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 */
409 unsigned long task_weight;
410 unsigned long shares;
412 * We need space to build a sched_domain wide view of the full task
413 * group tree, in order to avoid depending on dynamic memory allocation
414 * during the load balancing we place this in the per cpu task group
415 * hierarchy. This limits the load balancing to one instance per cpu,
416 * but more should not be needed anyway.
418 struct aggregate_struct {
420 * load = weight(cpus) * f(tg)
422 * Where f(tg) is the recursive weight fraction assigned to
428 * part of the group weight distributed to this span.
430 unsigned long shares;
433 * The sum of all runqueue weights within this span.
435 unsigned long rq_weight;
438 * Weight contributed by tasks; this is the part we can
439 * influence by moving tasks around.
441 unsigned long task_weight;
447 /* Real-Time classes' related field in a runqueue: */
449 struct rt_prio_array active;
450 unsigned long rt_nr_running;
451 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
452 int highest_prio; /* highest queued rt task prio */
455 unsigned long rt_nr_migratory;
461 /* Nests inside the rq lock: */
462 spinlock_t rt_runtime_lock;
464 #ifdef CONFIG_RT_GROUP_SCHED
465 unsigned long rt_nr_boosted;
468 struct list_head leaf_rt_rq_list;
469 struct task_group *tg;
470 struct sched_rt_entity *rt_se;
477 * We add the notion of a root-domain which will be used to define per-domain
478 * variables. Each exclusive cpuset essentially defines an island domain by
479 * fully partitioning the member cpus from any other cpuset. Whenever a new
480 * exclusive cpuset is created, we also create and attach a new root-domain
490 * The "RT overload" flag: it gets set if a CPU has more than
491 * one runnable RT task.
496 struct cpupri cpupri;
501 * By default the system creates a single root-domain with all cpus as
502 * members (mimicking the global state we have today).
504 static struct root_domain def_root_domain;
509 * This is the main, per-CPU runqueue data structure.
511 * Locking rule: those places that want to lock multiple runqueues
512 * (such as the load balancing or the thread migration code), lock
513 * acquire operations must be ordered by ascending &runqueue.
520 * nr_running and cpu_load should be in the same cacheline because
521 * remote CPUs use both these fields when doing load calculation.
523 unsigned long nr_running;
524 #define CPU_LOAD_IDX_MAX 5
525 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
526 unsigned char idle_at_tick;
528 unsigned long last_tick_seen;
529 unsigned char in_nohz_recently;
531 /* capture load from *all* tasks on this cpu: */
532 struct load_weight load;
533 unsigned long nr_load_updates;
539 #ifdef CONFIG_FAIR_GROUP_SCHED
540 /* list of leaf cfs_rq on this cpu: */
541 struct list_head leaf_cfs_rq_list;
543 #ifdef CONFIG_RT_GROUP_SCHED
544 struct list_head leaf_rt_rq_list;
548 * This is part of a global counter where only the total sum
549 * over all CPUs matters. A task can increase this counter on
550 * one CPU and if it got migrated afterwards it may decrease
551 * it on another CPU. Always updated under the runqueue lock:
553 unsigned long nr_uninterruptible;
555 struct task_struct *curr, *idle;
556 unsigned long next_balance;
557 struct mm_struct *prev_mm;
564 struct root_domain *rd;
565 struct sched_domain *sd;
567 /* For active balancing */
570 /* cpu of this runqueue: */
574 struct task_struct *migration_thread;
575 struct list_head migration_queue;
578 #ifdef CONFIG_SCHED_HRTICK
579 unsigned long hrtick_flags;
580 ktime_t hrtick_expire;
581 struct hrtimer hrtick_timer;
584 #ifdef CONFIG_SCHEDSTATS
586 struct sched_info rq_sched_info;
588 /* sys_sched_yield() stats */
589 unsigned int yld_exp_empty;
590 unsigned int yld_act_empty;
591 unsigned int yld_both_empty;
592 unsigned int yld_count;
594 /* schedule() stats */
595 unsigned int sched_switch;
596 unsigned int sched_count;
597 unsigned int sched_goidle;
599 /* try_to_wake_up() stats */
600 unsigned int ttwu_count;
601 unsigned int ttwu_local;
604 unsigned int bkl_count;
606 struct lock_class_key rq_lock_key;
609 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
611 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
613 rq->curr->sched_class->check_preempt_curr(rq, p);
616 static inline int cpu_of(struct rq *rq)
626 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
627 * See detach_destroy_domains: synchronize_sched for details.
629 * The domain tree of any CPU may only be accessed from within
630 * preempt-disabled sections.
632 #define for_each_domain(cpu, __sd) \
633 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
635 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
636 #define this_rq() (&__get_cpu_var(runqueues))
637 #define task_rq(p) cpu_rq(task_cpu(p))
638 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
640 static inline void update_rq_clock(struct rq *rq)
642 rq->clock = sched_clock_cpu(cpu_of(rq));
646 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
648 #ifdef CONFIG_SCHED_DEBUG
649 # define const_debug __read_mostly
651 # define const_debug static const
655 * Debugging: various feature bits
658 #define SCHED_FEAT(name, enabled) \
659 __SCHED_FEAT_##name ,
662 #include "sched_features.h"
667 #define SCHED_FEAT(name, enabled) \
668 (1UL << __SCHED_FEAT_##name) * enabled |
670 const_debug unsigned int sysctl_sched_features =
671 #include "sched_features.h"
676 #ifdef CONFIG_SCHED_DEBUG
677 #define SCHED_FEAT(name, enabled) \
680 static __read_mostly char *sched_feat_names[] = {
681 #include "sched_features.h"
687 static int sched_feat_open(struct inode *inode, struct file *filp)
689 filp->private_data = inode->i_private;
694 sched_feat_read(struct file *filp, char __user *ubuf,
695 size_t cnt, loff_t *ppos)
702 for (i = 0; sched_feat_names[i]; i++) {
703 len += strlen(sched_feat_names[i]);
707 buf = kmalloc(len + 2, GFP_KERNEL);
711 for (i = 0; sched_feat_names[i]; i++) {
712 if (sysctl_sched_features & (1UL << i))
713 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
715 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
718 r += sprintf(buf + r, "\n");
719 WARN_ON(r >= len + 2);
721 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
729 sched_feat_write(struct file *filp, const char __user *ubuf,
730 size_t cnt, loff_t *ppos)
740 if (copy_from_user(&buf, ubuf, cnt))
745 if (strncmp(buf, "NO_", 3) == 0) {
750 for (i = 0; sched_feat_names[i]; i++) {
751 int len = strlen(sched_feat_names[i]);
753 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
755 sysctl_sched_features &= ~(1UL << i);
757 sysctl_sched_features |= (1UL << i);
762 if (!sched_feat_names[i])
770 static struct file_operations sched_feat_fops = {
771 .open = sched_feat_open,
772 .read = sched_feat_read,
773 .write = sched_feat_write,
776 static __init int sched_init_debug(void)
778 debugfs_create_file("sched_features", 0644, NULL, NULL,
783 late_initcall(sched_init_debug);
787 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
790 * Number of tasks to iterate in a single balance run.
791 * Limited because this is done with IRQs disabled.
793 const_debug unsigned int sysctl_sched_nr_migrate = 32;
796 * period over which we measure -rt task cpu usage in us.
799 unsigned int sysctl_sched_rt_period = 1000000;
801 static __read_mostly int scheduler_running;
804 * part of the period that we allow rt tasks to run in us.
807 int sysctl_sched_rt_runtime = 950000;
809 static inline u64 global_rt_period(void)
811 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
814 static inline u64 global_rt_runtime(void)
816 if (sysctl_sched_rt_period < 0)
819 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
822 #ifndef prepare_arch_switch
823 # define prepare_arch_switch(next) do { } while (0)
825 #ifndef finish_arch_switch
826 # define finish_arch_switch(prev) do { } while (0)
829 static inline int task_current(struct rq *rq, struct task_struct *p)
831 return rq->curr == p;
834 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
835 static inline int task_running(struct rq *rq, struct task_struct *p)
837 return task_current(rq, p);
840 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
844 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
846 #ifdef CONFIG_DEBUG_SPINLOCK
847 /* this is a valid case when another task releases the spinlock */
848 rq->lock.owner = current;
851 * If we are tracking spinlock dependencies then we have to
852 * fix up the runqueue lock - which gets 'carried over' from
855 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
857 spin_unlock_irq(&rq->lock);
860 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
861 static inline int task_running(struct rq *rq, struct task_struct *p)
866 return task_current(rq, p);
870 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
874 * We can optimise this out completely for !SMP, because the
875 * SMP rebalancing from interrupt is the only thing that cares
880 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
881 spin_unlock_irq(&rq->lock);
883 spin_unlock(&rq->lock);
887 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
891 * After ->oncpu is cleared, the task can be moved to a different CPU.
892 * We must ensure this doesn't happen until the switch is completely
898 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
902 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
905 * __task_rq_lock - lock the runqueue a given task resides on.
906 * Must be called interrupts disabled.
908 static inline struct rq *__task_rq_lock(struct task_struct *p)
912 struct rq *rq = task_rq(p);
913 spin_lock(&rq->lock);
914 if (likely(rq == task_rq(p)))
916 spin_unlock(&rq->lock);
921 * task_rq_lock - lock the runqueue a given task resides on and disable
922 * interrupts. Note the ordering: we can safely lookup the task_rq without
923 * explicitly disabling preemption.
925 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
931 local_irq_save(*flags);
933 spin_lock(&rq->lock);
934 if (likely(rq == task_rq(p)))
936 spin_unlock_irqrestore(&rq->lock, *flags);
940 static void __task_rq_unlock(struct rq *rq)
943 spin_unlock(&rq->lock);
946 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
949 spin_unlock_irqrestore(&rq->lock, *flags);
953 * this_rq_lock - lock this runqueue and disable interrupts.
955 static struct rq *this_rq_lock(void)
962 spin_lock(&rq->lock);
967 static void __resched_task(struct task_struct *p, int tif_bit);
969 static inline void resched_task(struct task_struct *p)
971 __resched_task(p, TIF_NEED_RESCHED);
974 #ifdef CONFIG_SCHED_HRTICK
976 * Use HR-timers to deliver accurate preemption points.
978 * Its all a bit involved since we cannot program an hrt while holding the
979 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
982 * When we get rescheduled we reprogram the hrtick_timer outside of the
985 static inline void resched_hrt(struct task_struct *p)
987 __resched_task(p, TIF_HRTICK_RESCHED);
990 static inline void resched_rq(struct rq *rq)
994 spin_lock_irqsave(&rq->lock, flags);
995 resched_task(rq->curr);
996 spin_unlock_irqrestore(&rq->lock, flags);
1000 HRTICK_SET, /* re-programm hrtick_timer */
1001 HRTICK_RESET, /* not a new slice */
1002 HRTICK_BLOCK, /* stop hrtick operations */
1007 * - enabled by features
1008 * - hrtimer is actually high res
1010 static inline int hrtick_enabled(struct rq *rq)
1012 if (!sched_feat(HRTICK))
1014 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1016 return hrtimer_is_hres_active(&rq->hrtick_timer);
1020 * Called to set the hrtick timer state.
1022 * called with rq->lock held and irqs disabled
1024 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1026 assert_spin_locked(&rq->lock);
1029 * preempt at: now + delay
1032 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1034 * indicate we need to program the timer
1036 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1038 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1041 * New slices are called from the schedule path and don't need a
1042 * forced reschedule.
1045 resched_hrt(rq->curr);
1048 static void hrtick_clear(struct rq *rq)
1050 if (hrtimer_active(&rq->hrtick_timer))
1051 hrtimer_cancel(&rq->hrtick_timer);
1055 * Update the timer from the possible pending state.
1057 static void hrtick_set(struct rq *rq)
1061 unsigned long flags;
1063 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1065 spin_lock_irqsave(&rq->lock, flags);
1066 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1067 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1068 time = rq->hrtick_expire;
1069 clear_thread_flag(TIF_HRTICK_RESCHED);
1070 spin_unlock_irqrestore(&rq->lock, flags);
1073 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1074 if (reset && !hrtimer_active(&rq->hrtick_timer))
1081 * High-resolution timer tick.
1082 * Runs from hardirq context with interrupts disabled.
1084 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1086 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1088 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1090 spin_lock(&rq->lock);
1091 update_rq_clock(rq);
1092 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1093 spin_unlock(&rq->lock);
1095 return HRTIMER_NORESTART;
1099 static void hotplug_hrtick_disable(int cpu)
1101 struct rq *rq = cpu_rq(cpu);
1102 unsigned long flags;
1104 spin_lock_irqsave(&rq->lock, flags);
1105 rq->hrtick_flags = 0;
1106 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1107 spin_unlock_irqrestore(&rq->lock, flags);
1112 static void hotplug_hrtick_enable(int cpu)
1114 struct rq *rq = cpu_rq(cpu);
1115 unsigned long flags;
1117 spin_lock_irqsave(&rq->lock, flags);
1118 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1119 spin_unlock_irqrestore(&rq->lock, flags);
1123 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1125 int cpu = (int)(long)hcpu;
1128 case CPU_UP_CANCELED:
1129 case CPU_UP_CANCELED_FROZEN:
1130 case CPU_DOWN_PREPARE:
1131 case CPU_DOWN_PREPARE_FROZEN:
1133 case CPU_DEAD_FROZEN:
1134 hotplug_hrtick_disable(cpu);
1137 case CPU_UP_PREPARE:
1138 case CPU_UP_PREPARE_FROZEN:
1139 case CPU_DOWN_FAILED:
1140 case CPU_DOWN_FAILED_FROZEN:
1142 case CPU_ONLINE_FROZEN:
1143 hotplug_hrtick_enable(cpu);
1150 static void init_hrtick(void)
1152 hotcpu_notifier(hotplug_hrtick, 0);
1154 #endif /* CONFIG_SMP */
1156 static void init_rq_hrtick(struct rq *rq)
1158 rq->hrtick_flags = 0;
1159 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1160 rq->hrtick_timer.function = hrtick;
1161 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1164 void hrtick_resched(void)
1167 unsigned long flags;
1169 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1172 local_irq_save(flags);
1173 rq = cpu_rq(smp_processor_id());
1175 local_irq_restore(flags);
1178 static inline void hrtick_clear(struct rq *rq)
1182 static inline void hrtick_set(struct rq *rq)
1186 static inline void init_rq_hrtick(struct rq *rq)
1190 void hrtick_resched(void)
1194 static inline void init_hrtick(void)
1200 * resched_task - mark a task 'to be rescheduled now'.
1202 * On UP this means the setting of the need_resched flag, on SMP it
1203 * might also involve a cross-CPU call to trigger the scheduler on
1208 #ifndef tsk_is_polling
1209 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1212 static void __resched_task(struct task_struct *p, int tif_bit)
1216 assert_spin_locked(&task_rq(p)->lock);
1218 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1221 set_tsk_thread_flag(p, tif_bit);
1224 if (cpu == smp_processor_id())
1227 /* NEED_RESCHED must be visible before we test polling */
1229 if (!tsk_is_polling(p))
1230 smp_send_reschedule(cpu);
1233 static void resched_cpu(int cpu)
1235 struct rq *rq = cpu_rq(cpu);
1236 unsigned long flags;
1238 if (!spin_trylock_irqsave(&rq->lock, flags))
1240 resched_task(cpu_curr(cpu));
1241 spin_unlock_irqrestore(&rq->lock, flags);
1246 * When add_timer_on() enqueues a timer into the timer wheel of an
1247 * idle CPU then this timer might expire before the next timer event
1248 * which is scheduled to wake up that CPU. In case of a completely
1249 * idle system the next event might even be infinite time into the
1250 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1251 * leaves the inner idle loop so the newly added timer is taken into
1252 * account when the CPU goes back to idle and evaluates the timer
1253 * wheel for the next timer event.
1255 void wake_up_idle_cpu(int cpu)
1257 struct rq *rq = cpu_rq(cpu);
1259 if (cpu == smp_processor_id())
1263 * This is safe, as this function is called with the timer
1264 * wheel base lock of (cpu) held. When the CPU is on the way
1265 * to idle and has not yet set rq->curr to idle then it will
1266 * be serialized on the timer wheel base lock and take the new
1267 * timer into account automatically.
1269 if (rq->curr != rq->idle)
1273 * We can set TIF_RESCHED on the idle task of the other CPU
1274 * lockless. The worst case is that the other CPU runs the
1275 * idle task through an additional NOOP schedule()
1277 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1279 /* NEED_RESCHED must be visible before we test polling */
1281 if (!tsk_is_polling(rq->idle))
1282 smp_send_reschedule(cpu);
1284 #endif /* CONFIG_NO_HZ */
1286 #else /* !CONFIG_SMP */
1287 static void __resched_task(struct task_struct *p, int tif_bit)
1289 assert_spin_locked(&task_rq(p)->lock);
1290 set_tsk_thread_flag(p, tif_bit);
1292 #endif /* CONFIG_SMP */
1294 #if BITS_PER_LONG == 32
1295 # define WMULT_CONST (~0UL)
1297 # define WMULT_CONST (1UL << 32)
1300 #define WMULT_SHIFT 32
1303 * Shift right and round:
1305 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1308 * delta *= weight / lw
1310 static unsigned long
1311 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1312 struct load_weight *lw)
1316 if (!lw->inv_weight) {
1317 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1320 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1324 tmp = (u64)delta_exec * weight;
1326 * Check whether we'd overflow the 64-bit multiplication:
1328 if (unlikely(tmp > WMULT_CONST))
1329 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1332 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1334 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1337 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1343 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1350 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1351 * of tasks with abnormal "nice" values across CPUs the contribution that
1352 * each task makes to its run queue's load is weighted according to its
1353 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1354 * scaled version of the new time slice allocation that they receive on time
1358 #define WEIGHT_IDLEPRIO 2
1359 #define WMULT_IDLEPRIO (1 << 31)
1362 * Nice levels are multiplicative, with a gentle 10% change for every
1363 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1364 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1365 * that remained on nice 0.
1367 * The "10% effect" is relative and cumulative: from _any_ nice level,
1368 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1369 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1370 * If a task goes up by ~10% and another task goes down by ~10% then
1371 * the relative distance between them is ~25%.)
1373 static const int prio_to_weight[40] = {
1374 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1375 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1376 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1377 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1378 /* 0 */ 1024, 820, 655, 526, 423,
1379 /* 5 */ 335, 272, 215, 172, 137,
1380 /* 10 */ 110, 87, 70, 56, 45,
1381 /* 15 */ 36, 29, 23, 18, 15,
1385 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1387 * In cases where the weight does not change often, we can use the
1388 * precalculated inverse to speed up arithmetics by turning divisions
1389 * into multiplications:
1391 static const u32 prio_to_wmult[40] = {
1392 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1393 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1394 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1395 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1396 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1397 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1398 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1399 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1402 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1405 * runqueue iterator, to support SMP load-balancing between different
1406 * scheduling classes, without having to expose their internal data
1407 * structures to the load-balancing proper:
1409 struct rq_iterator {
1411 struct task_struct *(*start)(void *);
1412 struct task_struct *(*next)(void *);
1416 static unsigned long
1417 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1418 unsigned long max_load_move, struct sched_domain *sd,
1419 enum cpu_idle_type idle, int *all_pinned,
1420 int *this_best_prio, struct rq_iterator *iterator);
1423 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1424 struct sched_domain *sd, enum cpu_idle_type idle,
1425 struct rq_iterator *iterator);
1428 #ifdef CONFIG_CGROUP_CPUACCT
1429 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1431 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1434 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1436 update_load_add(&rq->load, load);
1439 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1441 update_load_sub(&rq->load, load);
1445 static unsigned long source_load(int cpu, int type);
1446 static unsigned long target_load(int cpu, int type);
1447 static unsigned long cpu_avg_load_per_task(int cpu);
1448 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1450 #ifdef CONFIG_FAIR_GROUP_SCHED
1453 * Group load balancing.
1455 * We calculate a few balance domain wide aggregate numbers; load and weight.
1456 * Given the pictures below, and assuming each item has equal weight:
1467 * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd,
1468 * which equals 1/9-th of the total load.
1471 * The weight of this group on the selected cpus.
1474 * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while
1478 * Part of the rq_weight contributed by tasks; all groups except B would
1482 static inline struct aggregate_struct *
1483 aggregate(struct task_group *tg, int cpu)
1485 return &tg->cfs_rq[cpu]->aggregate;
1488 typedef void (*aggregate_func)(struct task_group *, int, struct sched_domain *);
1491 * Iterate the full tree, calling @down when first entering a node and @up when
1492 * leaving it for the final time.
1495 void aggregate_walk_tree(aggregate_func down, aggregate_func up,
1496 int cpu, struct sched_domain *sd)
1498 struct task_group *parent, *child;
1501 parent = &root_task_group;
1503 (*down)(parent, cpu, sd);
1504 list_for_each_entry_rcu(child, &parent->children, siblings) {
1511 (*up)(parent, cpu, sd);
1514 parent = parent->parent;
1521 * Calculate the aggregate runqueue weight.
1524 aggregate_group_weight(struct task_group *tg, int cpu, struct sched_domain *sd)
1526 unsigned long rq_weight = 0;
1527 unsigned long task_weight = 0;
1530 for_each_cpu_mask(i, sd->span) {
1531 rq_weight += tg->cfs_rq[i]->load.weight;
1532 task_weight += tg->cfs_rq[i]->task_weight;
1535 aggregate(tg, cpu)->rq_weight = rq_weight;
1536 aggregate(tg, cpu)->task_weight = task_weight;
1540 * Compute the weight of this group on the given cpus.
1543 aggregate_group_shares(struct task_group *tg, int cpu, struct sched_domain *sd)
1545 unsigned long shares = 0;
1548 for_each_cpu_mask(i, sd->span)
1549 shares += tg->cfs_rq[i]->shares;
1551 if ((!shares && aggregate(tg, cpu)->rq_weight) || shares > tg->shares)
1552 shares = tg->shares;
1554 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1555 shares = tg->shares;
1557 aggregate(tg, cpu)->shares = shares;
1561 * Compute the load fraction assigned to this group, relies on the aggregate
1562 * weight and this group's parent's load, i.e. top-down.
1565 aggregate_group_load(struct task_group *tg, int cpu, struct sched_domain *sd)
1573 for_each_cpu_mask(i, sd->span)
1574 load += cpu_rq(i)->load.weight;
1577 load = aggregate(tg->parent, cpu)->load;
1580 * shares is our weight in the parent's rq so
1581 * shares/parent->rq_weight gives our fraction of the load
1583 load *= aggregate(tg, cpu)->shares;
1584 load /= aggregate(tg->parent, cpu)->rq_weight + 1;
1587 aggregate(tg, cpu)->load = load;
1590 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1593 * Calculate and set the cpu's group shares.
1596 __update_group_shares_cpu(struct task_group *tg, int cpu,
1597 struct sched_domain *sd, int tcpu)
1600 unsigned long shares;
1601 unsigned long rq_weight;
1606 rq_weight = tg->cfs_rq[tcpu]->load.weight;
1609 * If there are currently no tasks on the cpu pretend there is one of
1610 * average load so that when a new task gets to run here it will not
1611 * get delayed by group starvation.
1615 rq_weight = NICE_0_LOAD;
1619 * \Sum shares * rq_weight
1620 * shares = -----------------------
1624 shares = aggregate(tg, cpu)->shares * rq_weight;
1625 shares /= aggregate(tg, cpu)->rq_weight + 1;
1628 * record the actual number of shares, not the boosted amount.
1630 tg->cfs_rq[tcpu]->shares = boost ? 0 : shares;
1632 if (shares < MIN_SHARES)
1633 shares = MIN_SHARES;
1634 else if (shares > MAX_SHARES)
1635 shares = MAX_SHARES;
1637 __set_se_shares(tg->se[tcpu], shares);
1641 * Re-adjust the weights on the cpu the task came from and on the cpu the
1645 __move_group_shares(struct task_group *tg, int cpu, struct sched_domain *sd,
1648 __update_group_shares_cpu(tg, cpu, sd, scpu);
1649 __update_group_shares_cpu(tg, cpu, sd, dcpu);
1653 * Because changing a group's shares changes the weight of the super-group
1654 * we need to walk up the tree and change all shares until we hit the root.
1657 move_group_shares(struct task_group *tg, int cpu, struct sched_domain *sd,
1661 __move_group_shares(tg, cpu, sd, scpu, dcpu);
1667 aggregate_group_set_shares(struct task_group *tg, int cpu, struct sched_domain *sd)
1671 for_each_cpu_mask(i, sd->span) {
1672 struct rq *rq = cpu_rq(i);
1673 unsigned long flags;
1675 spin_lock_irqsave(&rq->lock, flags);
1676 __update_group_shares_cpu(tg, cpu, sd, i);
1677 spin_unlock_irqrestore(&rq->lock, flags);
1680 aggregate_group_shares(tg, cpu, sd);
1684 * Calculate the accumulative weight and recursive load of each task group
1685 * while walking down the tree.
1688 aggregate_get_down(struct task_group *tg, int cpu, struct sched_domain *sd)
1690 aggregate_group_weight(tg, cpu, sd);
1691 aggregate_group_shares(tg, cpu, sd);
1692 aggregate_group_load(tg, cpu, sd);
1696 * Rebalance the cpu shares while walking back up the tree.
1699 aggregate_get_up(struct task_group *tg, int cpu, struct sched_domain *sd)
1701 aggregate_group_set_shares(tg, cpu, sd);
1705 aggregate_get_nop(struct task_group *tg, int cpu, struct sched_domain *sd)
1709 static DEFINE_PER_CPU(spinlock_t, aggregate_lock);
1711 static void __init init_aggregate(void)
1715 for_each_possible_cpu(i)
1716 spin_lock_init(&per_cpu(aggregate_lock, i));
1719 static int get_aggregate(int cpu, struct sched_domain *sd)
1721 if (!spin_trylock(&per_cpu(aggregate_lock, cpu)))
1724 aggregate_walk_tree(aggregate_get_down, aggregate_get_up, cpu, sd);
1728 static void update_aggregate(int cpu, struct sched_domain *sd)
1730 aggregate_walk_tree(aggregate_get_down, aggregate_get_nop, cpu, sd);
1733 static void put_aggregate(int cpu, struct sched_domain *sd)
1735 spin_unlock(&per_cpu(aggregate_lock, cpu));
1738 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1740 cfs_rq->shares = shares;
1745 static inline void init_aggregate(void)
1749 static inline int get_aggregate(int cpu, struct sched_domain *sd)
1754 static inline void update_aggregate(int cpu, struct sched_domain *sd)
1758 static inline void put_aggregate(int cpu, struct sched_domain *sd)
1765 #include "sched_stats.h"
1766 #include "sched_idletask.c"
1767 #include "sched_fair.c"
1768 #include "sched_rt.c"
1769 #ifdef CONFIG_SCHED_DEBUG
1770 # include "sched_debug.c"
1773 #define sched_class_highest (&rt_sched_class)
1774 #define for_each_class(class) \
1775 for (class = sched_class_highest; class; class = class->next)
1777 static void inc_nr_running(struct rq *rq)
1782 static void dec_nr_running(struct rq *rq)
1787 static void set_load_weight(struct task_struct *p)
1789 if (task_has_rt_policy(p)) {
1790 p->se.load.weight = prio_to_weight[0] * 2;
1791 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1796 * SCHED_IDLE tasks get minimal weight:
1798 if (p->policy == SCHED_IDLE) {
1799 p->se.load.weight = WEIGHT_IDLEPRIO;
1800 p->se.load.inv_weight = WMULT_IDLEPRIO;
1804 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1805 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1808 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1810 sched_info_queued(p);
1811 p->sched_class->enqueue_task(rq, p, wakeup);
1815 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1817 p->sched_class->dequeue_task(rq, p, sleep);
1822 * __normal_prio - return the priority that is based on the static prio
1824 static inline int __normal_prio(struct task_struct *p)
1826 return p->static_prio;
1830 * Calculate the expected normal priority: i.e. priority
1831 * without taking RT-inheritance into account. Might be
1832 * boosted by interactivity modifiers. Changes upon fork,
1833 * setprio syscalls, and whenever the interactivity
1834 * estimator recalculates.
1836 static inline int normal_prio(struct task_struct *p)
1840 if (task_has_rt_policy(p))
1841 prio = MAX_RT_PRIO-1 - p->rt_priority;
1843 prio = __normal_prio(p);
1848 * Calculate the current priority, i.e. the priority
1849 * taken into account by the scheduler. This value might
1850 * be boosted by RT tasks, or might be boosted by
1851 * interactivity modifiers. Will be RT if the task got
1852 * RT-boosted. If not then it returns p->normal_prio.
1854 static int effective_prio(struct task_struct *p)
1856 p->normal_prio = normal_prio(p);
1858 * If we are RT tasks or we were boosted to RT priority,
1859 * keep the priority unchanged. Otherwise, update priority
1860 * to the normal priority:
1862 if (!rt_prio(p->prio))
1863 return p->normal_prio;
1868 * activate_task - move a task to the runqueue.
1870 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1872 if (task_contributes_to_load(p))
1873 rq->nr_uninterruptible--;
1875 enqueue_task(rq, p, wakeup);
1880 * deactivate_task - remove a task from the runqueue.
1882 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1884 if (task_contributes_to_load(p))
1885 rq->nr_uninterruptible++;
1887 dequeue_task(rq, p, sleep);
1892 * task_curr - is this task currently executing on a CPU?
1893 * @p: the task in question.
1895 inline int task_curr(const struct task_struct *p)
1897 return cpu_curr(task_cpu(p)) == p;
1900 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1902 set_task_rq(p, cpu);
1905 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1906 * successfuly executed on another CPU. We must ensure that updates of
1907 * per-task data have been completed by this moment.
1910 task_thread_info(p)->cpu = cpu;
1914 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1915 const struct sched_class *prev_class,
1916 int oldprio, int running)
1918 if (prev_class != p->sched_class) {
1919 if (prev_class->switched_from)
1920 prev_class->switched_from(rq, p, running);
1921 p->sched_class->switched_to(rq, p, running);
1923 p->sched_class->prio_changed(rq, p, oldprio, running);
1928 /* Used instead of source_load when we know the type == 0 */
1929 static unsigned long weighted_cpuload(const int cpu)
1931 return cpu_rq(cpu)->load.weight;
1935 * Is this task likely cache-hot:
1938 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1943 * Buddy candidates are cache hot:
1945 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1948 if (p->sched_class != &fair_sched_class)
1951 if (sysctl_sched_migration_cost == -1)
1953 if (sysctl_sched_migration_cost == 0)
1956 delta = now - p->se.exec_start;
1958 return delta < (s64)sysctl_sched_migration_cost;
1962 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1964 int old_cpu = task_cpu(p);
1965 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1966 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1967 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1970 clock_offset = old_rq->clock - new_rq->clock;
1972 #ifdef CONFIG_SCHEDSTATS
1973 if (p->se.wait_start)
1974 p->se.wait_start -= clock_offset;
1975 if (p->se.sleep_start)
1976 p->se.sleep_start -= clock_offset;
1977 if (p->se.block_start)
1978 p->se.block_start -= clock_offset;
1979 if (old_cpu != new_cpu) {
1980 schedstat_inc(p, se.nr_migrations);
1981 if (task_hot(p, old_rq->clock, NULL))
1982 schedstat_inc(p, se.nr_forced2_migrations);
1985 p->se.vruntime -= old_cfsrq->min_vruntime -
1986 new_cfsrq->min_vruntime;
1988 __set_task_cpu(p, new_cpu);
1991 struct migration_req {
1992 struct list_head list;
1994 struct task_struct *task;
1997 struct completion done;
2001 * The task's runqueue lock must be held.
2002 * Returns true if you have to wait for migration thread.
2005 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2007 struct rq *rq = task_rq(p);
2010 * If the task is not on a runqueue (and not running), then
2011 * it is sufficient to simply update the task's cpu field.
2013 if (!p->se.on_rq && !task_running(rq, p)) {
2014 set_task_cpu(p, dest_cpu);
2018 init_completion(&req->done);
2020 req->dest_cpu = dest_cpu;
2021 list_add(&req->list, &rq->migration_queue);
2027 * wait_task_inactive - wait for a thread to unschedule.
2029 * The caller must ensure that the task *will* unschedule sometime soon,
2030 * else this function might spin for a *long* time. This function can't
2031 * be called with interrupts off, or it may introduce deadlock with
2032 * smp_call_function() if an IPI is sent by the same process we are
2033 * waiting to become inactive.
2035 void wait_task_inactive(struct task_struct *p)
2037 unsigned long flags;
2043 * We do the initial early heuristics without holding
2044 * any task-queue locks at all. We'll only try to get
2045 * the runqueue lock when things look like they will
2051 * If the task is actively running on another CPU
2052 * still, just relax and busy-wait without holding
2055 * NOTE! Since we don't hold any locks, it's not
2056 * even sure that "rq" stays as the right runqueue!
2057 * But we don't care, since "task_running()" will
2058 * return false if the runqueue has changed and p
2059 * is actually now running somewhere else!
2061 while (task_running(rq, p))
2065 * Ok, time to look more closely! We need the rq
2066 * lock now, to be *sure*. If we're wrong, we'll
2067 * just go back and repeat.
2069 rq = task_rq_lock(p, &flags);
2070 running = task_running(rq, p);
2071 on_rq = p->se.on_rq;
2072 task_rq_unlock(rq, &flags);
2075 * Was it really running after all now that we
2076 * checked with the proper locks actually held?
2078 * Oops. Go back and try again..
2080 if (unlikely(running)) {
2086 * It's not enough that it's not actively running,
2087 * it must be off the runqueue _entirely_, and not
2090 * So if it wa still runnable (but just not actively
2091 * running right now), it's preempted, and we should
2092 * yield - it could be a while.
2094 if (unlikely(on_rq)) {
2095 schedule_timeout_uninterruptible(1);
2100 * Ahh, all good. It wasn't running, and it wasn't
2101 * runnable, which means that it will never become
2102 * running in the future either. We're all done!
2109 * kick_process - kick a running thread to enter/exit the kernel
2110 * @p: the to-be-kicked thread
2112 * Cause a process which is running on another CPU to enter
2113 * kernel-mode, without any delay. (to get signals handled.)
2115 * NOTE: this function doesnt have to take the runqueue lock,
2116 * because all it wants to ensure is that the remote task enters
2117 * the kernel. If the IPI races and the task has been migrated
2118 * to another CPU then no harm is done and the purpose has been
2121 void kick_process(struct task_struct *p)
2127 if ((cpu != smp_processor_id()) && task_curr(p))
2128 smp_send_reschedule(cpu);
2133 * Return a low guess at the load of a migration-source cpu weighted
2134 * according to the scheduling class and "nice" value.
2136 * We want to under-estimate the load of migration sources, to
2137 * balance conservatively.
2139 static unsigned long source_load(int cpu, int type)
2141 struct rq *rq = cpu_rq(cpu);
2142 unsigned long total = weighted_cpuload(cpu);
2147 return min(rq->cpu_load[type-1], total);
2151 * Return a high guess at the load of a migration-target cpu weighted
2152 * according to the scheduling class and "nice" value.
2154 static unsigned long target_load(int cpu, int type)
2156 struct rq *rq = cpu_rq(cpu);
2157 unsigned long total = weighted_cpuload(cpu);
2162 return max(rq->cpu_load[type-1], total);
2166 * Return the average load per task on the cpu's run queue
2168 static unsigned long cpu_avg_load_per_task(int cpu)
2170 struct rq *rq = cpu_rq(cpu);
2171 unsigned long total = weighted_cpuload(cpu);
2172 unsigned long n = rq->nr_running;
2174 return n ? total / n : SCHED_LOAD_SCALE;
2178 * find_idlest_group finds and returns the least busy CPU group within the
2181 static struct sched_group *
2182 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2184 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2185 unsigned long min_load = ULONG_MAX, this_load = 0;
2186 int load_idx = sd->forkexec_idx;
2187 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2190 * now that we have both rqs locked the rq weight won't change
2191 * anymore - so update the stats.
2193 update_aggregate(this_cpu, sd);
2196 unsigned long load, avg_load;
2200 /* Skip over this group if it has no CPUs allowed */
2201 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2204 local_group = cpu_isset(this_cpu, group->cpumask);
2206 /* Tally up the load of all CPUs in the group */
2209 for_each_cpu_mask(i, group->cpumask) {
2210 /* Bias balancing toward cpus of our domain */
2212 load = source_load(i, load_idx);
2214 load = target_load(i, load_idx);
2219 /* Adjust by relative CPU power of the group */
2220 avg_load = sg_div_cpu_power(group,
2221 avg_load * SCHED_LOAD_SCALE);
2224 this_load = avg_load;
2226 } else if (avg_load < min_load) {
2227 min_load = avg_load;
2230 } while (group = group->next, group != sd->groups);
2232 if (!idlest || 100*this_load < imbalance*min_load)
2238 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2241 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2244 unsigned long load, min_load = ULONG_MAX;
2248 /* Traverse only the allowed CPUs */
2249 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2251 for_each_cpu_mask(i, *tmp) {
2252 load = weighted_cpuload(i);
2254 if (load < min_load || (load == min_load && i == this_cpu)) {
2264 * sched_balance_self: balance the current task (running on cpu) in domains
2265 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2268 * Balance, ie. select the least loaded group.
2270 * Returns the target CPU number, or the same CPU if no balancing is needed.
2272 * preempt must be disabled.
2274 static int sched_balance_self(int cpu, int flag)
2276 struct task_struct *t = current;
2277 struct sched_domain *tmp, *sd = NULL;
2279 for_each_domain(cpu, tmp) {
2281 * If power savings logic is enabled for a domain, stop there.
2283 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2285 if (tmp->flags & flag)
2290 cpumask_t span, tmpmask;
2291 struct sched_group *group;
2292 int new_cpu, weight;
2294 if (!(sd->flags & flag)) {
2300 group = find_idlest_group(sd, t, cpu);
2306 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2307 if (new_cpu == -1 || new_cpu == cpu) {
2308 /* Now try balancing at a lower domain level of cpu */
2313 /* Now try balancing at a lower domain level of new_cpu */
2316 weight = cpus_weight(span);
2317 for_each_domain(cpu, tmp) {
2318 if (weight <= cpus_weight(tmp->span))
2320 if (tmp->flags & flag)
2323 /* while loop will break here if sd == NULL */
2329 #endif /* CONFIG_SMP */
2332 * try_to_wake_up - wake up a thread
2333 * @p: the to-be-woken-up thread
2334 * @state: the mask of task states that can be woken
2335 * @sync: do a synchronous wakeup?
2337 * Put it on the run-queue if it's not already there. The "current"
2338 * thread is always on the run-queue (except when the actual
2339 * re-schedule is in progress), and as such you're allowed to do
2340 * the simpler "current->state = TASK_RUNNING" to mark yourself
2341 * runnable without the overhead of this.
2343 * returns failure only if the task is already active.
2345 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2347 int cpu, orig_cpu, this_cpu, success = 0;
2348 unsigned long flags;
2352 if (!sched_feat(SYNC_WAKEUPS))
2356 rq = task_rq_lock(p, &flags);
2357 old_state = p->state;
2358 if (!(old_state & state))
2366 this_cpu = smp_processor_id();
2369 if (unlikely(task_running(rq, p)))
2372 cpu = p->sched_class->select_task_rq(p, sync);
2373 if (cpu != orig_cpu) {
2374 set_task_cpu(p, cpu);
2375 task_rq_unlock(rq, &flags);
2376 /* might preempt at this point */
2377 rq = task_rq_lock(p, &flags);
2378 old_state = p->state;
2379 if (!(old_state & state))
2384 this_cpu = smp_processor_id();
2388 #ifdef CONFIG_SCHEDSTATS
2389 schedstat_inc(rq, ttwu_count);
2390 if (cpu == this_cpu)
2391 schedstat_inc(rq, ttwu_local);
2393 struct sched_domain *sd;
2394 for_each_domain(this_cpu, sd) {
2395 if (cpu_isset(cpu, sd->span)) {
2396 schedstat_inc(sd, ttwu_wake_remote);
2401 #endif /* CONFIG_SCHEDSTATS */
2404 #endif /* CONFIG_SMP */
2405 schedstat_inc(p, se.nr_wakeups);
2407 schedstat_inc(p, se.nr_wakeups_sync);
2408 if (orig_cpu != cpu)
2409 schedstat_inc(p, se.nr_wakeups_migrate);
2410 if (cpu == this_cpu)
2411 schedstat_inc(p, se.nr_wakeups_local);
2413 schedstat_inc(p, se.nr_wakeups_remote);
2414 update_rq_clock(rq);
2415 activate_task(rq, p, 1);
2419 check_preempt_curr(rq, p);
2421 p->state = TASK_RUNNING;
2423 if (p->sched_class->task_wake_up)
2424 p->sched_class->task_wake_up(rq, p);
2427 task_rq_unlock(rq, &flags);
2432 int wake_up_process(struct task_struct *p)
2434 return try_to_wake_up(p, TASK_ALL, 0);
2436 EXPORT_SYMBOL(wake_up_process);
2438 int wake_up_state(struct task_struct *p, unsigned int state)
2440 return try_to_wake_up(p, state, 0);
2444 * Perform scheduler related setup for a newly forked process p.
2445 * p is forked by current.
2447 * __sched_fork() is basic setup used by init_idle() too:
2449 static void __sched_fork(struct task_struct *p)
2451 p->se.exec_start = 0;
2452 p->se.sum_exec_runtime = 0;
2453 p->se.prev_sum_exec_runtime = 0;
2454 p->se.last_wakeup = 0;
2455 p->se.avg_overlap = 0;
2457 #ifdef CONFIG_SCHEDSTATS
2458 p->se.wait_start = 0;
2459 p->se.sum_sleep_runtime = 0;
2460 p->se.sleep_start = 0;
2461 p->se.block_start = 0;
2462 p->se.sleep_max = 0;
2463 p->se.block_max = 0;
2465 p->se.slice_max = 0;
2469 INIT_LIST_HEAD(&p->rt.run_list);
2471 INIT_LIST_HEAD(&p->se.group_node);
2473 #ifdef CONFIG_PREEMPT_NOTIFIERS
2474 INIT_HLIST_HEAD(&p->preempt_notifiers);
2478 * We mark the process as running here, but have not actually
2479 * inserted it onto the runqueue yet. This guarantees that
2480 * nobody will actually run it, and a signal or other external
2481 * event cannot wake it up and insert it on the runqueue either.
2483 p->state = TASK_RUNNING;
2487 * fork()/clone()-time setup:
2489 void sched_fork(struct task_struct *p, int clone_flags)
2491 int cpu = get_cpu();
2496 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2498 set_task_cpu(p, cpu);
2501 * Make sure we do not leak PI boosting priority to the child:
2503 p->prio = current->normal_prio;
2504 if (!rt_prio(p->prio))
2505 p->sched_class = &fair_sched_class;
2507 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2508 if (likely(sched_info_on()))
2509 memset(&p->sched_info, 0, sizeof(p->sched_info));
2511 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2514 #ifdef CONFIG_PREEMPT
2515 /* Want to start with kernel preemption disabled. */
2516 task_thread_info(p)->preempt_count = 1;
2522 * wake_up_new_task - wake up a newly created task for the first time.
2524 * This function will do some initial scheduler statistics housekeeping
2525 * that must be done for every newly created context, then puts the task
2526 * on the runqueue and wakes it.
2528 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2530 unsigned long flags;
2533 rq = task_rq_lock(p, &flags);
2534 BUG_ON(p->state != TASK_RUNNING);
2535 update_rq_clock(rq);
2537 p->prio = effective_prio(p);
2539 if (!p->sched_class->task_new || !current->se.on_rq) {
2540 activate_task(rq, p, 0);
2543 * Let the scheduling class do new task startup
2544 * management (if any):
2546 p->sched_class->task_new(rq, p);
2549 check_preempt_curr(rq, p);
2551 if (p->sched_class->task_wake_up)
2552 p->sched_class->task_wake_up(rq, p);
2554 task_rq_unlock(rq, &flags);
2557 #ifdef CONFIG_PREEMPT_NOTIFIERS
2560 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2561 * @notifier: notifier struct to register
2563 void preempt_notifier_register(struct preempt_notifier *notifier)
2565 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2567 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2570 * preempt_notifier_unregister - no longer interested in preemption notifications
2571 * @notifier: notifier struct to unregister
2573 * This is safe to call from within a preemption notifier.
2575 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2577 hlist_del(¬ifier->link);
2579 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2581 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2583 struct preempt_notifier *notifier;
2584 struct hlist_node *node;
2586 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2587 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2591 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2592 struct task_struct *next)
2594 struct preempt_notifier *notifier;
2595 struct hlist_node *node;
2597 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2598 notifier->ops->sched_out(notifier, next);
2601 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2603 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2608 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2609 struct task_struct *next)
2613 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2616 * prepare_task_switch - prepare to switch tasks
2617 * @rq: the runqueue preparing to switch
2618 * @prev: the current task that is being switched out
2619 * @next: the task we are going to switch to.
2621 * This is called with the rq lock held and interrupts off. It must
2622 * be paired with a subsequent finish_task_switch after the context
2625 * prepare_task_switch sets up locking and calls architecture specific
2629 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2630 struct task_struct *next)
2632 fire_sched_out_preempt_notifiers(prev, next);
2633 prepare_lock_switch(rq, next);
2634 prepare_arch_switch(next);
2638 * finish_task_switch - clean up after a task-switch
2639 * @rq: runqueue associated with task-switch
2640 * @prev: the thread we just switched away from.
2642 * finish_task_switch must be called after the context switch, paired
2643 * with a prepare_task_switch call before the context switch.
2644 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2645 * and do any other architecture-specific cleanup actions.
2647 * Note that we may have delayed dropping an mm in context_switch(). If
2648 * so, we finish that here outside of the runqueue lock. (Doing it
2649 * with the lock held can cause deadlocks; see schedule() for
2652 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2653 __releases(rq->lock)
2655 struct mm_struct *mm = rq->prev_mm;
2661 * A task struct has one reference for the use as "current".
2662 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2663 * schedule one last time. The schedule call will never return, and
2664 * the scheduled task must drop that reference.
2665 * The test for TASK_DEAD must occur while the runqueue locks are
2666 * still held, otherwise prev could be scheduled on another cpu, die
2667 * there before we look at prev->state, and then the reference would
2669 * Manfred Spraul <manfred@colorfullife.com>
2671 prev_state = prev->state;
2672 finish_arch_switch(prev);
2673 finish_lock_switch(rq, prev);
2675 if (current->sched_class->post_schedule)
2676 current->sched_class->post_schedule(rq);
2679 fire_sched_in_preempt_notifiers(current);
2682 if (unlikely(prev_state == TASK_DEAD)) {
2684 * Remove function-return probe instances associated with this
2685 * task and put them back on the free list.
2687 kprobe_flush_task(prev);
2688 put_task_struct(prev);
2693 * schedule_tail - first thing a freshly forked thread must call.
2694 * @prev: the thread we just switched away from.
2696 asmlinkage void schedule_tail(struct task_struct *prev)
2697 __releases(rq->lock)
2699 struct rq *rq = this_rq();
2701 finish_task_switch(rq, prev);
2702 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2703 /* In this case, finish_task_switch does not reenable preemption */
2706 if (current->set_child_tid)
2707 put_user(task_pid_vnr(current), current->set_child_tid);
2711 * context_switch - switch to the new MM and the new
2712 * thread's register state.
2715 context_switch(struct rq *rq, struct task_struct *prev,
2716 struct task_struct *next)
2718 struct mm_struct *mm, *oldmm;
2720 prepare_task_switch(rq, prev, next);
2722 oldmm = prev->active_mm;
2724 * For paravirt, this is coupled with an exit in switch_to to
2725 * combine the page table reload and the switch backend into
2728 arch_enter_lazy_cpu_mode();
2730 if (unlikely(!mm)) {
2731 next->active_mm = oldmm;
2732 atomic_inc(&oldmm->mm_count);
2733 enter_lazy_tlb(oldmm, next);
2735 switch_mm(oldmm, mm, next);
2737 if (unlikely(!prev->mm)) {
2738 prev->active_mm = NULL;
2739 rq->prev_mm = oldmm;
2742 * Since the runqueue lock will be released by the next
2743 * task (which is an invalid locking op but in the case
2744 * of the scheduler it's an obvious special-case), so we
2745 * do an early lockdep release here:
2747 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2748 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2751 /* Here we just switch the register state and the stack. */
2752 switch_to(prev, next, prev);
2756 * this_rq must be evaluated again because prev may have moved
2757 * CPUs since it called schedule(), thus the 'rq' on its stack
2758 * frame will be invalid.
2760 finish_task_switch(this_rq(), prev);
2764 * nr_running, nr_uninterruptible and nr_context_switches:
2766 * externally visible scheduler statistics: current number of runnable
2767 * threads, current number of uninterruptible-sleeping threads, total
2768 * number of context switches performed since bootup.
2770 unsigned long nr_running(void)
2772 unsigned long i, sum = 0;
2774 for_each_online_cpu(i)
2775 sum += cpu_rq(i)->nr_running;
2780 unsigned long nr_uninterruptible(void)
2782 unsigned long i, sum = 0;
2784 for_each_possible_cpu(i)
2785 sum += cpu_rq(i)->nr_uninterruptible;
2788 * Since we read the counters lockless, it might be slightly
2789 * inaccurate. Do not allow it to go below zero though:
2791 if (unlikely((long)sum < 0))
2797 unsigned long long nr_context_switches(void)
2800 unsigned long long sum = 0;
2802 for_each_possible_cpu(i)
2803 sum += cpu_rq(i)->nr_switches;
2808 unsigned long nr_iowait(void)
2810 unsigned long i, sum = 0;
2812 for_each_possible_cpu(i)
2813 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2818 unsigned long nr_active(void)
2820 unsigned long i, running = 0, uninterruptible = 0;
2822 for_each_online_cpu(i) {
2823 running += cpu_rq(i)->nr_running;
2824 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2827 if (unlikely((long)uninterruptible < 0))
2828 uninterruptible = 0;
2830 return running + uninterruptible;
2834 * Update rq->cpu_load[] statistics. This function is usually called every
2835 * scheduler tick (TICK_NSEC).
2837 static void update_cpu_load(struct rq *this_rq)
2839 unsigned long this_load = this_rq->load.weight;
2842 this_rq->nr_load_updates++;
2844 /* Update our load: */
2845 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2846 unsigned long old_load, new_load;
2848 /* scale is effectively 1 << i now, and >> i divides by scale */
2850 old_load = this_rq->cpu_load[i];
2851 new_load = this_load;
2853 * Round up the averaging division if load is increasing. This
2854 * prevents us from getting stuck on 9 if the load is 10, for
2857 if (new_load > old_load)
2858 new_load += scale-1;
2859 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2866 * double_rq_lock - safely lock two runqueues
2868 * Note this does not disable interrupts like task_rq_lock,
2869 * you need to do so manually before calling.
2871 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2872 __acquires(rq1->lock)
2873 __acquires(rq2->lock)
2875 BUG_ON(!irqs_disabled());
2877 spin_lock(&rq1->lock);
2878 __acquire(rq2->lock); /* Fake it out ;) */
2881 spin_lock(&rq1->lock);
2882 spin_lock(&rq2->lock);
2884 spin_lock(&rq2->lock);
2885 spin_lock(&rq1->lock);
2888 update_rq_clock(rq1);
2889 update_rq_clock(rq2);
2893 * double_rq_unlock - safely unlock two runqueues
2895 * Note this does not restore interrupts like task_rq_unlock,
2896 * you need to do so manually after calling.
2898 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2899 __releases(rq1->lock)
2900 __releases(rq2->lock)
2902 spin_unlock(&rq1->lock);
2904 spin_unlock(&rq2->lock);
2906 __release(rq2->lock);
2910 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2912 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2913 __releases(this_rq->lock)
2914 __acquires(busiest->lock)
2915 __acquires(this_rq->lock)
2919 if (unlikely(!irqs_disabled())) {
2920 /* printk() doesn't work good under rq->lock */
2921 spin_unlock(&this_rq->lock);
2924 if (unlikely(!spin_trylock(&busiest->lock))) {
2925 if (busiest < this_rq) {
2926 spin_unlock(&this_rq->lock);
2927 spin_lock(&busiest->lock);
2928 spin_lock(&this_rq->lock);
2931 spin_lock(&busiest->lock);
2937 * If dest_cpu is allowed for this process, migrate the task to it.
2938 * This is accomplished by forcing the cpu_allowed mask to only
2939 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2940 * the cpu_allowed mask is restored.
2942 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2944 struct migration_req req;
2945 unsigned long flags;
2948 rq = task_rq_lock(p, &flags);
2949 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2950 || unlikely(cpu_is_offline(dest_cpu)))
2953 /* force the process onto the specified CPU */
2954 if (migrate_task(p, dest_cpu, &req)) {
2955 /* Need to wait for migration thread (might exit: take ref). */
2956 struct task_struct *mt = rq->migration_thread;
2958 get_task_struct(mt);
2959 task_rq_unlock(rq, &flags);
2960 wake_up_process(mt);
2961 put_task_struct(mt);
2962 wait_for_completion(&req.done);
2967 task_rq_unlock(rq, &flags);
2971 * sched_exec - execve() is a valuable balancing opportunity, because at
2972 * this point the task has the smallest effective memory and cache footprint.
2974 void sched_exec(void)
2976 int new_cpu, this_cpu = get_cpu();
2977 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2979 if (new_cpu != this_cpu)
2980 sched_migrate_task(current, new_cpu);
2984 * pull_task - move a task from a remote runqueue to the local runqueue.
2985 * Both runqueues must be locked.
2987 static void pull_task(struct rq *src_rq, struct task_struct *p,
2988 struct rq *this_rq, int this_cpu)
2990 deactivate_task(src_rq, p, 0);
2991 set_task_cpu(p, this_cpu);
2992 activate_task(this_rq, p, 0);
2994 * Note that idle threads have a prio of MAX_PRIO, for this test
2995 * to be always true for them.
2997 check_preempt_curr(this_rq, p);
3001 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3004 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3005 struct sched_domain *sd, enum cpu_idle_type idle,
3009 * We do not migrate tasks that are:
3010 * 1) running (obviously), or
3011 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3012 * 3) are cache-hot on their current CPU.
3014 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
3015 schedstat_inc(p, se.nr_failed_migrations_affine);
3020 if (task_running(rq, p)) {
3021 schedstat_inc(p, se.nr_failed_migrations_running);
3026 * Aggressive migration if:
3027 * 1) task is cache cold, or
3028 * 2) too many balance attempts have failed.
3031 if (!task_hot(p, rq->clock, sd) ||
3032 sd->nr_balance_failed > sd->cache_nice_tries) {
3033 #ifdef CONFIG_SCHEDSTATS
3034 if (task_hot(p, rq->clock, sd)) {
3035 schedstat_inc(sd, lb_hot_gained[idle]);
3036 schedstat_inc(p, se.nr_forced_migrations);
3042 if (task_hot(p, rq->clock, sd)) {
3043 schedstat_inc(p, se.nr_failed_migrations_hot);
3049 static unsigned long
3050 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3051 unsigned long max_load_move, struct sched_domain *sd,
3052 enum cpu_idle_type idle, int *all_pinned,
3053 int *this_best_prio, struct rq_iterator *iterator)
3055 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
3056 struct task_struct *p;
3057 long rem_load_move = max_load_move;
3059 if (max_load_move == 0)
3065 * Start the load-balancing iterator:
3067 p = iterator->start(iterator->arg);
3069 if (!p || loops++ > sysctl_sched_nr_migrate)
3072 * To help distribute high priority tasks across CPUs we don't
3073 * skip a task if it will be the highest priority task (i.e. smallest
3074 * prio value) on its new queue regardless of its load weight
3076 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
3077 SCHED_LOAD_SCALE_FUZZ;
3078 if ((skip_for_load && p->prio >= *this_best_prio) ||
3079 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3080 p = iterator->next(iterator->arg);
3084 pull_task(busiest, p, this_rq, this_cpu);
3086 rem_load_move -= p->se.load.weight;
3089 * We only want to steal up to the prescribed amount of weighted load.
3091 if (rem_load_move > 0) {
3092 if (p->prio < *this_best_prio)
3093 *this_best_prio = p->prio;
3094 p = iterator->next(iterator->arg);
3099 * Right now, this is one of only two places pull_task() is called,
3100 * so we can safely collect pull_task() stats here rather than
3101 * inside pull_task().
3103 schedstat_add(sd, lb_gained[idle], pulled);
3106 *all_pinned = pinned;
3108 return max_load_move - rem_load_move;
3112 * move_tasks tries to move up to max_load_move weighted load from busiest to
3113 * this_rq, as part of a balancing operation within domain "sd".
3114 * Returns 1 if successful and 0 otherwise.
3116 * Called with both runqueues locked.
3118 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3119 unsigned long max_load_move,
3120 struct sched_domain *sd, enum cpu_idle_type idle,
3123 const struct sched_class *class = sched_class_highest;
3124 unsigned long total_load_moved = 0;
3125 int this_best_prio = this_rq->curr->prio;
3129 class->load_balance(this_rq, this_cpu, busiest,
3130 max_load_move - total_load_moved,
3131 sd, idle, all_pinned, &this_best_prio);
3132 class = class->next;
3133 } while (class && max_load_move > total_load_moved);
3135 return total_load_moved > 0;
3139 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3140 struct sched_domain *sd, enum cpu_idle_type idle,
3141 struct rq_iterator *iterator)
3143 struct task_struct *p = iterator->start(iterator->arg);
3147 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3148 pull_task(busiest, p, this_rq, this_cpu);
3150 * Right now, this is only the second place pull_task()
3151 * is called, so we can safely collect pull_task()
3152 * stats here rather than inside pull_task().
3154 schedstat_inc(sd, lb_gained[idle]);
3158 p = iterator->next(iterator->arg);
3165 * move_one_task tries to move exactly one task from busiest to this_rq, as
3166 * part of active balancing operations within "domain".
3167 * Returns 1 if successful and 0 otherwise.
3169 * Called with both runqueues locked.
3171 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3172 struct sched_domain *sd, enum cpu_idle_type idle)
3174 const struct sched_class *class;
3176 for (class = sched_class_highest; class; class = class->next)
3177 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3184 * find_busiest_group finds and returns the busiest CPU group within the
3185 * domain. It calculates and returns the amount of weighted load which
3186 * should be moved to restore balance via the imbalance parameter.
3188 static struct sched_group *
3189 find_busiest_group(struct sched_domain *sd, int this_cpu,
3190 unsigned long *imbalance, enum cpu_idle_type idle,
3191 int *sd_idle, const cpumask_t *cpus, int *balance)
3193 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3194 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3195 unsigned long max_pull;
3196 unsigned long busiest_load_per_task, busiest_nr_running;
3197 unsigned long this_load_per_task, this_nr_running;
3198 int load_idx, group_imb = 0;
3199 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3200 int power_savings_balance = 1;
3201 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3202 unsigned long min_nr_running = ULONG_MAX;
3203 struct sched_group *group_min = NULL, *group_leader = NULL;
3206 max_load = this_load = total_load = total_pwr = 0;
3207 busiest_load_per_task = busiest_nr_running = 0;
3208 this_load_per_task = this_nr_running = 0;
3209 if (idle == CPU_NOT_IDLE)
3210 load_idx = sd->busy_idx;
3211 else if (idle == CPU_NEWLY_IDLE)
3212 load_idx = sd->newidle_idx;
3214 load_idx = sd->idle_idx;
3217 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3220 int __group_imb = 0;
3221 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3222 unsigned long sum_nr_running, sum_weighted_load;
3224 local_group = cpu_isset(this_cpu, group->cpumask);
3227 balance_cpu = first_cpu(group->cpumask);
3229 /* Tally up the load of all CPUs in the group */
3230 sum_weighted_load = sum_nr_running = avg_load = 0;
3232 min_cpu_load = ~0UL;
3234 for_each_cpu_mask(i, group->cpumask) {
3237 if (!cpu_isset(i, *cpus))
3242 if (*sd_idle && rq->nr_running)
3245 /* Bias balancing toward cpus of our domain */
3247 if (idle_cpu(i) && !first_idle_cpu) {
3252 load = target_load(i, load_idx);
3254 load = source_load(i, load_idx);
3255 if (load > max_cpu_load)
3256 max_cpu_load = load;
3257 if (min_cpu_load > load)
3258 min_cpu_load = load;
3262 sum_nr_running += rq->nr_running;
3263 sum_weighted_load += weighted_cpuload(i);
3267 * First idle cpu or the first cpu(busiest) in this sched group
3268 * is eligible for doing load balancing at this and above
3269 * domains. In the newly idle case, we will allow all the cpu's
3270 * to do the newly idle load balance.
3272 if (idle != CPU_NEWLY_IDLE && local_group &&
3273 balance_cpu != this_cpu && balance) {
3278 total_load += avg_load;
3279 total_pwr += group->__cpu_power;
3281 /* Adjust by relative CPU power of the group */
3282 avg_load = sg_div_cpu_power(group,
3283 avg_load * SCHED_LOAD_SCALE);
3285 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3288 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3291 this_load = avg_load;
3293 this_nr_running = sum_nr_running;
3294 this_load_per_task = sum_weighted_load;
3295 } else if (avg_load > max_load &&
3296 (sum_nr_running > group_capacity || __group_imb)) {
3297 max_load = avg_load;
3299 busiest_nr_running = sum_nr_running;
3300 busiest_load_per_task = sum_weighted_load;
3301 group_imb = __group_imb;
3304 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3306 * Busy processors will not participate in power savings
3309 if (idle == CPU_NOT_IDLE ||
3310 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3314 * If the local group is idle or completely loaded
3315 * no need to do power savings balance at this domain
3317 if (local_group && (this_nr_running >= group_capacity ||
3319 power_savings_balance = 0;
3322 * If a group is already running at full capacity or idle,
3323 * don't include that group in power savings calculations
3325 if (!power_savings_balance || sum_nr_running >= group_capacity
3330 * Calculate the group which has the least non-idle load.
3331 * This is the group from where we need to pick up the load
3334 if ((sum_nr_running < min_nr_running) ||
3335 (sum_nr_running == min_nr_running &&
3336 first_cpu(group->cpumask) <
3337 first_cpu(group_min->cpumask))) {
3339 min_nr_running = sum_nr_running;
3340 min_load_per_task = sum_weighted_load /
3345 * Calculate the group which is almost near its
3346 * capacity but still has some space to pick up some load
3347 * from other group and save more power
3349 if (sum_nr_running <= group_capacity - 1) {
3350 if (sum_nr_running > leader_nr_running ||
3351 (sum_nr_running == leader_nr_running &&
3352 first_cpu(group->cpumask) >
3353 first_cpu(group_leader->cpumask))) {
3354 group_leader = group;
3355 leader_nr_running = sum_nr_running;
3360 group = group->next;
3361 } while (group != sd->groups);
3363 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3366 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3368 if (this_load >= avg_load ||
3369 100*max_load <= sd->imbalance_pct*this_load)
3372 busiest_load_per_task /= busiest_nr_running;
3374 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3377 * We're trying to get all the cpus to the average_load, so we don't
3378 * want to push ourselves above the average load, nor do we wish to
3379 * reduce the max loaded cpu below the average load, as either of these
3380 * actions would just result in more rebalancing later, and ping-pong
3381 * tasks around. Thus we look for the minimum possible imbalance.
3382 * Negative imbalances (*we* are more loaded than anyone else) will
3383 * be counted as no imbalance for these purposes -- we can't fix that
3384 * by pulling tasks to us. Be careful of negative numbers as they'll
3385 * appear as very large values with unsigned longs.
3387 if (max_load <= busiest_load_per_task)
3391 * In the presence of smp nice balancing, certain scenarios can have
3392 * max load less than avg load(as we skip the groups at or below
3393 * its cpu_power, while calculating max_load..)
3395 if (max_load < avg_load) {
3397 goto small_imbalance;
3400 /* Don't want to pull so many tasks that a group would go idle */
3401 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3403 /* How much load to actually move to equalise the imbalance */
3404 *imbalance = min(max_pull * busiest->__cpu_power,
3405 (avg_load - this_load) * this->__cpu_power)
3409 * if *imbalance is less than the average load per runnable task
3410 * there is no gaurantee that any tasks will be moved so we'll have
3411 * a think about bumping its value to force at least one task to be
3414 if (*imbalance < busiest_load_per_task) {
3415 unsigned long tmp, pwr_now, pwr_move;
3419 pwr_move = pwr_now = 0;
3421 if (this_nr_running) {
3422 this_load_per_task /= this_nr_running;
3423 if (busiest_load_per_task > this_load_per_task)
3426 this_load_per_task = SCHED_LOAD_SCALE;
3428 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3429 busiest_load_per_task * imbn) {
3430 *imbalance = busiest_load_per_task;
3435 * OK, we don't have enough imbalance to justify moving tasks,
3436 * however we may be able to increase total CPU power used by
3440 pwr_now += busiest->__cpu_power *
3441 min(busiest_load_per_task, max_load);
3442 pwr_now += this->__cpu_power *
3443 min(this_load_per_task, this_load);
3444 pwr_now /= SCHED_LOAD_SCALE;
3446 /* Amount of load we'd subtract */
3447 tmp = sg_div_cpu_power(busiest,
3448 busiest_load_per_task * SCHED_LOAD_SCALE);
3450 pwr_move += busiest->__cpu_power *
3451 min(busiest_load_per_task, max_load - tmp);
3453 /* Amount of load we'd add */
3454 if (max_load * busiest->__cpu_power <
3455 busiest_load_per_task * SCHED_LOAD_SCALE)
3456 tmp = sg_div_cpu_power(this,
3457 max_load * busiest->__cpu_power);
3459 tmp = sg_div_cpu_power(this,
3460 busiest_load_per_task * SCHED_LOAD_SCALE);
3461 pwr_move += this->__cpu_power *
3462 min(this_load_per_task, this_load + tmp);
3463 pwr_move /= SCHED_LOAD_SCALE;
3465 /* Move if we gain throughput */
3466 if (pwr_move > pwr_now)
3467 *imbalance = busiest_load_per_task;
3473 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3474 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3477 if (this == group_leader && group_leader != group_min) {
3478 *imbalance = min_load_per_task;
3488 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3491 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3492 unsigned long imbalance, const cpumask_t *cpus)
3494 struct rq *busiest = NULL, *rq;
3495 unsigned long max_load = 0;
3498 for_each_cpu_mask(i, group->cpumask) {
3501 if (!cpu_isset(i, *cpus))
3505 wl = weighted_cpuload(i);
3507 if (rq->nr_running == 1 && wl > imbalance)
3510 if (wl > max_load) {
3520 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3521 * so long as it is large enough.
3523 #define MAX_PINNED_INTERVAL 512
3526 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3527 * tasks if there is an imbalance.
3529 static int load_balance(int this_cpu, struct rq *this_rq,
3530 struct sched_domain *sd, enum cpu_idle_type idle,
3531 int *balance, cpumask_t *cpus)
3533 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3534 struct sched_group *group;
3535 unsigned long imbalance;
3537 unsigned long flags;
3538 int unlock_aggregate;
3542 unlock_aggregate = get_aggregate(this_cpu, sd);
3545 * When power savings policy is enabled for the parent domain, idle
3546 * sibling can pick up load irrespective of busy siblings. In this case,
3547 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3548 * portraying it as CPU_NOT_IDLE.
3550 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3551 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3554 schedstat_inc(sd, lb_count[idle]);
3557 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3564 schedstat_inc(sd, lb_nobusyg[idle]);
3568 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3570 schedstat_inc(sd, lb_nobusyq[idle]);
3574 BUG_ON(busiest == this_rq);
3576 schedstat_add(sd, lb_imbalance[idle], imbalance);
3579 if (busiest->nr_running > 1) {
3581 * Attempt to move tasks. If find_busiest_group has found
3582 * an imbalance but busiest->nr_running <= 1, the group is
3583 * still unbalanced. ld_moved simply stays zero, so it is
3584 * correctly treated as an imbalance.
3586 local_irq_save(flags);
3587 double_rq_lock(this_rq, busiest);
3588 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3589 imbalance, sd, idle, &all_pinned);
3590 double_rq_unlock(this_rq, busiest);
3591 local_irq_restore(flags);
3594 * some other cpu did the load balance for us.
3596 if (ld_moved && this_cpu != smp_processor_id())
3597 resched_cpu(this_cpu);
3599 /* All tasks on this runqueue were pinned by CPU affinity */
3600 if (unlikely(all_pinned)) {
3601 cpu_clear(cpu_of(busiest), *cpus);
3602 if (!cpus_empty(*cpus))
3609 schedstat_inc(sd, lb_failed[idle]);
3610 sd->nr_balance_failed++;
3612 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3614 spin_lock_irqsave(&busiest->lock, flags);
3616 /* don't kick the migration_thread, if the curr
3617 * task on busiest cpu can't be moved to this_cpu
3619 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3620 spin_unlock_irqrestore(&busiest->lock, flags);
3622 goto out_one_pinned;
3625 if (!busiest->active_balance) {
3626 busiest->active_balance = 1;
3627 busiest->push_cpu = this_cpu;
3630 spin_unlock_irqrestore(&busiest->lock, flags);
3632 wake_up_process(busiest->migration_thread);
3635 * We've kicked active balancing, reset the failure
3638 sd->nr_balance_failed = sd->cache_nice_tries+1;
3641 sd->nr_balance_failed = 0;
3643 if (likely(!active_balance)) {
3644 /* We were unbalanced, so reset the balancing interval */
3645 sd->balance_interval = sd->min_interval;
3648 * If we've begun active balancing, start to back off. This
3649 * case may not be covered by the all_pinned logic if there
3650 * is only 1 task on the busy runqueue (because we don't call
3653 if (sd->balance_interval < sd->max_interval)
3654 sd->balance_interval *= 2;
3657 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3658 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3664 schedstat_inc(sd, lb_balanced[idle]);
3666 sd->nr_balance_failed = 0;
3669 /* tune up the balancing interval */
3670 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3671 (sd->balance_interval < sd->max_interval))
3672 sd->balance_interval *= 2;
3674 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3675 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3680 if (unlock_aggregate)
3681 put_aggregate(this_cpu, sd);
3686 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3687 * tasks if there is an imbalance.
3689 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3690 * this_rq is locked.
3693 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3696 struct sched_group *group;
3697 struct rq *busiest = NULL;
3698 unsigned long imbalance;
3706 * When power savings policy is enabled for the parent domain, idle
3707 * sibling can pick up load irrespective of busy siblings. In this case,
3708 * let the state of idle sibling percolate up as IDLE, instead of
3709 * portraying it as CPU_NOT_IDLE.
3711 if (sd->flags & SD_SHARE_CPUPOWER &&
3712 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3715 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3717 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3718 &sd_idle, cpus, NULL);
3720 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3724 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3726 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3730 BUG_ON(busiest == this_rq);
3732 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3735 if (busiest->nr_running > 1) {
3736 /* Attempt to move tasks */
3737 double_lock_balance(this_rq, busiest);
3738 /* this_rq->clock is already updated */
3739 update_rq_clock(busiest);
3740 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3741 imbalance, sd, CPU_NEWLY_IDLE,
3743 spin_unlock(&busiest->lock);
3745 if (unlikely(all_pinned)) {
3746 cpu_clear(cpu_of(busiest), *cpus);
3747 if (!cpus_empty(*cpus))
3753 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3754 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3755 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3758 sd->nr_balance_failed = 0;
3763 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3764 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3765 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3767 sd->nr_balance_failed = 0;
3773 * idle_balance is called by schedule() if this_cpu is about to become
3774 * idle. Attempts to pull tasks from other CPUs.
3776 static void idle_balance(int this_cpu, struct rq *this_rq)
3778 struct sched_domain *sd;
3779 int pulled_task = -1;
3780 unsigned long next_balance = jiffies + HZ;
3783 for_each_domain(this_cpu, sd) {
3784 unsigned long interval;
3786 if (!(sd->flags & SD_LOAD_BALANCE))
3789 if (sd->flags & SD_BALANCE_NEWIDLE)
3790 /* If we've pulled tasks over stop searching: */
3791 pulled_task = load_balance_newidle(this_cpu, this_rq,
3794 interval = msecs_to_jiffies(sd->balance_interval);
3795 if (time_after(next_balance, sd->last_balance + interval))
3796 next_balance = sd->last_balance + interval;
3800 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3802 * We are going idle. next_balance may be set based on
3803 * a busy processor. So reset next_balance.
3805 this_rq->next_balance = next_balance;
3810 * active_load_balance is run by migration threads. It pushes running tasks
3811 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3812 * running on each physical CPU where possible, and avoids physical /
3813 * logical imbalances.
3815 * Called with busiest_rq locked.
3817 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3819 int target_cpu = busiest_rq->push_cpu;
3820 struct sched_domain *sd;
3821 struct rq *target_rq;
3823 /* Is there any task to move? */
3824 if (busiest_rq->nr_running <= 1)
3827 target_rq = cpu_rq(target_cpu);
3830 * This condition is "impossible", if it occurs
3831 * we need to fix it. Originally reported by
3832 * Bjorn Helgaas on a 128-cpu setup.
3834 BUG_ON(busiest_rq == target_rq);
3836 /* move a task from busiest_rq to target_rq */
3837 double_lock_balance(busiest_rq, target_rq);
3838 update_rq_clock(busiest_rq);
3839 update_rq_clock(target_rq);
3841 /* Search for an sd spanning us and the target CPU. */
3842 for_each_domain(target_cpu, sd) {
3843 if ((sd->flags & SD_LOAD_BALANCE) &&
3844 cpu_isset(busiest_cpu, sd->span))
3849 schedstat_inc(sd, alb_count);
3851 if (move_one_task(target_rq, target_cpu, busiest_rq,
3853 schedstat_inc(sd, alb_pushed);
3855 schedstat_inc(sd, alb_failed);
3857 spin_unlock(&target_rq->lock);
3862 atomic_t load_balancer;
3864 } nohz ____cacheline_aligned = {
3865 .load_balancer = ATOMIC_INIT(-1),
3866 .cpu_mask = CPU_MASK_NONE,
3870 * This routine will try to nominate the ilb (idle load balancing)
3871 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3872 * load balancing on behalf of all those cpus. If all the cpus in the system
3873 * go into this tickless mode, then there will be no ilb owner (as there is
3874 * no need for one) and all the cpus will sleep till the next wakeup event
3877 * For the ilb owner, tick is not stopped. And this tick will be used
3878 * for idle load balancing. ilb owner will still be part of
3881 * While stopping the tick, this cpu will become the ilb owner if there
3882 * is no other owner. And will be the owner till that cpu becomes busy
3883 * or if all cpus in the system stop their ticks at which point
3884 * there is no need for ilb owner.
3886 * When the ilb owner becomes busy, it nominates another owner, during the
3887 * next busy scheduler_tick()
3889 int select_nohz_load_balancer(int stop_tick)
3891 int cpu = smp_processor_id();
3894 cpu_set(cpu, nohz.cpu_mask);
3895 cpu_rq(cpu)->in_nohz_recently = 1;
3898 * If we are going offline and still the leader, give up!
3900 if (cpu_is_offline(cpu) &&
3901 atomic_read(&nohz.load_balancer) == cpu) {
3902 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3907 /* time for ilb owner also to sleep */
3908 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3909 if (atomic_read(&nohz.load_balancer) == cpu)
3910 atomic_set(&nohz.load_balancer, -1);
3914 if (atomic_read(&nohz.load_balancer) == -1) {
3915 /* make me the ilb owner */
3916 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3918 } else if (atomic_read(&nohz.load_balancer) == cpu)
3921 if (!cpu_isset(cpu, nohz.cpu_mask))
3924 cpu_clear(cpu, nohz.cpu_mask);
3926 if (atomic_read(&nohz.load_balancer) == cpu)
3927 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3934 static DEFINE_SPINLOCK(balancing);
3937 * It checks each scheduling domain to see if it is due to be balanced,
3938 * and initiates a balancing operation if so.
3940 * Balancing parameters are set up in arch_init_sched_domains.
3942 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3945 struct rq *rq = cpu_rq(cpu);
3946 unsigned long interval;
3947 struct sched_domain *sd;
3948 /* Earliest time when we have to do rebalance again */
3949 unsigned long next_balance = jiffies + 60*HZ;
3950 int update_next_balance = 0;
3954 for_each_domain(cpu, sd) {
3955 if (!(sd->flags & SD_LOAD_BALANCE))
3958 interval = sd->balance_interval;
3959 if (idle != CPU_IDLE)
3960 interval *= sd->busy_factor;
3962 /* scale ms to jiffies */
3963 interval = msecs_to_jiffies(interval);
3964 if (unlikely(!interval))
3966 if (interval > HZ*NR_CPUS/10)
3967 interval = HZ*NR_CPUS/10;
3969 need_serialize = sd->flags & SD_SERIALIZE;
3971 if (need_serialize) {
3972 if (!spin_trylock(&balancing))
3976 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3977 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3979 * We've pulled tasks over so either we're no
3980 * longer idle, or one of our SMT siblings is
3983 idle = CPU_NOT_IDLE;
3985 sd->last_balance = jiffies;
3988 spin_unlock(&balancing);
3990 if (time_after(next_balance, sd->last_balance + interval)) {
3991 next_balance = sd->last_balance + interval;
3992 update_next_balance = 1;
3996 * Stop the load balance at this level. There is another
3997 * CPU in our sched group which is doing load balancing more
4005 * next_balance will be updated only when there is a need.
4006 * When the cpu is attached to null domain for ex, it will not be
4009 if (likely(update_next_balance))
4010 rq->next_balance = next_balance;
4014 * run_rebalance_domains is triggered when needed from the scheduler tick.
4015 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4016 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4018 static void run_rebalance_domains(struct softirq_action *h)
4020 int this_cpu = smp_processor_id();
4021 struct rq *this_rq = cpu_rq(this_cpu);
4022 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4023 CPU_IDLE : CPU_NOT_IDLE;
4025 rebalance_domains(this_cpu, idle);
4029 * If this cpu is the owner for idle load balancing, then do the
4030 * balancing on behalf of the other idle cpus whose ticks are
4033 if (this_rq->idle_at_tick &&
4034 atomic_read(&nohz.load_balancer) == this_cpu) {
4035 cpumask_t cpus = nohz.cpu_mask;
4039 cpu_clear(this_cpu, cpus);
4040 for_each_cpu_mask(balance_cpu, cpus) {
4042 * If this cpu gets work to do, stop the load balancing
4043 * work being done for other cpus. Next load
4044 * balancing owner will pick it up.
4049 rebalance_domains(balance_cpu, CPU_IDLE);
4051 rq = cpu_rq(balance_cpu);
4052 if (time_after(this_rq->next_balance, rq->next_balance))
4053 this_rq->next_balance = rq->next_balance;
4060 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4062 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4063 * idle load balancing owner or decide to stop the periodic load balancing,
4064 * if the whole system is idle.
4066 static inline void trigger_load_balance(struct rq *rq, int cpu)
4070 * If we were in the nohz mode recently and busy at the current
4071 * scheduler tick, then check if we need to nominate new idle
4074 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4075 rq->in_nohz_recently = 0;
4077 if (atomic_read(&nohz.load_balancer) == cpu) {
4078 cpu_clear(cpu, nohz.cpu_mask);
4079 atomic_set(&nohz.load_balancer, -1);
4082 if (atomic_read(&nohz.load_balancer) == -1) {
4084 * simple selection for now: Nominate the
4085 * first cpu in the nohz list to be the next
4088 * TBD: Traverse the sched domains and nominate
4089 * the nearest cpu in the nohz.cpu_mask.
4091 int ilb = first_cpu(nohz.cpu_mask);
4093 if (ilb < nr_cpu_ids)
4099 * If this cpu is idle and doing idle load balancing for all the
4100 * cpus with ticks stopped, is it time for that to stop?
4102 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4103 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4109 * If this cpu is idle and the idle load balancing is done by
4110 * someone else, then no need raise the SCHED_SOFTIRQ
4112 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4113 cpu_isset(cpu, nohz.cpu_mask))
4116 if (time_after_eq(jiffies, rq->next_balance))
4117 raise_softirq(SCHED_SOFTIRQ);
4120 #else /* CONFIG_SMP */
4123 * on UP we do not need to balance between CPUs:
4125 static inline void idle_balance(int cpu, struct rq *rq)
4131 DEFINE_PER_CPU(struct kernel_stat, kstat);
4133 EXPORT_PER_CPU_SYMBOL(kstat);
4136 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4137 * that have not yet been banked in case the task is currently running.
4139 unsigned long long task_sched_runtime(struct task_struct *p)
4141 unsigned long flags;
4145 rq = task_rq_lock(p, &flags);
4146 ns = p->se.sum_exec_runtime;
4147 if (task_current(rq, p)) {
4148 update_rq_clock(rq);
4149 delta_exec = rq->clock - p->se.exec_start;
4150 if ((s64)delta_exec > 0)
4153 task_rq_unlock(rq, &flags);
4159 * Account user cpu time to a process.
4160 * @p: the process that the cpu time gets accounted to
4161 * @cputime: the cpu time spent in user space since the last update
4163 void account_user_time(struct task_struct *p, cputime_t cputime)
4165 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4168 p->utime = cputime_add(p->utime, cputime);
4170 /* Add user time to cpustat. */
4171 tmp = cputime_to_cputime64(cputime);
4172 if (TASK_NICE(p) > 0)
4173 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4175 cpustat->user = cputime64_add(cpustat->user, tmp);
4179 * Account guest cpu time to a process.
4180 * @p: the process that the cpu time gets accounted to
4181 * @cputime: the cpu time spent in virtual machine since the last update
4183 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4186 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4188 tmp = cputime_to_cputime64(cputime);
4190 p->utime = cputime_add(p->utime, cputime);
4191 p->gtime = cputime_add(p->gtime, cputime);
4193 cpustat->user = cputime64_add(cpustat->user, tmp);
4194 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4198 * Account scaled user cpu time to a process.
4199 * @p: the process that the cpu time gets accounted to
4200 * @cputime: the cpu time spent in user space since the last update
4202 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4204 p->utimescaled = cputime_add(p->utimescaled, cputime);
4208 * Account system cpu time to a process.
4209 * @p: the process that the cpu time gets accounted to
4210 * @hardirq_offset: the offset to subtract from hardirq_count()
4211 * @cputime: the cpu time spent in kernel space since the last update
4213 void account_system_time(struct task_struct *p, int hardirq_offset,
4216 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4217 struct rq *rq = this_rq();
4220 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4221 account_guest_time(p, cputime);
4225 p->stime = cputime_add(p->stime, cputime);
4227 /* Add system time to cpustat. */
4228 tmp = cputime_to_cputime64(cputime);
4229 if (hardirq_count() - hardirq_offset)
4230 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4231 else if (softirq_count())
4232 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4233 else if (p != rq->idle)
4234 cpustat->system = cputime64_add(cpustat->system, tmp);
4235 else if (atomic_read(&rq->nr_iowait) > 0)
4236 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4238 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4239 /* Account for system time used */
4240 acct_update_integrals(p);
4244 * Account scaled system cpu time to a process.
4245 * @p: the process that the cpu time gets accounted to
4246 * @hardirq_offset: the offset to subtract from hardirq_count()
4247 * @cputime: the cpu time spent in kernel space since the last update
4249 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4251 p->stimescaled = cputime_add(p->stimescaled, cputime);
4255 * Account for involuntary wait time.
4256 * @p: the process from which the cpu time has been stolen
4257 * @steal: the cpu time spent in involuntary wait
4259 void account_steal_time(struct task_struct *p, cputime_t steal)
4261 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4262 cputime64_t tmp = cputime_to_cputime64(steal);
4263 struct rq *rq = this_rq();
4265 if (p == rq->idle) {
4266 p->stime = cputime_add(p->stime, steal);
4267 if (atomic_read(&rq->nr_iowait) > 0)
4268 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4270 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4272 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4276 * This function gets called by the timer code, with HZ frequency.
4277 * We call it with interrupts disabled.
4279 * It also gets called by the fork code, when changing the parent's
4282 void scheduler_tick(void)
4284 int cpu = smp_processor_id();
4285 struct rq *rq = cpu_rq(cpu);
4286 struct task_struct *curr = rq->curr;
4290 spin_lock(&rq->lock);
4291 update_rq_clock(rq);
4292 update_cpu_load(rq);
4293 curr->sched_class->task_tick(rq, curr, 0);
4294 spin_unlock(&rq->lock);
4297 rq->idle_at_tick = idle_cpu(cpu);
4298 trigger_load_balance(rq, cpu);
4302 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4304 void __kprobes add_preempt_count(int val)
4309 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4311 preempt_count() += val;
4313 * Spinlock count overflowing soon?
4315 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4318 EXPORT_SYMBOL(add_preempt_count);
4320 void __kprobes sub_preempt_count(int val)
4325 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4328 * Is the spinlock portion underflowing?
4330 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4331 !(preempt_count() & PREEMPT_MASK)))
4334 preempt_count() -= val;
4336 EXPORT_SYMBOL(sub_preempt_count);
4341 * Print scheduling while atomic bug:
4343 static noinline void __schedule_bug(struct task_struct *prev)
4345 struct pt_regs *regs = get_irq_regs();
4347 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4348 prev->comm, prev->pid, preempt_count());
4350 debug_show_held_locks(prev);
4352 if (irqs_disabled())
4353 print_irqtrace_events(prev);
4362 * Various schedule()-time debugging checks and statistics:
4364 static inline void schedule_debug(struct task_struct *prev)
4367 * Test if we are atomic. Since do_exit() needs to call into
4368 * schedule() atomically, we ignore that path for now.
4369 * Otherwise, whine if we are scheduling when we should not be.
4371 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4372 __schedule_bug(prev);
4374 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4376 schedstat_inc(this_rq(), sched_count);
4377 #ifdef CONFIG_SCHEDSTATS
4378 if (unlikely(prev->lock_depth >= 0)) {
4379 schedstat_inc(this_rq(), bkl_count);
4380 schedstat_inc(prev, sched_info.bkl_count);
4386 * Pick up the highest-prio task:
4388 static inline struct task_struct *
4389 pick_next_task(struct rq *rq, struct task_struct *prev)
4391 const struct sched_class *class;
4392 struct task_struct *p;
4395 * Optimization: we know that if all tasks are in
4396 * the fair class we can call that function directly:
4398 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4399 p = fair_sched_class.pick_next_task(rq);
4404 class = sched_class_highest;
4406 p = class->pick_next_task(rq);
4410 * Will never be NULL as the idle class always
4411 * returns a non-NULL p:
4413 class = class->next;
4418 * schedule() is the main scheduler function.
4420 asmlinkage void __sched schedule(void)
4422 struct task_struct *prev, *next;
4423 unsigned long *switch_count;
4425 int cpu, hrtick = sched_feat(HRTICK);
4429 cpu = smp_processor_id();
4433 switch_count = &prev->nivcsw;
4435 release_kernel_lock(prev);
4436 need_resched_nonpreemptible:
4438 schedule_debug(prev);
4444 * Do the rq-clock update outside the rq lock:
4446 local_irq_disable();
4447 update_rq_clock(rq);
4448 spin_lock(&rq->lock);
4449 clear_tsk_need_resched(prev);
4451 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4452 if (unlikely(signal_pending_state(prev->state, prev)))
4453 prev->state = TASK_RUNNING;
4455 deactivate_task(rq, prev, 1);
4456 switch_count = &prev->nvcsw;
4460 if (prev->sched_class->pre_schedule)
4461 prev->sched_class->pre_schedule(rq, prev);
4464 if (unlikely(!rq->nr_running))
4465 idle_balance(cpu, rq);
4467 prev->sched_class->put_prev_task(rq, prev);
4468 next = pick_next_task(rq, prev);
4470 if (likely(prev != next)) {
4471 sched_info_switch(prev, next);
4477 context_switch(rq, prev, next); /* unlocks the rq */
4479 * the context switch might have flipped the stack from under
4480 * us, hence refresh the local variables.
4482 cpu = smp_processor_id();
4485 spin_unlock_irq(&rq->lock);
4490 if (unlikely(reacquire_kernel_lock(current) < 0))
4491 goto need_resched_nonpreemptible;
4493 preempt_enable_no_resched();
4494 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4497 EXPORT_SYMBOL(schedule);
4499 #ifdef CONFIG_PREEMPT
4501 * this is the entry point to schedule() from in-kernel preemption
4502 * off of preempt_enable. Kernel preemptions off return from interrupt
4503 * occur there and call schedule directly.
4505 asmlinkage void __sched preempt_schedule(void)
4507 struct thread_info *ti = current_thread_info();
4510 * If there is a non-zero preempt_count or interrupts are disabled,
4511 * we do not want to preempt the current task. Just return..
4513 if (likely(ti->preempt_count || irqs_disabled()))
4517 add_preempt_count(PREEMPT_ACTIVE);
4519 sub_preempt_count(PREEMPT_ACTIVE);
4522 * Check again in case we missed a preemption opportunity
4523 * between schedule and now.
4526 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4528 EXPORT_SYMBOL(preempt_schedule);
4531 * this is the entry point to schedule() from kernel preemption
4532 * off of irq context.
4533 * Note, that this is called and return with irqs disabled. This will
4534 * protect us against recursive calling from irq.
4536 asmlinkage void __sched preempt_schedule_irq(void)
4538 struct thread_info *ti = current_thread_info();
4540 /* Catch callers which need to be fixed */
4541 BUG_ON(ti->preempt_count || !irqs_disabled());
4544 add_preempt_count(PREEMPT_ACTIVE);
4547 local_irq_disable();
4548 sub_preempt_count(PREEMPT_ACTIVE);
4551 * Check again in case we missed a preemption opportunity
4552 * between schedule and now.
4555 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4558 #endif /* CONFIG_PREEMPT */
4560 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4563 return try_to_wake_up(curr->private, mode, sync);
4565 EXPORT_SYMBOL(default_wake_function);
4568 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4569 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4570 * number) then we wake all the non-exclusive tasks and one exclusive task.
4572 * There are circumstances in which we can try to wake a task which has already
4573 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4574 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4576 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4577 int nr_exclusive, int sync, void *key)
4579 wait_queue_t *curr, *next;
4581 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4582 unsigned flags = curr->flags;
4584 if (curr->func(curr, mode, sync, key) &&
4585 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4591 * __wake_up - wake up threads blocked on a waitqueue.
4593 * @mode: which threads
4594 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4595 * @key: is directly passed to the wakeup function
4597 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4598 int nr_exclusive, void *key)
4600 unsigned long flags;
4602 spin_lock_irqsave(&q->lock, flags);
4603 __wake_up_common(q, mode, nr_exclusive, 0, key);
4604 spin_unlock_irqrestore(&q->lock, flags);
4606 EXPORT_SYMBOL(__wake_up);
4609 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4611 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4613 __wake_up_common(q, mode, 1, 0, NULL);
4617 * __wake_up_sync - wake up threads blocked on a waitqueue.
4619 * @mode: which threads
4620 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4622 * The sync wakeup differs that the waker knows that it will schedule
4623 * away soon, so while the target thread will be woken up, it will not
4624 * be migrated to another CPU - ie. the two threads are 'synchronized'
4625 * with each other. This can prevent needless bouncing between CPUs.
4627 * On UP it can prevent extra preemption.
4630 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4632 unsigned long flags;
4638 if (unlikely(!nr_exclusive))
4641 spin_lock_irqsave(&q->lock, flags);
4642 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4643 spin_unlock_irqrestore(&q->lock, flags);
4645 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4647 void complete(struct completion *x)
4649 unsigned long flags;
4651 spin_lock_irqsave(&x->wait.lock, flags);
4653 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4654 spin_unlock_irqrestore(&x->wait.lock, flags);
4656 EXPORT_SYMBOL(complete);
4658 void complete_all(struct completion *x)
4660 unsigned long flags;
4662 spin_lock_irqsave(&x->wait.lock, flags);
4663 x->done += UINT_MAX/2;
4664 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4665 spin_unlock_irqrestore(&x->wait.lock, flags);
4667 EXPORT_SYMBOL(complete_all);
4669 static inline long __sched
4670 do_wait_for_common(struct completion *x, long timeout, int state)
4673 DECLARE_WAITQUEUE(wait, current);
4675 wait.flags |= WQ_FLAG_EXCLUSIVE;
4676 __add_wait_queue_tail(&x->wait, &wait);
4678 if ((state == TASK_INTERRUPTIBLE &&
4679 signal_pending(current)) ||
4680 (state == TASK_KILLABLE &&
4681 fatal_signal_pending(current))) {
4682 timeout = -ERESTARTSYS;
4685 __set_current_state(state);
4686 spin_unlock_irq(&x->wait.lock);
4687 timeout = schedule_timeout(timeout);
4688 spin_lock_irq(&x->wait.lock);
4689 } while (!x->done && timeout);
4690 __remove_wait_queue(&x->wait, &wait);
4695 return timeout ?: 1;
4699 wait_for_common(struct completion *x, long timeout, int state)
4703 spin_lock_irq(&x->wait.lock);
4704 timeout = do_wait_for_common(x, timeout, state);
4705 spin_unlock_irq(&x->wait.lock);
4709 void __sched wait_for_completion(struct completion *x)
4711 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4713 EXPORT_SYMBOL(wait_for_completion);
4715 unsigned long __sched
4716 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4718 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4720 EXPORT_SYMBOL(wait_for_completion_timeout);
4722 int __sched wait_for_completion_interruptible(struct completion *x)
4724 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4725 if (t == -ERESTARTSYS)
4729 EXPORT_SYMBOL(wait_for_completion_interruptible);
4731 unsigned long __sched
4732 wait_for_completion_interruptible_timeout(struct completion *x,
4733 unsigned long timeout)
4735 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4737 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4739 int __sched wait_for_completion_killable(struct completion *x)
4741 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4742 if (t == -ERESTARTSYS)
4746 EXPORT_SYMBOL(wait_for_completion_killable);
4749 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4751 unsigned long flags;
4754 init_waitqueue_entry(&wait, current);
4756 __set_current_state(state);
4758 spin_lock_irqsave(&q->lock, flags);
4759 __add_wait_queue(q, &wait);
4760 spin_unlock(&q->lock);
4761 timeout = schedule_timeout(timeout);
4762 spin_lock_irq(&q->lock);
4763 __remove_wait_queue(q, &wait);
4764 spin_unlock_irqrestore(&q->lock, flags);
4769 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4771 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4773 EXPORT_SYMBOL(interruptible_sleep_on);
4776 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4778 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4780 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4782 void __sched sleep_on(wait_queue_head_t *q)
4784 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4786 EXPORT_SYMBOL(sleep_on);
4788 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4790 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4792 EXPORT_SYMBOL(sleep_on_timeout);
4794 #ifdef CONFIG_RT_MUTEXES
4797 * rt_mutex_setprio - set the current priority of a task
4799 * @prio: prio value (kernel-internal form)
4801 * This function changes the 'effective' priority of a task. It does
4802 * not touch ->normal_prio like __setscheduler().
4804 * Used by the rt_mutex code to implement priority inheritance logic.
4806 void rt_mutex_setprio(struct task_struct *p, int prio)
4808 unsigned long flags;
4809 int oldprio, on_rq, running;
4811 const struct sched_class *prev_class = p->sched_class;
4813 BUG_ON(prio < 0 || prio > MAX_PRIO);
4815 rq = task_rq_lock(p, &flags);
4816 update_rq_clock(rq);
4819 on_rq = p->se.on_rq;
4820 running = task_current(rq, p);
4822 dequeue_task(rq, p, 0);
4824 p->sched_class->put_prev_task(rq, p);
4827 p->sched_class = &rt_sched_class;
4829 p->sched_class = &fair_sched_class;
4834 p->sched_class->set_curr_task(rq);
4836 enqueue_task(rq, p, 0);
4838 check_class_changed(rq, p, prev_class, oldprio, running);
4840 task_rq_unlock(rq, &flags);
4845 void set_user_nice(struct task_struct *p, long nice)
4847 int old_prio, delta, on_rq;
4848 unsigned long flags;
4851 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4854 * We have to be careful, if called from sys_setpriority(),
4855 * the task might be in the middle of scheduling on another CPU.
4857 rq = task_rq_lock(p, &flags);
4858 update_rq_clock(rq);
4860 * The RT priorities are set via sched_setscheduler(), but we still
4861 * allow the 'normal' nice value to be set - but as expected
4862 * it wont have any effect on scheduling until the task is
4863 * SCHED_FIFO/SCHED_RR:
4865 if (task_has_rt_policy(p)) {
4866 p->static_prio = NICE_TO_PRIO(nice);
4869 on_rq = p->se.on_rq;
4871 dequeue_task(rq, p, 0);
4873 p->static_prio = NICE_TO_PRIO(nice);
4876 p->prio = effective_prio(p);
4877 delta = p->prio - old_prio;
4880 enqueue_task(rq, p, 0);
4882 * If the task increased its priority or is running and
4883 * lowered its priority, then reschedule its CPU:
4885 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4886 resched_task(rq->curr);
4889 task_rq_unlock(rq, &flags);
4891 EXPORT_SYMBOL(set_user_nice);
4894 * can_nice - check if a task can reduce its nice value
4898 int can_nice(const struct task_struct *p, const int nice)
4900 /* convert nice value [19,-20] to rlimit style value [1,40] */
4901 int nice_rlim = 20 - nice;
4903 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4904 capable(CAP_SYS_NICE));
4907 #ifdef __ARCH_WANT_SYS_NICE
4910 * sys_nice - change the priority of the current process.
4911 * @increment: priority increment
4913 * sys_setpriority is a more generic, but much slower function that
4914 * does similar things.
4916 asmlinkage long sys_nice(int increment)
4921 * Setpriority might change our priority at the same moment.
4922 * We don't have to worry. Conceptually one call occurs first
4923 * and we have a single winner.
4925 if (increment < -40)
4930 nice = PRIO_TO_NICE(current->static_prio) + increment;
4936 if (increment < 0 && !can_nice(current, nice))
4939 retval = security_task_setnice(current, nice);
4943 set_user_nice(current, nice);
4950 * task_prio - return the priority value of a given task.
4951 * @p: the task in question.
4953 * This is the priority value as seen by users in /proc.
4954 * RT tasks are offset by -200. Normal tasks are centered
4955 * around 0, value goes from -16 to +15.
4957 int task_prio(const struct task_struct *p)
4959 return p->prio - MAX_RT_PRIO;
4963 * task_nice - return the nice value of a given task.
4964 * @p: the task in question.
4966 int task_nice(const struct task_struct *p)
4968 return TASK_NICE(p);
4970 EXPORT_SYMBOL(task_nice);
4973 * idle_cpu - is a given cpu idle currently?
4974 * @cpu: the processor in question.
4976 int idle_cpu(int cpu)
4978 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4982 * idle_task - return the idle task for a given cpu.
4983 * @cpu: the processor in question.
4985 struct task_struct *idle_task(int cpu)
4987 return cpu_rq(cpu)->idle;
4991 * find_process_by_pid - find a process with a matching PID value.
4992 * @pid: the pid in question.
4994 static struct task_struct *find_process_by_pid(pid_t pid)
4996 return pid ? find_task_by_vpid(pid) : current;
4999 /* Actually do priority change: must hold rq lock. */
5001 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5003 BUG_ON(p->se.on_rq);
5006 switch (p->policy) {
5010 p->sched_class = &fair_sched_class;
5014 p->sched_class = &rt_sched_class;
5018 p->rt_priority = prio;
5019 p->normal_prio = normal_prio(p);
5020 /* we are holding p->pi_lock already */
5021 p->prio = rt_mutex_getprio(p);
5026 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5027 * @p: the task in question.
5028 * @policy: new policy.
5029 * @param: structure containing the new RT priority.
5031 * NOTE that the task may be already dead.
5033 int sched_setscheduler(struct task_struct *p, int policy,
5034 struct sched_param *param)
5036 int retval, oldprio, oldpolicy = -1, on_rq, running;
5037 unsigned long flags;
5038 const struct sched_class *prev_class = p->sched_class;
5041 /* may grab non-irq protected spin_locks */
5042 BUG_ON(in_interrupt());
5044 /* double check policy once rq lock held */
5046 policy = oldpolicy = p->policy;
5047 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5048 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5049 policy != SCHED_IDLE)
5052 * Valid priorities for SCHED_FIFO and SCHED_RR are
5053 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5054 * SCHED_BATCH and SCHED_IDLE is 0.
5056 if (param->sched_priority < 0 ||
5057 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5058 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5060 if (rt_policy(policy) != (param->sched_priority != 0))
5064 * Allow unprivileged RT tasks to decrease priority:
5066 if (!capable(CAP_SYS_NICE)) {
5067 if (rt_policy(policy)) {
5068 unsigned long rlim_rtprio;
5070 if (!lock_task_sighand(p, &flags))
5072 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5073 unlock_task_sighand(p, &flags);
5075 /* can't set/change the rt policy */
5076 if (policy != p->policy && !rlim_rtprio)
5079 /* can't increase priority */
5080 if (param->sched_priority > p->rt_priority &&
5081 param->sched_priority > rlim_rtprio)
5085 * Like positive nice levels, dont allow tasks to
5086 * move out of SCHED_IDLE either:
5088 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5091 /* can't change other user's priorities */
5092 if ((current->euid != p->euid) &&
5093 (current->euid != p->uid))
5097 #ifdef CONFIG_RT_GROUP_SCHED
5099 * Do not allow realtime tasks into groups that have no runtime
5102 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5106 retval = security_task_setscheduler(p, policy, param);
5110 * make sure no PI-waiters arrive (or leave) while we are
5111 * changing the priority of the task:
5113 spin_lock_irqsave(&p->pi_lock, flags);
5115 * To be able to change p->policy safely, the apropriate
5116 * runqueue lock must be held.
5118 rq = __task_rq_lock(p);
5119 /* recheck policy now with rq lock held */
5120 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5121 policy = oldpolicy = -1;
5122 __task_rq_unlock(rq);
5123 spin_unlock_irqrestore(&p->pi_lock, flags);
5126 update_rq_clock(rq);
5127 on_rq = p->se.on_rq;
5128 running = task_current(rq, p);
5130 deactivate_task(rq, p, 0);
5132 p->sched_class->put_prev_task(rq, p);
5135 __setscheduler(rq, p, policy, param->sched_priority);
5138 p->sched_class->set_curr_task(rq);
5140 activate_task(rq, p, 0);
5142 check_class_changed(rq, p, prev_class, oldprio, running);
5144 __task_rq_unlock(rq);
5145 spin_unlock_irqrestore(&p->pi_lock, flags);
5147 rt_mutex_adjust_pi(p);
5151 EXPORT_SYMBOL_GPL(sched_setscheduler);
5154 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5156 struct sched_param lparam;
5157 struct task_struct *p;
5160 if (!param || pid < 0)
5162 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5167 p = find_process_by_pid(pid);
5169 retval = sched_setscheduler(p, policy, &lparam);
5176 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5177 * @pid: the pid in question.
5178 * @policy: new policy.
5179 * @param: structure containing the new RT priority.
5182 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5184 /* negative values for policy are not valid */
5188 return do_sched_setscheduler(pid, policy, param);
5192 * sys_sched_setparam - set/change the RT priority of a thread
5193 * @pid: the pid in question.
5194 * @param: structure containing the new RT priority.
5196 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5198 return do_sched_setscheduler(pid, -1, param);
5202 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5203 * @pid: the pid in question.
5205 asmlinkage long sys_sched_getscheduler(pid_t pid)
5207 struct task_struct *p;
5214 read_lock(&tasklist_lock);
5215 p = find_process_by_pid(pid);
5217 retval = security_task_getscheduler(p);
5221 read_unlock(&tasklist_lock);
5226 * sys_sched_getscheduler - get the RT priority of a thread
5227 * @pid: the pid in question.
5228 * @param: structure containing the RT priority.
5230 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5232 struct sched_param lp;
5233 struct task_struct *p;
5236 if (!param || pid < 0)
5239 read_lock(&tasklist_lock);
5240 p = find_process_by_pid(pid);
5245 retval = security_task_getscheduler(p);
5249 lp.sched_priority = p->rt_priority;
5250 read_unlock(&tasklist_lock);
5253 * This one might sleep, we cannot do it with a spinlock held ...
5255 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5260 read_unlock(&tasklist_lock);
5264 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5266 cpumask_t cpus_allowed;
5267 cpumask_t new_mask = *in_mask;
5268 struct task_struct *p;
5272 read_lock(&tasklist_lock);
5274 p = find_process_by_pid(pid);
5276 read_unlock(&tasklist_lock);
5282 * It is not safe to call set_cpus_allowed with the
5283 * tasklist_lock held. We will bump the task_struct's
5284 * usage count and then drop tasklist_lock.
5287 read_unlock(&tasklist_lock);
5290 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5291 !capable(CAP_SYS_NICE))
5294 retval = security_task_setscheduler(p, 0, NULL);
5298 cpuset_cpus_allowed(p, &cpus_allowed);
5299 cpus_and(new_mask, new_mask, cpus_allowed);
5301 retval = set_cpus_allowed_ptr(p, &new_mask);
5304 cpuset_cpus_allowed(p, &cpus_allowed);
5305 if (!cpus_subset(new_mask, cpus_allowed)) {
5307 * We must have raced with a concurrent cpuset
5308 * update. Just reset the cpus_allowed to the
5309 * cpuset's cpus_allowed
5311 new_mask = cpus_allowed;
5321 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5322 cpumask_t *new_mask)
5324 if (len < sizeof(cpumask_t)) {
5325 memset(new_mask, 0, sizeof(cpumask_t));
5326 } else if (len > sizeof(cpumask_t)) {
5327 len = sizeof(cpumask_t);
5329 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5333 * sys_sched_setaffinity - set the cpu affinity of a process
5334 * @pid: pid of the process
5335 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5336 * @user_mask_ptr: user-space pointer to the new cpu mask
5338 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5339 unsigned long __user *user_mask_ptr)
5344 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5348 return sched_setaffinity(pid, &new_mask);
5351 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5353 struct task_struct *p;
5357 read_lock(&tasklist_lock);
5360 p = find_process_by_pid(pid);
5364 retval = security_task_getscheduler(p);
5368 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5371 read_unlock(&tasklist_lock);
5378 * sys_sched_getaffinity - get the cpu affinity of a process
5379 * @pid: pid of the process
5380 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5381 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5383 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5384 unsigned long __user *user_mask_ptr)
5389 if (len < sizeof(cpumask_t))
5392 ret = sched_getaffinity(pid, &mask);
5396 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5399 return sizeof(cpumask_t);
5403 * sys_sched_yield - yield the current processor to other threads.
5405 * This function yields the current CPU to other tasks. If there are no
5406 * other threads running on this CPU then this function will return.
5408 asmlinkage long sys_sched_yield(void)
5410 struct rq *rq = this_rq_lock();
5412 schedstat_inc(rq, yld_count);
5413 current->sched_class->yield_task(rq);
5416 * Since we are going to call schedule() anyway, there's
5417 * no need to preempt or enable interrupts:
5419 __release(rq->lock);
5420 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5421 _raw_spin_unlock(&rq->lock);
5422 preempt_enable_no_resched();
5429 static void __cond_resched(void)
5431 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5432 __might_sleep(__FILE__, __LINE__);
5435 * The BKS might be reacquired before we have dropped
5436 * PREEMPT_ACTIVE, which could trigger a second
5437 * cond_resched() call.
5440 add_preempt_count(PREEMPT_ACTIVE);
5442 sub_preempt_count(PREEMPT_ACTIVE);
5443 } while (need_resched());
5446 int __sched _cond_resched(void)
5448 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5449 system_state == SYSTEM_RUNNING) {
5455 EXPORT_SYMBOL(_cond_resched);
5458 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5459 * call schedule, and on return reacquire the lock.
5461 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5462 * operations here to prevent schedule() from being called twice (once via
5463 * spin_unlock(), once by hand).
5465 int cond_resched_lock(spinlock_t *lock)
5467 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5470 if (spin_needbreak(lock) || resched) {
5472 if (resched && need_resched())
5481 EXPORT_SYMBOL(cond_resched_lock);
5483 int __sched cond_resched_softirq(void)
5485 BUG_ON(!in_softirq());
5487 if (need_resched() && system_state == SYSTEM_RUNNING) {
5495 EXPORT_SYMBOL(cond_resched_softirq);
5498 * yield - yield the current processor to other threads.
5500 * This is a shortcut for kernel-space yielding - it marks the
5501 * thread runnable and calls sys_sched_yield().
5503 void __sched yield(void)
5505 set_current_state(TASK_RUNNING);
5508 EXPORT_SYMBOL(yield);
5511 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5512 * that process accounting knows that this is a task in IO wait state.
5514 * But don't do that if it is a deliberate, throttling IO wait (this task
5515 * has set its backing_dev_info: the queue against which it should throttle)
5517 void __sched io_schedule(void)
5519 struct rq *rq = &__raw_get_cpu_var(runqueues);
5521 delayacct_blkio_start();
5522 atomic_inc(&rq->nr_iowait);
5524 atomic_dec(&rq->nr_iowait);
5525 delayacct_blkio_end();
5527 EXPORT_SYMBOL(io_schedule);
5529 long __sched io_schedule_timeout(long timeout)
5531 struct rq *rq = &__raw_get_cpu_var(runqueues);
5534 delayacct_blkio_start();
5535 atomic_inc(&rq->nr_iowait);
5536 ret = schedule_timeout(timeout);
5537 atomic_dec(&rq->nr_iowait);
5538 delayacct_blkio_end();
5543 * sys_sched_get_priority_max - return maximum RT priority.
5544 * @policy: scheduling class.
5546 * this syscall returns the maximum rt_priority that can be used
5547 * by a given scheduling class.
5549 asmlinkage long sys_sched_get_priority_max(int policy)
5556 ret = MAX_USER_RT_PRIO-1;
5568 * sys_sched_get_priority_min - return minimum RT priority.
5569 * @policy: scheduling class.
5571 * this syscall returns the minimum rt_priority that can be used
5572 * by a given scheduling class.
5574 asmlinkage long sys_sched_get_priority_min(int policy)
5592 * sys_sched_rr_get_interval - return the default timeslice of a process.
5593 * @pid: pid of the process.
5594 * @interval: userspace pointer to the timeslice value.
5596 * this syscall writes the default timeslice value of a given process
5597 * into the user-space timespec buffer. A value of '0' means infinity.
5600 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5602 struct task_struct *p;
5603 unsigned int time_slice;
5611 read_lock(&tasklist_lock);
5612 p = find_process_by_pid(pid);
5616 retval = security_task_getscheduler(p);
5621 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5622 * tasks that are on an otherwise idle runqueue:
5625 if (p->policy == SCHED_RR) {
5626 time_slice = DEF_TIMESLICE;
5627 } else if (p->policy != SCHED_FIFO) {
5628 struct sched_entity *se = &p->se;
5629 unsigned long flags;
5632 rq = task_rq_lock(p, &flags);
5633 if (rq->cfs.load.weight)
5634 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5635 task_rq_unlock(rq, &flags);
5637 read_unlock(&tasklist_lock);
5638 jiffies_to_timespec(time_slice, &t);
5639 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5643 read_unlock(&tasklist_lock);
5647 static const char stat_nam[] = "RSDTtZX";
5649 void sched_show_task(struct task_struct *p)
5651 unsigned long free = 0;
5654 state = p->state ? __ffs(p->state) + 1 : 0;
5655 printk(KERN_INFO "%-13.13s %c", p->comm,
5656 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5657 #if BITS_PER_LONG == 32
5658 if (state == TASK_RUNNING)
5659 printk(KERN_CONT " running ");
5661 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5663 if (state == TASK_RUNNING)
5664 printk(KERN_CONT " running task ");
5666 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5668 #ifdef CONFIG_DEBUG_STACK_USAGE
5670 unsigned long *n = end_of_stack(p);
5673 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5676 printk(KERN_CONT "%5lu %5d %6d\n", free,
5677 task_pid_nr(p), task_pid_nr(p->real_parent));
5679 show_stack(p, NULL);
5682 void show_state_filter(unsigned long state_filter)
5684 struct task_struct *g, *p;
5686 #if BITS_PER_LONG == 32
5688 " task PC stack pid father\n");
5691 " task PC stack pid father\n");
5693 read_lock(&tasklist_lock);
5694 do_each_thread(g, p) {
5696 * reset the NMI-timeout, listing all files on a slow
5697 * console might take alot of time:
5699 touch_nmi_watchdog();
5700 if (!state_filter || (p->state & state_filter))
5702 } while_each_thread(g, p);
5704 touch_all_softlockup_watchdogs();
5706 #ifdef CONFIG_SCHED_DEBUG
5707 sysrq_sched_debug_show();
5709 read_unlock(&tasklist_lock);
5711 * Only show locks if all tasks are dumped:
5713 if (state_filter == -1)
5714 debug_show_all_locks();
5717 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5719 idle->sched_class = &idle_sched_class;
5723 * init_idle - set up an idle thread for a given CPU
5724 * @idle: task in question
5725 * @cpu: cpu the idle task belongs to
5727 * NOTE: this function does not set the idle thread's NEED_RESCHED
5728 * flag, to make booting more robust.
5730 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5732 struct rq *rq = cpu_rq(cpu);
5733 unsigned long flags;
5736 idle->se.exec_start = sched_clock();
5738 idle->prio = idle->normal_prio = MAX_PRIO;
5739 idle->cpus_allowed = cpumask_of_cpu(cpu);
5740 __set_task_cpu(idle, cpu);
5742 spin_lock_irqsave(&rq->lock, flags);
5743 rq->curr = rq->idle = idle;
5744 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5747 spin_unlock_irqrestore(&rq->lock, flags);
5749 /* Set the preempt count _outside_ the spinlocks! */
5750 #if defined(CONFIG_PREEMPT)
5751 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5753 task_thread_info(idle)->preempt_count = 0;
5756 * The idle tasks have their own, simple scheduling class:
5758 idle->sched_class = &idle_sched_class;
5762 * In a system that switches off the HZ timer nohz_cpu_mask
5763 * indicates which cpus entered this state. This is used
5764 * in the rcu update to wait only for active cpus. For system
5765 * which do not switch off the HZ timer nohz_cpu_mask should
5766 * always be CPU_MASK_NONE.
5768 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5771 * Increase the granularity value when there are more CPUs,
5772 * because with more CPUs the 'effective latency' as visible
5773 * to users decreases. But the relationship is not linear,
5774 * so pick a second-best guess by going with the log2 of the
5777 * This idea comes from the SD scheduler of Con Kolivas:
5779 static inline void sched_init_granularity(void)
5781 unsigned int factor = 1 + ilog2(num_online_cpus());
5782 const unsigned long limit = 200000000;
5784 sysctl_sched_min_granularity *= factor;
5785 if (sysctl_sched_min_granularity > limit)
5786 sysctl_sched_min_granularity = limit;
5788 sysctl_sched_latency *= factor;
5789 if (sysctl_sched_latency > limit)
5790 sysctl_sched_latency = limit;
5792 sysctl_sched_wakeup_granularity *= factor;
5797 * This is how migration works:
5799 * 1) we queue a struct migration_req structure in the source CPU's
5800 * runqueue and wake up that CPU's migration thread.
5801 * 2) we down() the locked semaphore => thread blocks.
5802 * 3) migration thread wakes up (implicitly it forces the migrated
5803 * thread off the CPU)
5804 * 4) it gets the migration request and checks whether the migrated
5805 * task is still in the wrong runqueue.
5806 * 5) if it's in the wrong runqueue then the migration thread removes
5807 * it and puts it into the right queue.
5808 * 6) migration thread up()s the semaphore.
5809 * 7) we wake up and the migration is done.
5813 * Change a given task's CPU affinity. Migrate the thread to a
5814 * proper CPU and schedule it away if the CPU it's executing on
5815 * is removed from the allowed bitmask.
5817 * NOTE: the caller must have a valid reference to the task, the
5818 * task must not exit() & deallocate itself prematurely. The
5819 * call is not atomic; no spinlocks may be held.
5821 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5823 struct migration_req req;
5824 unsigned long flags;
5828 rq = task_rq_lock(p, &flags);
5829 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5834 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5835 !cpus_equal(p->cpus_allowed, *new_mask))) {
5840 if (p->sched_class->set_cpus_allowed)
5841 p->sched_class->set_cpus_allowed(p, new_mask);
5843 p->cpus_allowed = *new_mask;
5844 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5847 /* Can the task run on the task's current CPU? If so, we're done */
5848 if (cpu_isset(task_cpu(p), *new_mask))
5851 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5852 /* Need help from migration thread: drop lock and wait. */
5853 task_rq_unlock(rq, &flags);
5854 wake_up_process(rq->migration_thread);
5855 wait_for_completion(&req.done);
5856 tlb_migrate_finish(p->mm);
5860 task_rq_unlock(rq, &flags);
5864 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5867 * Move (not current) task off this cpu, onto dest cpu. We're doing
5868 * this because either it can't run here any more (set_cpus_allowed()
5869 * away from this CPU, or CPU going down), or because we're
5870 * attempting to rebalance this task on exec (sched_exec).
5872 * So we race with normal scheduler movements, but that's OK, as long
5873 * as the task is no longer on this CPU.
5875 * Returns non-zero if task was successfully migrated.
5877 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5879 struct rq *rq_dest, *rq_src;
5882 if (unlikely(cpu_is_offline(dest_cpu)))
5885 rq_src = cpu_rq(src_cpu);
5886 rq_dest = cpu_rq(dest_cpu);
5888 double_rq_lock(rq_src, rq_dest);
5889 /* Already moved. */
5890 if (task_cpu(p) != src_cpu)
5892 /* Affinity changed (again). */
5893 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5896 on_rq = p->se.on_rq;
5898 deactivate_task(rq_src, p, 0);
5900 set_task_cpu(p, dest_cpu);
5902 activate_task(rq_dest, p, 0);
5903 check_preempt_curr(rq_dest, p);
5907 double_rq_unlock(rq_src, rq_dest);
5912 * migration_thread - this is a highprio system thread that performs
5913 * thread migration by bumping thread off CPU then 'pushing' onto
5916 static int migration_thread(void *data)
5918 int cpu = (long)data;
5922 BUG_ON(rq->migration_thread != current);
5924 set_current_state(TASK_INTERRUPTIBLE);
5925 while (!kthread_should_stop()) {
5926 struct migration_req *req;
5927 struct list_head *head;
5929 spin_lock_irq(&rq->lock);
5931 if (cpu_is_offline(cpu)) {
5932 spin_unlock_irq(&rq->lock);
5936 if (rq->active_balance) {
5937 active_load_balance(rq, cpu);
5938 rq->active_balance = 0;
5941 head = &rq->migration_queue;
5943 if (list_empty(head)) {
5944 spin_unlock_irq(&rq->lock);
5946 set_current_state(TASK_INTERRUPTIBLE);
5949 req = list_entry(head->next, struct migration_req, list);
5950 list_del_init(head->next);
5952 spin_unlock(&rq->lock);
5953 __migrate_task(req->task, cpu, req->dest_cpu);
5956 complete(&req->done);
5958 __set_current_state(TASK_RUNNING);
5962 /* Wait for kthread_stop */
5963 set_current_state(TASK_INTERRUPTIBLE);
5964 while (!kthread_should_stop()) {
5966 set_current_state(TASK_INTERRUPTIBLE);
5968 __set_current_state(TASK_RUNNING);
5972 #ifdef CONFIG_HOTPLUG_CPU
5974 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5978 local_irq_disable();
5979 ret = __migrate_task(p, src_cpu, dest_cpu);
5985 * Figure out where task on dead CPU should go, use force if necessary.
5986 * NOTE: interrupts should be disabled by the caller
5988 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5990 unsigned long flags;
5997 mask = node_to_cpumask(cpu_to_node(dead_cpu));
5998 cpus_and(mask, mask, p->cpus_allowed);
5999 dest_cpu = any_online_cpu(mask);
6001 /* On any allowed CPU? */
6002 if (dest_cpu >= nr_cpu_ids)
6003 dest_cpu = any_online_cpu(p->cpus_allowed);
6005 /* No more Mr. Nice Guy. */
6006 if (dest_cpu >= nr_cpu_ids) {
6007 cpumask_t cpus_allowed;
6009 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6011 * Try to stay on the same cpuset, where the
6012 * current cpuset may be a subset of all cpus.
6013 * The cpuset_cpus_allowed_locked() variant of
6014 * cpuset_cpus_allowed() will not block. It must be
6015 * called within calls to cpuset_lock/cpuset_unlock.
6017 rq = task_rq_lock(p, &flags);
6018 p->cpus_allowed = cpus_allowed;
6019 dest_cpu = any_online_cpu(p->cpus_allowed);
6020 task_rq_unlock(rq, &flags);
6023 * Don't tell them about moving exiting tasks or
6024 * kernel threads (both mm NULL), since they never
6027 if (p->mm && printk_ratelimit()) {
6028 printk(KERN_INFO "process %d (%s) no "
6029 "longer affine to cpu%d\n",
6030 task_pid_nr(p), p->comm, dead_cpu);
6033 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6037 * While a dead CPU has no uninterruptible tasks queued at this point,
6038 * it might still have a nonzero ->nr_uninterruptible counter, because
6039 * for performance reasons the counter is not stricly tracking tasks to
6040 * their home CPUs. So we just add the counter to another CPU's counter,
6041 * to keep the global sum constant after CPU-down:
6043 static void migrate_nr_uninterruptible(struct rq *rq_src)
6045 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6046 unsigned long flags;
6048 local_irq_save(flags);
6049 double_rq_lock(rq_src, rq_dest);
6050 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6051 rq_src->nr_uninterruptible = 0;
6052 double_rq_unlock(rq_src, rq_dest);
6053 local_irq_restore(flags);
6056 /* Run through task list and migrate tasks from the dead cpu. */
6057 static void migrate_live_tasks(int src_cpu)
6059 struct task_struct *p, *t;
6061 read_lock(&tasklist_lock);
6063 do_each_thread(t, p) {
6067 if (task_cpu(p) == src_cpu)
6068 move_task_off_dead_cpu(src_cpu, p);
6069 } while_each_thread(t, p);
6071 read_unlock(&tasklist_lock);
6075 * Schedules idle task to be the next runnable task on current CPU.
6076 * It does so by boosting its priority to highest possible.
6077 * Used by CPU offline code.
6079 void sched_idle_next(void)
6081 int this_cpu = smp_processor_id();
6082 struct rq *rq = cpu_rq(this_cpu);
6083 struct task_struct *p = rq->idle;
6084 unsigned long flags;
6086 /* cpu has to be offline */
6087 BUG_ON(cpu_online(this_cpu));
6090 * Strictly not necessary since rest of the CPUs are stopped by now
6091 * and interrupts disabled on the current cpu.
6093 spin_lock_irqsave(&rq->lock, flags);
6095 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6097 update_rq_clock(rq);
6098 activate_task(rq, p, 0);
6100 spin_unlock_irqrestore(&rq->lock, flags);
6104 * Ensures that the idle task is using init_mm right before its cpu goes
6107 void idle_task_exit(void)
6109 struct mm_struct *mm = current->active_mm;
6111 BUG_ON(cpu_online(smp_processor_id()));
6114 switch_mm(mm, &init_mm, current);
6118 /* called under rq->lock with disabled interrupts */
6119 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6121 struct rq *rq = cpu_rq(dead_cpu);
6123 /* Must be exiting, otherwise would be on tasklist. */
6124 BUG_ON(!p->exit_state);
6126 /* Cannot have done final schedule yet: would have vanished. */
6127 BUG_ON(p->state == TASK_DEAD);
6132 * Drop lock around migration; if someone else moves it,
6133 * that's OK. No task can be added to this CPU, so iteration is
6136 spin_unlock_irq(&rq->lock);
6137 move_task_off_dead_cpu(dead_cpu, p);
6138 spin_lock_irq(&rq->lock);
6143 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6144 static void migrate_dead_tasks(unsigned int dead_cpu)
6146 struct rq *rq = cpu_rq(dead_cpu);
6147 struct task_struct *next;
6150 if (!rq->nr_running)
6152 update_rq_clock(rq);
6153 next = pick_next_task(rq, rq->curr);
6156 migrate_dead(dead_cpu, next);
6160 #endif /* CONFIG_HOTPLUG_CPU */
6162 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6164 static struct ctl_table sd_ctl_dir[] = {
6166 .procname = "sched_domain",
6172 static struct ctl_table sd_ctl_root[] = {
6174 .ctl_name = CTL_KERN,
6175 .procname = "kernel",
6177 .child = sd_ctl_dir,
6182 static struct ctl_table *sd_alloc_ctl_entry(int n)
6184 struct ctl_table *entry =
6185 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6190 static void sd_free_ctl_entry(struct ctl_table **tablep)
6192 struct ctl_table *entry;
6195 * In the intermediate directories, both the child directory and
6196 * procname are dynamically allocated and could fail but the mode
6197 * will always be set. In the lowest directory the names are
6198 * static strings and all have proc handlers.
6200 for (entry = *tablep; entry->mode; entry++) {
6202 sd_free_ctl_entry(&entry->child);
6203 if (entry->proc_handler == NULL)
6204 kfree(entry->procname);
6212 set_table_entry(struct ctl_table *entry,
6213 const char *procname, void *data, int maxlen,
6214 mode_t mode, proc_handler *proc_handler)
6216 entry->procname = procname;
6218 entry->maxlen = maxlen;
6220 entry->proc_handler = proc_handler;
6223 static struct ctl_table *
6224 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6226 struct ctl_table *table = sd_alloc_ctl_entry(12);
6231 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6232 sizeof(long), 0644, proc_doulongvec_minmax);
6233 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6234 sizeof(long), 0644, proc_doulongvec_minmax);
6235 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6236 sizeof(int), 0644, proc_dointvec_minmax);
6237 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6238 sizeof(int), 0644, proc_dointvec_minmax);
6239 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6240 sizeof(int), 0644, proc_dointvec_minmax);
6241 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6242 sizeof(int), 0644, proc_dointvec_minmax);
6243 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6244 sizeof(int), 0644, proc_dointvec_minmax);
6245 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6246 sizeof(int), 0644, proc_dointvec_minmax);
6247 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6248 sizeof(int), 0644, proc_dointvec_minmax);
6249 set_table_entry(&table[9], "cache_nice_tries",
6250 &sd->cache_nice_tries,
6251 sizeof(int), 0644, proc_dointvec_minmax);
6252 set_table_entry(&table[10], "flags", &sd->flags,
6253 sizeof(int), 0644, proc_dointvec_minmax);
6254 /* &table[11] is terminator */
6259 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6261 struct ctl_table *entry, *table;
6262 struct sched_domain *sd;
6263 int domain_num = 0, i;
6266 for_each_domain(cpu, sd)
6268 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6273 for_each_domain(cpu, sd) {
6274 snprintf(buf, 32, "domain%d", i);
6275 entry->procname = kstrdup(buf, GFP_KERNEL);
6277 entry->child = sd_alloc_ctl_domain_table(sd);
6284 static struct ctl_table_header *sd_sysctl_header;
6285 static void register_sched_domain_sysctl(void)
6287 int i, cpu_num = num_online_cpus();
6288 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6291 WARN_ON(sd_ctl_dir[0].child);
6292 sd_ctl_dir[0].child = entry;
6297 for_each_online_cpu(i) {
6298 snprintf(buf, 32, "cpu%d", i);
6299 entry->procname = kstrdup(buf, GFP_KERNEL);
6301 entry->child = sd_alloc_ctl_cpu_table(i);
6305 WARN_ON(sd_sysctl_header);
6306 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6309 /* may be called multiple times per register */
6310 static void unregister_sched_domain_sysctl(void)
6312 if (sd_sysctl_header)
6313 unregister_sysctl_table(sd_sysctl_header);
6314 sd_sysctl_header = NULL;
6315 if (sd_ctl_dir[0].child)
6316 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6319 static void register_sched_domain_sysctl(void)
6322 static void unregister_sched_domain_sysctl(void)
6327 static void set_rq_online(struct rq *rq)
6330 const struct sched_class *class;
6332 cpu_set(rq->cpu, rq->rd->online);
6335 for_each_class(class) {
6336 if (class->rq_online)
6337 class->rq_online(rq);
6342 static void set_rq_offline(struct rq *rq)
6345 const struct sched_class *class;
6347 for_each_class(class) {
6348 if (class->rq_offline)
6349 class->rq_offline(rq);
6352 cpu_clear(rq->cpu, rq->rd->online);
6358 * migration_call - callback that gets triggered when a CPU is added.
6359 * Here we can start up the necessary migration thread for the new CPU.
6361 static int __cpuinit
6362 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6364 struct task_struct *p;
6365 int cpu = (long)hcpu;
6366 unsigned long flags;
6371 case CPU_UP_PREPARE:
6372 case CPU_UP_PREPARE_FROZEN:
6373 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6376 kthread_bind(p, cpu);
6377 /* Must be high prio: stop_machine expects to yield to it. */
6378 rq = task_rq_lock(p, &flags);
6379 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6380 task_rq_unlock(rq, &flags);
6381 cpu_rq(cpu)->migration_thread = p;
6385 case CPU_ONLINE_FROZEN:
6386 /* Strictly unnecessary, as first user will wake it. */
6387 wake_up_process(cpu_rq(cpu)->migration_thread);
6389 /* Update our root-domain */
6391 spin_lock_irqsave(&rq->lock, flags);
6393 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6397 spin_unlock_irqrestore(&rq->lock, flags);
6400 #ifdef CONFIG_HOTPLUG_CPU
6401 case CPU_UP_CANCELED:
6402 case CPU_UP_CANCELED_FROZEN:
6403 if (!cpu_rq(cpu)->migration_thread)
6405 /* Unbind it from offline cpu so it can run. Fall thru. */
6406 kthread_bind(cpu_rq(cpu)->migration_thread,
6407 any_online_cpu(cpu_online_map));
6408 kthread_stop(cpu_rq(cpu)->migration_thread);
6409 cpu_rq(cpu)->migration_thread = NULL;
6413 case CPU_DEAD_FROZEN:
6414 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6415 migrate_live_tasks(cpu);
6417 kthread_stop(rq->migration_thread);
6418 rq->migration_thread = NULL;
6419 /* Idle task back to normal (off runqueue, low prio) */
6420 spin_lock_irq(&rq->lock);
6421 update_rq_clock(rq);
6422 deactivate_task(rq, rq->idle, 0);
6423 rq->idle->static_prio = MAX_PRIO;
6424 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6425 rq->idle->sched_class = &idle_sched_class;
6426 migrate_dead_tasks(cpu);
6427 spin_unlock_irq(&rq->lock);
6429 migrate_nr_uninterruptible(rq);
6430 BUG_ON(rq->nr_running != 0);
6433 * No need to migrate the tasks: it was best-effort if
6434 * they didn't take sched_hotcpu_mutex. Just wake up
6437 spin_lock_irq(&rq->lock);
6438 while (!list_empty(&rq->migration_queue)) {
6439 struct migration_req *req;
6441 req = list_entry(rq->migration_queue.next,
6442 struct migration_req, list);
6443 list_del_init(&req->list);
6444 complete(&req->done);
6446 spin_unlock_irq(&rq->lock);
6450 case CPU_DYING_FROZEN:
6451 /* Update our root-domain */
6453 spin_lock_irqsave(&rq->lock, flags);
6455 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6458 spin_unlock_irqrestore(&rq->lock, flags);
6465 /* Register at highest priority so that task migration (migrate_all_tasks)
6466 * happens before everything else.
6468 static struct notifier_block __cpuinitdata migration_notifier = {
6469 .notifier_call = migration_call,
6473 void __init migration_init(void)
6475 void *cpu = (void *)(long)smp_processor_id();
6478 /* Start one for the boot CPU: */
6479 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6480 BUG_ON(err == NOTIFY_BAD);
6481 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6482 register_cpu_notifier(&migration_notifier);
6488 #ifdef CONFIG_SCHED_DEBUG
6490 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6503 case SD_LV_ALLNODES:
6512 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6513 cpumask_t *groupmask)
6515 struct sched_group *group = sd->groups;
6518 cpulist_scnprintf(str, sizeof(str), sd->span);
6519 cpus_clear(*groupmask);
6521 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6523 if (!(sd->flags & SD_LOAD_BALANCE)) {
6524 printk("does not load-balance\n");
6526 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6531 printk(KERN_CONT "span %s level %s\n",
6532 str, sd_level_to_string(sd->level));
6534 if (!cpu_isset(cpu, sd->span)) {
6535 printk(KERN_ERR "ERROR: domain->span does not contain "
6538 if (!cpu_isset(cpu, group->cpumask)) {
6539 printk(KERN_ERR "ERROR: domain->groups does not contain"
6543 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6547 printk(KERN_ERR "ERROR: group is NULL\n");
6551 if (!group->__cpu_power) {
6552 printk(KERN_CONT "\n");
6553 printk(KERN_ERR "ERROR: domain->cpu_power not "
6558 if (!cpus_weight(group->cpumask)) {
6559 printk(KERN_CONT "\n");
6560 printk(KERN_ERR "ERROR: empty group\n");
6564 if (cpus_intersects(*groupmask, group->cpumask)) {
6565 printk(KERN_CONT "\n");
6566 printk(KERN_ERR "ERROR: repeated CPUs\n");
6570 cpus_or(*groupmask, *groupmask, group->cpumask);
6572 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6573 printk(KERN_CONT " %s", str);
6575 group = group->next;
6576 } while (group != sd->groups);
6577 printk(KERN_CONT "\n");
6579 if (!cpus_equal(sd->span, *groupmask))
6580 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6582 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6583 printk(KERN_ERR "ERROR: parent span is not a superset "
6584 "of domain->span\n");
6588 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6590 cpumask_t *groupmask;
6594 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6598 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6600 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6602 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6607 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6616 #else /* !CONFIG_SCHED_DEBUG */
6617 # define sched_domain_debug(sd, cpu) do { } while (0)
6618 #endif /* CONFIG_SCHED_DEBUG */
6620 static int sd_degenerate(struct sched_domain *sd)
6622 if (cpus_weight(sd->span) == 1)
6625 /* Following flags need at least 2 groups */
6626 if (sd->flags & (SD_LOAD_BALANCE |
6627 SD_BALANCE_NEWIDLE |
6631 SD_SHARE_PKG_RESOURCES)) {
6632 if (sd->groups != sd->groups->next)
6636 /* Following flags don't use groups */
6637 if (sd->flags & (SD_WAKE_IDLE |
6646 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6648 unsigned long cflags = sd->flags, pflags = parent->flags;
6650 if (sd_degenerate(parent))
6653 if (!cpus_equal(sd->span, parent->span))
6656 /* Does parent contain flags not in child? */
6657 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6658 if (cflags & SD_WAKE_AFFINE)
6659 pflags &= ~SD_WAKE_BALANCE;
6660 /* Flags needing groups don't count if only 1 group in parent */
6661 if (parent->groups == parent->groups->next) {
6662 pflags &= ~(SD_LOAD_BALANCE |
6663 SD_BALANCE_NEWIDLE |
6667 SD_SHARE_PKG_RESOURCES);
6669 if (~cflags & pflags)
6675 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6677 unsigned long flags;
6679 spin_lock_irqsave(&rq->lock, flags);
6682 struct root_domain *old_rd = rq->rd;
6684 if (cpu_isset(rq->cpu, old_rd->online))
6687 cpu_clear(rq->cpu, old_rd->span);
6689 if (atomic_dec_and_test(&old_rd->refcount))
6693 atomic_inc(&rd->refcount);
6696 cpu_set(rq->cpu, rd->span);
6697 if (cpu_isset(rq->cpu, cpu_online_map))
6700 spin_unlock_irqrestore(&rq->lock, flags);
6703 static void init_rootdomain(struct root_domain *rd)
6705 memset(rd, 0, sizeof(*rd));
6707 cpus_clear(rd->span);
6708 cpus_clear(rd->online);
6710 cpupri_init(&rd->cpupri);
6713 static void init_defrootdomain(void)
6715 init_rootdomain(&def_root_domain);
6716 atomic_set(&def_root_domain.refcount, 1);
6719 static struct root_domain *alloc_rootdomain(void)
6721 struct root_domain *rd;
6723 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6727 init_rootdomain(rd);
6733 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6734 * hold the hotplug lock.
6737 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6739 struct rq *rq = cpu_rq(cpu);
6740 struct sched_domain *tmp;
6742 /* Remove the sched domains which do not contribute to scheduling. */
6743 for (tmp = sd; tmp; tmp = tmp->parent) {
6744 struct sched_domain *parent = tmp->parent;
6747 if (sd_parent_degenerate(tmp, parent)) {
6748 tmp->parent = parent->parent;
6750 parent->parent->child = tmp;
6754 if (sd && sd_degenerate(sd)) {
6760 sched_domain_debug(sd, cpu);
6762 rq_attach_root(rq, rd);
6763 rcu_assign_pointer(rq->sd, sd);
6766 /* cpus with isolated domains */
6767 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6769 /* Setup the mask of cpus configured for isolated domains */
6770 static int __init isolated_cpu_setup(char *str)
6772 int ints[NR_CPUS], i;
6774 str = get_options(str, ARRAY_SIZE(ints), ints);
6775 cpus_clear(cpu_isolated_map);
6776 for (i = 1; i <= ints[0]; i++)
6777 if (ints[i] < NR_CPUS)
6778 cpu_set(ints[i], cpu_isolated_map);
6782 __setup("isolcpus=", isolated_cpu_setup);
6785 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6786 * to a function which identifies what group(along with sched group) a CPU
6787 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6788 * (due to the fact that we keep track of groups covered with a cpumask_t).
6790 * init_sched_build_groups will build a circular linked list of the groups
6791 * covered by the given span, and will set each group's ->cpumask correctly,
6792 * and ->cpu_power to 0.
6795 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6796 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6797 struct sched_group **sg,
6798 cpumask_t *tmpmask),
6799 cpumask_t *covered, cpumask_t *tmpmask)
6801 struct sched_group *first = NULL, *last = NULL;
6804 cpus_clear(*covered);
6806 for_each_cpu_mask(i, *span) {
6807 struct sched_group *sg;
6808 int group = group_fn(i, cpu_map, &sg, tmpmask);
6811 if (cpu_isset(i, *covered))
6814 cpus_clear(sg->cpumask);
6815 sg->__cpu_power = 0;
6817 for_each_cpu_mask(j, *span) {
6818 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6821 cpu_set(j, *covered);
6822 cpu_set(j, sg->cpumask);
6833 #define SD_NODES_PER_DOMAIN 16
6838 * find_next_best_node - find the next node to include in a sched_domain
6839 * @node: node whose sched_domain we're building
6840 * @used_nodes: nodes already in the sched_domain
6842 * Find the next node to include in a given scheduling domain. Simply
6843 * finds the closest node not already in the @used_nodes map.
6845 * Should use nodemask_t.
6847 static int find_next_best_node(int node, nodemask_t *used_nodes)
6849 int i, n, val, min_val, best_node = 0;
6853 for (i = 0; i < MAX_NUMNODES; i++) {
6854 /* Start at @node */
6855 n = (node + i) % MAX_NUMNODES;
6857 if (!nr_cpus_node(n))
6860 /* Skip already used nodes */
6861 if (node_isset(n, *used_nodes))
6864 /* Simple min distance search */
6865 val = node_distance(node, n);
6867 if (val < min_val) {
6873 node_set(best_node, *used_nodes);
6878 * sched_domain_node_span - get a cpumask for a node's sched_domain
6879 * @node: node whose cpumask we're constructing
6880 * @span: resulting cpumask
6882 * Given a node, construct a good cpumask for its sched_domain to span. It
6883 * should be one that prevents unnecessary balancing, but also spreads tasks
6886 static void sched_domain_node_span(int node, cpumask_t *span)
6888 nodemask_t used_nodes;
6889 node_to_cpumask_ptr(nodemask, node);
6893 nodes_clear(used_nodes);
6895 cpus_or(*span, *span, *nodemask);
6896 node_set(node, used_nodes);
6898 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6899 int next_node = find_next_best_node(node, &used_nodes);
6901 node_to_cpumask_ptr_next(nodemask, next_node);
6902 cpus_or(*span, *span, *nodemask);
6905 #endif /* CONFIG_NUMA */
6907 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6910 * SMT sched-domains:
6912 #ifdef CONFIG_SCHED_SMT
6913 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6914 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6917 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6921 *sg = &per_cpu(sched_group_cpus, cpu);
6924 #endif /* CONFIG_SCHED_SMT */
6927 * multi-core sched-domains:
6929 #ifdef CONFIG_SCHED_MC
6930 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6931 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6932 #endif /* CONFIG_SCHED_MC */
6934 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6936 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6941 *mask = per_cpu(cpu_sibling_map, cpu);
6942 cpus_and(*mask, *mask, *cpu_map);
6943 group = first_cpu(*mask);
6945 *sg = &per_cpu(sched_group_core, group);
6948 #elif defined(CONFIG_SCHED_MC)
6950 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6954 *sg = &per_cpu(sched_group_core, cpu);
6959 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6960 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6963 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6967 #ifdef CONFIG_SCHED_MC
6968 *mask = cpu_coregroup_map(cpu);
6969 cpus_and(*mask, *mask, *cpu_map);
6970 group = first_cpu(*mask);
6971 #elif defined(CONFIG_SCHED_SMT)
6972 *mask = per_cpu(cpu_sibling_map, cpu);
6973 cpus_and(*mask, *mask, *cpu_map);
6974 group = first_cpu(*mask);
6979 *sg = &per_cpu(sched_group_phys, group);
6985 * The init_sched_build_groups can't handle what we want to do with node
6986 * groups, so roll our own. Now each node has its own list of groups which
6987 * gets dynamically allocated.
6989 static DEFINE_PER_CPU(struct sched_domain, node_domains);
6990 static struct sched_group ***sched_group_nodes_bycpu;
6992 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
6993 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
6995 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
6996 struct sched_group **sg, cpumask_t *nodemask)
7000 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7001 cpus_and(*nodemask, *nodemask, *cpu_map);
7002 group = first_cpu(*nodemask);
7005 *sg = &per_cpu(sched_group_allnodes, group);
7009 static void init_numa_sched_groups_power(struct sched_group *group_head)
7011 struct sched_group *sg = group_head;
7017 for_each_cpu_mask(j, sg->cpumask) {
7018 struct sched_domain *sd;
7020 sd = &per_cpu(phys_domains, j);
7021 if (j != first_cpu(sd->groups->cpumask)) {
7023 * Only add "power" once for each
7029 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7032 } while (sg != group_head);
7034 #endif /* CONFIG_NUMA */
7037 /* Free memory allocated for various sched_group structures */
7038 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7042 for_each_cpu_mask(cpu, *cpu_map) {
7043 struct sched_group **sched_group_nodes
7044 = sched_group_nodes_bycpu[cpu];
7046 if (!sched_group_nodes)
7049 for (i = 0; i < MAX_NUMNODES; i++) {
7050 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7052 *nodemask = node_to_cpumask(i);
7053 cpus_and(*nodemask, *nodemask, *cpu_map);
7054 if (cpus_empty(*nodemask))
7064 if (oldsg != sched_group_nodes[i])
7067 kfree(sched_group_nodes);
7068 sched_group_nodes_bycpu[cpu] = NULL;
7071 #else /* !CONFIG_NUMA */
7072 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7075 #endif /* CONFIG_NUMA */
7078 * Initialize sched groups cpu_power.
7080 * cpu_power indicates the capacity of sched group, which is used while
7081 * distributing the load between different sched groups in a sched domain.
7082 * Typically cpu_power for all the groups in a sched domain will be same unless
7083 * there are asymmetries in the topology. If there are asymmetries, group
7084 * having more cpu_power will pickup more load compared to the group having
7087 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7088 * the maximum number of tasks a group can handle in the presence of other idle
7089 * or lightly loaded groups in the same sched domain.
7091 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7093 struct sched_domain *child;
7094 struct sched_group *group;
7096 WARN_ON(!sd || !sd->groups);
7098 if (cpu != first_cpu(sd->groups->cpumask))
7103 sd->groups->__cpu_power = 0;
7106 * For perf policy, if the groups in child domain share resources
7107 * (for example cores sharing some portions of the cache hierarchy
7108 * or SMT), then set this domain groups cpu_power such that each group
7109 * can handle only one task, when there are other idle groups in the
7110 * same sched domain.
7112 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7114 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7115 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7120 * add cpu_power of each child group to this groups cpu_power
7122 group = child->groups;
7124 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7125 group = group->next;
7126 } while (group != child->groups);
7130 * Initializers for schedule domains
7131 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7134 #define SD_INIT(sd, type) sd_init_##type(sd)
7135 #define SD_INIT_FUNC(type) \
7136 static noinline void sd_init_##type(struct sched_domain *sd) \
7138 memset(sd, 0, sizeof(*sd)); \
7139 *sd = SD_##type##_INIT; \
7140 sd->level = SD_LV_##type; \
7145 SD_INIT_FUNC(ALLNODES)
7148 #ifdef CONFIG_SCHED_SMT
7149 SD_INIT_FUNC(SIBLING)
7151 #ifdef CONFIG_SCHED_MC
7156 * To minimize stack usage kmalloc room for cpumasks and share the
7157 * space as the usage in build_sched_domains() dictates. Used only
7158 * if the amount of space is significant.
7161 cpumask_t tmpmask; /* make this one first */
7164 cpumask_t this_sibling_map;
7165 cpumask_t this_core_map;
7167 cpumask_t send_covered;
7170 cpumask_t domainspan;
7172 cpumask_t notcovered;
7177 #define SCHED_CPUMASK_ALLOC 1
7178 #define SCHED_CPUMASK_FREE(v) kfree(v)
7179 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7181 #define SCHED_CPUMASK_ALLOC 0
7182 #define SCHED_CPUMASK_FREE(v)
7183 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7186 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7187 ((unsigned long)(a) + offsetof(struct allmasks, v))
7189 static int default_relax_domain_level = -1;
7191 static int __init setup_relax_domain_level(char *str)
7195 val = simple_strtoul(str, NULL, 0);
7196 if (val < SD_LV_MAX)
7197 default_relax_domain_level = val;
7201 __setup("relax_domain_level=", setup_relax_domain_level);
7203 static void set_domain_attribute(struct sched_domain *sd,
7204 struct sched_domain_attr *attr)
7208 if (!attr || attr->relax_domain_level < 0) {
7209 if (default_relax_domain_level < 0)
7212 request = default_relax_domain_level;
7214 request = attr->relax_domain_level;
7215 if (request < sd->level) {
7216 /* turn off idle balance on this domain */
7217 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7219 /* turn on idle balance on this domain */
7220 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7225 * Build sched domains for a given set of cpus and attach the sched domains
7226 * to the individual cpus
7228 static int __build_sched_domains(const cpumask_t *cpu_map,
7229 struct sched_domain_attr *attr)
7232 struct root_domain *rd;
7233 SCHED_CPUMASK_DECLARE(allmasks);
7236 struct sched_group **sched_group_nodes = NULL;
7237 int sd_allnodes = 0;
7240 * Allocate the per-node list of sched groups
7242 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7244 if (!sched_group_nodes) {
7245 printk(KERN_WARNING "Can not alloc sched group node list\n");
7250 rd = alloc_rootdomain();
7252 printk(KERN_WARNING "Cannot alloc root domain\n");
7254 kfree(sched_group_nodes);
7259 #if SCHED_CPUMASK_ALLOC
7260 /* get space for all scratch cpumask variables */
7261 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7263 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7266 kfree(sched_group_nodes);
7271 tmpmask = (cpumask_t *)allmasks;
7275 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7279 * Set up domains for cpus specified by the cpu_map.
7281 for_each_cpu_mask(i, *cpu_map) {
7282 struct sched_domain *sd = NULL, *p;
7283 SCHED_CPUMASK_VAR(nodemask, allmasks);
7285 *nodemask = node_to_cpumask(cpu_to_node(i));
7286 cpus_and(*nodemask, *nodemask, *cpu_map);
7289 if (cpus_weight(*cpu_map) >
7290 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7291 sd = &per_cpu(allnodes_domains, i);
7292 SD_INIT(sd, ALLNODES);
7293 set_domain_attribute(sd, attr);
7294 sd->span = *cpu_map;
7295 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7301 sd = &per_cpu(node_domains, i);
7303 set_domain_attribute(sd, attr);
7304 sched_domain_node_span(cpu_to_node(i), &sd->span);
7308 cpus_and(sd->span, sd->span, *cpu_map);
7312 sd = &per_cpu(phys_domains, i);
7314 set_domain_attribute(sd, attr);
7315 sd->span = *nodemask;
7319 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7321 #ifdef CONFIG_SCHED_MC
7323 sd = &per_cpu(core_domains, i);
7325 set_domain_attribute(sd, attr);
7326 sd->span = cpu_coregroup_map(i);
7327 cpus_and(sd->span, sd->span, *cpu_map);
7330 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7333 #ifdef CONFIG_SCHED_SMT
7335 sd = &per_cpu(cpu_domains, i);
7336 SD_INIT(sd, SIBLING);
7337 set_domain_attribute(sd, attr);
7338 sd->span = per_cpu(cpu_sibling_map, i);
7339 cpus_and(sd->span, sd->span, *cpu_map);
7342 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7346 #ifdef CONFIG_SCHED_SMT
7347 /* Set up CPU (sibling) groups */
7348 for_each_cpu_mask(i, *cpu_map) {
7349 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7350 SCHED_CPUMASK_VAR(send_covered, allmasks);
7352 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7353 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7354 if (i != first_cpu(*this_sibling_map))
7357 init_sched_build_groups(this_sibling_map, cpu_map,
7359 send_covered, tmpmask);
7363 #ifdef CONFIG_SCHED_MC
7364 /* Set up multi-core groups */
7365 for_each_cpu_mask(i, *cpu_map) {
7366 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7367 SCHED_CPUMASK_VAR(send_covered, allmasks);
7369 *this_core_map = cpu_coregroup_map(i);
7370 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7371 if (i != first_cpu(*this_core_map))
7374 init_sched_build_groups(this_core_map, cpu_map,
7376 send_covered, tmpmask);
7380 /* Set up physical groups */
7381 for (i = 0; i < MAX_NUMNODES; i++) {
7382 SCHED_CPUMASK_VAR(nodemask, allmasks);
7383 SCHED_CPUMASK_VAR(send_covered, allmasks);
7385 *nodemask = node_to_cpumask(i);
7386 cpus_and(*nodemask, *nodemask, *cpu_map);
7387 if (cpus_empty(*nodemask))
7390 init_sched_build_groups(nodemask, cpu_map,
7392 send_covered, tmpmask);
7396 /* Set up node groups */
7398 SCHED_CPUMASK_VAR(send_covered, allmasks);
7400 init_sched_build_groups(cpu_map, cpu_map,
7401 &cpu_to_allnodes_group,
7402 send_covered, tmpmask);
7405 for (i = 0; i < MAX_NUMNODES; i++) {
7406 /* Set up node groups */
7407 struct sched_group *sg, *prev;
7408 SCHED_CPUMASK_VAR(nodemask, allmasks);
7409 SCHED_CPUMASK_VAR(domainspan, allmasks);
7410 SCHED_CPUMASK_VAR(covered, allmasks);
7413 *nodemask = node_to_cpumask(i);
7414 cpus_clear(*covered);
7416 cpus_and(*nodemask, *nodemask, *cpu_map);
7417 if (cpus_empty(*nodemask)) {
7418 sched_group_nodes[i] = NULL;
7422 sched_domain_node_span(i, domainspan);
7423 cpus_and(*domainspan, *domainspan, *cpu_map);
7425 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7427 printk(KERN_WARNING "Can not alloc domain group for "
7431 sched_group_nodes[i] = sg;
7432 for_each_cpu_mask(j, *nodemask) {
7433 struct sched_domain *sd;
7435 sd = &per_cpu(node_domains, j);
7438 sg->__cpu_power = 0;
7439 sg->cpumask = *nodemask;
7441 cpus_or(*covered, *covered, *nodemask);
7444 for (j = 0; j < MAX_NUMNODES; j++) {
7445 SCHED_CPUMASK_VAR(notcovered, allmasks);
7446 int n = (i + j) % MAX_NUMNODES;
7447 node_to_cpumask_ptr(pnodemask, n);
7449 cpus_complement(*notcovered, *covered);
7450 cpus_and(*tmpmask, *notcovered, *cpu_map);
7451 cpus_and(*tmpmask, *tmpmask, *domainspan);
7452 if (cpus_empty(*tmpmask))
7455 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7456 if (cpus_empty(*tmpmask))
7459 sg = kmalloc_node(sizeof(struct sched_group),
7463 "Can not alloc domain group for node %d\n", j);
7466 sg->__cpu_power = 0;
7467 sg->cpumask = *tmpmask;
7468 sg->next = prev->next;
7469 cpus_or(*covered, *covered, *tmpmask);
7476 /* Calculate CPU power for physical packages and nodes */
7477 #ifdef CONFIG_SCHED_SMT
7478 for_each_cpu_mask(i, *cpu_map) {
7479 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7481 init_sched_groups_power(i, sd);
7484 #ifdef CONFIG_SCHED_MC
7485 for_each_cpu_mask(i, *cpu_map) {
7486 struct sched_domain *sd = &per_cpu(core_domains, i);
7488 init_sched_groups_power(i, sd);
7492 for_each_cpu_mask(i, *cpu_map) {
7493 struct sched_domain *sd = &per_cpu(phys_domains, i);
7495 init_sched_groups_power(i, sd);
7499 for (i = 0; i < MAX_NUMNODES; i++)
7500 init_numa_sched_groups_power(sched_group_nodes[i]);
7503 struct sched_group *sg;
7505 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7507 init_numa_sched_groups_power(sg);
7511 /* Attach the domains */
7512 for_each_cpu_mask(i, *cpu_map) {
7513 struct sched_domain *sd;
7514 #ifdef CONFIG_SCHED_SMT
7515 sd = &per_cpu(cpu_domains, i);
7516 #elif defined(CONFIG_SCHED_MC)
7517 sd = &per_cpu(core_domains, i);
7519 sd = &per_cpu(phys_domains, i);
7521 cpu_attach_domain(sd, rd, i);
7524 SCHED_CPUMASK_FREE((void *)allmasks);
7529 free_sched_groups(cpu_map, tmpmask);
7530 SCHED_CPUMASK_FREE((void *)allmasks);
7535 static int build_sched_domains(const cpumask_t *cpu_map)
7537 return __build_sched_domains(cpu_map, NULL);
7540 static cpumask_t *doms_cur; /* current sched domains */
7541 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7542 static struct sched_domain_attr *dattr_cur;
7543 /* attribues of custom domains in 'doms_cur' */
7546 * Special case: If a kmalloc of a doms_cur partition (array of
7547 * cpumask_t) fails, then fallback to a single sched domain,
7548 * as determined by the single cpumask_t fallback_doms.
7550 static cpumask_t fallback_doms;
7552 void __attribute__((weak)) arch_update_cpu_topology(void)
7557 * Free current domain masks.
7558 * Called after all cpus are attached to NULL domain.
7560 static void free_sched_domains(void)
7563 if (doms_cur != &fallback_doms)
7565 doms_cur = &fallback_doms;
7569 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7570 * For now this just excludes isolated cpus, but could be used to
7571 * exclude other special cases in the future.
7573 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7577 arch_update_cpu_topology();
7579 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7581 doms_cur = &fallback_doms;
7582 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7584 err = build_sched_domains(doms_cur);
7585 register_sched_domain_sysctl();
7590 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7593 free_sched_groups(cpu_map, tmpmask);
7597 * Detach sched domains from a group of cpus specified in cpu_map
7598 * These cpus will now be attached to the NULL domain
7600 static void detach_destroy_domains(const cpumask_t *cpu_map)
7605 unregister_sched_domain_sysctl();
7607 for_each_cpu_mask(i, *cpu_map)
7608 cpu_attach_domain(NULL, &def_root_domain, i);
7609 synchronize_sched();
7610 arch_destroy_sched_domains(cpu_map, &tmpmask);
7613 /* handle null as "default" */
7614 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7615 struct sched_domain_attr *new, int idx_new)
7617 struct sched_domain_attr tmp;
7624 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7625 new ? (new + idx_new) : &tmp,
7626 sizeof(struct sched_domain_attr));
7630 * Partition sched domains as specified by the 'ndoms_new'
7631 * cpumasks in the array doms_new[] of cpumasks. This compares
7632 * doms_new[] to the current sched domain partitioning, doms_cur[].
7633 * It destroys each deleted domain and builds each new domain.
7635 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7636 * The masks don't intersect (don't overlap.) We should setup one
7637 * sched domain for each mask. CPUs not in any of the cpumasks will
7638 * not be load balanced. If the same cpumask appears both in the
7639 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7642 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7643 * ownership of it and will kfree it when done with it. If the caller
7644 * failed the kmalloc call, then it can pass in doms_new == NULL,
7645 * and partition_sched_domains() will fallback to the single partition
7648 * Call with hotplug lock held
7650 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7651 struct sched_domain_attr *dattr_new)
7655 mutex_lock(&sched_domains_mutex);
7657 /* always unregister in case we don't destroy any domains */
7658 unregister_sched_domain_sysctl();
7660 if (doms_new == NULL) {
7662 doms_new = &fallback_doms;
7663 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7667 /* Destroy deleted domains */
7668 for (i = 0; i < ndoms_cur; i++) {
7669 for (j = 0; j < ndoms_new; j++) {
7670 if (cpus_equal(doms_cur[i], doms_new[j])
7671 && dattrs_equal(dattr_cur, i, dattr_new, j))
7674 /* no match - a current sched domain not in new doms_new[] */
7675 detach_destroy_domains(doms_cur + i);
7680 /* Build new domains */
7681 for (i = 0; i < ndoms_new; i++) {
7682 for (j = 0; j < ndoms_cur; j++) {
7683 if (cpus_equal(doms_new[i], doms_cur[j])
7684 && dattrs_equal(dattr_new, i, dattr_cur, j))
7687 /* no match - add a new doms_new */
7688 __build_sched_domains(doms_new + i,
7689 dattr_new ? dattr_new + i : NULL);
7694 /* Remember the new sched domains */
7695 if (doms_cur != &fallback_doms)
7697 kfree(dattr_cur); /* kfree(NULL) is safe */
7698 doms_cur = doms_new;
7699 dattr_cur = dattr_new;
7700 ndoms_cur = ndoms_new;
7702 register_sched_domain_sysctl();
7704 mutex_unlock(&sched_domains_mutex);
7707 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7708 int arch_reinit_sched_domains(void)
7713 mutex_lock(&sched_domains_mutex);
7714 detach_destroy_domains(&cpu_online_map);
7715 free_sched_domains();
7716 err = arch_init_sched_domains(&cpu_online_map);
7717 mutex_unlock(&sched_domains_mutex);
7723 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7727 if (buf[0] != '0' && buf[0] != '1')
7731 sched_smt_power_savings = (buf[0] == '1');
7733 sched_mc_power_savings = (buf[0] == '1');
7735 ret = arch_reinit_sched_domains();
7737 return ret ? ret : count;
7740 #ifdef CONFIG_SCHED_MC
7741 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7743 return sprintf(page, "%u\n", sched_mc_power_savings);
7745 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7746 const char *buf, size_t count)
7748 return sched_power_savings_store(buf, count, 0);
7750 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7751 sched_mc_power_savings_store);
7754 #ifdef CONFIG_SCHED_SMT
7755 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7757 return sprintf(page, "%u\n", sched_smt_power_savings);
7759 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7760 const char *buf, size_t count)
7762 return sched_power_savings_store(buf, count, 1);
7764 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7765 sched_smt_power_savings_store);
7768 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7772 #ifdef CONFIG_SCHED_SMT
7774 err = sysfs_create_file(&cls->kset.kobj,
7775 &attr_sched_smt_power_savings.attr);
7777 #ifdef CONFIG_SCHED_MC
7778 if (!err && mc_capable())
7779 err = sysfs_create_file(&cls->kset.kobj,
7780 &attr_sched_mc_power_savings.attr);
7784 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7787 * Force a reinitialization of the sched domains hierarchy. The domains
7788 * and groups cannot be updated in place without racing with the balancing
7789 * code, so we temporarily attach all running cpus to the NULL domain
7790 * which will prevent rebalancing while the sched domains are recalculated.
7792 static int update_sched_domains(struct notifier_block *nfb,
7793 unsigned long action, void *hcpu)
7795 int cpu = (int)(long)hcpu;
7798 case CPU_DOWN_PREPARE:
7799 case CPU_DOWN_PREPARE_FROZEN:
7800 disable_runtime(cpu_rq(cpu));
7802 case CPU_UP_PREPARE:
7803 case CPU_UP_PREPARE_FROZEN:
7804 detach_destroy_domains(&cpu_online_map);
7805 free_sched_domains();
7809 case CPU_DOWN_FAILED:
7810 case CPU_DOWN_FAILED_FROZEN:
7812 case CPU_ONLINE_FROZEN:
7813 enable_runtime(cpu_rq(cpu));
7815 case CPU_UP_CANCELED:
7816 case CPU_UP_CANCELED_FROZEN:
7818 case CPU_DEAD_FROZEN:
7820 * Fall through and re-initialise the domains.
7827 #ifndef CONFIG_CPUSETS
7829 * Create default domain partitioning if cpusets are disabled.
7830 * Otherwise we let cpusets rebuild the domains based on the
7834 /* The hotplug lock is already held by cpu_up/cpu_down */
7835 arch_init_sched_domains(&cpu_online_map);
7841 void __init sched_init_smp(void)
7843 cpumask_t non_isolated_cpus;
7845 #if defined(CONFIG_NUMA)
7846 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7848 BUG_ON(sched_group_nodes_bycpu == NULL);
7851 mutex_lock(&sched_domains_mutex);
7852 arch_init_sched_domains(&cpu_online_map);
7853 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7854 if (cpus_empty(non_isolated_cpus))
7855 cpu_set(smp_processor_id(), non_isolated_cpus);
7856 mutex_unlock(&sched_domains_mutex);
7858 /* XXX: Theoretical race here - CPU may be hotplugged now */
7859 hotcpu_notifier(update_sched_domains, 0);
7862 /* Move init over to a non-isolated CPU */
7863 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7865 sched_init_granularity();
7868 void __init sched_init_smp(void)
7870 sched_init_granularity();
7872 #endif /* CONFIG_SMP */
7874 int in_sched_functions(unsigned long addr)
7876 return in_lock_functions(addr) ||
7877 (addr >= (unsigned long)__sched_text_start
7878 && addr < (unsigned long)__sched_text_end);
7881 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7883 cfs_rq->tasks_timeline = RB_ROOT;
7884 INIT_LIST_HEAD(&cfs_rq->tasks);
7885 #ifdef CONFIG_FAIR_GROUP_SCHED
7888 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7891 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7893 struct rt_prio_array *array;
7896 array = &rt_rq->active;
7897 for (i = 0; i < MAX_RT_PRIO; i++) {
7898 INIT_LIST_HEAD(array->queue + i);
7899 __clear_bit(i, array->bitmap);
7901 /* delimiter for bitsearch: */
7902 __set_bit(MAX_RT_PRIO, array->bitmap);
7904 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7905 rt_rq->highest_prio = MAX_RT_PRIO;
7908 rt_rq->rt_nr_migratory = 0;
7909 rt_rq->overloaded = 0;
7913 rt_rq->rt_throttled = 0;
7914 rt_rq->rt_runtime = 0;
7915 spin_lock_init(&rt_rq->rt_runtime_lock);
7917 #ifdef CONFIG_RT_GROUP_SCHED
7918 rt_rq->rt_nr_boosted = 0;
7923 #ifdef CONFIG_FAIR_GROUP_SCHED
7924 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7925 struct sched_entity *se, int cpu, int add,
7926 struct sched_entity *parent)
7928 struct rq *rq = cpu_rq(cpu);
7929 tg->cfs_rq[cpu] = cfs_rq;
7930 init_cfs_rq(cfs_rq, rq);
7933 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7936 /* se could be NULL for init_task_group */
7941 se->cfs_rq = &rq->cfs;
7943 se->cfs_rq = parent->my_q;
7946 se->load.weight = tg->shares;
7947 se->load.inv_weight = 0;
7948 se->parent = parent;
7952 #ifdef CONFIG_RT_GROUP_SCHED
7953 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7954 struct sched_rt_entity *rt_se, int cpu, int add,
7955 struct sched_rt_entity *parent)
7957 struct rq *rq = cpu_rq(cpu);
7959 tg->rt_rq[cpu] = rt_rq;
7960 init_rt_rq(rt_rq, rq);
7962 rt_rq->rt_se = rt_se;
7963 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7965 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7967 tg->rt_se[cpu] = rt_se;
7972 rt_se->rt_rq = &rq->rt;
7974 rt_se->rt_rq = parent->my_q;
7976 rt_se->my_q = rt_rq;
7977 rt_se->parent = parent;
7978 INIT_LIST_HEAD(&rt_se->run_list);
7982 void __init sched_init(void)
7985 unsigned long alloc_size = 0, ptr;
7987 #ifdef CONFIG_FAIR_GROUP_SCHED
7988 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7990 #ifdef CONFIG_RT_GROUP_SCHED
7991 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7993 #ifdef CONFIG_USER_SCHED
7997 * As sched_init() is called before page_alloc is setup,
7998 * we use alloc_bootmem().
8001 ptr = (unsigned long)alloc_bootmem(alloc_size);
8003 #ifdef CONFIG_FAIR_GROUP_SCHED
8004 init_task_group.se = (struct sched_entity **)ptr;
8005 ptr += nr_cpu_ids * sizeof(void **);
8007 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8008 ptr += nr_cpu_ids * sizeof(void **);
8010 #ifdef CONFIG_USER_SCHED
8011 root_task_group.se = (struct sched_entity **)ptr;
8012 ptr += nr_cpu_ids * sizeof(void **);
8014 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8015 ptr += nr_cpu_ids * sizeof(void **);
8016 #endif /* CONFIG_USER_SCHED */
8017 #endif /* CONFIG_FAIR_GROUP_SCHED */
8018 #ifdef CONFIG_RT_GROUP_SCHED
8019 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8020 ptr += nr_cpu_ids * sizeof(void **);
8022 init_task_group.rt_rq = (struct rt_rq **)ptr;
8023 ptr += nr_cpu_ids * sizeof(void **);
8025 #ifdef CONFIG_USER_SCHED
8026 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8027 ptr += nr_cpu_ids * sizeof(void **);
8029 root_task_group.rt_rq = (struct rt_rq **)ptr;
8030 ptr += nr_cpu_ids * sizeof(void **);
8031 #endif /* CONFIG_USER_SCHED */
8032 #endif /* CONFIG_RT_GROUP_SCHED */
8037 init_defrootdomain();
8040 init_rt_bandwidth(&def_rt_bandwidth,
8041 global_rt_period(), global_rt_runtime());
8043 #ifdef CONFIG_RT_GROUP_SCHED
8044 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8045 global_rt_period(), global_rt_runtime());
8046 #ifdef CONFIG_USER_SCHED
8047 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8048 global_rt_period(), RUNTIME_INF);
8049 #endif /* CONFIG_USER_SCHED */
8050 #endif /* CONFIG_RT_GROUP_SCHED */
8052 #ifdef CONFIG_GROUP_SCHED
8053 list_add(&init_task_group.list, &task_groups);
8054 INIT_LIST_HEAD(&init_task_group.children);
8056 #ifdef CONFIG_USER_SCHED
8057 INIT_LIST_HEAD(&root_task_group.children);
8058 init_task_group.parent = &root_task_group;
8059 list_add(&init_task_group.siblings, &root_task_group.children);
8060 #endif /* CONFIG_USER_SCHED */
8061 #endif /* CONFIG_GROUP_SCHED */
8063 for_each_possible_cpu(i) {
8067 spin_lock_init(&rq->lock);
8068 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
8070 init_cfs_rq(&rq->cfs, rq);
8071 init_rt_rq(&rq->rt, rq);
8072 #ifdef CONFIG_FAIR_GROUP_SCHED
8073 init_task_group.shares = init_task_group_load;
8074 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8075 #ifdef CONFIG_CGROUP_SCHED
8077 * How much cpu bandwidth does init_task_group get?
8079 * In case of task-groups formed thr' the cgroup filesystem, it
8080 * gets 100% of the cpu resources in the system. This overall
8081 * system cpu resource is divided among the tasks of
8082 * init_task_group and its child task-groups in a fair manner,
8083 * based on each entity's (task or task-group's) weight
8084 * (se->load.weight).
8086 * In other words, if init_task_group has 10 tasks of weight
8087 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8088 * then A0's share of the cpu resource is:
8090 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8092 * We achieve this by letting init_task_group's tasks sit
8093 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8095 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8096 #elif defined CONFIG_USER_SCHED
8097 root_task_group.shares = NICE_0_LOAD;
8098 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8100 * In case of task-groups formed thr' the user id of tasks,
8101 * init_task_group represents tasks belonging to root user.
8102 * Hence it forms a sibling of all subsequent groups formed.
8103 * In this case, init_task_group gets only a fraction of overall
8104 * system cpu resource, based on the weight assigned to root
8105 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8106 * by letting tasks of init_task_group sit in a separate cfs_rq
8107 * (init_cfs_rq) and having one entity represent this group of
8108 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8110 init_tg_cfs_entry(&init_task_group,
8111 &per_cpu(init_cfs_rq, i),
8112 &per_cpu(init_sched_entity, i), i, 1,
8113 root_task_group.se[i]);
8116 #endif /* CONFIG_FAIR_GROUP_SCHED */
8118 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8119 #ifdef CONFIG_RT_GROUP_SCHED
8120 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8121 #ifdef CONFIG_CGROUP_SCHED
8122 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8123 #elif defined CONFIG_USER_SCHED
8124 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8125 init_tg_rt_entry(&init_task_group,
8126 &per_cpu(init_rt_rq, i),
8127 &per_cpu(init_sched_rt_entity, i), i, 1,
8128 root_task_group.rt_se[i]);
8132 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8133 rq->cpu_load[j] = 0;
8137 rq->active_balance = 0;
8138 rq->next_balance = jiffies;
8142 rq->migration_thread = NULL;
8143 INIT_LIST_HEAD(&rq->migration_queue);
8144 rq_attach_root(rq, &def_root_domain);
8147 atomic_set(&rq->nr_iowait, 0);
8150 set_load_weight(&init_task);
8152 #ifdef CONFIG_PREEMPT_NOTIFIERS
8153 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8157 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8160 #ifdef CONFIG_RT_MUTEXES
8161 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8165 * The boot idle thread does lazy MMU switching as well:
8167 atomic_inc(&init_mm.mm_count);
8168 enter_lazy_tlb(&init_mm, current);
8171 * Make us the idle thread. Technically, schedule() should not be
8172 * called from this thread, however somewhere below it might be,
8173 * but because we are the idle thread, we just pick up running again
8174 * when this runqueue becomes "idle".
8176 init_idle(current, smp_processor_id());
8178 * During early bootup we pretend to be a normal task:
8180 current->sched_class = &fair_sched_class;
8182 scheduler_running = 1;
8185 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8186 void __might_sleep(char *file, int line)
8189 static unsigned long prev_jiffy; /* ratelimiting */
8191 if ((in_atomic() || irqs_disabled()) &&
8192 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8193 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8195 prev_jiffy = jiffies;
8196 printk(KERN_ERR "BUG: sleeping function called from invalid"
8197 " context at %s:%d\n", file, line);
8198 printk("in_atomic():%d, irqs_disabled():%d\n",
8199 in_atomic(), irqs_disabled());
8200 debug_show_held_locks(current);
8201 if (irqs_disabled())
8202 print_irqtrace_events(current);
8207 EXPORT_SYMBOL(__might_sleep);
8210 #ifdef CONFIG_MAGIC_SYSRQ
8211 static void normalize_task(struct rq *rq, struct task_struct *p)
8215 update_rq_clock(rq);
8216 on_rq = p->se.on_rq;
8218 deactivate_task(rq, p, 0);
8219 __setscheduler(rq, p, SCHED_NORMAL, 0);
8221 activate_task(rq, p, 0);
8222 resched_task(rq->curr);
8226 void normalize_rt_tasks(void)
8228 struct task_struct *g, *p;
8229 unsigned long flags;
8232 read_lock_irqsave(&tasklist_lock, flags);
8233 do_each_thread(g, p) {
8235 * Only normalize user tasks:
8240 p->se.exec_start = 0;
8241 #ifdef CONFIG_SCHEDSTATS
8242 p->se.wait_start = 0;
8243 p->se.sleep_start = 0;
8244 p->se.block_start = 0;
8249 * Renice negative nice level userspace
8252 if (TASK_NICE(p) < 0 && p->mm)
8253 set_user_nice(p, 0);
8257 spin_lock(&p->pi_lock);
8258 rq = __task_rq_lock(p);
8260 normalize_task(rq, p);
8262 __task_rq_unlock(rq);
8263 spin_unlock(&p->pi_lock);
8264 } while_each_thread(g, p);
8266 read_unlock_irqrestore(&tasklist_lock, flags);
8269 #endif /* CONFIG_MAGIC_SYSRQ */
8273 * These functions are only useful for the IA64 MCA handling.
8275 * They can only be called when the whole system has been
8276 * stopped - every CPU needs to be quiescent, and no scheduling
8277 * activity can take place. Using them for anything else would
8278 * be a serious bug, and as a result, they aren't even visible
8279 * under any other configuration.
8283 * curr_task - return the current task for a given cpu.
8284 * @cpu: the processor in question.
8286 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8288 struct task_struct *curr_task(int cpu)
8290 return cpu_curr(cpu);
8294 * set_curr_task - set the current task for a given cpu.
8295 * @cpu: the processor in question.
8296 * @p: the task pointer to set.
8298 * Description: This function must only be used when non-maskable interrupts
8299 * are serviced on a separate stack. It allows the architecture to switch the
8300 * notion of the current task on a cpu in a non-blocking manner. This function
8301 * must be called with all CPU's synchronized, and interrupts disabled, the
8302 * and caller must save the original value of the current task (see
8303 * curr_task() above) and restore that value before reenabling interrupts and
8304 * re-starting the system.
8306 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8308 void set_curr_task(int cpu, struct task_struct *p)
8315 #ifdef CONFIG_FAIR_GROUP_SCHED
8316 static void free_fair_sched_group(struct task_group *tg)
8320 for_each_possible_cpu(i) {
8322 kfree(tg->cfs_rq[i]);
8332 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8334 struct cfs_rq *cfs_rq;
8335 struct sched_entity *se, *parent_se;
8339 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8342 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8346 tg->shares = NICE_0_LOAD;
8348 for_each_possible_cpu(i) {
8351 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8352 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8356 se = kmalloc_node(sizeof(struct sched_entity),
8357 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8361 parent_se = parent ? parent->se[i] : NULL;
8362 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8371 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8373 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8374 &cpu_rq(cpu)->leaf_cfs_rq_list);
8377 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8379 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8381 #else /* !CONFG_FAIR_GROUP_SCHED */
8382 static inline void free_fair_sched_group(struct task_group *tg)
8387 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8392 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8396 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8399 #endif /* CONFIG_FAIR_GROUP_SCHED */
8401 #ifdef CONFIG_RT_GROUP_SCHED
8402 static void free_rt_sched_group(struct task_group *tg)
8406 destroy_rt_bandwidth(&tg->rt_bandwidth);
8408 for_each_possible_cpu(i) {
8410 kfree(tg->rt_rq[i]);
8412 kfree(tg->rt_se[i]);
8420 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8422 struct rt_rq *rt_rq;
8423 struct sched_rt_entity *rt_se, *parent_se;
8427 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8430 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8434 init_rt_bandwidth(&tg->rt_bandwidth,
8435 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8437 for_each_possible_cpu(i) {
8440 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8441 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8445 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8446 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8450 parent_se = parent ? parent->rt_se[i] : NULL;
8451 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8460 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8462 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8463 &cpu_rq(cpu)->leaf_rt_rq_list);
8466 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8468 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8470 #else /* !CONFIG_RT_GROUP_SCHED */
8471 static inline void free_rt_sched_group(struct task_group *tg)
8476 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8481 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8485 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8488 #endif /* CONFIG_RT_GROUP_SCHED */
8490 #ifdef CONFIG_GROUP_SCHED
8491 static void free_sched_group(struct task_group *tg)
8493 free_fair_sched_group(tg);
8494 free_rt_sched_group(tg);
8498 /* allocate runqueue etc for a new task group */
8499 struct task_group *sched_create_group(struct task_group *parent)
8501 struct task_group *tg;
8502 unsigned long flags;
8505 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8507 return ERR_PTR(-ENOMEM);
8509 if (!alloc_fair_sched_group(tg, parent))
8512 if (!alloc_rt_sched_group(tg, parent))
8515 spin_lock_irqsave(&task_group_lock, flags);
8516 for_each_possible_cpu(i) {
8517 register_fair_sched_group(tg, i);
8518 register_rt_sched_group(tg, i);
8520 list_add_rcu(&tg->list, &task_groups);
8522 WARN_ON(!parent); /* root should already exist */
8524 tg->parent = parent;
8525 list_add_rcu(&tg->siblings, &parent->children);
8526 INIT_LIST_HEAD(&tg->children);
8527 spin_unlock_irqrestore(&task_group_lock, flags);
8532 free_sched_group(tg);
8533 return ERR_PTR(-ENOMEM);
8536 /* rcu callback to free various structures associated with a task group */
8537 static void free_sched_group_rcu(struct rcu_head *rhp)
8539 /* now it should be safe to free those cfs_rqs */
8540 free_sched_group(container_of(rhp, struct task_group, rcu));
8543 /* Destroy runqueue etc associated with a task group */
8544 void sched_destroy_group(struct task_group *tg)
8546 unsigned long flags;
8549 spin_lock_irqsave(&task_group_lock, flags);
8550 for_each_possible_cpu(i) {
8551 unregister_fair_sched_group(tg, i);
8552 unregister_rt_sched_group(tg, i);
8554 list_del_rcu(&tg->list);
8555 list_del_rcu(&tg->siblings);
8556 spin_unlock_irqrestore(&task_group_lock, flags);
8558 /* wait for possible concurrent references to cfs_rqs complete */
8559 call_rcu(&tg->rcu, free_sched_group_rcu);
8562 /* change task's runqueue when it moves between groups.
8563 * The caller of this function should have put the task in its new group
8564 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8565 * reflect its new group.
8567 void sched_move_task(struct task_struct *tsk)
8570 unsigned long flags;
8573 rq = task_rq_lock(tsk, &flags);
8575 update_rq_clock(rq);
8577 running = task_current(rq, tsk);
8578 on_rq = tsk->se.on_rq;
8581 dequeue_task(rq, tsk, 0);
8582 if (unlikely(running))
8583 tsk->sched_class->put_prev_task(rq, tsk);
8585 set_task_rq(tsk, task_cpu(tsk));
8587 #ifdef CONFIG_FAIR_GROUP_SCHED
8588 if (tsk->sched_class->moved_group)
8589 tsk->sched_class->moved_group(tsk);
8592 if (unlikely(running))
8593 tsk->sched_class->set_curr_task(rq);
8595 enqueue_task(rq, tsk, 0);
8597 task_rq_unlock(rq, &flags);
8599 #endif /* CONFIG_GROUP_SCHED */
8601 #ifdef CONFIG_FAIR_GROUP_SCHED
8602 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8604 struct cfs_rq *cfs_rq = se->cfs_rq;
8609 dequeue_entity(cfs_rq, se, 0);
8611 se->load.weight = shares;
8612 se->load.inv_weight = 0;
8615 enqueue_entity(cfs_rq, se, 0);
8618 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8620 struct cfs_rq *cfs_rq = se->cfs_rq;
8621 struct rq *rq = cfs_rq->rq;
8622 unsigned long flags;
8624 spin_lock_irqsave(&rq->lock, flags);
8625 __set_se_shares(se, shares);
8626 spin_unlock_irqrestore(&rq->lock, flags);
8629 static DEFINE_MUTEX(shares_mutex);
8631 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8634 unsigned long flags;
8637 * We can't change the weight of the root cgroup.
8642 if (shares < MIN_SHARES)
8643 shares = MIN_SHARES;
8644 else if (shares > MAX_SHARES)
8645 shares = MAX_SHARES;
8647 mutex_lock(&shares_mutex);
8648 if (tg->shares == shares)
8651 spin_lock_irqsave(&task_group_lock, flags);
8652 for_each_possible_cpu(i)
8653 unregister_fair_sched_group(tg, i);
8654 list_del_rcu(&tg->siblings);
8655 spin_unlock_irqrestore(&task_group_lock, flags);
8657 /* wait for any ongoing reference to this group to finish */
8658 synchronize_sched();
8661 * Now we are free to modify the group's share on each cpu
8662 * w/o tripping rebalance_share or load_balance_fair.
8664 tg->shares = shares;
8665 for_each_possible_cpu(i) {
8669 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8670 set_se_shares(tg->se[i], shares);
8674 * Enable load balance activity on this group, by inserting it back on
8675 * each cpu's rq->leaf_cfs_rq_list.
8677 spin_lock_irqsave(&task_group_lock, flags);
8678 for_each_possible_cpu(i)
8679 register_fair_sched_group(tg, i);
8680 list_add_rcu(&tg->siblings, &tg->parent->children);
8681 spin_unlock_irqrestore(&task_group_lock, flags);
8683 mutex_unlock(&shares_mutex);
8687 unsigned long sched_group_shares(struct task_group *tg)
8693 #ifdef CONFIG_RT_GROUP_SCHED
8695 * Ensure that the real time constraints are schedulable.
8697 static DEFINE_MUTEX(rt_constraints_mutex);
8699 static unsigned long to_ratio(u64 period, u64 runtime)
8701 if (runtime == RUNTIME_INF)
8704 return div64_u64(runtime << 16, period);
8707 #ifdef CONFIG_CGROUP_SCHED
8708 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8710 struct task_group *tgi, *parent = tg->parent;
8711 unsigned long total = 0;
8714 if (global_rt_period() < period)
8717 return to_ratio(period, runtime) <
8718 to_ratio(global_rt_period(), global_rt_runtime());
8721 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8725 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8729 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8730 tgi->rt_bandwidth.rt_runtime);
8734 return total + to_ratio(period, runtime) <=
8735 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8736 parent->rt_bandwidth.rt_runtime);
8738 #elif defined CONFIG_USER_SCHED
8739 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8741 struct task_group *tgi;
8742 unsigned long total = 0;
8743 unsigned long global_ratio =
8744 to_ratio(global_rt_period(), global_rt_runtime());
8747 list_for_each_entry_rcu(tgi, &task_groups, list) {
8751 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8752 tgi->rt_bandwidth.rt_runtime);
8756 return total + to_ratio(period, runtime) < global_ratio;
8760 /* Must be called with tasklist_lock held */
8761 static inline int tg_has_rt_tasks(struct task_group *tg)
8763 struct task_struct *g, *p;
8764 do_each_thread(g, p) {
8765 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8767 } while_each_thread(g, p);
8771 static int tg_set_bandwidth(struct task_group *tg,
8772 u64 rt_period, u64 rt_runtime)
8776 mutex_lock(&rt_constraints_mutex);
8777 read_lock(&tasklist_lock);
8778 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8782 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8787 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8788 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8789 tg->rt_bandwidth.rt_runtime = rt_runtime;
8791 for_each_possible_cpu(i) {
8792 struct rt_rq *rt_rq = tg->rt_rq[i];
8794 spin_lock(&rt_rq->rt_runtime_lock);
8795 rt_rq->rt_runtime = rt_runtime;
8796 spin_unlock(&rt_rq->rt_runtime_lock);
8798 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8800 read_unlock(&tasklist_lock);
8801 mutex_unlock(&rt_constraints_mutex);
8806 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8808 u64 rt_runtime, rt_period;
8810 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8811 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8812 if (rt_runtime_us < 0)
8813 rt_runtime = RUNTIME_INF;
8815 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8818 long sched_group_rt_runtime(struct task_group *tg)
8822 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8825 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8826 do_div(rt_runtime_us, NSEC_PER_USEC);
8827 return rt_runtime_us;
8830 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8832 u64 rt_runtime, rt_period;
8834 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8835 rt_runtime = tg->rt_bandwidth.rt_runtime;
8837 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8840 long sched_group_rt_period(struct task_group *tg)
8844 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8845 do_div(rt_period_us, NSEC_PER_USEC);
8846 return rt_period_us;
8849 static int sched_rt_global_constraints(void)
8851 struct task_group *tg = &root_task_group;
8852 u64 rt_runtime, rt_period;
8855 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8856 rt_runtime = tg->rt_bandwidth.rt_runtime;
8858 mutex_lock(&rt_constraints_mutex);
8859 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8861 mutex_unlock(&rt_constraints_mutex);
8865 #else /* !CONFIG_RT_GROUP_SCHED */
8866 static int sched_rt_global_constraints(void)
8868 unsigned long flags;
8871 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8872 for_each_possible_cpu(i) {
8873 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8875 spin_lock(&rt_rq->rt_runtime_lock);
8876 rt_rq->rt_runtime = global_rt_runtime();
8877 spin_unlock(&rt_rq->rt_runtime_lock);
8879 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8883 #endif /* CONFIG_RT_GROUP_SCHED */
8885 int sched_rt_handler(struct ctl_table *table, int write,
8886 struct file *filp, void __user *buffer, size_t *lenp,
8890 int old_period, old_runtime;
8891 static DEFINE_MUTEX(mutex);
8894 old_period = sysctl_sched_rt_period;
8895 old_runtime = sysctl_sched_rt_runtime;
8897 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8899 if (!ret && write) {
8900 ret = sched_rt_global_constraints();
8902 sysctl_sched_rt_period = old_period;
8903 sysctl_sched_rt_runtime = old_runtime;
8905 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8906 def_rt_bandwidth.rt_period =
8907 ns_to_ktime(global_rt_period());
8910 mutex_unlock(&mutex);
8915 #ifdef CONFIG_CGROUP_SCHED
8917 /* return corresponding task_group object of a cgroup */
8918 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8920 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8921 struct task_group, css);
8924 static struct cgroup_subsys_state *
8925 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8927 struct task_group *tg, *parent;
8929 if (!cgrp->parent) {
8930 /* This is early initialization for the top cgroup */
8931 init_task_group.css.cgroup = cgrp;
8932 return &init_task_group.css;
8935 parent = cgroup_tg(cgrp->parent);
8936 tg = sched_create_group(parent);
8938 return ERR_PTR(-ENOMEM);
8940 /* Bind the cgroup to task_group object we just created */
8941 tg->css.cgroup = cgrp;
8947 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8949 struct task_group *tg = cgroup_tg(cgrp);
8951 sched_destroy_group(tg);
8955 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8956 struct task_struct *tsk)
8958 #ifdef CONFIG_RT_GROUP_SCHED
8959 /* Don't accept realtime tasks when there is no way for them to run */
8960 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8963 /* We don't support RT-tasks being in separate groups */
8964 if (tsk->sched_class != &fair_sched_class)
8972 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8973 struct cgroup *old_cont, struct task_struct *tsk)
8975 sched_move_task(tsk);
8978 #ifdef CONFIG_FAIR_GROUP_SCHED
8979 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8982 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8985 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8987 struct task_group *tg = cgroup_tg(cgrp);
8989 return (u64) tg->shares;
8991 #endif /* CONFIG_FAIR_GROUP_SCHED */
8993 #ifdef CONFIG_RT_GROUP_SCHED
8994 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8997 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9000 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9002 return sched_group_rt_runtime(cgroup_tg(cgrp));
9005 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9008 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9011 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9013 return sched_group_rt_period(cgroup_tg(cgrp));
9015 #endif /* CONFIG_RT_GROUP_SCHED */
9017 static struct cftype cpu_files[] = {
9018 #ifdef CONFIG_FAIR_GROUP_SCHED
9021 .read_u64 = cpu_shares_read_u64,
9022 .write_u64 = cpu_shares_write_u64,
9025 #ifdef CONFIG_RT_GROUP_SCHED
9027 .name = "rt_runtime_us",
9028 .read_s64 = cpu_rt_runtime_read,
9029 .write_s64 = cpu_rt_runtime_write,
9032 .name = "rt_period_us",
9033 .read_u64 = cpu_rt_period_read_uint,
9034 .write_u64 = cpu_rt_period_write_uint,
9039 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9041 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9044 struct cgroup_subsys cpu_cgroup_subsys = {
9046 .create = cpu_cgroup_create,
9047 .destroy = cpu_cgroup_destroy,
9048 .can_attach = cpu_cgroup_can_attach,
9049 .attach = cpu_cgroup_attach,
9050 .populate = cpu_cgroup_populate,
9051 .subsys_id = cpu_cgroup_subsys_id,
9055 #endif /* CONFIG_CGROUP_SCHED */
9057 #ifdef CONFIG_CGROUP_CPUACCT
9060 * CPU accounting code for task groups.
9062 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9063 * (balbir@in.ibm.com).
9066 /* track cpu usage of a group of tasks */
9068 struct cgroup_subsys_state css;
9069 /* cpuusage holds pointer to a u64-type object on every cpu */
9073 struct cgroup_subsys cpuacct_subsys;
9075 /* return cpu accounting group corresponding to this container */
9076 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9078 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9079 struct cpuacct, css);
9082 /* return cpu accounting group to which this task belongs */
9083 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9085 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9086 struct cpuacct, css);
9089 /* create a new cpu accounting group */
9090 static struct cgroup_subsys_state *cpuacct_create(
9091 struct cgroup_subsys *ss, struct cgroup *cgrp)
9093 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9096 return ERR_PTR(-ENOMEM);
9098 ca->cpuusage = alloc_percpu(u64);
9099 if (!ca->cpuusage) {
9101 return ERR_PTR(-ENOMEM);
9107 /* destroy an existing cpu accounting group */
9109 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9111 struct cpuacct *ca = cgroup_ca(cgrp);
9113 free_percpu(ca->cpuusage);
9117 /* return total cpu usage (in nanoseconds) of a group */
9118 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9120 struct cpuacct *ca = cgroup_ca(cgrp);
9121 u64 totalcpuusage = 0;
9124 for_each_possible_cpu(i) {
9125 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9128 * Take rq->lock to make 64-bit addition safe on 32-bit
9131 spin_lock_irq(&cpu_rq(i)->lock);
9132 totalcpuusage += *cpuusage;
9133 spin_unlock_irq(&cpu_rq(i)->lock);
9136 return totalcpuusage;
9139 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9142 struct cpuacct *ca = cgroup_ca(cgrp);
9151 for_each_possible_cpu(i) {
9152 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9154 spin_lock_irq(&cpu_rq(i)->lock);
9156 spin_unlock_irq(&cpu_rq(i)->lock);
9162 static struct cftype files[] = {
9165 .read_u64 = cpuusage_read,
9166 .write_u64 = cpuusage_write,
9170 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9172 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9176 * charge this task's execution time to its accounting group.
9178 * called with rq->lock held.
9180 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9184 if (!cpuacct_subsys.active)
9189 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9191 *cpuusage += cputime;
9195 struct cgroup_subsys cpuacct_subsys = {
9197 .create = cpuacct_create,
9198 .destroy = cpuacct_destroy,
9199 .populate = cpuacct_populate,
9200 .subsys_id = cpuacct_subsys_id,
9202 #endif /* CONFIG_CGROUP_CPUACCT */