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 * period over which we measure -rt task cpu usage in us.
784 unsigned int sysctl_sched_rt_period = 1000000;
786 static __read_mostly int scheduler_running;
789 * part of the period that we allow rt tasks to run in us.
792 int sysctl_sched_rt_runtime = 950000;
794 static inline u64 global_rt_period(void)
796 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
799 static inline u64 global_rt_runtime(void)
801 if (sysctl_sched_rt_period < 0)
804 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
807 #ifndef prepare_arch_switch
808 # define prepare_arch_switch(next) do { } while (0)
810 #ifndef finish_arch_switch
811 # define finish_arch_switch(prev) do { } while (0)
814 static inline int task_current(struct rq *rq, struct task_struct *p)
816 return rq->curr == p;
819 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
820 static inline int task_running(struct rq *rq, struct task_struct *p)
822 return task_current(rq, p);
825 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
829 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
831 #ifdef CONFIG_DEBUG_SPINLOCK
832 /* this is a valid case when another task releases the spinlock */
833 rq->lock.owner = current;
836 * If we are tracking spinlock dependencies then we have to
837 * fix up the runqueue lock - which gets 'carried over' from
840 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
842 spin_unlock_irq(&rq->lock);
845 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
846 static inline int task_running(struct rq *rq, struct task_struct *p)
851 return task_current(rq, p);
855 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
859 * We can optimise this out completely for !SMP, because the
860 * SMP rebalancing from interrupt is the only thing that cares
865 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
866 spin_unlock_irq(&rq->lock);
868 spin_unlock(&rq->lock);
872 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
876 * After ->oncpu is cleared, the task can be moved to a different CPU.
877 * We must ensure this doesn't happen until the switch is completely
883 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
887 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
890 * __task_rq_lock - lock the runqueue a given task resides on.
891 * Must be called interrupts disabled.
893 static inline struct rq *__task_rq_lock(struct task_struct *p)
897 struct rq *rq = task_rq(p);
898 spin_lock(&rq->lock);
899 if (likely(rq == task_rq(p)))
901 spin_unlock(&rq->lock);
906 * task_rq_lock - lock the runqueue a given task resides on and disable
907 * interrupts. Note the ordering: we can safely lookup the task_rq without
908 * explicitly disabling preemption.
910 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
916 local_irq_save(*flags);
918 spin_lock(&rq->lock);
919 if (likely(rq == task_rq(p)))
921 spin_unlock_irqrestore(&rq->lock, *flags);
925 static void __task_rq_unlock(struct rq *rq)
928 spin_unlock(&rq->lock);
931 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
934 spin_unlock_irqrestore(&rq->lock, *flags);
938 * this_rq_lock - lock this runqueue and disable interrupts.
940 static struct rq *this_rq_lock(void)
947 spin_lock(&rq->lock);
952 static void __resched_task(struct task_struct *p, int tif_bit);
954 static inline void resched_task(struct task_struct *p)
956 __resched_task(p, TIF_NEED_RESCHED);
959 #ifdef CONFIG_SCHED_HRTICK
961 * Use HR-timers to deliver accurate preemption points.
963 * Its all a bit involved since we cannot program an hrt while holding the
964 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
967 * When we get rescheduled we reprogram the hrtick_timer outside of the
970 static inline void resched_hrt(struct task_struct *p)
972 __resched_task(p, TIF_HRTICK_RESCHED);
975 static inline void resched_rq(struct rq *rq)
979 spin_lock_irqsave(&rq->lock, flags);
980 resched_task(rq->curr);
981 spin_unlock_irqrestore(&rq->lock, flags);
985 HRTICK_SET, /* re-programm hrtick_timer */
986 HRTICK_RESET, /* not a new slice */
987 HRTICK_BLOCK, /* stop hrtick operations */
992 * - enabled by features
993 * - hrtimer is actually high res
995 static inline int hrtick_enabled(struct rq *rq)
997 if (!sched_feat(HRTICK))
999 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1001 return hrtimer_is_hres_active(&rq->hrtick_timer);
1005 * Called to set the hrtick timer state.
1007 * called with rq->lock held and irqs disabled
1009 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1011 assert_spin_locked(&rq->lock);
1014 * preempt at: now + delay
1017 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1019 * indicate we need to program the timer
1021 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1023 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1026 * New slices are called from the schedule path and don't need a
1027 * forced reschedule.
1030 resched_hrt(rq->curr);
1033 static void hrtick_clear(struct rq *rq)
1035 if (hrtimer_active(&rq->hrtick_timer))
1036 hrtimer_cancel(&rq->hrtick_timer);
1040 * Update the timer from the possible pending state.
1042 static void hrtick_set(struct rq *rq)
1046 unsigned long flags;
1048 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1050 spin_lock_irqsave(&rq->lock, flags);
1051 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1052 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1053 time = rq->hrtick_expire;
1054 clear_thread_flag(TIF_HRTICK_RESCHED);
1055 spin_unlock_irqrestore(&rq->lock, flags);
1058 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1059 if (reset && !hrtimer_active(&rq->hrtick_timer))
1066 * High-resolution timer tick.
1067 * Runs from hardirq context with interrupts disabled.
1069 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1071 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1073 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1075 spin_lock(&rq->lock);
1076 update_rq_clock(rq);
1077 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1078 spin_unlock(&rq->lock);
1080 return HRTIMER_NORESTART;
1084 static void hotplug_hrtick_disable(int cpu)
1086 struct rq *rq = cpu_rq(cpu);
1087 unsigned long flags;
1089 spin_lock_irqsave(&rq->lock, flags);
1090 rq->hrtick_flags = 0;
1091 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1092 spin_unlock_irqrestore(&rq->lock, flags);
1097 static void hotplug_hrtick_enable(int cpu)
1099 struct rq *rq = cpu_rq(cpu);
1100 unsigned long flags;
1102 spin_lock_irqsave(&rq->lock, flags);
1103 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1104 spin_unlock_irqrestore(&rq->lock, flags);
1108 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1110 int cpu = (int)(long)hcpu;
1113 case CPU_UP_CANCELED:
1114 case CPU_UP_CANCELED_FROZEN:
1115 case CPU_DOWN_PREPARE:
1116 case CPU_DOWN_PREPARE_FROZEN:
1118 case CPU_DEAD_FROZEN:
1119 hotplug_hrtick_disable(cpu);
1122 case CPU_UP_PREPARE:
1123 case CPU_UP_PREPARE_FROZEN:
1124 case CPU_DOWN_FAILED:
1125 case CPU_DOWN_FAILED_FROZEN:
1127 case CPU_ONLINE_FROZEN:
1128 hotplug_hrtick_enable(cpu);
1135 static void init_hrtick(void)
1137 hotcpu_notifier(hotplug_hrtick, 0);
1139 #endif /* CONFIG_SMP */
1141 static void init_rq_hrtick(struct rq *rq)
1143 rq->hrtick_flags = 0;
1144 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1145 rq->hrtick_timer.function = hrtick;
1146 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1149 void hrtick_resched(void)
1152 unsigned long flags;
1154 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1157 local_irq_save(flags);
1158 rq = cpu_rq(smp_processor_id());
1160 local_irq_restore(flags);
1163 static inline void hrtick_clear(struct rq *rq)
1167 static inline void hrtick_set(struct rq *rq)
1171 static inline void init_rq_hrtick(struct rq *rq)
1175 void hrtick_resched(void)
1179 static inline void init_hrtick(void)
1185 * resched_task - mark a task 'to be rescheduled now'.
1187 * On UP this means the setting of the need_resched flag, on SMP it
1188 * might also involve a cross-CPU call to trigger the scheduler on
1193 #ifndef tsk_is_polling
1194 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1197 static void __resched_task(struct task_struct *p, int tif_bit)
1201 assert_spin_locked(&task_rq(p)->lock);
1203 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1206 set_tsk_thread_flag(p, tif_bit);
1209 if (cpu == smp_processor_id())
1212 /* NEED_RESCHED must be visible before we test polling */
1214 if (!tsk_is_polling(p))
1215 smp_send_reschedule(cpu);
1218 static void resched_cpu(int cpu)
1220 struct rq *rq = cpu_rq(cpu);
1221 unsigned long flags;
1223 if (!spin_trylock_irqsave(&rq->lock, flags))
1225 resched_task(cpu_curr(cpu));
1226 spin_unlock_irqrestore(&rq->lock, flags);
1231 * When add_timer_on() enqueues a timer into the timer wheel of an
1232 * idle CPU then this timer might expire before the next timer event
1233 * which is scheduled to wake up that CPU. In case of a completely
1234 * idle system the next event might even be infinite time into the
1235 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1236 * leaves the inner idle loop so the newly added timer is taken into
1237 * account when the CPU goes back to idle and evaluates the timer
1238 * wheel for the next timer event.
1240 void wake_up_idle_cpu(int cpu)
1242 struct rq *rq = cpu_rq(cpu);
1244 if (cpu == smp_processor_id())
1248 * This is safe, as this function is called with the timer
1249 * wheel base lock of (cpu) held. When the CPU is on the way
1250 * to idle and has not yet set rq->curr to idle then it will
1251 * be serialized on the timer wheel base lock and take the new
1252 * timer into account automatically.
1254 if (rq->curr != rq->idle)
1258 * We can set TIF_RESCHED on the idle task of the other CPU
1259 * lockless. The worst case is that the other CPU runs the
1260 * idle task through an additional NOOP schedule()
1262 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1264 /* NEED_RESCHED must be visible before we test polling */
1266 if (!tsk_is_polling(rq->idle))
1267 smp_send_reschedule(cpu);
1269 #endif /* CONFIG_NO_HZ */
1271 #else /* !CONFIG_SMP */
1272 static void __resched_task(struct task_struct *p, int tif_bit)
1274 assert_spin_locked(&task_rq(p)->lock);
1275 set_tsk_thread_flag(p, tif_bit);
1277 #endif /* CONFIG_SMP */
1279 #if BITS_PER_LONG == 32
1280 # define WMULT_CONST (~0UL)
1282 # define WMULT_CONST (1UL << 32)
1285 #define WMULT_SHIFT 32
1288 * Shift right and round:
1290 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1293 * delta *= weight / lw
1295 static unsigned long
1296 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1297 struct load_weight *lw)
1301 if (!lw->inv_weight) {
1302 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1305 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1309 tmp = (u64)delta_exec * weight;
1311 * Check whether we'd overflow the 64-bit multiplication:
1313 if (unlikely(tmp > WMULT_CONST))
1314 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1317 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1319 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1322 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1328 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1335 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1336 * of tasks with abnormal "nice" values across CPUs the contribution that
1337 * each task makes to its run queue's load is weighted according to its
1338 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1339 * scaled version of the new time slice allocation that they receive on time
1343 #define WEIGHT_IDLEPRIO 2
1344 #define WMULT_IDLEPRIO (1 << 31)
1347 * Nice levels are multiplicative, with a gentle 10% change for every
1348 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1349 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1350 * that remained on nice 0.
1352 * The "10% effect" is relative and cumulative: from _any_ nice level,
1353 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1354 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1355 * If a task goes up by ~10% and another task goes down by ~10% then
1356 * the relative distance between them is ~25%.)
1358 static const int prio_to_weight[40] = {
1359 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1360 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1361 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1362 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1363 /* 0 */ 1024, 820, 655, 526, 423,
1364 /* 5 */ 335, 272, 215, 172, 137,
1365 /* 10 */ 110, 87, 70, 56, 45,
1366 /* 15 */ 36, 29, 23, 18, 15,
1370 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1372 * In cases where the weight does not change often, we can use the
1373 * precalculated inverse to speed up arithmetics by turning divisions
1374 * into multiplications:
1376 static const u32 prio_to_wmult[40] = {
1377 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1378 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1379 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1380 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1381 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1382 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1383 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1384 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1387 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1390 * runqueue iterator, to support SMP load-balancing between different
1391 * scheduling classes, without having to expose their internal data
1392 * structures to the load-balancing proper:
1394 struct rq_iterator {
1396 struct task_struct *(*start)(void *);
1397 struct task_struct *(*next)(void *);
1401 static unsigned long
1402 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1403 unsigned long max_load_move, struct sched_domain *sd,
1404 enum cpu_idle_type idle, int *all_pinned,
1405 int *this_best_prio, struct rq_iterator *iterator);
1408 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1409 struct sched_domain *sd, enum cpu_idle_type idle,
1410 struct rq_iterator *iterator);
1413 #ifdef CONFIG_CGROUP_CPUACCT
1414 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1416 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1419 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1421 update_load_add(&rq->load, load);
1424 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1426 update_load_sub(&rq->load, load);
1430 static unsigned long source_load(int cpu, int type);
1431 static unsigned long target_load(int cpu, int type);
1432 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1434 static unsigned long cpu_avg_load_per_task(int cpu)
1436 struct rq *rq = cpu_rq(cpu);
1439 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1441 return rq->avg_load_per_task;
1444 #ifdef CONFIG_FAIR_GROUP_SCHED
1446 typedef void (*tg_visitor)(struct task_group *, int, struct sched_domain *);
1449 * Iterate the full tree, calling @down when first entering a node and @up when
1450 * leaving it for the final time.
1453 walk_tg_tree(tg_visitor down, tg_visitor up, int cpu, struct sched_domain *sd)
1455 struct task_group *parent, *child;
1458 parent = &root_task_group;
1460 (*down)(parent, cpu, sd);
1461 list_for_each_entry_rcu(child, &parent->children, siblings) {
1468 (*up)(parent, cpu, sd);
1471 parent = parent->parent;
1477 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1480 * Calculate and set the cpu's group shares.
1483 __update_group_shares_cpu(struct task_group *tg, int cpu,
1484 unsigned long sd_shares, unsigned long sd_rq_weight)
1487 unsigned long shares;
1488 unsigned long rq_weight;
1493 rq_weight = tg->cfs_rq[cpu]->load.weight;
1496 * If there are currently no tasks on the cpu pretend there is one of
1497 * average load so that when a new task gets to run here it will not
1498 * get delayed by group starvation.
1502 rq_weight = NICE_0_LOAD;
1505 if (unlikely(rq_weight > sd_rq_weight))
1506 rq_weight = sd_rq_weight;
1509 * \Sum shares * rq_weight
1510 * shares = -----------------------
1514 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1517 * record the actual number of shares, not the boosted amount.
1519 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1521 if (shares < MIN_SHARES)
1522 shares = MIN_SHARES;
1523 else if (shares > MAX_SHARES)
1524 shares = MAX_SHARES;
1526 __set_se_shares(tg->se[cpu], shares);
1530 * Re-compute the task group their per cpu shares over the given domain.
1531 * This needs to be done in a bottom-up fashion because the rq weight of a
1532 * parent group depends on the shares of its child groups.
1535 tg_shares_up(struct task_group *tg, int cpu, struct sched_domain *sd)
1537 unsigned long rq_weight = 0;
1538 unsigned long shares = 0;
1541 for_each_cpu_mask(i, sd->span) {
1542 rq_weight += tg->cfs_rq[i]->load.weight;
1543 shares += tg->cfs_rq[i]->shares;
1546 if ((!shares && rq_weight) || shares > tg->shares)
1547 shares = tg->shares;
1549 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1550 shares = tg->shares;
1553 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1555 for_each_cpu_mask(i, sd->span) {
1556 struct rq *rq = cpu_rq(i);
1557 unsigned long flags;
1559 spin_lock_irqsave(&rq->lock, flags);
1560 __update_group_shares_cpu(tg, i, shares, rq_weight);
1561 spin_unlock_irqrestore(&rq->lock, flags);
1566 * Compute the cpu's hierarchical load factor for each task group.
1567 * This needs to be done in a top-down fashion because the load of a child
1568 * group is a fraction of its parents load.
1571 tg_load_down(struct task_group *tg, int cpu, struct sched_domain *sd)
1576 load = cpu_rq(cpu)->load.weight;
1578 load = tg->parent->cfs_rq[cpu]->h_load;
1579 load *= tg->cfs_rq[cpu]->shares;
1580 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1583 tg->cfs_rq[cpu]->h_load = load;
1587 tg_nop(struct task_group *tg, int cpu, struct sched_domain *sd)
1591 static void update_shares(struct sched_domain *sd)
1593 walk_tg_tree(tg_nop, tg_shares_up, 0, sd);
1596 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1598 spin_unlock(&rq->lock);
1600 spin_lock(&rq->lock);
1603 static void update_h_load(int cpu)
1605 walk_tg_tree(tg_load_down, tg_nop, cpu, NULL);
1608 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1610 cfs_rq->shares = shares;
1615 static inline void update_shares(struct sched_domain *sd)
1619 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1627 #include "sched_stats.h"
1628 #include "sched_idletask.c"
1629 #include "sched_fair.c"
1630 #include "sched_rt.c"
1631 #ifdef CONFIG_SCHED_DEBUG
1632 # include "sched_debug.c"
1635 #define sched_class_highest (&rt_sched_class)
1636 #define for_each_class(class) \
1637 for (class = sched_class_highest; class; class = class->next)
1639 static void inc_nr_running(struct rq *rq)
1644 static void dec_nr_running(struct rq *rq)
1649 static void set_load_weight(struct task_struct *p)
1651 if (task_has_rt_policy(p)) {
1652 p->se.load.weight = prio_to_weight[0] * 2;
1653 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1658 * SCHED_IDLE tasks get minimal weight:
1660 if (p->policy == SCHED_IDLE) {
1661 p->se.load.weight = WEIGHT_IDLEPRIO;
1662 p->se.load.inv_weight = WMULT_IDLEPRIO;
1666 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1667 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1670 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1672 sched_info_queued(p);
1673 p->sched_class->enqueue_task(rq, p, wakeup);
1677 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1679 p->sched_class->dequeue_task(rq, p, sleep);
1684 * __normal_prio - return the priority that is based on the static prio
1686 static inline int __normal_prio(struct task_struct *p)
1688 return p->static_prio;
1692 * Calculate the expected normal priority: i.e. priority
1693 * without taking RT-inheritance into account. Might be
1694 * boosted by interactivity modifiers. Changes upon fork,
1695 * setprio syscalls, and whenever the interactivity
1696 * estimator recalculates.
1698 static inline int normal_prio(struct task_struct *p)
1702 if (task_has_rt_policy(p))
1703 prio = MAX_RT_PRIO-1 - p->rt_priority;
1705 prio = __normal_prio(p);
1710 * Calculate the current priority, i.e. the priority
1711 * taken into account by the scheduler. This value might
1712 * be boosted by RT tasks, or might be boosted by
1713 * interactivity modifiers. Will be RT if the task got
1714 * RT-boosted. If not then it returns p->normal_prio.
1716 static int effective_prio(struct task_struct *p)
1718 p->normal_prio = normal_prio(p);
1720 * If we are RT tasks or we were boosted to RT priority,
1721 * keep the priority unchanged. Otherwise, update priority
1722 * to the normal priority:
1724 if (!rt_prio(p->prio))
1725 return p->normal_prio;
1730 * activate_task - move a task to the runqueue.
1732 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1734 if (task_contributes_to_load(p))
1735 rq->nr_uninterruptible--;
1737 enqueue_task(rq, p, wakeup);
1742 * deactivate_task - remove a task from the runqueue.
1744 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1746 if (task_contributes_to_load(p))
1747 rq->nr_uninterruptible++;
1749 dequeue_task(rq, p, sleep);
1754 * task_curr - is this task currently executing on a CPU?
1755 * @p: the task in question.
1757 inline int task_curr(const struct task_struct *p)
1759 return cpu_curr(task_cpu(p)) == p;
1762 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1764 set_task_rq(p, cpu);
1767 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1768 * successfuly executed on another CPU. We must ensure that updates of
1769 * per-task data have been completed by this moment.
1772 task_thread_info(p)->cpu = cpu;
1776 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1777 const struct sched_class *prev_class,
1778 int oldprio, int running)
1780 if (prev_class != p->sched_class) {
1781 if (prev_class->switched_from)
1782 prev_class->switched_from(rq, p, running);
1783 p->sched_class->switched_to(rq, p, running);
1785 p->sched_class->prio_changed(rq, p, oldprio, running);
1790 /* Used instead of source_load when we know the type == 0 */
1791 static unsigned long weighted_cpuload(const int cpu)
1793 return cpu_rq(cpu)->load.weight;
1797 * Is this task likely cache-hot:
1800 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1805 * Buddy candidates are cache hot:
1807 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1810 if (p->sched_class != &fair_sched_class)
1813 if (sysctl_sched_migration_cost == -1)
1815 if (sysctl_sched_migration_cost == 0)
1818 delta = now - p->se.exec_start;
1820 return delta < (s64)sysctl_sched_migration_cost;
1824 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1826 int old_cpu = task_cpu(p);
1827 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1828 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1829 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1832 clock_offset = old_rq->clock - new_rq->clock;
1834 #ifdef CONFIG_SCHEDSTATS
1835 if (p->se.wait_start)
1836 p->se.wait_start -= clock_offset;
1837 if (p->se.sleep_start)
1838 p->se.sleep_start -= clock_offset;
1839 if (p->se.block_start)
1840 p->se.block_start -= clock_offset;
1841 if (old_cpu != new_cpu) {
1842 schedstat_inc(p, se.nr_migrations);
1843 if (task_hot(p, old_rq->clock, NULL))
1844 schedstat_inc(p, se.nr_forced2_migrations);
1847 p->se.vruntime -= old_cfsrq->min_vruntime -
1848 new_cfsrq->min_vruntime;
1850 __set_task_cpu(p, new_cpu);
1853 struct migration_req {
1854 struct list_head list;
1856 struct task_struct *task;
1859 struct completion done;
1863 * The task's runqueue lock must be held.
1864 * Returns true if you have to wait for migration thread.
1867 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1869 struct rq *rq = task_rq(p);
1872 * If the task is not on a runqueue (and not running), then
1873 * it is sufficient to simply update the task's cpu field.
1875 if (!p->se.on_rq && !task_running(rq, p)) {
1876 set_task_cpu(p, dest_cpu);
1880 init_completion(&req->done);
1882 req->dest_cpu = dest_cpu;
1883 list_add(&req->list, &rq->migration_queue);
1889 * wait_task_inactive - wait for a thread to unschedule.
1891 * The caller must ensure that the task *will* unschedule sometime soon,
1892 * else this function might spin for a *long* time. This function can't
1893 * be called with interrupts off, or it may introduce deadlock with
1894 * smp_call_function() if an IPI is sent by the same process we are
1895 * waiting to become inactive.
1897 void wait_task_inactive(struct task_struct *p)
1899 unsigned long flags;
1905 * We do the initial early heuristics without holding
1906 * any task-queue locks at all. We'll only try to get
1907 * the runqueue lock when things look like they will
1913 * If the task is actively running on another CPU
1914 * still, just relax and busy-wait without holding
1917 * NOTE! Since we don't hold any locks, it's not
1918 * even sure that "rq" stays as the right runqueue!
1919 * But we don't care, since "task_running()" will
1920 * return false if the runqueue has changed and p
1921 * is actually now running somewhere else!
1923 while (task_running(rq, p))
1927 * Ok, time to look more closely! We need the rq
1928 * lock now, to be *sure*. If we're wrong, we'll
1929 * just go back and repeat.
1931 rq = task_rq_lock(p, &flags);
1932 running = task_running(rq, p);
1933 on_rq = p->se.on_rq;
1934 task_rq_unlock(rq, &flags);
1937 * Was it really running after all now that we
1938 * checked with the proper locks actually held?
1940 * Oops. Go back and try again..
1942 if (unlikely(running)) {
1948 * It's not enough that it's not actively running,
1949 * it must be off the runqueue _entirely_, and not
1952 * So if it wa still runnable (but just not actively
1953 * running right now), it's preempted, and we should
1954 * yield - it could be a while.
1956 if (unlikely(on_rq)) {
1957 schedule_timeout_uninterruptible(1);
1962 * Ahh, all good. It wasn't running, and it wasn't
1963 * runnable, which means that it will never become
1964 * running in the future either. We're all done!
1971 * kick_process - kick a running thread to enter/exit the kernel
1972 * @p: the to-be-kicked thread
1974 * Cause a process which is running on another CPU to enter
1975 * kernel-mode, without any delay. (to get signals handled.)
1977 * NOTE: this function doesnt have to take the runqueue lock,
1978 * because all it wants to ensure is that the remote task enters
1979 * the kernel. If the IPI races and the task has been migrated
1980 * to another CPU then no harm is done and the purpose has been
1983 void kick_process(struct task_struct *p)
1989 if ((cpu != smp_processor_id()) && task_curr(p))
1990 smp_send_reschedule(cpu);
1995 * Return a low guess at the load of a migration-source cpu weighted
1996 * according to the scheduling class and "nice" value.
1998 * We want to under-estimate the load of migration sources, to
1999 * balance conservatively.
2001 static unsigned long source_load(int cpu, int type)
2003 struct rq *rq = cpu_rq(cpu);
2004 unsigned long total = weighted_cpuload(cpu);
2006 if (type == 0 || !sched_feat(LB_BIAS))
2009 return min(rq->cpu_load[type-1], total);
2013 * Return a high guess at the load of a migration-target cpu weighted
2014 * according to the scheduling class and "nice" value.
2016 static unsigned long target_load(int cpu, int type)
2018 struct rq *rq = cpu_rq(cpu);
2019 unsigned long total = weighted_cpuload(cpu);
2021 if (type == 0 || !sched_feat(LB_BIAS))
2024 return max(rq->cpu_load[type-1], total);
2028 * find_idlest_group finds and returns the least busy CPU group within the
2031 static struct sched_group *
2032 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2034 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2035 unsigned long min_load = ULONG_MAX, this_load = 0;
2036 int load_idx = sd->forkexec_idx;
2037 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2040 unsigned long load, avg_load;
2044 /* Skip over this group if it has no CPUs allowed */
2045 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2048 local_group = cpu_isset(this_cpu, group->cpumask);
2050 /* Tally up the load of all CPUs in the group */
2053 for_each_cpu_mask(i, group->cpumask) {
2054 /* Bias balancing toward cpus of our domain */
2056 load = source_load(i, load_idx);
2058 load = target_load(i, load_idx);
2063 /* Adjust by relative CPU power of the group */
2064 avg_load = sg_div_cpu_power(group,
2065 avg_load * SCHED_LOAD_SCALE);
2068 this_load = avg_load;
2070 } else if (avg_load < min_load) {
2071 min_load = avg_load;
2074 } while (group = group->next, group != sd->groups);
2076 if (!idlest || 100*this_load < imbalance*min_load)
2082 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2085 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2088 unsigned long load, min_load = ULONG_MAX;
2092 /* Traverse only the allowed CPUs */
2093 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2095 for_each_cpu_mask(i, *tmp) {
2096 load = weighted_cpuload(i);
2098 if (load < min_load || (load == min_load && i == this_cpu)) {
2108 * sched_balance_self: balance the current task (running on cpu) in domains
2109 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2112 * Balance, ie. select the least loaded group.
2114 * Returns the target CPU number, or the same CPU if no balancing is needed.
2116 * preempt must be disabled.
2118 static int sched_balance_self(int cpu, int flag)
2120 struct task_struct *t = current;
2121 struct sched_domain *tmp, *sd = NULL;
2123 for_each_domain(cpu, tmp) {
2125 * If power savings logic is enabled for a domain, stop there.
2127 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2129 if (tmp->flags & flag)
2137 cpumask_t span, tmpmask;
2138 struct sched_group *group;
2139 int new_cpu, weight;
2141 if (!(sd->flags & flag)) {
2147 group = find_idlest_group(sd, t, cpu);
2153 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2154 if (new_cpu == -1 || new_cpu == cpu) {
2155 /* Now try balancing at a lower domain level of cpu */
2160 /* Now try balancing at a lower domain level of new_cpu */
2163 weight = cpus_weight(span);
2164 for_each_domain(cpu, tmp) {
2165 if (weight <= cpus_weight(tmp->span))
2167 if (tmp->flags & flag)
2170 /* while loop will break here if sd == NULL */
2176 #endif /* CONFIG_SMP */
2179 * try_to_wake_up - wake up a thread
2180 * @p: the to-be-woken-up thread
2181 * @state: the mask of task states that can be woken
2182 * @sync: do a synchronous wakeup?
2184 * Put it on the run-queue if it's not already there. The "current"
2185 * thread is always on the run-queue (except when the actual
2186 * re-schedule is in progress), and as such you're allowed to do
2187 * the simpler "current->state = TASK_RUNNING" to mark yourself
2188 * runnable without the overhead of this.
2190 * returns failure only if the task is already active.
2192 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2194 int cpu, orig_cpu, this_cpu, success = 0;
2195 unsigned long flags;
2199 if (!sched_feat(SYNC_WAKEUPS))
2203 rq = task_rq_lock(p, &flags);
2204 old_state = p->state;
2205 if (!(old_state & state))
2213 this_cpu = smp_processor_id();
2216 if (unlikely(task_running(rq, p)))
2219 cpu = p->sched_class->select_task_rq(p, sync);
2220 if (cpu != orig_cpu) {
2221 set_task_cpu(p, cpu);
2222 task_rq_unlock(rq, &flags);
2223 /* might preempt at this point */
2224 rq = task_rq_lock(p, &flags);
2225 old_state = p->state;
2226 if (!(old_state & state))
2231 this_cpu = smp_processor_id();
2235 #ifdef CONFIG_SCHEDSTATS
2236 schedstat_inc(rq, ttwu_count);
2237 if (cpu == this_cpu)
2238 schedstat_inc(rq, ttwu_local);
2240 struct sched_domain *sd;
2241 for_each_domain(this_cpu, sd) {
2242 if (cpu_isset(cpu, sd->span)) {
2243 schedstat_inc(sd, ttwu_wake_remote);
2248 #endif /* CONFIG_SCHEDSTATS */
2251 #endif /* CONFIG_SMP */
2252 schedstat_inc(p, se.nr_wakeups);
2254 schedstat_inc(p, se.nr_wakeups_sync);
2255 if (orig_cpu != cpu)
2256 schedstat_inc(p, se.nr_wakeups_migrate);
2257 if (cpu == this_cpu)
2258 schedstat_inc(p, se.nr_wakeups_local);
2260 schedstat_inc(p, se.nr_wakeups_remote);
2261 update_rq_clock(rq);
2262 activate_task(rq, p, 1);
2266 check_preempt_curr(rq, p);
2268 p->state = TASK_RUNNING;
2270 if (p->sched_class->task_wake_up)
2271 p->sched_class->task_wake_up(rq, p);
2274 task_rq_unlock(rq, &flags);
2279 int wake_up_process(struct task_struct *p)
2281 return try_to_wake_up(p, TASK_ALL, 0);
2283 EXPORT_SYMBOL(wake_up_process);
2285 int wake_up_state(struct task_struct *p, unsigned int state)
2287 return try_to_wake_up(p, state, 0);
2291 * Perform scheduler related setup for a newly forked process p.
2292 * p is forked by current.
2294 * __sched_fork() is basic setup used by init_idle() too:
2296 static void __sched_fork(struct task_struct *p)
2298 p->se.exec_start = 0;
2299 p->se.sum_exec_runtime = 0;
2300 p->se.prev_sum_exec_runtime = 0;
2301 p->se.last_wakeup = 0;
2302 p->se.avg_overlap = 0;
2304 #ifdef CONFIG_SCHEDSTATS
2305 p->se.wait_start = 0;
2306 p->se.sum_sleep_runtime = 0;
2307 p->se.sleep_start = 0;
2308 p->se.block_start = 0;
2309 p->se.sleep_max = 0;
2310 p->se.block_max = 0;
2312 p->se.slice_max = 0;
2316 INIT_LIST_HEAD(&p->rt.run_list);
2318 INIT_LIST_HEAD(&p->se.group_node);
2320 #ifdef CONFIG_PREEMPT_NOTIFIERS
2321 INIT_HLIST_HEAD(&p->preempt_notifiers);
2325 * We mark the process as running here, but have not actually
2326 * inserted it onto the runqueue yet. This guarantees that
2327 * nobody will actually run it, and a signal or other external
2328 * event cannot wake it up and insert it on the runqueue either.
2330 p->state = TASK_RUNNING;
2334 * fork()/clone()-time setup:
2336 void sched_fork(struct task_struct *p, int clone_flags)
2338 int cpu = get_cpu();
2343 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2345 set_task_cpu(p, cpu);
2348 * Make sure we do not leak PI boosting priority to the child:
2350 p->prio = current->normal_prio;
2351 if (!rt_prio(p->prio))
2352 p->sched_class = &fair_sched_class;
2354 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2355 if (likely(sched_info_on()))
2356 memset(&p->sched_info, 0, sizeof(p->sched_info));
2358 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2361 #ifdef CONFIG_PREEMPT
2362 /* Want to start with kernel preemption disabled. */
2363 task_thread_info(p)->preempt_count = 1;
2369 * wake_up_new_task - wake up a newly created task for the first time.
2371 * This function will do some initial scheduler statistics housekeeping
2372 * that must be done for every newly created context, then puts the task
2373 * on the runqueue and wakes it.
2375 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2377 unsigned long flags;
2380 rq = task_rq_lock(p, &flags);
2381 BUG_ON(p->state != TASK_RUNNING);
2382 update_rq_clock(rq);
2384 p->prio = effective_prio(p);
2386 if (!p->sched_class->task_new || !current->se.on_rq) {
2387 activate_task(rq, p, 0);
2390 * Let the scheduling class do new task startup
2391 * management (if any):
2393 p->sched_class->task_new(rq, p);
2396 check_preempt_curr(rq, p);
2398 if (p->sched_class->task_wake_up)
2399 p->sched_class->task_wake_up(rq, p);
2401 task_rq_unlock(rq, &flags);
2404 #ifdef CONFIG_PREEMPT_NOTIFIERS
2407 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2408 * @notifier: notifier struct to register
2410 void preempt_notifier_register(struct preempt_notifier *notifier)
2412 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2414 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2417 * preempt_notifier_unregister - no longer interested in preemption notifications
2418 * @notifier: notifier struct to unregister
2420 * This is safe to call from within a preemption notifier.
2422 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2424 hlist_del(¬ifier->link);
2426 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2428 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2430 struct preempt_notifier *notifier;
2431 struct hlist_node *node;
2433 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2434 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2438 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2439 struct task_struct *next)
2441 struct preempt_notifier *notifier;
2442 struct hlist_node *node;
2444 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2445 notifier->ops->sched_out(notifier, next);
2448 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2450 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2455 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2456 struct task_struct *next)
2460 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2463 * prepare_task_switch - prepare to switch tasks
2464 * @rq: the runqueue preparing to switch
2465 * @prev: the current task that is being switched out
2466 * @next: the task we are going to switch to.
2468 * This is called with the rq lock held and interrupts off. It must
2469 * be paired with a subsequent finish_task_switch after the context
2472 * prepare_task_switch sets up locking and calls architecture specific
2476 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2477 struct task_struct *next)
2479 fire_sched_out_preempt_notifiers(prev, next);
2480 prepare_lock_switch(rq, next);
2481 prepare_arch_switch(next);
2485 * finish_task_switch - clean up after a task-switch
2486 * @rq: runqueue associated with task-switch
2487 * @prev: the thread we just switched away from.
2489 * finish_task_switch must be called after the context switch, paired
2490 * with a prepare_task_switch call before the context switch.
2491 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2492 * and do any other architecture-specific cleanup actions.
2494 * Note that we may have delayed dropping an mm in context_switch(). If
2495 * so, we finish that here outside of the runqueue lock. (Doing it
2496 * with the lock held can cause deadlocks; see schedule() for
2499 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2500 __releases(rq->lock)
2502 struct mm_struct *mm = rq->prev_mm;
2508 * A task struct has one reference for the use as "current".
2509 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2510 * schedule one last time. The schedule call will never return, and
2511 * the scheduled task must drop that reference.
2512 * The test for TASK_DEAD must occur while the runqueue locks are
2513 * still held, otherwise prev could be scheduled on another cpu, die
2514 * there before we look at prev->state, and then the reference would
2516 * Manfred Spraul <manfred@colorfullife.com>
2518 prev_state = prev->state;
2519 finish_arch_switch(prev);
2520 finish_lock_switch(rq, prev);
2522 if (current->sched_class->post_schedule)
2523 current->sched_class->post_schedule(rq);
2526 fire_sched_in_preempt_notifiers(current);
2529 if (unlikely(prev_state == TASK_DEAD)) {
2531 * Remove function-return probe instances associated with this
2532 * task and put them back on the free list.
2534 kprobe_flush_task(prev);
2535 put_task_struct(prev);
2540 * schedule_tail - first thing a freshly forked thread must call.
2541 * @prev: the thread we just switched away from.
2543 asmlinkage void schedule_tail(struct task_struct *prev)
2544 __releases(rq->lock)
2546 struct rq *rq = this_rq();
2548 finish_task_switch(rq, prev);
2549 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2550 /* In this case, finish_task_switch does not reenable preemption */
2553 if (current->set_child_tid)
2554 put_user(task_pid_vnr(current), current->set_child_tid);
2558 * context_switch - switch to the new MM and the new
2559 * thread's register state.
2562 context_switch(struct rq *rq, struct task_struct *prev,
2563 struct task_struct *next)
2565 struct mm_struct *mm, *oldmm;
2567 prepare_task_switch(rq, prev, next);
2569 oldmm = prev->active_mm;
2571 * For paravirt, this is coupled with an exit in switch_to to
2572 * combine the page table reload and the switch backend into
2575 arch_enter_lazy_cpu_mode();
2577 if (unlikely(!mm)) {
2578 next->active_mm = oldmm;
2579 atomic_inc(&oldmm->mm_count);
2580 enter_lazy_tlb(oldmm, next);
2582 switch_mm(oldmm, mm, next);
2584 if (unlikely(!prev->mm)) {
2585 prev->active_mm = NULL;
2586 rq->prev_mm = oldmm;
2589 * Since the runqueue lock will be released by the next
2590 * task (which is an invalid locking op but in the case
2591 * of the scheduler it's an obvious special-case), so we
2592 * do an early lockdep release here:
2594 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2595 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2598 /* Here we just switch the register state and the stack. */
2599 switch_to(prev, next, prev);
2603 * this_rq must be evaluated again because prev may have moved
2604 * CPUs since it called schedule(), thus the 'rq' on its stack
2605 * frame will be invalid.
2607 finish_task_switch(this_rq(), prev);
2611 * nr_running, nr_uninterruptible and nr_context_switches:
2613 * externally visible scheduler statistics: current number of runnable
2614 * threads, current number of uninterruptible-sleeping threads, total
2615 * number of context switches performed since bootup.
2617 unsigned long nr_running(void)
2619 unsigned long i, sum = 0;
2621 for_each_online_cpu(i)
2622 sum += cpu_rq(i)->nr_running;
2627 unsigned long nr_uninterruptible(void)
2629 unsigned long i, sum = 0;
2631 for_each_possible_cpu(i)
2632 sum += cpu_rq(i)->nr_uninterruptible;
2635 * Since we read the counters lockless, it might be slightly
2636 * inaccurate. Do not allow it to go below zero though:
2638 if (unlikely((long)sum < 0))
2644 unsigned long long nr_context_switches(void)
2647 unsigned long long sum = 0;
2649 for_each_possible_cpu(i)
2650 sum += cpu_rq(i)->nr_switches;
2655 unsigned long nr_iowait(void)
2657 unsigned long i, sum = 0;
2659 for_each_possible_cpu(i)
2660 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2665 unsigned long nr_active(void)
2667 unsigned long i, running = 0, uninterruptible = 0;
2669 for_each_online_cpu(i) {
2670 running += cpu_rq(i)->nr_running;
2671 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2674 if (unlikely((long)uninterruptible < 0))
2675 uninterruptible = 0;
2677 return running + uninterruptible;
2681 * Update rq->cpu_load[] statistics. This function is usually called every
2682 * scheduler tick (TICK_NSEC).
2684 static void update_cpu_load(struct rq *this_rq)
2686 unsigned long this_load = this_rq->load.weight;
2689 this_rq->nr_load_updates++;
2691 /* Update our load: */
2692 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2693 unsigned long old_load, new_load;
2695 /* scale is effectively 1 << i now, and >> i divides by scale */
2697 old_load = this_rq->cpu_load[i];
2698 new_load = this_load;
2700 * Round up the averaging division if load is increasing. This
2701 * prevents us from getting stuck on 9 if the load is 10, for
2704 if (new_load > old_load)
2705 new_load += scale-1;
2706 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2713 * double_rq_lock - safely lock two runqueues
2715 * Note this does not disable interrupts like task_rq_lock,
2716 * you need to do so manually before calling.
2718 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2719 __acquires(rq1->lock)
2720 __acquires(rq2->lock)
2722 BUG_ON(!irqs_disabled());
2724 spin_lock(&rq1->lock);
2725 __acquire(rq2->lock); /* Fake it out ;) */
2728 spin_lock(&rq1->lock);
2729 spin_lock(&rq2->lock);
2731 spin_lock(&rq2->lock);
2732 spin_lock(&rq1->lock);
2735 update_rq_clock(rq1);
2736 update_rq_clock(rq2);
2740 * double_rq_unlock - safely unlock two runqueues
2742 * Note this does not restore interrupts like task_rq_unlock,
2743 * you need to do so manually after calling.
2745 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2746 __releases(rq1->lock)
2747 __releases(rq2->lock)
2749 spin_unlock(&rq1->lock);
2751 spin_unlock(&rq2->lock);
2753 __release(rq2->lock);
2757 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2759 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2760 __releases(this_rq->lock)
2761 __acquires(busiest->lock)
2762 __acquires(this_rq->lock)
2766 if (unlikely(!irqs_disabled())) {
2767 /* printk() doesn't work good under rq->lock */
2768 spin_unlock(&this_rq->lock);
2771 if (unlikely(!spin_trylock(&busiest->lock))) {
2772 if (busiest < this_rq) {
2773 spin_unlock(&this_rq->lock);
2774 spin_lock(&busiest->lock);
2775 spin_lock(&this_rq->lock);
2778 spin_lock(&busiest->lock);
2784 * If dest_cpu is allowed for this process, migrate the task to it.
2785 * This is accomplished by forcing the cpu_allowed mask to only
2786 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2787 * the cpu_allowed mask is restored.
2789 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2791 struct migration_req req;
2792 unsigned long flags;
2795 rq = task_rq_lock(p, &flags);
2796 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2797 || unlikely(cpu_is_offline(dest_cpu)))
2800 /* force the process onto the specified CPU */
2801 if (migrate_task(p, dest_cpu, &req)) {
2802 /* Need to wait for migration thread (might exit: take ref). */
2803 struct task_struct *mt = rq->migration_thread;
2805 get_task_struct(mt);
2806 task_rq_unlock(rq, &flags);
2807 wake_up_process(mt);
2808 put_task_struct(mt);
2809 wait_for_completion(&req.done);
2814 task_rq_unlock(rq, &flags);
2818 * sched_exec - execve() is a valuable balancing opportunity, because at
2819 * this point the task has the smallest effective memory and cache footprint.
2821 void sched_exec(void)
2823 int new_cpu, this_cpu = get_cpu();
2824 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2826 if (new_cpu != this_cpu)
2827 sched_migrate_task(current, new_cpu);
2831 * pull_task - move a task from a remote runqueue to the local runqueue.
2832 * Both runqueues must be locked.
2834 static void pull_task(struct rq *src_rq, struct task_struct *p,
2835 struct rq *this_rq, int this_cpu)
2837 deactivate_task(src_rq, p, 0);
2838 set_task_cpu(p, this_cpu);
2839 activate_task(this_rq, p, 0);
2841 * Note that idle threads have a prio of MAX_PRIO, for this test
2842 * to be always true for them.
2844 check_preempt_curr(this_rq, p);
2848 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2851 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2852 struct sched_domain *sd, enum cpu_idle_type idle,
2856 * We do not migrate tasks that are:
2857 * 1) running (obviously), or
2858 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2859 * 3) are cache-hot on their current CPU.
2861 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2862 schedstat_inc(p, se.nr_failed_migrations_affine);
2867 if (task_running(rq, p)) {
2868 schedstat_inc(p, se.nr_failed_migrations_running);
2873 * Aggressive migration if:
2874 * 1) task is cache cold, or
2875 * 2) too many balance attempts have failed.
2878 if (!task_hot(p, rq->clock, sd) ||
2879 sd->nr_balance_failed > sd->cache_nice_tries) {
2880 #ifdef CONFIG_SCHEDSTATS
2881 if (task_hot(p, rq->clock, sd)) {
2882 schedstat_inc(sd, lb_hot_gained[idle]);
2883 schedstat_inc(p, se.nr_forced_migrations);
2889 if (task_hot(p, rq->clock, sd)) {
2890 schedstat_inc(p, se.nr_failed_migrations_hot);
2896 static unsigned long
2897 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2898 unsigned long max_load_move, struct sched_domain *sd,
2899 enum cpu_idle_type idle, int *all_pinned,
2900 int *this_best_prio, struct rq_iterator *iterator)
2902 int loops = 0, pulled = 0, pinned = 0;
2903 struct task_struct *p;
2904 long rem_load_move = max_load_move;
2906 if (max_load_move == 0)
2912 * Start the load-balancing iterator:
2914 p = iterator->start(iterator->arg);
2916 if (!p || loops++ > sysctl_sched_nr_migrate)
2919 if ((p->se.load.weight >> 1) > rem_load_move ||
2920 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2921 p = iterator->next(iterator->arg);
2925 pull_task(busiest, p, this_rq, this_cpu);
2927 rem_load_move -= p->se.load.weight;
2930 * We only want to steal up to the prescribed amount of weighted load.
2932 if (rem_load_move > 0) {
2933 if (p->prio < *this_best_prio)
2934 *this_best_prio = p->prio;
2935 p = iterator->next(iterator->arg);
2940 * Right now, this is one of only two places pull_task() is called,
2941 * so we can safely collect pull_task() stats here rather than
2942 * inside pull_task().
2944 schedstat_add(sd, lb_gained[idle], pulled);
2947 *all_pinned = pinned;
2949 return max_load_move - rem_load_move;
2953 * move_tasks tries to move up to max_load_move weighted load from busiest to
2954 * this_rq, as part of a balancing operation within domain "sd".
2955 * Returns 1 if successful and 0 otherwise.
2957 * Called with both runqueues locked.
2959 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2960 unsigned long max_load_move,
2961 struct sched_domain *sd, enum cpu_idle_type idle,
2964 const struct sched_class *class = sched_class_highest;
2965 unsigned long total_load_moved = 0;
2966 int this_best_prio = this_rq->curr->prio;
2970 class->load_balance(this_rq, this_cpu, busiest,
2971 max_load_move - total_load_moved,
2972 sd, idle, all_pinned, &this_best_prio);
2973 class = class->next;
2974 } while (class && max_load_move > total_load_moved);
2976 return total_load_moved > 0;
2980 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
2981 struct sched_domain *sd, enum cpu_idle_type idle,
2982 struct rq_iterator *iterator)
2984 struct task_struct *p = iterator->start(iterator->arg);
2988 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2989 pull_task(busiest, p, this_rq, this_cpu);
2991 * Right now, this is only the second place pull_task()
2992 * is called, so we can safely collect pull_task()
2993 * stats here rather than inside pull_task().
2995 schedstat_inc(sd, lb_gained[idle]);
2999 p = iterator->next(iterator->arg);
3006 * move_one_task tries to move exactly one task from busiest to this_rq, as
3007 * part of active balancing operations within "domain".
3008 * Returns 1 if successful and 0 otherwise.
3010 * Called with both runqueues locked.
3012 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3013 struct sched_domain *sd, enum cpu_idle_type idle)
3015 const struct sched_class *class;
3017 for (class = sched_class_highest; class; class = class->next)
3018 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3025 * find_busiest_group finds and returns the busiest CPU group within the
3026 * domain. It calculates and returns the amount of weighted load which
3027 * should be moved to restore balance via the imbalance parameter.
3029 static struct sched_group *
3030 find_busiest_group(struct sched_domain *sd, int this_cpu,
3031 unsigned long *imbalance, enum cpu_idle_type idle,
3032 int *sd_idle, const cpumask_t *cpus, int *balance)
3034 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3035 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3036 unsigned long max_pull;
3037 unsigned long busiest_load_per_task, busiest_nr_running;
3038 unsigned long this_load_per_task, this_nr_running;
3039 int load_idx, group_imb = 0;
3040 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3041 int power_savings_balance = 1;
3042 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3043 unsigned long min_nr_running = ULONG_MAX;
3044 struct sched_group *group_min = NULL, *group_leader = NULL;
3047 max_load = this_load = total_load = total_pwr = 0;
3048 busiest_load_per_task = busiest_nr_running = 0;
3049 this_load_per_task = this_nr_running = 0;
3051 if (idle == CPU_NOT_IDLE)
3052 load_idx = sd->busy_idx;
3053 else if (idle == CPU_NEWLY_IDLE)
3054 load_idx = sd->newidle_idx;
3056 load_idx = sd->idle_idx;
3059 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3062 int __group_imb = 0;
3063 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3064 unsigned long sum_nr_running, sum_weighted_load;
3065 unsigned long sum_avg_load_per_task;
3066 unsigned long avg_load_per_task;
3068 local_group = cpu_isset(this_cpu, group->cpumask);
3071 balance_cpu = first_cpu(group->cpumask);
3073 /* Tally up the load of all CPUs in the group */
3074 sum_weighted_load = sum_nr_running = avg_load = 0;
3075 sum_avg_load_per_task = avg_load_per_task = 0;
3078 min_cpu_load = ~0UL;
3080 for_each_cpu_mask(i, group->cpumask) {
3083 if (!cpu_isset(i, *cpus))
3088 if (*sd_idle && rq->nr_running)
3091 /* Bias balancing toward cpus of our domain */
3093 if (idle_cpu(i) && !first_idle_cpu) {
3098 load = target_load(i, load_idx);
3100 load = source_load(i, load_idx);
3101 if (load > max_cpu_load)
3102 max_cpu_load = load;
3103 if (min_cpu_load > load)
3104 min_cpu_load = load;
3108 sum_nr_running += rq->nr_running;
3109 sum_weighted_load += weighted_cpuload(i);
3111 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3115 * First idle cpu or the first cpu(busiest) in this sched group
3116 * is eligible for doing load balancing at this and above
3117 * domains. In the newly idle case, we will allow all the cpu's
3118 * to do the newly idle load balance.
3120 if (idle != CPU_NEWLY_IDLE && local_group &&
3121 balance_cpu != this_cpu && balance) {
3126 total_load += avg_load;
3127 total_pwr += group->__cpu_power;
3129 /* Adjust by relative CPU power of the group */
3130 avg_load = sg_div_cpu_power(group,
3131 avg_load * SCHED_LOAD_SCALE);
3135 * Consider the group unbalanced when the imbalance is larger
3136 * than the average weight of two tasks.
3138 * APZ: with cgroup the avg task weight can vary wildly and
3139 * might not be a suitable number - should we keep a
3140 * normalized nr_running number somewhere that negates
3143 avg_load_per_task = sg_div_cpu_power(group,
3144 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3146 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3149 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3152 this_load = avg_load;
3154 this_nr_running = sum_nr_running;
3155 this_load_per_task = sum_weighted_load;
3156 } else if (avg_load > max_load &&
3157 (sum_nr_running > group_capacity || __group_imb)) {
3158 max_load = avg_load;
3160 busiest_nr_running = sum_nr_running;
3161 busiest_load_per_task = sum_weighted_load;
3162 group_imb = __group_imb;
3165 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3167 * Busy processors will not participate in power savings
3170 if (idle == CPU_NOT_IDLE ||
3171 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3175 * If the local group is idle or completely loaded
3176 * no need to do power savings balance at this domain
3178 if (local_group && (this_nr_running >= group_capacity ||
3180 power_savings_balance = 0;
3183 * If a group is already running at full capacity or idle,
3184 * don't include that group in power savings calculations
3186 if (!power_savings_balance || sum_nr_running >= group_capacity
3191 * Calculate the group which has the least non-idle load.
3192 * This is the group from where we need to pick up the load
3195 if ((sum_nr_running < min_nr_running) ||
3196 (sum_nr_running == min_nr_running &&
3197 first_cpu(group->cpumask) <
3198 first_cpu(group_min->cpumask))) {
3200 min_nr_running = sum_nr_running;
3201 min_load_per_task = sum_weighted_load /
3206 * Calculate the group which is almost near its
3207 * capacity but still has some space to pick up some load
3208 * from other group and save more power
3210 if (sum_nr_running <= group_capacity - 1) {
3211 if (sum_nr_running > leader_nr_running ||
3212 (sum_nr_running == leader_nr_running &&
3213 first_cpu(group->cpumask) >
3214 first_cpu(group_leader->cpumask))) {
3215 group_leader = group;
3216 leader_nr_running = sum_nr_running;
3221 group = group->next;
3222 } while (group != sd->groups);
3224 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3227 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3229 if (this_load >= avg_load ||
3230 100*max_load <= sd->imbalance_pct*this_load)
3233 busiest_load_per_task /= busiest_nr_running;
3235 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3238 * We're trying to get all the cpus to the average_load, so we don't
3239 * want to push ourselves above the average load, nor do we wish to
3240 * reduce the max loaded cpu below the average load, as either of these
3241 * actions would just result in more rebalancing later, and ping-pong
3242 * tasks around. Thus we look for the minimum possible imbalance.
3243 * Negative imbalances (*we* are more loaded than anyone else) will
3244 * be counted as no imbalance for these purposes -- we can't fix that
3245 * by pulling tasks to us. Be careful of negative numbers as they'll
3246 * appear as very large values with unsigned longs.
3248 if (max_load <= busiest_load_per_task)
3252 * In the presence of smp nice balancing, certain scenarios can have
3253 * max load less than avg load(as we skip the groups at or below
3254 * its cpu_power, while calculating max_load..)
3256 if (max_load < avg_load) {
3258 goto small_imbalance;
3261 /* Don't want to pull so many tasks that a group would go idle */
3262 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3264 /* How much load to actually move to equalise the imbalance */
3265 *imbalance = min(max_pull * busiest->__cpu_power,
3266 (avg_load - this_load) * this->__cpu_power)
3270 * if *imbalance is less than the average load per runnable task
3271 * there is no gaurantee that any tasks will be moved so we'll have
3272 * a think about bumping its value to force at least one task to be
3275 if (*imbalance < busiest_load_per_task) {
3276 unsigned long tmp, pwr_now, pwr_move;
3280 pwr_move = pwr_now = 0;
3282 if (this_nr_running) {
3283 this_load_per_task /= this_nr_running;
3284 if (busiest_load_per_task > this_load_per_task)
3287 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3289 if (max_load - this_load + 2*busiest_load_per_task >=
3290 busiest_load_per_task * imbn) {
3291 *imbalance = busiest_load_per_task;
3296 * OK, we don't have enough imbalance to justify moving tasks,
3297 * however we may be able to increase total CPU power used by
3301 pwr_now += busiest->__cpu_power *
3302 min(busiest_load_per_task, max_load);
3303 pwr_now += this->__cpu_power *
3304 min(this_load_per_task, this_load);
3305 pwr_now /= SCHED_LOAD_SCALE;
3307 /* Amount of load we'd subtract */
3308 tmp = sg_div_cpu_power(busiest,
3309 busiest_load_per_task * SCHED_LOAD_SCALE);
3311 pwr_move += busiest->__cpu_power *
3312 min(busiest_load_per_task, max_load - tmp);
3314 /* Amount of load we'd add */
3315 if (max_load * busiest->__cpu_power <
3316 busiest_load_per_task * SCHED_LOAD_SCALE)
3317 tmp = sg_div_cpu_power(this,
3318 max_load * busiest->__cpu_power);
3320 tmp = sg_div_cpu_power(this,
3321 busiest_load_per_task * SCHED_LOAD_SCALE);
3322 pwr_move += this->__cpu_power *
3323 min(this_load_per_task, this_load + tmp);
3324 pwr_move /= SCHED_LOAD_SCALE;
3326 /* Move if we gain throughput */
3327 if (pwr_move > pwr_now)
3328 *imbalance = busiest_load_per_task;
3334 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3335 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3338 if (this == group_leader && group_leader != group_min) {
3339 *imbalance = min_load_per_task;
3349 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3352 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3353 unsigned long imbalance, const cpumask_t *cpus)
3355 struct rq *busiest = NULL, *rq;
3356 unsigned long max_load = 0;
3359 for_each_cpu_mask(i, group->cpumask) {
3362 if (!cpu_isset(i, *cpus))
3366 wl = weighted_cpuload(i);
3368 if (rq->nr_running == 1 && wl > imbalance)
3371 if (wl > max_load) {
3381 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3382 * so long as it is large enough.
3384 #define MAX_PINNED_INTERVAL 512
3387 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3388 * tasks if there is an imbalance.
3390 static int load_balance(int this_cpu, struct rq *this_rq,
3391 struct sched_domain *sd, enum cpu_idle_type idle,
3392 int *balance, cpumask_t *cpus)
3394 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3395 struct sched_group *group;
3396 unsigned long imbalance;
3398 unsigned long flags;
3403 * When power savings policy is enabled for the parent domain, idle
3404 * sibling can pick up load irrespective of busy siblings. In this case,
3405 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3406 * portraying it as CPU_NOT_IDLE.
3408 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3409 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3412 schedstat_inc(sd, lb_count[idle]);
3416 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3423 schedstat_inc(sd, lb_nobusyg[idle]);
3427 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3429 schedstat_inc(sd, lb_nobusyq[idle]);
3433 BUG_ON(busiest == this_rq);
3435 schedstat_add(sd, lb_imbalance[idle], imbalance);
3438 if (busiest->nr_running > 1) {
3440 * Attempt to move tasks. If find_busiest_group has found
3441 * an imbalance but busiest->nr_running <= 1, the group is
3442 * still unbalanced. ld_moved simply stays zero, so it is
3443 * correctly treated as an imbalance.
3445 local_irq_save(flags);
3446 double_rq_lock(this_rq, busiest);
3447 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3448 imbalance, sd, idle, &all_pinned);
3449 double_rq_unlock(this_rq, busiest);
3450 local_irq_restore(flags);
3453 * some other cpu did the load balance for us.
3455 if (ld_moved && this_cpu != smp_processor_id())
3456 resched_cpu(this_cpu);
3458 /* All tasks on this runqueue were pinned by CPU affinity */
3459 if (unlikely(all_pinned)) {
3460 cpu_clear(cpu_of(busiest), *cpus);
3461 if (!cpus_empty(*cpus))
3468 schedstat_inc(sd, lb_failed[idle]);
3469 sd->nr_balance_failed++;
3471 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3473 spin_lock_irqsave(&busiest->lock, flags);
3475 /* don't kick the migration_thread, if the curr
3476 * task on busiest cpu can't be moved to this_cpu
3478 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3479 spin_unlock_irqrestore(&busiest->lock, flags);
3481 goto out_one_pinned;
3484 if (!busiest->active_balance) {
3485 busiest->active_balance = 1;
3486 busiest->push_cpu = this_cpu;
3489 spin_unlock_irqrestore(&busiest->lock, flags);
3491 wake_up_process(busiest->migration_thread);
3494 * We've kicked active balancing, reset the failure
3497 sd->nr_balance_failed = sd->cache_nice_tries+1;
3500 sd->nr_balance_failed = 0;
3502 if (likely(!active_balance)) {
3503 /* We were unbalanced, so reset the balancing interval */
3504 sd->balance_interval = sd->min_interval;
3507 * If we've begun active balancing, start to back off. This
3508 * case may not be covered by the all_pinned logic if there
3509 * is only 1 task on the busy runqueue (because we don't call
3512 if (sd->balance_interval < sd->max_interval)
3513 sd->balance_interval *= 2;
3516 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3517 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3523 schedstat_inc(sd, lb_balanced[idle]);
3525 sd->nr_balance_failed = 0;
3528 /* tune up the balancing interval */
3529 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3530 (sd->balance_interval < sd->max_interval))
3531 sd->balance_interval *= 2;
3533 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3534 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3545 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3546 * tasks if there is an imbalance.
3548 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3549 * this_rq is locked.
3552 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3555 struct sched_group *group;
3556 struct rq *busiest = NULL;
3557 unsigned long imbalance;
3565 * When power savings policy is enabled for the parent domain, idle
3566 * sibling can pick up load irrespective of busy siblings. In this case,
3567 * let the state of idle sibling percolate up as IDLE, instead of
3568 * portraying it as CPU_NOT_IDLE.
3570 if (sd->flags & SD_SHARE_CPUPOWER &&
3571 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3574 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3576 update_shares_locked(this_rq, sd);
3577 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3578 &sd_idle, cpus, NULL);
3580 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3584 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3586 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3590 BUG_ON(busiest == this_rq);
3592 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3595 if (busiest->nr_running > 1) {
3596 /* Attempt to move tasks */
3597 double_lock_balance(this_rq, busiest);
3598 /* this_rq->clock is already updated */
3599 update_rq_clock(busiest);
3600 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3601 imbalance, sd, CPU_NEWLY_IDLE,
3603 spin_unlock(&busiest->lock);
3605 if (unlikely(all_pinned)) {
3606 cpu_clear(cpu_of(busiest), *cpus);
3607 if (!cpus_empty(*cpus))
3613 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3614 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3615 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3618 sd->nr_balance_failed = 0;
3620 update_shares_locked(this_rq, sd);
3624 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3625 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3626 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3628 sd->nr_balance_failed = 0;
3634 * idle_balance is called by schedule() if this_cpu is about to become
3635 * idle. Attempts to pull tasks from other CPUs.
3637 static void idle_balance(int this_cpu, struct rq *this_rq)
3639 struct sched_domain *sd;
3640 int pulled_task = -1;
3641 unsigned long next_balance = jiffies + HZ;
3644 for_each_domain(this_cpu, sd) {
3645 unsigned long interval;
3647 if (!(sd->flags & SD_LOAD_BALANCE))
3650 if (sd->flags & SD_BALANCE_NEWIDLE)
3651 /* If we've pulled tasks over stop searching: */
3652 pulled_task = load_balance_newidle(this_cpu, this_rq,
3655 interval = msecs_to_jiffies(sd->balance_interval);
3656 if (time_after(next_balance, sd->last_balance + interval))
3657 next_balance = sd->last_balance + interval;
3661 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3663 * We are going idle. next_balance may be set based on
3664 * a busy processor. So reset next_balance.
3666 this_rq->next_balance = next_balance;
3671 * active_load_balance is run by migration threads. It pushes running tasks
3672 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3673 * running on each physical CPU where possible, and avoids physical /
3674 * logical imbalances.
3676 * Called with busiest_rq locked.
3678 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3680 int target_cpu = busiest_rq->push_cpu;
3681 struct sched_domain *sd;
3682 struct rq *target_rq;
3684 /* Is there any task to move? */
3685 if (busiest_rq->nr_running <= 1)
3688 target_rq = cpu_rq(target_cpu);
3691 * This condition is "impossible", if it occurs
3692 * we need to fix it. Originally reported by
3693 * Bjorn Helgaas on a 128-cpu setup.
3695 BUG_ON(busiest_rq == target_rq);
3697 /* move a task from busiest_rq to target_rq */
3698 double_lock_balance(busiest_rq, target_rq);
3699 update_rq_clock(busiest_rq);
3700 update_rq_clock(target_rq);
3702 /* Search for an sd spanning us and the target CPU. */
3703 for_each_domain(target_cpu, sd) {
3704 if ((sd->flags & SD_LOAD_BALANCE) &&
3705 cpu_isset(busiest_cpu, sd->span))
3710 schedstat_inc(sd, alb_count);
3712 if (move_one_task(target_rq, target_cpu, busiest_rq,
3714 schedstat_inc(sd, alb_pushed);
3716 schedstat_inc(sd, alb_failed);
3718 spin_unlock(&target_rq->lock);
3723 atomic_t load_balancer;
3725 } nohz ____cacheline_aligned = {
3726 .load_balancer = ATOMIC_INIT(-1),
3727 .cpu_mask = CPU_MASK_NONE,
3731 * This routine will try to nominate the ilb (idle load balancing)
3732 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3733 * load balancing on behalf of all those cpus. If all the cpus in the system
3734 * go into this tickless mode, then there will be no ilb owner (as there is
3735 * no need for one) and all the cpus will sleep till the next wakeup event
3738 * For the ilb owner, tick is not stopped. And this tick will be used
3739 * for idle load balancing. ilb owner will still be part of
3742 * While stopping the tick, this cpu will become the ilb owner if there
3743 * is no other owner. And will be the owner till that cpu becomes busy
3744 * or if all cpus in the system stop their ticks at which point
3745 * there is no need for ilb owner.
3747 * When the ilb owner becomes busy, it nominates another owner, during the
3748 * next busy scheduler_tick()
3750 int select_nohz_load_balancer(int stop_tick)
3752 int cpu = smp_processor_id();
3755 cpu_set(cpu, nohz.cpu_mask);
3756 cpu_rq(cpu)->in_nohz_recently = 1;
3759 * If we are going offline and still the leader, give up!
3761 if (cpu_is_offline(cpu) &&
3762 atomic_read(&nohz.load_balancer) == cpu) {
3763 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3768 /* time for ilb owner also to sleep */
3769 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3770 if (atomic_read(&nohz.load_balancer) == cpu)
3771 atomic_set(&nohz.load_balancer, -1);
3775 if (atomic_read(&nohz.load_balancer) == -1) {
3776 /* make me the ilb owner */
3777 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3779 } else if (atomic_read(&nohz.load_balancer) == cpu)
3782 if (!cpu_isset(cpu, nohz.cpu_mask))
3785 cpu_clear(cpu, nohz.cpu_mask);
3787 if (atomic_read(&nohz.load_balancer) == cpu)
3788 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3795 static DEFINE_SPINLOCK(balancing);
3798 * It checks each scheduling domain to see if it is due to be balanced,
3799 * and initiates a balancing operation if so.
3801 * Balancing parameters are set up in arch_init_sched_domains.
3803 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3806 struct rq *rq = cpu_rq(cpu);
3807 unsigned long interval;
3808 struct sched_domain *sd;
3809 /* Earliest time when we have to do rebalance again */
3810 unsigned long next_balance = jiffies + 60*HZ;
3811 int update_next_balance = 0;
3815 for_each_domain(cpu, sd) {
3816 if (!(sd->flags & SD_LOAD_BALANCE))
3819 interval = sd->balance_interval;
3820 if (idle != CPU_IDLE)
3821 interval *= sd->busy_factor;
3823 /* scale ms to jiffies */
3824 interval = msecs_to_jiffies(interval);
3825 if (unlikely(!interval))
3827 if (interval > HZ*NR_CPUS/10)
3828 interval = HZ*NR_CPUS/10;
3830 need_serialize = sd->flags & SD_SERIALIZE;
3832 if (need_serialize) {
3833 if (!spin_trylock(&balancing))
3837 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3838 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3840 * We've pulled tasks over so either we're no
3841 * longer idle, or one of our SMT siblings is
3844 idle = CPU_NOT_IDLE;
3846 sd->last_balance = jiffies;
3849 spin_unlock(&balancing);
3851 if (time_after(next_balance, sd->last_balance + interval)) {
3852 next_balance = sd->last_balance + interval;
3853 update_next_balance = 1;
3857 * Stop the load balance at this level. There is another
3858 * CPU in our sched group which is doing load balancing more
3866 * next_balance will be updated only when there is a need.
3867 * When the cpu is attached to null domain for ex, it will not be
3870 if (likely(update_next_balance))
3871 rq->next_balance = next_balance;
3875 * run_rebalance_domains is triggered when needed from the scheduler tick.
3876 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3877 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3879 static void run_rebalance_domains(struct softirq_action *h)
3881 int this_cpu = smp_processor_id();
3882 struct rq *this_rq = cpu_rq(this_cpu);
3883 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3884 CPU_IDLE : CPU_NOT_IDLE;
3886 rebalance_domains(this_cpu, idle);
3890 * If this cpu is the owner for idle load balancing, then do the
3891 * balancing on behalf of the other idle cpus whose ticks are
3894 if (this_rq->idle_at_tick &&
3895 atomic_read(&nohz.load_balancer) == this_cpu) {
3896 cpumask_t cpus = nohz.cpu_mask;
3900 cpu_clear(this_cpu, cpus);
3901 for_each_cpu_mask(balance_cpu, cpus) {
3903 * If this cpu gets work to do, stop the load balancing
3904 * work being done for other cpus. Next load
3905 * balancing owner will pick it up.
3910 rebalance_domains(balance_cpu, CPU_IDLE);
3912 rq = cpu_rq(balance_cpu);
3913 if (time_after(this_rq->next_balance, rq->next_balance))
3914 this_rq->next_balance = rq->next_balance;
3921 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3923 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3924 * idle load balancing owner or decide to stop the periodic load balancing,
3925 * if the whole system is idle.
3927 static inline void trigger_load_balance(struct rq *rq, int cpu)
3931 * If we were in the nohz mode recently and busy at the current
3932 * scheduler tick, then check if we need to nominate new idle
3935 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3936 rq->in_nohz_recently = 0;
3938 if (atomic_read(&nohz.load_balancer) == cpu) {
3939 cpu_clear(cpu, nohz.cpu_mask);
3940 atomic_set(&nohz.load_balancer, -1);
3943 if (atomic_read(&nohz.load_balancer) == -1) {
3945 * simple selection for now: Nominate the
3946 * first cpu in the nohz list to be the next
3949 * TBD: Traverse the sched domains and nominate
3950 * the nearest cpu in the nohz.cpu_mask.
3952 int ilb = first_cpu(nohz.cpu_mask);
3954 if (ilb < nr_cpu_ids)
3960 * If this cpu is idle and doing idle load balancing for all the
3961 * cpus with ticks stopped, is it time for that to stop?
3963 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
3964 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3970 * If this cpu is idle and the idle load balancing is done by
3971 * someone else, then no need raise the SCHED_SOFTIRQ
3973 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
3974 cpu_isset(cpu, nohz.cpu_mask))
3977 if (time_after_eq(jiffies, rq->next_balance))
3978 raise_softirq(SCHED_SOFTIRQ);
3981 #else /* CONFIG_SMP */
3984 * on UP we do not need to balance between CPUs:
3986 static inline void idle_balance(int cpu, struct rq *rq)
3992 DEFINE_PER_CPU(struct kernel_stat, kstat);
3994 EXPORT_PER_CPU_SYMBOL(kstat);
3997 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3998 * that have not yet been banked in case the task is currently running.
4000 unsigned long long task_sched_runtime(struct task_struct *p)
4002 unsigned long flags;
4006 rq = task_rq_lock(p, &flags);
4007 ns = p->se.sum_exec_runtime;
4008 if (task_current(rq, p)) {
4009 update_rq_clock(rq);
4010 delta_exec = rq->clock - p->se.exec_start;
4011 if ((s64)delta_exec > 0)
4014 task_rq_unlock(rq, &flags);
4020 * Account user cpu time to a process.
4021 * @p: the process that the cpu time gets accounted to
4022 * @cputime: the cpu time spent in user space since the last update
4024 void account_user_time(struct task_struct *p, cputime_t cputime)
4026 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4029 p->utime = cputime_add(p->utime, cputime);
4031 /* Add user time to cpustat. */
4032 tmp = cputime_to_cputime64(cputime);
4033 if (TASK_NICE(p) > 0)
4034 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4036 cpustat->user = cputime64_add(cpustat->user, tmp);
4040 * Account guest cpu time to a process.
4041 * @p: the process that the cpu time gets accounted to
4042 * @cputime: the cpu time spent in virtual machine since the last update
4044 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4047 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4049 tmp = cputime_to_cputime64(cputime);
4051 p->utime = cputime_add(p->utime, cputime);
4052 p->gtime = cputime_add(p->gtime, cputime);
4054 cpustat->user = cputime64_add(cpustat->user, tmp);
4055 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4059 * Account scaled user cpu time to a process.
4060 * @p: the process that the cpu time gets accounted to
4061 * @cputime: the cpu time spent in user space since the last update
4063 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4065 p->utimescaled = cputime_add(p->utimescaled, cputime);
4069 * Account system cpu time to a process.
4070 * @p: the process that the cpu time gets accounted to
4071 * @hardirq_offset: the offset to subtract from hardirq_count()
4072 * @cputime: the cpu time spent in kernel space since the last update
4074 void account_system_time(struct task_struct *p, int hardirq_offset,
4077 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4078 struct rq *rq = this_rq();
4081 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4082 account_guest_time(p, cputime);
4086 p->stime = cputime_add(p->stime, cputime);
4088 /* Add system time to cpustat. */
4089 tmp = cputime_to_cputime64(cputime);
4090 if (hardirq_count() - hardirq_offset)
4091 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4092 else if (softirq_count())
4093 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4094 else if (p != rq->idle)
4095 cpustat->system = cputime64_add(cpustat->system, tmp);
4096 else if (atomic_read(&rq->nr_iowait) > 0)
4097 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4099 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4100 /* Account for system time used */
4101 acct_update_integrals(p);
4105 * Account scaled system cpu time to a process.
4106 * @p: the process that the cpu time gets accounted to
4107 * @hardirq_offset: the offset to subtract from hardirq_count()
4108 * @cputime: the cpu time spent in kernel space since the last update
4110 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4112 p->stimescaled = cputime_add(p->stimescaled, cputime);
4116 * Account for involuntary wait time.
4117 * @p: the process from which the cpu time has been stolen
4118 * @steal: the cpu time spent in involuntary wait
4120 void account_steal_time(struct task_struct *p, cputime_t steal)
4122 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4123 cputime64_t tmp = cputime_to_cputime64(steal);
4124 struct rq *rq = this_rq();
4126 if (p == rq->idle) {
4127 p->stime = cputime_add(p->stime, steal);
4128 if (atomic_read(&rq->nr_iowait) > 0)
4129 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4131 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4133 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4137 * This function gets called by the timer code, with HZ frequency.
4138 * We call it with interrupts disabled.
4140 * It also gets called by the fork code, when changing the parent's
4143 void scheduler_tick(void)
4145 int cpu = smp_processor_id();
4146 struct rq *rq = cpu_rq(cpu);
4147 struct task_struct *curr = rq->curr;
4151 spin_lock(&rq->lock);
4152 update_rq_clock(rq);
4153 update_cpu_load(rq);
4154 curr->sched_class->task_tick(rq, curr, 0);
4155 spin_unlock(&rq->lock);
4158 rq->idle_at_tick = idle_cpu(cpu);
4159 trigger_load_balance(rq, cpu);
4163 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4165 void __kprobes add_preempt_count(int val)
4170 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4172 preempt_count() += val;
4174 * Spinlock count overflowing soon?
4176 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4179 EXPORT_SYMBOL(add_preempt_count);
4181 void __kprobes sub_preempt_count(int val)
4186 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4189 * Is the spinlock portion underflowing?
4191 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4192 !(preempt_count() & PREEMPT_MASK)))
4195 preempt_count() -= val;
4197 EXPORT_SYMBOL(sub_preempt_count);
4202 * Print scheduling while atomic bug:
4204 static noinline void __schedule_bug(struct task_struct *prev)
4206 struct pt_regs *regs = get_irq_regs();
4208 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4209 prev->comm, prev->pid, preempt_count());
4211 debug_show_held_locks(prev);
4213 if (irqs_disabled())
4214 print_irqtrace_events(prev);
4223 * Various schedule()-time debugging checks and statistics:
4225 static inline void schedule_debug(struct task_struct *prev)
4228 * Test if we are atomic. Since do_exit() needs to call into
4229 * schedule() atomically, we ignore that path for now.
4230 * Otherwise, whine if we are scheduling when we should not be.
4232 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4233 __schedule_bug(prev);
4235 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4237 schedstat_inc(this_rq(), sched_count);
4238 #ifdef CONFIG_SCHEDSTATS
4239 if (unlikely(prev->lock_depth >= 0)) {
4240 schedstat_inc(this_rq(), bkl_count);
4241 schedstat_inc(prev, sched_info.bkl_count);
4247 * Pick up the highest-prio task:
4249 static inline struct task_struct *
4250 pick_next_task(struct rq *rq, struct task_struct *prev)
4252 const struct sched_class *class;
4253 struct task_struct *p;
4256 * Optimization: we know that if all tasks are in
4257 * the fair class we can call that function directly:
4259 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4260 p = fair_sched_class.pick_next_task(rq);
4265 class = sched_class_highest;
4267 p = class->pick_next_task(rq);
4271 * Will never be NULL as the idle class always
4272 * returns a non-NULL p:
4274 class = class->next;
4279 * schedule() is the main scheduler function.
4281 asmlinkage void __sched schedule(void)
4283 struct task_struct *prev, *next;
4284 unsigned long *switch_count;
4286 int cpu, hrtick = sched_feat(HRTICK);
4290 cpu = smp_processor_id();
4294 switch_count = &prev->nivcsw;
4296 release_kernel_lock(prev);
4297 need_resched_nonpreemptible:
4299 schedule_debug(prev);
4305 * Do the rq-clock update outside the rq lock:
4307 local_irq_disable();
4308 update_rq_clock(rq);
4309 spin_lock(&rq->lock);
4310 clear_tsk_need_resched(prev);
4312 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4313 if (unlikely(signal_pending_state(prev->state, prev)))
4314 prev->state = TASK_RUNNING;
4316 deactivate_task(rq, prev, 1);
4317 switch_count = &prev->nvcsw;
4321 if (prev->sched_class->pre_schedule)
4322 prev->sched_class->pre_schedule(rq, prev);
4325 if (unlikely(!rq->nr_running))
4326 idle_balance(cpu, rq);
4328 prev->sched_class->put_prev_task(rq, prev);
4329 next = pick_next_task(rq, prev);
4331 if (likely(prev != next)) {
4332 sched_info_switch(prev, next);
4338 context_switch(rq, prev, next); /* unlocks the rq */
4340 * the context switch might have flipped the stack from under
4341 * us, hence refresh the local variables.
4343 cpu = smp_processor_id();
4346 spin_unlock_irq(&rq->lock);
4351 if (unlikely(reacquire_kernel_lock(current) < 0))
4352 goto need_resched_nonpreemptible;
4354 preempt_enable_no_resched();
4355 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4358 EXPORT_SYMBOL(schedule);
4360 #ifdef CONFIG_PREEMPT
4362 * this is the entry point to schedule() from in-kernel preemption
4363 * off of preempt_enable. Kernel preemptions off return from interrupt
4364 * occur there and call schedule directly.
4366 asmlinkage void __sched preempt_schedule(void)
4368 struct thread_info *ti = current_thread_info();
4371 * If there is a non-zero preempt_count or interrupts are disabled,
4372 * we do not want to preempt the current task. Just return..
4374 if (likely(ti->preempt_count || irqs_disabled()))
4378 add_preempt_count(PREEMPT_ACTIVE);
4380 sub_preempt_count(PREEMPT_ACTIVE);
4383 * Check again in case we missed a preemption opportunity
4384 * between schedule and now.
4387 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4389 EXPORT_SYMBOL(preempt_schedule);
4392 * this is the entry point to schedule() from kernel preemption
4393 * off of irq context.
4394 * Note, that this is called and return with irqs disabled. This will
4395 * protect us against recursive calling from irq.
4397 asmlinkage void __sched preempt_schedule_irq(void)
4399 struct thread_info *ti = current_thread_info();
4401 /* Catch callers which need to be fixed */
4402 BUG_ON(ti->preempt_count || !irqs_disabled());
4405 add_preempt_count(PREEMPT_ACTIVE);
4408 local_irq_disable();
4409 sub_preempt_count(PREEMPT_ACTIVE);
4412 * Check again in case we missed a preemption opportunity
4413 * between schedule and now.
4416 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4419 #endif /* CONFIG_PREEMPT */
4421 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4424 return try_to_wake_up(curr->private, mode, sync);
4426 EXPORT_SYMBOL(default_wake_function);
4429 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4430 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4431 * number) then we wake all the non-exclusive tasks and one exclusive task.
4433 * There are circumstances in which we can try to wake a task which has already
4434 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4435 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4437 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4438 int nr_exclusive, int sync, void *key)
4440 wait_queue_t *curr, *next;
4442 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4443 unsigned flags = curr->flags;
4445 if (curr->func(curr, mode, sync, key) &&
4446 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4452 * __wake_up - wake up threads blocked on a waitqueue.
4454 * @mode: which threads
4455 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4456 * @key: is directly passed to the wakeup function
4458 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4459 int nr_exclusive, void *key)
4461 unsigned long flags;
4463 spin_lock_irqsave(&q->lock, flags);
4464 __wake_up_common(q, mode, nr_exclusive, 0, key);
4465 spin_unlock_irqrestore(&q->lock, flags);
4467 EXPORT_SYMBOL(__wake_up);
4470 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4472 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4474 __wake_up_common(q, mode, 1, 0, NULL);
4478 * __wake_up_sync - wake up threads blocked on a waitqueue.
4480 * @mode: which threads
4481 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4483 * The sync wakeup differs that the waker knows that it will schedule
4484 * away soon, so while the target thread will be woken up, it will not
4485 * be migrated to another CPU - ie. the two threads are 'synchronized'
4486 * with each other. This can prevent needless bouncing between CPUs.
4488 * On UP it can prevent extra preemption.
4491 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4493 unsigned long flags;
4499 if (unlikely(!nr_exclusive))
4502 spin_lock_irqsave(&q->lock, flags);
4503 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4504 spin_unlock_irqrestore(&q->lock, flags);
4506 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4508 void complete(struct completion *x)
4510 unsigned long flags;
4512 spin_lock_irqsave(&x->wait.lock, flags);
4514 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4515 spin_unlock_irqrestore(&x->wait.lock, flags);
4517 EXPORT_SYMBOL(complete);
4519 void complete_all(struct completion *x)
4521 unsigned long flags;
4523 spin_lock_irqsave(&x->wait.lock, flags);
4524 x->done += UINT_MAX/2;
4525 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4526 spin_unlock_irqrestore(&x->wait.lock, flags);
4528 EXPORT_SYMBOL(complete_all);
4530 static inline long __sched
4531 do_wait_for_common(struct completion *x, long timeout, int state)
4534 DECLARE_WAITQUEUE(wait, current);
4536 wait.flags |= WQ_FLAG_EXCLUSIVE;
4537 __add_wait_queue_tail(&x->wait, &wait);
4539 if ((state == TASK_INTERRUPTIBLE &&
4540 signal_pending(current)) ||
4541 (state == TASK_KILLABLE &&
4542 fatal_signal_pending(current))) {
4543 timeout = -ERESTARTSYS;
4546 __set_current_state(state);
4547 spin_unlock_irq(&x->wait.lock);
4548 timeout = schedule_timeout(timeout);
4549 spin_lock_irq(&x->wait.lock);
4550 } while (!x->done && timeout);
4551 __remove_wait_queue(&x->wait, &wait);
4556 return timeout ?: 1;
4560 wait_for_common(struct completion *x, long timeout, int state)
4564 spin_lock_irq(&x->wait.lock);
4565 timeout = do_wait_for_common(x, timeout, state);
4566 spin_unlock_irq(&x->wait.lock);
4570 void __sched wait_for_completion(struct completion *x)
4572 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4574 EXPORT_SYMBOL(wait_for_completion);
4576 unsigned long __sched
4577 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4579 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4581 EXPORT_SYMBOL(wait_for_completion_timeout);
4583 int __sched wait_for_completion_interruptible(struct completion *x)
4585 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4586 if (t == -ERESTARTSYS)
4590 EXPORT_SYMBOL(wait_for_completion_interruptible);
4592 unsigned long __sched
4593 wait_for_completion_interruptible_timeout(struct completion *x,
4594 unsigned long timeout)
4596 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4598 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4600 int __sched wait_for_completion_killable(struct completion *x)
4602 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4603 if (t == -ERESTARTSYS)
4607 EXPORT_SYMBOL(wait_for_completion_killable);
4610 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4612 unsigned long flags;
4615 init_waitqueue_entry(&wait, current);
4617 __set_current_state(state);
4619 spin_lock_irqsave(&q->lock, flags);
4620 __add_wait_queue(q, &wait);
4621 spin_unlock(&q->lock);
4622 timeout = schedule_timeout(timeout);
4623 spin_lock_irq(&q->lock);
4624 __remove_wait_queue(q, &wait);
4625 spin_unlock_irqrestore(&q->lock, flags);
4630 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4632 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4634 EXPORT_SYMBOL(interruptible_sleep_on);
4637 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4639 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4641 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4643 void __sched sleep_on(wait_queue_head_t *q)
4645 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4647 EXPORT_SYMBOL(sleep_on);
4649 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4651 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4653 EXPORT_SYMBOL(sleep_on_timeout);
4655 #ifdef CONFIG_RT_MUTEXES
4658 * rt_mutex_setprio - set the current priority of a task
4660 * @prio: prio value (kernel-internal form)
4662 * This function changes the 'effective' priority of a task. It does
4663 * not touch ->normal_prio like __setscheduler().
4665 * Used by the rt_mutex code to implement priority inheritance logic.
4667 void rt_mutex_setprio(struct task_struct *p, int prio)
4669 unsigned long flags;
4670 int oldprio, on_rq, running;
4672 const struct sched_class *prev_class = p->sched_class;
4674 BUG_ON(prio < 0 || prio > MAX_PRIO);
4676 rq = task_rq_lock(p, &flags);
4677 update_rq_clock(rq);
4680 on_rq = p->se.on_rq;
4681 running = task_current(rq, p);
4683 dequeue_task(rq, p, 0);
4685 p->sched_class->put_prev_task(rq, p);
4688 p->sched_class = &rt_sched_class;
4690 p->sched_class = &fair_sched_class;
4695 p->sched_class->set_curr_task(rq);
4697 enqueue_task(rq, p, 0);
4699 check_class_changed(rq, p, prev_class, oldprio, running);
4701 task_rq_unlock(rq, &flags);
4706 void set_user_nice(struct task_struct *p, long nice)
4708 int old_prio, delta, on_rq;
4709 unsigned long flags;
4712 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4715 * We have to be careful, if called from sys_setpriority(),
4716 * the task might be in the middle of scheduling on another CPU.
4718 rq = task_rq_lock(p, &flags);
4719 update_rq_clock(rq);
4721 * The RT priorities are set via sched_setscheduler(), but we still
4722 * allow the 'normal' nice value to be set - but as expected
4723 * it wont have any effect on scheduling until the task is
4724 * SCHED_FIFO/SCHED_RR:
4726 if (task_has_rt_policy(p)) {
4727 p->static_prio = NICE_TO_PRIO(nice);
4730 on_rq = p->se.on_rq;
4732 dequeue_task(rq, p, 0);
4734 p->static_prio = NICE_TO_PRIO(nice);
4737 p->prio = effective_prio(p);
4738 delta = p->prio - old_prio;
4741 enqueue_task(rq, p, 0);
4743 * If the task increased its priority or is running and
4744 * lowered its priority, then reschedule its CPU:
4746 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4747 resched_task(rq->curr);
4750 task_rq_unlock(rq, &flags);
4752 EXPORT_SYMBOL(set_user_nice);
4755 * can_nice - check if a task can reduce its nice value
4759 int can_nice(const struct task_struct *p, const int nice)
4761 /* convert nice value [19,-20] to rlimit style value [1,40] */
4762 int nice_rlim = 20 - nice;
4764 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4765 capable(CAP_SYS_NICE));
4768 #ifdef __ARCH_WANT_SYS_NICE
4771 * sys_nice - change the priority of the current process.
4772 * @increment: priority increment
4774 * sys_setpriority is a more generic, but much slower function that
4775 * does similar things.
4777 asmlinkage long sys_nice(int increment)
4782 * Setpriority might change our priority at the same moment.
4783 * We don't have to worry. Conceptually one call occurs first
4784 * and we have a single winner.
4786 if (increment < -40)
4791 nice = PRIO_TO_NICE(current->static_prio) + increment;
4797 if (increment < 0 && !can_nice(current, nice))
4800 retval = security_task_setnice(current, nice);
4804 set_user_nice(current, nice);
4811 * task_prio - return the priority value of a given task.
4812 * @p: the task in question.
4814 * This is the priority value as seen by users in /proc.
4815 * RT tasks are offset by -200. Normal tasks are centered
4816 * around 0, value goes from -16 to +15.
4818 int task_prio(const struct task_struct *p)
4820 return p->prio - MAX_RT_PRIO;
4824 * task_nice - return the nice value of a given task.
4825 * @p: the task in question.
4827 int task_nice(const struct task_struct *p)
4829 return TASK_NICE(p);
4831 EXPORT_SYMBOL(task_nice);
4834 * idle_cpu - is a given cpu idle currently?
4835 * @cpu: the processor in question.
4837 int idle_cpu(int cpu)
4839 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4843 * idle_task - return the idle task for a given cpu.
4844 * @cpu: the processor in question.
4846 struct task_struct *idle_task(int cpu)
4848 return cpu_rq(cpu)->idle;
4852 * find_process_by_pid - find a process with a matching PID value.
4853 * @pid: the pid in question.
4855 static struct task_struct *find_process_by_pid(pid_t pid)
4857 return pid ? find_task_by_vpid(pid) : current;
4860 /* Actually do priority change: must hold rq lock. */
4862 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4864 BUG_ON(p->se.on_rq);
4867 switch (p->policy) {
4871 p->sched_class = &fair_sched_class;
4875 p->sched_class = &rt_sched_class;
4879 p->rt_priority = prio;
4880 p->normal_prio = normal_prio(p);
4881 /* we are holding p->pi_lock already */
4882 p->prio = rt_mutex_getprio(p);
4887 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4888 * @p: the task in question.
4889 * @policy: new policy.
4890 * @param: structure containing the new RT priority.
4892 * NOTE that the task may be already dead.
4894 int sched_setscheduler(struct task_struct *p, int policy,
4895 struct sched_param *param)
4897 int retval, oldprio, oldpolicy = -1, on_rq, running;
4898 unsigned long flags;
4899 const struct sched_class *prev_class = p->sched_class;
4902 /* may grab non-irq protected spin_locks */
4903 BUG_ON(in_interrupt());
4905 /* double check policy once rq lock held */
4907 policy = oldpolicy = p->policy;
4908 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
4909 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4910 policy != SCHED_IDLE)
4913 * Valid priorities for SCHED_FIFO and SCHED_RR are
4914 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4915 * SCHED_BATCH and SCHED_IDLE is 0.
4917 if (param->sched_priority < 0 ||
4918 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4919 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4921 if (rt_policy(policy) != (param->sched_priority != 0))
4925 * Allow unprivileged RT tasks to decrease priority:
4927 if (!capable(CAP_SYS_NICE)) {
4928 if (rt_policy(policy)) {
4929 unsigned long rlim_rtprio;
4931 if (!lock_task_sighand(p, &flags))
4933 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4934 unlock_task_sighand(p, &flags);
4936 /* can't set/change the rt policy */
4937 if (policy != p->policy && !rlim_rtprio)
4940 /* can't increase priority */
4941 if (param->sched_priority > p->rt_priority &&
4942 param->sched_priority > rlim_rtprio)
4946 * Like positive nice levels, dont allow tasks to
4947 * move out of SCHED_IDLE either:
4949 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4952 /* can't change other user's priorities */
4953 if ((current->euid != p->euid) &&
4954 (current->euid != p->uid))
4958 #ifdef CONFIG_RT_GROUP_SCHED
4960 * Do not allow realtime tasks into groups that have no runtime
4963 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
4967 retval = security_task_setscheduler(p, policy, param);
4971 * make sure no PI-waiters arrive (or leave) while we are
4972 * changing the priority of the task:
4974 spin_lock_irqsave(&p->pi_lock, flags);
4976 * To be able to change p->policy safely, the apropriate
4977 * runqueue lock must be held.
4979 rq = __task_rq_lock(p);
4980 /* recheck policy now with rq lock held */
4981 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4982 policy = oldpolicy = -1;
4983 __task_rq_unlock(rq);
4984 spin_unlock_irqrestore(&p->pi_lock, flags);
4987 update_rq_clock(rq);
4988 on_rq = p->se.on_rq;
4989 running = task_current(rq, p);
4991 deactivate_task(rq, p, 0);
4993 p->sched_class->put_prev_task(rq, p);
4996 __setscheduler(rq, p, policy, param->sched_priority);
4999 p->sched_class->set_curr_task(rq);
5001 activate_task(rq, p, 0);
5003 check_class_changed(rq, p, prev_class, oldprio, running);
5005 __task_rq_unlock(rq);
5006 spin_unlock_irqrestore(&p->pi_lock, flags);
5008 rt_mutex_adjust_pi(p);
5012 EXPORT_SYMBOL_GPL(sched_setscheduler);
5015 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5017 struct sched_param lparam;
5018 struct task_struct *p;
5021 if (!param || pid < 0)
5023 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5028 p = find_process_by_pid(pid);
5030 retval = sched_setscheduler(p, policy, &lparam);
5037 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5038 * @pid: the pid in question.
5039 * @policy: new policy.
5040 * @param: structure containing the new RT priority.
5043 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5045 /* negative values for policy are not valid */
5049 return do_sched_setscheduler(pid, policy, param);
5053 * sys_sched_setparam - set/change the RT priority of a thread
5054 * @pid: the pid in question.
5055 * @param: structure containing the new RT priority.
5057 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5059 return do_sched_setscheduler(pid, -1, param);
5063 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5064 * @pid: the pid in question.
5066 asmlinkage long sys_sched_getscheduler(pid_t pid)
5068 struct task_struct *p;
5075 read_lock(&tasklist_lock);
5076 p = find_process_by_pid(pid);
5078 retval = security_task_getscheduler(p);
5082 read_unlock(&tasklist_lock);
5087 * sys_sched_getscheduler - get the RT priority of a thread
5088 * @pid: the pid in question.
5089 * @param: structure containing the RT priority.
5091 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5093 struct sched_param lp;
5094 struct task_struct *p;
5097 if (!param || pid < 0)
5100 read_lock(&tasklist_lock);
5101 p = find_process_by_pid(pid);
5106 retval = security_task_getscheduler(p);
5110 lp.sched_priority = p->rt_priority;
5111 read_unlock(&tasklist_lock);
5114 * This one might sleep, we cannot do it with a spinlock held ...
5116 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5121 read_unlock(&tasklist_lock);
5125 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5127 cpumask_t cpus_allowed;
5128 cpumask_t new_mask = *in_mask;
5129 struct task_struct *p;
5133 read_lock(&tasklist_lock);
5135 p = find_process_by_pid(pid);
5137 read_unlock(&tasklist_lock);
5143 * It is not safe to call set_cpus_allowed with the
5144 * tasklist_lock held. We will bump the task_struct's
5145 * usage count and then drop tasklist_lock.
5148 read_unlock(&tasklist_lock);
5151 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5152 !capable(CAP_SYS_NICE))
5155 retval = security_task_setscheduler(p, 0, NULL);
5159 cpuset_cpus_allowed(p, &cpus_allowed);
5160 cpus_and(new_mask, new_mask, cpus_allowed);
5162 retval = set_cpus_allowed_ptr(p, &new_mask);
5165 cpuset_cpus_allowed(p, &cpus_allowed);
5166 if (!cpus_subset(new_mask, cpus_allowed)) {
5168 * We must have raced with a concurrent cpuset
5169 * update. Just reset the cpus_allowed to the
5170 * cpuset's cpus_allowed
5172 new_mask = cpus_allowed;
5182 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5183 cpumask_t *new_mask)
5185 if (len < sizeof(cpumask_t)) {
5186 memset(new_mask, 0, sizeof(cpumask_t));
5187 } else if (len > sizeof(cpumask_t)) {
5188 len = sizeof(cpumask_t);
5190 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5194 * sys_sched_setaffinity - set the cpu affinity of a process
5195 * @pid: pid of the process
5196 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5197 * @user_mask_ptr: user-space pointer to the new cpu mask
5199 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5200 unsigned long __user *user_mask_ptr)
5205 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5209 return sched_setaffinity(pid, &new_mask);
5212 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5214 struct task_struct *p;
5218 read_lock(&tasklist_lock);
5221 p = find_process_by_pid(pid);
5225 retval = security_task_getscheduler(p);
5229 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5232 read_unlock(&tasklist_lock);
5239 * sys_sched_getaffinity - get the cpu affinity of a process
5240 * @pid: pid of the process
5241 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5242 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5244 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5245 unsigned long __user *user_mask_ptr)
5250 if (len < sizeof(cpumask_t))
5253 ret = sched_getaffinity(pid, &mask);
5257 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5260 return sizeof(cpumask_t);
5264 * sys_sched_yield - yield the current processor to other threads.
5266 * This function yields the current CPU to other tasks. If there are no
5267 * other threads running on this CPU then this function will return.
5269 asmlinkage long sys_sched_yield(void)
5271 struct rq *rq = this_rq_lock();
5273 schedstat_inc(rq, yld_count);
5274 current->sched_class->yield_task(rq);
5277 * Since we are going to call schedule() anyway, there's
5278 * no need to preempt or enable interrupts:
5280 __release(rq->lock);
5281 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5282 _raw_spin_unlock(&rq->lock);
5283 preempt_enable_no_resched();
5290 static void __cond_resched(void)
5292 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5293 __might_sleep(__FILE__, __LINE__);
5296 * The BKS might be reacquired before we have dropped
5297 * PREEMPT_ACTIVE, which could trigger a second
5298 * cond_resched() call.
5301 add_preempt_count(PREEMPT_ACTIVE);
5303 sub_preempt_count(PREEMPT_ACTIVE);
5304 } while (need_resched());
5307 int __sched _cond_resched(void)
5309 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5310 system_state == SYSTEM_RUNNING) {
5316 EXPORT_SYMBOL(_cond_resched);
5319 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5320 * call schedule, and on return reacquire the lock.
5322 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5323 * operations here to prevent schedule() from being called twice (once via
5324 * spin_unlock(), once by hand).
5326 int cond_resched_lock(spinlock_t *lock)
5328 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5331 if (spin_needbreak(lock) || resched) {
5333 if (resched && need_resched())
5342 EXPORT_SYMBOL(cond_resched_lock);
5344 int __sched cond_resched_softirq(void)
5346 BUG_ON(!in_softirq());
5348 if (need_resched() && system_state == SYSTEM_RUNNING) {
5356 EXPORT_SYMBOL(cond_resched_softirq);
5359 * yield - yield the current processor to other threads.
5361 * This is a shortcut for kernel-space yielding - it marks the
5362 * thread runnable and calls sys_sched_yield().
5364 void __sched yield(void)
5366 set_current_state(TASK_RUNNING);
5369 EXPORT_SYMBOL(yield);
5372 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5373 * that process accounting knows that this is a task in IO wait state.
5375 * But don't do that if it is a deliberate, throttling IO wait (this task
5376 * has set its backing_dev_info: the queue against which it should throttle)
5378 void __sched io_schedule(void)
5380 struct rq *rq = &__raw_get_cpu_var(runqueues);
5382 delayacct_blkio_start();
5383 atomic_inc(&rq->nr_iowait);
5385 atomic_dec(&rq->nr_iowait);
5386 delayacct_blkio_end();
5388 EXPORT_SYMBOL(io_schedule);
5390 long __sched io_schedule_timeout(long timeout)
5392 struct rq *rq = &__raw_get_cpu_var(runqueues);
5395 delayacct_blkio_start();
5396 atomic_inc(&rq->nr_iowait);
5397 ret = schedule_timeout(timeout);
5398 atomic_dec(&rq->nr_iowait);
5399 delayacct_blkio_end();
5404 * sys_sched_get_priority_max - return maximum RT priority.
5405 * @policy: scheduling class.
5407 * this syscall returns the maximum rt_priority that can be used
5408 * by a given scheduling class.
5410 asmlinkage long sys_sched_get_priority_max(int policy)
5417 ret = MAX_USER_RT_PRIO-1;
5429 * sys_sched_get_priority_min - return minimum RT priority.
5430 * @policy: scheduling class.
5432 * this syscall returns the minimum rt_priority that can be used
5433 * by a given scheduling class.
5435 asmlinkage long sys_sched_get_priority_min(int policy)
5453 * sys_sched_rr_get_interval - return the default timeslice of a process.
5454 * @pid: pid of the process.
5455 * @interval: userspace pointer to the timeslice value.
5457 * this syscall writes the default timeslice value of a given process
5458 * into the user-space timespec buffer. A value of '0' means infinity.
5461 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5463 struct task_struct *p;
5464 unsigned int time_slice;
5472 read_lock(&tasklist_lock);
5473 p = find_process_by_pid(pid);
5477 retval = security_task_getscheduler(p);
5482 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5483 * tasks that are on an otherwise idle runqueue:
5486 if (p->policy == SCHED_RR) {
5487 time_slice = DEF_TIMESLICE;
5488 } else if (p->policy != SCHED_FIFO) {
5489 struct sched_entity *se = &p->se;
5490 unsigned long flags;
5493 rq = task_rq_lock(p, &flags);
5494 if (rq->cfs.load.weight)
5495 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5496 task_rq_unlock(rq, &flags);
5498 read_unlock(&tasklist_lock);
5499 jiffies_to_timespec(time_slice, &t);
5500 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5504 read_unlock(&tasklist_lock);
5508 static const char stat_nam[] = "RSDTtZX";
5510 void sched_show_task(struct task_struct *p)
5512 unsigned long free = 0;
5515 state = p->state ? __ffs(p->state) + 1 : 0;
5516 printk(KERN_INFO "%-13.13s %c", p->comm,
5517 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5518 #if BITS_PER_LONG == 32
5519 if (state == TASK_RUNNING)
5520 printk(KERN_CONT " running ");
5522 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5524 if (state == TASK_RUNNING)
5525 printk(KERN_CONT " running task ");
5527 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5529 #ifdef CONFIG_DEBUG_STACK_USAGE
5531 unsigned long *n = end_of_stack(p);
5534 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5537 printk(KERN_CONT "%5lu %5d %6d\n", free,
5538 task_pid_nr(p), task_pid_nr(p->real_parent));
5540 show_stack(p, NULL);
5543 void show_state_filter(unsigned long state_filter)
5545 struct task_struct *g, *p;
5547 #if BITS_PER_LONG == 32
5549 " task PC stack pid father\n");
5552 " task PC stack pid father\n");
5554 read_lock(&tasklist_lock);
5555 do_each_thread(g, p) {
5557 * reset the NMI-timeout, listing all files on a slow
5558 * console might take alot of time:
5560 touch_nmi_watchdog();
5561 if (!state_filter || (p->state & state_filter))
5563 } while_each_thread(g, p);
5565 touch_all_softlockup_watchdogs();
5567 #ifdef CONFIG_SCHED_DEBUG
5568 sysrq_sched_debug_show();
5570 read_unlock(&tasklist_lock);
5572 * Only show locks if all tasks are dumped:
5574 if (state_filter == -1)
5575 debug_show_all_locks();
5578 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5580 idle->sched_class = &idle_sched_class;
5584 * init_idle - set up an idle thread for a given CPU
5585 * @idle: task in question
5586 * @cpu: cpu the idle task belongs to
5588 * NOTE: this function does not set the idle thread's NEED_RESCHED
5589 * flag, to make booting more robust.
5591 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5593 struct rq *rq = cpu_rq(cpu);
5594 unsigned long flags;
5597 idle->se.exec_start = sched_clock();
5599 idle->prio = idle->normal_prio = MAX_PRIO;
5600 idle->cpus_allowed = cpumask_of_cpu(cpu);
5601 __set_task_cpu(idle, cpu);
5603 spin_lock_irqsave(&rq->lock, flags);
5604 rq->curr = rq->idle = idle;
5605 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5608 spin_unlock_irqrestore(&rq->lock, flags);
5610 /* Set the preempt count _outside_ the spinlocks! */
5611 #if defined(CONFIG_PREEMPT)
5612 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5614 task_thread_info(idle)->preempt_count = 0;
5617 * The idle tasks have their own, simple scheduling class:
5619 idle->sched_class = &idle_sched_class;
5623 * In a system that switches off the HZ timer nohz_cpu_mask
5624 * indicates which cpus entered this state. This is used
5625 * in the rcu update to wait only for active cpus. For system
5626 * which do not switch off the HZ timer nohz_cpu_mask should
5627 * always be CPU_MASK_NONE.
5629 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5632 * Increase the granularity value when there are more CPUs,
5633 * because with more CPUs the 'effective latency' as visible
5634 * to users decreases. But the relationship is not linear,
5635 * so pick a second-best guess by going with the log2 of the
5638 * This idea comes from the SD scheduler of Con Kolivas:
5640 static inline void sched_init_granularity(void)
5642 unsigned int factor = 1 + ilog2(num_online_cpus());
5643 const unsigned long limit = 200000000;
5645 sysctl_sched_min_granularity *= factor;
5646 if (sysctl_sched_min_granularity > limit)
5647 sysctl_sched_min_granularity = limit;
5649 sysctl_sched_latency *= factor;
5650 if (sysctl_sched_latency > limit)
5651 sysctl_sched_latency = limit;
5653 sysctl_sched_wakeup_granularity *= factor;
5658 * This is how migration works:
5660 * 1) we queue a struct migration_req structure in the source CPU's
5661 * runqueue and wake up that CPU's migration thread.
5662 * 2) we down() the locked semaphore => thread blocks.
5663 * 3) migration thread wakes up (implicitly it forces the migrated
5664 * thread off the CPU)
5665 * 4) it gets the migration request and checks whether the migrated
5666 * task is still in the wrong runqueue.
5667 * 5) if it's in the wrong runqueue then the migration thread removes
5668 * it and puts it into the right queue.
5669 * 6) migration thread up()s the semaphore.
5670 * 7) we wake up and the migration is done.
5674 * Change a given task's CPU affinity. Migrate the thread to a
5675 * proper CPU and schedule it away if the CPU it's executing on
5676 * is removed from the allowed bitmask.
5678 * NOTE: the caller must have a valid reference to the task, the
5679 * task must not exit() & deallocate itself prematurely. The
5680 * call is not atomic; no spinlocks may be held.
5682 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5684 struct migration_req req;
5685 unsigned long flags;
5689 rq = task_rq_lock(p, &flags);
5690 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5695 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5696 !cpus_equal(p->cpus_allowed, *new_mask))) {
5701 if (p->sched_class->set_cpus_allowed)
5702 p->sched_class->set_cpus_allowed(p, new_mask);
5704 p->cpus_allowed = *new_mask;
5705 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5708 /* Can the task run on the task's current CPU? If so, we're done */
5709 if (cpu_isset(task_cpu(p), *new_mask))
5712 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5713 /* Need help from migration thread: drop lock and wait. */
5714 task_rq_unlock(rq, &flags);
5715 wake_up_process(rq->migration_thread);
5716 wait_for_completion(&req.done);
5717 tlb_migrate_finish(p->mm);
5721 task_rq_unlock(rq, &flags);
5725 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5728 * Move (not current) task off this cpu, onto dest cpu. We're doing
5729 * this because either it can't run here any more (set_cpus_allowed()
5730 * away from this CPU, or CPU going down), or because we're
5731 * attempting to rebalance this task on exec (sched_exec).
5733 * So we race with normal scheduler movements, but that's OK, as long
5734 * as the task is no longer on this CPU.
5736 * Returns non-zero if task was successfully migrated.
5738 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5740 struct rq *rq_dest, *rq_src;
5743 if (unlikely(cpu_is_offline(dest_cpu)))
5746 rq_src = cpu_rq(src_cpu);
5747 rq_dest = cpu_rq(dest_cpu);
5749 double_rq_lock(rq_src, rq_dest);
5750 /* Already moved. */
5751 if (task_cpu(p) != src_cpu)
5753 /* Affinity changed (again). */
5754 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5757 on_rq = p->se.on_rq;
5759 deactivate_task(rq_src, p, 0);
5761 set_task_cpu(p, dest_cpu);
5763 activate_task(rq_dest, p, 0);
5764 check_preempt_curr(rq_dest, p);
5768 double_rq_unlock(rq_src, rq_dest);
5773 * migration_thread - this is a highprio system thread that performs
5774 * thread migration by bumping thread off CPU then 'pushing' onto
5777 static int migration_thread(void *data)
5779 int cpu = (long)data;
5783 BUG_ON(rq->migration_thread != current);
5785 set_current_state(TASK_INTERRUPTIBLE);
5786 while (!kthread_should_stop()) {
5787 struct migration_req *req;
5788 struct list_head *head;
5790 spin_lock_irq(&rq->lock);
5792 if (cpu_is_offline(cpu)) {
5793 spin_unlock_irq(&rq->lock);
5797 if (rq->active_balance) {
5798 active_load_balance(rq, cpu);
5799 rq->active_balance = 0;
5802 head = &rq->migration_queue;
5804 if (list_empty(head)) {
5805 spin_unlock_irq(&rq->lock);
5807 set_current_state(TASK_INTERRUPTIBLE);
5810 req = list_entry(head->next, struct migration_req, list);
5811 list_del_init(head->next);
5813 spin_unlock(&rq->lock);
5814 __migrate_task(req->task, cpu, req->dest_cpu);
5817 complete(&req->done);
5819 __set_current_state(TASK_RUNNING);
5823 /* Wait for kthread_stop */
5824 set_current_state(TASK_INTERRUPTIBLE);
5825 while (!kthread_should_stop()) {
5827 set_current_state(TASK_INTERRUPTIBLE);
5829 __set_current_state(TASK_RUNNING);
5833 #ifdef CONFIG_HOTPLUG_CPU
5835 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5839 local_irq_disable();
5840 ret = __migrate_task(p, src_cpu, dest_cpu);
5846 * Figure out where task on dead CPU should go, use force if necessary.
5847 * NOTE: interrupts should be disabled by the caller
5849 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5851 unsigned long flags;
5858 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5859 cpus_and(mask, mask, p->cpus_allowed);
5860 dest_cpu = any_online_cpu(mask);
5862 /* On any allowed CPU? */
5863 if (dest_cpu >= nr_cpu_ids)
5864 dest_cpu = any_online_cpu(p->cpus_allowed);
5866 /* No more Mr. Nice Guy. */
5867 if (dest_cpu >= nr_cpu_ids) {
5868 cpumask_t cpus_allowed;
5870 cpuset_cpus_allowed_locked(p, &cpus_allowed);
5872 * Try to stay on the same cpuset, where the
5873 * current cpuset may be a subset of all cpus.
5874 * The cpuset_cpus_allowed_locked() variant of
5875 * cpuset_cpus_allowed() will not block. It must be
5876 * called within calls to cpuset_lock/cpuset_unlock.
5878 rq = task_rq_lock(p, &flags);
5879 p->cpus_allowed = cpus_allowed;
5880 dest_cpu = any_online_cpu(p->cpus_allowed);
5881 task_rq_unlock(rq, &flags);
5884 * Don't tell them about moving exiting tasks or
5885 * kernel threads (both mm NULL), since they never
5888 if (p->mm && printk_ratelimit()) {
5889 printk(KERN_INFO "process %d (%s) no "
5890 "longer affine to cpu%d\n",
5891 task_pid_nr(p), p->comm, dead_cpu);
5894 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
5898 * While a dead CPU has no uninterruptible tasks queued at this point,
5899 * it might still have a nonzero ->nr_uninterruptible counter, because
5900 * for performance reasons the counter is not stricly tracking tasks to
5901 * their home CPUs. So we just add the counter to another CPU's counter,
5902 * to keep the global sum constant after CPU-down:
5904 static void migrate_nr_uninterruptible(struct rq *rq_src)
5906 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
5907 unsigned long flags;
5909 local_irq_save(flags);
5910 double_rq_lock(rq_src, rq_dest);
5911 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5912 rq_src->nr_uninterruptible = 0;
5913 double_rq_unlock(rq_src, rq_dest);
5914 local_irq_restore(flags);
5917 /* Run through task list and migrate tasks from the dead cpu. */
5918 static void migrate_live_tasks(int src_cpu)
5920 struct task_struct *p, *t;
5922 read_lock(&tasklist_lock);
5924 do_each_thread(t, p) {
5928 if (task_cpu(p) == src_cpu)
5929 move_task_off_dead_cpu(src_cpu, p);
5930 } while_each_thread(t, p);
5932 read_unlock(&tasklist_lock);
5936 * Schedules idle task to be the next runnable task on current CPU.
5937 * It does so by boosting its priority to highest possible.
5938 * Used by CPU offline code.
5940 void sched_idle_next(void)
5942 int this_cpu = smp_processor_id();
5943 struct rq *rq = cpu_rq(this_cpu);
5944 struct task_struct *p = rq->idle;
5945 unsigned long flags;
5947 /* cpu has to be offline */
5948 BUG_ON(cpu_online(this_cpu));
5951 * Strictly not necessary since rest of the CPUs are stopped by now
5952 * and interrupts disabled on the current cpu.
5954 spin_lock_irqsave(&rq->lock, flags);
5956 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5958 update_rq_clock(rq);
5959 activate_task(rq, p, 0);
5961 spin_unlock_irqrestore(&rq->lock, flags);
5965 * Ensures that the idle task is using init_mm right before its cpu goes
5968 void idle_task_exit(void)
5970 struct mm_struct *mm = current->active_mm;
5972 BUG_ON(cpu_online(smp_processor_id()));
5975 switch_mm(mm, &init_mm, current);
5979 /* called under rq->lock with disabled interrupts */
5980 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5982 struct rq *rq = cpu_rq(dead_cpu);
5984 /* Must be exiting, otherwise would be on tasklist. */
5985 BUG_ON(!p->exit_state);
5987 /* Cannot have done final schedule yet: would have vanished. */
5988 BUG_ON(p->state == TASK_DEAD);
5993 * Drop lock around migration; if someone else moves it,
5994 * that's OK. No task can be added to this CPU, so iteration is
5997 spin_unlock_irq(&rq->lock);
5998 move_task_off_dead_cpu(dead_cpu, p);
5999 spin_lock_irq(&rq->lock);
6004 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6005 static void migrate_dead_tasks(unsigned int dead_cpu)
6007 struct rq *rq = cpu_rq(dead_cpu);
6008 struct task_struct *next;
6011 if (!rq->nr_running)
6013 update_rq_clock(rq);
6014 next = pick_next_task(rq, rq->curr);
6017 migrate_dead(dead_cpu, next);
6021 #endif /* CONFIG_HOTPLUG_CPU */
6023 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6025 static struct ctl_table sd_ctl_dir[] = {
6027 .procname = "sched_domain",
6033 static struct ctl_table sd_ctl_root[] = {
6035 .ctl_name = CTL_KERN,
6036 .procname = "kernel",
6038 .child = sd_ctl_dir,
6043 static struct ctl_table *sd_alloc_ctl_entry(int n)
6045 struct ctl_table *entry =
6046 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6051 static void sd_free_ctl_entry(struct ctl_table **tablep)
6053 struct ctl_table *entry;
6056 * In the intermediate directories, both the child directory and
6057 * procname are dynamically allocated and could fail but the mode
6058 * will always be set. In the lowest directory the names are
6059 * static strings and all have proc handlers.
6061 for (entry = *tablep; entry->mode; entry++) {
6063 sd_free_ctl_entry(&entry->child);
6064 if (entry->proc_handler == NULL)
6065 kfree(entry->procname);
6073 set_table_entry(struct ctl_table *entry,
6074 const char *procname, void *data, int maxlen,
6075 mode_t mode, proc_handler *proc_handler)
6077 entry->procname = procname;
6079 entry->maxlen = maxlen;
6081 entry->proc_handler = proc_handler;
6084 static struct ctl_table *
6085 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6087 struct ctl_table *table = sd_alloc_ctl_entry(12);
6092 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6093 sizeof(long), 0644, proc_doulongvec_minmax);
6094 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6095 sizeof(long), 0644, proc_doulongvec_minmax);
6096 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6097 sizeof(int), 0644, proc_dointvec_minmax);
6098 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6099 sizeof(int), 0644, proc_dointvec_minmax);
6100 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6101 sizeof(int), 0644, proc_dointvec_minmax);
6102 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6103 sizeof(int), 0644, proc_dointvec_minmax);
6104 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6105 sizeof(int), 0644, proc_dointvec_minmax);
6106 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6107 sizeof(int), 0644, proc_dointvec_minmax);
6108 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6109 sizeof(int), 0644, proc_dointvec_minmax);
6110 set_table_entry(&table[9], "cache_nice_tries",
6111 &sd->cache_nice_tries,
6112 sizeof(int), 0644, proc_dointvec_minmax);
6113 set_table_entry(&table[10], "flags", &sd->flags,
6114 sizeof(int), 0644, proc_dointvec_minmax);
6115 /* &table[11] is terminator */
6120 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6122 struct ctl_table *entry, *table;
6123 struct sched_domain *sd;
6124 int domain_num = 0, i;
6127 for_each_domain(cpu, sd)
6129 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6134 for_each_domain(cpu, sd) {
6135 snprintf(buf, 32, "domain%d", i);
6136 entry->procname = kstrdup(buf, GFP_KERNEL);
6138 entry->child = sd_alloc_ctl_domain_table(sd);
6145 static struct ctl_table_header *sd_sysctl_header;
6146 static void register_sched_domain_sysctl(void)
6148 int i, cpu_num = num_online_cpus();
6149 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6152 WARN_ON(sd_ctl_dir[0].child);
6153 sd_ctl_dir[0].child = entry;
6158 for_each_online_cpu(i) {
6159 snprintf(buf, 32, "cpu%d", i);
6160 entry->procname = kstrdup(buf, GFP_KERNEL);
6162 entry->child = sd_alloc_ctl_cpu_table(i);
6166 WARN_ON(sd_sysctl_header);
6167 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6170 /* may be called multiple times per register */
6171 static void unregister_sched_domain_sysctl(void)
6173 if (sd_sysctl_header)
6174 unregister_sysctl_table(sd_sysctl_header);
6175 sd_sysctl_header = NULL;
6176 if (sd_ctl_dir[0].child)
6177 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6180 static void register_sched_domain_sysctl(void)
6183 static void unregister_sched_domain_sysctl(void)
6188 static void set_rq_online(struct rq *rq)
6191 const struct sched_class *class;
6193 cpu_set(rq->cpu, rq->rd->online);
6196 for_each_class(class) {
6197 if (class->rq_online)
6198 class->rq_online(rq);
6203 static void set_rq_offline(struct rq *rq)
6206 const struct sched_class *class;
6208 for_each_class(class) {
6209 if (class->rq_offline)
6210 class->rq_offline(rq);
6213 cpu_clear(rq->cpu, rq->rd->online);
6219 * migration_call - callback that gets triggered when a CPU is added.
6220 * Here we can start up the necessary migration thread for the new CPU.
6222 static int __cpuinit
6223 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6225 struct task_struct *p;
6226 int cpu = (long)hcpu;
6227 unsigned long flags;
6232 case CPU_UP_PREPARE:
6233 case CPU_UP_PREPARE_FROZEN:
6234 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6237 kthread_bind(p, cpu);
6238 /* Must be high prio: stop_machine expects to yield to it. */
6239 rq = task_rq_lock(p, &flags);
6240 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6241 task_rq_unlock(rq, &flags);
6242 cpu_rq(cpu)->migration_thread = p;
6246 case CPU_ONLINE_FROZEN:
6247 /* Strictly unnecessary, as first user will wake it. */
6248 wake_up_process(cpu_rq(cpu)->migration_thread);
6250 /* Update our root-domain */
6252 spin_lock_irqsave(&rq->lock, flags);
6254 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6258 spin_unlock_irqrestore(&rq->lock, flags);
6261 #ifdef CONFIG_HOTPLUG_CPU
6262 case CPU_UP_CANCELED:
6263 case CPU_UP_CANCELED_FROZEN:
6264 if (!cpu_rq(cpu)->migration_thread)
6266 /* Unbind it from offline cpu so it can run. Fall thru. */
6267 kthread_bind(cpu_rq(cpu)->migration_thread,
6268 any_online_cpu(cpu_online_map));
6269 kthread_stop(cpu_rq(cpu)->migration_thread);
6270 cpu_rq(cpu)->migration_thread = NULL;
6274 case CPU_DEAD_FROZEN:
6275 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6276 migrate_live_tasks(cpu);
6278 kthread_stop(rq->migration_thread);
6279 rq->migration_thread = NULL;
6280 /* Idle task back to normal (off runqueue, low prio) */
6281 spin_lock_irq(&rq->lock);
6282 update_rq_clock(rq);
6283 deactivate_task(rq, rq->idle, 0);
6284 rq->idle->static_prio = MAX_PRIO;
6285 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6286 rq->idle->sched_class = &idle_sched_class;
6287 migrate_dead_tasks(cpu);
6288 spin_unlock_irq(&rq->lock);
6290 migrate_nr_uninterruptible(rq);
6291 BUG_ON(rq->nr_running != 0);
6294 * No need to migrate the tasks: it was best-effort if
6295 * they didn't take sched_hotcpu_mutex. Just wake up
6298 spin_lock_irq(&rq->lock);
6299 while (!list_empty(&rq->migration_queue)) {
6300 struct migration_req *req;
6302 req = list_entry(rq->migration_queue.next,
6303 struct migration_req, list);
6304 list_del_init(&req->list);
6305 complete(&req->done);
6307 spin_unlock_irq(&rq->lock);
6311 case CPU_DYING_FROZEN:
6312 /* Update our root-domain */
6314 spin_lock_irqsave(&rq->lock, flags);
6316 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6319 spin_unlock_irqrestore(&rq->lock, flags);
6326 /* Register at highest priority so that task migration (migrate_all_tasks)
6327 * happens before everything else.
6329 static struct notifier_block __cpuinitdata migration_notifier = {
6330 .notifier_call = migration_call,
6334 void __init migration_init(void)
6336 void *cpu = (void *)(long)smp_processor_id();
6339 /* Start one for the boot CPU: */
6340 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6341 BUG_ON(err == NOTIFY_BAD);
6342 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6343 register_cpu_notifier(&migration_notifier);
6349 #ifdef CONFIG_SCHED_DEBUG
6351 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6364 case SD_LV_ALLNODES:
6373 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6374 cpumask_t *groupmask)
6376 struct sched_group *group = sd->groups;
6379 cpulist_scnprintf(str, sizeof(str), sd->span);
6380 cpus_clear(*groupmask);
6382 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6384 if (!(sd->flags & SD_LOAD_BALANCE)) {
6385 printk("does not load-balance\n");
6387 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6392 printk(KERN_CONT "span %s level %s\n",
6393 str, sd_level_to_string(sd->level));
6395 if (!cpu_isset(cpu, sd->span)) {
6396 printk(KERN_ERR "ERROR: domain->span does not contain "
6399 if (!cpu_isset(cpu, group->cpumask)) {
6400 printk(KERN_ERR "ERROR: domain->groups does not contain"
6404 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6408 printk(KERN_ERR "ERROR: group is NULL\n");
6412 if (!group->__cpu_power) {
6413 printk(KERN_CONT "\n");
6414 printk(KERN_ERR "ERROR: domain->cpu_power not "
6419 if (!cpus_weight(group->cpumask)) {
6420 printk(KERN_CONT "\n");
6421 printk(KERN_ERR "ERROR: empty group\n");
6425 if (cpus_intersects(*groupmask, group->cpumask)) {
6426 printk(KERN_CONT "\n");
6427 printk(KERN_ERR "ERROR: repeated CPUs\n");
6431 cpus_or(*groupmask, *groupmask, group->cpumask);
6433 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6434 printk(KERN_CONT " %s", str);
6436 group = group->next;
6437 } while (group != sd->groups);
6438 printk(KERN_CONT "\n");
6440 if (!cpus_equal(sd->span, *groupmask))
6441 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6443 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6444 printk(KERN_ERR "ERROR: parent span is not a superset "
6445 "of domain->span\n");
6449 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6451 cpumask_t *groupmask;
6455 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6459 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6461 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6463 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6468 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6477 #else /* !CONFIG_SCHED_DEBUG */
6478 # define sched_domain_debug(sd, cpu) do { } while (0)
6479 #endif /* CONFIG_SCHED_DEBUG */
6481 static int sd_degenerate(struct sched_domain *sd)
6483 if (cpus_weight(sd->span) == 1)
6486 /* Following flags need at least 2 groups */
6487 if (sd->flags & (SD_LOAD_BALANCE |
6488 SD_BALANCE_NEWIDLE |
6492 SD_SHARE_PKG_RESOURCES)) {
6493 if (sd->groups != sd->groups->next)
6497 /* Following flags don't use groups */
6498 if (sd->flags & (SD_WAKE_IDLE |
6507 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6509 unsigned long cflags = sd->flags, pflags = parent->flags;
6511 if (sd_degenerate(parent))
6514 if (!cpus_equal(sd->span, parent->span))
6517 /* Does parent contain flags not in child? */
6518 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6519 if (cflags & SD_WAKE_AFFINE)
6520 pflags &= ~SD_WAKE_BALANCE;
6521 /* Flags needing groups don't count if only 1 group in parent */
6522 if (parent->groups == parent->groups->next) {
6523 pflags &= ~(SD_LOAD_BALANCE |
6524 SD_BALANCE_NEWIDLE |
6528 SD_SHARE_PKG_RESOURCES);
6530 if (~cflags & pflags)
6536 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6538 unsigned long flags;
6540 spin_lock_irqsave(&rq->lock, flags);
6543 struct root_domain *old_rd = rq->rd;
6545 if (cpu_isset(rq->cpu, old_rd->online))
6548 cpu_clear(rq->cpu, old_rd->span);
6550 if (atomic_dec_and_test(&old_rd->refcount))
6554 atomic_inc(&rd->refcount);
6557 cpu_set(rq->cpu, rd->span);
6558 if (cpu_isset(rq->cpu, cpu_online_map))
6561 spin_unlock_irqrestore(&rq->lock, flags);
6564 static void init_rootdomain(struct root_domain *rd)
6566 memset(rd, 0, sizeof(*rd));
6568 cpus_clear(rd->span);
6569 cpus_clear(rd->online);
6571 cpupri_init(&rd->cpupri);
6574 static void init_defrootdomain(void)
6576 init_rootdomain(&def_root_domain);
6577 atomic_set(&def_root_domain.refcount, 1);
6580 static struct root_domain *alloc_rootdomain(void)
6582 struct root_domain *rd;
6584 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6588 init_rootdomain(rd);
6594 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6595 * hold the hotplug lock.
6598 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6600 struct rq *rq = cpu_rq(cpu);
6601 struct sched_domain *tmp;
6603 /* Remove the sched domains which do not contribute to scheduling. */
6604 for (tmp = sd; tmp; tmp = tmp->parent) {
6605 struct sched_domain *parent = tmp->parent;
6608 if (sd_parent_degenerate(tmp, parent)) {
6609 tmp->parent = parent->parent;
6611 parent->parent->child = tmp;
6615 if (sd && sd_degenerate(sd)) {
6621 sched_domain_debug(sd, cpu);
6623 rq_attach_root(rq, rd);
6624 rcu_assign_pointer(rq->sd, sd);
6627 /* cpus with isolated domains */
6628 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6630 /* Setup the mask of cpus configured for isolated domains */
6631 static int __init isolated_cpu_setup(char *str)
6633 int ints[NR_CPUS], i;
6635 str = get_options(str, ARRAY_SIZE(ints), ints);
6636 cpus_clear(cpu_isolated_map);
6637 for (i = 1; i <= ints[0]; i++)
6638 if (ints[i] < NR_CPUS)
6639 cpu_set(ints[i], cpu_isolated_map);
6643 __setup("isolcpus=", isolated_cpu_setup);
6646 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6647 * to a function which identifies what group(along with sched group) a CPU
6648 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6649 * (due to the fact that we keep track of groups covered with a cpumask_t).
6651 * init_sched_build_groups will build a circular linked list of the groups
6652 * covered by the given span, and will set each group's ->cpumask correctly,
6653 * and ->cpu_power to 0.
6656 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6657 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6658 struct sched_group **sg,
6659 cpumask_t *tmpmask),
6660 cpumask_t *covered, cpumask_t *tmpmask)
6662 struct sched_group *first = NULL, *last = NULL;
6665 cpus_clear(*covered);
6667 for_each_cpu_mask(i, *span) {
6668 struct sched_group *sg;
6669 int group = group_fn(i, cpu_map, &sg, tmpmask);
6672 if (cpu_isset(i, *covered))
6675 cpus_clear(sg->cpumask);
6676 sg->__cpu_power = 0;
6678 for_each_cpu_mask(j, *span) {
6679 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6682 cpu_set(j, *covered);
6683 cpu_set(j, sg->cpumask);
6694 #define SD_NODES_PER_DOMAIN 16
6699 * find_next_best_node - find the next node to include in a sched_domain
6700 * @node: node whose sched_domain we're building
6701 * @used_nodes: nodes already in the sched_domain
6703 * Find the next node to include in a given scheduling domain. Simply
6704 * finds the closest node not already in the @used_nodes map.
6706 * Should use nodemask_t.
6708 static int find_next_best_node(int node, nodemask_t *used_nodes)
6710 int i, n, val, min_val, best_node = 0;
6714 for (i = 0; i < MAX_NUMNODES; i++) {
6715 /* Start at @node */
6716 n = (node + i) % MAX_NUMNODES;
6718 if (!nr_cpus_node(n))
6721 /* Skip already used nodes */
6722 if (node_isset(n, *used_nodes))
6725 /* Simple min distance search */
6726 val = node_distance(node, n);
6728 if (val < min_val) {
6734 node_set(best_node, *used_nodes);
6739 * sched_domain_node_span - get a cpumask for a node's sched_domain
6740 * @node: node whose cpumask we're constructing
6741 * @span: resulting cpumask
6743 * Given a node, construct a good cpumask for its sched_domain to span. It
6744 * should be one that prevents unnecessary balancing, but also spreads tasks
6747 static void sched_domain_node_span(int node, cpumask_t *span)
6749 nodemask_t used_nodes;
6750 node_to_cpumask_ptr(nodemask, node);
6754 nodes_clear(used_nodes);
6756 cpus_or(*span, *span, *nodemask);
6757 node_set(node, used_nodes);
6759 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6760 int next_node = find_next_best_node(node, &used_nodes);
6762 node_to_cpumask_ptr_next(nodemask, next_node);
6763 cpus_or(*span, *span, *nodemask);
6766 #endif /* CONFIG_NUMA */
6768 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6771 * SMT sched-domains:
6773 #ifdef CONFIG_SCHED_SMT
6774 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6775 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6778 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6782 *sg = &per_cpu(sched_group_cpus, cpu);
6785 #endif /* CONFIG_SCHED_SMT */
6788 * multi-core sched-domains:
6790 #ifdef CONFIG_SCHED_MC
6791 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6792 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6793 #endif /* CONFIG_SCHED_MC */
6795 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6797 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6802 *mask = per_cpu(cpu_sibling_map, cpu);
6803 cpus_and(*mask, *mask, *cpu_map);
6804 group = first_cpu(*mask);
6806 *sg = &per_cpu(sched_group_core, group);
6809 #elif defined(CONFIG_SCHED_MC)
6811 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6815 *sg = &per_cpu(sched_group_core, cpu);
6820 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6821 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6824 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6828 #ifdef CONFIG_SCHED_MC
6829 *mask = cpu_coregroup_map(cpu);
6830 cpus_and(*mask, *mask, *cpu_map);
6831 group = first_cpu(*mask);
6832 #elif defined(CONFIG_SCHED_SMT)
6833 *mask = per_cpu(cpu_sibling_map, cpu);
6834 cpus_and(*mask, *mask, *cpu_map);
6835 group = first_cpu(*mask);
6840 *sg = &per_cpu(sched_group_phys, group);
6846 * The init_sched_build_groups can't handle what we want to do with node
6847 * groups, so roll our own. Now each node has its own list of groups which
6848 * gets dynamically allocated.
6850 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6851 static struct sched_group ***sched_group_nodes_bycpu;
6853 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6854 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6856 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6857 struct sched_group **sg, cpumask_t *nodemask)
6861 *nodemask = node_to_cpumask(cpu_to_node(cpu));
6862 cpus_and(*nodemask, *nodemask, *cpu_map);
6863 group = first_cpu(*nodemask);
6866 *sg = &per_cpu(sched_group_allnodes, group);
6870 static void init_numa_sched_groups_power(struct sched_group *group_head)
6872 struct sched_group *sg = group_head;
6878 for_each_cpu_mask(j, sg->cpumask) {
6879 struct sched_domain *sd;
6881 sd = &per_cpu(phys_domains, j);
6882 if (j != first_cpu(sd->groups->cpumask)) {
6884 * Only add "power" once for each
6890 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
6893 } while (sg != group_head);
6895 #endif /* CONFIG_NUMA */
6898 /* Free memory allocated for various sched_group structures */
6899 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6903 for_each_cpu_mask(cpu, *cpu_map) {
6904 struct sched_group **sched_group_nodes
6905 = sched_group_nodes_bycpu[cpu];
6907 if (!sched_group_nodes)
6910 for (i = 0; i < MAX_NUMNODES; i++) {
6911 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6913 *nodemask = node_to_cpumask(i);
6914 cpus_and(*nodemask, *nodemask, *cpu_map);
6915 if (cpus_empty(*nodemask))
6925 if (oldsg != sched_group_nodes[i])
6928 kfree(sched_group_nodes);
6929 sched_group_nodes_bycpu[cpu] = NULL;
6932 #else /* !CONFIG_NUMA */
6933 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
6936 #endif /* CONFIG_NUMA */
6939 * Initialize sched groups cpu_power.
6941 * cpu_power indicates the capacity of sched group, which is used while
6942 * distributing the load between different sched groups in a sched domain.
6943 * Typically cpu_power for all the groups in a sched domain will be same unless
6944 * there are asymmetries in the topology. If there are asymmetries, group
6945 * having more cpu_power will pickup more load compared to the group having
6948 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6949 * the maximum number of tasks a group can handle in the presence of other idle
6950 * or lightly loaded groups in the same sched domain.
6952 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6954 struct sched_domain *child;
6955 struct sched_group *group;
6957 WARN_ON(!sd || !sd->groups);
6959 if (cpu != first_cpu(sd->groups->cpumask))
6964 sd->groups->__cpu_power = 0;
6967 * For perf policy, if the groups in child domain share resources
6968 * (for example cores sharing some portions of the cache hierarchy
6969 * or SMT), then set this domain groups cpu_power such that each group
6970 * can handle only one task, when there are other idle groups in the
6971 * same sched domain.
6973 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
6975 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
6976 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
6981 * add cpu_power of each child group to this groups cpu_power
6983 group = child->groups;
6985 sg_inc_cpu_power(sd->groups, group->__cpu_power);
6986 group = group->next;
6987 } while (group != child->groups);
6991 * Initializers for schedule domains
6992 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6995 #define SD_INIT(sd, type) sd_init_##type(sd)
6996 #define SD_INIT_FUNC(type) \
6997 static noinline void sd_init_##type(struct sched_domain *sd) \
6999 memset(sd, 0, sizeof(*sd)); \
7000 *sd = SD_##type##_INIT; \
7001 sd->level = SD_LV_##type; \
7006 SD_INIT_FUNC(ALLNODES)
7009 #ifdef CONFIG_SCHED_SMT
7010 SD_INIT_FUNC(SIBLING)
7012 #ifdef CONFIG_SCHED_MC
7017 * To minimize stack usage kmalloc room for cpumasks and share the
7018 * space as the usage in build_sched_domains() dictates. Used only
7019 * if the amount of space is significant.
7022 cpumask_t tmpmask; /* make this one first */
7025 cpumask_t this_sibling_map;
7026 cpumask_t this_core_map;
7028 cpumask_t send_covered;
7031 cpumask_t domainspan;
7033 cpumask_t notcovered;
7038 #define SCHED_CPUMASK_ALLOC 1
7039 #define SCHED_CPUMASK_FREE(v) kfree(v)
7040 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7042 #define SCHED_CPUMASK_ALLOC 0
7043 #define SCHED_CPUMASK_FREE(v)
7044 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7047 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7048 ((unsigned long)(a) + offsetof(struct allmasks, v))
7050 static int default_relax_domain_level = -1;
7052 static int __init setup_relax_domain_level(char *str)
7056 val = simple_strtoul(str, NULL, 0);
7057 if (val < SD_LV_MAX)
7058 default_relax_domain_level = val;
7062 __setup("relax_domain_level=", setup_relax_domain_level);
7064 static void set_domain_attribute(struct sched_domain *sd,
7065 struct sched_domain_attr *attr)
7069 if (!attr || attr->relax_domain_level < 0) {
7070 if (default_relax_domain_level < 0)
7073 request = default_relax_domain_level;
7075 request = attr->relax_domain_level;
7076 if (request < sd->level) {
7077 /* turn off idle balance on this domain */
7078 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7080 /* turn on idle balance on this domain */
7081 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7086 * Build sched domains for a given set of cpus and attach the sched domains
7087 * to the individual cpus
7089 static int __build_sched_domains(const cpumask_t *cpu_map,
7090 struct sched_domain_attr *attr)
7093 struct root_domain *rd;
7094 SCHED_CPUMASK_DECLARE(allmasks);
7097 struct sched_group **sched_group_nodes = NULL;
7098 int sd_allnodes = 0;
7101 * Allocate the per-node list of sched groups
7103 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7105 if (!sched_group_nodes) {
7106 printk(KERN_WARNING "Can not alloc sched group node list\n");
7111 rd = alloc_rootdomain();
7113 printk(KERN_WARNING "Cannot alloc root domain\n");
7115 kfree(sched_group_nodes);
7120 #if SCHED_CPUMASK_ALLOC
7121 /* get space for all scratch cpumask variables */
7122 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7124 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7127 kfree(sched_group_nodes);
7132 tmpmask = (cpumask_t *)allmasks;
7136 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7140 * Set up domains for cpus specified by the cpu_map.
7142 for_each_cpu_mask(i, *cpu_map) {
7143 struct sched_domain *sd = NULL, *p;
7144 SCHED_CPUMASK_VAR(nodemask, allmasks);
7146 *nodemask = node_to_cpumask(cpu_to_node(i));
7147 cpus_and(*nodemask, *nodemask, *cpu_map);
7150 if (cpus_weight(*cpu_map) >
7151 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7152 sd = &per_cpu(allnodes_domains, i);
7153 SD_INIT(sd, ALLNODES);
7154 set_domain_attribute(sd, attr);
7155 sd->span = *cpu_map;
7156 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7162 sd = &per_cpu(node_domains, i);
7164 set_domain_attribute(sd, attr);
7165 sched_domain_node_span(cpu_to_node(i), &sd->span);
7169 cpus_and(sd->span, sd->span, *cpu_map);
7173 sd = &per_cpu(phys_domains, i);
7175 set_domain_attribute(sd, attr);
7176 sd->span = *nodemask;
7180 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7182 #ifdef CONFIG_SCHED_MC
7184 sd = &per_cpu(core_domains, i);
7186 set_domain_attribute(sd, attr);
7187 sd->span = cpu_coregroup_map(i);
7188 cpus_and(sd->span, sd->span, *cpu_map);
7191 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7194 #ifdef CONFIG_SCHED_SMT
7196 sd = &per_cpu(cpu_domains, i);
7197 SD_INIT(sd, SIBLING);
7198 set_domain_attribute(sd, attr);
7199 sd->span = per_cpu(cpu_sibling_map, i);
7200 cpus_and(sd->span, sd->span, *cpu_map);
7203 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7207 #ifdef CONFIG_SCHED_SMT
7208 /* Set up CPU (sibling) groups */
7209 for_each_cpu_mask(i, *cpu_map) {
7210 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7211 SCHED_CPUMASK_VAR(send_covered, allmasks);
7213 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7214 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7215 if (i != first_cpu(*this_sibling_map))
7218 init_sched_build_groups(this_sibling_map, cpu_map,
7220 send_covered, tmpmask);
7224 #ifdef CONFIG_SCHED_MC
7225 /* Set up multi-core groups */
7226 for_each_cpu_mask(i, *cpu_map) {
7227 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7228 SCHED_CPUMASK_VAR(send_covered, allmasks);
7230 *this_core_map = cpu_coregroup_map(i);
7231 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7232 if (i != first_cpu(*this_core_map))
7235 init_sched_build_groups(this_core_map, cpu_map,
7237 send_covered, tmpmask);
7241 /* Set up physical groups */
7242 for (i = 0; i < MAX_NUMNODES; i++) {
7243 SCHED_CPUMASK_VAR(nodemask, allmasks);
7244 SCHED_CPUMASK_VAR(send_covered, allmasks);
7246 *nodemask = node_to_cpumask(i);
7247 cpus_and(*nodemask, *nodemask, *cpu_map);
7248 if (cpus_empty(*nodemask))
7251 init_sched_build_groups(nodemask, cpu_map,
7253 send_covered, tmpmask);
7257 /* Set up node groups */
7259 SCHED_CPUMASK_VAR(send_covered, allmasks);
7261 init_sched_build_groups(cpu_map, cpu_map,
7262 &cpu_to_allnodes_group,
7263 send_covered, tmpmask);
7266 for (i = 0; i < MAX_NUMNODES; i++) {
7267 /* Set up node groups */
7268 struct sched_group *sg, *prev;
7269 SCHED_CPUMASK_VAR(nodemask, allmasks);
7270 SCHED_CPUMASK_VAR(domainspan, allmasks);
7271 SCHED_CPUMASK_VAR(covered, allmasks);
7274 *nodemask = node_to_cpumask(i);
7275 cpus_clear(*covered);
7277 cpus_and(*nodemask, *nodemask, *cpu_map);
7278 if (cpus_empty(*nodemask)) {
7279 sched_group_nodes[i] = NULL;
7283 sched_domain_node_span(i, domainspan);
7284 cpus_and(*domainspan, *domainspan, *cpu_map);
7286 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7288 printk(KERN_WARNING "Can not alloc domain group for "
7292 sched_group_nodes[i] = sg;
7293 for_each_cpu_mask(j, *nodemask) {
7294 struct sched_domain *sd;
7296 sd = &per_cpu(node_domains, j);
7299 sg->__cpu_power = 0;
7300 sg->cpumask = *nodemask;
7302 cpus_or(*covered, *covered, *nodemask);
7305 for (j = 0; j < MAX_NUMNODES; j++) {
7306 SCHED_CPUMASK_VAR(notcovered, allmasks);
7307 int n = (i + j) % MAX_NUMNODES;
7308 node_to_cpumask_ptr(pnodemask, n);
7310 cpus_complement(*notcovered, *covered);
7311 cpus_and(*tmpmask, *notcovered, *cpu_map);
7312 cpus_and(*tmpmask, *tmpmask, *domainspan);
7313 if (cpus_empty(*tmpmask))
7316 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7317 if (cpus_empty(*tmpmask))
7320 sg = kmalloc_node(sizeof(struct sched_group),
7324 "Can not alloc domain group for node %d\n", j);
7327 sg->__cpu_power = 0;
7328 sg->cpumask = *tmpmask;
7329 sg->next = prev->next;
7330 cpus_or(*covered, *covered, *tmpmask);
7337 /* Calculate CPU power for physical packages and nodes */
7338 #ifdef CONFIG_SCHED_SMT
7339 for_each_cpu_mask(i, *cpu_map) {
7340 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7342 init_sched_groups_power(i, sd);
7345 #ifdef CONFIG_SCHED_MC
7346 for_each_cpu_mask(i, *cpu_map) {
7347 struct sched_domain *sd = &per_cpu(core_domains, i);
7349 init_sched_groups_power(i, sd);
7353 for_each_cpu_mask(i, *cpu_map) {
7354 struct sched_domain *sd = &per_cpu(phys_domains, i);
7356 init_sched_groups_power(i, sd);
7360 for (i = 0; i < MAX_NUMNODES; i++)
7361 init_numa_sched_groups_power(sched_group_nodes[i]);
7364 struct sched_group *sg;
7366 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7368 init_numa_sched_groups_power(sg);
7372 /* Attach the domains */
7373 for_each_cpu_mask(i, *cpu_map) {
7374 struct sched_domain *sd;
7375 #ifdef CONFIG_SCHED_SMT
7376 sd = &per_cpu(cpu_domains, i);
7377 #elif defined(CONFIG_SCHED_MC)
7378 sd = &per_cpu(core_domains, i);
7380 sd = &per_cpu(phys_domains, i);
7382 cpu_attach_domain(sd, rd, i);
7385 SCHED_CPUMASK_FREE((void *)allmasks);
7390 free_sched_groups(cpu_map, tmpmask);
7391 SCHED_CPUMASK_FREE((void *)allmasks);
7396 static int build_sched_domains(const cpumask_t *cpu_map)
7398 return __build_sched_domains(cpu_map, NULL);
7401 static cpumask_t *doms_cur; /* current sched domains */
7402 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7403 static struct sched_domain_attr *dattr_cur;
7404 /* attribues of custom domains in 'doms_cur' */
7407 * Special case: If a kmalloc of a doms_cur partition (array of
7408 * cpumask_t) fails, then fallback to a single sched domain,
7409 * as determined by the single cpumask_t fallback_doms.
7411 static cpumask_t fallback_doms;
7413 void __attribute__((weak)) arch_update_cpu_topology(void)
7418 * Free current domain masks.
7419 * Called after all cpus are attached to NULL domain.
7421 static void free_sched_domains(void)
7424 if (doms_cur != &fallback_doms)
7426 doms_cur = &fallback_doms;
7430 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7431 * For now this just excludes isolated cpus, but could be used to
7432 * exclude other special cases in the future.
7434 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7438 arch_update_cpu_topology();
7440 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7442 doms_cur = &fallback_doms;
7443 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7445 err = build_sched_domains(doms_cur);
7446 register_sched_domain_sysctl();
7451 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7454 free_sched_groups(cpu_map, tmpmask);
7458 * Detach sched domains from a group of cpus specified in cpu_map
7459 * These cpus will now be attached to the NULL domain
7461 static void detach_destroy_domains(const cpumask_t *cpu_map)
7466 unregister_sched_domain_sysctl();
7468 for_each_cpu_mask(i, *cpu_map)
7469 cpu_attach_domain(NULL, &def_root_domain, i);
7470 synchronize_sched();
7471 arch_destroy_sched_domains(cpu_map, &tmpmask);
7474 /* handle null as "default" */
7475 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7476 struct sched_domain_attr *new, int idx_new)
7478 struct sched_domain_attr tmp;
7485 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7486 new ? (new + idx_new) : &tmp,
7487 sizeof(struct sched_domain_attr));
7491 * Partition sched domains as specified by the 'ndoms_new'
7492 * cpumasks in the array doms_new[] of cpumasks. This compares
7493 * doms_new[] to the current sched domain partitioning, doms_cur[].
7494 * It destroys each deleted domain and builds each new domain.
7496 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7497 * The masks don't intersect (don't overlap.) We should setup one
7498 * sched domain for each mask. CPUs not in any of the cpumasks will
7499 * not be load balanced. If the same cpumask appears both in the
7500 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7503 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7504 * ownership of it and will kfree it when done with it. If the caller
7505 * failed the kmalloc call, then it can pass in doms_new == NULL,
7506 * and partition_sched_domains() will fallback to the single partition
7509 * Call with hotplug lock held
7511 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7512 struct sched_domain_attr *dattr_new)
7516 mutex_lock(&sched_domains_mutex);
7518 /* always unregister in case we don't destroy any domains */
7519 unregister_sched_domain_sysctl();
7521 if (doms_new == NULL) {
7523 doms_new = &fallback_doms;
7524 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7528 /* Destroy deleted domains */
7529 for (i = 0; i < ndoms_cur; i++) {
7530 for (j = 0; j < ndoms_new; j++) {
7531 if (cpus_equal(doms_cur[i], doms_new[j])
7532 && dattrs_equal(dattr_cur, i, dattr_new, j))
7535 /* no match - a current sched domain not in new doms_new[] */
7536 detach_destroy_domains(doms_cur + i);
7541 /* Build new domains */
7542 for (i = 0; i < ndoms_new; i++) {
7543 for (j = 0; j < ndoms_cur; j++) {
7544 if (cpus_equal(doms_new[i], doms_cur[j])
7545 && dattrs_equal(dattr_new, i, dattr_cur, j))
7548 /* no match - add a new doms_new */
7549 __build_sched_domains(doms_new + i,
7550 dattr_new ? dattr_new + i : NULL);
7555 /* Remember the new sched domains */
7556 if (doms_cur != &fallback_doms)
7558 kfree(dattr_cur); /* kfree(NULL) is safe */
7559 doms_cur = doms_new;
7560 dattr_cur = dattr_new;
7561 ndoms_cur = ndoms_new;
7563 register_sched_domain_sysctl();
7565 mutex_unlock(&sched_domains_mutex);
7568 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7569 int arch_reinit_sched_domains(void)
7574 mutex_lock(&sched_domains_mutex);
7575 detach_destroy_domains(&cpu_online_map);
7576 free_sched_domains();
7577 err = arch_init_sched_domains(&cpu_online_map);
7578 mutex_unlock(&sched_domains_mutex);
7584 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7588 if (buf[0] != '0' && buf[0] != '1')
7592 sched_smt_power_savings = (buf[0] == '1');
7594 sched_mc_power_savings = (buf[0] == '1');
7596 ret = arch_reinit_sched_domains();
7598 return ret ? ret : count;
7601 #ifdef CONFIG_SCHED_MC
7602 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7604 return sprintf(page, "%u\n", sched_mc_power_savings);
7606 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7607 const char *buf, size_t count)
7609 return sched_power_savings_store(buf, count, 0);
7611 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7612 sched_mc_power_savings_store);
7615 #ifdef CONFIG_SCHED_SMT
7616 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7618 return sprintf(page, "%u\n", sched_smt_power_savings);
7620 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7621 const char *buf, size_t count)
7623 return sched_power_savings_store(buf, count, 1);
7625 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7626 sched_smt_power_savings_store);
7629 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7633 #ifdef CONFIG_SCHED_SMT
7635 err = sysfs_create_file(&cls->kset.kobj,
7636 &attr_sched_smt_power_savings.attr);
7638 #ifdef CONFIG_SCHED_MC
7639 if (!err && mc_capable())
7640 err = sysfs_create_file(&cls->kset.kobj,
7641 &attr_sched_mc_power_savings.attr);
7645 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7648 * Force a reinitialization of the sched domains hierarchy. The domains
7649 * and groups cannot be updated in place without racing with the balancing
7650 * code, so we temporarily attach all running cpus to the NULL domain
7651 * which will prevent rebalancing while the sched domains are recalculated.
7653 static int update_sched_domains(struct notifier_block *nfb,
7654 unsigned long action, void *hcpu)
7656 int cpu = (int)(long)hcpu;
7659 case CPU_DOWN_PREPARE:
7660 case CPU_DOWN_PREPARE_FROZEN:
7661 disable_runtime(cpu_rq(cpu));
7663 case CPU_UP_PREPARE:
7664 case CPU_UP_PREPARE_FROZEN:
7665 detach_destroy_domains(&cpu_online_map);
7666 free_sched_domains();
7670 case CPU_DOWN_FAILED:
7671 case CPU_DOWN_FAILED_FROZEN:
7673 case CPU_ONLINE_FROZEN:
7674 enable_runtime(cpu_rq(cpu));
7676 case CPU_UP_CANCELED:
7677 case CPU_UP_CANCELED_FROZEN:
7679 case CPU_DEAD_FROZEN:
7681 * Fall through and re-initialise the domains.
7688 #ifndef CONFIG_CPUSETS
7690 * Create default domain partitioning if cpusets are disabled.
7691 * Otherwise we let cpusets rebuild the domains based on the
7695 /* The hotplug lock is already held by cpu_up/cpu_down */
7696 arch_init_sched_domains(&cpu_online_map);
7702 void __init sched_init_smp(void)
7704 cpumask_t non_isolated_cpus;
7706 #if defined(CONFIG_NUMA)
7707 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7709 BUG_ON(sched_group_nodes_bycpu == NULL);
7712 mutex_lock(&sched_domains_mutex);
7713 arch_init_sched_domains(&cpu_online_map);
7714 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7715 if (cpus_empty(non_isolated_cpus))
7716 cpu_set(smp_processor_id(), non_isolated_cpus);
7717 mutex_unlock(&sched_domains_mutex);
7719 /* XXX: Theoretical race here - CPU may be hotplugged now */
7720 hotcpu_notifier(update_sched_domains, 0);
7723 /* Move init over to a non-isolated CPU */
7724 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7726 sched_init_granularity();
7729 void __init sched_init_smp(void)
7731 sched_init_granularity();
7733 #endif /* CONFIG_SMP */
7735 int in_sched_functions(unsigned long addr)
7737 return in_lock_functions(addr) ||
7738 (addr >= (unsigned long)__sched_text_start
7739 && addr < (unsigned long)__sched_text_end);
7742 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7744 cfs_rq->tasks_timeline = RB_ROOT;
7745 INIT_LIST_HEAD(&cfs_rq->tasks);
7746 #ifdef CONFIG_FAIR_GROUP_SCHED
7749 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7752 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7754 struct rt_prio_array *array;
7757 array = &rt_rq->active;
7758 for (i = 0; i < MAX_RT_PRIO; i++) {
7759 INIT_LIST_HEAD(array->queue + i);
7760 __clear_bit(i, array->bitmap);
7762 /* delimiter for bitsearch: */
7763 __set_bit(MAX_RT_PRIO, array->bitmap);
7765 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7766 rt_rq->highest_prio = MAX_RT_PRIO;
7769 rt_rq->rt_nr_migratory = 0;
7770 rt_rq->overloaded = 0;
7774 rt_rq->rt_throttled = 0;
7775 rt_rq->rt_runtime = 0;
7776 spin_lock_init(&rt_rq->rt_runtime_lock);
7778 #ifdef CONFIG_RT_GROUP_SCHED
7779 rt_rq->rt_nr_boosted = 0;
7784 #ifdef CONFIG_FAIR_GROUP_SCHED
7785 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7786 struct sched_entity *se, int cpu, int add,
7787 struct sched_entity *parent)
7789 struct rq *rq = cpu_rq(cpu);
7790 tg->cfs_rq[cpu] = cfs_rq;
7791 init_cfs_rq(cfs_rq, rq);
7794 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7797 /* se could be NULL for init_task_group */
7802 se->cfs_rq = &rq->cfs;
7804 se->cfs_rq = parent->my_q;
7807 se->load.weight = tg->shares;
7808 se->load.inv_weight = 0;
7809 se->parent = parent;
7813 #ifdef CONFIG_RT_GROUP_SCHED
7814 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7815 struct sched_rt_entity *rt_se, int cpu, int add,
7816 struct sched_rt_entity *parent)
7818 struct rq *rq = cpu_rq(cpu);
7820 tg->rt_rq[cpu] = rt_rq;
7821 init_rt_rq(rt_rq, rq);
7823 rt_rq->rt_se = rt_se;
7824 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7826 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7828 tg->rt_se[cpu] = rt_se;
7833 rt_se->rt_rq = &rq->rt;
7835 rt_se->rt_rq = parent->my_q;
7837 rt_se->my_q = rt_rq;
7838 rt_se->parent = parent;
7839 INIT_LIST_HEAD(&rt_se->run_list);
7843 void __init sched_init(void)
7846 unsigned long alloc_size = 0, ptr;
7848 #ifdef CONFIG_FAIR_GROUP_SCHED
7849 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7851 #ifdef CONFIG_RT_GROUP_SCHED
7852 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7854 #ifdef CONFIG_USER_SCHED
7858 * As sched_init() is called before page_alloc is setup,
7859 * we use alloc_bootmem().
7862 ptr = (unsigned long)alloc_bootmem(alloc_size);
7864 #ifdef CONFIG_FAIR_GROUP_SCHED
7865 init_task_group.se = (struct sched_entity **)ptr;
7866 ptr += nr_cpu_ids * sizeof(void **);
7868 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7869 ptr += nr_cpu_ids * sizeof(void **);
7871 #ifdef CONFIG_USER_SCHED
7872 root_task_group.se = (struct sched_entity **)ptr;
7873 ptr += nr_cpu_ids * sizeof(void **);
7875 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
7876 ptr += nr_cpu_ids * sizeof(void **);
7877 #endif /* CONFIG_USER_SCHED */
7878 #endif /* CONFIG_FAIR_GROUP_SCHED */
7879 #ifdef CONFIG_RT_GROUP_SCHED
7880 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7881 ptr += nr_cpu_ids * sizeof(void **);
7883 init_task_group.rt_rq = (struct rt_rq **)ptr;
7884 ptr += nr_cpu_ids * sizeof(void **);
7886 #ifdef CONFIG_USER_SCHED
7887 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
7888 ptr += nr_cpu_ids * sizeof(void **);
7890 root_task_group.rt_rq = (struct rt_rq **)ptr;
7891 ptr += nr_cpu_ids * sizeof(void **);
7892 #endif /* CONFIG_USER_SCHED */
7893 #endif /* CONFIG_RT_GROUP_SCHED */
7897 init_defrootdomain();
7900 init_rt_bandwidth(&def_rt_bandwidth,
7901 global_rt_period(), global_rt_runtime());
7903 #ifdef CONFIG_RT_GROUP_SCHED
7904 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7905 global_rt_period(), global_rt_runtime());
7906 #ifdef CONFIG_USER_SCHED
7907 init_rt_bandwidth(&root_task_group.rt_bandwidth,
7908 global_rt_period(), RUNTIME_INF);
7909 #endif /* CONFIG_USER_SCHED */
7910 #endif /* CONFIG_RT_GROUP_SCHED */
7912 #ifdef CONFIG_GROUP_SCHED
7913 list_add(&init_task_group.list, &task_groups);
7914 INIT_LIST_HEAD(&init_task_group.children);
7916 #ifdef CONFIG_USER_SCHED
7917 INIT_LIST_HEAD(&root_task_group.children);
7918 init_task_group.parent = &root_task_group;
7919 list_add(&init_task_group.siblings, &root_task_group.children);
7920 #endif /* CONFIG_USER_SCHED */
7921 #endif /* CONFIG_GROUP_SCHED */
7923 for_each_possible_cpu(i) {
7927 spin_lock_init(&rq->lock);
7928 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
7930 init_cfs_rq(&rq->cfs, rq);
7931 init_rt_rq(&rq->rt, rq);
7932 #ifdef CONFIG_FAIR_GROUP_SCHED
7933 init_task_group.shares = init_task_group_load;
7934 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7935 #ifdef CONFIG_CGROUP_SCHED
7937 * How much cpu bandwidth does init_task_group get?
7939 * In case of task-groups formed thr' the cgroup filesystem, it
7940 * gets 100% of the cpu resources in the system. This overall
7941 * system cpu resource is divided among the tasks of
7942 * init_task_group and its child task-groups in a fair manner,
7943 * based on each entity's (task or task-group's) weight
7944 * (se->load.weight).
7946 * In other words, if init_task_group has 10 tasks of weight
7947 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7948 * then A0's share of the cpu resource is:
7950 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7952 * We achieve this by letting init_task_group's tasks sit
7953 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7955 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7956 #elif defined CONFIG_USER_SCHED
7957 root_task_group.shares = NICE_0_LOAD;
7958 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
7960 * In case of task-groups formed thr' the user id of tasks,
7961 * init_task_group represents tasks belonging to root user.
7962 * Hence it forms a sibling of all subsequent groups formed.
7963 * In this case, init_task_group gets only a fraction of overall
7964 * system cpu resource, based on the weight assigned to root
7965 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7966 * by letting tasks of init_task_group sit in a separate cfs_rq
7967 * (init_cfs_rq) and having one entity represent this group of
7968 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7970 init_tg_cfs_entry(&init_task_group,
7971 &per_cpu(init_cfs_rq, i),
7972 &per_cpu(init_sched_entity, i), i, 1,
7973 root_task_group.se[i]);
7976 #endif /* CONFIG_FAIR_GROUP_SCHED */
7978 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7979 #ifdef CONFIG_RT_GROUP_SCHED
7980 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7981 #ifdef CONFIG_CGROUP_SCHED
7982 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7983 #elif defined CONFIG_USER_SCHED
7984 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
7985 init_tg_rt_entry(&init_task_group,
7986 &per_cpu(init_rt_rq, i),
7987 &per_cpu(init_sched_rt_entity, i), i, 1,
7988 root_task_group.rt_se[i]);
7992 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7993 rq->cpu_load[j] = 0;
7997 rq->active_balance = 0;
7998 rq->next_balance = jiffies;
8002 rq->migration_thread = NULL;
8003 INIT_LIST_HEAD(&rq->migration_queue);
8004 rq_attach_root(rq, &def_root_domain);
8007 atomic_set(&rq->nr_iowait, 0);
8010 set_load_weight(&init_task);
8012 #ifdef CONFIG_PREEMPT_NOTIFIERS
8013 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8017 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8020 #ifdef CONFIG_RT_MUTEXES
8021 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8025 * The boot idle thread does lazy MMU switching as well:
8027 atomic_inc(&init_mm.mm_count);
8028 enter_lazy_tlb(&init_mm, current);
8031 * Make us the idle thread. Technically, schedule() should not be
8032 * called from this thread, however somewhere below it might be,
8033 * but because we are the idle thread, we just pick up running again
8034 * when this runqueue becomes "idle".
8036 init_idle(current, smp_processor_id());
8038 * During early bootup we pretend to be a normal task:
8040 current->sched_class = &fair_sched_class;
8042 scheduler_running = 1;
8045 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8046 void __might_sleep(char *file, int line)
8049 static unsigned long prev_jiffy; /* ratelimiting */
8051 if ((in_atomic() || irqs_disabled()) &&
8052 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8053 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8055 prev_jiffy = jiffies;
8056 printk(KERN_ERR "BUG: sleeping function called from invalid"
8057 " context at %s:%d\n", file, line);
8058 printk("in_atomic():%d, irqs_disabled():%d\n",
8059 in_atomic(), irqs_disabled());
8060 debug_show_held_locks(current);
8061 if (irqs_disabled())
8062 print_irqtrace_events(current);
8067 EXPORT_SYMBOL(__might_sleep);
8070 #ifdef CONFIG_MAGIC_SYSRQ
8071 static void normalize_task(struct rq *rq, struct task_struct *p)
8075 update_rq_clock(rq);
8076 on_rq = p->se.on_rq;
8078 deactivate_task(rq, p, 0);
8079 __setscheduler(rq, p, SCHED_NORMAL, 0);
8081 activate_task(rq, p, 0);
8082 resched_task(rq->curr);
8086 void normalize_rt_tasks(void)
8088 struct task_struct *g, *p;
8089 unsigned long flags;
8092 read_lock_irqsave(&tasklist_lock, flags);
8093 do_each_thread(g, p) {
8095 * Only normalize user tasks:
8100 p->se.exec_start = 0;
8101 #ifdef CONFIG_SCHEDSTATS
8102 p->se.wait_start = 0;
8103 p->se.sleep_start = 0;
8104 p->se.block_start = 0;
8109 * Renice negative nice level userspace
8112 if (TASK_NICE(p) < 0 && p->mm)
8113 set_user_nice(p, 0);
8117 spin_lock(&p->pi_lock);
8118 rq = __task_rq_lock(p);
8120 normalize_task(rq, p);
8122 __task_rq_unlock(rq);
8123 spin_unlock(&p->pi_lock);
8124 } while_each_thread(g, p);
8126 read_unlock_irqrestore(&tasklist_lock, flags);
8129 #endif /* CONFIG_MAGIC_SYSRQ */
8133 * These functions are only useful for the IA64 MCA handling.
8135 * They can only be called when the whole system has been
8136 * stopped - every CPU needs to be quiescent, and no scheduling
8137 * activity can take place. Using them for anything else would
8138 * be a serious bug, and as a result, they aren't even visible
8139 * under any other configuration.
8143 * curr_task - return the current task for a given cpu.
8144 * @cpu: the processor in question.
8146 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8148 struct task_struct *curr_task(int cpu)
8150 return cpu_curr(cpu);
8154 * set_curr_task - set the current task for a given cpu.
8155 * @cpu: the processor in question.
8156 * @p: the task pointer to set.
8158 * Description: This function must only be used when non-maskable interrupts
8159 * are serviced on a separate stack. It allows the architecture to switch the
8160 * notion of the current task on a cpu in a non-blocking manner. This function
8161 * must be called with all CPU's synchronized, and interrupts disabled, the
8162 * and caller must save the original value of the current task (see
8163 * curr_task() above) and restore that value before reenabling interrupts and
8164 * re-starting the system.
8166 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8168 void set_curr_task(int cpu, struct task_struct *p)
8175 #ifdef CONFIG_FAIR_GROUP_SCHED
8176 static void free_fair_sched_group(struct task_group *tg)
8180 for_each_possible_cpu(i) {
8182 kfree(tg->cfs_rq[i]);
8192 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8194 struct cfs_rq *cfs_rq;
8195 struct sched_entity *se, *parent_se;
8199 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8202 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8206 tg->shares = NICE_0_LOAD;
8208 for_each_possible_cpu(i) {
8211 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8212 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8216 se = kmalloc_node(sizeof(struct sched_entity),
8217 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8221 parent_se = parent ? parent->se[i] : NULL;
8222 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8231 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8233 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8234 &cpu_rq(cpu)->leaf_cfs_rq_list);
8237 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8239 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8241 #else /* !CONFG_FAIR_GROUP_SCHED */
8242 static inline void free_fair_sched_group(struct task_group *tg)
8247 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8252 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8256 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8259 #endif /* CONFIG_FAIR_GROUP_SCHED */
8261 #ifdef CONFIG_RT_GROUP_SCHED
8262 static void free_rt_sched_group(struct task_group *tg)
8266 destroy_rt_bandwidth(&tg->rt_bandwidth);
8268 for_each_possible_cpu(i) {
8270 kfree(tg->rt_rq[i]);
8272 kfree(tg->rt_se[i]);
8280 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8282 struct rt_rq *rt_rq;
8283 struct sched_rt_entity *rt_se, *parent_se;
8287 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8290 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8294 init_rt_bandwidth(&tg->rt_bandwidth,
8295 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8297 for_each_possible_cpu(i) {
8300 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8301 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8305 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8306 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8310 parent_se = parent ? parent->rt_se[i] : NULL;
8311 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8320 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8322 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8323 &cpu_rq(cpu)->leaf_rt_rq_list);
8326 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8328 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8330 #else /* !CONFIG_RT_GROUP_SCHED */
8331 static inline void free_rt_sched_group(struct task_group *tg)
8336 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8341 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8345 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8348 #endif /* CONFIG_RT_GROUP_SCHED */
8350 #ifdef CONFIG_GROUP_SCHED
8351 static void free_sched_group(struct task_group *tg)
8353 free_fair_sched_group(tg);
8354 free_rt_sched_group(tg);
8358 /* allocate runqueue etc for a new task group */
8359 struct task_group *sched_create_group(struct task_group *parent)
8361 struct task_group *tg;
8362 unsigned long flags;
8365 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8367 return ERR_PTR(-ENOMEM);
8369 if (!alloc_fair_sched_group(tg, parent))
8372 if (!alloc_rt_sched_group(tg, parent))
8375 spin_lock_irqsave(&task_group_lock, flags);
8376 for_each_possible_cpu(i) {
8377 register_fair_sched_group(tg, i);
8378 register_rt_sched_group(tg, i);
8380 list_add_rcu(&tg->list, &task_groups);
8382 WARN_ON(!parent); /* root should already exist */
8384 tg->parent = parent;
8385 list_add_rcu(&tg->siblings, &parent->children);
8386 INIT_LIST_HEAD(&tg->children);
8387 spin_unlock_irqrestore(&task_group_lock, flags);
8392 free_sched_group(tg);
8393 return ERR_PTR(-ENOMEM);
8396 /* rcu callback to free various structures associated with a task group */
8397 static void free_sched_group_rcu(struct rcu_head *rhp)
8399 /* now it should be safe to free those cfs_rqs */
8400 free_sched_group(container_of(rhp, struct task_group, rcu));
8403 /* Destroy runqueue etc associated with a task group */
8404 void sched_destroy_group(struct task_group *tg)
8406 unsigned long flags;
8409 spin_lock_irqsave(&task_group_lock, flags);
8410 for_each_possible_cpu(i) {
8411 unregister_fair_sched_group(tg, i);
8412 unregister_rt_sched_group(tg, i);
8414 list_del_rcu(&tg->list);
8415 list_del_rcu(&tg->siblings);
8416 spin_unlock_irqrestore(&task_group_lock, flags);
8418 /* wait for possible concurrent references to cfs_rqs complete */
8419 call_rcu(&tg->rcu, free_sched_group_rcu);
8422 /* change task's runqueue when it moves between groups.
8423 * The caller of this function should have put the task in its new group
8424 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8425 * reflect its new group.
8427 void sched_move_task(struct task_struct *tsk)
8430 unsigned long flags;
8433 rq = task_rq_lock(tsk, &flags);
8435 update_rq_clock(rq);
8437 running = task_current(rq, tsk);
8438 on_rq = tsk->se.on_rq;
8441 dequeue_task(rq, tsk, 0);
8442 if (unlikely(running))
8443 tsk->sched_class->put_prev_task(rq, tsk);
8445 set_task_rq(tsk, task_cpu(tsk));
8447 #ifdef CONFIG_FAIR_GROUP_SCHED
8448 if (tsk->sched_class->moved_group)
8449 tsk->sched_class->moved_group(tsk);
8452 if (unlikely(running))
8453 tsk->sched_class->set_curr_task(rq);
8455 enqueue_task(rq, tsk, 0);
8457 task_rq_unlock(rq, &flags);
8459 #endif /* CONFIG_GROUP_SCHED */
8461 #ifdef CONFIG_FAIR_GROUP_SCHED
8462 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8464 struct cfs_rq *cfs_rq = se->cfs_rq;
8469 dequeue_entity(cfs_rq, se, 0);
8471 se->load.weight = shares;
8472 se->load.inv_weight = 0;
8475 enqueue_entity(cfs_rq, se, 0);
8478 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8480 struct cfs_rq *cfs_rq = se->cfs_rq;
8481 struct rq *rq = cfs_rq->rq;
8482 unsigned long flags;
8484 spin_lock_irqsave(&rq->lock, flags);
8485 __set_se_shares(se, shares);
8486 spin_unlock_irqrestore(&rq->lock, flags);
8489 static DEFINE_MUTEX(shares_mutex);
8491 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8494 unsigned long flags;
8497 * We can't change the weight of the root cgroup.
8502 if (shares < MIN_SHARES)
8503 shares = MIN_SHARES;
8504 else if (shares > MAX_SHARES)
8505 shares = MAX_SHARES;
8507 mutex_lock(&shares_mutex);
8508 if (tg->shares == shares)
8511 spin_lock_irqsave(&task_group_lock, flags);
8512 for_each_possible_cpu(i)
8513 unregister_fair_sched_group(tg, i);
8514 list_del_rcu(&tg->siblings);
8515 spin_unlock_irqrestore(&task_group_lock, flags);
8517 /* wait for any ongoing reference to this group to finish */
8518 synchronize_sched();
8521 * Now we are free to modify the group's share on each cpu
8522 * w/o tripping rebalance_share or load_balance_fair.
8524 tg->shares = shares;
8525 for_each_possible_cpu(i) {
8529 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8530 set_se_shares(tg->se[i], shares);
8534 * Enable load balance activity on this group, by inserting it back on
8535 * each cpu's rq->leaf_cfs_rq_list.
8537 spin_lock_irqsave(&task_group_lock, flags);
8538 for_each_possible_cpu(i)
8539 register_fair_sched_group(tg, i);
8540 list_add_rcu(&tg->siblings, &tg->parent->children);
8541 spin_unlock_irqrestore(&task_group_lock, flags);
8543 mutex_unlock(&shares_mutex);
8547 unsigned long sched_group_shares(struct task_group *tg)
8553 #ifdef CONFIG_RT_GROUP_SCHED
8555 * Ensure that the real time constraints are schedulable.
8557 static DEFINE_MUTEX(rt_constraints_mutex);
8559 static unsigned long to_ratio(u64 period, u64 runtime)
8561 if (runtime == RUNTIME_INF)
8564 return div64_u64(runtime << 16, period);
8567 #ifdef CONFIG_CGROUP_SCHED
8568 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8570 struct task_group *tgi, *parent = tg->parent;
8571 unsigned long total = 0;
8574 if (global_rt_period() < period)
8577 return to_ratio(period, runtime) <
8578 to_ratio(global_rt_period(), global_rt_runtime());
8581 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8585 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8589 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8590 tgi->rt_bandwidth.rt_runtime);
8594 return total + to_ratio(period, runtime) <=
8595 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8596 parent->rt_bandwidth.rt_runtime);
8598 #elif defined CONFIG_USER_SCHED
8599 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8601 struct task_group *tgi;
8602 unsigned long total = 0;
8603 unsigned long global_ratio =
8604 to_ratio(global_rt_period(), global_rt_runtime());
8607 list_for_each_entry_rcu(tgi, &task_groups, list) {
8611 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8612 tgi->rt_bandwidth.rt_runtime);
8616 return total + to_ratio(period, runtime) < global_ratio;
8620 /* Must be called with tasklist_lock held */
8621 static inline int tg_has_rt_tasks(struct task_group *tg)
8623 struct task_struct *g, *p;
8624 do_each_thread(g, p) {
8625 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8627 } while_each_thread(g, p);
8631 static int tg_set_bandwidth(struct task_group *tg,
8632 u64 rt_period, u64 rt_runtime)
8636 mutex_lock(&rt_constraints_mutex);
8637 read_lock(&tasklist_lock);
8638 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8642 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8647 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8648 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8649 tg->rt_bandwidth.rt_runtime = rt_runtime;
8651 for_each_possible_cpu(i) {
8652 struct rt_rq *rt_rq = tg->rt_rq[i];
8654 spin_lock(&rt_rq->rt_runtime_lock);
8655 rt_rq->rt_runtime = rt_runtime;
8656 spin_unlock(&rt_rq->rt_runtime_lock);
8658 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8660 read_unlock(&tasklist_lock);
8661 mutex_unlock(&rt_constraints_mutex);
8666 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8668 u64 rt_runtime, rt_period;
8670 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8671 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8672 if (rt_runtime_us < 0)
8673 rt_runtime = RUNTIME_INF;
8675 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8678 long sched_group_rt_runtime(struct task_group *tg)
8682 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8685 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8686 do_div(rt_runtime_us, NSEC_PER_USEC);
8687 return rt_runtime_us;
8690 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8692 u64 rt_runtime, rt_period;
8694 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8695 rt_runtime = tg->rt_bandwidth.rt_runtime;
8697 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8700 long sched_group_rt_period(struct task_group *tg)
8704 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8705 do_div(rt_period_us, NSEC_PER_USEC);
8706 return rt_period_us;
8709 static int sched_rt_global_constraints(void)
8711 struct task_group *tg = &root_task_group;
8712 u64 rt_runtime, rt_period;
8715 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8716 rt_runtime = tg->rt_bandwidth.rt_runtime;
8718 mutex_lock(&rt_constraints_mutex);
8719 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8721 mutex_unlock(&rt_constraints_mutex);
8725 #else /* !CONFIG_RT_GROUP_SCHED */
8726 static int sched_rt_global_constraints(void)
8728 unsigned long flags;
8731 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8732 for_each_possible_cpu(i) {
8733 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8735 spin_lock(&rt_rq->rt_runtime_lock);
8736 rt_rq->rt_runtime = global_rt_runtime();
8737 spin_unlock(&rt_rq->rt_runtime_lock);
8739 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8743 #endif /* CONFIG_RT_GROUP_SCHED */
8745 int sched_rt_handler(struct ctl_table *table, int write,
8746 struct file *filp, void __user *buffer, size_t *lenp,
8750 int old_period, old_runtime;
8751 static DEFINE_MUTEX(mutex);
8754 old_period = sysctl_sched_rt_period;
8755 old_runtime = sysctl_sched_rt_runtime;
8757 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8759 if (!ret && write) {
8760 ret = sched_rt_global_constraints();
8762 sysctl_sched_rt_period = old_period;
8763 sysctl_sched_rt_runtime = old_runtime;
8765 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8766 def_rt_bandwidth.rt_period =
8767 ns_to_ktime(global_rt_period());
8770 mutex_unlock(&mutex);
8775 #ifdef CONFIG_CGROUP_SCHED
8777 /* return corresponding task_group object of a cgroup */
8778 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8780 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8781 struct task_group, css);
8784 static struct cgroup_subsys_state *
8785 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8787 struct task_group *tg, *parent;
8789 if (!cgrp->parent) {
8790 /* This is early initialization for the top cgroup */
8791 init_task_group.css.cgroup = cgrp;
8792 return &init_task_group.css;
8795 parent = cgroup_tg(cgrp->parent);
8796 tg = sched_create_group(parent);
8798 return ERR_PTR(-ENOMEM);
8800 /* Bind the cgroup to task_group object we just created */
8801 tg->css.cgroup = cgrp;
8807 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8809 struct task_group *tg = cgroup_tg(cgrp);
8811 sched_destroy_group(tg);
8815 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8816 struct task_struct *tsk)
8818 #ifdef CONFIG_RT_GROUP_SCHED
8819 /* Don't accept realtime tasks when there is no way for them to run */
8820 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8823 /* We don't support RT-tasks being in separate groups */
8824 if (tsk->sched_class != &fair_sched_class)
8832 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8833 struct cgroup *old_cont, struct task_struct *tsk)
8835 sched_move_task(tsk);
8838 #ifdef CONFIG_FAIR_GROUP_SCHED
8839 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8842 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8845 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8847 struct task_group *tg = cgroup_tg(cgrp);
8849 return (u64) tg->shares;
8851 #endif /* CONFIG_FAIR_GROUP_SCHED */
8853 #ifdef CONFIG_RT_GROUP_SCHED
8854 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8857 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8860 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8862 return sched_group_rt_runtime(cgroup_tg(cgrp));
8865 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8868 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8871 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8873 return sched_group_rt_period(cgroup_tg(cgrp));
8875 #endif /* CONFIG_RT_GROUP_SCHED */
8877 static struct cftype cpu_files[] = {
8878 #ifdef CONFIG_FAIR_GROUP_SCHED
8881 .read_u64 = cpu_shares_read_u64,
8882 .write_u64 = cpu_shares_write_u64,
8885 #ifdef CONFIG_RT_GROUP_SCHED
8887 .name = "rt_runtime_us",
8888 .read_s64 = cpu_rt_runtime_read,
8889 .write_s64 = cpu_rt_runtime_write,
8892 .name = "rt_period_us",
8893 .read_u64 = cpu_rt_period_read_uint,
8894 .write_u64 = cpu_rt_period_write_uint,
8899 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8901 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8904 struct cgroup_subsys cpu_cgroup_subsys = {
8906 .create = cpu_cgroup_create,
8907 .destroy = cpu_cgroup_destroy,
8908 .can_attach = cpu_cgroup_can_attach,
8909 .attach = cpu_cgroup_attach,
8910 .populate = cpu_cgroup_populate,
8911 .subsys_id = cpu_cgroup_subsys_id,
8915 #endif /* CONFIG_CGROUP_SCHED */
8917 #ifdef CONFIG_CGROUP_CPUACCT
8920 * CPU accounting code for task groups.
8922 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8923 * (balbir@in.ibm.com).
8926 /* track cpu usage of a group of tasks */
8928 struct cgroup_subsys_state css;
8929 /* cpuusage holds pointer to a u64-type object on every cpu */
8933 struct cgroup_subsys cpuacct_subsys;
8935 /* return cpu accounting group corresponding to this container */
8936 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8938 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8939 struct cpuacct, css);
8942 /* return cpu accounting group to which this task belongs */
8943 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8945 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8946 struct cpuacct, css);
8949 /* create a new cpu accounting group */
8950 static struct cgroup_subsys_state *cpuacct_create(
8951 struct cgroup_subsys *ss, struct cgroup *cgrp)
8953 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8956 return ERR_PTR(-ENOMEM);
8958 ca->cpuusage = alloc_percpu(u64);
8959 if (!ca->cpuusage) {
8961 return ERR_PTR(-ENOMEM);
8967 /* destroy an existing cpu accounting group */
8969 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8971 struct cpuacct *ca = cgroup_ca(cgrp);
8973 free_percpu(ca->cpuusage);
8977 /* return total cpu usage (in nanoseconds) of a group */
8978 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8980 struct cpuacct *ca = cgroup_ca(cgrp);
8981 u64 totalcpuusage = 0;
8984 for_each_possible_cpu(i) {
8985 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
8988 * Take rq->lock to make 64-bit addition safe on 32-bit
8991 spin_lock_irq(&cpu_rq(i)->lock);
8992 totalcpuusage += *cpuusage;
8993 spin_unlock_irq(&cpu_rq(i)->lock);
8996 return totalcpuusage;
8999 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9002 struct cpuacct *ca = cgroup_ca(cgrp);
9011 for_each_possible_cpu(i) {
9012 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9014 spin_lock_irq(&cpu_rq(i)->lock);
9016 spin_unlock_irq(&cpu_rq(i)->lock);
9022 static struct cftype files[] = {
9025 .read_u64 = cpuusage_read,
9026 .write_u64 = cpuusage_write,
9030 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9032 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9036 * charge this task's execution time to its accounting group.
9038 * called with rq->lock held.
9040 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9044 if (!cpuacct_subsys.active)
9049 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9051 *cpuusage += cputime;
9055 struct cgroup_subsys cpuacct_subsys = {
9057 .create = cpuacct_create,
9058 .destroy = cpuacct_destroy,
9059 .populate = cpuacct_populate,
9060 .subsys_id = cpuacct_subsys_id,
9062 #endif /* CONFIG_CGROUP_CPUACCT */