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 aggregate(tg, cpu)->shares = shares;
1558 * Compute the load fraction assigned to this group, relies on the aggregate
1559 * weight and this group's parent's load, i.e. top-down.
1562 aggregate_group_load(struct task_group *tg, int cpu, struct sched_domain *sd)
1570 for_each_cpu_mask(i, sd->span)
1571 load += cpu_rq(i)->load.weight;
1574 load = aggregate(tg->parent, cpu)->load;
1577 * shares is our weight in the parent's rq so
1578 * shares/parent->rq_weight gives our fraction of the load
1580 load *= aggregate(tg, cpu)->shares;
1581 load /= aggregate(tg->parent, cpu)->rq_weight + 1;
1584 aggregate(tg, cpu)->load = load;
1587 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1590 * Calculate and set the cpu's group shares.
1593 __update_group_shares_cpu(struct task_group *tg, int cpu,
1594 struct sched_domain *sd, int tcpu)
1597 unsigned long shares;
1598 unsigned long rq_weight;
1603 rq_weight = tg->cfs_rq[tcpu]->load.weight;
1606 * If there are currently no tasks on the cpu pretend there is one of
1607 * average load so that when a new task gets to run here it will not
1608 * get delayed by group starvation.
1612 rq_weight = NICE_0_LOAD;
1616 * \Sum shares * rq_weight
1617 * shares = -----------------------
1621 shares = aggregate(tg, cpu)->shares * rq_weight;
1622 shares /= aggregate(tg, cpu)->rq_weight + 1;
1625 * record the actual number of shares, not the boosted amount.
1627 tg->cfs_rq[tcpu]->shares = boost ? 0 : shares;
1629 if (shares < MIN_SHARES)
1630 shares = MIN_SHARES;
1631 else if (shares > MAX_SHARES)
1632 shares = MAX_SHARES;
1634 __set_se_shares(tg->se[tcpu], shares);
1638 * Re-adjust the weights on the cpu the task came from and on the cpu the
1642 __move_group_shares(struct task_group *tg, int cpu, struct sched_domain *sd,
1645 unsigned long shares;
1647 shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1649 __update_group_shares_cpu(tg, cpu, sd, scpu);
1650 __update_group_shares_cpu(tg, cpu, sd, dcpu);
1653 * ensure we never loose shares due to rounding errors in the
1654 * above redistribution.
1656 shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1658 tg->cfs_rq[dcpu]->shares += shares;
1662 * Because changing a group's shares changes the weight of the super-group
1663 * we need to walk up the tree and change all shares until we hit the root.
1666 move_group_shares(struct task_group *tg, int cpu, struct sched_domain *sd,
1670 __move_group_shares(tg, cpu, sd, scpu, dcpu);
1676 aggregate_group_set_shares(struct task_group *tg, int cpu, struct sched_domain *sd)
1678 unsigned long shares = aggregate(tg, cpu)->shares;
1681 for_each_cpu_mask(i, sd->span) {
1682 struct rq *rq = cpu_rq(i);
1683 unsigned long flags;
1685 spin_lock_irqsave(&rq->lock, flags);
1686 __update_group_shares_cpu(tg, cpu, sd, i);
1687 spin_unlock_irqrestore(&rq->lock, flags);
1690 aggregate_group_shares(tg, cpu, sd);
1693 * ensure we never loose shares due to rounding errors in the
1694 * above redistribution.
1696 shares -= aggregate(tg, cpu)->shares;
1698 tg->cfs_rq[cpu]->shares += shares;
1699 aggregate(tg, cpu)->shares += shares;
1704 * Calculate the accumulative weight and recursive load of each task group
1705 * while walking down the tree.
1708 aggregate_get_down(struct task_group *tg, int cpu, struct sched_domain *sd)
1710 aggregate_group_weight(tg, cpu, sd);
1711 aggregate_group_shares(tg, cpu, sd);
1712 aggregate_group_load(tg, cpu, sd);
1716 * Rebalance the cpu shares while walking back up the tree.
1719 aggregate_get_up(struct task_group *tg, int cpu, struct sched_domain *sd)
1721 aggregate_group_set_shares(tg, cpu, sd);
1725 aggregate_get_nop(struct task_group *tg, int cpu, struct sched_domain *sd)
1729 static DEFINE_PER_CPU(spinlock_t, aggregate_lock);
1731 static void __init init_aggregate(void)
1735 for_each_possible_cpu(i)
1736 spin_lock_init(&per_cpu(aggregate_lock, i));
1739 static int get_aggregate(int cpu, struct sched_domain *sd)
1741 if (!spin_trylock(&per_cpu(aggregate_lock, cpu)))
1744 aggregate_walk_tree(aggregate_get_down, aggregate_get_up, cpu, sd);
1748 static void update_aggregate(int cpu, struct sched_domain *sd)
1750 aggregate_walk_tree(aggregate_get_down, aggregate_get_nop, cpu, sd);
1753 static void put_aggregate(int cpu, struct sched_domain *sd)
1755 spin_unlock(&per_cpu(aggregate_lock, cpu));
1758 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1760 cfs_rq->shares = shares;
1765 static inline void init_aggregate(void)
1769 static inline int get_aggregate(int cpu, struct sched_domain *sd)
1774 static inline void update_aggregate(int cpu, struct sched_domain *sd)
1778 static inline void put_aggregate(int cpu, struct sched_domain *sd)
1785 #include "sched_stats.h"
1786 #include "sched_idletask.c"
1787 #include "sched_fair.c"
1788 #include "sched_rt.c"
1789 #ifdef CONFIG_SCHED_DEBUG
1790 # include "sched_debug.c"
1793 #define sched_class_highest (&rt_sched_class)
1794 #define for_each_class(class) \
1795 for (class = sched_class_highest; class; class = class->next)
1797 static void inc_nr_running(struct rq *rq)
1802 static void dec_nr_running(struct rq *rq)
1807 static void set_load_weight(struct task_struct *p)
1809 if (task_has_rt_policy(p)) {
1810 p->se.load.weight = prio_to_weight[0] * 2;
1811 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1816 * SCHED_IDLE tasks get minimal weight:
1818 if (p->policy == SCHED_IDLE) {
1819 p->se.load.weight = WEIGHT_IDLEPRIO;
1820 p->se.load.inv_weight = WMULT_IDLEPRIO;
1824 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1825 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1828 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1830 sched_info_queued(p);
1831 p->sched_class->enqueue_task(rq, p, wakeup);
1835 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1837 p->sched_class->dequeue_task(rq, p, sleep);
1842 * __normal_prio - return the priority that is based on the static prio
1844 static inline int __normal_prio(struct task_struct *p)
1846 return p->static_prio;
1850 * Calculate the expected normal priority: i.e. priority
1851 * without taking RT-inheritance into account. Might be
1852 * boosted by interactivity modifiers. Changes upon fork,
1853 * setprio syscalls, and whenever the interactivity
1854 * estimator recalculates.
1856 static inline int normal_prio(struct task_struct *p)
1860 if (task_has_rt_policy(p))
1861 prio = MAX_RT_PRIO-1 - p->rt_priority;
1863 prio = __normal_prio(p);
1868 * Calculate the current priority, i.e. the priority
1869 * taken into account by the scheduler. This value might
1870 * be boosted by RT tasks, or might be boosted by
1871 * interactivity modifiers. Will be RT if the task got
1872 * RT-boosted. If not then it returns p->normal_prio.
1874 static int effective_prio(struct task_struct *p)
1876 p->normal_prio = normal_prio(p);
1878 * If we are RT tasks or we were boosted to RT priority,
1879 * keep the priority unchanged. Otherwise, update priority
1880 * to the normal priority:
1882 if (!rt_prio(p->prio))
1883 return p->normal_prio;
1888 * activate_task - move a task to the runqueue.
1890 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1892 if (task_contributes_to_load(p))
1893 rq->nr_uninterruptible--;
1895 enqueue_task(rq, p, wakeup);
1900 * deactivate_task - remove a task from the runqueue.
1902 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1904 if (task_contributes_to_load(p))
1905 rq->nr_uninterruptible++;
1907 dequeue_task(rq, p, sleep);
1912 * task_curr - is this task currently executing on a CPU?
1913 * @p: the task in question.
1915 inline int task_curr(const struct task_struct *p)
1917 return cpu_curr(task_cpu(p)) == p;
1920 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1922 set_task_rq(p, cpu);
1925 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1926 * successfuly executed on another CPU. We must ensure that updates of
1927 * per-task data have been completed by this moment.
1930 task_thread_info(p)->cpu = cpu;
1934 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1935 const struct sched_class *prev_class,
1936 int oldprio, int running)
1938 if (prev_class != p->sched_class) {
1939 if (prev_class->switched_from)
1940 prev_class->switched_from(rq, p, running);
1941 p->sched_class->switched_to(rq, p, running);
1943 p->sched_class->prio_changed(rq, p, oldprio, running);
1948 /* Used instead of source_load when we know the type == 0 */
1949 static unsigned long weighted_cpuload(const int cpu)
1951 return cpu_rq(cpu)->load.weight;
1955 * Is this task likely cache-hot:
1958 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1963 * Buddy candidates are cache hot:
1965 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
1968 if (p->sched_class != &fair_sched_class)
1971 if (sysctl_sched_migration_cost == -1)
1973 if (sysctl_sched_migration_cost == 0)
1976 delta = now - p->se.exec_start;
1978 return delta < (s64)sysctl_sched_migration_cost;
1982 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1984 int old_cpu = task_cpu(p);
1985 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1986 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1987 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1990 clock_offset = old_rq->clock - new_rq->clock;
1992 #ifdef CONFIG_SCHEDSTATS
1993 if (p->se.wait_start)
1994 p->se.wait_start -= clock_offset;
1995 if (p->se.sleep_start)
1996 p->se.sleep_start -= clock_offset;
1997 if (p->se.block_start)
1998 p->se.block_start -= clock_offset;
1999 if (old_cpu != new_cpu) {
2000 schedstat_inc(p, se.nr_migrations);
2001 if (task_hot(p, old_rq->clock, NULL))
2002 schedstat_inc(p, se.nr_forced2_migrations);
2005 p->se.vruntime -= old_cfsrq->min_vruntime -
2006 new_cfsrq->min_vruntime;
2008 __set_task_cpu(p, new_cpu);
2011 struct migration_req {
2012 struct list_head list;
2014 struct task_struct *task;
2017 struct completion done;
2021 * The task's runqueue lock must be held.
2022 * Returns true if you have to wait for migration thread.
2025 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2027 struct rq *rq = task_rq(p);
2030 * If the task is not on a runqueue (and not running), then
2031 * it is sufficient to simply update the task's cpu field.
2033 if (!p->se.on_rq && !task_running(rq, p)) {
2034 set_task_cpu(p, dest_cpu);
2038 init_completion(&req->done);
2040 req->dest_cpu = dest_cpu;
2041 list_add(&req->list, &rq->migration_queue);
2047 * wait_task_inactive - wait for a thread to unschedule.
2049 * The caller must ensure that the task *will* unschedule sometime soon,
2050 * else this function might spin for a *long* time. This function can't
2051 * be called with interrupts off, or it may introduce deadlock with
2052 * smp_call_function() if an IPI is sent by the same process we are
2053 * waiting to become inactive.
2055 void wait_task_inactive(struct task_struct *p)
2057 unsigned long flags;
2063 * We do the initial early heuristics without holding
2064 * any task-queue locks at all. We'll only try to get
2065 * the runqueue lock when things look like they will
2071 * If the task is actively running on another CPU
2072 * still, just relax and busy-wait without holding
2075 * NOTE! Since we don't hold any locks, it's not
2076 * even sure that "rq" stays as the right runqueue!
2077 * But we don't care, since "task_running()" will
2078 * return false if the runqueue has changed and p
2079 * is actually now running somewhere else!
2081 while (task_running(rq, p))
2085 * Ok, time to look more closely! We need the rq
2086 * lock now, to be *sure*. If we're wrong, we'll
2087 * just go back and repeat.
2089 rq = task_rq_lock(p, &flags);
2090 running = task_running(rq, p);
2091 on_rq = p->se.on_rq;
2092 task_rq_unlock(rq, &flags);
2095 * Was it really running after all now that we
2096 * checked with the proper locks actually held?
2098 * Oops. Go back and try again..
2100 if (unlikely(running)) {
2106 * It's not enough that it's not actively running,
2107 * it must be off the runqueue _entirely_, and not
2110 * So if it wa still runnable (but just not actively
2111 * running right now), it's preempted, and we should
2112 * yield - it could be a while.
2114 if (unlikely(on_rq)) {
2115 schedule_timeout_uninterruptible(1);
2120 * Ahh, all good. It wasn't running, and it wasn't
2121 * runnable, which means that it will never become
2122 * running in the future either. We're all done!
2129 * kick_process - kick a running thread to enter/exit the kernel
2130 * @p: the to-be-kicked thread
2132 * Cause a process which is running on another CPU to enter
2133 * kernel-mode, without any delay. (to get signals handled.)
2135 * NOTE: this function doesnt have to take the runqueue lock,
2136 * because all it wants to ensure is that the remote task enters
2137 * the kernel. If the IPI races and the task has been migrated
2138 * to another CPU then no harm is done and the purpose has been
2141 void kick_process(struct task_struct *p)
2147 if ((cpu != smp_processor_id()) && task_curr(p))
2148 smp_send_reschedule(cpu);
2153 * Return a low guess at the load of a migration-source cpu weighted
2154 * according to the scheduling class and "nice" value.
2156 * We want to under-estimate the load of migration sources, to
2157 * balance conservatively.
2159 static unsigned long source_load(int cpu, int type)
2161 struct rq *rq = cpu_rq(cpu);
2162 unsigned long total = weighted_cpuload(cpu);
2167 return min(rq->cpu_load[type-1], total);
2171 * Return a high guess at the load of a migration-target cpu weighted
2172 * according to the scheduling class and "nice" value.
2174 static unsigned long target_load(int cpu, int type)
2176 struct rq *rq = cpu_rq(cpu);
2177 unsigned long total = weighted_cpuload(cpu);
2182 return max(rq->cpu_load[type-1], total);
2186 * Return the average load per task on the cpu's run queue
2188 static unsigned long cpu_avg_load_per_task(int cpu)
2190 struct rq *rq = cpu_rq(cpu);
2191 unsigned long total = weighted_cpuload(cpu);
2192 unsigned long n = rq->nr_running;
2194 return n ? total / n : SCHED_LOAD_SCALE;
2198 * find_idlest_group finds and returns the least busy CPU group within the
2201 static struct sched_group *
2202 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2204 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2205 unsigned long min_load = ULONG_MAX, this_load = 0;
2206 int load_idx = sd->forkexec_idx;
2207 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2210 * now that we have both rqs locked the rq weight won't change
2211 * anymore - so update the stats.
2213 update_aggregate(this_cpu, sd);
2216 unsigned long load, avg_load;
2220 /* Skip over this group if it has no CPUs allowed */
2221 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2224 local_group = cpu_isset(this_cpu, group->cpumask);
2226 /* Tally up the load of all CPUs in the group */
2229 for_each_cpu_mask(i, group->cpumask) {
2230 /* Bias balancing toward cpus of our domain */
2232 load = source_load(i, load_idx);
2234 load = target_load(i, load_idx);
2239 /* Adjust by relative CPU power of the group */
2240 avg_load = sg_div_cpu_power(group,
2241 avg_load * SCHED_LOAD_SCALE);
2244 this_load = avg_load;
2246 } else if (avg_load < min_load) {
2247 min_load = avg_load;
2250 } while (group = group->next, group != sd->groups);
2252 if (!idlest || 100*this_load < imbalance*min_load)
2258 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2261 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2264 unsigned long load, min_load = ULONG_MAX;
2268 /* Traverse only the allowed CPUs */
2269 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2271 for_each_cpu_mask(i, *tmp) {
2272 load = weighted_cpuload(i);
2274 if (load < min_load || (load == min_load && i == this_cpu)) {
2284 * sched_balance_self: balance the current task (running on cpu) in domains
2285 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2288 * Balance, ie. select the least loaded group.
2290 * Returns the target CPU number, or the same CPU if no balancing is needed.
2292 * preempt must be disabled.
2294 static int sched_balance_self(int cpu, int flag)
2296 struct task_struct *t = current;
2297 struct sched_domain *tmp, *sd = NULL;
2299 for_each_domain(cpu, tmp) {
2301 * If power savings logic is enabled for a domain, stop there.
2303 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2305 if (tmp->flags & flag)
2310 cpumask_t span, tmpmask;
2311 struct sched_group *group;
2312 int new_cpu, weight;
2314 if (!(sd->flags & flag)) {
2320 group = find_idlest_group(sd, t, cpu);
2326 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2327 if (new_cpu == -1 || new_cpu == cpu) {
2328 /* Now try balancing at a lower domain level of cpu */
2333 /* Now try balancing at a lower domain level of new_cpu */
2336 weight = cpus_weight(span);
2337 for_each_domain(cpu, tmp) {
2338 if (weight <= cpus_weight(tmp->span))
2340 if (tmp->flags & flag)
2343 /* while loop will break here if sd == NULL */
2349 #endif /* CONFIG_SMP */
2352 * try_to_wake_up - wake up a thread
2353 * @p: the to-be-woken-up thread
2354 * @state: the mask of task states that can be woken
2355 * @sync: do a synchronous wakeup?
2357 * Put it on the run-queue if it's not already there. The "current"
2358 * thread is always on the run-queue (except when the actual
2359 * re-schedule is in progress), and as such you're allowed to do
2360 * the simpler "current->state = TASK_RUNNING" to mark yourself
2361 * runnable without the overhead of this.
2363 * returns failure only if the task is already active.
2365 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2367 int cpu, orig_cpu, this_cpu, success = 0;
2368 unsigned long flags;
2372 if (!sched_feat(SYNC_WAKEUPS))
2376 rq = task_rq_lock(p, &flags);
2377 old_state = p->state;
2378 if (!(old_state & state))
2386 this_cpu = smp_processor_id();
2389 if (unlikely(task_running(rq, p)))
2392 cpu = p->sched_class->select_task_rq(p, sync);
2393 if (cpu != orig_cpu) {
2394 set_task_cpu(p, cpu);
2395 task_rq_unlock(rq, &flags);
2396 /* might preempt at this point */
2397 rq = task_rq_lock(p, &flags);
2398 old_state = p->state;
2399 if (!(old_state & state))
2404 this_cpu = smp_processor_id();
2408 #ifdef CONFIG_SCHEDSTATS
2409 schedstat_inc(rq, ttwu_count);
2410 if (cpu == this_cpu)
2411 schedstat_inc(rq, ttwu_local);
2413 struct sched_domain *sd;
2414 for_each_domain(this_cpu, sd) {
2415 if (cpu_isset(cpu, sd->span)) {
2416 schedstat_inc(sd, ttwu_wake_remote);
2421 #endif /* CONFIG_SCHEDSTATS */
2424 #endif /* CONFIG_SMP */
2425 schedstat_inc(p, se.nr_wakeups);
2427 schedstat_inc(p, se.nr_wakeups_sync);
2428 if (orig_cpu != cpu)
2429 schedstat_inc(p, se.nr_wakeups_migrate);
2430 if (cpu == this_cpu)
2431 schedstat_inc(p, se.nr_wakeups_local);
2433 schedstat_inc(p, se.nr_wakeups_remote);
2434 update_rq_clock(rq);
2435 activate_task(rq, p, 1);
2439 check_preempt_curr(rq, p);
2441 p->state = TASK_RUNNING;
2443 if (p->sched_class->task_wake_up)
2444 p->sched_class->task_wake_up(rq, p);
2447 task_rq_unlock(rq, &flags);
2452 int wake_up_process(struct task_struct *p)
2454 return try_to_wake_up(p, TASK_ALL, 0);
2456 EXPORT_SYMBOL(wake_up_process);
2458 int wake_up_state(struct task_struct *p, unsigned int state)
2460 return try_to_wake_up(p, state, 0);
2464 * Perform scheduler related setup for a newly forked process p.
2465 * p is forked by current.
2467 * __sched_fork() is basic setup used by init_idle() too:
2469 static void __sched_fork(struct task_struct *p)
2471 p->se.exec_start = 0;
2472 p->se.sum_exec_runtime = 0;
2473 p->se.prev_sum_exec_runtime = 0;
2474 p->se.last_wakeup = 0;
2475 p->se.avg_overlap = 0;
2477 #ifdef CONFIG_SCHEDSTATS
2478 p->se.wait_start = 0;
2479 p->se.sum_sleep_runtime = 0;
2480 p->se.sleep_start = 0;
2481 p->se.block_start = 0;
2482 p->se.sleep_max = 0;
2483 p->se.block_max = 0;
2485 p->se.slice_max = 0;
2489 INIT_LIST_HEAD(&p->rt.run_list);
2491 INIT_LIST_HEAD(&p->se.group_node);
2493 #ifdef CONFIG_PREEMPT_NOTIFIERS
2494 INIT_HLIST_HEAD(&p->preempt_notifiers);
2498 * We mark the process as running here, but have not actually
2499 * inserted it onto the runqueue yet. This guarantees that
2500 * nobody will actually run it, and a signal or other external
2501 * event cannot wake it up and insert it on the runqueue either.
2503 p->state = TASK_RUNNING;
2507 * fork()/clone()-time setup:
2509 void sched_fork(struct task_struct *p, int clone_flags)
2511 int cpu = get_cpu();
2516 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2518 set_task_cpu(p, cpu);
2521 * Make sure we do not leak PI boosting priority to the child:
2523 p->prio = current->normal_prio;
2524 if (!rt_prio(p->prio))
2525 p->sched_class = &fair_sched_class;
2527 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2528 if (likely(sched_info_on()))
2529 memset(&p->sched_info, 0, sizeof(p->sched_info));
2531 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2534 #ifdef CONFIG_PREEMPT
2535 /* Want to start with kernel preemption disabled. */
2536 task_thread_info(p)->preempt_count = 1;
2542 * wake_up_new_task - wake up a newly created task for the first time.
2544 * This function will do some initial scheduler statistics housekeeping
2545 * that must be done for every newly created context, then puts the task
2546 * on the runqueue and wakes it.
2548 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2550 unsigned long flags;
2553 rq = task_rq_lock(p, &flags);
2554 BUG_ON(p->state != TASK_RUNNING);
2555 update_rq_clock(rq);
2557 p->prio = effective_prio(p);
2559 if (!p->sched_class->task_new || !current->se.on_rq) {
2560 activate_task(rq, p, 0);
2563 * Let the scheduling class do new task startup
2564 * management (if any):
2566 p->sched_class->task_new(rq, p);
2569 check_preempt_curr(rq, p);
2571 if (p->sched_class->task_wake_up)
2572 p->sched_class->task_wake_up(rq, p);
2574 task_rq_unlock(rq, &flags);
2577 #ifdef CONFIG_PREEMPT_NOTIFIERS
2580 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2581 * @notifier: notifier struct to register
2583 void preempt_notifier_register(struct preempt_notifier *notifier)
2585 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2587 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2590 * preempt_notifier_unregister - no longer interested in preemption notifications
2591 * @notifier: notifier struct to unregister
2593 * This is safe to call from within a preemption notifier.
2595 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2597 hlist_del(¬ifier->link);
2599 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2601 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2603 struct preempt_notifier *notifier;
2604 struct hlist_node *node;
2606 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2607 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2611 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2612 struct task_struct *next)
2614 struct preempt_notifier *notifier;
2615 struct hlist_node *node;
2617 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2618 notifier->ops->sched_out(notifier, next);
2621 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2623 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2628 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2629 struct task_struct *next)
2633 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2636 * prepare_task_switch - prepare to switch tasks
2637 * @rq: the runqueue preparing to switch
2638 * @prev: the current task that is being switched out
2639 * @next: the task we are going to switch to.
2641 * This is called with the rq lock held and interrupts off. It must
2642 * be paired with a subsequent finish_task_switch after the context
2645 * prepare_task_switch sets up locking and calls architecture specific
2649 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2650 struct task_struct *next)
2652 fire_sched_out_preempt_notifiers(prev, next);
2653 prepare_lock_switch(rq, next);
2654 prepare_arch_switch(next);
2658 * finish_task_switch - clean up after a task-switch
2659 * @rq: runqueue associated with task-switch
2660 * @prev: the thread we just switched away from.
2662 * finish_task_switch must be called after the context switch, paired
2663 * with a prepare_task_switch call before the context switch.
2664 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2665 * and do any other architecture-specific cleanup actions.
2667 * Note that we may have delayed dropping an mm in context_switch(). If
2668 * so, we finish that here outside of the runqueue lock. (Doing it
2669 * with the lock held can cause deadlocks; see schedule() for
2672 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2673 __releases(rq->lock)
2675 struct mm_struct *mm = rq->prev_mm;
2681 * A task struct has one reference for the use as "current".
2682 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2683 * schedule one last time. The schedule call will never return, and
2684 * the scheduled task must drop that reference.
2685 * The test for TASK_DEAD must occur while the runqueue locks are
2686 * still held, otherwise prev could be scheduled on another cpu, die
2687 * there before we look at prev->state, and then the reference would
2689 * Manfred Spraul <manfred@colorfullife.com>
2691 prev_state = prev->state;
2692 finish_arch_switch(prev);
2693 finish_lock_switch(rq, prev);
2695 if (current->sched_class->post_schedule)
2696 current->sched_class->post_schedule(rq);
2699 fire_sched_in_preempt_notifiers(current);
2702 if (unlikely(prev_state == TASK_DEAD)) {
2704 * Remove function-return probe instances associated with this
2705 * task and put them back on the free list.
2707 kprobe_flush_task(prev);
2708 put_task_struct(prev);
2713 * schedule_tail - first thing a freshly forked thread must call.
2714 * @prev: the thread we just switched away from.
2716 asmlinkage void schedule_tail(struct task_struct *prev)
2717 __releases(rq->lock)
2719 struct rq *rq = this_rq();
2721 finish_task_switch(rq, prev);
2722 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2723 /* In this case, finish_task_switch does not reenable preemption */
2726 if (current->set_child_tid)
2727 put_user(task_pid_vnr(current), current->set_child_tid);
2731 * context_switch - switch to the new MM and the new
2732 * thread's register state.
2735 context_switch(struct rq *rq, struct task_struct *prev,
2736 struct task_struct *next)
2738 struct mm_struct *mm, *oldmm;
2740 prepare_task_switch(rq, prev, next);
2742 oldmm = prev->active_mm;
2744 * For paravirt, this is coupled with an exit in switch_to to
2745 * combine the page table reload and the switch backend into
2748 arch_enter_lazy_cpu_mode();
2750 if (unlikely(!mm)) {
2751 next->active_mm = oldmm;
2752 atomic_inc(&oldmm->mm_count);
2753 enter_lazy_tlb(oldmm, next);
2755 switch_mm(oldmm, mm, next);
2757 if (unlikely(!prev->mm)) {
2758 prev->active_mm = NULL;
2759 rq->prev_mm = oldmm;
2762 * Since the runqueue lock will be released by the next
2763 * task (which is an invalid locking op but in the case
2764 * of the scheduler it's an obvious special-case), so we
2765 * do an early lockdep release here:
2767 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2768 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2771 /* Here we just switch the register state and the stack. */
2772 switch_to(prev, next, prev);
2776 * this_rq must be evaluated again because prev may have moved
2777 * CPUs since it called schedule(), thus the 'rq' on its stack
2778 * frame will be invalid.
2780 finish_task_switch(this_rq(), prev);
2784 * nr_running, nr_uninterruptible and nr_context_switches:
2786 * externally visible scheduler statistics: current number of runnable
2787 * threads, current number of uninterruptible-sleeping threads, total
2788 * number of context switches performed since bootup.
2790 unsigned long nr_running(void)
2792 unsigned long i, sum = 0;
2794 for_each_online_cpu(i)
2795 sum += cpu_rq(i)->nr_running;
2800 unsigned long nr_uninterruptible(void)
2802 unsigned long i, sum = 0;
2804 for_each_possible_cpu(i)
2805 sum += cpu_rq(i)->nr_uninterruptible;
2808 * Since we read the counters lockless, it might be slightly
2809 * inaccurate. Do not allow it to go below zero though:
2811 if (unlikely((long)sum < 0))
2817 unsigned long long nr_context_switches(void)
2820 unsigned long long sum = 0;
2822 for_each_possible_cpu(i)
2823 sum += cpu_rq(i)->nr_switches;
2828 unsigned long nr_iowait(void)
2830 unsigned long i, sum = 0;
2832 for_each_possible_cpu(i)
2833 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2838 unsigned long nr_active(void)
2840 unsigned long i, running = 0, uninterruptible = 0;
2842 for_each_online_cpu(i) {
2843 running += cpu_rq(i)->nr_running;
2844 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2847 if (unlikely((long)uninterruptible < 0))
2848 uninterruptible = 0;
2850 return running + uninterruptible;
2854 * Update rq->cpu_load[] statistics. This function is usually called every
2855 * scheduler tick (TICK_NSEC).
2857 static void update_cpu_load(struct rq *this_rq)
2859 unsigned long this_load = this_rq->load.weight;
2862 this_rq->nr_load_updates++;
2864 /* Update our load: */
2865 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2866 unsigned long old_load, new_load;
2868 /* scale is effectively 1 << i now, and >> i divides by scale */
2870 old_load = this_rq->cpu_load[i];
2871 new_load = this_load;
2873 * Round up the averaging division if load is increasing. This
2874 * prevents us from getting stuck on 9 if the load is 10, for
2877 if (new_load > old_load)
2878 new_load += scale-1;
2879 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2886 * double_rq_lock - safely lock two runqueues
2888 * Note this does not disable interrupts like task_rq_lock,
2889 * you need to do so manually before calling.
2891 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2892 __acquires(rq1->lock)
2893 __acquires(rq2->lock)
2895 BUG_ON(!irqs_disabled());
2897 spin_lock(&rq1->lock);
2898 __acquire(rq2->lock); /* Fake it out ;) */
2901 spin_lock(&rq1->lock);
2902 spin_lock(&rq2->lock);
2904 spin_lock(&rq2->lock);
2905 spin_lock(&rq1->lock);
2908 update_rq_clock(rq1);
2909 update_rq_clock(rq2);
2913 * double_rq_unlock - safely unlock two runqueues
2915 * Note this does not restore interrupts like task_rq_unlock,
2916 * you need to do so manually after calling.
2918 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2919 __releases(rq1->lock)
2920 __releases(rq2->lock)
2922 spin_unlock(&rq1->lock);
2924 spin_unlock(&rq2->lock);
2926 __release(rq2->lock);
2930 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2932 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2933 __releases(this_rq->lock)
2934 __acquires(busiest->lock)
2935 __acquires(this_rq->lock)
2939 if (unlikely(!irqs_disabled())) {
2940 /* printk() doesn't work good under rq->lock */
2941 spin_unlock(&this_rq->lock);
2944 if (unlikely(!spin_trylock(&busiest->lock))) {
2945 if (busiest < this_rq) {
2946 spin_unlock(&this_rq->lock);
2947 spin_lock(&busiest->lock);
2948 spin_lock(&this_rq->lock);
2951 spin_lock(&busiest->lock);
2957 * If dest_cpu is allowed for this process, migrate the task to it.
2958 * This is accomplished by forcing the cpu_allowed mask to only
2959 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2960 * the cpu_allowed mask is restored.
2962 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2964 struct migration_req req;
2965 unsigned long flags;
2968 rq = task_rq_lock(p, &flags);
2969 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2970 || unlikely(cpu_is_offline(dest_cpu)))
2973 /* force the process onto the specified CPU */
2974 if (migrate_task(p, dest_cpu, &req)) {
2975 /* Need to wait for migration thread (might exit: take ref). */
2976 struct task_struct *mt = rq->migration_thread;
2978 get_task_struct(mt);
2979 task_rq_unlock(rq, &flags);
2980 wake_up_process(mt);
2981 put_task_struct(mt);
2982 wait_for_completion(&req.done);
2987 task_rq_unlock(rq, &flags);
2991 * sched_exec - execve() is a valuable balancing opportunity, because at
2992 * this point the task has the smallest effective memory and cache footprint.
2994 void sched_exec(void)
2996 int new_cpu, this_cpu = get_cpu();
2997 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2999 if (new_cpu != this_cpu)
3000 sched_migrate_task(current, new_cpu);
3004 * pull_task - move a task from a remote runqueue to the local runqueue.
3005 * Both runqueues must be locked.
3007 static void pull_task(struct rq *src_rq, struct task_struct *p,
3008 struct rq *this_rq, int this_cpu)
3010 deactivate_task(src_rq, p, 0);
3011 set_task_cpu(p, this_cpu);
3012 activate_task(this_rq, p, 0);
3014 * Note that idle threads have a prio of MAX_PRIO, for this test
3015 * to be always true for them.
3017 check_preempt_curr(this_rq, p);
3021 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3024 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3025 struct sched_domain *sd, enum cpu_idle_type idle,
3029 * We do not migrate tasks that are:
3030 * 1) running (obviously), or
3031 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3032 * 3) are cache-hot on their current CPU.
3034 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
3035 schedstat_inc(p, se.nr_failed_migrations_affine);
3040 if (task_running(rq, p)) {
3041 schedstat_inc(p, se.nr_failed_migrations_running);
3046 * Aggressive migration if:
3047 * 1) task is cache cold, or
3048 * 2) too many balance attempts have failed.
3051 if (!task_hot(p, rq->clock, sd) ||
3052 sd->nr_balance_failed > sd->cache_nice_tries) {
3053 #ifdef CONFIG_SCHEDSTATS
3054 if (task_hot(p, rq->clock, sd)) {
3055 schedstat_inc(sd, lb_hot_gained[idle]);
3056 schedstat_inc(p, se.nr_forced_migrations);
3062 if (task_hot(p, rq->clock, sd)) {
3063 schedstat_inc(p, se.nr_failed_migrations_hot);
3069 static unsigned long
3070 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3071 unsigned long max_load_move, struct sched_domain *sd,
3072 enum cpu_idle_type idle, int *all_pinned,
3073 int *this_best_prio, struct rq_iterator *iterator)
3075 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
3076 struct task_struct *p;
3077 long rem_load_move = max_load_move;
3079 if (max_load_move == 0)
3085 * Start the load-balancing iterator:
3087 p = iterator->start(iterator->arg);
3089 if (!p || loops++ > sysctl_sched_nr_migrate)
3092 * To help distribute high priority tasks across CPUs we don't
3093 * skip a task if it will be the highest priority task (i.e. smallest
3094 * prio value) on its new queue regardless of its load weight
3096 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
3097 SCHED_LOAD_SCALE_FUZZ;
3098 if ((skip_for_load && p->prio >= *this_best_prio) ||
3099 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3100 p = iterator->next(iterator->arg);
3104 pull_task(busiest, p, this_rq, this_cpu);
3106 rem_load_move -= p->se.load.weight;
3109 * We only want to steal up to the prescribed amount of weighted load.
3111 if (rem_load_move > 0) {
3112 if (p->prio < *this_best_prio)
3113 *this_best_prio = p->prio;
3114 p = iterator->next(iterator->arg);
3119 * Right now, this is one of only two places pull_task() is called,
3120 * so we can safely collect pull_task() stats here rather than
3121 * inside pull_task().
3123 schedstat_add(sd, lb_gained[idle], pulled);
3126 *all_pinned = pinned;
3128 return max_load_move - rem_load_move;
3132 * move_tasks tries to move up to max_load_move weighted load from busiest to
3133 * this_rq, as part of a balancing operation within domain "sd".
3134 * Returns 1 if successful and 0 otherwise.
3136 * Called with both runqueues locked.
3138 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3139 unsigned long max_load_move,
3140 struct sched_domain *sd, enum cpu_idle_type idle,
3143 const struct sched_class *class = sched_class_highest;
3144 unsigned long total_load_moved = 0;
3145 int this_best_prio = this_rq->curr->prio;
3149 class->load_balance(this_rq, this_cpu, busiest,
3150 max_load_move - total_load_moved,
3151 sd, idle, all_pinned, &this_best_prio);
3152 class = class->next;
3153 } while (class && max_load_move > total_load_moved);
3155 return total_load_moved > 0;
3159 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3160 struct sched_domain *sd, enum cpu_idle_type idle,
3161 struct rq_iterator *iterator)
3163 struct task_struct *p = iterator->start(iterator->arg);
3167 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3168 pull_task(busiest, p, this_rq, this_cpu);
3170 * Right now, this is only the second place pull_task()
3171 * is called, so we can safely collect pull_task()
3172 * stats here rather than inside pull_task().
3174 schedstat_inc(sd, lb_gained[idle]);
3178 p = iterator->next(iterator->arg);
3185 * move_one_task tries to move exactly one task from busiest to this_rq, as
3186 * part of active balancing operations within "domain".
3187 * Returns 1 if successful and 0 otherwise.
3189 * Called with both runqueues locked.
3191 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3192 struct sched_domain *sd, enum cpu_idle_type idle)
3194 const struct sched_class *class;
3196 for (class = sched_class_highest; class; class = class->next)
3197 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3204 * find_busiest_group finds and returns the busiest CPU group within the
3205 * domain. It calculates and returns the amount of weighted load which
3206 * should be moved to restore balance via the imbalance parameter.
3208 static struct sched_group *
3209 find_busiest_group(struct sched_domain *sd, int this_cpu,
3210 unsigned long *imbalance, enum cpu_idle_type idle,
3211 int *sd_idle, const cpumask_t *cpus, int *balance)
3213 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3214 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3215 unsigned long max_pull;
3216 unsigned long busiest_load_per_task, busiest_nr_running;
3217 unsigned long this_load_per_task, this_nr_running;
3218 int load_idx, group_imb = 0;
3219 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3220 int power_savings_balance = 1;
3221 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3222 unsigned long min_nr_running = ULONG_MAX;
3223 struct sched_group *group_min = NULL, *group_leader = NULL;
3226 max_load = this_load = total_load = total_pwr = 0;
3227 busiest_load_per_task = busiest_nr_running = 0;
3228 this_load_per_task = this_nr_running = 0;
3229 if (idle == CPU_NOT_IDLE)
3230 load_idx = sd->busy_idx;
3231 else if (idle == CPU_NEWLY_IDLE)
3232 load_idx = sd->newidle_idx;
3234 load_idx = sd->idle_idx;
3237 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3240 int __group_imb = 0;
3241 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3242 unsigned long sum_nr_running, sum_weighted_load;
3244 local_group = cpu_isset(this_cpu, group->cpumask);
3247 balance_cpu = first_cpu(group->cpumask);
3249 /* Tally up the load of all CPUs in the group */
3250 sum_weighted_load = sum_nr_running = avg_load = 0;
3252 min_cpu_load = ~0UL;
3254 for_each_cpu_mask(i, group->cpumask) {
3257 if (!cpu_isset(i, *cpus))
3262 if (*sd_idle && rq->nr_running)
3265 /* Bias balancing toward cpus of our domain */
3267 if (idle_cpu(i) && !first_idle_cpu) {
3272 load = target_load(i, load_idx);
3274 load = source_load(i, load_idx);
3275 if (load > max_cpu_load)
3276 max_cpu_load = load;
3277 if (min_cpu_load > load)
3278 min_cpu_load = load;
3282 sum_nr_running += rq->nr_running;
3283 sum_weighted_load += weighted_cpuload(i);
3287 * First idle cpu or the first cpu(busiest) in this sched group
3288 * is eligible for doing load balancing at this and above
3289 * domains. In the newly idle case, we will allow all the cpu's
3290 * to do the newly idle load balance.
3292 if (idle != CPU_NEWLY_IDLE && local_group &&
3293 balance_cpu != this_cpu && balance) {
3298 total_load += avg_load;
3299 total_pwr += group->__cpu_power;
3301 /* Adjust by relative CPU power of the group */
3302 avg_load = sg_div_cpu_power(group,
3303 avg_load * SCHED_LOAD_SCALE);
3305 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3308 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3311 this_load = avg_load;
3313 this_nr_running = sum_nr_running;
3314 this_load_per_task = sum_weighted_load;
3315 } else if (avg_load > max_load &&
3316 (sum_nr_running > group_capacity || __group_imb)) {
3317 max_load = avg_load;
3319 busiest_nr_running = sum_nr_running;
3320 busiest_load_per_task = sum_weighted_load;
3321 group_imb = __group_imb;
3324 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3326 * Busy processors will not participate in power savings
3329 if (idle == CPU_NOT_IDLE ||
3330 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3334 * If the local group is idle or completely loaded
3335 * no need to do power savings balance at this domain
3337 if (local_group && (this_nr_running >= group_capacity ||
3339 power_savings_balance = 0;
3342 * If a group is already running at full capacity or idle,
3343 * don't include that group in power savings calculations
3345 if (!power_savings_balance || sum_nr_running >= group_capacity
3350 * Calculate the group which has the least non-idle load.
3351 * This is the group from where we need to pick up the load
3354 if ((sum_nr_running < min_nr_running) ||
3355 (sum_nr_running == min_nr_running &&
3356 first_cpu(group->cpumask) <
3357 first_cpu(group_min->cpumask))) {
3359 min_nr_running = sum_nr_running;
3360 min_load_per_task = sum_weighted_load /
3365 * Calculate the group which is almost near its
3366 * capacity but still has some space to pick up some load
3367 * from other group and save more power
3369 if (sum_nr_running <= group_capacity - 1) {
3370 if (sum_nr_running > leader_nr_running ||
3371 (sum_nr_running == leader_nr_running &&
3372 first_cpu(group->cpumask) >
3373 first_cpu(group_leader->cpumask))) {
3374 group_leader = group;
3375 leader_nr_running = sum_nr_running;
3380 group = group->next;
3381 } while (group != sd->groups);
3383 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3386 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3388 if (this_load >= avg_load ||
3389 100*max_load <= sd->imbalance_pct*this_load)
3392 busiest_load_per_task /= busiest_nr_running;
3394 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3397 * We're trying to get all the cpus to the average_load, so we don't
3398 * want to push ourselves above the average load, nor do we wish to
3399 * reduce the max loaded cpu below the average load, as either of these
3400 * actions would just result in more rebalancing later, and ping-pong
3401 * tasks around. Thus we look for the minimum possible imbalance.
3402 * Negative imbalances (*we* are more loaded than anyone else) will
3403 * be counted as no imbalance for these purposes -- we can't fix that
3404 * by pulling tasks to us. Be careful of negative numbers as they'll
3405 * appear as very large values with unsigned longs.
3407 if (max_load <= busiest_load_per_task)
3411 * In the presence of smp nice balancing, certain scenarios can have
3412 * max load less than avg load(as we skip the groups at or below
3413 * its cpu_power, while calculating max_load..)
3415 if (max_load < avg_load) {
3417 goto small_imbalance;
3420 /* Don't want to pull so many tasks that a group would go idle */
3421 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3423 /* How much load to actually move to equalise the imbalance */
3424 *imbalance = min(max_pull * busiest->__cpu_power,
3425 (avg_load - this_load) * this->__cpu_power)
3429 * if *imbalance is less than the average load per runnable task
3430 * there is no gaurantee that any tasks will be moved so we'll have
3431 * a think about bumping its value to force at least one task to be
3434 if (*imbalance < busiest_load_per_task) {
3435 unsigned long tmp, pwr_now, pwr_move;
3439 pwr_move = pwr_now = 0;
3441 if (this_nr_running) {
3442 this_load_per_task /= this_nr_running;
3443 if (busiest_load_per_task > this_load_per_task)
3446 this_load_per_task = SCHED_LOAD_SCALE;
3448 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3449 busiest_load_per_task * imbn) {
3450 *imbalance = busiest_load_per_task;
3455 * OK, we don't have enough imbalance to justify moving tasks,
3456 * however we may be able to increase total CPU power used by
3460 pwr_now += busiest->__cpu_power *
3461 min(busiest_load_per_task, max_load);
3462 pwr_now += this->__cpu_power *
3463 min(this_load_per_task, this_load);
3464 pwr_now /= SCHED_LOAD_SCALE;
3466 /* Amount of load we'd subtract */
3467 tmp = sg_div_cpu_power(busiest,
3468 busiest_load_per_task * SCHED_LOAD_SCALE);
3470 pwr_move += busiest->__cpu_power *
3471 min(busiest_load_per_task, max_load - tmp);
3473 /* Amount of load we'd add */
3474 if (max_load * busiest->__cpu_power <
3475 busiest_load_per_task * SCHED_LOAD_SCALE)
3476 tmp = sg_div_cpu_power(this,
3477 max_load * busiest->__cpu_power);
3479 tmp = sg_div_cpu_power(this,
3480 busiest_load_per_task * SCHED_LOAD_SCALE);
3481 pwr_move += this->__cpu_power *
3482 min(this_load_per_task, this_load + tmp);
3483 pwr_move /= SCHED_LOAD_SCALE;
3485 /* Move if we gain throughput */
3486 if (pwr_move > pwr_now)
3487 *imbalance = busiest_load_per_task;
3493 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3494 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3497 if (this == group_leader && group_leader != group_min) {
3498 *imbalance = min_load_per_task;
3508 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3511 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3512 unsigned long imbalance, const cpumask_t *cpus)
3514 struct rq *busiest = NULL, *rq;
3515 unsigned long max_load = 0;
3518 for_each_cpu_mask(i, group->cpumask) {
3521 if (!cpu_isset(i, *cpus))
3525 wl = weighted_cpuload(i);
3527 if (rq->nr_running == 1 && wl > imbalance)
3530 if (wl > max_load) {
3540 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3541 * so long as it is large enough.
3543 #define MAX_PINNED_INTERVAL 512
3546 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3547 * tasks if there is an imbalance.
3549 static int load_balance(int this_cpu, struct rq *this_rq,
3550 struct sched_domain *sd, enum cpu_idle_type idle,
3551 int *balance, cpumask_t *cpus)
3553 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3554 struct sched_group *group;
3555 unsigned long imbalance;
3557 unsigned long flags;
3558 int unlock_aggregate;
3562 unlock_aggregate = get_aggregate(this_cpu, sd);
3565 * When power savings policy is enabled for the parent domain, idle
3566 * sibling can pick up load irrespective of busy siblings. In this case,
3567 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3568 * portraying it as CPU_NOT_IDLE.
3570 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3571 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3574 schedstat_inc(sd, lb_count[idle]);
3577 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3584 schedstat_inc(sd, lb_nobusyg[idle]);
3588 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3590 schedstat_inc(sd, lb_nobusyq[idle]);
3594 BUG_ON(busiest == this_rq);
3596 schedstat_add(sd, lb_imbalance[idle], imbalance);
3599 if (busiest->nr_running > 1) {
3601 * Attempt to move tasks. If find_busiest_group has found
3602 * an imbalance but busiest->nr_running <= 1, the group is
3603 * still unbalanced. ld_moved simply stays zero, so it is
3604 * correctly treated as an imbalance.
3606 local_irq_save(flags);
3607 double_rq_lock(this_rq, busiest);
3608 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3609 imbalance, sd, idle, &all_pinned);
3610 double_rq_unlock(this_rq, busiest);
3611 local_irq_restore(flags);
3614 * some other cpu did the load balance for us.
3616 if (ld_moved && this_cpu != smp_processor_id())
3617 resched_cpu(this_cpu);
3619 /* All tasks on this runqueue were pinned by CPU affinity */
3620 if (unlikely(all_pinned)) {
3621 cpu_clear(cpu_of(busiest), *cpus);
3622 if (!cpus_empty(*cpus))
3629 schedstat_inc(sd, lb_failed[idle]);
3630 sd->nr_balance_failed++;
3632 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3634 spin_lock_irqsave(&busiest->lock, flags);
3636 /* don't kick the migration_thread, if the curr
3637 * task on busiest cpu can't be moved to this_cpu
3639 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3640 spin_unlock_irqrestore(&busiest->lock, flags);
3642 goto out_one_pinned;
3645 if (!busiest->active_balance) {
3646 busiest->active_balance = 1;
3647 busiest->push_cpu = this_cpu;
3650 spin_unlock_irqrestore(&busiest->lock, flags);
3652 wake_up_process(busiest->migration_thread);
3655 * We've kicked active balancing, reset the failure
3658 sd->nr_balance_failed = sd->cache_nice_tries+1;
3661 sd->nr_balance_failed = 0;
3663 if (likely(!active_balance)) {
3664 /* We were unbalanced, so reset the balancing interval */
3665 sd->balance_interval = sd->min_interval;
3668 * If we've begun active balancing, start to back off. This
3669 * case may not be covered by the all_pinned logic if there
3670 * is only 1 task on the busy runqueue (because we don't call
3673 if (sd->balance_interval < sd->max_interval)
3674 sd->balance_interval *= 2;
3677 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3678 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3684 schedstat_inc(sd, lb_balanced[idle]);
3686 sd->nr_balance_failed = 0;
3689 /* tune up the balancing interval */
3690 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3691 (sd->balance_interval < sd->max_interval))
3692 sd->balance_interval *= 2;
3694 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3695 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3700 if (unlock_aggregate)
3701 put_aggregate(this_cpu, sd);
3706 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3707 * tasks if there is an imbalance.
3709 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3710 * this_rq is locked.
3713 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3716 struct sched_group *group;
3717 struct rq *busiest = NULL;
3718 unsigned long imbalance;
3726 * When power savings policy is enabled for the parent domain, idle
3727 * sibling can pick up load irrespective of busy siblings. In this case,
3728 * let the state of idle sibling percolate up as IDLE, instead of
3729 * portraying it as CPU_NOT_IDLE.
3731 if (sd->flags & SD_SHARE_CPUPOWER &&
3732 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3735 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3737 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3738 &sd_idle, cpus, NULL);
3740 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3744 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3746 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3750 BUG_ON(busiest == this_rq);
3752 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3755 if (busiest->nr_running > 1) {
3756 /* Attempt to move tasks */
3757 double_lock_balance(this_rq, busiest);
3758 /* this_rq->clock is already updated */
3759 update_rq_clock(busiest);
3760 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3761 imbalance, sd, CPU_NEWLY_IDLE,
3763 spin_unlock(&busiest->lock);
3765 if (unlikely(all_pinned)) {
3766 cpu_clear(cpu_of(busiest), *cpus);
3767 if (!cpus_empty(*cpus))
3773 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3774 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3775 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3778 sd->nr_balance_failed = 0;
3783 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3784 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3785 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3787 sd->nr_balance_failed = 0;
3793 * idle_balance is called by schedule() if this_cpu is about to become
3794 * idle. Attempts to pull tasks from other CPUs.
3796 static void idle_balance(int this_cpu, struct rq *this_rq)
3798 struct sched_domain *sd;
3799 int pulled_task = -1;
3800 unsigned long next_balance = jiffies + HZ;
3803 for_each_domain(this_cpu, sd) {
3804 unsigned long interval;
3806 if (!(sd->flags & SD_LOAD_BALANCE))
3809 if (sd->flags & SD_BALANCE_NEWIDLE)
3810 /* If we've pulled tasks over stop searching: */
3811 pulled_task = load_balance_newidle(this_cpu, this_rq,
3814 interval = msecs_to_jiffies(sd->balance_interval);
3815 if (time_after(next_balance, sd->last_balance + interval))
3816 next_balance = sd->last_balance + interval;
3820 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3822 * We are going idle. next_balance may be set based on
3823 * a busy processor. So reset next_balance.
3825 this_rq->next_balance = next_balance;
3830 * active_load_balance is run by migration threads. It pushes running tasks
3831 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3832 * running on each physical CPU where possible, and avoids physical /
3833 * logical imbalances.
3835 * Called with busiest_rq locked.
3837 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3839 int target_cpu = busiest_rq->push_cpu;
3840 struct sched_domain *sd;
3841 struct rq *target_rq;
3843 /* Is there any task to move? */
3844 if (busiest_rq->nr_running <= 1)
3847 target_rq = cpu_rq(target_cpu);
3850 * This condition is "impossible", if it occurs
3851 * we need to fix it. Originally reported by
3852 * Bjorn Helgaas on a 128-cpu setup.
3854 BUG_ON(busiest_rq == target_rq);
3856 /* move a task from busiest_rq to target_rq */
3857 double_lock_balance(busiest_rq, target_rq);
3858 update_rq_clock(busiest_rq);
3859 update_rq_clock(target_rq);
3861 /* Search for an sd spanning us and the target CPU. */
3862 for_each_domain(target_cpu, sd) {
3863 if ((sd->flags & SD_LOAD_BALANCE) &&
3864 cpu_isset(busiest_cpu, sd->span))
3869 schedstat_inc(sd, alb_count);
3871 if (move_one_task(target_rq, target_cpu, busiest_rq,
3873 schedstat_inc(sd, alb_pushed);
3875 schedstat_inc(sd, alb_failed);
3877 spin_unlock(&target_rq->lock);
3882 atomic_t load_balancer;
3884 } nohz ____cacheline_aligned = {
3885 .load_balancer = ATOMIC_INIT(-1),
3886 .cpu_mask = CPU_MASK_NONE,
3890 * This routine will try to nominate the ilb (idle load balancing)
3891 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3892 * load balancing on behalf of all those cpus. If all the cpus in the system
3893 * go into this tickless mode, then there will be no ilb owner (as there is
3894 * no need for one) and all the cpus will sleep till the next wakeup event
3897 * For the ilb owner, tick is not stopped. And this tick will be used
3898 * for idle load balancing. ilb owner will still be part of
3901 * While stopping the tick, this cpu will become the ilb owner if there
3902 * is no other owner. And will be the owner till that cpu becomes busy
3903 * or if all cpus in the system stop their ticks at which point
3904 * there is no need for ilb owner.
3906 * When the ilb owner becomes busy, it nominates another owner, during the
3907 * next busy scheduler_tick()
3909 int select_nohz_load_balancer(int stop_tick)
3911 int cpu = smp_processor_id();
3914 cpu_set(cpu, nohz.cpu_mask);
3915 cpu_rq(cpu)->in_nohz_recently = 1;
3918 * If we are going offline and still the leader, give up!
3920 if (cpu_is_offline(cpu) &&
3921 atomic_read(&nohz.load_balancer) == cpu) {
3922 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3927 /* time for ilb owner also to sleep */
3928 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3929 if (atomic_read(&nohz.load_balancer) == cpu)
3930 atomic_set(&nohz.load_balancer, -1);
3934 if (atomic_read(&nohz.load_balancer) == -1) {
3935 /* make me the ilb owner */
3936 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3938 } else if (atomic_read(&nohz.load_balancer) == cpu)
3941 if (!cpu_isset(cpu, nohz.cpu_mask))
3944 cpu_clear(cpu, nohz.cpu_mask);
3946 if (atomic_read(&nohz.load_balancer) == cpu)
3947 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3954 static DEFINE_SPINLOCK(balancing);
3957 * It checks each scheduling domain to see if it is due to be balanced,
3958 * and initiates a balancing operation if so.
3960 * Balancing parameters are set up in arch_init_sched_domains.
3962 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3965 struct rq *rq = cpu_rq(cpu);
3966 unsigned long interval;
3967 struct sched_domain *sd;
3968 /* Earliest time when we have to do rebalance again */
3969 unsigned long next_balance = jiffies + 60*HZ;
3970 int update_next_balance = 0;
3974 for_each_domain(cpu, sd) {
3975 if (!(sd->flags & SD_LOAD_BALANCE))
3978 interval = sd->balance_interval;
3979 if (idle != CPU_IDLE)
3980 interval *= sd->busy_factor;
3982 /* scale ms to jiffies */
3983 interval = msecs_to_jiffies(interval);
3984 if (unlikely(!interval))
3986 if (interval > HZ*NR_CPUS/10)
3987 interval = HZ*NR_CPUS/10;
3989 need_serialize = sd->flags & SD_SERIALIZE;
3991 if (need_serialize) {
3992 if (!spin_trylock(&balancing))
3996 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3997 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3999 * We've pulled tasks over so either we're no
4000 * longer idle, or one of our SMT siblings is
4003 idle = CPU_NOT_IDLE;
4005 sd->last_balance = jiffies;
4008 spin_unlock(&balancing);
4010 if (time_after(next_balance, sd->last_balance + interval)) {
4011 next_balance = sd->last_balance + interval;
4012 update_next_balance = 1;
4016 * Stop the load balance at this level. There is another
4017 * CPU in our sched group which is doing load balancing more
4025 * next_balance will be updated only when there is a need.
4026 * When the cpu is attached to null domain for ex, it will not be
4029 if (likely(update_next_balance))
4030 rq->next_balance = next_balance;
4034 * run_rebalance_domains is triggered when needed from the scheduler tick.
4035 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4036 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4038 static void run_rebalance_domains(struct softirq_action *h)
4040 int this_cpu = smp_processor_id();
4041 struct rq *this_rq = cpu_rq(this_cpu);
4042 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4043 CPU_IDLE : CPU_NOT_IDLE;
4045 rebalance_domains(this_cpu, idle);
4049 * If this cpu is the owner for idle load balancing, then do the
4050 * balancing on behalf of the other idle cpus whose ticks are
4053 if (this_rq->idle_at_tick &&
4054 atomic_read(&nohz.load_balancer) == this_cpu) {
4055 cpumask_t cpus = nohz.cpu_mask;
4059 cpu_clear(this_cpu, cpus);
4060 for_each_cpu_mask(balance_cpu, cpus) {
4062 * If this cpu gets work to do, stop the load balancing
4063 * work being done for other cpus. Next load
4064 * balancing owner will pick it up.
4069 rebalance_domains(balance_cpu, CPU_IDLE);
4071 rq = cpu_rq(balance_cpu);
4072 if (time_after(this_rq->next_balance, rq->next_balance))
4073 this_rq->next_balance = rq->next_balance;
4080 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4082 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4083 * idle load balancing owner or decide to stop the periodic load balancing,
4084 * if the whole system is idle.
4086 static inline void trigger_load_balance(struct rq *rq, int cpu)
4090 * If we were in the nohz mode recently and busy at the current
4091 * scheduler tick, then check if we need to nominate new idle
4094 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4095 rq->in_nohz_recently = 0;
4097 if (atomic_read(&nohz.load_balancer) == cpu) {
4098 cpu_clear(cpu, nohz.cpu_mask);
4099 atomic_set(&nohz.load_balancer, -1);
4102 if (atomic_read(&nohz.load_balancer) == -1) {
4104 * simple selection for now: Nominate the
4105 * first cpu in the nohz list to be the next
4108 * TBD: Traverse the sched domains and nominate
4109 * the nearest cpu in the nohz.cpu_mask.
4111 int ilb = first_cpu(nohz.cpu_mask);
4113 if (ilb < nr_cpu_ids)
4119 * If this cpu is idle and doing idle load balancing for all the
4120 * cpus with ticks stopped, is it time for that to stop?
4122 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4123 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4129 * If this cpu is idle and the idle load balancing is done by
4130 * someone else, then no need raise the SCHED_SOFTIRQ
4132 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4133 cpu_isset(cpu, nohz.cpu_mask))
4136 if (time_after_eq(jiffies, rq->next_balance))
4137 raise_softirq(SCHED_SOFTIRQ);
4140 #else /* CONFIG_SMP */
4143 * on UP we do not need to balance between CPUs:
4145 static inline void idle_balance(int cpu, struct rq *rq)
4151 DEFINE_PER_CPU(struct kernel_stat, kstat);
4153 EXPORT_PER_CPU_SYMBOL(kstat);
4156 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4157 * that have not yet been banked in case the task is currently running.
4159 unsigned long long task_sched_runtime(struct task_struct *p)
4161 unsigned long flags;
4165 rq = task_rq_lock(p, &flags);
4166 ns = p->se.sum_exec_runtime;
4167 if (task_current(rq, p)) {
4168 update_rq_clock(rq);
4169 delta_exec = rq->clock - p->se.exec_start;
4170 if ((s64)delta_exec > 0)
4173 task_rq_unlock(rq, &flags);
4179 * Account user cpu time to a process.
4180 * @p: the process that the cpu time gets accounted to
4181 * @cputime: the cpu time spent in user space since the last update
4183 void account_user_time(struct task_struct *p, cputime_t cputime)
4185 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4188 p->utime = cputime_add(p->utime, cputime);
4190 /* Add user time to cpustat. */
4191 tmp = cputime_to_cputime64(cputime);
4192 if (TASK_NICE(p) > 0)
4193 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4195 cpustat->user = cputime64_add(cpustat->user, tmp);
4199 * Account guest cpu time to a process.
4200 * @p: the process that the cpu time gets accounted to
4201 * @cputime: the cpu time spent in virtual machine since the last update
4203 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4206 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4208 tmp = cputime_to_cputime64(cputime);
4210 p->utime = cputime_add(p->utime, cputime);
4211 p->gtime = cputime_add(p->gtime, cputime);
4213 cpustat->user = cputime64_add(cpustat->user, tmp);
4214 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4218 * Account scaled user cpu time to a process.
4219 * @p: the process that the cpu time gets accounted to
4220 * @cputime: the cpu time spent in user space since the last update
4222 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4224 p->utimescaled = cputime_add(p->utimescaled, cputime);
4228 * Account system cpu time to a process.
4229 * @p: the process that the cpu time gets accounted to
4230 * @hardirq_offset: the offset to subtract from hardirq_count()
4231 * @cputime: the cpu time spent in kernel space since the last update
4233 void account_system_time(struct task_struct *p, int hardirq_offset,
4236 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4237 struct rq *rq = this_rq();
4240 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4241 account_guest_time(p, cputime);
4245 p->stime = cputime_add(p->stime, cputime);
4247 /* Add system time to cpustat. */
4248 tmp = cputime_to_cputime64(cputime);
4249 if (hardirq_count() - hardirq_offset)
4250 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4251 else if (softirq_count())
4252 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4253 else if (p != rq->idle)
4254 cpustat->system = cputime64_add(cpustat->system, tmp);
4255 else if (atomic_read(&rq->nr_iowait) > 0)
4256 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4258 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4259 /* Account for system time used */
4260 acct_update_integrals(p);
4264 * Account scaled system cpu time to a process.
4265 * @p: the process that the cpu time gets accounted to
4266 * @hardirq_offset: the offset to subtract from hardirq_count()
4267 * @cputime: the cpu time spent in kernel space since the last update
4269 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4271 p->stimescaled = cputime_add(p->stimescaled, cputime);
4275 * Account for involuntary wait time.
4276 * @p: the process from which the cpu time has been stolen
4277 * @steal: the cpu time spent in involuntary wait
4279 void account_steal_time(struct task_struct *p, cputime_t steal)
4281 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4282 cputime64_t tmp = cputime_to_cputime64(steal);
4283 struct rq *rq = this_rq();
4285 if (p == rq->idle) {
4286 p->stime = cputime_add(p->stime, steal);
4287 if (atomic_read(&rq->nr_iowait) > 0)
4288 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4290 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4292 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4296 * This function gets called by the timer code, with HZ frequency.
4297 * We call it with interrupts disabled.
4299 * It also gets called by the fork code, when changing the parent's
4302 void scheduler_tick(void)
4304 int cpu = smp_processor_id();
4305 struct rq *rq = cpu_rq(cpu);
4306 struct task_struct *curr = rq->curr;
4310 spin_lock(&rq->lock);
4311 update_rq_clock(rq);
4312 update_cpu_load(rq);
4313 curr->sched_class->task_tick(rq, curr, 0);
4314 spin_unlock(&rq->lock);
4317 rq->idle_at_tick = idle_cpu(cpu);
4318 trigger_load_balance(rq, cpu);
4322 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4324 void __kprobes add_preempt_count(int val)
4329 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4331 preempt_count() += val;
4333 * Spinlock count overflowing soon?
4335 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4338 EXPORT_SYMBOL(add_preempt_count);
4340 void __kprobes sub_preempt_count(int val)
4345 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4348 * Is the spinlock portion underflowing?
4350 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4351 !(preempt_count() & PREEMPT_MASK)))
4354 preempt_count() -= val;
4356 EXPORT_SYMBOL(sub_preempt_count);
4361 * Print scheduling while atomic bug:
4363 static noinline void __schedule_bug(struct task_struct *prev)
4365 struct pt_regs *regs = get_irq_regs();
4367 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4368 prev->comm, prev->pid, preempt_count());
4370 debug_show_held_locks(prev);
4372 if (irqs_disabled())
4373 print_irqtrace_events(prev);
4382 * Various schedule()-time debugging checks and statistics:
4384 static inline void schedule_debug(struct task_struct *prev)
4387 * Test if we are atomic. Since do_exit() needs to call into
4388 * schedule() atomically, we ignore that path for now.
4389 * Otherwise, whine if we are scheduling when we should not be.
4391 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4392 __schedule_bug(prev);
4394 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4396 schedstat_inc(this_rq(), sched_count);
4397 #ifdef CONFIG_SCHEDSTATS
4398 if (unlikely(prev->lock_depth >= 0)) {
4399 schedstat_inc(this_rq(), bkl_count);
4400 schedstat_inc(prev, sched_info.bkl_count);
4406 * Pick up the highest-prio task:
4408 static inline struct task_struct *
4409 pick_next_task(struct rq *rq, struct task_struct *prev)
4411 const struct sched_class *class;
4412 struct task_struct *p;
4415 * Optimization: we know that if all tasks are in
4416 * the fair class we can call that function directly:
4418 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4419 p = fair_sched_class.pick_next_task(rq);
4424 class = sched_class_highest;
4426 p = class->pick_next_task(rq);
4430 * Will never be NULL as the idle class always
4431 * returns a non-NULL p:
4433 class = class->next;
4438 * schedule() is the main scheduler function.
4440 asmlinkage void __sched schedule(void)
4442 struct task_struct *prev, *next;
4443 unsigned long *switch_count;
4445 int cpu, hrtick = sched_feat(HRTICK);
4449 cpu = smp_processor_id();
4453 switch_count = &prev->nivcsw;
4455 release_kernel_lock(prev);
4456 need_resched_nonpreemptible:
4458 schedule_debug(prev);
4464 * Do the rq-clock update outside the rq lock:
4466 local_irq_disable();
4467 update_rq_clock(rq);
4468 spin_lock(&rq->lock);
4469 clear_tsk_need_resched(prev);
4471 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4472 if (unlikely(signal_pending_state(prev->state, prev)))
4473 prev->state = TASK_RUNNING;
4475 deactivate_task(rq, prev, 1);
4476 switch_count = &prev->nvcsw;
4480 if (prev->sched_class->pre_schedule)
4481 prev->sched_class->pre_schedule(rq, prev);
4484 if (unlikely(!rq->nr_running))
4485 idle_balance(cpu, rq);
4487 prev->sched_class->put_prev_task(rq, prev);
4488 next = pick_next_task(rq, prev);
4490 if (likely(prev != next)) {
4491 sched_info_switch(prev, next);
4497 context_switch(rq, prev, next); /* unlocks the rq */
4499 * the context switch might have flipped the stack from under
4500 * us, hence refresh the local variables.
4502 cpu = smp_processor_id();
4505 spin_unlock_irq(&rq->lock);
4510 if (unlikely(reacquire_kernel_lock(current) < 0))
4511 goto need_resched_nonpreemptible;
4513 preempt_enable_no_resched();
4514 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4517 EXPORT_SYMBOL(schedule);
4519 #ifdef CONFIG_PREEMPT
4521 * this is the entry point to schedule() from in-kernel preemption
4522 * off of preempt_enable. Kernel preemptions off return from interrupt
4523 * occur there and call schedule directly.
4525 asmlinkage void __sched preempt_schedule(void)
4527 struct thread_info *ti = current_thread_info();
4530 * If there is a non-zero preempt_count or interrupts are disabled,
4531 * we do not want to preempt the current task. Just return..
4533 if (likely(ti->preempt_count || irqs_disabled()))
4537 add_preempt_count(PREEMPT_ACTIVE);
4539 sub_preempt_count(PREEMPT_ACTIVE);
4542 * Check again in case we missed a preemption opportunity
4543 * between schedule and now.
4546 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4548 EXPORT_SYMBOL(preempt_schedule);
4551 * this is the entry point to schedule() from kernel preemption
4552 * off of irq context.
4553 * Note, that this is called and return with irqs disabled. This will
4554 * protect us against recursive calling from irq.
4556 asmlinkage void __sched preempt_schedule_irq(void)
4558 struct thread_info *ti = current_thread_info();
4560 /* Catch callers which need to be fixed */
4561 BUG_ON(ti->preempt_count || !irqs_disabled());
4564 add_preempt_count(PREEMPT_ACTIVE);
4567 local_irq_disable();
4568 sub_preempt_count(PREEMPT_ACTIVE);
4571 * Check again in case we missed a preemption opportunity
4572 * between schedule and now.
4575 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4578 #endif /* CONFIG_PREEMPT */
4580 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4583 return try_to_wake_up(curr->private, mode, sync);
4585 EXPORT_SYMBOL(default_wake_function);
4588 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4589 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4590 * number) then we wake all the non-exclusive tasks and one exclusive task.
4592 * There are circumstances in which we can try to wake a task which has already
4593 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4594 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4596 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4597 int nr_exclusive, int sync, void *key)
4599 wait_queue_t *curr, *next;
4601 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4602 unsigned flags = curr->flags;
4604 if (curr->func(curr, mode, sync, key) &&
4605 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4611 * __wake_up - wake up threads blocked on a waitqueue.
4613 * @mode: which threads
4614 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4615 * @key: is directly passed to the wakeup function
4617 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4618 int nr_exclusive, void *key)
4620 unsigned long flags;
4622 spin_lock_irqsave(&q->lock, flags);
4623 __wake_up_common(q, mode, nr_exclusive, 0, key);
4624 spin_unlock_irqrestore(&q->lock, flags);
4626 EXPORT_SYMBOL(__wake_up);
4629 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4631 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4633 __wake_up_common(q, mode, 1, 0, NULL);
4637 * __wake_up_sync - wake up threads blocked on a waitqueue.
4639 * @mode: which threads
4640 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4642 * The sync wakeup differs that the waker knows that it will schedule
4643 * away soon, so while the target thread will be woken up, it will not
4644 * be migrated to another CPU - ie. the two threads are 'synchronized'
4645 * with each other. This can prevent needless bouncing between CPUs.
4647 * On UP it can prevent extra preemption.
4650 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4652 unsigned long flags;
4658 if (unlikely(!nr_exclusive))
4661 spin_lock_irqsave(&q->lock, flags);
4662 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4663 spin_unlock_irqrestore(&q->lock, flags);
4665 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4667 void complete(struct completion *x)
4669 unsigned long flags;
4671 spin_lock_irqsave(&x->wait.lock, flags);
4673 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4674 spin_unlock_irqrestore(&x->wait.lock, flags);
4676 EXPORT_SYMBOL(complete);
4678 void complete_all(struct completion *x)
4680 unsigned long flags;
4682 spin_lock_irqsave(&x->wait.lock, flags);
4683 x->done += UINT_MAX/2;
4684 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4685 spin_unlock_irqrestore(&x->wait.lock, flags);
4687 EXPORT_SYMBOL(complete_all);
4689 static inline long __sched
4690 do_wait_for_common(struct completion *x, long timeout, int state)
4693 DECLARE_WAITQUEUE(wait, current);
4695 wait.flags |= WQ_FLAG_EXCLUSIVE;
4696 __add_wait_queue_tail(&x->wait, &wait);
4698 if ((state == TASK_INTERRUPTIBLE &&
4699 signal_pending(current)) ||
4700 (state == TASK_KILLABLE &&
4701 fatal_signal_pending(current))) {
4702 timeout = -ERESTARTSYS;
4705 __set_current_state(state);
4706 spin_unlock_irq(&x->wait.lock);
4707 timeout = schedule_timeout(timeout);
4708 spin_lock_irq(&x->wait.lock);
4709 } while (!x->done && timeout);
4710 __remove_wait_queue(&x->wait, &wait);
4715 return timeout ?: 1;
4719 wait_for_common(struct completion *x, long timeout, int state)
4723 spin_lock_irq(&x->wait.lock);
4724 timeout = do_wait_for_common(x, timeout, state);
4725 spin_unlock_irq(&x->wait.lock);
4729 void __sched wait_for_completion(struct completion *x)
4731 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4733 EXPORT_SYMBOL(wait_for_completion);
4735 unsigned long __sched
4736 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4738 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4740 EXPORT_SYMBOL(wait_for_completion_timeout);
4742 int __sched wait_for_completion_interruptible(struct completion *x)
4744 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4745 if (t == -ERESTARTSYS)
4749 EXPORT_SYMBOL(wait_for_completion_interruptible);
4751 unsigned long __sched
4752 wait_for_completion_interruptible_timeout(struct completion *x,
4753 unsigned long timeout)
4755 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4757 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4759 int __sched wait_for_completion_killable(struct completion *x)
4761 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4762 if (t == -ERESTARTSYS)
4766 EXPORT_SYMBOL(wait_for_completion_killable);
4769 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4771 unsigned long flags;
4774 init_waitqueue_entry(&wait, current);
4776 __set_current_state(state);
4778 spin_lock_irqsave(&q->lock, flags);
4779 __add_wait_queue(q, &wait);
4780 spin_unlock(&q->lock);
4781 timeout = schedule_timeout(timeout);
4782 spin_lock_irq(&q->lock);
4783 __remove_wait_queue(q, &wait);
4784 spin_unlock_irqrestore(&q->lock, flags);
4789 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4791 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4793 EXPORT_SYMBOL(interruptible_sleep_on);
4796 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4798 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4800 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4802 void __sched sleep_on(wait_queue_head_t *q)
4804 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4806 EXPORT_SYMBOL(sleep_on);
4808 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4810 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4812 EXPORT_SYMBOL(sleep_on_timeout);
4814 #ifdef CONFIG_RT_MUTEXES
4817 * rt_mutex_setprio - set the current priority of a task
4819 * @prio: prio value (kernel-internal form)
4821 * This function changes the 'effective' priority of a task. It does
4822 * not touch ->normal_prio like __setscheduler().
4824 * Used by the rt_mutex code to implement priority inheritance logic.
4826 void rt_mutex_setprio(struct task_struct *p, int prio)
4828 unsigned long flags;
4829 int oldprio, on_rq, running;
4831 const struct sched_class *prev_class = p->sched_class;
4833 BUG_ON(prio < 0 || prio > MAX_PRIO);
4835 rq = task_rq_lock(p, &flags);
4836 update_rq_clock(rq);
4839 on_rq = p->se.on_rq;
4840 running = task_current(rq, p);
4842 dequeue_task(rq, p, 0);
4844 p->sched_class->put_prev_task(rq, p);
4847 p->sched_class = &rt_sched_class;
4849 p->sched_class = &fair_sched_class;
4854 p->sched_class->set_curr_task(rq);
4856 enqueue_task(rq, p, 0);
4858 check_class_changed(rq, p, prev_class, oldprio, running);
4860 task_rq_unlock(rq, &flags);
4865 void set_user_nice(struct task_struct *p, long nice)
4867 int old_prio, delta, on_rq;
4868 unsigned long flags;
4871 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4874 * We have to be careful, if called from sys_setpriority(),
4875 * the task might be in the middle of scheduling on another CPU.
4877 rq = task_rq_lock(p, &flags);
4878 update_rq_clock(rq);
4880 * The RT priorities are set via sched_setscheduler(), but we still
4881 * allow the 'normal' nice value to be set - but as expected
4882 * it wont have any effect on scheduling until the task is
4883 * SCHED_FIFO/SCHED_RR:
4885 if (task_has_rt_policy(p)) {
4886 p->static_prio = NICE_TO_PRIO(nice);
4889 on_rq = p->se.on_rq;
4891 dequeue_task(rq, p, 0);
4893 p->static_prio = NICE_TO_PRIO(nice);
4896 p->prio = effective_prio(p);
4897 delta = p->prio - old_prio;
4900 enqueue_task(rq, p, 0);
4902 * If the task increased its priority or is running and
4903 * lowered its priority, then reschedule its CPU:
4905 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4906 resched_task(rq->curr);
4909 task_rq_unlock(rq, &flags);
4911 EXPORT_SYMBOL(set_user_nice);
4914 * can_nice - check if a task can reduce its nice value
4918 int can_nice(const struct task_struct *p, const int nice)
4920 /* convert nice value [19,-20] to rlimit style value [1,40] */
4921 int nice_rlim = 20 - nice;
4923 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4924 capable(CAP_SYS_NICE));
4927 #ifdef __ARCH_WANT_SYS_NICE
4930 * sys_nice - change the priority of the current process.
4931 * @increment: priority increment
4933 * sys_setpriority is a more generic, but much slower function that
4934 * does similar things.
4936 asmlinkage long sys_nice(int increment)
4941 * Setpriority might change our priority at the same moment.
4942 * We don't have to worry. Conceptually one call occurs first
4943 * and we have a single winner.
4945 if (increment < -40)
4950 nice = PRIO_TO_NICE(current->static_prio) + increment;
4956 if (increment < 0 && !can_nice(current, nice))
4959 retval = security_task_setnice(current, nice);
4963 set_user_nice(current, nice);
4970 * task_prio - return the priority value of a given task.
4971 * @p: the task in question.
4973 * This is the priority value as seen by users in /proc.
4974 * RT tasks are offset by -200. Normal tasks are centered
4975 * around 0, value goes from -16 to +15.
4977 int task_prio(const struct task_struct *p)
4979 return p->prio - MAX_RT_PRIO;
4983 * task_nice - return the nice value of a given task.
4984 * @p: the task in question.
4986 int task_nice(const struct task_struct *p)
4988 return TASK_NICE(p);
4990 EXPORT_SYMBOL(task_nice);
4993 * idle_cpu - is a given cpu idle currently?
4994 * @cpu: the processor in question.
4996 int idle_cpu(int cpu)
4998 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5002 * idle_task - return the idle task for a given cpu.
5003 * @cpu: the processor in question.
5005 struct task_struct *idle_task(int cpu)
5007 return cpu_rq(cpu)->idle;
5011 * find_process_by_pid - find a process with a matching PID value.
5012 * @pid: the pid in question.
5014 static struct task_struct *find_process_by_pid(pid_t pid)
5016 return pid ? find_task_by_vpid(pid) : current;
5019 /* Actually do priority change: must hold rq lock. */
5021 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5023 BUG_ON(p->se.on_rq);
5026 switch (p->policy) {
5030 p->sched_class = &fair_sched_class;
5034 p->sched_class = &rt_sched_class;
5038 p->rt_priority = prio;
5039 p->normal_prio = normal_prio(p);
5040 /* we are holding p->pi_lock already */
5041 p->prio = rt_mutex_getprio(p);
5046 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5047 * @p: the task in question.
5048 * @policy: new policy.
5049 * @param: structure containing the new RT priority.
5051 * NOTE that the task may be already dead.
5053 int sched_setscheduler(struct task_struct *p, int policy,
5054 struct sched_param *param)
5056 int retval, oldprio, oldpolicy = -1, on_rq, running;
5057 unsigned long flags;
5058 const struct sched_class *prev_class = p->sched_class;
5061 /* may grab non-irq protected spin_locks */
5062 BUG_ON(in_interrupt());
5064 /* double check policy once rq lock held */
5066 policy = oldpolicy = p->policy;
5067 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5068 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5069 policy != SCHED_IDLE)
5072 * Valid priorities for SCHED_FIFO and SCHED_RR are
5073 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5074 * SCHED_BATCH and SCHED_IDLE is 0.
5076 if (param->sched_priority < 0 ||
5077 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5078 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5080 if (rt_policy(policy) != (param->sched_priority != 0))
5084 * Allow unprivileged RT tasks to decrease priority:
5086 if (!capable(CAP_SYS_NICE)) {
5087 if (rt_policy(policy)) {
5088 unsigned long rlim_rtprio;
5090 if (!lock_task_sighand(p, &flags))
5092 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5093 unlock_task_sighand(p, &flags);
5095 /* can't set/change the rt policy */
5096 if (policy != p->policy && !rlim_rtprio)
5099 /* can't increase priority */
5100 if (param->sched_priority > p->rt_priority &&
5101 param->sched_priority > rlim_rtprio)
5105 * Like positive nice levels, dont allow tasks to
5106 * move out of SCHED_IDLE either:
5108 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5111 /* can't change other user's priorities */
5112 if ((current->euid != p->euid) &&
5113 (current->euid != p->uid))
5117 #ifdef CONFIG_RT_GROUP_SCHED
5119 * Do not allow realtime tasks into groups that have no runtime
5122 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5126 retval = security_task_setscheduler(p, policy, param);
5130 * make sure no PI-waiters arrive (or leave) while we are
5131 * changing the priority of the task:
5133 spin_lock_irqsave(&p->pi_lock, flags);
5135 * To be able to change p->policy safely, the apropriate
5136 * runqueue lock must be held.
5138 rq = __task_rq_lock(p);
5139 /* recheck policy now with rq lock held */
5140 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5141 policy = oldpolicy = -1;
5142 __task_rq_unlock(rq);
5143 spin_unlock_irqrestore(&p->pi_lock, flags);
5146 update_rq_clock(rq);
5147 on_rq = p->se.on_rq;
5148 running = task_current(rq, p);
5150 deactivate_task(rq, p, 0);
5152 p->sched_class->put_prev_task(rq, p);
5155 __setscheduler(rq, p, policy, param->sched_priority);
5158 p->sched_class->set_curr_task(rq);
5160 activate_task(rq, p, 0);
5162 check_class_changed(rq, p, prev_class, oldprio, running);
5164 __task_rq_unlock(rq);
5165 spin_unlock_irqrestore(&p->pi_lock, flags);
5167 rt_mutex_adjust_pi(p);
5171 EXPORT_SYMBOL_GPL(sched_setscheduler);
5174 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5176 struct sched_param lparam;
5177 struct task_struct *p;
5180 if (!param || pid < 0)
5182 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5187 p = find_process_by_pid(pid);
5189 retval = sched_setscheduler(p, policy, &lparam);
5196 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5197 * @pid: the pid in question.
5198 * @policy: new policy.
5199 * @param: structure containing the new RT priority.
5202 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5204 /* negative values for policy are not valid */
5208 return do_sched_setscheduler(pid, policy, param);
5212 * sys_sched_setparam - set/change the RT priority of a thread
5213 * @pid: the pid in question.
5214 * @param: structure containing the new RT priority.
5216 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5218 return do_sched_setscheduler(pid, -1, param);
5222 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5223 * @pid: the pid in question.
5225 asmlinkage long sys_sched_getscheduler(pid_t pid)
5227 struct task_struct *p;
5234 read_lock(&tasklist_lock);
5235 p = find_process_by_pid(pid);
5237 retval = security_task_getscheduler(p);
5241 read_unlock(&tasklist_lock);
5246 * sys_sched_getscheduler - get the RT priority of a thread
5247 * @pid: the pid in question.
5248 * @param: structure containing the RT priority.
5250 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5252 struct sched_param lp;
5253 struct task_struct *p;
5256 if (!param || pid < 0)
5259 read_lock(&tasklist_lock);
5260 p = find_process_by_pid(pid);
5265 retval = security_task_getscheduler(p);
5269 lp.sched_priority = p->rt_priority;
5270 read_unlock(&tasklist_lock);
5273 * This one might sleep, we cannot do it with a spinlock held ...
5275 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5280 read_unlock(&tasklist_lock);
5284 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5286 cpumask_t cpus_allowed;
5287 cpumask_t new_mask = *in_mask;
5288 struct task_struct *p;
5292 read_lock(&tasklist_lock);
5294 p = find_process_by_pid(pid);
5296 read_unlock(&tasklist_lock);
5302 * It is not safe to call set_cpus_allowed with the
5303 * tasklist_lock held. We will bump the task_struct's
5304 * usage count and then drop tasklist_lock.
5307 read_unlock(&tasklist_lock);
5310 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5311 !capable(CAP_SYS_NICE))
5314 retval = security_task_setscheduler(p, 0, NULL);
5318 cpuset_cpus_allowed(p, &cpus_allowed);
5319 cpus_and(new_mask, new_mask, cpus_allowed);
5321 retval = set_cpus_allowed_ptr(p, &new_mask);
5324 cpuset_cpus_allowed(p, &cpus_allowed);
5325 if (!cpus_subset(new_mask, cpus_allowed)) {
5327 * We must have raced with a concurrent cpuset
5328 * update. Just reset the cpus_allowed to the
5329 * cpuset's cpus_allowed
5331 new_mask = cpus_allowed;
5341 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5342 cpumask_t *new_mask)
5344 if (len < sizeof(cpumask_t)) {
5345 memset(new_mask, 0, sizeof(cpumask_t));
5346 } else if (len > sizeof(cpumask_t)) {
5347 len = sizeof(cpumask_t);
5349 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5353 * sys_sched_setaffinity - set the cpu affinity of a process
5354 * @pid: pid of the process
5355 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5356 * @user_mask_ptr: user-space pointer to the new cpu mask
5358 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5359 unsigned long __user *user_mask_ptr)
5364 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5368 return sched_setaffinity(pid, &new_mask);
5371 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5373 struct task_struct *p;
5377 read_lock(&tasklist_lock);
5380 p = find_process_by_pid(pid);
5384 retval = security_task_getscheduler(p);
5388 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5391 read_unlock(&tasklist_lock);
5398 * sys_sched_getaffinity - get the cpu affinity of a process
5399 * @pid: pid of the process
5400 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5401 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5403 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5404 unsigned long __user *user_mask_ptr)
5409 if (len < sizeof(cpumask_t))
5412 ret = sched_getaffinity(pid, &mask);
5416 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5419 return sizeof(cpumask_t);
5423 * sys_sched_yield - yield the current processor to other threads.
5425 * This function yields the current CPU to other tasks. If there are no
5426 * other threads running on this CPU then this function will return.
5428 asmlinkage long sys_sched_yield(void)
5430 struct rq *rq = this_rq_lock();
5432 schedstat_inc(rq, yld_count);
5433 current->sched_class->yield_task(rq);
5436 * Since we are going to call schedule() anyway, there's
5437 * no need to preempt or enable interrupts:
5439 __release(rq->lock);
5440 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5441 _raw_spin_unlock(&rq->lock);
5442 preempt_enable_no_resched();
5449 static void __cond_resched(void)
5451 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5452 __might_sleep(__FILE__, __LINE__);
5455 * The BKS might be reacquired before we have dropped
5456 * PREEMPT_ACTIVE, which could trigger a second
5457 * cond_resched() call.
5460 add_preempt_count(PREEMPT_ACTIVE);
5462 sub_preempt_count(PREEMPT_ACTIVE);
5463 } while (need_resched());
5466 int __sched _cond_resched(void)
5468 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5469 system_state == SYSTEM_RUNNING) {
5475 EXPORT_SYMBOL(_cond_resched);
5478 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5479 * call schedule, and on return reacquire the lock.
5481 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5482 * operations here to prevent schedule() from being called twice (once via
5483 * spin_unlock(), once by hand).
5485 int cond_resched_lock(spinlock_t *lock)
5487 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5490 if (spin_needbreak(lock) || resched) {
5492 if (resched && need_resched())
5501 EXPORT_SYMBOL(cond_resched_lock);
5503 int __sched cond_resched_softirq(void)
5505 BUG_ON(!in_softirq());
5507 if (need_resched() && system_state == SYSTEM_RUNNING) {
5515 EXPORT_SYMBOL(cond_resched_softirq);
5518 * yield - yield the current processor to other threads.
5520 * This is a shortcut for kernel-space yielding - it marks the
5521 * thread runnable and calls sys_sched_yield().
5523 void __sched yield(void)
5525 set_current_state(TASK_RUNNING);
5528 EXPORT_SYMBOL(yield);
5531 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5532 * that process accounting knows that this is a task in IO wait state.
5534 * But don't do that if it is a deliberate, throttling IO wait (this task
5535 * has set its backing_dev_info: the queue against which it should throttle)
5537 void __sched io_schedule(void)
5539 struct rq *rq = &__raw_get_cpu_var(runqueues);
5541 delayacct_blkio_start();
5542 atomic_inc(&rq->nr_iowait);
5544 atomic_dec(&rq->nr_iowait);
5545 delayacct_blkio_end();
5547 EXPORT_SYMBOL(io_schedule);
5549 long __sched io_schedule_timeout(long timeout)
5551 struct rq *rq = &__raw_get_cpu_var(runqueues);
5554 delayacct_blkio_start();
5555 atomic_inc(&rq->nr_iowait);
5556 ret = schedule_timeout(timeout);
5557 atomic_dec(&rq->nr_iowait);
5558 delayacct_blkio_end();
5563 * sys_sched_get_priority_max - return maximum RT priority.
5564 * @policy: scheduling class.
5566 * this syscall returns the maximum rt_priority that can be used
5567 * by a given scheduling class.
5569 asmlinkage long sys_sched_get_priority_max(int policy)
5576 ret = MAX_USER_RT_PRIO-1;
5588 * sys_sched_get_priority_min - return minimum RT priority.
5589 * @policy: scheduling class.
5591 * this syscall returns the minimum rt_priority that can be used
5592 * by a given scheduling class.
5594 asmlinkage long sys_sched_get_priority_min(int policy)
5612 * sys_sched_rr_get_interval - return the default timeslice of a process.
5613 * @pid: pid of the process.
5614 * @interval: userspace pointer to the timeslice value.
5616 * this syscall writes the default timeslice value of a given process
5617 * into the user-space timespec buffer. A value of '0' means infinity.
5620 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5622 struct task_struct *p;
5623 unsigned int time_slice;
5631 read_lock(&tasklist_lock);
5632 p = find_process_by_pid(pid);
5636 retval = security_task_getscheduler(p);
5641 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5642 * tasks that are on an otherwise idle runqueue:
5645 if (p->policy == SCHED_RR) {
5646 time_slice = DEF_TIMESLICE;
5647 } else if (p->policy != SCHED_FIFO) {
5648 struct sched_entity *se = &p->se;
5649 unsigned long flags;
5652 rq = task_rq_lock(p, &flags);
5653 if (rq->cfs.load.weight)
5654 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5655 task_rq_unlock(rq, &flags);
5657 read_unlock(&tasklist_lock);
5658 jiffies_to_timespec(time_slice, &t);
5659 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5663 read_unlock(&tasklist_lock);
5667 static const char stat_nam[] = "RSDTtZX";
5669 void sched_show_task(struct task_struct *p)
5671 unsigned long free = 0;
5674 state = p->state ? __ffs(p->state) + 1 : 0;
5675 printk(KERN_INFO "%-13.13s %c", p->comm,
5676 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5677 #if BITS_PER_LONG == 32
5678 if (state == TASK_RUNNING)
5679 printk(KERN_CONT " running ");
5681 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5683 if (state == TASK_RUNNING)
5684 printk(KERN_CONT " running task ");
5686 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5688 #ifdef CONFIG_DEBUG_STACK_USAGE
5690 unsigned long *n = end_of_stack(p);
5693 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5696 printk(KERN_CONT "%5lu %5d %6d\n", free,
5697 task_pid_nr(p), task_pid_nr(p->real_parent));
5699 show_stack(p, NULL);
5702 void show_state_filter(unsigned long state_filter)
5704 struct task_struct *g, *p;
5706 #if BITS_PER_LONG == 32
5708 " task PC stack pid father\n");
5711 " task PC stack pid father\n");
5713 read_lock(&tasklist_lock);
5714 do_each_thread(g, p) {
5716 * reset the NMI-timeout, listing all files on a slow
5717 * console might take alot of time:
5719 touch_nmi_watchdog();
5720 if (!state_filter || (p->state & state_filter))
5722 } while_each_thread(g, p);
5724 touch_all_softlockup_watchdogs();
5726 #ifdef CONFIG_SCHED_DEBUG
5727 sysrq_sched_debug_show();
5729 read_unlock(&tasklist_lock);
5731 * Only show locks if all tasks are dumped:
5733 if (state_filter == -1)
5734 debug_show_all_locks();
5737 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5739 idle->sched_class = &idle_sched_class;
5743 * init_idle - set up an idle thread for a given CPU
5744 * @idle: task in question
5745 * @cpu: cpu the idle task belongs to
5747 * NOTE: this function does not set the idle thread's NEED_RESCHED
5748 * flag, to make booting more robust.
5750 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5752 struct rq *rq = cpu_rq(cpu);
5753 unsigned long flags;
5756 idle->se.exec_start = sched_clock();
5758 idle->prio = idle->normal_prio = MAX_PRIO;
5759 idle->cpus_allowed = cpumask_of_cpu(cpu);
5760 __set_task_cpu(idle, cpu);
5762 spin_lock_irqsave(&rq->lock, flags);
5763 rq->curr = rq->idle = idle;
5764 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5767 spin_unlock_irqrestore(&rq->lock, flags);
5769 /* Set the preempt count _outside_ the spinlocks! */
5770 #if defined(CONFIG_PREEMPT)
5771 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5773 task_thread_info(idle)->preempt_count = 0;
5776 * The idle tasks have their own, simple scheduling class:
5778 idle->sched_class = &idle_sched_class;
5782 * In a system that switches off the HZ timer nohz_cpu_mask
5783 * indicates which cpus entered this state. This is used
5784 * in the rcu update to wait only for active cpus. For system
5785 * which do not switch off the HZ timer nohz_cpu_mask should
5786 * always be CPU_MASK_NONE.
5788 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5791 * Increase the granularity value when there are more CPUs,
5792 * because with more CPUs the 'effective latency' as visible
5793 * to users decreases. But the relationship is not linear,
5794 * so pick a second-best guess by going with the log2 of the
5797 * This idea comes from the SD scheduler of Con Kolivas:
5799 static inline void sched_init_granularity(void)
5801 unsigned int factor = 1 + ilog2(num_online_cpus());
5802 const unsigned long limit = 200000000;
5804 sysctl_sched_min_granularity *= factor;
5805 if (sysctl_sched_min_granularity > limit)
5806 sysctl_sched_min_granularity = limit;
5808 sysctl_sched_latency *= factor;
5809 if (sysctl_sched_latency > limit)
5810 sysctl_sched_latency = limit;
5812 sysctl_sched_wakeup_granularity *= factor;
5817 * This is how migration works:
5819 * 1) we queue a struct migration_req structure in the source CPU's
5820 * runqueue and wake up that CPU's migration thread.
5821 * 2) we down() the locked semaphore => thread blocks.
5822 * 3) migration thread wakes up (implicitly it forces the migrated
5823 * thread off the CPU)
5824 * 4) it gets the migration request and checks whether the migrated
5825 * task is still in the wrong runqueue.
5826 * 5) if it's in the wrong runqueue then the migration thread removes
5827 * it and puts it into the right queue.
5828 * 6) migration thread up()s the semaphore.
5829 * 7) we wake up and the migration is done.
5833 * Change a given task's CPU affinity. Migrate the thread to a
5834 * proper CPU and schedule it away if the CPU it's executing on
5835 * is removed from the allowed bitmask.
5837 * NOTE: the caller must have a valid reference to the task, the
5838 * task must not exit() & deallocate itself prematurely. The
5839 * call is not atomic; no spinlocks may be held.
5841 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5843 struct migration_req req;
5844 unsigned long flags;
5848 rq = task_rq_lock(p, &flags);
5849 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5854 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5855 !cpus_equal(p->cpus_allowed, *new_mask))) {
5860 if (p->sched_class->set_cpus_allowed)
5861 p->sched_class->set_cpus_allowed(p, new_mask);
5863 p->cpus_allowed = *new_mask;
5864 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5867 /* Can the task run on the task's current CPU? If so, we're done */
5868 if (cpu_isset(task_cpu(p), *new_mask))
5871 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5872 /* Need help from migration thread: drop lock and wait. */
5873 task_rq_unlock(rq, &flags);
5874 wake_up_process(rq->migration_thread);
5875 wait_for_completion(&req.done);
5876 tlb_migrate_finish(p->mm);
5880 task_rq_unlock(rq, &flags);
5884 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5887 * Move (not current) task off this cpu, onto dest cpu. We're doing
5888 * this because either it can't run here any more (set_cpus_allowed()
5889 * away from this CPU, or CPU going down), or because we're
5890 * attempting to rebalance this task on exec (sched_exec).
5892 * So we race with normal scheduler movements, but that's OK, as long
5893 * as the task is no longer on this CPU.
5895 * Returns non-zero if task was successfully migrated.
5897 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5899 struct rq *rq_dest, *rq_src;
5902 if (unlikely(cpu_is_offline(dest_cpu)))
5905 rq_src = cpu_rq(src_cpu);
5906 rq_dest = cpu_rq(dest_cpu);
5908 double_rq_lock(rq_src, rq_dest);
5909 /* Already moved. */
5910 if (task_cpu(p) != src_cpu)
5912 /* Affinity changed (again). */
5913 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5916 on_rq = p->se.on_rq;
5918 deactivate_task(rq_src, p, 0);
5920 set_task_cpu(p, dest_cpu);
5922 activate_task(rq_dest, p, 0);
5923 check_preempt_curr(rq_dest, p);
5927 double_rq_unlock(rq_src, rq_dest);
5932 * migration_thread - this is a highprio system thread that performs
5933 * thread migration by bumping thread off CPU then 'pushing' onto
5936 static int migration_thread(void *data)
5938 int cpu = (long)data;
5942 BUG_ON(rq->migration_thread != current);
5944 set_current_state(TASK_INTERRUPTIBLE);
5945 while (!kthread_should_stop()) {
5946 struct migration_req *req;
5947 struct list_head *head;
5949 spin_lock_irq(&rq->lock);
5951 if (cpu_is_offline(cpu)) {
5952 spin_unlock_irq(&rq->lock);
5956 if (rq->active_balance) {
5957 active_load_balance(rq, cpu);
5958 rq->active_balance = 0;
5961 head = &rq->migration_queue;
5963 if (list_empty(head)) {
5964 spin_unlock_irq(&rq->lock);
5966 set_current_state(TASK_INTERRUPTIBLE);
5969 req = list_entry(head->next, struct migration_req, list);
5970 list_del_init(head->next);
5972 spin_unlock(&rq->lock);
5973 __migrate_task(req->task, cpu, req->dest_cpu);
5976 complete(&req->done);
5978 __set_current_state(TASK_RUNNING);
5982 /* Wait for kthread_stop */
5983 set_current_state(TASK_INTERRUPTIBLE);
5984 while (!kthread_should_stop()) {
5986 set_current_state(TASK_INTERRUPTIBLE);
5988 __set_current_state(TASK_RUNNING);
5992 #ifdef CONFIG_HOTPLUG_CPU
5994 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5998 local_irq_disable();
5999 ret = __migrate_task(p, src_cpu, dest_cpu);
6005 * Figure out where task on dead CPU should go, use force if necessary.
6006 * NOTE: interrupts should be disabled by the caller
6008 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6010 unsigned long flags;
6017 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6018 cpus_and(mask, mask, p->cpus_allowed);
6019 dest_cpu = any_online_cpu(mask);
6021 /* On any allowed CPU? */
6022 if (dest_cpu >= nr_cpu_ids)
6023 dest_cpu = any_online_cpu(p->cpus_allowed);
6025 /* No more Mr. Nice Guy. */
6026 if (dest_cpu >= nr_cpu_ids) {
6027 cpumask_t cpus_allowed;
6029 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6031 * Try to stay on the same cpuset, where the
6032 * current cpuset may be a subset of all cpus.
6033 * The cpuset_cpus_allowed_locked() variant of
6034 * cpuset_cpus_allowed() will not block. It must be
6035 * called within calls to cpuset_lock/cpuset_unlock.
6037 rq = task_rq_lock(p, &flags);
6038 p->cpus_allowed = cpus_allowed;
6039 dest_cpu = any_online_cpu(p->cpus_allowed);
6040 task_rq_unlock(rq, &flags);
6043 * Don't tell them about moving exiting tasks or
6044 * kernel threads (both mm NULL), since they never
6047 if (p->mm && printk_ratelimit()) {
6048 printk(KERN_INFO "process %d (%s) no "
6049 "longer affine to cpu%d\n",
6050 task_pid_nr(p), p->comm, dead_cpu);
6053 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6057 * While a dead CPU has no uninterruptible tasks queued at this point,
6058 * it might still have a nonzero ->nr_uninterruptible counter, because
6059 * for performance reasons the counter is not stricly tracking tasks to
6060 * their home CPUs. So we just add the counter to another CPU's counter,
6061 * to keep the global sum constant after CPU-down:
6063 static void migrate_nr_uninterruptible(struct rq *rq_src)
6065 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6066 unsigned long flags;
6068 local_irq_save(flags);
6069 double_rq_lock(rq_src, rq_dest);
6070 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6071 rq_src->nr_uninterruptible = 0;
6072 double_rq_unlock(rq_src, rq_dest);
6073 local_irq_restore(flags);
6076 /* Run through task list and migrate tasks from the dead cpu. */
6077 static void migrate_live_tasks(int src_cpu)
6079 struct task_struct *p, *t;
6081 read_lock(&tasklist_lock);
6083 do_each_thread(t, p) {
6087 if (task_cpu(p) == src_cpu)
6088 move_task_off_dead_cpu(src_cpu, p);
6089 } while_each_thread(t, p);
6091 read_unlock(&tasklist_lock);
6095 * Schedules idle task to be the next runnable task on current CPU.
6096 * It does so by boosting its priority to highest possible.
6097 * Used by CPU offline code.
6099 void sched_idle_next(void)
6101 int this_cpu = smp_processor_id();
6102 struct rq *rq = cpu_rq(this_cpu);
6103 struct task_struct *p = rq->idle;
6104 unsigned long flags;
6106 /* cpu has to be offline */
6107 BUG_ON(cpu_online(this_cpu));
6110 * Strictly not necessary since rest of the CPUs are stopped by now
6111 * and interrupts disabled on the current cpu.
6113 spin_lock_irqsave(&rq->lock, flags);
6115 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6117 update_rq_clock(rq);
6118 activate_task(rq, p, 0);
6120 spin_unlock_irqrestore(&rq->lock, flags);
6124 * Ensures that the idle task is using init_mm right before its cpu goes
6127 void idle_task_exit(void)
6129 struct mm_struct *mm = current->active_mm;
6131 BUG_ON(cpu_online(smp_processor_id()));
6134 switch_mm(mm, &init_mm, current);
6138 /* called under rq->lock with disabled interrupts */
6139 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6141 struct rq *rq = cpu_rq(dead_cpu);
6143 /* Must be exiting, otherwise would be on tasklist. */
6144 BUG_ON(!p->exit_state);
6146 /* Cannot have done final schedule yet: would have vanished. */
6147 BUG_ON(p->state == TASK_DEAD);
6152 * Drop lock around migration; if someone else moves it,
6153 * that's OK. No task can be added to this CPU, so iteration is
6156 spin_unlock_irq(&rq->lock);
6157 move_task_off_dead_cpu(dead_cpu, p);
6158 spin_lock_irq(&rq->lock);
6163 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6164 static void migrate_dead_tasks(unsigned int dead_cpu)
6166 struct rq *rq = cpu_rq(dead_cpu);
6167 struct task_struct *next;
6170 if (!rq->nr_running)
6172 update_rq_clock(rq);
6173 next = pick_next_task(rq, rq->curr);
6176 migrate_dead(dead_cpu, next);
6180 #endif /* CONFIG_HOTPLUG_CPU */
6182 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6184 static struct ctl_table sd_ctl_dir[] = {
6186 .procname = "sched_domain",
6192 static struct ctl_table sd_ctl_root[] = {
6194 .ctl_name = CTL_KERN,
6195 .procname = "kernel",
6197 .child = sd_ctl_dir,
6202 static struct ctl_table *sd_alloc_ctl_entry(int n)
6204 struct ctl_table *entry =
6205 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6210 static void sd_free_ctl_entry(struct ctl_table **tablep)
6212 struct ctl_table *entry;
6215 * In the intermediate directories, both the child directory and
6216 * procname are dynamically allocated and could fail but the mode
6217 * will always be set. In the lowest directory the names are
6218 * static strings and all have proc handlers.
6220 for (entry = *tablep; entry->mode; entry++) {
6222 sd_free_ctl_entry(&entry->child);
6223 if (entry->proc_handler == NULL)
6224 kfree(entry->procname);
6232 set_table_entry(struct ctl_table *entry,
6233 const char *procname, void *data, int maxlen,
6234 mode_t mode, proc_handler *proc_handler)
6236 entry->procname = procname;
6238 entry->maxlen = maxlen;
6240 entry->proc_handler = proc_handler;
6243 static struct ctl_table *
6244 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6246 struct ctl_table *table = sd_alloc_ctl_entry(12);
6251 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6252 sizeof(long), 0644, proc_doulongvec_minmax);
6253 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6254 sizeof(long), 0644, proc_doulongvec_minmax);
6255 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6256 sizeof(int), 0644, proc_dointvec_minmax);
6257 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6258 sizeof(int), 0644, proc_dointvec_minmax);
6259 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6260 sizeof(int), 0644, proc_dointvec_minmax);
6261 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6262 sizeof(int), 0644, proc_dointvec_minmax);
6263 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6264 sizeof(int), 0644, proc_dointvec_minmax);
6265 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6266 sizeof(int), 0644, proc_dointvec_minmax);
6267 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6268 sizeof(int), 0644, proc_dointvec_minmax);
6269 set_table_entry(&table[9], "cache_nice_tries",
6270 &sd->cache_nice_tries,
6271 sizeof(int), 0644, proc_dointvec_minmax);
6272 set_table_entry(&table[10], "flags", &sd->flags,
6273 sizeof(int), 0644, proc_dointvec_minmax);
6274 /* &table[11] is terminator */
6279 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6281 struct ctl_table *entry, *table;
6282 struct sched_domain *sd;
6283 int domain_num = 0, i;
6286 for_each_domain(cpu, sd)
6288 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6293 for_each_domain(cpu, sd) {
6294 snprintf(buf, 32, "domain%d", i);
6295 entry->procname = kstrdup(buf, GFP_KERNEL);
6297 entry->child = sd_alloc_ctl_domain_table(sd);
6304 static struct ctl_table_header *sd_sysctl_header;
6305 static void register_sched_domain_sysctl(void)
6307 int i, cpu_num = num_online_cpus();
6308 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6311 WARN_ON(sd_ctl_dir[0].child);
6312 sd_ctl_dir[0].child = entry;
6317 for_each_online_cpu(i) {
6318 snprintf(buf, 32, "cpu%d", i);
6319 entry->procname = kstrdup(buf, GFP_KERNEL);
6321 entry->child = sd_alloc_ctl_cpu_table(i);
6325 WARN_ON(sd_sysctl_header);
6326 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6329 /* may be called multiple times per register */
6330 static void unregister_sched_domain_sysctl(void)
6332 if (sd_sysctl_header)
6333 unregister_sysctl_table(sd_sysctl_header);
6334 sd_sysctl_header = NULL;
6335 if (sd_ctl_dir[0].child)
6336 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6339 static void register_sched_domain_sysctl(void)
6342 static void unregister_sched_domain_sysctl(void)
6347 static void set_rq_online(struct rq *rq)
6350 const struct sched_class *class;
6352 cpu_set(rq->cpu, rq->rd->online);
6355 for_each_class(class) {
6356 if (class->rq_online)
6357 class->rq_online(rq);
6362 static void set_rq_offline(struct rq *rq)
6365 const struct sched_class *class;
6367 for_each_class(class) {
6368 if (class->rq_offline)
6369 class->rq_offline(rq);
6372 cpu_clear(rq->cpu, rq->rd->online);
6378 * migration_call - callback that gets triggered when a CPU is added.
6379 * Here we can start up the necessary migration thread for the new CPU.
6381 static int __cpuinit
6382 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6384 struct task_struct *p;
6385 int cpu = (long)hcpu;
6386 unsigned long flags;
6391 case CPU_UP_PREPARE:
6392 case CPU_UP_PREPARE_FROZEN:
6393 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6396 kthread_bind(p, cpu);
6397 /* Must be high prio: stop_machine expects to yield to it. */
6398 rq = task_rq_lock(p, &flags);
6399 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6400 task_rq_unlock(rq, &flags);
6401 cpu_rq(cpu)->migration_thread = p;
6405 case CPU_ONLINE_FROZEN:
6406 /* Strictly unnecessary, as first user will wake it. */
6407 wake_up_process(cpu_rq(cpu)->migration_thread);
6409 /* Update our root-domain */
6411 spin_lock_irqsave(&rq->lock, flags);
6413 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6417 spin_unlock_irqrestore(&rq->lock, flags);
6420 #ifdef CONFIG_HOTPLUG_CPU
6421 case CPU_UP_CANCELED:
6422 case CPU_UP_CANCELED_FROZEN:
6423 if (!cpu_rq(cpu)->migration_thread)
6425 /* Unbind it from offline cpu so it can run. Fall thru. */
6426 kthread_bind(cpu_rq(cpu)->migration_thread,
6427 any_online_cpu(cpu_online_map));
6428 kthread_stop(cpu_rq(cpu)->migration_thread);
6429 cpu_rq(cpu)->migration_thread = NULL;
6433 case CPU_DEAD_FROZEN:
6434 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6435 migrate_live_tasks(cpu);
6437 kthread_stop(rq->migration_thread);
6438 rq->migration_thread = NULL;
6439 /* Idle task back to normal (off runqueue, low prio) */
6440 spin_lock_irq(&rq->lock);
6441 update_rq_clock(rq);
6442 deactivate_task(rq, rq->idle, 0);
6443 rq->idle->static_prio = MAX_PRIO;
6444 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6445 rq->idle->sched_class = &idle_sched_class;
6446 migrate_dead_tasks(cpu);
6447 spin_unlock_irq(&rq->lock);
6449 migrate_nr_uninterruptible(rq);
6450 BUG_ON(rq->nr_running != 0);
6453 * No need to migrate the tasks: it was best-effort if
6454 * they didn't take sched_hotcpu_mutex. Just wake up
6457 spin_lock_irq(&rq->lock);
6458 while (!list_empty(&rq->migration_queue)) {
6459 struct migration_req *req;
6461 req = list_entry(rq->migration_queue.next,
6462 struct migration_req, list);
6463 list_del_init(&req->list);
6464 complete(&req->done);
6466 spin_unlock_irq(&rq->lock);
6470 case CPU_DYING_FROZEN:
6471 /* Update our root-domain */
6473 spin_lock_irqsave(&rq->lock, flags);
6475 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6478 spin_unlock_irqrestore(&rq->lock, flags);
6485 /* Register at highest priority so that task migration (migrate_all_tasks)
6486 * happens before everything else.
6488 static struct notifier_block __cpuinitdata migration_notifier = {
6489 .notifier_call = migration_call,
6493 void __init migration_init(void)
6495 void *cpu = (void *)(long)smp_processor_id();
6498 /* Start one for the boot CPU: */
6499 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6500 BUG_ON(err == NOTIFY_BAD);
6501 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6502 register_cpu_notifier(&migration_notifier);
6508 #ifdef CONFIG_SCHED_DEBUG
6510 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6523 case SD_LV_ALLNODES:
6532 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6533 cpumask_t *groupmask)
6535 struct sched_group *group = sd->groups;
6538 cpulist_scnprintf(str, sizeof(str), sd->span);
6539 cpus_clear(*groupmask);
6541 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6543 if (!(sd->flags & SD_LOAD_BALANCE)) {
6544 printk("does not load-balance\n");
6546 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6551 printk(KERN_CONT "span %s level %s\n",
6552 str, sd_level_to_string(sd->level));
6554 if (!cpu_isset(cpu, sd->span)) {
6555 printk(KERN_ERR "ERROR: domain->span does not contain "
6558 if (!cpu_isset(cpu, group->cpumask)) {
6559 printk(KERN_ERR "ERROR: domain->groups does not contain"
6563 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6567 printk(KERN_ERR "ERROR: group is NULL\n");
6571 if (!group->__cpu_power) {
6572 printk(KERN_CONT "\n");
6573 printk(KERN_ERR "ERROR: domain->cpu_power not "
6578 if (!cpus_weight(group->cpumask)) {
6579 printk(KERN_CONT "\n");
6580 printk(KERN_ERR "ERROR: empty group\n");
6584 if (cpus_intersects(*groupmask, group->cpumask)) {
6585 printk(KERN_CONT "\n");
6586 printk(KERN_ERR "ERROR: repeated CPUs\n");
6590 cpus_or(*groupmask, *groupmask, group->cpumask);
6592 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6593 printk(KERN_CONT " %s", str);
6595 group = group->next;
6596 } while (group != sd->groups);
6597 printk(KERN_CONT "\n");
6599 if (!cpus_equal(sd->span, *groupmask))
6600 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6602 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6603 printk(KERN_ERR "ERROR: parent span is not a superset "
6604 "of domain->span\n");
6608 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6610 cpumask_t *groupmask;
6614 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6618 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6620 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6622 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6627 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6636 #else /* !CONFIG_SCHED_DEBUG */
6637 # define sched_domain_debug(sd, cpu) do { } while (0)
6638 #endif /* CONFIG_SCHED_DEBUG */
6640 static int sd_degenerate(struct sched_domain *sd)
6642 if (cpus_weight(sd->span) == 1)
6645 /* Following flags need at least 2 groups */
6646 if (sd->flags & (SD_LOAD_BALANCE |
6647 SD_BALANCE_NEWIDLE |
6651 SD_SHARE_PKG_RESOURCES)) {
6652 if (sd->groups != sd->groups->next)
6656 /* Following flags don't use groups */
6657 if (sd->flags & (SD_WAKE_IDLE |
6666 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6668 unsigned long cflags = sd->flags, pflags = parent->flags;
6670 if (sd_degenerate(parent))
6673 if (!cpus_equal(sd->span, parent->span))
6676 /* Does parent contain flags not in child? */
6677 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6678 if (cflags & SD_WAKE_AFFINE)
6679 pflags &= ~SD_WAKE_BALANCE;
6680 /* Flags needing groups don't count if only 1 group in parent */
6681 if (parent->groups == parent->groups->next) {
6682 pflags &= ~(SD_LOAD_BALANCE |
6683 SD_BALANCE_NEWIDLE |
6687 SD_SHARE_PKG_RESOURCES);
6689 if (~cflags & pflags)
6695 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6697 unsigned long flags;
6699 spin_lock_irqsave(&rq->lock, flags);
6702 struct root_domain *old_rd = rq->rd;
6704 if (cpu_isset(rq->cpu, old_rd->online))
6707 cpu_clear(rq->cpu, old_rd->span);
6709 if (atomic_dec_and_test(&old_rd->refcount))
6713 atomic_inc(&rd->refcount);
6716 cpu_set(rq->cpu, rd->span);
6717 if (cpu_isset(rq->cpu, cpu_online_map))
6720 spin_unlock_irqrestore(&rq->lock, flags);
6723 static void init_rootdomain(struct root_domain *rd)
6725 memset(rd, 0, sizeof(*rd));
6727 cpus_clear(rd->span);
6728 cpus_clear(rd->online);
6730 cpupri_init(&rd->cpupri);
6733 static void init_defrootdomain(void)
6735 init_rootdomain(&def_root_domain);
6736 atomic_set(&def_root_domain.refcount, 1);
6739 static struct root_domain *alloc_rootdomain(void)
6741 struct root_domain *rd;
6743 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6747 init_rootdomain(rd);
6753 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6754 * hold the hotplug lock.
6757 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6759 struct rq *rq = cpu_rq(cpu);
6760 struct sched_domain *tmp;
6762 /* Remove the sched domains which do not contribute to scheduling. */
6763 for (tmp = sd; tmp; tmp = tmp->parent) {
6764 struct sched_domain *parent = tmp->parent;
6767 if (sd_parent_degenerate(tmp, parent)) {
6768 tmp->parent = parent->parent;
6770 parent->parent->child = tmp;
6774 if (sd && sd_degenerate(sd)) {
6780 sched_domain_debug(sd, cpu);
6782 rq_attach_root(rq, rd);
6783 rcu_assign_pointer(rq->sd, sd);
6786 /* cpus with isolated domains */
6787 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6789 /* Setup the mask of cpus configured for isolated domains */
6790 static int __init isolated_cpu_setup(char *str)
6792 int ints[NR_CPUS], i;
6794 str = get_options(str, ARRAY_SIZE(ints), ints);
6795 cpus_clear(cpu_isolated_map);
6796 for (i = 1; i <= ints[0]; i++)
6797 if (ints[i] < NR_CPUS)
6798 cpu_set(ints[i], cpu_isolated_map);
6802 __setup("isolcpus=", isolated_cpu_setup);
6805 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6806 * to a function which identifies what group(along with sched group) a CPU
6807 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6808 * (due to the fact that we keep track of groups covered with a cpumask_t).
6810 * init_sched_build_groups will build a circular linked list of the groups
6811 * covered by the given span, and will set each group's ->cpumask correctly,
6812 * and ->cpu_power to 0.
6815 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6816 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6817 struct sched_group **sg,
6818 cpumask_t *tmpmask),
6819 cpumask_t *covered, cpumask_t *tmpmask)
6821 struct sched_group *first = NULL, *last = NULL;
6824 cpus_clear(*covered);
6826 for_each_cpu_mask(i, *span) {
6827 struct sched_group *sg;
6828 int group = group_fn(i, cpu_map, &sg, tmpmask);
6831 if (cpu_isset(i, *covered))
6834 cpus_clear(sg->cpumask);
6835 sg->__cpu_power = 0;
6837 for_each_cpu_mask(j, *span) {
6838 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6841 cpu_set(j, *covered);
6842 cpu_set(j, sg->cpumask);
6853 #define SD_NODES_PER_DOMAIN 16
6858 * find_next_best_node - find the next node to include in a sched_domain
6859 * @node: node whose sched_domain we're building
6860 * @used_nodes: nodes already in the sched_domain
6862 * Find the next node to include in a given scheduling domain. Simply
6863 * finds the closest node not already in the @used_nodes map.
6865 * Should use nodemask_t.
6867 static int find_next_best_node(int node, nodemask_t *used_nodes)
6869 int i, n, val, min_val, best_node = 0;
6873 for (i = 0; i < MAX_NUMNODES; i++) {
6874 /* Start at @node */
6875 n = (node + i) % MAX_NUMNODES;
6877 if (!nr_cpus_node(n))
6880 /* Skip already used nodes */
6881 if (node_isset(n, *used_nodes))
6884 /* Simple min distance search */
6885 val = node_distance(node, n);
6887 if (val < min_val) {
6893 node_set(best_node, *used_nodes);
6898 * sched_domain_node_span - get a cpumask for a node's sched_domain
6899 * @node: node whose cpumask we're constructing
6900 * @span: resulting cpumask
6902 * Given a node, construct a good cpumask for its sched_domain to span. It
6903 * should be one that prevents unnecessary balancing, but also spreads tasks
6906 static void sched_domain_node_span(int node, cpumask_t *span)
6908 nodemask_t used_nodes;
6909 node_to_cpumask_ptr(nodemask, node);
6913 nodes_clear(used_nodes);
6915 cpus_or(*span, *span, *nodemask);
6916 node_set(node, used_nodes);
6918 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6919 int next_node = find_next_best_node(node, &used_nodes);
6921 node_to_cpumask_ptr_next(nodemask, next_node);
6922 cpus_or(*span, *span, *nodemask);
6925 #endif /* CONFIG_NUMA */
6927 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6930 * SMT sched-domains:
6932 #ifdef CONFIG_SCHED_SMT
6933 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6934 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6937 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6941 *sg = &per_cpu(sched_group_cpus, cpu);
6944 #endif /* CONFIG_SCHED_SMT */
6947 * multi-core sched-domains:
6949 #ifdef CONFIG_SCHED_MC
6950 static DEFINE_PER_CPU(struct sched_domain, core_domains);
6951 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
6952 #endif /* CONFIG_SCHED_MC */
6954 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6956 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6961 *mask = per_cpu(cpu_sibling_map, cpu);
6962 cpus_and(*mask, *mask, *cpu_map);
6963 group = first_cpu(*mask);
6965 *sg = &per_cpu(sched_group_core, group);
6968 #elif defined(CONFIG_SCHED_MC)
6970 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6974 *sg = &per_cpu(sched_group_core, cpu);
6979 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
6980 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
6983 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6987 #ifdef CONFIG_SCHED_MC
6988 *mask = cpu_coregroup_map(cpu);
6989 cpus_and(*mask, *mask, *cpu_map);
6990 group = first_cpu(*mask);
6991 #elif defined(CONFIG_SCHED_SMT)
6992 *mask = per_cpu(cpu_sibling_map, cpu);
6993 cpus_and(*mask, *mask, *cpu_map);
6994 group = first_cpu(*mask);
6999 *sg = &per_cpu(sched_group_phys, group);
7005 * The init_sched_build_groups can't handle what we want to do with node
7006 * groups, so roll our own. Now each node has its own list of groups which
7007 * gets dynamically allocated.
7009 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7010 static struct sched_group ***sched_group_nodes_bycpu;
7012 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7013 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7015 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7016 struct sched_group **sg, cpumask_t *nodemask)
7020 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7021 cpus_and(*nodemask, *nodemask, *cpu_map);
7022 group = first_cpu(*nodemask);
7025 *sg = &per_cpu(sched_group_allnodes, group);
7029 static void init_numa_sched_groups_power(struct sched_group *group_head)
7031 struct sched_group *sg = group_head;
7037 for_each_cpu_mask(j, sg->cpumask) {
7038 struct sched_domain *sd;
7040 sd = &per_cpu(phys_domains, j);
7041 if (j != first_cpu(sd->groups->cpumask)) {
7043 * Only add "power" once for each
7049 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7052 } while (sg != group_head);
7054 #endif /* CONFIG_NUMA */
7057 /* Free memory allocated for various sched_group structures */
7058 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7062 for_each_cpu_mask(cpu, *cpu_map) {
7063 struct sched_group **sched_group_nodes
7064 = sched_group_nodes_bycpu[cpu];
7066 if (!sched_group_nodes)
7069 for (i = 0; i < MAX_NUMNODES; i++) {
7070 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7072 *nodemask = node_to_cpumask(i);
7073 cpus_and(*nodemask, *nodemask, *cpu_map);
7074 if (cpus_empty(*nodemask))
7084 if (oldsg != sched_group_nodes[i])
7087 kfree(sched_group_nodes);
7088 sched_group_nodes_bycpu[cpu] = NULL;
7091 #else /* !CONFIG_NUMA */
7092 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7095 #endif /* CONFIG_NUMA */
7098 * Initialize sched groups cpu_power.
7100 * cpu_power indicates the capacity of sched group, which is used while
7101 * distributing the load between different sched groups in a sched domain.
7102 * Typically cpu_power for all the groups in a sched domain will be same unless
7103 * there are asymmetries in the topology. If there are asymmetries, group
7104 * having more cpu_power will pickup more load compared to the group having
7107 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7108 * the maximum number of tasks a group can handle in the presence of other idle
7109 * or lightly loaded groups in the same sched domain.
7111 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7113 struct sched_domain *child;
7114 struct sched_group *group;
7116 WARN_ON(!sd || !sd->groups);
7118 if (cpu != first_cpu(sd->groups->cpumask))
7123 sd->groups->__cpu_power = 0;
7126 * For perf policy, if the groups in child domain share resources
7127 * (for example cores sharing some portions of the cache hierarchy
7128 * or SMT), then set this domain groups cpu_power such that each group
7129 * can handle only one task, when there are other idle groups in the
7130 * same sched domain.
7132 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7134 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7135 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7140 * add cpu_power of each child group to this groups cpu_power
7142 group = child->groups;
7144 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7145 group = group->next;
7146 } while (group != child->groups);
7150 * Initializers for schedule domains
7151 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7154 #define SD_INIT(sd, type) sd_init_##type(sd)
7155 #define SD_INIT_FUNC(type) \
7156 static noinline void sd_init_##type(struct sched_domain *sd) \
7158 memset(sd, 0, sizeof(*sd)); \
7159 *sd = SD_##type##_INIT; \
7160 sd->level = SD_LV_##type; \
7165 SD_INIT_FUNC(ALLNODES)
7168 #ifdef CONFIG_SCHED_SMT
7169 SD_INIT_FUNC(SIBLING)
7171 #ifdef CONFIG_SCHED_MC
7176 * To minimize stack usage kmalloc room for cpumasks and share the
7177 * space as the usage in build_sched_domains() dictates. Used only
7178 * if the amount of space is significant.
7181 cpumask_t tmpmask; /* make this one first */
7184 cpumask_t this_sibling_map;
7185 cpumask_t this_core_map;
7187 cpumask_t send_covered;
7190 cpumask_t domainspan;
7192 cpumask_t notcovered;
7197 #define SCHED_CPUMASK_ALLOC 1
7198 #define SCHED_CPUMASK_FREE(v) kfree(v)
7199 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7201 #define SCHED_CPUMASK_ALLOC 0
7202 #define SCHED_CPUMASK_FREE(v)
7203 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7206 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7207 ((unsigned long)(a) + offsetof(struct allmasks, v))
7209 static int default_relax_domain_level = -1;
7211 static int __init setup_relax_domain_level(char *str)
7215 val = simple_strtoul(str, NULL, 0);
7216 if (val < SD_LV_MAX)
7217 default_relax_domain_level = val;
7221 __setup("relax_domain_level=", setup_relax_domain_level);
7223 static void set_domain_attribute(struct sched_domain *sd,
7224 struct sched_domain_attr *attr)
7228 if (!attr || attr->relax_domain_level < 0) {
7229 if (default_relax_domain_level < 0)
7232 request = default_relax_domain_level;
7234 request = attr->relax_domain_level;
7235 if (request < sd->level) {
7236 /* turn off idle balance on this domain */
7237 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7239 /* turn on idle balance on this domain */
7240 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7245 * Build sched domains for a given set of cpus and attach the sched domains
7246 * to the individual cpus
7248 static int __build_sched_domains(const cpumask_t *cpu_map,
7249 struct sched_domain_attr *attr)
7252 struct root_domain *rd;
7253 SCHED_CPUMASK_DECLARE(allmasks);
7256 struct sched_group **sched_group_nodes = NULL;
7257 int sd_allnodes = 0;
7260 * Allocate the per-node list of sched groups
7262 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7264 if (!sched_group_nodes) {
7265 printk(KERN_WARNING "Can not alloc sched group node list\n");
7270 rd = alloc_rootdomain();
7272 printk(KERN_WARNING "Cannot alloc root domain\n");
7274 kfree(sched_group_nodes);
7279 #if SCHED_CPUMASK_ALLOC
7280 /* get space for all scratch cpumask variables */
7281 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7283 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7286 kfree(sched_group_nodes);
7291 tmpmask = (cpumask_t *)allmasks;
7295 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7299 * Set up domains for cpus specified by the cpu_map.
7301 for_each_cpu_mask(i, *cpu_map) {
7302 struct sched_domain *sd = NULL, *p;
7303 SCHED_CPUMASK_VAR(nodemask, allmasks);
7305 *nodemask = node_to_cpumask(cpu_to_node(i));
7306 cpus_and(*nodemask, *nodemask, *cpu_map);
7309 if (cpus_weight(*cpu_map) >
7310 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7311 sd = &per_cpu(allnodes_domains, i);
7312 SD_INIT(sd, ALLNODES);
7313 set_domain_attribute(sd, attr);
7314 sd->span = *cpu_map;
7315 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7321 sd = &per_cpu(node_domains, i);
7323 set_domain_attribute(sd, attr);
7324 sched_domain_node_span(cpu_to_node(i), &sd->span);
7328 cpus_and(sd->span, sd->span, *cpu_map);
7332 sd = &per_cpu(phys_domains, i);
7334 set_domain_attribute(sd, attr);
7335 sd->span = *nodemask;
7339 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7341 #ifdef CONFIG_SCHED_MC
7343 sd = &per_cpu(core_domains, i);
7345 set_domain_attribute(sd, attr);
7346 sd->span = cpu_coregroup_map(i);
7347 cpus_and(sd->span, sd->span, *cpu_map);
7350 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7353 #ifdef CONFIG_SCHED_SMT
7355 sd = &per_cpu(cpu_domains, i);
7356 SD_INIT(sd, SIBLING);
7357 set_domain_attribute(sd, attr);
7358 sd->span = per_cpu(cpu_sibling_map, i);
7359 cpus_and(sd->span, sd->span, *cpu_map);
7362 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7366 #ifdef CONFIG_SCHED_SMT
7367 /* Set up CPU (sibling) groups */
7368 for_each_cpu_mask(i, *cpu_map) {
7369 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7370 SCHED_CPUMASK_VAR(send_covered, allmasks);
7372 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7373 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7374 if (i != first_cpu(*this_sibling_map))
7377 init_sched_build_groups(this_sibling_map, cpu_map,
7379 send_covered, tmpmask);
7383 #ifdef CONFIG_SCHED_MC
7384 /* Set up multi-core groups */
7385 for_each_cpu_mask(i, *cpu_map) {
7386 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7387 SCHED_CPUMASK_VAR(send_covered, allmasks);
7389 *this_core_map = cpu_coregroup_map(i);
7390 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7391 if (i != first_cpu(*this_core_map))
7394 init_sched_build_groups(this_core_map, cpu_map,
7396 send_covered, tmpmask);
7400 /* Set up physical groups */
7401 for (i = 0; i < MAX_NUMNODES; i++) {
7402 SCHED_CPUMASK_VAR(nodemask, allmasks);
7403 SCHED_CPUMASK_VAR(send_covered, allmasks);
7405 *nodemask = node_to_cpumask(i);
7406 cpus_and(*nodemask, *nodemask, *cpu_map);
7407 if (cpus_empty(*nodemask))
7410 init_sched_build_groups(nodemask, cpu_map,
7412 send_covered, tmpmask);
7416 /* Set up node groups */
7418 SCHED_CPUMASK_VAR(send_covered, allmasks);
7420 init_sched_build_groups(cpu_map, cpu_map,
7421 &cpu_to_allnodes_group,
7422 send_covered, tmpmask);
7425 for (i = 0; i < MAX_NUMNODES; i++) {
7426 /* Set up node groups */
7427 struct sched_group *sg, *prev;
7428 SCHED_CPUMASK_VAR(nodemask, allmasks);
7429 SCHED_CPUMASK_VAR(domainspan, allmasks);
7430 SCHED_CPUMASK_VAR(covered, allmasks);
7433 *nodemask = node_to_cpumask(i);
7434 cpus_clear(*covered);
7436 cpus_and(*nodemask, *nodemask, *cpu_map);
7437 if (cpus_empty(*nodemask)) {
7438 sched_group_nodes[i] = NULL;
7442 sched_domain_node_span(i, domainspan);
7443 cpus_and(*domainspan, *domainspan, *cpu_map);
7445 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7447 printk(KERN_WARNING "Can not alloc domain group for "
7451 sched_group_nodes[i] = sg;
7452 for_each_cpu_mask(j, *nodemask) {
7453 struct sched_domain *sd;
7455 sd = &per_cpu(node_domains, j);
7458 sg->__cpu_power = 0;
7459 sg->cpumask = *nodemask;
7461 cpus_or(*covered, *covered, *nodemask);
7464 for (j = 0; j < MAX_NUMNODES; j++) {
7465 SCHED_CPUMASK_VAR(notcovered, allmasks);
7466 int n = (i + j) % MAX_NUMNODES;
7467 node_to_cpumask_ptr(pnodemask, n);
7469 cpus_complement(*notcovered, *covered);
7470 cpus_and(*tmpmask, *notcovered, *cpu_map);
7471 cpus_and(*tmpmask, *tmpmask, *domainspan);
7472 if (cpus_empty(*tmpmask))
7475 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7476 if (cpus_empty(*tmpmask))
7479 sg = kmalloc_node(sizeof(struct sched_group),
7483 "Can not alloc domain group for node %d\n", j);
7486 sg->__cpu_power = 0;
7487 sg->cpumask = *tmpmask;
7488 sg->next = prev->next;
7489 cpus_or(*covered, *covered, *tmpmask);
7496 /* Calculate CPU power for physical packages and nodes */
7497 #ifdef CONFIG_SCHED_SMT
7498 for_each_cpu_mask(i, *cpu_map) {
7499 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7501 init_sched_groups_power(i, sd);
7504 #ifdef CONFIG_SCHED_MC
7505 for_each_cpu_mask(i, *cpu_map) {
7506 struct sched_domain *sd = &per_cpu(core_domains, i);
7508 init_sched_groups_power(i, sd);
7512 for_each_cpu_mask(i, *cpu_map) {
7513 struct sched_domain *sd = &per_cpu(phys_domains, i);
7515 init_sched_groups_power(i, sd);
7519 for (i = 0; i < MAX_NUMNODES; i++)
7520 init_numa_sched_groups_power(sched_group_nodes[i]);
7523 struct sched_group *sg;
7525 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7527 init_numa_sched_groups_power(sg);
7531 /* Attach the domains */
7532 for_each_cpu_mask(i, *cpu_map) {
7533 struct sched_domain *sd;
7534 #ifdef CONFIG_SCHED_SMT
7535 sd = &per_cpu(cpu_domains, i);
7536 #elif defined(CONFIG_SCHED_MC)
7537 sd = &per_cpu(core_domains, i);
7539 sd = &per_cpu(phys_domains, i);
7541 cpu_attach_domain(sd, rd, i);
7544 SCHED_CPUMASK_FREE((void *)allmasks);
7549 free_sched_groups(cpu_map, tmpmask);
7550 SCHED_CPUMASK_FREE((void *)allmasks);
7555 static int build_sched_domains(const cpumask_t *cpu_map)
7557 return __build_sched_domains(cpu_map, NULL);
7560 static cpumask_t *doms_cur; /* current sched domains */
7561 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7562 static struct sched_domain_attr *dattr_cur;
7563 /* attribues of custom domains in 'doms_cur' */
7566 * Special case: If a kmalloc of a doms_cur partition (array of
7567 * cpumask_t) fails, then fallback to a single sched domain,
7568 * as determined by the single cpumask_t fallback_doms.
7570 static cpumask_t fallback_doms;
7572 void __attribute__((weak)) arch_update_cpu_topology(void)
7577 * Free current domain masks.
7578 * Called after all cpus are attached to NULL domain.
7580 static void free_sched_domains(void)
7583 if (doms_cur != &fallback_doms)
7585 doms_cur = &fallback_doms;
7589 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7590 * For now this just excludes isolated cpus, but could be used to
7591 * exclude other special cases in the future.
7593 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7597 arch_update_cpu_topology();
7599 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7601 doms_cur = &fallback_doms;
7602 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7604 err = build_sched_domains(doms_cur);
7605 register_sched_domain_sysctl();
7610 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7613 free_sched_groups(cpu_map, tmpmask);
7617 * Detach sched domains from a group of cpus specified in cpu_map
7618 * These cpus will now be attached to the NULL domain
7620 static void detach_destroy_domains(const cpumask_t *cpu_map)
7625 unregister_sched_domain_sysctl();
7627 for_each_cpu_mask(i, *cpu_map)
7628 cpu_attach_domain(NULL, &def_root_domain, i);
7629 synchronize_sched();
7630 arch_destroy_sched_domains(cpu_map, &tmpmask);
7633 /* handle null as "default" */
7634 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7635 struct sched_domain_attr *new, int idx_new)
7637 struct sched_domain_attr tmp;
7644 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7645 new ? (new + idx_new) : &tmp,
7646 sizeof(struct sched_domain_attr));
7650 * Partition sched domains as specified by the 'ndoms_new'
7651 * cpumasks in the array doms_new[] of cpumasks. This compares
7652 * doms_new[] to the current sched domain partitioning, doms_cur[].
7653 * It destroys each deleted domain and builds each new domain.
7655 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7656 * The masks don't intersect (don't overlap.) We should setup one
7657 * sched domain for each mask. CPUs not in any of the cpumasks will
7658 * not be load balanced. If the same cpumask appears both in the
7659 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7662 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7663 * ownership of it and will kfree it when done with it. If the caller
7664 * failed the kmalloc call, then it can pass in doms_new == NULL,
7665 * and partition_sched_domains() will fallback to the single partition
7668 * Call with hotplug lock held
7670 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7671 struct sched_domain_attr *dattr_new)
7675 mutex_lock(&sched_domains_mutex);
7677 /* always unregister in case we don't destroy any domains */
7678 unregister_sched_domain_sysctl();
7680 if (doms_new == NULL) {
7682 doms_new = &fallback_doms;
7683 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7687 /* Destroy deleted domains */
7688 for (i = 0; i < ndoms_cur; i++) {
7689 for (j = 0; j < ndoms_new; j++) {
7690 if (cpus_equal(doms_cur[i], doms_new[j])
7691 && dattrs_equal(dattr_cur, i, dattr_new, j))
7694 /* no match - a current sched domain not in new doms_new[] */
7695 detach_destroy_domains(doms_cur + i);
7700 /* Build new domains */
7701 for (i = 0; i < ndoms_new; i++) {
7702 for (j = 0; j < ndoms_cur; j++) {
7703 if (cpus_equal(doms_new[i], doms_cur[j])
7704 && dattrs_equal(dattr_new, i, dattr_cur, j))
7707 /* no match - add a new doms_new */
7708 __build_sched_domains(doms_new + i,
7709 dattr_new ? dattr_new + i : NULL);
7714 /* Remember the new sched domains */
7715 if (doms_cur != &fallback_doms)
7717 kfree(dattr_cur); /* kfree(NULL) is safe */
7718 doms_cur = doms_new;
7719 dattr_cur = dattr_new;
7720 ndoms_cur = ndoms_new;
7722 register_sched_domain_sysctl();
7724 mutex_unlock(&sched_domains_mutex);
7727 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7728 int arch_reinit_sched_domains(void)
7733 mutex_lock(&sched_domains_mutex);
7734 detach_destroy_domains(&cpu_online_map);
7735 free_sched_domains();
7736 err = arch_init_sched_domains(&cpu_online_map);
7737 mutex_unlock(&sched_domains_mutex);
7743 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7747 if (buf[0] != '0' && buf[0] != '1')
7751 sched_smt_power_savings = (buf[0] == '1');
7753 sched_mc_power_savings = (buf[0] == '1');
7755 ret = arch_reinit_sched_domains();
7757 return ret ? ret : count;
7760 #ifdef CONFIG_SCHED_MC
7761 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7763 return sprintf(page, "%u\n", sched_mc_power_savings);
7765 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7766 const char *buf, size_t count)
7768 return sched_power_savings_store(buf, count, 0);
7770 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7771 sched_mc_power_savings_store);
7774 #ifdef CONFIG_SCHED_SMT
7775 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7777 return sprintf(page, "%u\n", sched_smt_power_savings);
7779 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7780 const char *buf, size_t count)
7782 return sched_power_savings_store(buf, count, 1);
7784 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7785 sched_smt_power_savings_store);
7788 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7792 #ifdef CONFIG_SCHED_SMT
7794 err = sysfs_create_file(&cls->kset.kobj,
7795 &attr_sched_smt_power_savings.attr);
7797 #ifdef CONFIG_SCHED_MC
7798 if (!err && mc_capable())
7799 err = sysfs_create_file(&cls->kset.kobj,
7800 &attr_sched_mc_power_savings.attr);
7804 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7807 * Force a reinitialization of the sched domains hierarchy. The domains
7808 * and groups cannot be updated in place without racing with the balancing
7809 * code, so we temporarily attach all running cpus to the NULL domain
7810 * which will prevent rebalancing while the sched domains are recalculated.
7812 static int update_sched_domains(struct notifier_block *nfb,
7813 unsigned long action, void *hcpu)
7815 int cpu = (int)(long)hcpu;
7818 case CPU_DOWN_PREPARE:
7819 case CPU_DOWN_PREPARE_FROZEN:
7820 disable_runtime(cpu_rq(cpu));
7822 case CPU_UP_PREPARE:
7823 case CPU_UP_PREPARE_FROZEN:
7824 detach_destroy_domains(&cpu_online_map);
7825 free_sched_domains();
7829 case CPU_DOWN_FAILED:
7830 case CPU_DOWN_FAILED_FROZEN:
7832 case CPU_ONLINE_FROZEN:
7833 enable_runtime(cpu_rq(cpu));
7835 case CPU_UP_CANCELED:
7836 case CPU_UP_CANCELED_FROZEN:
7838 case CPU_DEAD_FROZEN:
7840 * Fall through and re-initialise the domains.
7847 #ifndef CONFIG_CPUSETS
7849 * Create default domain partitioning if cpusets are disabled.
7850 * Otherwise we let cpusets rebuild the domains based on the
7854 /* The hotplug lock is already held by cpu_up/cpu_down */
7855 arch_init_sched_domains(&cpu_online_map);
7861 void __init sched_init_smp(void)
7863 cpumask_t non_isolated_cpus;
7865 #if defined(CONFIG_NUMA)
7866 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7868 BUG_ON(sched_group_nodes_bycpu == NULL);
7871 mutex_lock(&sched_domains_mutex);
7872 arch_init_sched_domains(&cpu_online_map);
7873 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7874 if (cpus_empty(non_isolated_cpus))
7875 cpu_set(smp_processor_id(), non_isolated_cpus);
7876 mutex_unlock(&sched_domains_mutex);
7878 /* XXX: Theoretical race here - CPU may be hotplugged now */
7879 hotcpu_notifier(update_sched_domains, 0);
7882 /* Move init over to a non-isolated CPU */
7883 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7885 sched_init_granularity();
7888 void __init sched_init_smp(void)
7890 sched_init_granularity();
7892 #endif /* CONFIG_SMP */
7894 int in_sched_functions(unsigned long addr)
7896 return in_lock_functions(addr) ||
7897 (addr >= (unsigned long)__sched_text_start
7898 && addr < (unsigned long)__sched_text_end);
7901 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7903 cfs_rq->tasks_timeline = RB_ROOT;
7904 INIT_LIST_HEAD(&cfs_rq->tasks);
7905 #ifdef CONFIG_FAIR_GROUP_SCHED
7908 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7911 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7913 struct rt_prio_array *array;
7916 array = &rt_rq->active;
7917 for (i = 0; i < MAX_RT_PRIO; i++) {
7918 INIT_LIST_HEAD(array->queue + i);
7919 __clear_bit(i, array->bitmap);
7921 /* delimiter for bitsearch: */
7922 __set_bit(MAX_RT_PRIO, array->bitmap);
7924 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7925 rt_rq->highest_prio = MAX_RT_PRIO;
7928 rt_rq->rt_nr_migratory = 0;
7929 rt_rq->overloaded = 0;
7933 rt_rq->rt_throttled = 0;
7934 rt_rq->rt_runtime = 0;
7935 spin_lock_init(&rt_rq->rt_runtime_lock);
7937 #ifdef CONFIG_RT_GROUP_SCHED
7938 rt_rq->rt_nr_boosted = 0;
7943 #ifdef CONFIG_FAIR_GROUP_SCHED
7944 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7945 struct sched_entity *se, int cpu, int add,
7946 struct sched_entity *parent)
7948 struct rq *rq = cpu_rq(cpu);
7949 tg->cfs_rq[cpu] = cfs_rq;
7950 init_cfs_rq(cfs_rq, rq);
7953 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7956 /* se could be NULL for init_task_group */
7961 se->cfs_rq = &rq->cfs;
7963 se->cfs_rq = parent->my_q;
7966 se->load.weight = tg->shares;
7967 se->load.inv_weight = 0;
7968 se->parent = parent;
7972 #ifdef CONFIG_RT_GROUP_SCHED
7973 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7974 struct sched_rt_entity *rt_se, int cpu, int add,
7975 struct sched_rt_entity *parent)
7977 struct rq *rq = cpu_rq(cpu);
7979 tg->rt_rq[cpu] = rt_rq;
7980 init_rt_rq(rt_rq, rq);
7982 rt_rq->rt_se = rt_se;
7983 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7985 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7987 tg->rt_se[cpu] = rt_se;
7992 rt_se->rt_rq = &rq->rt;
7994 rt_se->rt_rq = parent->my_q;
7996 rt_se->my_q = rt_rq;
7997 rt_se->parent = parent;
7998 INIT_LIST_HEAD(&rt_se->run_list);
8002 void __init sched_init(void)
8005 unsigned long alloc_size = 0, ptr;
8007 #ifdef CONFIG_FAIR_GROUP_SCHED
8008 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8010 #ifdef CONFIG_RT_GROUP_SCHED
8011 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8013 #ifdef CONFIG_USER_SCHED
8017 * As sched_init() is called before page_alloc is setup,
8018 * we use alloc_bootmem().
8021 ptr = (unsigned long)alloc_bootmem(alloc_size);
8023 #ifdef CONFIG_FAIR_GROUP_SCHED
8024 init_task_group.se = (struct sched_entity **)ptr;
8025 ptr += nr_cpu_ids * sizeof(void **);
8027 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8028 ptr += nr_cpu_ids * sizeof(void **);
8030 #ifdef CONFIG_USER_SCHED
8031 root_task_group.se = (struct sched_entity **)ptr;
8032 ptr += nr_cpu_ids * sizeof(void **);
8034 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8035 ptr += nr_cpu_ids * sizeof(void **);
8036 #endif /* CONFIG_USER_SCHED */
8037 #endif /* CONFIG_FAIR_GROUP_SCHED */
8038 #ifdef CONFIG_RT_GROUP_SCHED
8039 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8040 ptr += nr_cpu_ids * sizeof(void **);
8042 init_task_group.rt_rq = (struct rt_rq **)ptr;
8043 ptr += nr_cpu_ids * sizeof(void **);
8045 #ifdef CONFIG_USER_SCHED
8046 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8047 ptr += nr_cpu_ids * sizeof(void **);
8049 root_task_group.rt_rq = (struct rt_rq **)ptr;
8050 ptr += nr_cpu_ids * sizeof(void **);
8051 #endif /* CONFIG_USER_SCHED */
8052 #endif /* CONFIG_RT_GROUP_SCHED */
8057 init_defrootdomain();
8060 init_rt_bandwidth(&def_rt_bandwidth,
8061 global_rt_period(), global_rt_runtime());
8063 #ifdef CONFIG_RT_GROUP_SCHED
8064 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8065 global_rt_period(), global_rt_runtime());
8066 #ifdef CONFIG_USER_SCHED
8067 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8068 global_rt_period(), RUNTIME_INF);
8069 #endif /* CONFIG_USER_SCHED */
8070 #endif /* CONFIG_RT_GROUP_SCHED */
8072 #ifdef CONFIG_GROUP_SCHED
8073 list_add(&init_task_group.list, &task_groups);
8074 INIT_LIST_HEAD(&init_task_group.children);
8076 #ifdef CONFIG_USER_SCHED
8077 INIT_LIST_HEAD(&root_task_group.children);
8078 init_task_group.parent = &root_task_group;
8079 list_add(&init_task_group.siblings, &root_task_group.children);
8080 #endif /* CONFIG_USER_SCHED */
8081 #endif /* CONFIG_GROUP_SCHED */
8083 for_each_possible_cpu(i) {
8087 spin_lock_init(&rq->lock);
8088 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
8090 init_cfs_rq(&rq->cfs, rq);
8091 init_rt_rq(&rq->rt, rq);
8092 #ifdef CONFIG_FAIR_GROUP_SCHED
8093 init_task_group.shares = init_task_group_load;
8094 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8095 #ifdef CONFIG_CGROUP_SCHED
8097 * How much cpu bandwidth does init_task_group get?
8099 * In case of task-groups formed thr' the cgroup filesystem, it
8100 * gets 100% of the cpu resources in the system. This overall
8101 * system cpu resource is divided among the tasks of
8102 * init_task_group and its child task-groups in a fair manner,
8103 * based on each entity's (task or task-group's) weight
8104 * (se->load.weight).
8106 * In other words, if init_task_group has 10 tasks of weight
8107 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8108 * then A0's share of the cpu resource is:
8110 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8112 * We achieve this by letting init_task_group's tasks sit
8113 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8115 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8116 #elif defined CONFIG_USER_SCHED
8117 root_task_group.shares = NICE_0_LOAD;
8118 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8120 * In case of task-groups formed thr' the user id of tasks,
8121 * init_task_group represents tasks belonging to root user.
8122 * Hence it forms a sibling of all subsequent groups formed.
8123 * In this case, init_task_group gets only a fraction of overall
8124 * system cpu resource, based on the weight assigned to root
8125 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8126 * by letting tasks of init_task_group sit in a separate cfs_rq
8127 * (init_cfs_rq) and having one entity represent this group of
8128 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8130 init_tg_cfs_entry(&init_task_group,
8131 &per_cpu(init_cfs_rq, i),
8132 &per_cpu(init_sched_entity, i), i, 1,
8133 root_task_group.se[i]);
8136 #endif /* CONFIG_FAIR_GROUP_SCHED */
8138 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8139 #ifdef CONFIG_RT_GROUP_SCHED
8140 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8141 #ifdef CONFIG_CGROUP_SCHED
8142 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8143 #elif defined CONFIG_USER_SCHED
8144 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8145 init_tg_rt_entry(&init_task_group,
8146 &per_cpu(init_rt_rq, i),
8147 &per_cpu(init_sched_rt_entity, i), i, 1,
8148 root_task_group.rt_se[i]);
8152 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8153 rq->cpu_load[j] = 0;
8157 rq->active_balance = 0;
8158 rq->next_balance = jiffies;
8162 rq->migration_thread = NULL;
8163 INIT_LIST_HEAD(&rq->migration_queue);
8164 rq_attach_root(rq, &def_root_domain);
8167 atomic_set(&rq->nr_iowait, 0);
8170 set_load_weight(&init_task);
8172 #ifdef CONFIG_PREEMPT_NOTIFIERS
8173 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8177 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8180 #ifdef CONFIG_RT_MUTEXES
8181 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8185 * The boot idle thread does lazy MMU switching as well:
8187 atomic_inc(&init_mm.mm_count);
8188 enter_lazy_tlb(&init_mm, current);
8191 * Make us the idle thread. Technically, schedule() should not be
8192 * called from this thread, however somewhere below it might be,
8193 * but because we are the idle thread, we just pick up running again
8194 * when this runqueue becomes "idle".
8196 init_idle(current, smp_processor_id());
8198 * During early bootup we pretend to be a normal task:
8200 current->sched_class = &fair_sched_class;
8202 scheduler_running = 1;
8205 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8206 void __might_sleep(char *file, int line)
8209 static unsigned long prev_jiffy; /* ratelimiting */
8211 if ((in_atomic() || irqs_disabled()) &&
8212 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8213 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8215 prev_jiffy = jiffies;
8216 printk(KERN_ERR "BUG: sleeping function called from invalid"
8217 " context at %s:%d\n", file, line);
8218 printk("in_atomic():%d, irqs_disabled():%d\n",
8219 in_atomic(), irqs_disabled());
8220 debug_show_held_locks(current);
8221 if (irqs_disabled())
8222 print_irqtrace_events(current);
8227 EXPORT_SYMBOL(__might_sleep);
8230 #ifdef CONFIG_MAGIC_SYSRQ
8231 static void normalize_task(struct rq *rq, struct task_struct *p)
8235 update_rq_clock(rq);
8236 on_rq = p->se.on_rq;
8238 deactivate_task(rq, p, 0);
8239 __setscheduler(rq, p, SCHED_NORMAL, 0);
8241 activate_task(rq, p, 0);
8242 resched_task(rq->curr);
8246 void normalize_rt_tasks(void)
8248 struct task_struct *g, *p;
8249 unsigned long flags;
8252 read_lock_irqsave(&tasklist_lock, flags);
8253 do_each_thread(g, p) {
8255 * Only normalize user tasks:
8260 p->se.exec_start = 0;
8261 #ifdef CONFIG_SCHEDSTATS
8262 p->se.wait_start = 0;
8263 p->se.sleep_start = 0;
8264 p->se.block_start = 0;
8269 * Renice negative nice level userspace
8272 if (TASK_NICE(p) < 0 && p->mm)
8273 set_user_nice(p, 0);
8277 spin_lock(&p->pi_lock);
8278 rq = __task_rq_lock(p);
8280 normalize_task(rq, p);
8282 __task_rq_unlock(rq);
8283 spin_unlock(&p->pi_lock);
8284 } while_each_thread(g, p);
8286 read_unlock_irqrestore(&tasklist_lock, flags);
8289 #endif /* CONFIG_MAGIC_SYSRQ */
8293 * These functions are only useful for the IA64 MCA handling.
8295 * They can only be called when the whole system has been
8296 * stopped - every CPU needs to be quiescent, and no scheduling
8297 * activity can take place. Using them for anything else would
8298 * be a serious bug, and as a result, they aren't even visible
8299 * under any other configuration.
8303 * curr_task - return the current task for a given cpu.
8304 * @cpu: the processor in question.
8306 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8308 struct task_struct *curr_task(int cpu)
8310 return cpu_curr(cpu);
8314 * set_curr_task - set the current task for a given cpu.
8315 * @cpu: the processor in question.
8316 * @p: the task pointer to set.
8318 * Description: This function must only be used when non-maskable interrupts
8319 * are serviced on a separate stack. It allows the architecture to switch the
8320 * notion of the current task on a cpu in a non-blocking manner. This function
8321 * must be called with all CPU's synchronized, and interrupts disabled, the
8322 * and caller must save the original value of the current task (see
8323 * curr_task() above) and restore that value before reenabling interrupts and
8324 * re-starting the system.
8326 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8328 void set_curr_task(int cpu, struct task_struct *p)
8335 #ifdef CONFIG_FAIR_GROUP_SCHED
8336 static void free_fair_sched_group(struct task_group *tg)
8340 for_each_possible_cpu(i) {
8342 kfree(tg->cfs_rq[i]);
8352 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8354 struct cfs_rq *cfs_rq;
8355 struct sched_entity *se, *parent_se;
8359 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8362 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8366 tg->shares = NICE_0_LOAD;
8368 for_each_possible_cpu(i) {
8371 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8372 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8376 se = kmalloc_node(sizeof(struct sched_entity),
8377 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8381 parent_se = parent ? parent->se[i] : NULL;
8382 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8391 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8393 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8394 &cpu_rq(cpu)->leaf_cfs_rq_list);
8397 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8399 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8401 #else /* !CONFG_FAIR_GROUP_SCHED */
8402 static inline void free_fair_sched_group(struct task_group *tg)
8407 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8412 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8416 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8419 #endif /* CONFIG_FAIR_GROUP_SCHED */
8421 #ifdef CONFIG_RT_GROUP_SCHED
8422 static void free_rt_sched_group(struct task_group *tg)
8426 destroy_rt_bandwidth(&tg->rt_bandwidth);
8428 for_each_possible_cpu(i) {
8430 kfree(tg->rt_rq[i]);
8432 kfree(tg->rt_se[i]);
8440 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8442 struct rt_rq *rt_rq;
8443 struct sched_rt_entity *rt_se, *parent_se;
8447 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8450 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8454 init_rt_bandwidth(&tg->rt_bandwidth,
8455 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8457 for_each_possible_cpu(i) {
8460 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8461 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8465 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8466 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8470 parent_se = parent ? parent->rt_se[i] : NULL;
8471 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8480 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8482 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8483 &cpu_rq(cpu)->leaf_rt_rq_list);
8486 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8488 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8490 #else /* !CONFIG_RT_GROUP_SCHED */
8491 static inline void free_rt_sched_group(struct task_group *tg)
8496 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8501 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8505 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8508 #endif /* CONFIG_RT_GROUP_SCHED */
8510 #ifdef CONFIG_GROUP_SCHED
8511 static void free_sched_group(struct task_group *tg)
8513 free_fair_sched_group(tg);
8514 free_rt_sched_group(tg);
8518 /* allocate runqueue etc for a new task group */
8519 struct task_group *sched_create_group(struct task_group *parent)
8521 struct task_group *tg;
8522 unsigned long flags;
8525 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8527 return ERR_PTR(-ENOMEM);
8529 if (!alloc_fair_sched_group(tg, parent))
8532 if (!alloc_rt_sched_group(tg, parent))
8535 spin_lock_irqsave(&task_group_lock, flags);
8536 for_each_possible_cpu(i) {
8537 register_fair_sched_group(tg, i);
8538 register_rt_sched_group(tg, i);
8540 list_add_rcu(&tg->list, &task_groups);
8542 WARN_ON(!parent); /* root should already exist */
8544 tg->parent = parent;
8545 list_add_rcu(&tg->siblings, &parent->children);
8546 INIT_LIST_HEAD(&tg->children);
8547 spin_unlock_irqrestore(&task_group_lock, flags);
8552 free_sched_group(tg);
8553 return ERR_PTR(-ENOMEM);
8556 /* rcu callback to free various structures associated with a task group */
8557 static void free_sched_group_rcu(struct rcu_head *rhp)
8559 /* now it should be safe to free those cfs_rqs */
8560 free_sched_group(container_of(rhp, struct task_group, rcu));
8563 /* Destroy runqueue etc associated with a task group */
8564 void sched_destroy_group(struct task_group *tg)
8566 unsigned long flags;
8569 spin_lock_irqsave(&task_group_lock, flags);
8570 for_each_possible_cpu(i) {
8571 unregister_fair_sched_group(tg, i);
8572 unregister_rt_sched_group(tg, i);
8574 list_del_rcu(&tg->list);
8575 list_del_rcu(&tg->siblings);
8576 spin_unlock_irqrestore(&task_group_lock, flags);
8578 /* wait for possible concurrent references to cfs_rqs complete */
8579 call_rcu(&tg->rcu, free_sched_group_rcu);
8582 /* change task's runqueue when it moves between groups.
8583 * The caller of this function should have put the task in its new group
8584 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8585 * reflect its new group.
8587 void sched_move_task(struct task_struct *tsk)
8590 unsigned long flags;
8593 rq = task_rq_lock(tsk, &flags);
8595 update_rq_clock(rq);
8597 running = task_current(rq, tsk);
8598 on_rq = tsk->se.on_rq;
8601 dequeue_task(rq, tsk, 0);
8602 if (unlikely(running))
8603 tsk->sched_class->put_prev_task(rq, tsk);
8605 set_task_rq(tsk, task_cpu(tsk));
8607 #ifdef CONFIG_FAIR_GROUP_SCHED
8608 if (tsk->sched_class->moved_group)
8609 tsk->sched_class->moved_group(tsk);
8612 if (unlikely(running))
8613 tsk->sched_class->set_curr_task(rq);
8615 enqueue_task(rq, tsk, 0);
8617 task_rq_unlock(rq, &flags);
8619 #endif /* CONFIG_GROUP_SCHED */
8621 #ifdef CONFIG_FAIR_GROUP_SCHED
8622 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8624 struct cfs_rq *cfs_rq = se->cfs_rq;
8629 dequeue_entity(cfs_rq, se, 0);
8631 se->load.weight = shares;
8632 se->load.inv_weight = 0;
8635 enqueue_entity(cfs_rq, se, 0);
8638 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8640 struct cfs_rq *cfs_rq = se->cfs_rq;
8641 struct rq *rq = cfs_rq->rq;
8642 unsigned long flags;
8644 spin_lock_irqsave(&rq->lock, flags);
8645 __set_se_shares(se, shares);
8646 spin_unlock_irqrestore(&rq->lock, flags);
8649 static DEFINE_MUTEX(shares_mutex);
8651 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8654 unsigned long flags;
8657 * We can't change the weight of the root cgroup.
8662 if (shares < MIN_SHARES)
8663 shares = MIN_SHARES;
8664 else if (shares > MAX_SHARES)
8665 shares = MAX_SHARES;
8667 mutex_lock(&shares_mutex);
8668 if (tg->shares == shares)
8671 spin_lock_irqsave(&task_group_lock, flags);
8672 for_each_possible_cpu(i)
8673 unregister_fair_sched_group(tg, i);
8674 list_del_rcu(&tg->siblings);
8675 spin_unlock_irqrestore(&task_group_lock, flags);
8677 /* wait for any ongoing reference to this group to finish */
8678 synchronize_sched();
8681 * Now we are free to modify the group's share on each cpu
8682 * w/o tripping rebalance_share or load_balance_fair.
8684 tg->shares = shares;
8685 for_each_possible_cpu(i) {
8689 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8690 set_se_shares(tg->se[i], shares);
8694 * Enable load balance activity on this group, by inserting it back on
8695 * each cpu's rq->leaf_cfs_rq_list.
8697 spin_lock_irqsave(&task_group_lock, flags);
8698 for_each_possible_cpu(i)
8699 register_fair_sched_group(tg, i);
8700 list_add_rcu(&tg->siblings, &tg->parent->children);
8701 spin_unlock_irqrestore(&task_group_lock, flags);
8703 mutex_unlock(&shares_mutex);
8707 unsigned long sched_group_shares(struct task_group *tg)
8713 #ifdef CONFIG_RT_GROUP_SCHED
8715 * Ensure that the real time constraints are schedulable.
8717 static DEFINE_MUTEX(rt_constraints_mutex);
8719 static unsigned long to_ratio(u64 period, u64 runtime)
8721 if (runtime == RUNTIME_INF)
8724 return div64_u64(runtime << 16, period);
8727 #ifdef CONFIG_CGROUP_SCHED
8728 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8730 struct task_group *tgi, *parent = tg->parent;
8731 unsigned long total = 0;
8734 if (global_rt_period() < period)
8737 return to_ratio(period, runtime) <
8738 to_ratio(global_rt_period(), global_rt_runtime());
8741 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8745 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8749 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8750 tgi->rt_bandwidth.rt_runtime);
8754 return total + to_ratio(period, runtime) <=
8755 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8756 parent->rt_bandwidth.rt_runtime);
8758 #elif defined CONFIG_USER_SCHED
8759 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8761 struct task_group *tgi;
8762 unsigned long total = 0;
8763 unsigned long global_ratio =
8764 to_ratio(global_rt_period(), global_rt_runtime());
8767 list_for_each_entry_rcu(tgi, &task_groups, list) {
8771 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8772 tgi->rt_bandwidth.rt_runtime);
8776 return total + to_ratio(period, runtime) < global_ratio;
8780 /* Must be called with tasklist_lock held */
8781 static inline int tg_has_rt_tasks(struct task_group *tg)
8783 struct task_struct *g, *p;
8784 do_each_thread(g, p) {
8785 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8787 } while_each_thread(g, p);
8791 static int tg_set_bandwidth(struct task_group *tg,
8792 u64 rt_period, u64 rt_runtime)
8796 mutex_lock(&rt_constraints_mutex);
8797 read_lock(&tasklist_lock);
8798 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8802 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8807 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8808 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8809 tg->rt_bandwidth.rt_runtime = rt_runtime;
8811 for_each_possible_cpu(i) {
8812 struct rt_rq *rt_rq = tg->rt_rq[i];
8814 spin_lock(&rt_rq->rt_runtime_lock);
8815 rt_rq->rt_runtime = rt_runtime;
8816 spin_unlock(&rt_rq->rt_runtime_lock);
8818 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8820 read_unlock(&tasklist_lock);
8821 mutex_unlock(&rt_constraints_mutex);
8826 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8828 u64 rt_runtime, rt_period;
8830 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8831 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8832 if (rt_runtime_us < 0)
8833 rt_runtime = RUNTIME_INF;
8835 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8838 long sched_group_rt_runtime(struct task_group *tg)
8842 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8845 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8846 do_div(rt_runtime_us, NSEC_PER_USEC);
8847 return rt_runtime_us;
8850 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8852 u64 rt_runtime, rt_period;
8854 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8855 rt_runtime = tg->rt_bandwidth.rt_runtime;
8857 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8860 long sched_group_rt_period(struct task_group *tg)
8864 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8865 do_div(rt_period_us, NSEC_PER_USEC);
8866 return rt_period_us;
8869 static int sched_rt_global_constraints(void)
8871 struct task_group *tg = &root_task_group;
8872 u64 rt_runtime, rt_period;
8875 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8876 rt_runtime = tg->rt_bandwidth.rt_runtime;
8878 mutex_lock(&rt_constraints_mutex);
8879 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8881 mutex_unlock(&rt_constraints_mutex);
8885 #else /* !CONFIG_RT_GROUP_SCHED */
8886 static int sched_rt_global_constraints(void)
8888 unsigned long flags;
8891 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8892 for_each_possible_cpu(i) {
8893 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8895 spin_lock(&rt_rq->rt_runtime_lock);
8896 rt_rq->rt_runtime = global_rt_runtime();
8897 spin_unlock(&rt_rq->rt_runtime_lock);
8899 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8903 #endif /* CONFIG_RT_GROUP_SCHED */
8905 int sched_rt_handler(struct ctl_table *table, int write,
8906 struct file *filp, void __user *buffer, size_t *lenp,
8910 int old_period, old_runtime;
8911 static DEFINE_MUTEX(mutex);
8914 old_period = sysctl_sched_rt_period;
8915 old_runtime = sysctl_sched_rt_runtime;
8917 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8919 if (!ret && write) {
8920 ret = sched_rt_global_constraints();
8922 sysctl_sched_rt_period = old_period;
8923 sysctl_sched_rt_runtime = old_runtime;
8925 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8926 def_rt_bandwidth.rt_period =
8927 ns_to_ktime(global_rt_period());
8930 mutex_unlock(&mutex);
8935 #ifdef CONFIG_CGROUP_SCHED
8937 /* return corresponding task_group object of a cgroup */
8938 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8940 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8941 struct task_group, css);
8944 static struct cgroup_subsys_state *
8945 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8947 struct task_group *tg, *parent;
8949 if (!cgrp->parent) {
8950 /* This is early initialization for the top cgroup */
8951 init_task_group.css.cgroup = cgrp;
8952 return &init_task_group.css;
8955 parent = cgroup_tg(cgrp->parent);
8956 tg = sched_create_group(parent);
8958 return ERR_PTR(-ENOMEM);
8960 /* Bind the cgroup to task_group object we just created */
8961 tg->css.cgroup = cgrp;
8967 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8969 struct task_group *tg = cgroup_tg(cgrp);
8971 sched_destroy_group(tg);
8975 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8976 struct task_struct *tsk)
8978 #ifdef CONFIG_RT_GROUP_SCHED
8979 /* Don't accept realtime tasks when there is no way for them to run */
8980 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
8983 /* We don't support RT-tasks being in separate groups */
8984 if (tsk->sched_class != &fair_sched_class)
8992 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8993 struct cgroup *old_cont, struct task_struct *tsk)
8995 sched_move_task(tsk);
8998 #ifdef CONFIG_FAIR_GROUP_SCHED
8999 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9002 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9005 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9007 struct task_group *tg = cgroup_tg(cgrp);
9009 return (u64) tg->shares;
9011 #endif /* CONFIG_FAIR_GROUP_SCHED */
9013 #ifdef CONFIG_RT_GROUP_SCHED
9014 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9017 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9020 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9022 return sched_group_rt_runtime(cgroup_tg(cgrp));
9025 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9028 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9031 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9033 return sched_group_rt_period(cgroup_tg(cgrp));
9035 #endif /* CONFIG_RT_GROUP_SCHED */
9037 static struct cftype cpu_files[] = {
9038 #ifdef CONFIG_FAIR_GROUP_SCHED
9041 .read_u64 = cpu_shares_read_u64,
9042 .write_u64 = cpu_shares_write_u64,
9045 #ifdef CONFIG_RT_GROUP_SCHED
9047 .name = "rt_runtime_us",
9048 .read_s64 = cpu_rt_runtime_read,
9049 .write_s64 = cpu_rt_runtime_write,
9052 .name = "rt_period_us",
9053 .read_u64 = cpu_rt_period_read_uint,
9054 .write_u64 = cpu_rt_period_write_uint,
9059 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9061 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9064 struct cgroup_subsys cpu_cgroup_subsys = {
9066 .create = cpu_cgroup_create,
9067 .destroy = cpu_cgroup_destroy,
9068 .can_attach = cpu_cgroup_can_attach,
9069 .attach = cpu_cgroup_attach,
9070 .populate = cpu_cgroup_populate,
9071 .subsys_id = cpu_cgroup_subsys_id,
9075 #endif /* CONFIG_CGROUP_SCHED */
9077 #ifdef CONFIG_CGROUP_CPUACCT
9080 * CPU accounting code for task groups.
9082 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9083 * (balbir@in.ibm.com).
9086 /* track cpu usage of a group of tasks */
9088 struct cgroup_subsys_state css;
9089 /* cpuusage holds pointer to a u64-type object on every cpu */
9093 struct cgroup_subsys cpuacct_subsys;
9095 /* return cpu accounting group corresponding to this container */
9096 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9098 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9099 struct cpuacct, css);
9102 /* return cpu accounting group to which this task belongs */
9103 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9105 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9106 struct cpuacct, css);
9109 /* create a new cpu accounting group */
9110 static struct cgroup_subsys_state *cpuacct_create(
9111 struct cgroup_subsys *ss, struct cgroup *cgrp)
9113 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9116 return ERR_PTR(-ENOMEM);
9118 ca->cpuusage = alloc_percpu(u64);
9119 if (!ca->cpuusage) {
9121 return ERR_PTR(-ENOMEM);
9127 /* destroy an existing cpu accounting group */
9129 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9131 struct cpuacct *ca = cgroup_ca(cgrp);
9133 free_percpu(ca->cpuusage);
9137 /* return total cpu usage (in nanoseconds) of a group */
9138 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9140 struct cpuacct *ca = cgroup_ca(cgrp);
9141 u64 totalcpuusage = 0;
9144 for_each_possible_cpu(i) {
9145 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9148 * Take rq->lock to make 64-bit addition safe on 32-bit
9151 spin_lock_irq(&cpu_rq(i)->lock);
9152 totalcpuusage += *cpuusage;
9153 spin_unlock_irq(&cpu_rq(i)->lock);
9156 return totalcpuusage;
9159 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9162 struct cpuacct *ca = cgroup_ca(cgrp);
9171 for_each_possible_cpu(i) {
9172 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9174 spin_lock_irq(&cpu_rq(i)->lock);
9176 spin_unlock_irq(&cpu_rq(i)->lock);
9182 static struct cftype files[] = {
9185 .read_u64 = cpuusage_read,
9186 .write_u64 = cpuusage_write,
9190 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9192 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9196 * charge this task's execution time to its accounting group.
9198 * called with rq->lock held.
9200 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9204 if (!cpuacct_subsys.active)
9209 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9211 *cpuusage += cputime;
9215 struct cgroup_subsys cpuacct_subsys = {
9217 .create = cpuacct_create,
9218 .destroy = cpuacct_destroy,
9219 .populate = cpuacct_populate,
9220 .subsys_id = cpuacct_subsys_id,
9222 #endif /* CONFIG_CGROUP_CPUACCT */