq *rq)
1149 rq->hrtick_flags = 0;
1150 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1151 rq->hrtick_timer.function = hrtick;
1152 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1155 void hrtick_resched(void)
1158 unsigned long flags;
1160 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1163 local_irq_save(flags);
1164 rq = cpu_rq(smp_processor_id());
1166 local_irq_restore(flags);
1169 static inline void hrtick_clear(struct rq *rq)
1173 static inline void hrtick_set(struct rq *rq)
1177 static inline void init_rq_hrtick(struct rq *rq)
1181 void hrtick_resched(void)
1185 static inline void init_hrtick(void)
1191 * resched_task - mark a task 'to be rescheduled now'.
1193 * On UP this means the setting of the need_resched flag, on SMP it
1194 * might also involve a cross-CPU call to trigger the scheduler on
1199 #ifndef tsk_is_polling
1200 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1203 static void __resched_task(struct task_struct *p, int tif_bit)
1207 assert_spin_locked(&task_rq(p)->lock);
1209 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1212 set_tsk_thread_flag(p, tif_bit);
1215 if (cpu == smp_processor_id())
1218 /* NEED_RESCHED must be visible before we test polling */
1220 if (!tsk_is_polling(p))
1221 smp_send_reschedule(cpu);
1224 static void resched_cpu(int cpu)
1226 struct rq *rq = cpu_rq(cpu);
1227 unsigned long flags;
1229 if (!spin_trylock_irqsave(&rq->lock, flags))
1231 resched_task(cpu_curr(cpu));
1232 spin_unlock_irqrestore(&rq->lock, flags);
1237 * When add_timer_on() enqueues a timer into the timer wheel of an
1238 * idle CPU then this timer might expire before the next timer event
1239 * which is scheduled to wake up that CPU. In case of a completely
1240 * idle system the next event might even be infinite time into the
1241 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1242 * leaves the inner idle loop so the newly added timer is taken into
1243 * account when the CPU goes back to idle and evaluates the timer
1244 * wheel for the next timer event.
1246 void wake_up_idle_cpu(int cpu)
1248 struct rq *rq = cpu_rq(cpu);
1250 if (cpu == smp_processor_id())
1254 * This is safe, as this function is called with the timer
1255 * wheel base lock of (cpu) held. When the CPU is on the way
1256 * to idle and has not yet set rq->curr to idle then it will
1257 * be serialized on the timer wheel base lock and take the new
1258 * timer into account automatically.
1260 if (rq->curr != rq->idle)
1264 * We can set TIF_RESCHED on the idle task of the other CPU
1265 * lockless. The worst case is that the other CPU runs the
1266 * idle task through an additional NOOP schedule()
1268 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1270 /* NEED_RESCHED must be visible before we test polling */
1272 if (!tsk_is_polling(rq->idle))
1273 smp_send_reschedule(cpu);
1275 #endif /* CONFIG_NO_HZ */
1277 #else /* !CONFIG_SMP */
1278 static void __resched_task(struct task_struct *p, int tif_bit)
1280 assert_spin_locked(&task_rq(p)->lock);
1281 set_tsk_thread_flag(p, tif_bit);
1283 #endif /* CONFIG_SMP */
1285 #if BITS_PER_LONG == 32
1286 # define WMULT_CONST (~0UL)
1288 # define WMULT_CONST (1UL << 32)
1291 #define WMULT_SHIFT 32
1294 * Shift right and round:
1296 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1299 * delta *= weight / lw
1301 static unsigned long
1302 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1303 struct load_weight *lw)
1307 if (!lw->inv_weight) {
1308 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1311 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1315 tmp = (u64)delta_exec * weight;
1317 * Check whether we'd overflow the 64-bit multiplication:
1319 if (unlikely(tmp > WMULT_CONST))
1320 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1323 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1325 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1328 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1334 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1341 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1342 * of tasks with abnormal "nice" values across CPUs the contribution that
1343 * each task makes to its run queue's load is weighted according to its
1344 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1345 * scaled version of the new time slice allocation that they receive on time
1349 #define WEIGHT_IDLEPRIO 2
1350 #define WMULT_IDLEPRIO (1 << 31)
1353 * Nice levels are multiplicative, with a gentle 10% change for every
1354 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1355 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1356 * that remained on nice 0.
1358 * The "10% effect" is relative and cumulative: from _any_ nice level,
1359 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1360 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1361 * If a task goes up by ~10% and another task goes down by ~10% then
1362 * the relative distance between them is ~25%.)
1364 static const int prio_to_weight[40] = {
1365 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1366 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1367 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1368 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1369 /* 0 */ 1024, 820, 655, 526, 423,
1370 /* 5 */ 335, 272, 215, 172, 137,
1371 /* 10 */ 110, 87, 70, 56, 45,
1372 /* 15 */ 36, 29, 23, 18, 15,
1376 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1378 * In cases where the weight does not change often, we can use the
1379 * precalculated inverse to speed up arithmetics by turning divisions
1380 * into multiplications:
1382 static const u32 prio_to_wmult[40] = {
1383 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1384 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1385 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1386 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1387 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1388 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1389 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1390 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1393 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1396 * runqueue iterator, to support SMP load-balancing between different
1397 * scheduling classes, without having to expose their internal data
1398 * structures to the load-balancing proper:
1400 struct rq_iterator {
1402 struct task_struct *(*start)(void *);
1403 struct task_struct *(*next)(void *);
1407 static unsigned long
1408 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1409 unsigned long max_load_move, struct sched_domain *sd,
1410 enum cpu_idle_type idle, int *all_pinned,
1411 int *this_best_prio, struct rq_iterator *iterator);
1414 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1415 struct sched_domain *sd, enum cpu_idle_type idle,
1416 struct rq_iterator *iterator);
1419 #ifdef CONFIG_CGROUP_CPUACCT
1420 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1422 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1425 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1427 update_load_add(&rq->load, load);
1430 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1432 update_load_sub(&rq->load, load);
1436 static unsigned long source_load(int cpu, int type);
1437 static unsigned long target_load(int cpu, int type);
1438 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1440 static unsigned long cpu_avg_load_per_task(int cpu)
1442 struct rq *rq = cpu_rq(cpu);
1445 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1447 return rq->avg_load_per_task;
1450 #ifdef CONFIG_FAIR_GROUP_SCHED
1452 typedef void (*tg_visitor)(struct task_group *, int, struct sched_domain *);
1455 * Iterate the full tree, calling @down when first entering a node and @up when
1456 * leaving it for the final time.
1459 walk_tg_tree(tg_visitor down, tg_visitor up, int cpu, struct sched_domain *sd)
1461 struct task_group *parent, *child;
1464 parent = &root_task_group;
1466 (*down)(parent, cpu, sd);
1467 list_for_each_entry_rcu(child, &parent->children, siblings) {
1474 (*up)(parent, cpu, sd);
1477 parent = parent->parent;
1483 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1486 * Calculate and set the cpu's group shares.
1489 __update_group_shares_cpu(struct task_group *tg, int cpu,
1490 unsigned long sd_shares, unsigned long sd_rq_weight)
1493 unsigned long shares;
1494 unsigned long rq_weight;
1499 rq_weight = tg->cfs_rq[cpu]->load.weight;
1502 * If there are currently no tasks on the cpu pretend there is one of
1503 * average load so that when a new task gets to run here it will not
1504 * get delayed by group starvation.
1508 rq_weight = NICE_0_LOAD;
1511 if (unlikely(rq_weight > sd_rq_weight))
1512 rq_weight = sd_rq_weight;
1515 * \Sum shares * rq_weight
1516 * shares = -----------------------
1520 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1523 * record the actual number of shares, not the boosted amount.
1525 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1527 if (shares < MIN_SHARES)
1528 shares = MIN_SHARES;
1529 else if (shares > MAX_SHARES)
1530 shares = MAX_SHARES;
1532 __set_se_shares(tg->se[cpu], shares);
1536 * Re-compute the task group their per cpu shares over the given domain.
1537 * This needs to be done in a bottom-up fashion because the rq weight of a
1538 * parent group depends on the shares of its child groups.
1541 tg_shares_up(struct task_group *tg, int cpu, struct sched_domain *sd)
1543 unsigned long rq_weight = 0;
1544 unsigned long shares = 0;
1547 for_each_cpu_mask(i, sd->span) {
1548 rq_weight += tg->cfs_rq[i]->load.weight;
1549 shares += tg->cfs_rq[i]->shares;
1552 if ((!shares && rq_weight) || shares > tg->shares)
1553 shares = tg->shares;
1555 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1556 shares = tg->shares;
1559 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1561 for_each_cpu_mask(i, sd->span) {
1562 struct rq *rq = cpu_rq(i);
1563 unsigned long flags;
1565 spin_lock_irqsave(&rq->lock, flags);
1566 __update_group_shares_cpu(tg, i, shares, rq_weight);
1567 spin_unlock_irqrestore(&rq->lock, flags);
1572 * Compute the cpu's hierarchical load factor for each task group.
1573 * This needs to be done in a top-down fashion because the load of a child
1574 * group is a fraction of its parents load.
1577 tg_load_down(struct task_group *tg, int cpu, struct sched_domain *sd)
1582 load = cpu_rq(cpu)->load.weight;
1584 load = tg->parent->cfs_rq[cpu]->h_load;
1585 load *= tg->cfs_rq[cpu]->shares;
1586 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1589 tg->cfs_rq[cpu]->h_load = load;
1593 tg_nop(struct task_group *tg, int cpu, struct sched_domain *sd)
1597 static void update_shares(struct sched_domain *sd)
1599 u64 now = cpu_clock(raw_smp_processor_id());
1600 s64 elapsed = now - sd->last_update;
1602 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1603 sd->last_update = now;
1604 walk_tg_tree(tg_nop, tg_shares_up, 0, sd);
1608 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1610 spin_unlock(&rq->lock);
1612 spin_lock(&rq->lock);
1615 static void update_h_load(int cpu)
1617 walk_tg_tree(tg_load_down, tg_nop, cpu, NULL);
1620 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1622 cfs_rq->shares = shares;
1627 static inline void update_shares(struct sched_domain *sd)
1631 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1639 #include "sched_stats.h"
1640 #include "sched_idletask.c"
1641 #include "sched_fair.c"
1642 #include "sched_rt.c"
1643 #ifdef CONFIG_SCHED_DEBUG
1644 # include "sched_debug.c"
1647 #define sched_class_highest (&rt_sched_class)
1648 #define for_each_class(class) \
1649 for (class = sched_class_highest; class; class = class->next)
1651 static void inc_nr_running(struct rq *rq)
1656 static void dec_nr_running(struct rq *rq)
1661 static void set_load_weight(struct task_struct *p)
1663 if (task_has_rt_policy(p)) {
1664 p->se.load.weight = prio_to_weight[0] * 2;
1665 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1670 * SCHED_IDLE tasks get minimal weight:
1672 if (p->policy == SCHED_IDLE) {
1673 p->se.load.weight = WEIGHT_IDLEPRIO;
1674 p->se.load.inv_weight = WMULT_IDLEPRIO;
1678 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1679 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1682 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1684 sched_info_queued(p);
1685 p->sched_class->enqueue_task(rq, p, wakeup);
1689 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1691 p->sched_class->dequeue_task(rq, p, sleep);
1696 * __normal_prio - return the priority that is based on the static prio
1698 static inline int __normal_prio(struct task_struct *p)
1700 return p->static_prio;
1704 * Calculate the expected normal priority: i.e. priority
1705 * without taking RT-inheritance into account. Might be
1706 * boosted by interactivity modifiers. Changes upon fork,
1707 * setprio syscalls, and whenever the interactivity
1708 * estimator recalculates.
1710 static inline int normal_prio(struct task_struct *p)
1714 if (task_has_rt_policy(p))
1715 prio = MAX_RT_PRIO-1 - p->rt_priority;
1717 prio = __normal_prio(p);
1722 * Calculate the current priority, i.e. the priority
1723 * taken into account by the scheduler. This value might
1724 * be boosted by RT tasks, or might be boosted by
1725 * interactivity modifiers. Will be RT if the task got
1726 * RT-boosted. If not then it returns p->normal_prio.
1728 static int effective_prio(struct task_struct *p)
1730 p->normal_prio = normal_prio(p);
1732 * If we are RT tasks or we were boosted to RT priority,
1733 * keep the priority unchanged. Otherwise, update priority
1734 * to the normal priority:
1736 if (!rt_prio(p->prio))
1737 return p->normal_prio;
1742 * activate_task - move a task to the runqueue.
1744 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1746 if (task_contributes_to_load(p))
1747 rq->nr_uninterruptible--;
1749 enqueue_task(rq, p, wakeup);
1754 * deactivate_task - remove a task from the runqueue.
1756 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1758 if (task_contributes_to_load(p))
1759 rq->nr_uninterruptible++;
1761 dequeue_task(rq, p, sleep);
1766 * task_curr - is this task currently executing on a CPU?
1767 * @p: the task in question.
1769 inline int task_curr(const struct task_struct *p)
1771 return cpu_curr(task_cpu(p)) == p;
1774 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1776 set_task_rq(p, cpu);
1779 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1780 * successfuly executed on another CPU. We must ensure that updates of
1781 * per-task data have been completed by this moment.
1784 task_thread_info(p)->cpu = cpu;
1788 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1789 const struct sched_class *prev_class,
1790 int oldprio, int running)
1792 if (prev_class != p->sched_class) {
1793 if (prev_class->switched_from)
1794 prev_class->switched_from(rq, p, running);
1795 p->sched_class->switched_to(rq, p, running);
1797 p->sched_class->prio_changed(rq, p, oldprio, running);
1802 /* Used instead of source_load when we know the type == 0 */
1803 static unsigned long weighted_cpuload(const int cpu)
1805 return cpu_rq(cpu)->load.weight;
1809 * Is this task likely cache-hot:
1812 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1817 * Buddy candidates are cache hot:
1819 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1822 if (p->sched_class != &fair_sched_class)
1825 if (sysctl_sched_migration_cost == -1)
1827 if (sysctl_sched_migration_cost == 0)
1830 delta = now - p->se.exec_start;
1832 return delta < (s64)sysctl_sched_migration_cost;
1836 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1838 int old_cpu = task_cpu(p);
1839 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1840 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1841 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1844 clock_offset = old_rq->clock - new_rq->clock;
1846 #ifdef CONFIG_SCHEDSTATS
1847 if (p->se.wait_start)
1848 p->se.wait_start -= clock_offset;
1849 if (p->se.sleep_start)
1850 p->se.sleep_start -= clock_offset;
1851 if (p->se.block_start)
1852 p->se.block_start -= clock_offset;
1853 if (old_cpu != new_cpu) {
1854 schedstat_inc(p, se.nr_migrations);
1855 if (task_hot(p, old_rq->clock, NULL))
1856 schedstat_inc(p, se.nr_forced2_migrations);
1859 p->se.vruntime -= old_cfsrq->min_vruntime -
1860 new_cfsrq->min_vruntime;
1862 __set_task_cpu(p, new_cpu);
1865 struct migration_req {
1866 struct list_head list;
1868 struct task_struct *task;
1871 struct completion done;
1875 * The task's runqueue lock must be held.
1876 * Returns true if you have to wait for migration thread.
1879 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1881 struct rq *rq = task_rq(p);
1884 * If the task is not on a runqueue (and not running), then
1885 * it is sufficient to simply update the task's cpu field.
1887 if (!p->se.on_rq && !task_running(rq, p)) {
1888 set_task_cpu(p, dest_cpu);
1892 init_completion(&req->done);
1894 req->dest_cpu = dest_cpu;
1895 list_add(&req->list, &rq->migration_queue);
1901 * wait_task_inactive - wait for a thread to unschedule.
1903 * The caller must ensure that the task *will* unschedule sometime soon,
1904 * else this function might spin for a *long* time. This function can't
1905 * be called with interrupts off, or it may introduce deadlock with
1906 * smp_call_function() if an IPI is sent by the same process we are
1907 * waiting to become inactive.
1909 void wait_task_inactive(struct task_struct *p)
1911 unsigned long flags;
1917 * We do the initial early heuristics without holding
1918 * any task-queue locks at all. We'll only try to get
1919 * the runqueue lock when things look like they will
1925 * If the task is actively running on another CPU
1926 * still, just relax and busy-wait without holding
1929 * NOTE! Since we don't hold any locks, it's not
1930 * even sure that "rq" stays as the right runqueue!
1931 * But we don't care, since "task_running()" will
1932 * return false if the runqueue has changed and p
1933 * is actually now running somewhere else!
1935 while (task_running(rq, p))
1939 * Ok, time to look more closely! We need the rq
1940 * lock now, to be *sure*. If we're wrong, we'll
1941 * just go back and repeat.
1943 rq = task_rq_lock(p, &flags);
1944 running = task_running(rq, p);
1945 on_rq = p->se.on_rq;
1946 task_rq_unlock(rq, &flags);
1949 * Was it really running after all now that we
1950 * checked with the proper locks actually held?
1952 * Oops. Go back and try again..
1954 if (unlikely(running)) {
1960 * It's not enough that it's not actively running,
1961 * it must be off the runqueue _entirely_, and not
1964 * So if it wa still runnable (but just not actively
1965 * running right now), it's preempted, and we should
1966 * yield - it could be a while.
1968 if (unlikely(on_rq)) {
1969 schedule_timeout_uninterruptible(1);
1974 * Ahh, all good. It wasn't running, and it wasn't
1975 * runnable, which means that it will never become
1976 * running in the future either. We're all done!
1983 * kick_process - kick a running thread to enter/exit the kernel
1984 * @p: the to-be-kicked thread
1986 * Cause a process which is running on another CPU to enter
1987 * kernel-mode, without any delay. (to get signals handled.)
1989 * NOTE: this function doesnt have to take the runqueue lock,
1990 * because all it wants to ensure is that the remote task enters
1991 * the kernel. If the IPI races and the task has been migrated
1992 * to another CPU then no harm is done and the purpose has been
1995 void kick_process(struct task_struct *p)
2001 if ((cpu != smp_processor_id()) && task_curr(p))
2002 smp_send_reschedule(cpu);
2007 * Return a low guess at the load of a migration-source cpu weighted
2008 * according to the scheduling class and "nice" value.
2010 * We want to under-estimate the load of migration sources, to
2011 * balance conservatively.
2013 static unsigned long source_load(int cpu, int type)
2015 struct rq *rq = cpu_rq(cpu);
2016 unsigned long total = weighted_cpuload(cpu);
2018 if (type == 0 || !sched_feat(LB_BIAS))
2021 return min(rq->cpu_load[type-1], total);
2025 * Return a high guess at the load of a migration-target cpu weighted
2026 * according to the scheduling class and "nice" value.
2028 static unsigned long target_load(int cpu, int type)
2030 struct rq *rq = cpu_rq(cpu);
2031 unsigned long total = weighted_cpuload(cpu);
2033 if (type == 0 || !sched_feat(LB_BIAS))
2036 return max(rq->cpu_load[type-1], total);
2040 * find_idlest_group finds and returns the least busy CPU group within the
2043 static struct sched_group *
2044 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2046 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2047 unsigned long min_load = ULONG_MAX, this_load = 0;
2048 int load_idx = sd->forkexec_idx;
2049 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2052 unsigned long load, avg_load;
2056 /* Skip over this group if it has no CPUs allowed */
2057 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2060 local_group = cpu_isset(this_cpu, group->cpumask);
2062 /* Tally up the load of all CPUs in the group */
2065 for_each_cpu_mask(i, group->cpumask) {
2066 /* Bias balancing toward cpus of our domain */
2068 load = source_load(i, load_idx);
2070 load = target_load(i, load_idx);
2075 /* Adjust by relative CPU power of the group */
2076 avg_load = sg_div_cpu_power(group,
2077 avg_load * SCHED_LOAD_SCALE);
2080 this_load = avg_load;
2082 } else if (avg_load < min_load) {
2083 min_load = avg_load;
2086 } while (group = group->next, group != sd->groups);
2088 if (!idlest || 100*this_load < imbalance*min_load)
2094 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2097 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2100 unsigned long load, min_load = ULONG_MAX;
2104 /* Traverse only the allowed CPUs */
2105 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2107 for_each_cpu_mask(i, *tmp) {
2108 load = weighted_cpuload(i);
2110 if (load < min_load || (load == min_load && i == this_cpu)) {
2120 * sched_balance_self: balance the current task (running on cpu) in domains
2121 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2124 * Balance, ie. select the least loaded group.
2126 * Returns the target CPU number, or the same CPU if no balancing is needed.
2128 * preempt must be disabled.
2130 static int sched_balance_self(int cpu, int flag)
2132 struct task_struct *t = current;
2133 struct sched_domain *tmp, *sd = NULL;
2135 for_each_domain(cpu, tmp) {
2137 * If power savings logic is enabled for a domain, stop there.
2139 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2141 if (tmp->flags & flag)
2149 cpumask_t span, tmpmask;
2150 struct sched_group *group;
2151 int new_cpu, weight;
2153 if (!(sd->flags & flag)) {
2159 group = find_idlest_group(sd, t, cpu);
2165 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2166 if (new_cpu == -1 || new_cpu == cpu) {
2167 /* Now try balancing at a lower domain level of cpu */
2172 /* Now try balancing at a lower domain level of new_cpu */
2175 weight = cpus_weight(span);
2176 for_each_domain(cpu, tmp) {
2177 if (weight <= cpus_weight(tmp->span))
2179 if (tmp->flags & flag)
2182 /* while loop will break here if sd == NULL */
2188 #endif /* CONFIG_SMP */
2191 * try_to_wake_up - wake up a thread
2192 * @p: the to-be-woken-up thread
2193 * @state: the mask of task states that can be woken
2194 * @sync: do a synchronous wakeup?
2196 * Put it on the run-queue if it's not already there. The "current"
2197 * thread is always on the run-queue (except when the actual
2198 * re-schedule is in progress), and as such you're allowed to do
2199 * the simpler "current->state = TASK_RUNNING" to mark yourself
2200 * runnable without the overhead of this.
2202 * returns failure only if the task is already active.
2204 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2206 int cpu, orig_cpu, this_cpu, success = 0;
2207 unsigned long flags;
2211 if (!sched_feat(SYNC_WAKEUPS))
2215 if (sched_feat(LB_WAKEUP_UPDATE)) {
2216 struct sched_domain *sd;
2218 this_cpu = raw_smp_processor_id();
2221 for_each_domain(this_cpu, sd) {
2222 if (cpu_isset(cpu, sd->span)) {
2231 rq = task_rq_lock(p, &flags);
2232 old_state = p->state;
2233 if (!(old_state & state))
2241 this_cpu = smp_processor_id();
2244 if (unlikely(task_running(rq, p)))
2247 cpu = p->sched_class->select_task_rq(p, sync);
2248 if (cpu != orig_cpu) {
2249 set_task_cpu(p, cpu);
2250 task_rq_unlock(rq, &flags);
2251 /* might preempt at this point */
2252 rq = task_rq_lock(p, &flags);
2253 old_state = p->state;
2254 if (!(old_state & state))
2259 this_cpu = smp_processor_id();
2263 #ifdef CONFIG_SCHEDSTATS
2264 schedstat_inc(rq, ttwu_count);
2265 if (cpu == this_cpu)
2266 schedstat_inc(rq, ttwu_local);
2268 struct sched_domain *sd;
2269 for_each_domain(this_cpu, sd) {
2270 if (cpu_isset(cpu, sd->span)) {
2271 schedstat_inc(sd, ttwu_wake_remote);
2276 #endif /* CONFIG_SCHEDSTATS */
2279 #endif /* CONFIG_SMP */
2280 schedstat_inc(p, se.nr_wakeups);
2282 schedstat_inc(p, se.nr_wakeups_sync);
2283 if (orig_cpu != cpu)
2284 schedstat_inc(p, se.nr_wakeups_migrate);
2285 if (cpu == this_cpu)
2286 schedstat_inc(p, se.nr_wakeups_local);
2288 schedstat_inc(p, se.nr_wakeups_remote);
2289 update_rq_clock(rq);
2290 activate_task(rq, p, 1);
2294 check_preempt_curr(rq, p);
2296 p->state = TASK_RUNNING;
2298 if (p->sched_class->task_wake_up)
2299 p->sched_class->task_wake_up(rq, p);
2302 task_rq_unlock(rq, &flags);
2307 int wake_up_process(struct task_struct *p)
2309 return try_to_wake_up(p, TASK_ALL, 0);
2311 EXPORT_SYMBOL(wake_up_process);
2313 int wake_up_state(struct task_struct *p, unsigned int state)
2315 return try_to_wake_up(p, state, 0);
2319 * Perform scheduler related setup for a newly forked process p.
2320 * p is forked by current.
2322 * __sched_fork() is basic setup used by init_idle() too:
2324 static void __sched_fork(struct task_struct *p)
2326 p->se.exec_start = 0;
2327 p->se.sum_exec_runtime = 0;
2328 p->se.prev_sum_exec_runtime = 0;
2329 p->se.last_wakeup = 0;
2330 p->se.avg_overlap = 0;
2332 #ifdef CONFIG_SCHEDSTATS
2333 p->se.wait_start = 0;
2334 p->se.sum_sleep_runtime = 0;
2335 p->se.sleep_start = 0;
2336 p->se.block_start = 0;
2337 p->se.sleep_max = 0;
2338 p->se.block_max = 0;
2340 p->se.slice_max = 0;
2344 INIT_LIST_HEAD(&p->rt.run_list);
2346 INIT_LIST_HEAD(&p->se.group_node);
2348 #ifdef CONFIG_PREEMPT_NOTIFIERS
2349 INIT_HLIST_HEAD(&p->preempt_notifiers);
2353 * We mark the process as running here, but have not actually
2354 * inserted it onto the runqueue yet. This guarantees that
2355 * nobody will actually run it, and a signal or other external
2356 * event cannot wake it up and insert it on the runqueue either.
2358 p->state = TASK_RUNNING;
2362 * fork()/clone()-time setup:
2364 void sched_fork(struct task_struct *p, int clone_flags)
2366 int cpu = get_cpu();
2371 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2373 set_task_cpu(p, cpu);
2376 * Make sure we do not leak PI boosting priority to the child:
2378 p->prio = current->normal_prio;
2379 if (!rt_prio(p->prio))
2380 p->sched_class = &fair_sched_class;
2382 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2383 if (likely(sched_info_on()))
2384 memset(&p->sched_info, 0, sizeof(p->sched_info));
2386 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2389 #ifdef CONFIG_PREEMPT
2390 /* Want to start with kernel preemption disabled. */
2391 task_thread_info(p)->preempt_count = 1;
2397 * wake_up_new_task - wake up a newly created task for the first time.
2399 * This function will do some initial scheduler statistics housekeeping
2400 * that must be done for every newly created context, then puts the task
2401 * on the runqueue and wakes it.
2403 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2405 unsigned long flags;
2408 rq = task_rq_lock(p, &flags);
2409 BUG_ON(p->state != TASK_RUNNING);
2410 update_rq_clock(rq);
2412 p->prio = effective_prio(p);
2414 if (!p->sched_class->task_new || !current->se.on_rq) {
2415 activate_task(rq, p, 0);
2418 * Let the scheduling class do new task startup
2419 * management (if any):
2421 p->sched_class->task_new(rq, p);
2424 check_preempt_curr(rq, p);
2426 if (p->sched_class->task_wake_up)
2427 p->sched_class->task_wake_up(rq, p);
2429 task_rq_unlock(rq, &flags);
2432 #ifdef CONFIG_PREEMPT_NOTIFIERS
2435 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2436 * @notifier: notifier struct to register
2438 void preempt_notifier_register(struct preempt_notifier *notifier)
2440 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2442 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2445 * preempt_notifier_unregister - no longer interested in preemption notifications
2446 * @notifier: notifier struct to unregister
2448 * This is safe to call from within a preemption notifier.
2450 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2452 hlist_del(¬ifier->link);
2454 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2456 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2458 struct preempt_notifier *notifier;
2459 struct hlist_node *node;
2461 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2462 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2466 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2467 struct task_struct *next)
2469 struct preempt_notifier *notifier;
2470 struct hlist_node *node;
2472 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2473 notifier->ops->sched_out(notifier, next);
2476 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2478 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2483 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2484 struct task_struct *next)
2488 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2491 * prepare_task_switch - prepare to switch tasks
2492 * @rq: the runqueue preparing to switch
2493 * @prev: the current task that is being switched out
2494 * @next: the task we are going to switch to.
2496 * This is called with the rq lock held and interrupts off. It must
2497 * be paired with a subsequent finish_task_switch after the context
2500 * prepare_task_switch sets up locking and calls architecture specific
2504 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2505 struct task_struct *next)
2507 fire_sched_out_preempt_notifiers(prev, next);
2508 prepare_lock_switch(rq, next);
2509 prepare_arch_switch(next);
2513 * finish_task_switch - clean up after a task-switch
2514 * @rq: runqueue associated with task-switch
2515 * @prev: the thread we just switched away from.
2517 * finish_task_switch must be called after the context switch, paired
2518 * with a prepare_task_switch call before the context switch.
2519 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2520 * and do any other architecture-specific cleanup actions.
2522 * Note that we may have delayed dropping an mm in context_switch(). If
2523 * so, we finish that here outside of the runqueue lock. (Doing it
2524 * with the lock held can cause deadlocks; see schedule() for
2527 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2528 __releases(rq->lock)
2530 struct mm_struct *mm = rq->prev_mm;
2536 * A task struct has one reference for the use as "current".
2537 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2538 * schedule one last time. The schedule call will never return, and
2539 * the scheduled task must drop that reference.
2540 * The test for TASK_DEAD must occur while the runqueue locks are
2541 * still held, otherwise prev could be scheduled on another cpu, die
2542 * there before we look at prev->state, and then the reference would
2544 * Manfred Spraul <manfred@colorfullife.com>
2546 prev_state = prev->state;
2547 finish_arch_switch(prev);
2548 finish_lock_switch(rq, prev);
2550 if (current->sched_class->post_schedule)
2551 current->sched_class->post_schedule(rq);
2554 fire_sched_in_preempt_notifiers(current);
2557 if (unlikely(prev_state == TASK_DEAD)) {
2559 * Remove function-return probe instances associated with this
2560 * task and put them back on the free list.
2562 kprobe_flush_task(prev);
2563 put_task_struct(prev);
2568 * schedule_tail - first thing a freshly forked thread must call.
2569 * @prev: the thread we just switched away from.
2571 asmlinkage void schedule_tail(struct task_struct *prev)
2572 __releases(rq->lock)
2574 struct rq *rq = this_rq();
2576 finish_task_switch(rq, prev);
2577 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2578 /* In this case, finish_task_switch does not reenable preemption */
2581 if (current->set_child_tid)
2582 put_user(task_pid_vnr(current), current->set_child_tid);
2586 * context_switch - switch to the new MM and the new
2587 * thread's register state.
2590 context_switch(struct rq *rq, struct task_struct *prev,
2591 struct task_struct *next)
2593 struct mm_struct *mm, *oldmm;
2595 prepare_task_switch(rq, prev, next);
2597 oldmm = prev->active_mm;
2599 * For paravirt, this is coupled with an exit in switch_to to
2600 * combine the page table reload and the switch backend into
2603 arch_enter_lazy_cpu_mode();
2605 if (unlikely(!mm)) {
2606 next->active_mm = oldmm;
2607 atomic_inc(&oldmm->mm_count);
2608 enter_lazy_tlb(oldmm, next);
2610 switch_mm(oldmm, mm, next);
2612 if (unlikely(!prev->mm)) {
2613 prev->active_mm = NULL;
2614 rq->prev_mm = oldmm;
2617 * Since the runqueue lock will be released by the next
2618 * task (which is an invalid locking op but in the case
2619 * of the scheduler it's an obvious special-case), so we
2620 * do an early lockdep release here:
2622 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2623 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2626 /* Here we just switch the register state and the stack. */
2627 switch_to(prev, next, prev);
2631 * this_rq must be evaluated again because prev may have moved
2632 * CPUs since it called schedule(), thus the 'rq' on its stack
2633 * frame will be invalid.
2635 finish_task_switch(this_rq(), prev);
2639 * nr_running, nr_uninterruptible and nr_context_switches:
2641 * externally visible scheduler statistics: current number of runnable
2642 * threads, current number of uninterruptible-sleeping threads, total
2643 * number of context switches performed since bootup.
2645 unsigned long nr_running(void)
2647 unsigned long i, sum = 0;
2649 for_each_online_cpu(i)
2650 sum += cpu_rq(i)->nr_running;
2655 unsigned long nr_uninterruptible(void)
2657 unsigned long i, sum = 0;
2659 for_each_possible_cpu(i)
2660 sum += cpu_rq(i)->nr_uninterruptible;
2663 * Since we read the counters lockless, it might be slightly
2664 * inaccurate. Do not allow it to go below zero though:
2666 if (unlikely((long)sum < 0))
2672 unsigned long long nr_context_switches(void)
2675 unsigned long long sum = 0;
2677 for_each_possible_cpu(i)
2678 sum += cpu_rq(i)->nr_switches;
2683 unsigned long nr_iowait(void)
2685 unsigned long i, sum = 0;
2687 for_each_possible_cpu(i)
2688 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2693 unsigned long nr_active(void)
2695 unsigned long i, running = 0, uninterruptible = 0;
2697 for_each_online_cpu(i) {
2698 running += cpu_rq(i)->nr_running;
2699 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2702 if (unlikely((long)uninterruptible < 0))
2703 uninterruptible = 0;
2705 return running + uninterruptible;
2709 * Update rq->cpu_load[] statistics. This function is usually called every
2710 * scheduler tick (TICK_NSEC).
2712 static void update_cpu_load(struct rq *this_rq)
2714 unsigned long this_load = this_rq->load.weight;
2717 this_rq->nr_load_updates++;
2719 /* Update our load: */
2720 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2721 unsigned long old_load, new_load;
2723 /* scale is effectively 1 << i now, and >> i divides by scale */
2725 old_load = this_rq->cpu_load[i];
2726 new_load = this_load;
2728 * Round up the averaging division if load is increasing. This
2729 * prevents us from getting stuck on 9 if the load is 10, for
2732 if (new_load > old_load)
2733 new_load += scale-1;
2734 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2741 * double_rq_lock - safely lock two runqueues
2743 * Note this does not disable interrupts like task_rq_lock,
2744 * you need to do so manually before calling.
2746 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2747 __acquires(rq1->lock)
2748 __acquires(rq2->lock)
2750 BUG_ON(!irqs_disabled());
2752 spin_lock(&rq1->lock);
2753 __acquire(rq2->lock); /* Fake it out ;) */
2756 spin_lock(&rq1->lock);
2757 spin_lock(&rq2->lock);
2759 spin_lock(&rq2->lock);
2760 spin_lock(&rq1->lock);
2763 update_rq_clock(rq1);
2764 update_rq_clock(rq2);
2768 * double_rq_unlock - safely unlock two runqueues
2770 * Note this does not restore interrupts like task_rq_unlock,
2771 * you need to do so manually after calling.
2773 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2774 __releases(rq1->lock)
2775 __releases(rq2->lock)
2777 spin_unlock(&rq1->lock);
2779 spin_unlock(&rq2->lock);
2781 __release(rq2->lock);
2785 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2787 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2788 __releases(this_rq->lock)
2789 __acquires(busiest->lock)
2790 __acquires(this_rq->lock)
2794 if (unlikely(!irqs_disabled())) {
2795 /* printk() doesn't work good under rq->lock */
2796 spin_unlock(&this_rq->lock);
2799 if (unlikely(!spin_trylock(&busiest->lock))) {
2800 if (busiest < this_rq) {
2801 spin_unlock(&this_rq->lock);
2802 spin_lock(&busiest->lock);
2803 spin_lock(&this_rq->lock);
2806 spin_lock(&busiest->lock);
2812 * If dest_cpu is allowed for this process, migrate the task to it.
2813 * This is accomplished by forcing the cpu_allowed mask to only
2814 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2815 * the cpu_allowed mask is restored.
2817 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2819 struct migration_req req;
2820 unsigned long flags;
2823 rq = task_rq_lock(p, &flags);
2824 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2825 || unlikely(cpu_is_offline(dest_cpu)))
2828 /* force the process onto the specified CPU */
2829 if (migrate_task(p, dest_cpu, &req)) {
2830 /* Need to wait for migration thread (might exit: take ref). */
2831 struct task_struct *mt = rq->migration_thread;
2833 get_task_struct(mt);
2834 task_rq_unlock(rq, &flags);
2835 wake_up_process(mt);
2836 put_task_struct(mt);
2837 wait_for_completion(&req.done);
2842 task_rq_unlock(rq, &flags);
2846 * sched_exec - execve() is a valuable balancing opportunity, because at
2847 * this point the task has the smallest effective memory and cache footprint.
2849 void sched_exec(void)
2851 int new_cpu, this_cpu = get_cpu();
2852 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2854 if (new_cpu != this_cpu)
2855 sched_migrate_task(current, new_cpu);
2859 * pull_task - move a task from a remote runqueue to the local runqueue.
2860 * Both runqueues must be locked.
2862 static void pull_task(struct rq *src_rq, struct task_struct *p,
2863 struct rq *this_rq, int this_cpu)
2865 deactivate_task(src_rq, p, 0);
2866 set_task_cpu(p, this_cpu);
2867 activate_task(this_rq, p, 0);
2869 * Note that idle threads have a prio of MAX_PRIO, for this test
2870 * to be always true for them.
2872 check_preempt_curr(this_rq, p);
2876 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2879 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2880 struct sched_domain *sd, enum cpu_idle_type idle,
2884 * We do not migrate tasks that are:
2885 * 1) running (obviously), or
2886 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2887 * 3) are cache-hot on their current CPU.
2889 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2890 schedstat_inc(p, se.nr_failed_migrations_affine);
2895 if (task_running(rq, p)) {
2896 schedstat_inc(p, se.nr_failed_migrations_running);
2901 * Aggressive migration if:
2902 * 1) task is cache cold, or
2903 * 2) too many balance attempts have failed.
2906 if (!task_hot(p, rq->clock, sd) ||
2907 sd->nr_balance_failed > sd->cache_nice_tries) {
2908 #ifdef CONFIG_SCHEDSTATS
2909 if (task_hot(p, rq->clock, sd)) {
2910 schedstat_inc(sd, lb_hot_gained[idle]);
2911 schedstat_inc(p, se.nr_forced_migrations);
2917 if (task_hot(p, rq->clock, sd)) {
2918 schedstat_inc(p, se.nr_failed_migrations_hot);
2924 static unsigned long
2925 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2926 unsigned long max_load_move, struct sched_domain *sd,
2927 enum cpu_idle_type idle, int *all_pinned,
2928 int *this_best_prio, struct rq_iterator *iterator)
2930 int loops = 0, pulled = 0, pinned = 0;
2931 struct task_struct *p;
2932 long rem_load_move = max_load_move;
2934 if (max_load_move == 0)
2940 * Start the load-balancing iterator:
2942 p = iterator->start(iterator->arg);
2944 if (!p || loops++ > sysctl_sched_nr_migrate)
2947 if ((p->se.load.weight >> 1) > rem_load_move ||
2948 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2949 p = iterator->next(iterator->arg);
2953 pull_task(busiest, p, this_rq, this_cpu);
2955 rem_load_move -= p->se.load.weight;
2958 * We only want to steal up to the prescribed amount of weighted load.
2960 if (rem_load_move > 0) {
2961 if (p->prio < *this_best_prio)
2962 *this_best_prio = p->prio;
2963 p = iterator->next(iterator->arg);
2968 * Right now, this is one of only two places pull_task() is called,
2969 * so we can safely collect pull_task() stats here rather than
2970 * inside pull_task().
2972 schedstat_add(sd, lb_gained[idle], pulled);
2975 *all_pinned = pinned;
2977 return max_load_move - rem_load_move;
2981 * move_tasks tries to move up to max_load_move weighted load from busiest to
2982 * this_rq, as part of a balancing operation within domain "sd".
2983 * Returns 1 if successful and 0 otherwise.
2985 * Called with both runqueues locked.
2987 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2988 unsigned long max_load_move,
2989 struct sched_domain *sd, enum cpu_idle_type idle,
2992 const struct sched_class *class = sched_class_highest;
2993 unsigned long total_load_moved = 0;
2994 int this_best_prio = this_rq->curr->prio;
2998 class->load_balance(this_rq, this_cpu, busiest,
2999 max_load_move - total_load_moved,
3000 sd, idle, all_pinned, &this_best_prio);
3001 class = class->next;
3002 } while (class && max_load_move > total_load_moved);
3004 return total_load_moved > 0;
3008 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3009 struct sched_domain *sd, enum cpu_idle_type idle,
3010 struct rq_iterator *iterator)
3012 struct task_struct *p = iterator->start(iterator->arg);
3016 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3017 pull_task(busiest, p, this_rq, this_cpu);
3019 * Right now, this is only the second place pull_task()
3020 * is called, so we can safely collect pull_task()
3021 * stats here rather than inside pull_task().
3023 schedstat_inc(sd, lb_gained[idle]);
3027 p = iterator->next(iterator->arg);
3034 * move_one_task tries to move exactly one task from busiest to this_rq, as
3035 * part of active balancing operations within "domain".
3036 * Returns 1 if successful and 0 otherwise.
3038 * Called with both runqueues locked.
3040 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3041 struct sched_domain *sd, enum cpu_idle_type idle)
3043 const struct sched_class *class;
3045 for (class = sched_class_highest; class; class = class->next)
3046 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3053 * find_busiest_group finds and returns the busiest CPU group within the
3054 * domain. It calculates and returns the amount of weighted load which
3055 * should be moved to restore balance via the imbalance parameter.
3057 static struct sched_group *
3058 find_busiest_group(struct sched_domain *sd, int this_cpu,
3059 unsigned long *imbalance, enum cpu_idle_type idle,
3060 int *sd_idle, const cpumask_t *cpus, int *balance)
3062 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3063 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3064 unsigned long max_pull;
3065 unsigned long busiest_load_per_task, busiest_nr_running;
3066 unsigned long this_load_per_task, this_nr_running;
3067 int load_idx, group_imb = 0;
3068 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3069 int power_savings_balance = 1;
3070 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3071 unsigned long min_nr_running = ULONG_MAX;
3072 struct sched_group *group_min = NULL, *group_leader = NULL;
3075 max_load = this_load = total_load = total_pwr = 0;
3076 busiest_load_per_task = busiest_nr_running = 0;
3077 this_load_per_task = this_nr_running = 0;
3079 if (idle == CPU_NOT_IDLE)
3080 load_idx = sd->busy_idx;
3081 else if (idle == CPU_NEWLY_IDLE)
3082 load_idx = sd->newidle_idx;
3084 load_idx = sd->idle_idx;
3087 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3090 int __group_imb = 0;
3091 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3092 unsigned long sum_nr_running, sum_weighted_load;
3093 unsigned long sum_avg_load_per_task;
3094 unsigned long avg_load_per_task;
3096 local_group = cpu_isset(this_cpu, group->cpumask);
3099 balance_cpu = first_cpu(group->cpumask);
3101 /* Tally up the load of all CPUs in the group */
3102 sum_weighted_load = sum_nr_running = avg_load = 0;
3103 sum_avg_load_per_task = avg_load_per_task = 0;
3106 min_cpu_load = ~0UL;
3108 for_each_cpu_mask(i, group->cpumask) {
3111 if (!cpu_isset(i, *cpus))
3116 if (*sd_idle && rq->nr_running)
3119 /* Bias balancing toward cpus of our domain */
3121 if (idle_cpu(i) && !first_idle_cpu) {
3126 load = target_load(i, load_idx);
3128 load = source_load(i, load_idx);
3129 if (load > max_cpu_load)
3130 max_cpu_load = load;
3131 if (min_cpu_load > load)
3132 min_cpu_load = load;
3136 sum_nr_running += rq->nr_running;
3137 sum_weighted_load += weighted_cpuload(i);
3139 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3143 * First idle cpu or the first cpu(busiest) in this sched group
3144 * is eligible for doing load balancing at this and above
3145 * domains. In the newly idle case, we will allow all the cpu's
3146 * to do the newly idle load balance.
3148 if (idle != CPU_NEWLY_IDLE && local_group &&
3149 balance_cpu != this_cpu && balance) {
3154 total_load += avg_load;
3155 total_pwr += group->__cpu_power;
3157 /* Adjust by relative CPU power of the group */
3158 avg_load = sg_div_cpu_power(group,
3159 avg_load * SCHED_LOAD_SCALE);
3163 * Consider the group unbalanced when the imbalance is larger
3164 * than the average weight of two tasks.
3166 * APZ: with cgroup the avg task weight can vary wildly and
3167 * might not be a suitable number - should we keep a
3168 * normalized nr_running number somewhere that negates
3171 avg_load_per_task = sg_div_cpu_power(group,
3172 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3174 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3177 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3180 this_load = avg_load;
3182 this_nr_running = sum_nr_running;
3183 this_load_per_task = sum_weighted_load;
3184 } else if (avg_load > max_load &&
3185 (sum_nr_running > group_capacity || __group_imb)) {
3186 max_load = avg_load;
3188 busiest_nr_running = sum_nr_running;
3189 busiest_load_per_task = sum_weighted_load;
3190 group_imb = __group_imb;
3193 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3195 * Busy processors will not participate in power savings
3198 if (idle == CPU_NOT_IDLE ||
3199 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3203 * If the local group is idle or completely loaded
3204 * no need to do power savings balance at this domain
3206 if (local_group && (this_nr_running >= group_capacity ||
3208 power_savings_balance = 0;
3211 * If a group is already running at full capacity or idle,
3212 * don't include that group in power savings calculations
3214 if (!power_savings_balance || sum_nr_running >= group_capacity
3219 * Calculate the group which has the least non-idle load.
3220 * This is the group from where we need to pick up the load
3223 if ((sum_nr_running < min_nr_running) ||
3224 (sum_nr_running == min_nr_running &&
3225 first_cpu(group->cpumask) <
3226 first_cpu(group_min->cpumask))) {
3228 min_nr_running = sum_nr_running;
3229 min_load_per_task = sum_weighted_load /
3234 * Calculate the group which is almost near its
3235 * capacity but still has some space to pick up some load
3236 * from other group and save more power
3238 if (sum_nr_running <= group_capacity - 1) {
3239 if (sum_nr_running > leader_nr_running ||
3240 (sum_nr_running == leader_nr_running &&
3241 first_cpu(group->cpumask) >
3242 first_cpu(group_leader->cpumask))) {
3243 group_leader = group;
3244 leader_nr_running = sum_nr_running;
3249 group = group->next;
3250 } while (group != sd->groups);
3252 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3255 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3257 if (this_load >= avg_load ||
3258 100*max_load <= sd->imbalance_pct*this_load)
3261 busiest_load_per_task /= busiest_nr_running;
3263 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3266 * We're trying to get all the cpus to the average_load, so we don't
3267 * want to push ourselves above the average load, nor do we wish to
3268 * reduce the max loaded cpu below the average load, as either of these
3269 * actions would just result in more rebalancing later, and ping-pong
3270 * tasks around. Thus we look for the minimum possible imbalance.
3271 * Negative imbalances (*we* are more loaded than anyone else) will
3272 * be counted as no imbalance for these purposes -- we can't fix that
3273 * by pulling tasks to us. Be careful of negative numbers as they'll
3274 * appear as very large values with unsigned longs.
3276 if (max_load <= busiest_load_per_task)
3280 * In the presence of smp nice balancing, certain scenarios can have
3281 * max load less than avg load(as we skip the groups at or below
3282 * its cpu_power, while calculating max_load..)
3284 if (max_load < avg_load) {
3286 goto small_imbalance;
3289 /* Don't want to pull so many tasks that a group would go idle */
3290 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3292 /* How much load to actually move to equalise the imbalance */
3293 *imbalance = min(max_pull * busiest->__cpu_power,
3294 (avg_load - this_load) * this->__cpu_power)
3298 * if *imbalance is less than the average load per runnable task
3299 * there is no gaurantee that any tasks will be moved so we'll have
3300 * a think about bumping its value to force at least one task to be
3303 if (*imbalance < busiest_load_per_task) {
3304 unsigned long tmp, pwr_now, pwr_move;
3308 pwr_move = pwr_now = 0;
3310 if (this_nr_running) {
3311 this_load_per_task /= this_nr_running;
3312 if (busiest_load_per_task > this_load_per_task)
3315 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3317 if (max_load - this_load + 2*busiest_load_per_task >=
3318 busiest_load_per_task * imbn) {
3319 *imbalance = busiest_load_per_task;
3324 * OK, we don't have enough imbalance to justify moving tasks,
3325 * however we may be able to increase total CPU power used by
3329 pwr_now += busiest->__cpu_power *
3330 min(busiest_load_per_task, max_load);
3331 pwr_now += this->__cpu_power *
3332 min(this_load_per_task, this_load);
3333 pwr_now /= SCHED_LOAD_SCALE;
3335 /* Amount of load we'd subtract */
3336 tmp = sg_div_cpu_power(busiest,
3337 busiest_load_per_task * SCHED_LOAD_SCALE);
3339 pwr_move += busiest->__cpu_power *
3340 min(busiest_load_per_task, max_load - tmp);
3342 /* Amount of load we'd add */
3343 if (max_load * busiest->__cpu_power <
3344 busiest_load_per_task * SCHED_LOAD_SCALE)
3345 tmp = sg_div_cpu_power(this,
3346 max_load * busiest->__cpu_power);
3348 tmp = sg_div_cpu_power(this,
3349 busiest_load_per_task * SCHED_LOAD_SCALE);
3350 pwr_move += this->__cpu_power *
3351 min(this_load_per_task, this_load + tmp);
3352 pwr_move /= SCHED_LOAD_SCALE;
3354 /* Move if we gain throughput */
3355 if (pwr_move > pwr_now)
3356 *imbalance = busiest_load_per_task;
3362 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3363 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3366 if (this == group_leader && group_leader != group_min) {
3367 *imbalance = min_load_per_task;
3377 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3380 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3381 unsigned long imbalance, const cpumask_t *cpus)
3383 struct rq *busiest = NULL, *rq;
3384 unsigned long max_load = 0;
3387 for_each_cpu_mask(i, group->cpumask) {
3390 if (!cpu_isset(i, *cpus))
3394 wl = weighted_cpuload(i);
3396 if (rq->nr_running == 1 && wl > imbalance)
3399 if (wl > max_load) {
3409 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3410 * so long as it is large enough.
3412 #define MAX_PINNED_INTERVAL 512
3415 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3416 * tasks if there is an imbalance.
3418 static int load_balance(int this_cpu, struct rq *this_rq,
3419 struct sched_domain *sd, enum cpu_idle_type idle,
3420 int *balance, cpumask_t *cpus)
3422 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3423 struct sched_group *group;
3424 unsigned long imbalance;
3426 unsigned long flags;
3431 * When power savings policy is enabled for the parent domain, idle
3432 * sibling can pick up load irrespective of busy siblings. In this case,
3433 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3434 * portraying it as CPU_NOT_IDLE.
3436 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3437 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3440 schedstat_inc(sd, lb_count[idle]);
3444 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3451 schedstat_inc(sd, lb_nobusyg[idle]);
3455 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3457 schedstat_inc(sd, lb_nobusyq[idle]);
3461 BUG_ON(busiest == this_rq);
3463 schedstat_add(sd, lb_imbalance[idle], imbalance);
3466 if (busiest->nr_running > 1) {
3468 * Attempt to move tasks. If find_busiest_group has found
3469 * an imbalance but busiest->nr_running <= 1, the group is
3470 * still unbalanced. ld_moved simply stays zero, so it is
3471 * correctly treated as an imbalance.
3473 local_irq_save(flags);
3474 double_rq_lock(this_rq, busiest);
3475 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3476 imbalance, sd, idle, &all_pinned);
3477 double_rq_unlock(this_rq, busiest);
3478 local_irq_restore(flags);
3481 * some other cpu did the load balance for us.
3483 if (ld_moved && this_cpu != smp_processor_id())
3484 resched_cpu(this_cpu);
3486 /* All tasks on this runqueue were pinned by CPU affinity */
3487 if (unlikely(all_pinned)) {
3488 cpu_clear(cpu_of(busiest), *cpus);
3489 if (!cpus_empty(*cpus))
3496 schedstat_inc(sd, lb_failed[idle]);
3497 sd->nr_balance_failed++;
3499 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3501 spin_lock_irqsave(&busiest->lock, flags);
3503 /* don't kick the migration_thread, if the curr
3504 * task on busiest cpu can't be moved to this_cpu
3506 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3507 spin_unlock_irqrestore(&busiest->lock, flags);
3509 goto out_one_pinned;
3512 if (!busiest->active_balance) {
3513 busiest->active_balance = 1;
3514 busiest->push_cpu = this_cpu;
3517 spin_unlock_irqrestore(&busiest->lock, flags);
3519 wake_up_process(busiest->migration_thread);
3522 * We've kicked active balancing, reset the failure
3525 sd->nr_balance_failed = sd->cache_nice_tries+1;
3528 sd->nr_balance_failed = 0;
3530 if (likely(!active_balance)) {
3531 /* We were unbalanced, so reset the balancing interval */
3532 sd->balance_interval = sd->min_interval;
3535 * If we've begun active balancing, start to back off. This
3536 * case may not be covered by the all_pinned logic if there
3537 * is only 1 task on the busy runqueue (because we don't call
3540 if (sd->balance_interval < sd->max_interval)
3541 sd->balance_interval *= 2;
3544 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3545 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3551 schedstat_inc(sd, lb_balanced[idle]);
3553 sd->nr_balance_failed = 0;
3556 /* tune up the balancing interval */
3557 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3558 (sd->balance_interval < sd->max_interval))
3559 sd->balance_interval *= 2;
3561 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3562 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3573 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3574 * tasks if there is an imbalance.
3576 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3577 * this_rq is locked.
3580 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3583 struct sched_group *group;
3584 struct rq *busiest = NULL;
3585 unsigned long imbalance;
3593 * When power savings policy is enabled for the parent domain, idle
3594 * sibling can pick up load irrespective of busy siblings. In this case,
3595 * let the state of idle sibling percolate up as IDLE, instead of
3596 * portraying it as CPU_NOT_IDLE.
3598 if (sd->flags & SD_SHARE_CPUPOWER &&
3599 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3602 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3604 update_shares_locked(this_rq, sd);
3605 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3606 &sd_idle, cpus, NULL);
3608 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3612 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3614 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3618 BUG_ON(busiest == this_rq);
3620 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3623 if (busiest->nr_running > 1) {
3624 /* Attempt to move tasks */
3625 double_lock_balance(this_rq, busiest);
3626 /* this_rq->clock is already updated */
3627 update_rq_clock(busiest);
3628 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3629 imbalance, sd, CPU_NEWLY_IDLE,
3631 spin_unlock(&busiest->lock);
3633 if (unlikely(all_pinned)) {
3634 cpu_clear(cpu_of(busiest), *cpus);
3635 if (!cpus_empty(*cpus))
3641 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3642 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3643 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3646 sd->nr_balance_failed = 0;
3648 update_shares_locked(this_rq, sd);
3652 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3653 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3654 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3656 sd->nr_balance_failed = 0;
3662 * idle_balance is called by schedule() if this_cpu is about to become
3663 * idle. Attempts to pull tasks from other CPUs.
3665 static void idle_balance(int this_cpu, struct rq *this_rq)
3667 struct sched_domain *sd;
3668 int pulled_task = -1;
3669 unsigned long next_balance = jiffies + HZ;
3672 for_each_domain(this_cpu, sd) {
3673 unsigned long interval;
3675 if (!(sd->flags & SD_LOAD_BALANCE))
3678 if (sd->flags & SD_BALANCE_NEWIDLE)
3679 /* If we've pulled tasks over stop searching: */
3680 pulled_task = load_balance_newidle(this_cpu, this_rq,
3683 interval = msecs_to_jiffies(sd->balance_interval);
3684 if (time_after(next_balance, sd->last_balance + interval))
3685 next_balance = sd->last_balance + interval;
3689 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3691 * We are going idle. next_balance may be set based on
3692 * a busy processor. So reset next_balance.
3694 this_rq->next_balance = next_balance;
3699 * active_load_balance is run by migration threads. It pushes running tasks
3700 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3701 * running on each physical CPU where possible, and avoids physical /
3702 * logical imbalances.
3704 * Called with busiest_rq locked.
3706 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3708 int target_cpu = busiest_rq->push_cpu;
3709 struct sched_domain *sd;
3710 struct rq *target_rq;
3712 /* Is there any task to move? */
3713 if (busiest_rq->nr_running <= 1)
3716 target_rq = cpu_rq(target_cpu);
3719 * This condition is "impossible", if it occurs
3720 * we need to fix it. Originally reported by
3721 * Bjorn Helgaas on a 128-cpu setup.
3723 BUG_ON(busiest_rq == target_rq);
3725 /* move a task from busiest_rq to target_rq */
3726 double_lock_balance(busiest_rq, target_rq);
3727 update_rq_clock(busiest_rq);
3728 update_rq_clock(target_rq);
3730 /* Search for an sd spanning us and the target CPU. */
3731 for_each_domain(target_cpu, sd) {
3732 if ((sd->flags & SD_LOAD_BALANCE) &&
3733 cpu_isset(busiest_cpu, sd->span))
3738 schedstat_inc(sd, alb_count);
3740 if (move_one_task(target_rq, target_cpu, busiest_rq,
3742 schedstat_inc(sd, alb_pushed);
3744 schedstat_inc(sd, alb_failed);
3746 spin_unlock(&target_rq->lock);
3751 atomic_t load_balancer;
3753 } nohz ____cacheline_aligned = {
3754 .load_balancer = ATOMIC_INIT(-1),
3755 .cpu_mask = CPU_MASK_NONE,
3759 * This routine will try to nominate the ilb (idle load balancing)
3760 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3761 * load balancing on behalf of all those cpus. If all the cpus in the system
3762 * go into this tickless mode, then there will be no ilb owner (as there is
3763 * no need for one) and all the cpus will sleep till the next wakeup event
3766 * For the ilb owner, tick is not stopped. And this tick will be used
3767 * for idle load balancing. ilb owner will still be part of
3770 * While stopping the tick, this cpu will become the ilb owner if there
3771 * is no other owner. And will be the owner till that cpu becomes busy
3772 * or if all cpus in the system stop their ticks at which point
3773 * there is no need for ilb owner.
3775 * When the ilb owner becomes busy, it nominates another owner, during the
3776 * next busy scheduler_tick()
3778 int select_nohz_load_balancer(int stop_tick)
3780 int cpu = smp_processor_id();
3783 cpu_set(cpu, nohz.cpu_mask);
3784 cpu_rq(cpu)->in_nohz_recently = 1;
3787 * If we are going offline and still the leader, give up!
3789 if (cpu_is_offline(cpu) &&
3790 atomic_read(&nohz.load_balancer) == cpu) {
3791 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3796 /* time for ilb owner also to sleep */
3797 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3798 if (atomic_read(&nohz.load_balancer) == cpu)
3799 atomic_set(&nohz.load_balancer, -1);
3803 if (atomic_read(&nohz.load_balancer) == -1) {
3804 /* make me the ilb owner */
3805 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3807 } else if (atomic_read(&nohz.load_balancer) == cpu)
3810 if (!cpu_isset(cpu, nohz.cpu_mask))
3813 cpu_clear(cpu, nohz.cpu_mask);
3815 if (atomic_read(&nohz.load_balancer) == cpu)
3816 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3823 static DEFINE_SPINLOCK(balancing);
3826 * It checks each scheduling domain to see if it is due to be balanced,
3827 * and initiates a balancing operation if so.
3829 * Balancing parameters are set up in arch_init_sched_domains.
3831 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3834 struct rq *rq = cpu_rq(cpu);
3835 unsigned long interval;
3836 struct sched_domain *sd;
3837 /* Earliest time when we have to do rebalance again */
3838 unsigned long next_balance = jiffies + 60*HZ;
3839 int update_next_balance = 0;
3843 for_each_domain(cpu, sd) {
3844 if (!(sd->flags & SD_LOAD_BALANCE))
3847 interval = sd->balance_interval;
3848 if (idle != CPU_IDLE)
3849 interval *= sd->busy_factor;
3851 /* scale ms to jiffies */
3852 interval = msecs_to_jiffies(interval);
3853 if (unlikely(!interval))
3855 if (interval > HZ*NR_CPUS/10)
3856 interval = HZ*NR_CPUS/10;
3858 need_serialize = sd->flags & SD_SERIALIZE;
3860 if (need_serialize) {
3861 if (!spin_trylock(&balancing))
3865 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3866 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3868 * We've pulled tasks over so either we're no
3869 * longer idle, or one of our SMT siblings is
3872 idle = CPU_NOT_IDLE;
3874 sd->last_balance = jiffies;
3877 spin_unlock(&balancing);
3879 if (time_after(next_balance, sd->last_balance + interval)) {
3880 next_balance = sd->last_balance + interval;
3881 update_next_balance = 1;
3885 * Stop the load balance at this level. There is another
3886 * CPU in our sched group which is doing load balancing more
3894 * next_balance will be updated only when there is a need.
3895 * When the cpu is attached to null domain for ex, it will not be
3898 if (likely(update_next_balance))
3899 rq->next_balance = next_balance;
3903 * run_rebalance_domains is triggered when needed from the scheduler tick.
3904 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3905 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3907 static void run_rebalance_domains(struct softirq_action *h)
3909 int this_cpu = smp_processor_id();
3910 struct rq *this_rq = cpu_rq(this_cpu);
3911 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3912 CPU_IDLE : CPU_NOT_IDLE;
3914 rebalance_domains(this_cpu, idle);
3918 * If this cpu is the owner for idle load balancing, then do the
3919 * balancing on behalf of the other idle cpus whose ticks are
3922 if (this_rq->idle_at_tick &&
3923 atomic_read(&nohz.load_balancer) == this_cpu) {
3924 cpumask_t cpus = nohz.cpu_mask;
3928 cpu_clear(this_cpu, cpus);
3929 for_each_cpu_mask(balance_cpu, cpus) {
3931 * If this cpu gets work to do, stop the load balancing
3932 * work being done for other cpus. Next load
3933 * balancing owner will pick it up.
3938 rebalance_domains(balance_cpu, CPU_IDLE);
3940 rq = cpu_rq(balance_cpu);
3941 if (time_after(this_rq->next_balance, rq->next_balance))
3942 this_rq->next_balance = rq->next_balance;
3949 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3951 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3952 * idle load balancing owner or decide to stop the periodic load balancing,
3953 * if the whole system is idle.
3955 static inline void trigger_load_balance(struct rq *rq, int cpu)
3959 * If we were in the nohz mode recently and busy at the current
3960 * scheduler tick, then check if we need to nominate new idle
3963 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3964 rq->in_nohz_recently = 0;
3966 if (atomic_read(&nohz.load_balancer) == cpu) {
3967 cpu_clear(cpu, nohz.cpu_mask);
3968 atomic_set(&nohz.load_balancer, -1);
3971 if (atomic_read(&nohz.load_balancer) == -1) {
3973 * simple selection for now: Nominate the
3974 * first cpu in the nohz list to be the next
3977 * TBD: Traverse the sched domains and nominate
3978 * the nearest cpu in the nohz.cpu_mask.
3980 int ilb = first_cpu(nohz.cpu_mask);
3982 if (ilb < nr_cpu_ids)
3988 * If this cpu is idle and doing idle load balancing for all the
3989 * cpus with ticks stopped, is it time for that to stop?
3991 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3992 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3998 * If this cpu is idle and the idle load balancing is done by
3999 * someone else, then no need raise the SCHED_SOFTIRQ
4001 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4002 cpu_isset(cpu, nohz.cpu_mask))
4005 if (time_after_eq(jiffies, rq->next_balance))
4006 raise_softirq(SCHED_SOFTIRQ);
4009 #else /* CONFIG_SMP */
4012 * on UP we do not need to balance between CPUs:
4014 static inline void idle_balance(int cpu, struct rq *rq)
4020 DEFINE_PER_CPU(struct kernel_stat, kstat);
4022 EXPORT_PER_CPU_SYMBOL(kstat);
4025 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4026 * that have not yet been banked in case the task is currently running.
4028 unsigned long long task_sched_runtime(struct task_struct *p)
4030 unsigned long flags;
4034 rq = task_rq_lock(p, &flags);
4035 ns = p->se.sum_exec_runtime;
4036 if (task_current(rq, p)) {
4037 update_rq_clock(rq);
4038 delta_exec = rq->clock - p->se.exec_start;
4039 if ((s64)delta_exec > 0)
4042 task_rq_unlock(rq, &flags);
4048 * Account user cpu time to a process.
4049 * @p: the process that the cpu time gets accounted to
4050 * @cputime: the cpu time spent in user space since the last update
4052 void account_user_time(struct task_struct *p, cputime_t cputime)
4054 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4057 p->utime = cputime_add(p->utime, cputime);
4059 /* Add user time to cpustat. */
4060 tmp = cputime_to_cputime64(cputime);
4061 if (TASK_NICE(p) > 0)
4062 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4064 cpustat->user = cputime64_add(cpustat->user, tmp);
4068 * Account guest cpu time to a process.
4069 * @p: the process that the cpu time gets accounted to
4070 * @cputime: the cpu time spent in virtual machine since the last update
4072 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4075 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4077 tmp = cputime_to_cputime64(cputime);
4079 p->utime = cputime_add(p->utime, cputime);
4080 p->gtime = cputime_add(p->gtime, cputime);
4082 cpustat->user = cputime64_add(cpustat->user, tmp);
4083 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4087 * Account scaled user cpu time to a process.
4088 * @p: the process that the cpu time gets accounted to
4089 * @cputime: the cpu time spent in user space since the last update
4091 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4093 p->utimescaled = cputime_add(p->utimescaled, cputime);
4097 * Account system cpu time to a process.
4098 * @p: the process that the cpu time gets accounted to
4099 * @hardirq_offset: the offset to subtract from hardirq_count()
4100 * @cputime: the cpu time spent in kernel space since the last update
4102 void account_system_time(struct task_struct *p, int hardirq_offset,
4105 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4106 struct rq *rq = this_rq();
4109 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4110 account_guest_time(p, cputime);
4114 p->stime = cputime_add(p->stime, cputime);
4116 /* Add system time to cpustat. */
4117 tmp = cputime_to_cputime64(cputime);
4118 if (hardirq_count() - hardirq_offset)
4119 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4120 else if (softirq_count())
4121 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4122 else if (p != rq->idle)
4123 cpustat->system = cputime64_add(cpustat->system, tmp);
4124 else if (atomic_read(&rq->nr_iowait) > 0)
4125 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4127 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4128 /* Account for system time used */
4129 acct_update_integrals(p);
4133 * Account scaled system cpu time to a process.
4134 * @p: the process that the cpu time gets accounted to
4135 * @hardirq_offset: the offset to subtract from hardirq_count()
4136 * @cputime: the cpu time spent in kernel space since the last update
4138 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4140 p->stimescaled = cputime_add(p->stimescaled, cputime);
4144 * Account for involuntary wait time.
4145 * @p: the process from which the cpu time has been stolen
4146 * @steal: the cpu time spent in involuntary wait
4148 void account_steal_time(struct task_struct *p, cputime_t steal)
4150 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4151 cputime64_t tmp = cputime_to_cputime64(steal);
4152 struct rq *rq = this_rq();
4154 if (p == rq->idle) {
4155 p->stime = cputime_add(p->stime, steal);
4156 if (atomic_read(&rq->nr_iowait) > 0)
4157 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4159 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4161 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4165 * This function gets called by the timer code, with HZ frequency.
4166 * We call it with interrupts disabled.
4168 * It also gets called by the fork code, when changing the parent's
4171 void scheduler_tick(void)
4173 int cpu = smp_processor_id();
4174 struct rq *rq = cpu_rq(cpu);
4175 struct task_struct *curr = rq->curr;
4179 spin_lock(&rq->lock);
4180 update_rq_clock(rq);
4181 update_cpu_load(rq);
4182 curr->sched_class->task_tick(rq, curr, 0);
4183 spin_unlock(&rq->lock);
4186 rq->idle_at_tick = idle_cpu(cpu);
4187 trigger_load_balance(rq, cpu);
4191 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4193 void __kprobes add_preempt_count(int val)
4198 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4200 preempt_count() += val;
4202 * Spinlock count overflowing soon?
4204 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4207 EXPORT_SYMBOL(add_preempt_count);
4209 void __kprobes sub_preempt_count(int val)
4214 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4217 * Is the spinlock portion underflowing?
4219 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4220 !(preempt_count() & PREEMPT_MASK)))
4223 preempt_count() -= val;
4225 EXPORT_SYMBOL(sub_preempt_count);
4230 * Print scheduling while atomic bug:
4232 static noinline void __schedule_bug(struct task_struct *prev)
4234 struct pt_regs *regs = get_irq_regs();
4236 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4237 prev->comm, prev->pid, preempt_count());
4239 debug_show_held_locks(prev);
4241 if (irqs_disabled())
4242 print_irqtrace_events(prev);
4251 * Various schedule()-time debugging checks and statistics:
4253 static inline void schedule_debug(struct task_struct *prev)
4256 * Test if we are atomic. Since do_exit() needs to call into
4257 * schedule() atomically, we ignore that path for now.
4258 * Otherwise, whine if we are scheduling when we should not be.
4260 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4261 __schedule_bug(prev);
4263 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4265 schedstat_inc(this_rq(), sched_count);
4266 #ifdef CONFIG_SCHEDSTATS
4267 if (unlikely(prev->lock_depth >= 0)) {
4268 schedstat_inc(this_rq(), bkl_count);
4269 schedstat_inc(prev, sched_info.bkl_count);
4275 * Pick up the highest-prio task:
4277 static inline struct task_struct *
4278 pick_next_task(struct rq *rq, struct task_struct *prev)
4280 const struct sched_class *class;
4281 struct task_struct *p;
4284 * Optimization: we know that if all tasks are in
4285 * the fair class we can call that function directly:
4287 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4288 p = fair_sched_class.pick_next_task(rq);
4293 class = sched_class_highest;
4295 p = class->pick_next_task(rq);
4299 * Will never be NULL as the idle class always
4300 * returns a non-NULL p:
4302 class = class->next;
4307 * schedule() is the main scheduler function.
4309 asmlinkage void __sched schedule(void)
4311 struct task_struct *prev, *next;
4312 unsigned long *switch_count;
4314 int cpu, hrtick = sched_feat(HRTICK);
4318 cpu = smp_processor_id();
4322 switch_count = &prev->nivcsw;
4324 release_kernel_lock(prev);
4325 need_resched_nonpreemptible:
4327 schedule_debug(prev);
4333 * Do the rq-clock update outside the rq lock:
4335 local_irq_disable();
4336 update_rq_clock(rq);
4337 spin_lock(&rq->lock);
4338 clear_tsk_need_resched(prev);
4340 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4341 if (unlikely(signal_pending_state(prev->state, prev)))
4342 prev->state = TASK_RUNNING;
4344 deactivate_task(rq, prev, 1);
4345 switch_count = &prev->nvcsw;
4349 if (prev->sched_class->pre_schedule)
4350 prev->sched_class->pre_schedule(rq, prev);
4353 if (unlikely(!rq->nr_running))
4354 idle_balance(cpu, rq);
4356 prev->sched_class->put_prev_task(rq, prev);
4357 next = pick_next_task(rq, prev);
4359 if (likely(prev != next)) {
4360 sched_info_switch(prev, next);
4366 context_switch(rq, prev, next); /* unlocks the rq */
4368 * the context switch might have flipped the stack from under
4369 * us, hence refresh the local variables.
4371 cpu = smp_processor_id();
4374 spin_unlock_irq(&rq->lock);
4379 if (unlikely(reacquire_kernel_lock(current) < 0))
4380 goto need_resched_nonpreemptible;
4382 preempt_enable_no_resched();
4383 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4386 EXPORT_SYMBOL(schedule);
4388 #ifdef CONFIG_PREEMPT
4390 * this is the entry point to schedule() from in-kernel preemption
4391 * off of preempt_enable. Kernel preemptions off return from interrupt
4392 * occur there and call schedule directly.
4394 asmlinkage void __sched preempt_schedule(void)
4396 struct thread_info *ti = current_thread_info();
4399 * If there is a non-zero preempt_count or interrupts are disabled,
4400 * we do not want to preempt the current task. Just return..
4402 if (likely(ti->preempt_count || irqs_disabled()))
4406 add_preempt_count(PREEMPT_ACTIVE);
4408 sub_preempt_count(PREEMPT_ACTIVE);
4411 * Check again in case we missed a preemption opportunity
4412 * between schedule and now.
4415 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4417 EXPORT_SYMBOL(preempt_schedule);
4420 * this is the entry point to schedule() from kernel preemption
4421 * off of irq context.
4422 * Note, that this is called and return with irqs disabled. This will
4423 * protect us against recursive calling from irq.
4425 asmlinkage void __sched preempt_schedule_irq(void)
4427 struct thread_info *ti = current_thread_info();
4429 /* Catch callers which need to be fixed */
4430 BUG_ON(ti->preempt_count || !irqs_disabled());
4433 add_preempt_count(PREEMPT_ACTIVE);
4436 local_irq_disable();
4437 sub_preempt_count(PREEMPT_ACTIVE);
4440 * Check again in case we missed a preemption opportunity
4441 * between schedule and now.
4444 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4447 #endif /* CONFIG_PREEMPT */
4449 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4452 return try_to_wake_up(curr->private, mode, sync);
4454 EXPORT_SYMBOL(default_wake_function);
4457 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4458 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4459 * number) then we wake all the non-exclusive tasks and one exclusive task.
4461 * There are circumstances in which we can try to wake a task which has already
4462 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4463 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4465 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4466 int nr_exclusive, int sync, void *key)
4468 wait_queue_t *curr, *next;
4470 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4471 unsigned flags = curr->flags;
4473 if (curr->func(curr, mode, sync, key) &&
4474 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4480 * __wake_up - wake up threads blocked on a waitqueue.
4482 * @mode: which threads
4483 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4484 * @key: is directly passed to the wakeup function
4486 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4487 int nr_exclusive, void *key)
4489 unsigned long flags;
4491 spin_lock_irqsave(&q->lock, flags);
4492 __wake_up_common(q, mode, nr_exclusive, 0, key);
4493 spin_unlock_irqrestore(&q->lock, flags);
4495 EXPORT_SYMBOL(__wake_up);
4498 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4500 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4502 __wake_up_common(q, mode, 1, 0, NULL);
4506 * __wake_up_sync - wake up threads blocked on a waitqueue.
4508 * @mode: which threads
4509 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4511 * The sync wakeup differs that the waker knows that it will schedule
4512 * away soon, so while the target thread will be woken up, it will not
4513 * be migrated to another CPU - ie. the two threads are 'synchronized'
4514 * with each other. This can prevent needless bouncing between CPUs.
4516 * On UP it can prevent extra preemption.
4519 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4521 unsigned long flags;
4527 if (unlikely(!nr_exclusive))
4530 spin_lock_irqsave(&q->lock, flags);
4531 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4532 spin_unlock_irqrestore(&q->lock, flags);
4534 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4536 void complete(struct completion *x)
4538 unsigned long flags;
4540 spin_lock_irqsave(&x->wait.lock, flags);
4542 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4543 spin_unlock_irqrestore(&x->wait.lock, flags);
4545 EXPORT_SYMBOL(complete);
4547 void complete_all(struct completion *x)
4549 unsigned long flags;
4551 spin_lock_irqsave(&x->wait.lock, flags);
4552 x->done += UINT_MAX/2;
4553 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4554 spin_unlock_irqrestore(&x->wait.lock, flags);
4556 EXPORT_SYMBOL(complete_all);
4558 static inline long __sched
4559 do_wait_for_common(struct completion *x, long timeout, int state)
4562 DECLARE_WAITQUEUE(wait, current);
4564 wait.flags |= WQ_FLAG_EXCLUSIVE;
4565 __add_wait_queue_tail(&x->wait, &wait);
4567 if ((state == TASK_INTERRUPTIBLE &&
4568 signal_pending(current)) ||
4569 (state == TASK_KILLABLE &&
4570 fatal_signal_pending(current))) {
4571 timeout = -ERESTARTSYS;
4574 __set_current_state(state);
4575 spin_unlock_irq(&x->wait.lock);
4576 timeout = schedule_timeout(timeout);
4577 spin_lock_irq(&x->wait.lock);
4578 } while (!x->done && timeout);
4579 __remove_wait_queue(&x->wait, &wait);
4584 return timeout ?: 1;
4588 wait_for_common(struct completion *x, long timeout, int state)
4592 spin_lock_irq(&x->wait.lock);
4593 timeout = do_wait_for_common(x, timeout, state);
4594 spin_unlock_irq(&x->wait.lock);
4598 void __sched wait_for_completion(struct completion *x)
4600 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4602 EXPORT_SYMBOL(wait_for_completion);
4604 unsigned long __sched
4605 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4607 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4609 EXPORT_SYMBOL(wait_for_completion_timeout);
4611 int __sched wait_for_completion_interruptible(struct completion *x)
4613 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4614 if (t == -ERESTARTSYS)
4618 EXPORT_SYMBOL(wait_for_completion_interruptible);
4620 unsigned long __sched
4621 wait_for_completion_interruptible_timeout(struct completion *x,
4622 unsigned long timeout)
4624 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4626 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4628 int __sched wait_for_completion_killable(struct completion *x)
4630 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4631 if (t == -ERESTARTSYS)
4635 EXPORT_SYMBOL(wait_for_completion_killable);
4638 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4640 unsigned long flags;
4643 init_waitqueue_entry(&wait, current);
4645 __set_current_state(state);
4647 spin_lock_irqsave(&q->lock, flags);
4648 __add_wait_queue(q, &wait);
4649 spin_unlock(&q->lock);
4650 timeout = schedule_timeout(timeout);
4651 spin_lock_irq(&q->lock);
4652 __remove_wait_queue(q, &wait);
4653 spin_unlock_irqrestore(&q->lock, flags);
4658 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4660 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4662 EXPORT_SYMBOL(interruptible_sleep_on);
4665 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4667 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4669 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4671 void __sched sleep_on(wait_queue_head_t *q)
4673 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4675 EXPORT_SYMBOL(sleep_on);
4677 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4679 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4681 EXPORT_SYMBOL(sleep_on_timeout);
4683 #ifdef CONFIG_RT_MUTEXES
4686 * rt_mutex_setprio - set the current priority of a task
4688 * @prio: prio value (kernel-internal form)
4690 * This function changes the 'effective' priority of a task. It does
4691 * not touch ->normal_prio like __setscheduler().
4693 * Used by the rt_mutex code to implement priority inheritance logic.
4695 void rt_mutex_setprio(struct task_struct *p, int prio)
4697 unsigned long flags;
4698 int oldprio, on_rq, running;
4700 const struct sched_class *prev_class = p->sched_class;
4702 BUG_ON(prio < 0 || prio > MAX_PRIO);
4704 rq = task_rq_lock(p, &flags);
4705 update_rq_clock(rq);
4708 on_rq = p->se.on_rq;
4709 running = task_current(rq, p);
4711 dequeue_task(rq, p, 0);
4713 p->sched_class->put_prev_task(rq, p);
4716 p->sched_class = &rt_sched_class;
4718 p->sched_class = &fair_sched_class;
4723 p->sched_class->set_curr_task(rq);
4725 enqueue_task(rq, p, 0);
4727 check_class_changed(rq, p, prev_class, oldprio, running);
4729 task_rq_unlock(rq, &flags);
4734 void set_user_nice(struct task_struct *p, long nice)
4736 int old_prio, delta, on_rq;
4737 unsigned long flags;
4740 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4743 * We have to be careful, if called from sys_setpriority(),
4744 * the task might be in the middle of scheduling on another CPU.
4746 rq = task_rq_lock(p, &flags);
4747 update_rq_clock(rq);
4749 * The RT priorities are set via sched_setscheduler(), but we still
4750 * allow the 'normal' nice value to be set - but as expected
4751 * it wont have any effect on scheduling until the task is
4752 * SCHED_FIFO/SCHED_RR:
4754 if (task_has_rt_policy(p)) {
4755 p->static_prio = NICE_TO_PRIO(nice);
4758 on_rq = p->se.on_rq;
4760 dequeue_task(rq, p, 0);
4762 p->static_prio = NICE_TO_PRIO(nice);
4765 p->prio = effective_prio(p);
4766 delta = p->prio - old_prio;
4769 enqueue_task(rq, p, 0);
4771 * If the task increased its priority or is running and
4772 * lowered its priority, then reschedule its CPU:
4774 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4775 resched_task(rq->curr);
4778 task_rq_unlock(rq, &flags);
4780 EXPORT_SYMBOL(set_user_nice);
4783 * can_nice - check if a task can reduce its nice value
4787 int can_nice(const struct task_struct *p, const int nice)
4789 /* convert nice value [19,-20] to rlimit style value [1,40] */
4790 int nice_rlim = 20 - nice;
4792 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4793 capable(CAP_SYS_NICE));
4796 #ifdef __ARCH_WANT_SYS_NICE
4799 * sys_nice - change the priority of the current process.
4800 * @increment: priority increment
4802 * sys_setpriority is a more generic, but much slower function that
4803 * does similar things.
4805 asmlinkage long sys_nice(int increment)
4810 * Setpriority might change our priority at the same moment.
4811 * We don't have to worry. Conceptually one call occurs first
4812 * and we have a single winner.
4814 if (increment < -40)
4819 nice = PRIO_TO_NICE(current->static_prio) + increment;
4825 if (increment < 0 && !can_nice(current, nice))
4828 retval = security_task_setnice(current, nice);
4832 set_user_nice(current, nice);
4839 * task_prio - return the priority value of a given task.
4840 * @p: the task in question.
4842 * This is the priority value as seen by users in /proc.
4843 * RT tasks are offset by -200. Normal tasks are centered
4844 * around 0, value goes from -16 to +15.
4846 int task_prio(const struct task_struct *p)
4848 return p->prio - MAX_RT_PRIO;
4852 * task_nice - return the nice value of a given task.
4853 * @p: the task in question.
4855 int task_nice(const struct task_struct *p)
4857 return TASK_NICE(p);
4859 EXPORT_SYMBOL(task_nice);
4862 * idle_cpu - is a given cpu idle currently?
4863 * @cpu: the processor in question.
4865 int idle_cpu(int cpu)
4867 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4871 * idle_task - return the idle task for a given cpu.
4872 * @cpu: the processor in question.
4874 struct task_struct *idle_task(int cpu)
4876 return cpu_rq(cpu)->idle;
4880 * find_process_by_pid - find a process with a matching PID value.
4881 * @pid: the pid in question.
4883 static struct task_struct *find_process_by_pid(pid_t pid)
4885 return pid ? find_task_by_vpid(pid) : current;
4888 /* Actually do priority change: must hold rq lock. */
4890 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4892 BUG_ON(p->se.on_rq);
4895 switch (p->policy) {
4899 p->sched_class = &fair_sched_class;
4903 p->sched_class = &rt_sched_class;
4907 p->rt_priority = prio;
4908 p->normal_prio = normal_prio(p);
4909 /* we are holding p->pi_lock already */
4910 p->prio = rt_mutex_getprio(p);
4915 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4916 * @p: the task in question.
4917 * @policy: new policy.
4918 * @param: structure containing the new RT priority.
4920 * NOTE that the task may be already dead.
4922 int sched_setscheduler(struct task_struct *p, int policy,
4923 struct sched_param *param)
4925 int retval, oldprio, oldpolicy = -1, on_rq, running;
4926 unsigned long flags;
4927 const struct sched_class *prev_class = p->sched_class;
4930 /* may grab non-irq protected spin_locks */
4931 BUG_ON(in_interrupt());
4933 /* double check policy once rq lock held */
4935 policy = oldpolicy = p->policy;
4936 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4937 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4938 policy != SCHED_IDLE)
4941 * Valid priorities for SCHED_FIFO and SCHED_RR are
4942 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4943 * SCHED_BATCH and SCHED_IDLE is 0.
4945 if (param->sched_priority < 0 ||
4946 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4947 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4949 if (rt_policy(policy) != (param->sched_priority != 0))
4953 * Allow unprivileged RT tasks to decrease priority:
4955 if (!capable(CAP_SYS_NICE)) {
4956 if (rt_policy(policy)) {
4957 unsigned long rlim_rtprio;
4959 if (!lock_task_sighand(p, &flags))
4961 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4962 unlock_task_sighand(p, &flags);
4964 /* can't set/change the rt policy */
4965 if (policy != p->policy && !rlim_rtprio)
4968 /* can't increase priority */
4969 if (param->sched_priority > p->rt_priority &&
4970 param->sched_priority > rlim_rtprio)
4974 * Like positive nice levels, dont allow tasks to
4975 * move out of SCHED_IDLE either:
4977 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4980 /* can't change other user's priorities */
4981 if ((current->euid != p->euid) &&
4982 (current->euid != p->uid))
4986 #ifdef CONFIG_RT_GROUP_SCHED
4988 * Do not allow realtime tasks into groups that have no runtime
4991 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4995 retval = security_task_setscheduler(p, policy, param);
4999 * make sure no PI-waiters arrive (or leave) while we are
5000 * changing the priority of the task:
5002 spin_lock_irqsave(&p->pi_lock, flags);
5004 * To be able to change p->policy safely, the apropriate
5005 * runqueue lock must be held.
5007 rq = __task_rq_lock(p);
5008 /* recheck policy now with rq lock held */
5009 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5010 policy = oldpolicy = -1;
5011 __task_rq_unlock(rq);
5012 spin_unlock_irqrestore(&p->pi_lock, flags);
5015 update_rq_clock(rq);
5016 on_rq = p->se.on_rq;
5017 running = task_current(rq, p);
5019 deactivate_task(rq, p, 0);
5021 p->sched_class->put_prev_task(rq, p);
5024 __setscheduler(rq, p, policy, param->sched_priority);
5027 p->sched_class->set_curr_task(rq);
5029 activate_task(rq, p, 0);
5031 check_class_changed(rq, p, prev_class, oldprio, running);
5033 __task_rq_unlock(rq);
5034 spin_unlock_irqrestore(&p->pi_lock, flags);
5036 rt_mutex_adjust_pi(p);
5040 EXPORT_SYMBOL_GPL(sched_setscheduler);
5043 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5045 struct sched_param lparam;
5046 struct task_struct *p;
5049 if (!param || pid < 0)
5051 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5056 p = find_process_by_pid(pid);
5058 retval = sched_setscheduler(p, policy, &lparam);
5065 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5066 * @pid: the pid in question.
5067 * @policy: new policy.
5068 * @param: structure containing the new RT priority.
5071 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5073 /* negative values for policy are not valid */
5077 return do_sched_setscheduler(pid, policy, param);
5081 * sys_sched_setparam - set/change the RT priority of a thread
5082 * @pid: the pid in question.
5083 * @param: structure containing the new RT priority.
5085 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5087 return do_sched_setscheduler(pid, -1, param);
5091 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5092 * @pid: the pid in question.
5094 asmlinkage long sys_sched_getscheduler(pid_t pid)
5096 struct task_struct *p;
5103 read_lock(&tasklist_lock);
5104 p = find_process_by_pid(pid);
5106 retval = security_task_getscheduler(p);
5110 read_unlock(&tasklist_lock);
5115 * sys_sched_getscheduler - get the RT priority of a thread
5116 * @pid: the pid in question.
5117 * @param: structure containing the RT priority.
5119 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5121 struct sched_param lp;
5122 struct task_struct *p;
5125 if (!param || pid < 0)
5128 read_lock(&tasklist_lock);
5129 p = find_process_by_pid(pid);
5134 retval = security_task_getscheduler(p);
5138 lp.sched_priority = p->rt_priority;
5139 read_unlock(&tasklist_lock);
5142 * This one might sleep, we cannot do it with a spinlock held ...
5144 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5149 read_unlock(&tasklist_lock);
5153 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5155 cpumask_t cpus_allowed;
5156 cpumask_t new_mask = *in_mask;
5157 struct task_struct *p;
5161 read_lock(&tasklist_lock);
5163 p = find_process_by_pid(pid);
5165 read_unlock(&tasklist_lock);
5171 * It is not safe to call set_cpus_allowed with the
5172 * tasklist_lock held. We will bump the task_struct's
5173 * usage count and then drop tasklist_lock.
5176 read_unlock(&tasklist_lock);
5179 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5180 !capable(CAP_SYS_NICE))
5183 retval = security_task_setscheduler(p, 0, NULL);
5187 cpuset_cpus_allowed(p, &cpus_allowed);
5188 cpus_and(new_mask, new_mask, cpus_allowed);
5190 retval = set_cpus_allowed_ptr(p, &new_mask);
5193 cpuset_cpus_allowed(p, &cpus_allowed);
5194 if (!cpus_subset(new_mask, cpus_allowed)) {
5196 * We must have raced with a concurrent cpuset
5197 * update. Just reset the cpus_allowed to the
5198 * cpuset's cpus_allowed
5200 new_mask = cpus_allowed;
5210 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5211 cpumask_t *new_mask)
5213 if (len < sizeof(cpumask_t)) {
5214 memset(new_mask, 0, sizeof(cpumask_t));
5215 } else if (len > sizeof(cpumask_t)) {
5216 len = sizeof(cpumask_t);
5218 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5222 * sys_sched_setaffinity - set the cpu affinity of a process
5223 * @pid: pid of the process
5224 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5225 * @user_mask_ptr: user-space pointer to the new cpu mask
5227 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5228 unsigned long __user *user_mask_ptr)
5233 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5237 return sched_setaffinity(pid, &new_mask);
5240 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5242 struct task_struct *p;
5246 read_lock(&tasklist_lock);
5249 p = find_process_by_pid(pid);
5253 retval = security_task_getscheduler(p);
5257 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5260 read_unlock(&tasklist_lock);
5267 * sys_sched_getaffinity - get the cpu affinity of a process
5268 * @pid: pid of the process
5269 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5270 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5272 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5273 unsigned long __user *user_mask_ptr)
5278 if (len < sizeof(cpumask_t))
5281 ret = sched_getaffinity(pid, &mask);
5285 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5288 return sizeof(cpumask_t);
5292 * sys_sched_yield - yield the current processor to other threads.
5294 * This function yields the current CPU to other tasks. If there are no
5295 * other threads running on this CPU then this function will return.
5297 asmlinkage long sys_sched_yield(void)
5299 struct rq *rq = this_rq_lock();
5301 schedstat_inc(rq, yld_count);
5302 current->sched_class->yield_task(rq);
5305 * Since we are going to call schedule() anyway, there's
5306 * no need to preempt or enable interrupts:
5308 __release(rq->lock);
5309 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5310 _raw_spin_unlock(&rq->lock);
5311 preempt_enable_no_resched();
5318 static void __cond_resched(void)
5320 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5321 __might_sleep(__FILE__, __LINE__);
5324 * The BKS might be reacquired before we have dropped
5325 * PREEMPT_ACTIVE, which could trigger a second
5326 * cond_resched() call.
5329 add_preempt_count(PREEMPT_ACTIVE);
5331 sub_preempt_count(PREEMPT_ACTIVE);
5332 } while (need_resched());
5335 int __sched _cond_resched(void)
5337 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5338 system_state == SYSTEM_RUNNING) {
5344 EXPORT_SYMBOL(_cond_resched);
5347 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5348 * call schedule, and on return reacquire the lock.
5350 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5351 * operations here to prevent schedule() from being called twice (once via
5352 * spin_unlock(), once by hand).
5354 int cond_resched_lock(spinlock_t *lock)
5356 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5359 if (spin_needbreak(lock) || resched) {
5361 if (resched && need_resched())
5370 EXPORT_SYMBOL(cond_resched_lock);
5372 int __sched cond_resched_softirq(void)
5374 BUG_ON(!in_softirq());
5376 if (need_resched() && system_state == SYSTEM_RUNNING) {
5384 EXPORT_SYMBOL(cond_resched_softirq);
5387 * yield - yield the current processor to other threads.
5389 * This is a shortcut for kernel-space yielding - it marks the
5390 * thread runnable and calls sys_sched_yield().
5392 void __sched yield(void)
5394 set_current_state(TASK_RUNNING);
5397 EXPORT_SYMBOL(yield);
5400 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5401 * that process accounting knows that this is a task in IO wait state.
5403 * But don't do that if it is a deliberate, throttling IO wait (this task
5404 * has set its backing_dev_info: the queue against which it should throttle)
5406 void __sched io_schedule(void)
5408 struct rq *rq = &__raw_get_cpu_var(runqueues);
5410 delayacct_blkio_start();
5411 atomic_inc(&rq->nr_iowait);
5413 atomic_dec(&rq->nr_iowait);
5414 delayacct_blkio_end();
5416 EXPORT_SYMBOL(io_schedule);
5418 long __sched io_schedule_timeout(long timeout)
5420 struct rq *rq = &__raw_get_cpu_var(runqueues);
5423 delayacct_blkio_start();
5424 atomic_inc(&rq->nr_iowait);
5425 ret = schedule_timeout(timeout);
5426 atomic_dec(&rq->nr_iowait);
5427 delayacct_blkio_end();
5432 * sys_sched_get_priority_max - return maximum RT priority.
5433 * @policy: scheduling class.
5435 * this syscall returns the maximum rt_priority that can be used
5436 * by a given scheduling class.
5438 asmlinkage long sys_sched_get_priority_max(int policy)
5445 ret = MAX_USER_RT_PRIO-1;
5457 * sys_sched_get_priority_min - return minimum RT priority.
5458 * @policy: scheduling class.
5460 * this syscall returns the minimum rt_priority that can be used
5461 * by a given scheduling class.
5463 asmlinkage long sys_sched_get_priority_min(int policy)
5481 * sys_sched_rr_get_interval - return the default timeslice of a process.
5482 * @pid: pid of the process.
5483 * @interval: userspace pointer to the timeslice value.
5485 * this syscall writes the default timeslice value of a given process
5486 * into the user-space timespec buffer. A value of '0' means infinity.
5489 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5491 struct task_struct *p;
5492 unsigned int time_slice;
5500 read_lock(&tasklist_lock);
5501 p = find_process_by_pid(pid);
5505 retval = security_task_getscheduler(p);
5510 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5511 * tasks that are on an otherwise idle runqueue:
5514 if (p->policy == SCHED_RR) {
5515 time_slice = DEF_TIMESLICE;
5516 } else if (p->policy != SCHED_FIFO) {
5517 struct sched_entity *se = &p->se;
5518 unsigned long flags;
5521 rq = task_rq_lock(p, &flags);
5522 if (rq->cfs.load.weight)
5523 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5524 task_rq_unlock(rq, &flags);
5526 read_unlock(&tasklist_lock);
5527 jiffies_to_timespec(time_slice, &t);
5528 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5532 read_unlock(&tasklist_lock);
5536 static const char stat_nam[] = "RSDTtZX";
5538 void sched_show_task(struct task_struct *p)
5540 unsigned long free = 0;
5543 state = p->state ? __ffs(p->state) + 1 : 0;
5544 printk(KERN_INFO "%-13.13s %c", p->comm,
5545 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5546 #if BITS_PER_LONG == 32
5547 if (state == TASK_RUNNING)
5548 printk(KERN_CONT " running ");
5550 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5552 if (state == TASK_RUNNING)
5553 printk(KERN_CONT " running task ");
5555 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5557 #ifdef CONFIG_DEBUG_STACK_USAGE
5559 unsigned long *n = end_of_stack(p);
5562 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5565 printk(KERN_CONT "%5lu %5d %6d\n", free,
5566 task_pid_nr(p), task_pid_nr(p->real_parent));
5568 show_stack(p, NULL);
5571 void show_state_filter(unsigned long state_filter)
5573 struct task_struct *g, *p;
5575 #if BITS_PER_LONG == 32
5577 " task PC stack pid father\n");
5580 " task PC stack pid father\n");
5582 read_lock(&tasklist_lock);
5583 do_each_thread(g, p) {
5585 * reset the NMI-timeout, listing all files on a slow
5586 * console might take alot of time:
5588 touch_nmi_watchdog();
5589 if (!state_filter || (p->state & state_filter))
5591 } while_each_thread(g, p);
5593 touch_all_softlockup_watchdogs();
5595 #ifdef CONFIG_SCHED_DEBUG
5596 sysrq_sched_debug_show();
5598 read_unlock(&tasklist_lock);
5600 * Only show locks if all tasks are dumped:
5602 if (state_filter == -1)
5603 debug_show_all_locks();
5606 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5608 idle->sched_class = &idle_sched_class;
5612 * init_idle - set up an idle thread for a given CPU
5613 * @idle: task in question
5614 * @cpu: cpu the idle task belongs to
5616 * NOTE: this function does not set the idle thread's NEED_RESCHED
5617 * flag, to make booting more robust.
5619 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5621 struct rq *rq = cpu_rq(cpu);
5622 unsigned long flags;
5625 idle->se.exec_start = sched_clock();
5627 idle->prio = idle->normal_prio = MAX_PRIO;
5628 idle->cpus_allowed = cpumask_of_cpu(cpu);
5629 __set_task_cpu(idle, cpu);
5631 spin_lock_irqsave(&rq->lock, flags);
5632 rq->curr = rq->idle = idle;
5633 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5636 spin_unlock_irqrestore(&rq->lock, flags);
5638 /* Set the preempt count _outside_ the spinlocks! */
5639 #if defined(CONFIG_PREEMPT)
5640 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5642 task_thread_info(idle)->preempt_count = 0;
5645 * The idle tasks have their own, simple scheduling class:
5647 idle->sched_class = &idle_sched_class;
5651 * In a system that switches off the HZ timer nohz_cpu_mask
5652 * indicates which cpus entered this state. This is used
5653 * in the rcu update to wait only for active cpus. For system
5654 * which do not switch off the HZ timer nohz_cpu_mask should
5655 * always be CPU_MASK_NONE.
5657 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5660 * Increase the granularity value when there are more CPUs,
5661 * because with more CPUs the 'effective latency' as visible
5662 * to users decreases. But the relationship is not linear,
5663 * so pick a second-best guess by going with the log2 of the
5666 * This idea comes from the SD scheduler of Con Kolivas:
5668 static inline void sched_init_granularity(void)
5670 unsigned int factor = 1 + ilog2(num_online_cpus());
5671 const unsigned long limit = 200000000;
5673 sysctl_sched_min_granularity *= factor;
5674 if (sysctl_sched_min_granularity > limit)
5675 sysctl_sched_min_granularity = limit;
5677 sysctl_sched_latency *= factor;
5678 if (sysctl_sched_latency > limit)
5679 sysctl_sched_latency = limit;
5681 sysctl_sched_wakeup_granularity *= factor;
5686 * This is how migration works:
5688 * 1) we queue a struct migration_req structure in the source CPU's
5689 * runqueue and wake up that CPU's migration thread.
5690 * 2) we down() the locked semaphore => thread blocks.
5691 * 3) migration thread wakes up (implicitly it forces the migrated
5692 * thread off the CPU)
5693 * 4) it gets the migration request and checks whether the migrated
5694 * task is still in the wrong runqueue.
5695 * 5) if it's in the wrong runqueue then the migration thread removes
5696 * it and puts it into the right queue.
5697 * 6) migration thread up()s the semaphore.
5698 * 7) we wake up and the migration is done.
5702 * Change a given task's CPU affinity. Migrate the thread to a
5703 * proper CPU and schedule it away if the CPU it's executing on
5704 * is removed from the allowed bitmask.
5706 * NOTE: the caller must have a valid reference to the task, the
5707 * task must not exit() & deallocate itself prematurely. The
5708 * call is not atomic; no spinlocks may be held.
5710 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5712 struct migration_req req;
5713 unsigned long flags;
5717 rq = task_rq_lock(p, &flags);
5718 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5723 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5724 !cpus_equal(p->cpus_allowed, *new_mask))) {
5729 if (p->sched_class->set_cpus_allowed)
5730 p->sched_class->set_cpus_allowed(p, new_mask);
5732 p->cpus_allowed = *new_mask;
5733 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5736 /* Can the task run on the task's current CPU? If so, we're done */
5737 if (cpu_isset(task_cpu(p), *new_mask))
5740 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5741 /* Need help from migration thread: drop lock and wait. */
5742 task_rq_unlock(rq, &flags);
5743 wake_up_process(rq->migration_thread);
5744 wait_for_completion(&req.done);
5745 tlb_migrate_finish(p->mm);
5749 task_rq_unlock(rq, &flags);
5753 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5756 * Move (not current) task off this cpu, onto dest cpu. We're doing
5757 * this because either it can't run here any more (set_cpus_allowed()
5758 * away from this CPU, or CPU going down), or because we're
5759 * attempting to rebalance this task on exec (sched_exec).
5761 * So we race with normal scheduler movements, but that's OK, as long
5762 * as the task is no longer on this CPU.
5764 * Returns non-zero if task was successfully migrated.
5766 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5768 struct rq *rq_dest, *rq_src;
5771 if (unlikely(cpu_is_offline(dest_cpu)))
5774 rq_src = cpu_rq(src_cpu);
5775 rq_dest = cpu_rq(dest_cpu);
5777 double_rq_lock(rq_src, rq_dest);
5778 /* Already moved. */
5779 if (task_cpu(p) != src_cpu)
5781 /* Affinity changed (again). */
5782 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5785 on_rq = p->se.on_rq;
5787 deactivate_task(rq_src, p, 0);
5789 set_task_cpu(p, dest_cpu);
5791 activate_task(rq_dest, p, 0);
5792 check_preempt_curr(rq_dest, p);
5796 double_rq_unlock(rq_src, rq_dest);
5801 * migration_thread - this is a highprio system thread that performs
5802 * thread migration by bumping thread off CPU then 'pushing' onto
5805 static int migration_thread(void *data)
5807 int cpu = (long)data;
5811 BUG_ON(rq->migration_thread != current);
5813 set_current_state(TASK_INTERRUPTIBLE);
5814 while (!kthread_should_stop()) {
5815 struct migration_req *req;
5816 struct list_head *head;
5818 spin_lock_irq(&rq->lock);
5820 if (cpu_is_offline(cpu)) {
5821 spin_unlock_irq(&rq->lock);
5825 if (rq->active_balance) {
5826 active_load_balance(rq, cpu);
5827 rq->active_balance = 0;
5830 head = &rq->migration_queue;
5832 if (list_empty(head)) {
5833 spin_unlock_irq(&rq->lock);
5835 set_current_state(TASK_INTERRUPTIBLE);
5838 req = list_entry(head->next, struct migration_req, list);
5839 list_del_init(head->next);
5841 spin_unlock(&rq->lock);
5842 __migrate_task(req->task, cpu, req->dest_cpu);
5845 complete(&req->done);
5847 __set_current_state(TASK_RUNNING);
5851 /* Wait for kthread_stop */
5852 set_current_state(TASK_INTERRUPTIBLE);
5853 while (!kthread_should_stop()) {
5855 set_current_state(TASK_INTERRUPTIBLE);
5857 __set_current_state(TASK_RUNNING);
5861 #ifdef CONFIG_HOTPLUG_CPU
5863 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5867 local_irq_disable();
5868 ret = __migrate_task(p, src_cpu, dest_cpu);
5874 * Figure out where task on dead CPU should go, use force if necessary.
5875 * NOTE: interrupts should be disabled by the caller
5877 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5879 unsigned long flags;
5886 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5887 cpus_and(mask, mask, p->cpus_allowed);
5888 dest_cpu = any_online_cpu(mask);
5890 /* On any allowed CPU? */
5891 if (dest_cpu >= nr_cpu_ids)
5892 dest_cpu = any_online_cpu(p->cpus_allowed);
5894 /* No more Mr. Nice Guy. */
5895 if (dest_cpu >= nr_cpu_ids) {
5896 cpumask_t cpus_allowed;
5898 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5900 * Try to stay on the same cpuset, where the
5901 * current cpuset may be a subset of all cpus.
5902 * The cpuset_cpus_allowed_locked() variant of
5903 * cpuset_cpus_allowed() will not block. It must be
5904 * called within calls to cpuset_lock/cpuset_unlock.
5906 rq = task_rq_lock(p, &flags);
5907 p->cpus_allowed = cpus_allowed;
5908 dest_cpu = any_online_cpu(p->cpus_allowed);
5909 task_rq_unlock(rq, &flags);
5912 * Don't tell them about moving exiting tasks or
5913 * kernel threads (both mm NULL), since they never
5916 if (p->mm && printk_ratelimit()) {
5917 printk(KERN_INFO "process %d (%s) no "
5918 "longer affine to cpu%d\n",
5919 task_pid_nr(p), p->comm, dead_cpu);
5922 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5926 * While a dead CPU has no uninterruptible tasks queued at this point,
5927 * it might still have a nonzero ->nr_uninterruptible counter, because
5928 * for performance reasons the counter is not stricly tracking tasks to
5929 * their home CPUs. So we just add the counter to another CPU's counter,
5930 * to keep the global sum constant after CPU-down:
5932 static void migrate_nr_uninterruptible(struct rq *rq_src)
5934 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5935 unsigned long flags;
5937 local_irq_save(flags);
5938 double_rq_lock(rq_src, rq_dest);
5939 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5940 rq_src->nr_uninterruptible = 0;
5941 double_rq_unlock(rq_src, rq_dest);
5942 local_irq_restore(flags);
5945 /* Run through task list and migrate tasks from the dead cpu. */
5946 static void migrate_live_tasks(int src_cpu)
5948 struct task_struct *p, *t;
5950 read_lock(&tasklist_lock);
5952 do_each_thread(t, p) {
5956 if (task_cpu(p) == src_cpu)
5957 move_task_off_dead_cpu(src_cpu, p);
5958 } while_each_thread(t, p);
5960 read_unlock(&tasklist_lock);
5964 * Schedules idle task to be the next runnable task on current CPU.
5965 * It does so by boosting its priority to highest possible.
5966 * Used by CPU offline code.
5968 void sched_idle_next(void)
5970 int this_cpu = smp_processor_id();
5971 struct rq *rq = cpu_rq(this_cpu);
5972 struct task_struct *p = rq->idle;
5973 unsigned long flags;
5975 /* cpu has to be offline */
5976 BUG_ON(cpu_online(this_cpu));
5979 * Strictly not necessary since rest of the CPUs are stopped by now
5980 * and interrupts disabled on the current cpu.
5982 spin_lock_irqsave(&rq->lock, flags);
5984 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5986 update_rq_clock(rq);
5987 activate_task(rq, p, 0);
5989 spin_unlock_irqrestore(&rq->lock, flags);
5993 * Ensures that the idle task is using init_mm right before its cpu goes
5996 void idle_task_exit(void)
5998 struct mm_struct *mm = current->active_mm;
6000 BUG_ON(cpu_online(smp_processor_id()));
6003 switch_mm(mm, &init_mm, current);
6007 /* called under rq->lock with disabled interrupts */
6008 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6010 struct rq *rq = cpu_rq(dead_cpu);
6012 /* Must be exiting, otherwise would be on tasklist. */
6013 BUG_ON(!p->exit_state);
6015 /* Cannot have done final schedule yet: would have vanished. */
6016 BUG_ON(p->state == TASK_DEAD);
6021 * Drop lock around migration; if someone else moves it,
6022 * that's OK. No task can be added to this CPU, so iteration is
6025 spin_unlock_irq(&rq->lock);
6026 move_task_off_dead_cpu(dead_cpu, p);
6027 spin_lock_irq(&rq->lock);
6032 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6033 static void migrate_dead_tasks(unsigned int dead_cpu)
6035 struct rq *rq = cpu_rq(dead_cpu);
6036 struct task_struct *next;
6039 if (!rq->nr_running)
6041 update_rq_clock(rq);
6042 next = pick_next_task(rq, rq->curr);
6045 migrate_dead(dead_cpu, next);
6049 #endif /* CONFIG_HOTPLUG_CPU */
6051 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6053 static struct ctl_table sd_ctl_dir[] = {
6055 .procname = "sched_domain",
6061 static struct ctl_table sd_ctl_root[] = {
6063 .ctl_name = CTL_KERN,
6064 .procname = "kernel",
6066 .child = sd_ctl_dir,
6071 static struct ctl_table *sd_alloc_ctl_entry(int n)
6073 struct ctl_table *entry =
6074 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6079 static void sd_free_ctl_entry(struct ctl_table **tablep)
6081 struct ctl_table *entry;
6084 * In the intermediate directories, both the child directory and
6085 * procname are dynamically allocated and could fail but the mode
6086 * will always be set. In the lowest directory the names are
6087 * static strings and all have proc handlers.
6089 for (entry = *tablep; entry->mode; entry++) {
6091 sd_free_ctl_entry(&entry->child);
6092 if (entry->proc_handler == NULL)
6093 kfree(entry->procname);
6101 set_table_entry(struct ctl_table *entry,
6102 const char *procname, void *data, int maxlen,
6103 mode_t mode, proc_handler *proc_handler)
6105 entry->procname = procname;
6107 entry->maxlen = maxlen;
6109 entry->proc_handler = proc_handler;
6112 static struct ctl_table *
6113 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6115 struct ctl_table *table = sd_alloc_ctl_entry(12);
6120 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6121 sizeof(long), 0644, proc_doulongvec_minmax);
6122 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6123 sizeof(long), 0644, proc_doulongvec_minmax);
6124 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6125 sizeof(int), 0644, proc_dointvec_minmax);
6126 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6127 sizeof(int), 0644, proc_dointvec_minmax);
6128 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6129 sizeof(int), 0644, proc_dointvec_minmax);
6130 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6131 sizeof(int), 0644, proc_dointvec_minmax);
6132 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6133 sizeof(int), 0644, proc_dointvec_minmax);
6134 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6135 sizeof(int), 0644, proc_dointvec_minmax);
6136 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6137 sizeof(int), 0644, proc_dointvec_minmax);
6138 set_table_entry(&table[9], "cache_nice_tries",
6139 &sd->cache_nice_tries,
6140 sizeof(int), 0644, proc_dointvec_minmax);
6141 set_table_entry(&table[10], "flags", &sd->flags,
6142 sizeof(int), 0644, proc_dointvec_minmax);
6143 /* &table[11] is terminator */
6148 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6150 struct ctl_table *entry, *table;
6151 struct sched_domain *sd;
6152 int domain_num = 0, i;
6155 for_each_domain(cpu, sd)
6157 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6162 for_each_domain(cpu, sd) {
6163 snprintf(buf, 32, "domain%d", i);
6164 entry->procname = kstrdup(buf, GFP_KERNEL);
6166 entry->child = sd_alloc_ctl_domain_table(sd);
6173 static struct ctl_table_header *sd_sysctl_header;
6174 static void register_sched_domain_sysctl(void)
6176 int i, cpu_num = num_online_cpus();
6177 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6180 WARN_ON(sd_ctl_dir[0].child);
6181 sd_ctl_dir[0].child = entry;
6186 for_each_online_cpu(i) {
6187 snprintf(buf, 32, "cpu%d", i);
6188 entry->procname = kstrdup(buf, GFP_KERNEL);
6190 entry->child = sd_alloc_ctl_cpu_table(i);
6194 WARN_ON(sd_sysctl_header);
6195 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6198 /* may be called multiple times per register */
6199 static void unregister_sched_domain_sysctl(void)
6201 if (sd_sysctl_header)
6202 unregister_sysctl_table(sd_sysctl_header);
6203 sd_sysctl_header = NULL;
6204 if (sd_ctl_dir[0].child)
6205 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6208 static void register_sched_domain_sysctl(void)
6211 static void unregister_sched_domain_sysctl(void)
6216 static void set_rq_online(struct rq *rq)
6219 const struct sched_class *class;
6221 cpu_set(rq->cpu, rq->rd->online);
6224 for_each_class(class) {
6225 if (class->rq_online)
6226 class->rq_online(rq);
6231 static void set_rq_offline(struct rq *rq)
6234 const struct sched_class *class;
6236 for_each_class(class) {
6237 if (class->rq_offline)
6238 class->rq_offline(rq);
6241 cpu_clear(rq->cpu, rq->rd->online);
6247 * migration_call - callback that gets triggered when a CPU is added.
6248 * Here we can start up the necessary migration thread for the new CPU.
6250 static int __cpuinit
6251 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6253 struct task_struct *p;
6254 int cpu = (long)hcpu;
6255 unsigned long flags;
6260 case CPU_UP_PREPARE:
6261 case CPU_UP_PREPARE_FROZEN:
6262 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6265 kthread_bind(p, cpu);
6266 /* Must be high prio: stop_machine expects to yield to it. */
6267 rq = task_rq_lock(p, &flags);
6268 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6269 task_rq_unlock(rq, &flags);
6270 cpu_rq(cpu)->migration_thread = p;
6274 case CPU_ONLINE_FROZEN:
6275 /* Strictly unnecessary, as first user will wake it. */
6276 wake_up_process(cpu_rq(cpu)->migration_thread);
6278 /* Update our root-domain */
6280 spin_lock_irqsave(&rq->lock, flags);
6282 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6286 spin_unlock_irqrestore(&rq->lock, flags);
6289 #ifdef CONFIG_HOTPLUG_CPU
6290 case CPU_UP_CANCELED:
6291 case CPU_UP_CANCELED_FROZEN:
6292 if (!cpu_rq(cpu)->migration_thread)
6294 /* Unbind it from offline cpu so it can run. Fall thru. */
6295 kthread_bind(cpu_rq(cpu)->migration_thread,
6296 any_online_cpu(cpu_online_map));
6297 kthread_stop(cpu_rq(cpu)->migration_thread);
6298 cpu_rq(cpu)->migration_thread = NULL;
6302 case CPU_DEAD_FROZEN:
6303 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6304 migrate_live_tasks(cpu);
6306 kthread_stop(rq->migration_thread);
6307 rq->migration_thread = NULL;
6308 /* Idle task back to normal (off runqueue, low prio) */
6309 spin_lock_irq(&rq->lock);
6310 update_rq_clock(rq);
6311 deactivate_task(rq, rq->idle, 0);
6312 rq->idle->static_prio = MAX_PRIO;
6313 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6314 rq->idle->sched_class = &idle_sched_class;
6315 migrate_dead_tasks(cpu);
6316 spin_unlock_irq(&rq->lock);
6318 migrate_nr_uninterruptible(rq);
6319 BUG_ON(rq->nr_running != 0);
6322 * No need to migrate the tasks: it was best-effort if
6323 * they didn't take sched_hotcpu_mutex. Just wake up
6326 spin_lock_irq(&rq->lock);
6327 while (!list_empty(&rq->migration_queue)) {
6328 struct migration_req *req;
6330 req = list_entry(rq->migration_queue.next,
6331 struct migration_req, list);
6332 list_del_init(&req->list);
6333 complete(&req->done);
6335 spin_unlock_irq(&rq->lock);
6339 case CPU_DYING_FROZEN:
6340 /* Update our root-domain */
6342 spin_lock_irqsave(&rq->lock, flags);
6344 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6347 spin_unlock_irqrestore(&rq->lock, flags);
6354 /* Register at highest priority so that task migration (migrate_all_tasks)
6355 * happens before everything else.
6357 static struct notifier_block __cpuinitdata migration_notifier = {
6358 .notifier_call = migration_call,
6362 void __init migration_init(void)
6364 void *cpu = (void *)(long)smp_processor_id();
6367 /* Start one for the boot CPU: */
6368 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6369 BUG_ON(err == NOTIFY_BAD);
6370 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6371 register_cpu_notifier(&migration_notifier);
6377 #ifdef CONFIG_SCHED_DEBUG
6379 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6392 case SD_LV_ALLNODES:
6401 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6402 cpumask_t *groupmask)
6404 struct sched_group *group = sd->groups;
6407 cpulist_scnprintf(str, sizeof(str), sd->span);
6408 cpus_clear(*groupmask);
6410 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6412 if (!(sd->flags & SD_LOAD_BALANCE)) {
6413 printk("does not load-balance\n");
6415 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6420 printk(KERN_CONT "span %s level %s\n",
6421 str, sd_level_to_string(sd->level));
6423 if (!cpu_isset(cpu, sd->span)) {
6424 printk(KERN_ERR "ERROR: domain->span does not contain "
6427 if (!cpu_isset(cpu, group->cpumask)) {
6428 printk(KERN_ERR "ERROR: domain->groups does not contain"
6432 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6436 printk(KERN_ERR "ERROR: group is NULL\n");
6440 if (!group->__cpu_power) {
6441 printk(KERN_CONT "\n");
6442 printk(KERN_ERR "ERROR: domain->cpu_power not "
6447 if (!cpus_weight(group->cpumask)) {
6448 printk(KERN_CONT "\n");
6449 printk(KERN_ERR "ERROR: empty group\n");
6453 if (cpus_intersects(*groupmask, group->cpumask)) {
6454 printk(KERN_CONT "\n");
6455 printk(KERN_ERR "ERROR: repeated CPUs\n");
6459 cpus_or(*groupmask, *groupmask, group->cpumask);
6461 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6462 printk(KERN_CONT " %s", str);
6464 group = group->next;
6465 } while (group != sd->groups);
6466 printk(KERN_CONT "\n");
6468 if (!cpus_equal(sd->span, *groupmask))
6469 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6471 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6472 printk(KERN_ERR "ERROR: parent span is not a superset "
6473 "of domain->span\n");
6477 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6479 cpumask_t *groupmask;
6483 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6487 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6489 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6491 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6496 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6505 #else /* !CONFIG_SCHED_DEBUG */
6506 # define sched_domain_debug(sd, cpu) do { } while (0)
6507 #endif /* CONFIG_SCHED_DEBUG */
6509 static int sd_degenerate(struct sched_domain *sd)
6511 if (cpus_weight(sd->span) == 1)
6514 /* Following flags need at least 2 groups */
6515 if (sd->flags & (SD_LOAD_BALANCE |
6516 SD_BALANCE_NEWIDLE |
6520 SD_SHARE_PKG_RESOURCES)) {
6521 if (sd->groups != sd->groups->next)
6525 /* Following flags don't use groups */
6526 if (sd->flags & (SD_WAKE_IDLE |
6535 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6537 unsigned long cflags = sd->flags, pflags = parent->flags;
6539 if (sd_degenerate(parent))
6542 if (!cpus_equal(sd->span, parent->span))
6545 /* Does parent contain flags not in child? */
6546 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6547 if (cflags & SD_WAKE_AFFINE)
6548 pflags &= ~SD_WAKE_BALANCE;
6549 /* Flags needing groups don't count if only 1 group in parent */
6550 if (parent->groups == parent->groups->next) {
6551 pflags &= ~(SD_LOAD_BALANCE |
6552 SD_BALANCE_NEWIDLE |
6556 SD_SHARE_PKG_RESOURCES);
6558 if (~cflags & pflags)
6564 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6566 unsigned long flags;
6568 spin_lock_irqsave(&rq->lock, flags);
6571 struct root_domain *old_rd = rq->rd;
6573 if (cpu_isset(rq->cpu, old_rd->online))
6576 cpu_clear(rq->cpu, old_rd->span);
6578 if (atomic_dec_and_test(&old_rd->refcount))
6582 atomic_inc(&rd->refcount);
6585 cpu_set(rq->cpu, rd->span);
6586 if (cpu_isset(rq->cpu, cpu_online_map))
6589 spin_unlock_irqrestore(&rq->lock, flags);
6592 static void init_rootdomain(struct root_domain *rd)
6594 memset(rd, 0, sizeof(*rd));
6596 cpus_clear(rd->span);
6597 cpus_clear(rd->online);
6599 cpupri_init(&rd->cpupri);
6602 static void init_defrootdomain(void)
6604 init_rootdomain(&def_root_domain);
6605 atomic_set(&def_root_domain.refcount, 1);
6608 static struct root_domain *alloc_rootdomain(void)
6610 struct root_domain *rd;
6612 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6616 init_rootdomain(rd);
6622 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6623 * hold the hotplug lock.
6626 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6628 struct rq *rq = cpu_rq(cpu);
6629 struct sched_domain *tmp;
6631 /* Remove the sched domains which do not contribute to scheduling. */
6632 for (tmp = sd; tmp; tmp = tmp->parent) {
6633 struct sched_domain *parent = tmp->parent;
6636 if (sd_parent_degenerate(tmp, parent)) {
6637 tmp->parent = parent->parent;
6639 parent->parent->child = tmp;
6643 if (sd && sd_degenerate(sd)) {
6649 sched_domain_debug(sd, cpu);
6651 rq_attach_root(rq, rd);
6652 rcu_assign_pointer(rq->sd, sd);
6655 /* cpus with isolated domains */
6656 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6658 /* Setup the mask of cpus configured for isolated domains */
6659 static int __init isolated_cpu_setup(char *str)
6661 int ints[NR_CPUS], i;
6663 str = get_options(str, ARRAY_SIZE(ints), ints);
6664 cpus_clear(cpu_isolated_map);
6665 for (i = 1; i <= ints[0]; i++)
6666 if (ints[i] < NR_CPUS)
6667 cpu_set(ints[i], cpu_isolated_map);
6671 __setup("isolcpus=", isolated_cpu_setup);
6674 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6675 * to a function which identifies what group(along with sched group) a CPU
6676 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6677 * (due to the fact that we keep track of groups covered with a cpumask_t).
6679 * init_sched_build_groups will build a circular linked list of the groups
6680 * covered by the given span, and will set each group's ->cpumask correctly,
6681 * and ->cpu_power to 0.
6684 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6685 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6686 struct sched_group **sg,
6687 cpumask_t *tmpmask),
6688 cpumask_t *covered, cpumask_t *tmpmask)
6690 struct sched_group *first = NULL, *last = NULL;
6693 cpus_clear(*covered);
6695 for_each_cpu_mask(i, *span) {
6696 struct sched_group *sg;
6697 int group = group_fn(i, cpu_map, &sg, tmpmask);
6700 if (cpu_isset(i, *covered))
6703 cpus_clear(sg->cpumask);
6704 sg->__cpu_power = 0;
6706 for_each_cpu_mask(j, *span) {
6707 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6710 cpu_set(j, *covered);
6711 cpu_set(j, sg->cpumask);
6722 #define SD_NODES_PER_DOMAIN 16
6727 * find_next_best_node - find the next node to include in a sched_domain
6728 * @node: node whose sched_domain we're building
6729 * @used_nodes: nodes already in the sched_domain
6731 * Find the next node to include in a given scheduling domain. Simply
6732 * finds the closest node not already in the @used_nodes map.
6734 * Should use nodemask_t.
6736 static int find_next_best_node(int node, nodemask_t *used_nodes)
6738 int i, n, val, min_val, best_node = 0;
6742 for (i = 0; i < MAX_NUMNODES; i++) {
6743 /* Start at @node */
6744 n = (node + i) % MAX_NUMNODES;
6746 if (!nr_cpus_node(n))
6749 /* Skip already used nodes */
6750 if (node_isset(n, *used_nodes))
6753 /* Simple min distance search */
6754 val = node_distance(node, n);
6756 if (val < min_val) {
6762 node_set(best_node, *used_nodes);
6767 * sched_domain_node_span - get a cpumask for a node's sched_domain
6768 * @node: node whose cpumask we're constructing
6769 * @span: resulting cpumask
6771 * Given a node, construct a good cpumask for its sched_domain to span. It
6772 * should be one that prevents unnecessary balancing, but also spreads tasks
6775 static void sched_domain_node_span(int node, cpumask_t *span)
6777 nodemask_t used_nodes;
6778 node_to_cpumask_ptr(nodemask, node);
6782 nodes_clear(used_nodes);
6784 cpus_or(*span, *span, *nodemask);
6785 node_set(node, used_nodes);
6787 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6788 int next_node = find_next_best_node(node, &used_nodes);
6790 node_to_cpumask_ptr_next(nodemask, next_node);
6791 cpus_or(*span, *span, *nodemask);
6794 #endif /* CONFIG_NUMA */
6796 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6799 * SMT sched-domains:
6801 #ifdef CONFIG_SCHED_SMT
6802 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6803 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6806 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6810 *sg = &per_cpu(sched_group_cpus, cpu);
6813 #endif /* CONFIG_SCHED_SMT */
6816 * multi-core sched-domains:
6818 #ifdef CONFIG_SCHED_MC
6819 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6820 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6821 #endif /* CONFIG_SCHED_MC */
6823 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6825 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6830 *mask = per_cpu(cpu_sibling_map, cpu);
6831 cpus_and(*mask, *mask, *cpu_map);
6832 group = first_cpu(*mask);
6834 *sg = &per_cpu(sched_group_core, group);
6837 #elif defined(CONFIG_SCHED_MC)
6839 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6843 *sg = &per_cpu(sched_group_core, cpu);
6848 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6849 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6852 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6856 #ifdef CONFIG_SCHED_MC
6857 *mask = cpu_coregroup_map(cpu);
6858 cpus_and(*mask, *mask, *cpu_map);
6859 group = first_cpu(*mask);
6860 #elif defined(CONFIG_SCHED_SMT)
6861 *mask = per_cpu(cpu_sibling_map, cpu);
6862 cpus_and(*mask, *mask, *cpu_map);
6863 group = first_cpu(*mask);
6868 *sg = &per_cpu(sched_group_phys, group);
6874 * The init_sched_build_groups can't handle what we want to do with node
6875 * groups, so roll our own. Now each node has its own list of groups which
6876 * gets dynamically allocated.
6878 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6879 static struct sched_group ***sched_group_nodes_bycpu;
6881 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6882 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6884 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6885 struct sched_group **sg, cpumask_t *nodemask)
6889 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6890 cpus_and(*nodemask, *nodemask, *cpu_map);
6891 group = first_cpu(*nodemask);
6894 *sg = &per_cpu(sched_group_allnodes, group);
6898 static void init_numa_sched_groups_power(struct sched_group *group_head)
6900 struct sched_group *sg = group_head;
6906 for_each_cpu_mask(j, sg->cpumask) {
6907 struct sched_domain *sd;
6909 sd = &per_cpu(phys_domains, j);
6910 if (j != first_cpu(sd->groups->cpumask)) {
6912 * Only add "power" once for each
6918 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6921 } while (sg != group_head);
6923 #endif /* CONFIG_NUMA */
6926 /* Free memory allocated for various sched_group structures */
6927 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6931 for_each_cpu_mask(cpu, *cpu_map) {
6932 struct sched_group **sched_group_nodes
6933 = sched_group_nodes_bycpu[cpu];
6935 if (!sched_group_nodes)
6938 for (i = 0; i < MAX_NUMNODES; i++) {
6939 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6941 *nodemask = node_to_cpumask(i);
6942 cpus_and(*nodemask, *nodemask, *cpu_map);
6943 if (cpus_empty(*nodemask))
6953 if (oldsg != sched_group_nodes[i])
6956 kfree(sched_group_nodes);
6957 sched_group_nodes_bycpu[cpu] = NULL;
6960 #else /* !CONFIG_NUMA */
6961 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6964 #endif /* CONFIG_NUMA */
6967 * Initialize sched groups cpu_power.
6969 * cpu_power indicates the capacity of sched group, which is used while
6970 * distributing the load between different sched groups in a sched domain.
6971 * Typically cpu_power for all the groups in a sched domain will be same unless
6972 * there are asymmetries in the topology. If there are asymmetries, group
6973 * having more cpu_power will pickup more load compared to the group having
6976 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6977 * the maximum number of tasks a group can handle in the presence of other idle
6978 * or lightly loaded groups in the same sched domain.
6980 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6982 struct sched_domain *child;
6983 struct sched_group *group;
6985 WARN_ON(!sd || !sd->groups);
6987 if (cpu != first_cpu(sd->groups->cpumask))
6992 sd->groups->__cpu_power = 0;
6995 * For perf policy, if the groups in child domain share resources
6996 * (for example cores sharing some portions of the cache hierarchy
6997 * or SMT), then set this domain groups cpu_power such that each group
6998 * can handle only one task, when there are other idle groups in the
6999 * same sched domain.
7001 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7003 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7004 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7009 * add cpu_power of each child group to this groups cpu_power
7011 group = child->groups;
7013 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7014 group = group->next;
7015 } while (group != child->groups);
7019 * Initializers for schedule domains
7020 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7023 #define SD_INIT(sd, type) sd_init_##type(sd)
7024 #define SD_INIT_FUNC(type) \
7025 static noinline void sd_init_##type(struct sched_domain *sd) \
7027 memset(sd, 0, sizeof(*sd)); \
7028 *sd = SD_##type##_INIT; \
7029 sd->level = SD_LV_##type; \
7034 SD_INIT_FUNC(ALLNODES)
7037 #ifdef CONFIG_SCHED_SMT
7038 SD_INIT_FUNC(SIBLING)
7040 #ifdef CONFIG_SCHED_MC
7045 * To minimize stack usage kmalloc room for cpumasks and share the
7046 * space as the usage in build_sched_domains() dictates. Used only
7047 * if the amount of space is significant.
7050 cpumask_t tmpmask; /* make this one first */
7053 cpumask_t this_sibling_map;
7054 cpumask_t this_core_map;
7056 cpumask_t send_covered;
7059 cpumask_t domainspan;
7061 cpumask_t notcovered;
7066 #define SCHED_CPUMASK_ALLOC 1
7067 #define SCHED_CPUMASK_FREE(v) kfree(v)
7068 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7070 #define SCHED_CPUMASK_ALLOC 0
7071 #define SCHED_CPUMASK_FREE(v)
7072 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7075 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7076 ((unsigned long)(a) + offsetof(struct allmasks, v))
7078 static int default_relax_domain_level = -1;
7080 static int __init setup_relax_domain_level(char *str)
7084 val = simple_strtoul(str, NULL, 0);
7085 if (val < SD_LV_MAX)
7086 default_relax_domain_level = val;
7090 __setup("relax_domain_level=", setup_relax_domain_level);
7092 static void set_domain_attribute(struct sched_domain *sd,
7093 struct sched_domain_attr *attr)
7097 if (!attr || attr->relax_domain_level < 0) {
7098 if (default_relax_domain_level < 0)
7101 request = default_relax_domain_level;
7103 request = attr->relax_domain_level;
7104 if (request < sd->level) {
7105 /* turn off idle balance on this domain */
7106 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7108 /* turn on idle balance on this domain */
7109 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7114 * Build sched domains for a given set of cpus and attach the sched domains
7115 * to the individual cpus
7117 static int __build_sched_domains(const cpumask_t *cpu_map,
7118 struct sched_domain_attr *attr)
7121 struct root_domain *rd;
7122 SCHED_CPUMASK_DECLARE(allmasks);
7125 struct sched_group **sched_group_nodes = NULL;
7126 int sd_allnodes = 0;
7129 * Allocate the per-node list of sched groups
7131 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7133 if (!sched_group_nodes) {
7134 printk(KERN_WARNING "Can not alloc sched group node list\n");
7139 rd = alloc_rootdomain();
7141 printk(KERN_WARNING "Cannot alloc root domain\n");
7143 kfree(sched_group_nodes);
7148 #if SCHED_CPUMASK_ALLOC
7149 /* get space for all scratch cpumask variables */
7150 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7152 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7155 kfree(sched_group_nodes);
7160 tmpmask = (cpumask_t *)allmasks;
7164 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7168 * Set up domains for cpus specified by the cpu_map.
7170 for_each_cpu_mask(i, *cpu_map) {
7171 struct sched_domain *sd = NULL, *p;
7172 SCHED_CPUMASK_VAR(nodemask, allmasks);
7174 *nodemask = node_to_cpumask(cpu_to_node(i));
7175 cpus_and(*nodemask, *nodemask, *cpu_map);
7178 if (cpus_weight(*cpu_map) >
7179 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7180 sd = &per_cpu(allnodes_domains, i);
7181 SD_INIT(sd, ALLNODES);
7182 set_domain_attribute(sd, attr);
7183 sd->span = *cpu_map;
7184 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7190 sd = &per_cpu(node_domains, i);
7192 set_domain_attribute(sd, attr);
7193 sched_domain_node_span(cpu_to_node(i), &sd->span);
7197 cpus_and(sd->span, sd->span, *cpu_map);
7201 sd = &per_cpu(phys_domains, i);
7203 set_domain_attribute(sd, attr);
7204 sd->span = *nodemask;
7208 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7210 #ifdef CONFIG_SCHED_MC
7212 sd = &per_cpu(core_domains, i);
7214 set_domain_attribute(sd, attr);
7215 sd->span = cpu_coregroup_map(i);
7216 cpus_and(sd->span, sd->span, *cpu_map);
7219 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7222 #ifdef CONFIG_SCHED_SMT
7224 sd = &per_cpu(cpu_domains, i);
7225 SD_INIT(sd, SIBLING);
7226 set_domain_attribute(sd, attr);
7227 sd->span = per_cpu(cpu_sibling_map, i);
7228 cpus_and(sd->span, sd->span, *cpu_map);
7231 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7235 #ifdef CONFIG_SCHED_SMT
7236 /* Set up CPU (sibling) groups */
7237 for_each_cpu_mask(i, *cpu_map) {
7238 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7239 SCHED_CPUMASK_VAR(send_covered, allmasks);
7241 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7242 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7243 if (i != first_cpu(*this_sibling_map))
7246 init_sched_build_groups(this_sibling_map, cpu_map,
7248 send_covered, tmpmask);
7252 #ifdef CONFIG_SCHED_MC
7253 /* Set up multi-core groups */
7254 for_each_cpu_mask(i, *cpu_map) {
7255 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7256 SCHED_CPUMASK_VAR(send_covered, allmasks);
7258 *this_core_map = cpu_coregroup_map(i);
7259 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7260 if (i != first_cpu(*this_core_map))
7263 init_sched_build_groups(this_core_map, cpu_map,
7265 send_covered, tmpmask);
7269 /* Set up physical groups */
7270 for (i = 0; i < MAX_NUMNODES; i++) {
7271 SCHED_CPUMASK_VAR(nodemask, allmasks);
7272 SCHED_CPUMASK_VAR(send_covered, allmasks);
7274 *nodemask = node_to_cpumask(i);
7275 cpus_and(*nodemask, *nodemask, *cpu_map);
7276 if (cpus_empty(*nodemask))
7279 init_sched_build_groups(nodemask, cpu_map,
7281 send_covered, tmpmask);
7285 /* Set up node groups */
7287 SCHED_CPUMASK_VAR(send_covered, allmasks);
7289 init_sched_build_groups(cpu_map, cpu_map,
7290 &cpu_to_allnodes_group,
7291 send_covered, tmpmask);
7294 for (i = 0; i < MAX_NUMNODES; i++) {
7295 /* Set up node groups */
7296 struct sched_group *sg, *prev;
7297 SCHED_CPUMASK_VAR(nodemask, allmasks);
7298 SCHED_CPUMASK_VAR(domainspan, allmasks);
7299 SCHED_CPUMASK_VAR(covered, allmasks);
7302 *nodemask = node_to_cpumask(i);
7303 cpus_clear(*covered);
7305 cpus_and(*nodemask, *nodemask, *cpu_map);
7306 if (cpus_empty(*nodemask)) {
7307 sched_group_nodes[i] = NULL;
7311 sched_domain_node_span(i, domainspan);
7312 cpus_and(*domainspan, *domainspan, *cpu_map);
7314 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7316 printk(KERN_WARNING "Can not alloc domain group for "
7320 sched_group_nodes[i] = sg;
7321 for_each_cpu_mask(j, *nodemask) {
7322 struct sched_domain *sd;
7324 sd = &per_cpu(node_domains, j);
7327 sg->__cpu_power = 0;
7328 sg->cpumask = *nodemask;
7330 cpus_or(*covered, *covered, *nodemask);
7333 for (j = 0; j < MAX_NUMNODES; j++) {
7334 SCHED_CPUMASK_VAR(notcovered, allmasks);
7335 int n = (i + j) % MAX_NUMNODES;
7336 node_to_cpumask_ptr(pnodemask, n);
7338 cpus_complement(*notcovered, *covered);
7339 cpus_and(*tmpmask, *notcovered, *cpu_map);
7340 cpus_and(*tmpmask, *tmpmask, *domainspan);
7341 if (cpus_empty(*tmpmask))
7344 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7345 if (cpus_empty(*tmpmask))
7348 sg = kmalloc_node(sizeof(struct sched_group),
7352 "Can not alloc domain group for node %d\n", j);
7355 sg->__cpu_power = 0;
7356 sg->cpumask = *tmpmask;
7357 sg->next = prev->next;
7358 cpus_or(*covered, *covered, *tmpmask);
7365 /* Calculate CPU power for physical packages and nodes */
7366 #ifdef CONFIG_SCHED_SMT
7367 for_each_cpu_mask(i, *cpu_map) {
7368 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7370 init_sched_groups_power(i, sd);
7373 #ifdef CONFIG_SCHED_MC
7374 for_each_cpu_mask(i, *cpu_map) {
7375 struct sched_domain *sd = &per_cpu(core_domains, i);
7377 init_sched_groups_power(i, sd);
7381 for_each_cpu_mask(i, *cpu_map) {
7382 struct sched_domain *sd = &per_cpu(phys_domains, i);
7384 init_sched_groups_power(i, sd);
7388 for (i = 0; i < MAX_NUMNODES; i++)
7389 init_numa_sched_groups_power(sched_group_nodes[i]);
7392 struct sched_group *sg;
7394 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7396 init_numa_sched_groups_power(sg);
7400 /* Attach the domains */
7401 for_each_cpu_mask(i, *cpu_map) {
7402 struct sched_domain *sd;
7403 #ifdef CONFIG_SCHED_SMT
7404 sd = &per_cpu(cpu_domains, i);
7405 #elif defined(CONFIG_SCHED_MC)
7406 sd = &per_cpu(core_domains, i);
7408 sd = &per_cpu(phys_domains, i);
7410 cpu_attach_domain(sd, rd, i);
7413 SCHED_CPUMASK_FREE((void *)allmasks);
7418 free_sched_groups(cpu_map, tmpmask);
7419 SCHED_CPUMASK_FREE((void *)allmasks);
7424 static int build_sched_domains(const cpumask_t *cpu_map)
7426 return __build_sched_domains(cpu_map, NULL);
7429 static cpumask_t *doms_cur; /* current sched domains */
7430 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7431 static struct sched_domain_attr *dattr_cur;
7432 /* attribues of custom domains in 'doms_cur' */
7435 * Special case: If a kmalloc of a doms_cur partition (array of
7436 * cpumask_t) fails, then fallback to a single sched domain,
7437 * as determined by the single cpumask_t fallback_doms.
7439 static cpumask_t fallback_doms;
7441 void __attribute__((weak)) arch_update_cpu_topology(void)
7446 * Free current domain masks.
7447 * Called after all cpus are attached to NULL domain.
7449 static void free_sched_domains(void)
7452 if (doms_cur != &fallback_doms)
7454 doms_cur = &fallback_doms;
7458 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7459 * For now this just excludes isolated cpus, but could be used to
7460 * exclude other special cases in the future.
7462 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7466 arch_update_cpu_topology();
7468 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7470 doms_cur = &fallback_doms;
7471 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7473 err = build_sched_domains(doms_cur);
7474 register_sched_domain_sysctl();
7479 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7482 free_sched_groups(cpu_map, tmpmask);
7486 * Detach sched domains from a group of cpus specified in cpu_map
7487 * These cpus will now be attached to the NULL domain
7489 static void detach_destroy_domains(const cpumask_t *cpu_map)
7494 unregister_sched_domain_sysctl();
7496 for_each_cpu_mask(i, *cpu_map)
7497 cpu_attach_domain(NULL, &def_root_domain, i);
7498 synchronize_sched();
7499 arch_destroy_sched_domains(cpu_map, &tmpmask);
7502 /* handle null as "default" */
7503 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7504 struct sched_domain_attr *new, int idx_new)
7506 struct sched_domain_attr tmp;
7513 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7514 new ? (new + idx_new) : &tmp,
7515 sizeof(struct sched_domain_attr));
7519 * Partition sched domains as specified by the 'ndoms_new'
7520 * cpumasks in the array doms_new[] of cpumasks. This compares
7521 * doms_new[] to the current sched domain partitioning, doms_cur[].
7522 * It destroys each deleted domain and builds each new domain.
7524 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7525 * The masks don't intersect (don't overlap.) We should setup one
7526 * sched domain for each mask. CPUs not in any of the cpumasks will
7527 * not be load balanced. If the same cpumask appears both in the
7528 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7531 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7532 * ownership of it and will kfree it when done with it. If the caller
7533 * failed the kmalloc call, then it can pass in doms_new == NULL,
7534 * and partition_sched_domains() will fallback to the single partition
7537 * Call with hotplug lock held
7539 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7540 struct sched_domain_attr *dattr_new)
7544 mutex_lock(&sched_domains_mutex);
7546 /* always unregister in case we don't destroy any domains */
7547 unregister_sched_domain_sysctl();
7549 if (doms_new == NULL) {
7551 doms_new = &fallback_doms;
7552 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7556 /* Destroy deleted domains */
7557 for (i = 0; i < ndoms_cur; i++) {
7558 for (j = 0; j < ndoms_new; j++) {
7559 if (cpus_equal(doms_cur[i], doms_new[j])
7560 && dattrs_equal(dattr_cur, i, dattr_new, j))
7563 /* no match - a current sched domain not in new doms_new[] */
7564 detach_destroy_domains(doms_cur + i);
7569 /* Build new domains */
7570 for (i = 0; i < ndoms_new; i++) {
7571 for (j = 0; j < ndoms_cur; j++) {
7572 if (cpus_equal(doms_new[i], doms_cur[j])
7573 && dattrs_equal(dattr_new, i, dattr_cur, j))
7576 /* no match - add a new doms_new */
7577 __build_sched_domains(doms_new + i,
7578 dattr_new ? dattr_new + i : NULL);
7583 /* Remember the new sched domains */
7584 if (doms_cur != &fallback_doms)
7586 kfree(dattr_cur); /* kfree(NULL) is safe */
7587 doms_cur = doms_new;
7588 dattr_cur = dattr_new;
7589 ndoms_cur = ndoms_new;
7591 register_sched_domain_sysctl();
7593 mutex_unlock(&sched_domains_mutex);
7596 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7597 int arch_reinit_sched_domains(void)
7602 mutex_lock(&sched_domains_mutex);
7603 detach_destroy_domains(&cpu_online_map);
7604 free_sched_domains();
7605 err = arch_init_sched_domains(&cpu_online_map);
7606 mutex_unlock(&sched_domains_mutex);
7612 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7616 if (buf[0] != '0' && buf[0] != '1')
7620 sched_smt_power_savings = (buf[0] == '1');
7622 sched_mc_power_savings = (buf[0] == '1');
7624 ret = arch_reinit_sched_domains();
7626 return ret ? ret : count;
7629 #ifdef CONFIG_SCHED_MC
7630 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7632 return sprintf(page, "%u\n", sched_mc_power_savings);
7634 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7635 const char *buf, size_t count)
7637 return sched_power_savings_store(buf, count, 0);
7639 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7640 sched_mc_power_savings_store);
7643 #ifdef CONFIG_SCHED_SMT
7644 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7646 return sprintf(page, "%u\n", sched_smt_power_savings);
7648 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7649 const char *buf, size_t count)
7651 return sched_power_savings_store(buf, count, 1);
7653 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7654 sched_smt_power_savings_store);
7657 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7661 #ifdef CONFIG_SCHED_SMT
7663 err = sysfs_create_file(&cls->kset.kobj,
7664 &attr_sched_smt_power_savings.attr);
7666 #ifdef CONFIG_SCHED_MC
7667 if (!err && mc_capable())
7668 err = sysfs_create_file(&cls->kset.kobj,
7669 &attr_sched_mc_power_savings.attr);
7673 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7676 * Force a reinitialization of the sched domains hierarchy. The domains
7677 * and groups cannot be updated in place without racing with the balancing
7678 * code, so we temporarily attach all running cpus to the NULL domain
7679 * which will prevent rebalancing while the sched domains are recalculated.
7681 static int update_sched_domains(struct notifier_block *nfb,
7682 unsigned long action, void *hcpu)
7684 int cpu = (int)(long)hcpu;
7687 case CPU_DOWN_PREPARE:
7688 case CPU_DOWN_PREPARE_FROZEN:
7689 disable_runtime(cpu_rq(cpu));
7691 case CPU_UP_PREPARE:
7692 case CPU_UP_PREPARE_FROZEN:
7693 detach_destroy_domains(&cpu_online_map);
7694 free_sched_domains();
7698 case CPU_DOWN_FAILED:
7699 case CPU_DOWN_FAILED_FROZEN:
7701 case CPU_ONLINE_FROZEN:
7702 enable_runtime(cpu_rq(cpu));
7704 case CPU_UP_CANCELED:
7705 case CPU_UP_CANCELED_FROZEN:
7707 case CPU_DEAD_FROZEN:
7709 * Fall through and re-initialise the domains.
7716 #ifndef CONFIG_CPUSETS
7718 * Create default domain partitioning if cpusets are disabled.
7719 * Otherwise we let cpusets rebuild the domains based on the
7723 /* The hotplug lock is already held by cpu_up/cpu_down */
7724 arch_init_sched_domains(&cpu_online_map);
7730 void __init sched_init_smp(void)
7732 cpumask_t non_isolated_cpus;
7734 #if defined(CONFIG_NUMA)
7735 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7737 BUG_ON(sched_group_nodes_bycpu == NULL);
7740 mutex_lock(&sched_domains_mutex);
7741 arch_init_sched_domains(&cpu_online_map);
7742 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7743 if (cpus_empty(non_isolated_cpus))
7744 cpu_set(smp_processor_id(), non_isolated_cpus);
7745 mutex_unlock(&sched_domains_mutex);
7747 /* XXX: Theoretical race here - CPU may be hotplugged now */
7748 hotcpu_notifier(update_sched_domains, 0);
7751 /* Move init over to a non-isolated CPU */
7752 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7754 sched_init_granularity();
7757 void __init sched_init_smp(void)
7759 sched_init_granularity();
7761 #endif /* CONFIG_SMP */
7763 int in_sched_functions(unsigned long addr)
7765 return in_lock_functions(addr) ||
7766 (addr >= (unsigned long)__sched_text_start
7767 && addr < (unsigned long)__sched_text_end);
7770 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7772 cfs_rq->tasks_timeline = RB_ROOT;
7773 INIT_LIST_HEAD(&cfs_rq->tasks);
7774 #ifdef CONFIG_FAIR_GROUP_SCHED
7777 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7780 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7782 struct rt_prio_array *array;
7785 array = &rt_rq->active;
7786 for (i = 0; i < MAX_RT_PRIO; i++) {
7787 INIT_LIST_HEAD(array->queue + i);
7788 __clear_bit(i, array->bitmap);
7790 /* delimiter for bitsearch: */
7791 __set_bit(MAX_RT_PRIO, array->bitmap);
7793 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7794 rt_rq->highest_prio = MAX_RT_PRIO;
7797 rt_rq->rt_nr_migratory = 0;
7798 rt_rq->overloaded = 0;
7802 rt_rq->rt_throttled = 0;
7803 rt_rq->rt_runtime = 0;
7804 spin_lock_init(&rt_rq->rt_runtime_lock);
7806 #ifdef CONFIG_RT_GROUP_SCHED
7807 rt_rq->rt_nr_boosted = 0;
7812 #ifdef CONFIG_FAIR_GROUP_SCHED
7813 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7814 struct sched_entity *se, int cpu, int add,
7815 struct sched_entity *parent)
7817 struct rq *rq = cpu_rq(cpu);
7818 tg->cfs_rq[cpu] = cfs_rq;
7819 init_cfs_rq(cfs_rq, rq);
7822 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7825 /* se could be NULL for init_task_group */
7830 se->cfs_rq = &rq->cfs;
7832 se->cfs_rq = parent->my_q;
7835 se->load.weight = tg->shares;
7836 se->load.inv_weight = 0;
7837 se->parent = parent;
7841 #ifdef CONFIG_RT_GROUP_SCHED
7842 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7843 struct sched_rt_entity *rt_se, int cpu, int add,
7844 struct sched_rt_entity *parent)
7846 struct rq *rq = cpu_rq(cpu);
7848 tg->rt_rq[cpu] = rt_rq;
7849 init_rt_rq(rt_rq, rq);
7851 rt_rq->rt_se = rt_se;
7852 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7854 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7856 tg->rt_se[cpu] = rt_se;
7861 rt_se->rt_rq = &rq->rt;
7863 rt_se->rt_rq = parent->my_q;
7865 rt_se->my_q = rt_rq;
7866 rt_se->parent = parent;
7867 INIT_LIST_HEAD(&rt_se->run_list);
7871 void __init sched_init(void)
7874 unsigned long alloc_size = 0, ptr;
7876 #ifdef CONFIG_FAIR_GROUP_SCHED
7877 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7879 #ifdef CONFIG_RT_GROUP_SCHED
7880 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7882 #ifdef CONFIG_USER_SCHED
7886 * As sched_init() is called before page_alloc is setup,
7887 * we use alloc_bootmem().
7890 ptr = (unsigned long)alloc_bootmem(alloc_size);
7892 #ifdef CONFIG_FAIR_GROUP_SCHED
7893 init_task_group.se = (struct sched_entity **)ptr;
7894 ptr += nr_cpu_ids * sizeof(void **);
7896 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7897 ptr += nr_cpu_ids * sizeof(void **);
7899 #ifdef CONFIG_USER_SCHED
7900 root_task_group.se = (struct sched_entity **)ptr;
7901 ptr += nr_cpu_ids * sizeof(void **);
7903 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7904 ptr += nr_cpu_ids * sizeof(void **);
7905 #endif /* CONFIG_USER_SCHED */
7906 #endif /* CONFIG_FAIR_GROUP_SCHED */
7907 #ifdef CONFIG_RT_GROUP_SCHED
7908 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7909 ptr += nr_cpu_ids * sizeof(void **);
7911 init_task_group.rt_rq = (struct rt_rq **)ptr;
7912 ptr += nr_cpu_ids * sizeof(void **);
7914 #ifdef CONFIG_USER_SCHED
7915 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7916 ptr += nr_cpu_ids * sizeof(void **);
7918 root_task_group.rt_rq = (struct rt_rq **)ptr;
7919 ptr += nr_cpu_ids * sizeof(void **);
7920 #endif /* CONFIG_USER_SCHED */
7921 #endif /* CONFIG_RT_GROUP_SCHED */
7925 init_defrootdomain();
7928 init_rt_bandwidth(&def_rt_bandwidth,
7929 global_rt_period(), global_rt_runtime());
7931 #ifdef CONFIG_RT_GROUP_SCHED
7932 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7933 global_rt_period(), global_rt_runtime());
7934 #ifdef CONFIG_USER_SCHED
7935 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7936 global_rt_period(), RUNTIME_INF);
7937 #endif /* CONFIG_USER_SCHED */
7938 #endif /* CONFIG_RT_GROUP_SCHED */
7940 #ifdef CONFIG_GROUP_SCHED
7941 list_add(&init_task_group.list, &task_groups);
7942 INIT_LIST_HEAD(&init_task_group.children);
7944 #ifdef CONFIG_USER_SCHED
7945 INIT_LIST_HEAD(&root_task_group.children);
7946 init_task_group.parent = &root_task_group;
7947 list_add(&init_task_group.siblings, &root_task_group.children);
7948 #endif /* CONFIG_USER_SCHED */
7949 #endif /* CONFIG_GROUP_SCHED */
7951 for_each_possible_cpu(i) {
7955 spin_lock_init(&rq->lock);
7956 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7958 init_cfs_rq(&rq->cfs, rq);
7959 init_rt_rq(&rq->rt, rq);
7960 #ifdef CONFIG_FAIR_GROUP_SCHED
7961 init_task_group.shares = init_task_group_load;
7962 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7963 #ifdef CONFIG_CGROUP_SCHED
7965 * How much cpu bandwidth does init_task_group get?
7967 * In case of task-groups formed thr' the cgroup filesystem, it
7968 * gets 100% of the cpu resources in the system. This overall
7969 * system cpu resource is divided among the tasks of
7970 * init_task_group and its child task-groups in a fair manner,
7971 * based on each entity's (task or task-group's) weight
7972 * (se->load.weight).
7974 * In other words, if init_task_group has 10 tasks of weight
7975 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7976 * then A0's share of the cpu resource is:
7978 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7980 * We achieve this by letting init_task_group's tasks sit
7981 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7983 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7984 #elif defined CONFIG_USER_SCHED
7985 root_task_group.shares = NICE_0_LOAD;
7986 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
7988 * In case of task-groups formed thr' the user id of tasks,
7989 * init_task_group represents tasks belonging to root user.
7990 * Hence it forms a sibling of all subsequent groups formed.
7991 * In this case, init_task_group gets only a fraction of overall
7992 * system cpu resource, based on the weight assigned to root
7993 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7994 * by letting tasks of init_task_group sit in a separate cfs_rq
7995 * (init_cfs_rq) and having one entity represent this group of
7996 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7998 init_tg_cfs_entry(&init_task_group,
7999 &per_cpu(init_cfs_rq, i),
8000 &per_cpu(init_sched_entity, i), i, 1,
8001 root_task_group.se[i]);
8004 #endif /* CONFIG_FAIR_GROUP_SCHED */
8006 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8007 #ifdef CONFIG_RT_GROUP_SCHED
8008 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8009 #ifdef CONFIG_CGROUP_SCHED
8010 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8011 #elif defined CONFIG_USER_SCHED
8012 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8013 init_tg_rt_entry(&init_task_group,
8014 &per_cpu(init_rt_rq, i),
8015 &per_cpu(init_sched_rt_entity, i), i, 1,
8016 root_task_group.rt_se[i]);
8020 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8021 rq->cpu_load[j] = 0;
8025 rq->active_balance = 0;
8026 rq->next_balance = jiffies;
8030 rq->migration_thread = NULL;
8031 INIT_LIST_HEAD(&rq->migration_queue);
8032 rq_attach_root(rq, &def_root_domain);
8035 atomic_set(&rq->nr_iowait, 0);
8038 set_load_weight(&init_task);
8040 #ifdef CONFIG_PREEMPT_NOTIFIERS
8041 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8045 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8048 #ifdef CONFIG_RT_MUTEXES
8049 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8053 * The boot idle thread does lazy MMU switching as well:
8055 atomic_inc(&init_mm.mm_count);
8056 enter_lazy_tlb(&init_mm, current);
8059 * Make us the idle thread. Technically, schedule() should not be
8060 * called from this thread, however somewhere below it might be,
8061 * but because we are the idle thread, we just pick up running again
8062 * when this runqueue becomes "idle".
8064 init_idle(current, smp_processor_id());
8066 * During early bootup we pretend to be a normal task:
8068 current->sched_class = &fair_sched_class;
8070 scheduler_running = 1;
8073 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8074 void __might_sleep(char *file, int line)
8077 static unsigned long prev_jiffy; /* ratelimiting */
8079 if ((in_atomic() || irqs_disabled()) &&
8080 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8081 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8083 prev_jiffy = jiffies;
8084 printk(KERN_ERR "BUG: sleeping function called from invalid"
8085 " context at %s:%d\n", file, line);
8086 printk("in_atomic():%d, irqs_disabled():%d\n",
8087 in_atomic(), irqs_disabled());
8088 debug_show_held_locks(current);
8089 if (irqs_disabled())
8090 print_irqtrace_events(current);
8095 EXPORT_SYMBOL(__might_sleep);
8098 #ifdef CONFIG_MAGIC_SYSRQ
8099 static void normalize_task(struct rq *rq, struct task_struct *p)
8103 update_rq_clock(rq);
8104 on_rq = p->se.on_rq;
8106 deactivate_task(rq, p, 0);
8107 __setscheduler(rq, p, SCHED_NORMAL, 0);
8109 activate_task(rq, p, 0);
8110 resched_task(rq->curr);
8114 void normalize_rt_tasks(void)
8116 struct task_struct *g, *p;
8117 unsigned long flags;
8120 read_lock_irqsave(&tasklist_lock, flags);
8121 do_each_thread(g, p) {
8123 * Only normalize user tasks:
8128 p->se.exec_start = 0;
8129 #ifdef CONFIG_SCHEDSTATS
8130 p->se.wait_start = 0;
8131 p->se.sleep_start = 0;
8132 p->se.block_start = 0;
8137 * Renice negative nice level userspace
8140 if (TASK_NICE(p) < 0 && p->mm)
8141 set_user_nice(p, 0);
8145 spin_lock(&p->pi_lock);
8146 rq = __task_rq_lock(p);
8148 normalize_task(rq, p);
8150 __task_rq_unlock(rq);
8151 spin_unlock(&p->pi_lock);
8152 } while_each_thread(g, p);
8154 read_unlock_irqrestore(&tasklist_lock, flags);
8157 #endif /* CONFIG_MAGIC_SYSRQ */
8161 * These functions are only useful for the IA64 MCA handling.
8163 * They can only be called when the whole system has been
8164 * stopped - every CPU needs to be quiescent, and no scheduling
8165 * activity can take place. Using them for anything else would
8166 * be a serious bug, and as a result, they aren't even visible
8167 * under any other configuration.
8171 * curr_task - return the current task for a given cpu.
8172 * @cpu: the processor in question.
8174 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8176 struct task_struct *curr_task(int cpu)
8178 return cpu_curr(cpu);
8182 * set_curr_task - set the current task for a given cpu.
8183 * @cpu: the processor in question.
8184 * @p: the task pointer to set.
8186 * Description: This function must only be used when non-maskable interrupts
8187 * are serviced on a separate stack. It allows the architecture to switch the
8188 * notion of the current task on a cpu in a non-blocking manner. This function
8189 * must be called with all CPU's synchronized, and interrupts disabled, the
8190 * and caller must save the original value of the current task (see
8191 * curr_task() above) and restore that value before reenabling interrupts and
8192 * re-starting the system.
8194 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8196 void set_curr_task(int cpu, struct task_struct *p)
8203 #ifdef CONFIG_FAIR_GROUP_SCHED
8204 static void free_fair_sched_group(struct task_group *tg)
8208 for_each_possible_cpu(i) {
8210 kfree(tg->cfs_rq[i]);
8220 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8222 struct cfs_rq *cfs_rq;
8223 struct sched_entity *se, *parent_se;
8227 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8230 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8234 tg->shares = NICE_0_LOAD;
8236 for_each_possible_cpu(i) {
8239 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8240 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8244 se = kmalloc_node(sizeof(struct sched_entity),
8245 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8249 parent_se = parent ? parent->se[i] : NULL;
8250 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8259 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8261 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8262 &cpu_rq(cpu)->leaf_cfs_rq_list);
8265 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8267 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8269 #else /* !CONFG_FAIR_GROUP_SCHED */
8270 static inline void free_fair_sched_group(struct task_group *tg)
8275 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8280 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8284 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8287 #endif /* CONFIG_FAIR_GROUP_SCHED */
8289 #ifdef CONFIG_RT_GROUP_SCHED
8290 static void free_rt_sched_group(struct task_group *tg)
8294 destroy_rt_bandwidth(&tg->rt_bandwidth);
8296 for_each_possible_cpu(i) {
8298 kfree(tg->rt_rq[i]);
8300 kfree(tg->rt_se[i]);
8308 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8310 struct rt_rq *rt_rq;
8311 struct sched_rt_entity *rt_se, *parent_se;
8315 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8318 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8322 init_rt_bandwidth(&tg->rt_bandwidth,
8323 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8325 for_each_possible_cpu(i) {
8328 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8329 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8333 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8334 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8338 parent_se = parent ? parent->rt_se[i] : NULL;
8339 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8348 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8350 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8351 &cpu_rq(cpu)->leaf_rt_rq_list);
8354 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8356 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8358 #else /* !CONFIG_RT_GROUP_SCHED */
8359 static inline void free_rt_sched_group(struct task_group *tg)
8364 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8369 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8373 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8376 #endif /* CONFIG_RT_GROUP_SCHED */
8378 #ifdef CONFIG_GROUP_SCHED
8379 static void free_sched_group(struct task_group *tg)
8381 free_fair_sched_group(tg);
8382 free_rt_sched_group(tg);
8386 /* allocate runqueue etc for a new task group */
8387 struct task_group *sched_create_group(struct task_group *parent)
8389 struct task_group *tg;
8390 unsigned long flags;
8393 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8395 return ERR_PTR(-ENOMEM);
8397 if (!alloc_fair_sched_group(tg, parent))
8400 if (!alloc_rt_sched_group(tg, parent))
8403 spin_lock_irqsave(&task_group_lock, flags);
8404 for_each_possible_cpu(i) {
8405 register_fair_sched_group(tg, i);
8406 register_rt_sched_group(tg, i);
8408 list_add_rcu(&tg->list, &task_groups);
8410 WARN_ON(!parent); /* root should already exist */
8412 tg->parent = parent;
8413 list_add_rcu(&tg->siblings, &parent->children);
8414 INIT_LIST_HEAD(&tg->children);
8415 spin_unlock_irqrestore(&task_group_lock, flags);
8420 free_sched_group(tg);
8421 return ERR_PTR(-ENOMEM);
8424 /* rcu callback to free various structures associated with a task group */
8425 static void free_sched_group_rcu(struct rcu_head *rhp)
8427 /* now it should be safe to free those cfs_rqs */
8428 free_sched_group(container_of(rhp, struct task_group, rcu));
8431 /* Destroy runqueue etc associated with a task group */
8432 void sched_destroy_group(struct task_group *tg)
8434 unsigned long flags;
8437 spin_lock_irqsave(&task_group_lock, flags);
8438 for_each_possible_cpu(i) {
8439 unregister_fair_sched_group(tg, i);
8440 unregister_rt_sched_group(tg, i);
8442 list_del_rcu(&tg->list);
8443 list_del_rcu(&tg->siblings);
8444 spin_unlock_irqrestore(&task_group_lock, flags);
8446 /* wait for possible concurrent references to cfs_rqs complete */
8447 call_rcu(&tg->rcu, free_sched_group_rcu);
8450 /* change task's runqueue when it moves between groups.
8451 * The caller of this function should have put the task in its new group
8452 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8453 * reflect its new group.
8455 void sched_move_task(struct task_struct *tsk)
8458 unsigned long flags;
8461 rq = task_rq_lock(tsk, &flags);
8463 update_rq_clock(rq);
8465 running = task_current(rq, tsk);
8466 on_rq = tsk->se.on_rq;
8469 dequeue_task(rq, tsk, 0);
8470 if (unlikely(running))
8471 tsk->sched_class->put_prev_task(rq, tsk);
8473 set_task_rq(tsk, task_cpu(tsk));
8475 #ifdef CONFIG_FAIR_GROUP_SCHED
8476 if (tsk->sched_class->moved_group)
8477 tsk->sched_class->moved_group(tsk);
8480 if (unlikely(running))
8481 tsk->sched_class->set_curr_task(rq);
8483 enqueue_task(rq, tsk, 0);
8485 task_rq_unlock(rq, &flags);
8487 #endif /* CONFIG_GROUP_SCHED */
8489 #ifdef CONFIG_FAIR_GROUP_SCHED
8490 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8492 struct cfs_rq *cfs_rq = se->cfs_rq;
8497 dequeue_entity(cfs_rq, se, 0);
8499 se->load.weight = shares;
8500 se->load.inv_weight = 0;
8503 enqueue_entity(cfs_rq, se, 0);
8506 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8508 struct cfs_rq *cfs_rq = se->cfs_rq;
8509 struct rq *rq = cfs_rq->rq;
8510 unsigned long flags;
8512 spin_lock_irqsave(&rq->lock, flags);
8513 __set_se_shares(se, shares);
8514 spin_unlock_irqrestore(&rq->lock, flags);
8517 static DEFINE_MUTEX(shares_mutex);
8519 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8522 unsigned long flags;
8525 * We can't change the weight of the root cgroup.
8530 if (shares < MIN_SHARES)
8531 shares = MIN_SHARES;
8532 else if (shares > MAX_SHARES)
8533 shares = MAX_SHARES;
8535 mutex_lock(&shares_mutex);
8536 if (tg->shares == shares)
8539 spin_lock_irqsave(&task_group_lock, flags);
8540 for_each_possible_cpu(i)
8541 unregister_fair_sched_group(tg, i);
8542 list_del_rcu(&tg->siblings);
8543 spin_unlock_irqrestore(&task_group_lock, flags);
8545 /* wait for any ongoing reference to this group to finish */
8546 synchronize_sched();
8549 * Now we are free to modify the group's share on each cpu
8550 * w/o tripping rebalance_share or load_balance_fair.
8552 tg->shares = shares;
8553 for_each_possible_cpu(i) {
8557 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8558 set_se_shares(tg->se[i], shares);
8562 * Enable load balance activity on this group, by inserting it back on
8563 * each cpu's rq->leaf_cfs_rq_list.
8565 spin_lock_irqsave(&task_group_lock, flags);
8566 for_each_possible_cpu(i)
8567 register_fair_sched_group(tg, i);
8568 list_add_rcu(&tg->siblings, &tg->parent->children);
8569 spin_unlock_irqrestore(&task_group_lock, flags);
8571 mutex_unlock(&shares_mutex);
8575 unsigned long sched_group_shares(struct task_group *tg)
8581 #ifdef CONFIG_RT_GROUP_SCHED
8583 * Ensure that the real time constraints are schedulable.
8585 static DEFINE_MUTEX(rt_constraints_mutex);
8587 static unsigned long to_ratio(u64 period, u64 runtime)
8589 if (runtime == RUNTIME_INF)
8592 return div64_u64(runtime << 16, period);
8595 #ifdef CONFIG_CGROUP_SCHED
8596 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8598 struct task_group *tgi, *parent = tg->parent;
8599 unsigned long total = 0;
8602 if (global_rt_period() < period)
8605 return to_ratio(period, runtime) <
8606 to_ratio(global_rt_period(), global_rt_runtime());
8609 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8613 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8617 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8618 tgi->rt_bandwidth.rt_runtime);
8622 return total + to_ratio(period, runtime) <=
8623 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8624 parent->rt_bandwidth.rt_runtime);
8626 #elif defined CONFIG_USER_SCHED
8627 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8629 struct task_group *tgi;
8630 unsigned long total = 0;
8631 unsigned long global_ratio =
8632 to_ratio(global_rt_period(), global_rt_runtime());
8635 list_for_each_entry_rcu(tgi, &task_groups, list) {
8639 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8640 tgi->rt_bandwidth.rt_runtime);
8644 return total + to_ratio(period, runtime) < global_ratio;
8648 /* Must be called with tasklist_lock held */
8649 static inline int tg_has_rt_tasks(struct task_group *tg)
8651 struct task_struct *g, *p;
8652 do_each_thread(g, p) {
8653 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8655 } while_each_thread(g, p);
8659 static int tg_set_bandwidth(struct task_group *tg,
8660 u64 rt_period, u64 rt_runtime)
8664 mutex_lock(&rt_constraints_mutex);
8665 read_lock(&tasklist_lock);
8666 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8670 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8675 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8676 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8677 tg->rt_bandwidth.rt_runtime = rt_runtime;
8679 for_each_possible_cpu(i) {
8680 struct rt_rq *rt_rq = tg->rt_rq[i];
8682 spin_lock(&rt_rq->rt_runtime_lock);
8683 rt_rq->rt_runtime = rt_runtime;
8684 spin_unlock(&rt_rq->rt_runtime_lock);
8686 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8688 read_unlock(&tasklist_lock);
8689 mutex_unlock(&rt_constraints_mutex);
8694 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8696 u64 rt_runtime, rt_period;
8698 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8699 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8700 if (rt_runtime_us < 0)
8701 rt_runtime = RUNTIME_INF;
8703 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8706 long sched_group_rt_runtime(struct task_group *tg)
8710 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8713 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8714 do_div(rt_runtime_us, NSEC_PER_USEC);
8715 return rt_runtime_us;
8718 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8720 u64 rt_runtime, rt_period;
8722 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8723 rt_runtime = tg->rt_bandwidth.rt_runtime;
8725 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8728 long sched_group_rt_period(struct task_group *tg)
8732 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8733 do_div(rt_period_us, NSEC_PER_USEC);
8734 return rt_period_us;
8737 static int sched_rt_global_constraints(void)
8739 struct task_group *tg = &root_task_group;
8740 u64 rt_runtime, rt_period;
8743 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8744 rt_runtime = tg->rt_bandwidth.rt_runtime;
8746 mutex_lock(&rt_constraints_mutex);
8747 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8749 mutex_unlock(&rt_constraints_mutex);
8753 #else /* !CONFIG_RT_GROUP_SCHED */
8754 static int sched_rt_global_constraints(void)
8756 unsigned long flags;
8759 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8760 for_each_possible_cpu(i) {
8761 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8763 spin_lock(&rt_rq->rt_runtime_lock);
8764 rt_rq->rt_runtime = global_rt_runtime();
8765 spin_unlock(&rt_rq->rt_runtime_lock);
8767 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8771 #endif /* CONFIG_RT_GROUP_SCHED */
8773 int sched_rt_handler(struct ctl_table *table, int write,
8774 struct file *filp, void __user *buffer, size_t *lenp,
8778 int old_period, old_runtime;
8779 static DEFINE_MUTEX(mutex);
8782 old_period = sysctl_sched_rt_period;
8783 old_runtime = sysctl_sched_rt_runtime;
8785 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8787 if (!ret && write) {
8788 ret = sched_rt_global_constraints();
8790 sysctl_sched_rt_period = old_period;
8791 sysctl_sched_rt_runtime = old_runtime;
8793 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8794 def_rt_bandwidth.rt_period =
8795 ns_to_ktime(global_rt_period());
8798 mutex_unlock(&mutex);
8803 #ifdef CONFIG_CGROUP_SCHED
8805 /* return corresponding task_group object of a cgroup */
8806 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8808 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8809 struct task_group, css);
8812 static struct cgroup_subsys_state *
8813 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8815 struct task_group *tg, *parent;
8817 if (!cgrp->parent) {
8818 /* This is early initialization for the top cgroup */
8819 init_task_group.css.cgroup = cgrp;
8820 return &init_task_group.css;
8823 parent = cgroup_tg(cgrp->parent);
8824 tg = sched_create_group(parent);
8826 return ERR_PTR(-ENOMEM);
8828 /* Bind the cgroup to task_group object we just created */
8829 tg->css.cgroup = cgrp;
8835 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8837 struct task_group *tg = cgroup_tg(cgrp);
8839 sched_destroy_group(tg);
8843 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8844 struct task_struct *tsk)
8846 #ifdef CONFIG_RT_GROUP_SCHED
8847 /* Don't accept realtime tasks when there is no way for them to run */
8848 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8851 /* We don't support RT-tasks being in separate groups */
8852 if (tsk->sched_class != &fair_sched_class)
8860 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8861 struct cgroup *old_cont, struct task_struct *tsk)
8863 sched_move_task(tsk);
8866 #ifdef CONFIG_FAIR_GROUP_SCHED
8867 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8870 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8873 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8875 struct task_group *tg = cgroup_tg(cgrp);
8877 return (u64) tg->shares;
8879 #endif /* CONFIG_FAIR_GROUP_SCHED */
8881 #ifdef CONFIG_RT_GROUP_SCHED
8882 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8885 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8888 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8890 return sched_group_rt_runtime(cgroup_tg(cgrp));
8893 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8896 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8899 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8901 return sched_group_rt_period(cgroup_tg(cgrp));
8903 #endif /* CONFIG_RT_GROUP_SCHED */
8905 static struct cftype cpu_files[] = {
8906 #ifdef CONFIG_FAIR_GROUP_SCHED
8909 .read_u64 = cpu_shares_read_u64,
8910 .write_u64 = cpu_shares_write_u64,
8913 #ifdef CONFIG_RT_GROUP_SCHED
8915 .name = "rt_runtime_us",
8916 .read_s64 = cpu_rt_runtime_read,
8917 .write_s64 = cpu_rt_runtime_write,
8920 .name = "rt_period_us",
8921 .read_u64 = cpu_rt_period_read_uint,
8922 .write_u64 = cpu_rt_period_write_uint,
8927 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8929 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8932 struct cgroup_subsys cpu_cgroup_subsys = {
8934 .create = cpu_cgroup_create,
8935 .destroy = cpu_cgroup_destroy,
8936 .can_attach = cpu_cgroup_can_attach,
8937 .attach = cpu_cgroup_attach,
8938 .populate = cpu_cgroup_populate,
8939 .subsys_id = cpu_cgroup_subsys_id,
8943 #endif /* CONFIG_CGROUP_SCHED */
8945 #ifdef CONFIG_CGROUP_CPUACCT
8948 * CPU accounting code for task groups.
8950 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8951 * (balbir@in.ibm.com).
8954 /* track cpu usage of a group of tasks */
8956 struct cgroup_subsys_state css;
8957 /* cpuusage holds pointer to a u64-type object on every cpu */
8961 struct cgroup_subsys cpuacct_subsys;
8963 /* return cpu accounting group corresponding to this container */
8964 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8966 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8967 struct cpuacct, css);
8970 /* return cpu accounting group to which this task belongs */
8971 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8973 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8974 struct cpuacct, css);
8977 /* create a new cpu accounting group */
8978 static struct cgroup_subsys_state *cpuacct_create(
8979 struct cgroup_subsys *ss, struct cgroup *cgrp)
8981 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8984 return ERR_PTR(-ENOMEM);
8986 ca->cpuusage = alloc_percpu(u64);
8987 if (!ca->cpuusage) {
8989 return ERR_PTR(-ENOMEM);
8995 /* destroy an existing cpu accounting group */
8997 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8999 struct cpuacct *ca = cgroup_ca(cgrp);
9001 free_percpu(ca->cpuusage);
9005 /* return total cpu usage (in nanoseconds) of a group */
9006 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9008 struct cpuacct *ca = cgroup_ca(cgrp);
9009 u64 totalcpuusage = 0;
9012 for_each_possible_cpu(i) {
9013 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9016 * Take rq->lock to make 64-bit addition safe on 32-bit
9019 spin_lock_irq(&cpu_rq(i)->lock);
9020 totalcpuusage += *cpuusage;
9021 spin_unlock_irq(&cpu_rq(i)->lock);
9024 return totalcpuusage;
9027 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9030 struct cpuacct *ca = cgroup_ca(cgrp);
9039 for_each_possible_cpu(i) {
9040 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9042 spin_lock_irq(&cpu_rq(i)->lock);
9044 spin_unlock_irq(&cpu_rq(i)->lock);
9050 static struct cftype files[] = {
9053 .read_u64 = cpuusage_read,
9054 .write_u64 = cpuusage_write,
9058 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9060 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9064 * charge this task's execution time to its accounting group.
9066 * called with rq->lock held.
9068 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9072 if (!cpuacct_subsys.active)
9077 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9079 *cpuusage += cputime;
9083 struct cgroup_subsys cpuacct_subsys = {
9085 .create = cpuacct_create,
9086 .destroy = cpuacct_destroy,
9087 .populate = cpuacct_populate,
9088 .subsys_id = cpuacct_subsys_id,
9090 #endif /* CONFIG_CGROUP_CPUACCT */