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/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
134 return reciprocal_divide(load, sg->reciprocal_cpu_power);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
143 sg->__cpu_power += val;
144 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
148 static inline int rt_policy(int policy)
150 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
155 static inline int task_has_rt_policy(struct task_struct *p)
157 return rt_policy(p->policy);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array {
164 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
165 struct list_head queue[MAX_RT_PRIO];
168 struct rt_bandwidth {
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock;
173 struct hrtimer rt_period_timer;
176 static struct rt_bandwidth def_rt_bandwidth;
178 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
180 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
182 struct rt_bandwidth *rt_b =
183 container_of(timer, struct rt_bandwidth, rt_period_timer);
189 now = hrtimer_cb_get_time(timer);
190 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
195 idle = do_sched_rt_period_timer(rt_b, overrun);
198 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
202 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
204 rt_b->rt_period = ns_to_ktime(period);
205 rt_b->rt_runtime = runtime;
207 spin_lock_init(&rt_b->rt_runtime_lock);
209 hrtimer_init(&rt_b->rt_period_timer,
210 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
211 rt_b->rt_period_timer.function = sched_rt_period_timer;
212 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
215 static inline int rt_bandwidth_enabled(void)
217 return sysctl_sched_rt_runtime >= 0;
220 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
224 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
227 if (hrtimer_active(&rt_b->rt_period_timer))
230 spin_lock(&rt_b->rt_runtime_lock);
232 if (hrtimer_active(&rt_b->rt_period_timer))
235 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
236 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
237 hrtimer_start_expires(&rt_b->rt_period_timer,
240 spin_unlock(&rt_b->rt_runtime_lock);
243 #ifdef CONFIG_RT_GROUP_SCHED
244 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
246 hrtimer_cancel(&rt_b->rt_period_timer);
251 * sched_domains_mutex serializes calls to arch_init_sched_domains,
252 * detach_destroy_domains and partition_sched_domains.
254 static DEFINE_MUTEX(sched_domains_mutex);
256 #ifdef CONFIG_GROUP_SCHED
258 #include <linux/cgroup.h>
262 static LIST_HEAD(task_groups);
264 /* task group related information */
266 #ifdef CONFIG_CGROUP_SCHED
267 struct cgroup_subsys_state css;
270 #ifdef CONFIG_FAIR_GROUP_SCHED
271 /* schedulable entities of this group on each cpu */
272 struct sched_entity **se;
273 /* runqueue "owned" by this group on each cpu */
274 struct cfs_rq **cfs_rq;
275 unsigned long shares;
278 #ifdef CONFIG_RT_GROUP_SCHED
279 struct sched_rt_entity **rt_se;
280 struct rt_rq **rt_rq;
282 struct rt_bandwidth rt_bandwidth;
286 struct list_head list;
288 struct task_group *parent;
289 struct list_head siblings;
290 struct list_head children;
293 #ifdef CONFIG_USER_SCHED
297 * Every UID task group (including init_task_group aka UID-0) will
298 * be a child to this group.
300 struct task_group root_task_group;
302 #ifdef CONFIG_FAIR_GROUP_SCHED
303 /* Default task group's sched entity on each cpu */
304 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
305 /* Default task group's cfs_rq on each cpu */
306 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
307 #endif /* CONFIG_FAIR_GROUP_SCHED */
309 #ifdef CONFIG_RT_GROUP_SCHED
310 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
311 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
312 #endif /* CONFIG_RT_GROUP_SCHED */
313 #else /* !CONFIG_USER_SCHED */
314 #define root_task_group init_task_group
315 #endif /* CONFIG_USER_SCHED */
317 /* task_group_lock serializes add/remove of task groups and also changes to
318 * a task group's cpu shares.
320 static DEFINE_SPINLOCK(task_group_lock);
322 #ifdef CONFIG_FAIR_GROUP_SCHED
323 #ifdef CONFIG_USER_SCHED
324 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
325 #else /* !CONFIG_USER_SCHED */
326 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
327 #endif /* CONFIG_USER_SCHED */
330 * A weight of 0 or 1 can cause arithmetics problems.
331 * A weight of a cfs_rq is the sum of weights of which entities
332 * are queued on this cfs_rq, so a weight of a entity should not be
333 * too large, so as the shares value of a task group.
334 * (The default weight is 1024 - so there's no practical
335 * limitation from this.)
338 #define MAX_SHARES (1UL << 18)
340 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
343 /* Default task group.
344 * Every task in system belong to this group at bootup.
346 struct task_group init_task_group;
348 /* return group to which a task belongs */
349 static inline struct task_group *task_group(struct task_struct *p)
351 struct task_group *tg;
353 #ifdef CONFIG_USER_SCHED
355 tg = __task_cred(p)->user->tg;
357 #elif defined(CONFIG_CGROUP_SCHED)
358 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
359 struct task_group, css);
361 tg = &init_task_group;
366 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
367 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
369 #ifdef CONFIG_FAIR_GROUP_SCHED
370 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
371 p->se.parent = task_group(p)->se[cpu];
374 #ifdef CONFIG_RT_GROUP_SCHED
375 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
376 p->rt.parent = task_group(p)->rt_se[cpu];
382 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
383 static inline struct task_group *task_group(struct task_struct *p)
388 #endif /* CONFIG_GROUP_SCHED */
390 /* CFS-related fields in a runqueue */
392 struct load_weight load;
393 unsigned long nr_running;
398 struct rb_root tasks_timeline;
399 struct rb_node *rb_leftmost;
401 struct list_head tasks;
402 struct list_head *balance_iterator;
405 * 'curr' points to currently running entity on this cfs_rq.
406 * It is set to NULL otherwise (i.e when none are currently running).
408 struct sched_entity *curr, *next, *last;
410 unsigned int nr_spread_over;
412 #ifdef CONFIG_FAIR_GROUP_SCHED
413 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
416 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
417 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
418 * (like users, containers etc.)
420 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
421 * list is used during load balance.
423 struct list_head leaf_cfs_rq_list;
424 struct task_group *tg; /* group that "owns" this runqueue */
428 * the part of load.weight contributed by tasks
430 unsigned long task_weight;
433 * h_load = weight * f(tg)
435 * Where f(tg) is the recursive weight fraction assigned to
438 unsigned long h_load;
441 * this cpu's part of tg->shares
443 unsigned long shares;
446 * load.weight at the time we set shares
448 unsigned long rq_weight;
453 /* Real-Time classes' related field in a runqueue: */
455 struct rt_prio_array active;
456 unsigned long rt_nr_running;
457 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
458 int highest_prio; /* highest queued rt task prio */
461 unsigned long rt_nr_migratory;
467 /* Nests inside the rq lock: */
468 spinlock_t rt_runtime_lock;
470 #ifdef CONFIG_RT_GROUP_SCHED
471 unsigned long rt_nr_boosted;
474 struct list_head leaf_rt_rq_list;
475 struct task_group *tg;
476 struct sched_rt_entity *rt_se;
483 * We add the notion of a root-domain which will be used to define per-domain
484 * variables. Each exclusive cpuset essentially defines an island domain by
485 * fully partitioning the member cpus from any other cpuset. Whenever a new
486 * exclusive cpuset is created, we also create and attach a new root-domain
496 * The "RT overload" flag: it gets set if a CPU has more than
497 * one runnable RT task.
502 struct cpupri cpupri;
507 * By default the system creates a single root-domain with all cpus as
508 * members (mimicking the global state we have today).
510 static struct root_domain def_root_domain;
515 * This is the main, per-CPU runqueue data structure.
517 * Locking rule: those places that want to lock multiple runqueues
518 * (such as the load balancing or the thread migration code), lock
519 * acquire operations must be ordered by ascending &runqueue.
526 * nr_running and cpu_load should be in the same cacheline because
527 * remote CPUs use both these fields when doing load calculation.
529 unsigned long nr_running;
530 #define CPU_LOAD_IDX_MAX 5
531 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
532 unsigned char idle_at_tick;
534 unsigned long last_tick_seen;
535 unsigned char in_nohz_recently;
537 /* capture load from *all* tasks on this cpu: */
538 struct load_weight load;
539 unsigned long nr_load_updates;
545 #ifdef CONFIG_FAIR_GROUP_SCHED
546 /* list of leaf cfs_rq on this cpu: */
547 struct list_head leaf_cfs_rq_list;
549 #ifdef CONFIG_RT_GROUP_SCHED
550 struct list_head leaf_rt_rq_list;
554 * This is part of a global counter where only the total sum
555 * over all CPUs matters. A task can increase this counter on
556 * one CPU and if it got migrated afterwards it may decrease
557 * it on another CPU. Always updated under the runqueue lock:
559 unsigned long nr_uninterruptible;
561 struct task_struct *curr, *idle;
562 unsigned long next_balance;
563 struct mm_struct *prev_mm;
570 struct root_domain *rd;
571 struct sched_domain *sd;
573 /* For active balancing */
576 /* cpu of this runqueue: */
580 unsigned long avg_load_per_task;
582 struct task_struct *migration_thread;
583 struct list_head migration_queue;
586 #ifdef CONFIG_SCHED_HRTICK
588 int hrtick_csd_pending;
589 struct call_single_data hrtick_csd;
591 struct hrtimer hrtick_timer;
594 #ifdef CONFIG_SCHEDSTATS
596 struct sched_info rq_sched_info;
598 /* sys_sched_yield() stats */
599 unsigned int yld_exp_empty;
600 unsigned int yld_act_empty;
601 unsigned int yld_both_empty;
602 unsigned int yld_count;
604 /* schedule() stats */
605 unsigned int sched_switch;
606 unsigned int sched_count;
607 unsigned int sched_goidle;
609 /* try_to_wake_up() stats */
610 unsigned int ttwu_count;
611 unsigned int ttwu_local;
614 unsigned int bkl_count;
618 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
620 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
622 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
625 static inline int cpu_of(struct rq *rq)
635 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
636 * See detach_destroy_domains: synchronize_sched for details.
638 * The domain tree of any CPU may only be accessed from within
639 * preempt-disabled sections.
641 #define for_each_domain(cpu, __sd) \
642 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
644 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
645 #define this_rq() (&__get_cpu_var(runqueues))
646 #define task_rq(p) cpu_rq(task_cpu(p))
647 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
649 static inline void update_rq_clock(struct rq *rq)
651 rq->clock = sched_clock_cpu(cpu_of(rq));
655 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
657 #ifdef CONFIG_SCHED_DEBUG
658 # define const_debug __read_mostly
660 # define const_debug static const
666 * Returns true if the current cpu runqueue is locked.
667 * This interface allows printk to be called with the runqueue lock
668 * held and know whether or not it is OK to wake up the klogd.
670 int runqueue_is_locked(void)
673 struct rq *rq = cpu_rq(cpu);
676 ret = spin_is_locked(&rq->lock);
682 * Debugging: various feature bits
685 #define SCHED_FEAT(name, enabled) \
686 __SCHED_FEAT_##name ,
689 #include "sched_features.h"
694 #define SCHED_FEAT(name, enabled) \
695 (1UL << __SCHED_FEAT_##name) * enabled |
697 const_debug unsigned int sysctl_sched_features =
698 #include "sched_features.h"
703 #ifdef CONFIG_SCHED_DEBUG
704 #define SCHED_FEAT(name, enabled) \
707 static __read_mostly char *sched_feat_names[] = {
708 #include "sched_features.h"
714 static int sched_feat_open(struct inode *inode, struct file *filp)
716 filp->private_data = inode->i_private;
721 sched_feat_read(struct file *filp, char __user *ubuf,
722 size_t cnt, loff_t *ppos)
729 for (i = 0; sched_feat_names[i]; i++) {
730 len += strlen(sched_feat_names[i]);
734 buf = kmalloc(len + 2, GFP_KERNEL);
738 for (i = 0; sched_feat_names[i]; i++) {
739 if (sysctl_sched_features & (1UL << i))
740 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
742 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
745 r += sprintf(buf + r, "\n");
746 WARN_ON(r >= len + 2);
748 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
756 sched_feat_write(struct file *filp, const char __user *ubuf,
757 size_t cnt, loff_t *ppos)
767 if (copy_from_user(&buf, ubuf, cnt))
772 if (strncmp(buf, "NO_", 3) == 0) {
777 for (i = 0; sched_feat_names[i]; i++) {
778 int len = strlen(sched_feat_names[i]);
780 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
782 sysctl_sched_features &= ~(1UL << i);
784 sysctl_sched_features |= (1UL << i);
789 if (!sched_feat_names[i])
797 static struct file_operations sched_feat_fops = {
798 .open = sched_feat_open,
799 .read = sched_feat_read,
800 .write = sched_feat_write,
803 static __init int sched_init_debug(void)
805 debugfs_create_file("sched_features", 0644, NULL, NULL,
810 late_initcall(sched_init_debug);
814 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
817 * Number of tasks to iterate in a single balance run.
818 * Limited because this is done with IRQs disabled.
820 const_debug unsigned int sysctl_sched_nr_migrate = 32;
823 * ratelimit for updating the group shares.
826 unsigned int sysctl_sched_shares_ratelimit = 250000;
829 * Inject some fuzzyness into changing the per-cpu group shares
830 * this avoids remote rq-locks at the expense of fairness.
833 unsigned int sysctl_sched_shares_thresh = 4;
836 * period over which we measure -rt task cpu usage in us.
839 unsigned int sysctl_sched_rt_period = 1000000;
841 static __read_mostly int scheduler_running;
844 * part of the period that we allow rt tasks to run in us.
847 int sysctl_sched_rt_runtime = 950000;
849 static inline u64 global_rt_period(void)
851 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
854 static inline u64 global_rt_runtime(void)
856 if (sysctl_sched_rt_runtime < 0)
859 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
862 #ifndef prepare_arch_switch
863 # define prepare_arch_switch(next) do { } while (0)
865 #ifndef finish_arch_switch
866 # define finish_arch_switch(prev) do { } while (0)
869 static inline int task_current(struct rq *rq, struct task_struct *p)
871 return rq->curr == p;
874 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
875 static inline int task_running(struct rq *rq, struct task_struct *p)
877 return task_current(rq, p);
880 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
884 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
886 #ifdef CONFIG_DEBUG_SPINLOCK
887 /* this is a valid case when another task releases the spinlock */
888 rq->lock.owner = current;
891 * If we are tracking spinlock dependencies then we have to
892 * fix up the runqueue lock - which gets 'carried over' from
895 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
897 spin_unlock_irq(&rq->lock);
900 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
901 static inline int task_running(struct rq *rq, struct task_struct *p)
906 return task_current(rq, p);
910 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
914 * We can optimise this out completely for !SMP, because the
915 * SMP rebalancing from interrupt is the only thing that cares
920 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 spin_unlock_irq(&rq->lock);
923 spin_unlock(&rq->lock);
927 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
931 * After ->oncpu is cleared, the task can be moved to a different CPU.
932 * We must ensure this doesn't happen until the switch is completely
938 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
942 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
945 * __task_rq_lock - lock the runqueue a given task resides on.
946 * Must be called interrupts disabled.
948 static inline struct rq *__task_rq_lock(struct task_struct *p)
952 struct rq *rq = task_rq(p);
953 spin_lock(&rq->lock);
954 if (likely(rq == task_rq(p)))
956 spin_unlock(&rq->lock);
961 * task_rq_lock - lock the runqueue a given task resides on and disable
962 * interrupts. Note the ordering: we can safely lookup the task_rq without
963 * explicitly disabling preemption.
965 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
971 local_irq_save(*flags);
973 spin_lock(&rq->lock);
974 if (likely(rq == task_rq(p)))
976 spin_unlock_irqrestore(&rq->lock, *flags);
980 void task_rq_unlock_wait(struct task_struct *p)
982 struct rq *rq = task_rq(p);
984 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
985 spin_unlock_wait(&rq->lock);
988 static void __task_rq_unlock(struct rq *rq)
991 spin_unlock(&rq->lock);
994 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
997 spin_unlock_irqrestore(&rq->lock, *flags);
1001 * this_rq_lock - lock this runqueue and disable interrupts.
1003 static struct rq *this_rq_lock(void)
1004 __acquires(rq->lock)
1008 local_irq_disable();
1010 spin_lock(&rq->lock);
1015 #ifdef CONFIG_SCHED_HRTICK
1017 * Use HR-timers to deliver accurate preemption points.
1019 * Its all a bit involved since we cannot program an hrt while holding the
1020 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1023 * When we get rescheduled we reprogram the hrtick_timer outside of the
1029 * - enabled by features
1030 * - hrtimer is actually high res
1032 static inline int hrtick_enabled(struct rq *rq)
1034 if (!sched_feat(HRTICK))
1036 if (!cpu_active(cpu_of(rq)))
1038 return hrtimer_is_hres_active(&rq->hrtick_timer);
1041 static void hrtick_clear(struct rq *rq)
1043 if (hrtimer_active(&rq->hrtick_timer))
1044 hrtimer_cancel(&rq->hrtick_timer);
1048 * High-resolution timer tick.
1049 * Runs from hardirq context with interrupts disabled.
1051 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1053 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1055 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1057 spin_lock(&rq->lock);
1058 update_rq_clock(rq);
1059 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1060 spin_unlock(&rq->lock);
1062 return HRTIMER_NORESTART;
1067 * called from hardirq (IPI) context
1069 static void __hrtick_start(void *arg)
1071 struct rq *rq = arg;
1073 spin_lock(&rq->lock);
1074 hrtimer_restart(&rq->hrtick_timer);
1075 rq->hrtick_csd_pending = 0;
1076 spin_unlock(&rq->lock);
1080 * Called to set the hrtick timer state.
1082 * called with rq->lock held and irqs disabled
1084 static void hrtick_start(struct rq *rq, u64 delay)
1086 struct hrtimer *timer = &rq->hrtick_timer;
1087 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1089 hrtimer_set_expires(timer, time);
1091 if (rq == this_rq()) {
1092 hrtimer_restart(timer);
1093 } else if (!rq->hrtick_csd_pending) {
1094 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1095 rq->hrtick_csd_pending = 1;
1100 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1102 int cpu = (int)(long)hcpu;
1105 case CPU_UP_CANCELED:
1106 case CPU_UP_CANCELED_FROZEN:
1107 case CPU_DOWN_PREPARE:
1108 case CPU_DOWN_PREPARE_FROZEN:
1110 case CPU_DEAD_FROZEN:
1111 hrtick_clear(cpu_rq(cpu));
1118 static __init void init_hrtick(void)
1120 hotcpu_notifier(hotplug_hrtick, 0);
1124 * Called to set the hrtick timer state.
1126 * called with rq->lock held and irqs disabled
1128 static void hrtick_start(struct rq *rq, u64 delay)
1130 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1133 static inline void init_hrtick(void)
1136 #endif /* CONFIG_SMP */
1138 static void init_rq_hrtick(struct rq *rq)
1141 rq->hrtick_csd_pending = 0;
1143 rq->hrtick_csd.flags = 0;
1144 rq->hrtick_csd.func = __hrtick_start;
1145 rq->hrtick_csd.info = rq;
1148 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1149 rq->hrtick_timer.function = hrtick;
1150 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
1152 #else /* CONFIG_SCHED_HRTICK */
1153 static inline void hrtick_clear(struct rq *rq)
1157 static inline void init_rq_hrtick(struct rq *rq)
1161 static inline void init_hrtick(void)
1164 #endif /* CONFIG_SCHED_HRTICK */
1167 * resched_task - mark a task 'to be rescheduled now'.
1169 * On UP this means the setting of the need_resched flag, on SMP it
1170 * might also involve a cross-CPU call to trigger the scheduler on
1175 #ifndef tsk_is_polling
1176 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1179 static void resched_task(struct task_struct *p)
1183 assert_spin_locked(&task_rq(p)->lock);
1185 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1188 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1191 if (cpu == smp_processor_id())
1194 /* NEED_RESCHED must be visible before we test polling */
1196 if (!tsk_is_polling(p))
1197 smp_send_reschedule(cpu);
1200 static void resched_cpu(int cpu)
1202 struct rq *rq = cpu_rq(cpu);
1203 unsigned long flags;
1205 if (!spin_trylock_irqsave(&rq->lock, flags))
1207 resched_task(cpu_curr(cpu));
1208 spin_unlock_irqrestore(&rq->lock, flags);
1213 * When add_timer_on() enqueues a timer into the timer wheel of an
1214 * idle CPU then this timer might expire before the next timer event
1215 * which is scheduled to wake up that CPU. In case of a completely
1216 * idle system the next event might even be infinite time into the
1217 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1218 * leaves the inner idle loop so the newly added timer is taken into
1219 * account when the CPU goes back to idle and evaluates the timer
1220 * wheel for the next timer event.
1222 void wake_up_idle_cpu(int cpu)
1224 struct rq *rq = cpu_rq(cpu);
1226 if (cpu == smp_processor_id())
1230 * This is safe, as this function is called with the timer
1231 * wheel base lock of (cpu) held. When the CPU is on the way
1232 * to idle and has not yet set rq->curr to idle then it will
1233 * be serialized on the timer wheel base lock and take the new
1234 * timer into account automatically.
1236 if (rq->curr != rq->idle)
1240 * We can set TIF_RESCHED on the idle task of the other CPU
1241 * lockless. The worst case is that the other CPU runs the
1242 * idle task through an additional NOOP schedule()
1244 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1246 /* NEED_RESCHED must be visible before we test polling */
1248 if (!tsk_is_polling(rq->idle))
1249 smp_send_reschedule(cpu);
1251 #endif /* CONFIG_NO_HZ */
1253 #else /* !CONFIG_SMP */
1254 static void resched_task(struct task_struct *p)
1256 assert_spin_locked(&task_rq(p)->lock);
1257 set_tsk_need_resched(p);
1259 #endif /* CONFIG_SMP */
1261 #if BITS_PER_LONG == 32
1262 # define WMULT_CONST (~0UL)
1264 # define WMULT_CONST (1UL << 32)
1267 #define WMULT_SHIFT 32
1270 * Shift right and round:
1272 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1275 * delta *= weight / lw
1277 static unsigned long
1278 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1279 struct load_weight *lw)
1283 if (!lw->inv_weight) {
1284 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1287 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1291 tmp = (u64)delta_exec * weight;
1293 * Check whether we'd overflow the 64-bit multiplication:
1295 if (unlikely(tmp > WMULT_CONST))
1296 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1299 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1301 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1304 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1310 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1317 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1318 * of tasks with abnormal "nice" values across CPUs the contribution that
1319 * each task makes to its run queue's load is weighted according to its
1320 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1321 * scaled version of the new time slice allocation that they receive on time
1325 #define WEIGHT_IDLEPRIO 2
1326 #define WMULT_IDLEPRIO (1 << 31)
1329 * Nice levels are multiplicative, with a gentle 10% change for every
1330 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1331 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1332 * that remained on nice 0.
1334 * The "10% effect" is relative and cumulative: from _any_ nice level,
1335 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1336 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1337 * If a task goes up by ~10% and another task goes down by ~10% then
1338 * the relative distance between them is ~25%.)
1340 static const int prio_to_weight[40] = {
1341 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1342 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1343 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1344 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1345 /* 0 */ 1024, 820, 655, 526, 423,
1346 /* 5 */ 335, 272, 215, 172, 137,
1347 /* 10 */ 110, 87, 70, 56, 45,
1348 /* 15 */ 36, 29, 23, 18, 15,
1352 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1354 * In cases where the weight does not change often, we can use the
1355 * precalculated inverse to speed up arithmetics by turning divisions
1356 * into multiplications:
1358 static const u32 prio_to_wmult[40] = {
1359 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1360 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1361 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1362 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1363 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1364 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1365 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1366 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1369 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1372 * runqueue iterator, to support SMP load-balancing between different
1373 * scheduling classes, without having to expose their internal data
1374 * structures to the load-balancing proper:
1376 struct rq_iterator {
1378 struct task_struct *(*start)(void *);
1379 struct task_struct *(*next)(void *);
1383 static unsigned long
1384 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1385 unsigned long max_load_move, struct sched_domain *sd,
1386 enum cpu_idle_type idle, int *all_pinned,
1387 int *this_best_prio, struct rq_iterator *iterator);
1390 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1391 struct sched_domain *sd, enum cpu_idle_type idle,
1392 struct rq_iterator *iterator);
1395 #ifdef CONFIG_CGROUP_CPUACCT
1396 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1398 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1401 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1403 update_load_add(&rq->load, load);
1406 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1408 update_load_sub(&rq->load, load);
1411 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1412 typedef int (*tg_visitor)(struct task_group *, void *);
1415 * Iterate the full tree, calling @down when first entering a node and @up when
1416 * leaving it for the final time.
1418 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1420 struct task_group *parent, *child;
1424 parent = &root_task_group;
1426 ret = (*down)(parent, data);
1429 list_for_each_entry_rcu(child, &parent->children, siblings) {
1436 ret = (*up)(parent, data);
1441 parent = parent->parent;
1450 static int tg_nop(struct task_group *tg, void *data)
1457 static unsigned long source_load(int cpu, int type);
1458 static unsigned long target_load(int cpu, int type);
1459 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1461 static unsigned long cpu_avg_load_per_task(int cpu)
1463 struct rq *rq = cpu_rq(cpu);
1464 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1467 rq->avg_load_per_task = rq->load.weight / nr_running;
1469 rq->avg_load_per_task = 0;
1471 return rq->avg_load_per_task;
1474 #ifdef CONFIG_FAIR_GROUP_SCHED
1476 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1479 * Calculate and set the cpu's group shares.
1482 update_group_shares_cpu(struct task_group *tg, int cpu,
1483 unsigned long sd_shares, unsigned long sd_rq_weight)
1486 unsigned long shares;
1487 unsigned long rq_weight;
1492 rq_weight = tg->cfs_rq[cpu]->load.weight;
1495 * If there are currently no tasks on the cpu pretend there is one of
1496 * average load so that when a new task gets to run here it will not
1497 * get delayed by group starvation.
1501 rq_weight = NICE_0_LOAD;
1504 if (unlikely(rq_weight > sd_rq_weight))
1505 rq_weight = sd_rq_weight;
1508 * \Sum shares * rq_weight
1509 * shares = -----------------------
1513 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1514 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1516 if (abs(shares - tg->se[cpu]->load.weight) >
1517 sysctl_sched_shares_thresh) {
1518 struct rq *rq = cpu_rq(cpu);
1519 unsigned long flags;
1521 spin_lock_irqsave(&rq->lock, flags);
1523 * record the actual number of shares, not the boosted amount.
1525 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1526 tg->cfs_rq[cpu]->rq_weight = rq_weight;
1528 __set_se_shares(tg->se[cpu], shares);
1529 spin_unlock_irqrestore(&rq->lock, flags);
1534 * Re-compute the task group their per cpu shares over the given domain.
1535 * This needs to be done in a bottom-up fashion because the rq weight of a
1536 * parent group depends on the shares of its child groups.
1538 static int tg_shares_up(struct task_group *tg, void *data)
1540 unsigned long rq_weight = 0;
1541 unsigned long shares = 0;
1542 struct sched_domain *sd = data;
1545 for_each_cpu_mask(i, sd->span) {
1546 rq_weight += tg->cfs_rq[i]->load.weight;
1547 shares += tg->cfs_rq[i]->shares;
1550 if ((!shares && rq_weight) || shares > tg->shares)
1551 shares = tg->shares;
1553 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1554 shares = tg->shares;
1557 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1559 for_each_cpu_mask(i, sd->span)
1560 update_group_shares_cpu(tg, i, shares, rq_weight);
1566 * Compute the cpu's hierarchical load factor for each task group.
1567 * This needs to be done in a top-down fashion because the load of a child
1568 * group is a fraction of its parents load.
1570 static int tg_load_down(struct task_group *tg, void *data)
1573 long cpu = (long)data;
1576 load = cpu_rq(cpu)->load.weight;
1578 load = tg->parent->cfs_rq[cpu]->h_load;
1579 load *= tg->cfs_rq[cpu]->shares;
1580 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1583 tg->cfs_rq[cpu]->h_load = load;
1588 static void update_shares(struct sched_domain *sd)
1590 u64 now = cpu_clock(raw_smp_processor_id());
1591 s64 elapsed = now - sd->last_update;
1593 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1594 sd->last_update = now;
1595 walk_tg_tree(tg_nop, tg_shares_up, sd);
1599 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1601 spin_unlock(&rq->lock);
1603 spin_lock(&rq->lock);
1606 static void update_h_load(long cpu)
1608 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1613 static inline void update_shares(struct sched_domain *sd)
1617 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1625 #ifdef CONFIG_FAIR_GROUP_SCHED
1626 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1629 cfs_rq->shares = shares;
1634 #include "sched_stats.h"
1635 #include "sched_idletask.c"
1636 #include "sched_fair.c"
1637 #include "sched_rt.c"
1638 #ifdef CONFIG_SCHED_DEBUG
1639 # include "sched_debug.c"
1642 #define sched_class_highest (&rt_sched_class)
1643 #define for_each_class(class) \
1644 for (class = sched_class_highest; class; class = class->next)
1646 static void inc_nr_running(struct rq *rq)
1651 static void dec_nr_running(struct rq *rq)
1656 static void set_load_weight(struct task_struct *p)
1658 if (task_has_rt_policy(p)) {
1659 p->se.load.weight = prio_to_weight[0] * 2;
1660 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1665 * SCHED_IDLE tasks get minimal weight:
1667 if (p->policy == SCHED_IDLE) {
1668 p->se.load.weight = WEIGHT_IDLEPRIO;
1669 p->se.load.inv_weight = WMULT_IDLEPRIO;
1673 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1674 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1677 static void update_avg(u64 *avg, u64 sample)
1679 s64 diff = sample - *avg;
1683 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1685 sched_info_queued(p);
1686 p->sched_class->enqueue_task(rq, p, wakeup);
1690 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1692 if (sleep && p->se.last_wakeup) {
1693 update_avg(&p->se.avg_overlap,
1694 p->se.sum_exec_runtime - p->se.last_wakeup);
1695 p->se.last_wakeup = 0;
1698 sched_info_dequeued(p);
1699 p->sched_class->dequeue_task(rq, p, sleep);
1704 * __normal_prio - return the priority that is based on the static prio
1706 static inline int __normal_prio(struct task_struct *p)
1708 return p->static_prio;
1712 * Calculate the expected normal priority: i.e. priority
1713 * without taking RT-inheritance into account. Might be
1714 * boosted by interactivity modifiers. Changes upon fork,
1715 * setprio syscalls, and whenever the interactivity
1716 * estimator recalculates.
1718 static inline int normal_prio(struct task_struct *p)
1722 if (task_has_rt_policy(p))
1723 prio = MAX_RT_PRIO-1 - p->rt_priority;
1725 prio = __normal_prio(p);
1730 * Calculate the current priority, i.e. the priority
1731 * taken into account by the scheduler. This value might
1732 * be boosted by RT tasks, or might be boosted by
1733 * interactivity modifiers. Will be RT if the task got
1734 * RT-boosted. If not then it returns p->normal_prio.
1736 static int effective_prio(struct task_struct *p)
1738 p->normal_prio = normal_prio(p);
1740 * If we are RT tasks or we were boosted to RT priority,
1741 * keep the priority unchanged. Otherwise, update priority
1742 * to the normal priority:
1744 if (!rt_prio(p->prio))
1745 return p->normal_prio;
1750 * activate_task - move a task to the runqueue.
1752 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1754 if (task_contributes_to_load(p))
1755 rq->nr_uninterruptible--;
1757 enqueue_task(rq, p, wakeup);
1762 * deactivate_task - remove a task from the runqueue.
1764 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1766 if (task_contributes_to_load(p))
1767 rq->nr_uninterruptible++;
1769 dequeue_task(rq, p, sleep);
1774 * task_curr - is this task currently executing on a CPU?
1775 * @p: the task in question.
1777 inline int task_curr(const struct task_struct *p)
1779 return cpu_curr(task_cpu(p)) == p;
1782 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1784 set_task_rq(p, cpu);
1787 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1788 * successfuly executed on another CPU. We must ensure that updates of
1789 * per-task data have been completed by this moment.
1792 task_thread_info(p)->cpu = cpu;
1796 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1797 const struct sched_class *prev_class,
1798 int oldprio, int running)
1800 if (prev_class != p->sched_class) {
1801 if (prev_class->switched_from)
1802 prev_class->switched_from(rq, p, running);
1803 p->sched_class->switched_to(rq, p, running);
1805 p->sched_class->prio_changed(rq, p, oldprio, running);
1810 /* Used instead of source_load when we know the type == 0 */
1811 static unsigned long weighted_cpuload(const int cpu)
1813 return cpu_rq(cpu)->load.weight;
1817 * Is this task likely cache-hot:
1820 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1825 * Buddy candidates are cache hot:
1827 if (sched_feat(CACHE_HOT_BUDDY) &&
1828 (&p->se == cfs_rq_of(&p->se)->next ||
1829 &p->se == cfs_rq_of(&p->se)->last))
1832 if (p->sched_class != &fair_sched_class)
1835 if (sysctl_sched_migration_cost == -1)
1837 if (sysctl_sched_migration_cost == 0)
1840 delta = now - p->se.exec_start;
1842 return delta < (s64)sysctl_sched_migration_cost;
1846 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1848 int old_cpu = task_cpu(p);
1849 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1850 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1851 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1854 clock_offset = old_rq->clock - new_rq->clock;
1856 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1858 #ifdef CONFIG_SCHEDSTATS
1859 if (p->se.wait_start)
1860 p->se.wait_start -= clock_offset;
1861 if (p->se.sleep_start)
1862 p->se.sleep_start -= clock_offset;
1863 if (p->se.block_start)
1864 p->se.block_start -= clock_offset;
1865 if (old_cpu != new_cpu) {
1866 schedstat_inc(p, se.nr_migrations);
1867 if (task_hot(p, old_rq->clock, NULL))
1868 schedstat_inc(p, se.nr_forced2_migrations);
1871 p->se.vruntime -= old_cfsrq->min_vruntime -
1872 new_cfsrq->min_vruntime;
1874 __set_task_cpu(p, new_cpu);
1877 struct migration_req {
1878 struct list_head list;
1880 struct task_struct *task;
1883 struct completion done;
1887 * The task's runqueue lock must be held.
1888 * Returns true if you have to wait for migration thread.
1891 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1893 struct rq *rq = task_rq(p);
1896 * If the task is not on a runqueue (and not running), then
1897 * it is sufficient to simply update the task's cpu field.
1899 if (!p->se.on_rq && !task_running(rq, p)) {
1900 set_task_cpu(p, dest_cpu);
1904 init_completion(&req->done);
1906 req->dest_cpu = dest_cpu;
1907 list_add(&req->list, &rq->migration_queue);
1913 * wait_task_inactive - wait for a thread to unschedule.
1915 * If @match_state is nonzero, it's the @p->state value just checked and
1916 * not expected to change. If it changes, i.e. @p might have woken up,
1917 * then return zero. When we succeed in waiting for @p to be off its CPU,
1918 * we return a positive number (its total switch count). If a second call
1919 * a short while later returns the same number, the caller can be sure that
1920 * @p has remained unscheduled the whole time.
1922 * The caller must ensure that the task *will* unschedule sometime soon,
1923 * else this function might spin for a *long* time. This function can't
1924 * be called with interrupts off, or it may introduce deadlock with
1925 * smp_call_function() if an IPI is sent by the same process we are
1926 * waiting to become inactive.
1928 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1930 unsigned long flags;
1937 * We do the initial early heuristics without holding
1938 * any task-queue locks at all. We'll only try to get
1939 * the runqueue lock when things look like they will
1945 * If the task is actively running on another CPU
1946 * still, just relax and busy-wait without holding
1949 * NOTE! Since we don't hold any locks, it's not
1950 * even sure that "rq" stays as the right runqueue!
1951 * But we don't care, since "task_running()" will
1952 * return false if the runqueue has changed and p
1953 * is actually now running somewhere else!
1955 while (task_running(rq, p)) {
1956 if (match_state && unlikely(p->state != match_state))
1962 * Ok, time to look more closely! We need the rq
1963 * lock now, to be *sure*. If we're wrong, we'll
1964 * just go back and repeat.
1966 rq = task_rq_lock(p, &flags);
1967 trace_sched_wait_task(rq, p);
1968 running = task_running(rq, p);
1969 on_rq = p->se.on_rq;
1971 if (!match_state || p->state == match_state)
1972 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1973 task_rq_unlock(rq, &flags);
1976 * If it changed from the expected state, bail out now.
1978 if (unlikely(!ncsw))
1982 * Was it really running after all now that we
1983 * checked with the proper locks actually held?
1985 * Oops. Go back and try again..
1987 if (unlikely(running)) {
1993 * It's not enough that it's not actively running,
1994 * it must be off the runqueue _entirely_, and not
1997 * So if it wa still runnable (but just not actively
1998 * running right now), it's preempted, and we should
1999 * yield - it could be a while.
2001 if (unlikely(on_rq)) {
2002 schedule_timeout_uninterruptible(1);
2007 * Ahh, all good. It wasn't running, and it wasn't
2008 * runnable, which means that it will never become
2009 * running in the future either. We're all done!
2018 * kick_process - kick a running thread to enter/exit the kernel
2019 * @p: the to-be-kicked thread
2021 * Cause a process which is running on another CPU to enter
2022 * kernel-mode, without any delay. (to get signals handled.)
2024 * NOTE: this function doesnt have to take the runqueue lock,
2025 * because all it wants to ensure is that the remote task enters
2026 * the kernel. If the IPI races and the task has been migrated
2027 * to another CPU then no harm is done and the purpose has been
2030 void kick_process(struct task_struct *p)
2036 if ((cpu != smp_processor_id()) && task_curr(p))
2037 smp_send_reschedule(cpu);
2042 * Return a low guess at the load of a migration-source cpu weighted
2043 * according to the scheduling class and "nice" value.
2045 * We want to under-estimate the load of migration sources, to
2046 * balance conservatively.
2048 static unsigned long source_load(int cpu, int type)
2050 struct rq *rq = cpu_rq(cpu);
2051 unsigned long total = weighted_cpuload(cpu);
2053 if (type == 0 || !sched_feat(LB_BIAS))
2056 return min(rq->cpu_load[type-1], total);
2060 * Return a high guess at the load of a migration-target cpu weighted
2061 * according to the scheduling class and "nice" value.
2063 static unsigned long target_load(int cpu, int type)
2065 struct rq *rq = cpu_rq(cpu);
2066 unsigned long total = weighted_cpuload(cpu);
2068 if (type == 0 || !sched_feat(LB_BIAS))
2071 return max(rq->cpu_load[type-1], total);
2075 * find_idlest_group finds and returns the least busy CPU group within the
2078 static struct sched_group *
2079 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2081 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2082 unsigned long min_load = ULONG_MAX, this_load = 0;
2083 int load_idx = sd->forkexec_idx;
2084 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2087 unsigned long load, avg_load;
2091 /* Skip over this group if it has no CPUs allowed */
2092 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2095 local_group = cpu_isset(this_cpu, group->cpumask);
2097 /* Tally up the load of all CPUs in the group */
2100 for_each_cpu_mask_nr(i, group->cpumask) {
2101 /* Bias balancing toward cpus of our domain */
2103 load = source_load(i, load_idx);
2105 load = target_load(i, load_idx);
2110 /* Adjust by relative CPU power of the group */
2111 avg_load = sg_div_cpu_power(group,
2112 avg_load * SCHED_LOAD_SCALE);
2115 this_load = avg_load;
2117 } else if (avg_load < min_load) {
2118 min_load = avg_load;
2121 } while (group = group->next, group != sd->groups);
2123 if (!idlest || 100*this_load < imbalance*min_load)
2129 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2132 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2135 unsigned long load, min_load = ULONG_MAX;
2139 /* Traverse only the allowed CPUs */
2140 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2142 for_each_cpu_mask_nr(i, *tmp) {
2143 load = weighted_cpuload(i);
2145 if (load < min_load || (load == min_load && i == this_cpu)) {
2155 * sched_balance_self: balance the current task (running on cpu) in domains
2156 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2159 * Balance, ie. select the least loaded group.
2161 * Returns the target CPU number, or the same CPU if no balancing is needed.
2163 * preempt must be disabled.
2165 static int sched_balance_self(int cpu, int flag)
2167 struct task_struct *t = current;
2168 struct sched_domain *tmp, *sd = NULL;
2170 for_each_domain(cpu, tmp) {
2172 * If power savings logic is enabled for a domain, stop there.
2174 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2176 if (tmp->flags & flag)
2184 cpumask_t span, tmpmask;
2185 struct sched_group *group;
2186 int new_cpu, weight;
2188 if (!(sd->flags & flag)) {
2194 group = find_idlest_group(sd, t, cpu);
2200 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2201 if (new_cpu == -1 || new_cpu == cpu) {
2202 /* Now try balancing at a lower domain level of cpu */
2207 /* Now try balancing at a lower domain level of new_cpu */
2210 weight = cpus_weight(span);
2211 for_each_domain(cpu, tmp) {
2212 if (weight <= cpus_weight(tmp->span))
2214 if (tmp->flags & flag)
2217 /* while loop will break here if sd == NULL */
2223 #endif /* CONFIG_SMP */
2226 * try_to_wake_up - wake up a thread
2227 * @p: the to-be-woken-up thread
2228 * @state: the mask of task states that can be woken
2229 * @sync: do a synchronous wakeup?
2231 * Put it on the run-queue if it's not already there. The "current"
2232 * thread is always on the run-queue (except when the actual
2233 * re-schedule is in progress), and as such you're allowed to do
2234 * the simpler "current->state = TASK_RUNNING" to mark yourself
2235 * runnable without the overhead of this.
2237 * returns failure only if the task is already active.
2239 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2241 int cpu, orig_cpu, this_cpu, success = 0;
2242 unsigned long flags;
2246 if (!sched_feat(SYNC_WAKEUPS))
2250 if (sched_feat(LB_WAKEUP_UPDATE)) {
2251 struct sched_domain *sd;
2253 this_cpu = raw_smp_processor_id();
2256 for_each_domain(this_cpu, sd) {
2257 if (cpu_isset(cpu, sd->span)) {
2266 rq = task_rq_lock(p, &flags);
2267 old_state = p->state;
2268 if (!(old_state & state))
2276 this_cpu = smp_processor_id();
2279 if (unlikely(task_running(rq, p)))
2282 cpu = p->sched_class->select_task_rq(p, sync);
2283 if (cpu != orig_cpu) {
2284 set_task_cpu(p, cpu);
2285 task_rq_unlock(rq, &flags);
2286 /* might preempt at this point */
2287 rq = task_rq_lock(p, &flags);
2288 old_state = p->state;
2289 if (!(old_state & state))
2294 this_cpu = smp_processor_id();
2298 #ifdef CONFIG_SCHEDSTATS
2299 schedstat_inc(rq, ttwu_count);
2300 if (cpu == this_cpu)
2301 schedstat_inc(rq, ttwu_local);
2303 struct sched_domain *sd;
2304 for_each_domain(this_cpu, sd) {
2305 if (cpu_isset(cpu, sd->span)) {
2306 schedstat_inc(sd, ttwu_wake_remote);
2311 #endif /* CONFIG_SCHEDSTATS */
2314 #endif /* CONFIG_SMP */
2315 schedstat_inc(p, se.nr_wakeups);
2317 schedstat_inc(p, se.nr_wakeups_sync);
2318 if (orig_cpu != cpu)
2319 schedstat_inc(p, se.nr_wakeups_migrate);
2320 if (cpu == this_cpu)
2321 schedstat_inc(p, se.nr_wakeups_local);
2323 schedstat_inc(p, se.nr_wakeups_remote);
2324 update_rq_clock(rq);
2325 activate_task(rq, p, 1);
2329 trace_sched_wakeup(rq, p, success);
2330 check_preempt_curr(rq, p, sync);
2332 p->state = TASK_RUNNING;
2334 if (p->sched_class->task_wake_up)
2335 p->sched_class->task_wake_up(rq, p);
2338 current->se.last_wakeup = current->se.sum_exec_runtime;
2340 task_rq_unlock(rq, &flags);
2345 int wake_up_process(struct task_struct *p)
2347 return try_to_wake_up(p, TASK_ALL, 0);
2349 EXPORT_SYMBOL(wake_up_process);
2351 int wake_up_state(struct task_struct *p, unsigned int state)
2353 return try_to_wake_up(p, state, 0);
2357 * Perform scheduler related setup for a newly forked process p.
2358 * p is forked by current.
2360 * __sched_fork() is basic setup used by init_idle() too:
2362 static void __sched_fork(struct task_struct *p)
2364 p->se.exec_start = 0;
2365 p->se.sum_exec_runtime = 0;
2366 p->se.prev_sum_exec_runtime = 0;
2367 p->se.last_wakeup = 0;
2368 p->se.avg_overlap = 0;
2370 #ifdef CONFIG_SCHEDSTATS
2371 p->se.wait_start = 0;
2372 p->se.sum_sleep_runtime = 0;
2373 p->se.sleep_start = 0;
2374 p->se.block_start = 0;
2375 p->se.sleep_max = 0;
2376 p->se.block_max = 0;
2378 p->se.slice_max = 0;
2382 INIT_LIST_HEAD(&p->rt.run_list);
2384 INIT_LIST_HEAD(&p->se.group_node);
2386 #ifdef CONFIG_PREEMPT_NOTIFIERS
2387 INIT_HLIST_HEAD(&p->preempt_notifiers);
2391 * We mark the process as running here, but have not actually
2392 * inserted it onto the runqueue yet. This guarantees that
2393 * nobody will actually run it, and a signal or other external
2394 * event cannot wake it up and insert it on the runqueue either.
2396 p->state = TASK_RUNNING;
2400 * fork()/clone()-time setup:
2402 void sched_fork(struct task_struct *p, int clone_flags)
2404 int cpu = get_cpu();
2409 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2411 set_task_cpu(p, cpu);
2414 * Make sure we do not leak PI boosting priority to the child:
2416 p->prio = current->normal_prio;
2417 if (!rt_prio(p->prio))
2418 p->sched_class = &fair_sched_class;
2420 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2421 if (likely(sched_info_on()))
2422 memset(&p->sched_info, 0, sizeof(p->sched_info));
2424 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2427 #ifdef CONFIG_PREEMPT
2428 /* Want to start with kernel preemption disabled. */
2429 task_thread_info(p)->preempt_count = 1;
2435 * wake_up_new_task - wake up a newly created task for the first time.
2437 * This function will do some initial scheduler statistics housekeeping
2438 * that must be done for every newly created context, then puts the task
2439 * on the runqueue and wakes it.
2441 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2443 unsigned long flags;
2446 rq = task_rq_lock(p, &flags);
2447 BUG_ON(p->state != TASK_RUNNING);
2448 update_rq_clock(rq);
2450 p->prio = effective_prio(p);
2452 if (!p->sched_class->task_new || !current->se.on_rq) {
2453 activate_task(rq, p, 0);
2456 * Let the scheduling class do new task startup
2457 * management (if any):
2459 p->sched_class->task_new(rq, p);
2462 trace_sched_wakeup_new(rq, p, 1);
2463 check_preempt_curr(rq, p, 0);
2465 if (p->sched_class->task_wake_up)
2466 p->sched_class->task_wake_up(rq, p);
2468 task_rq_unlock(rq, &flags);
2471 #ifdef CONFIG_PREEMPT_NOTIFIERS
2474 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2475 * @notifier: notifier struct to register
2477 void preempt_notifier_register(struct preempt_notifier *notifier)
2479 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2481 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2484 * preempt_notifier_unregister - no longer interested in preemption notifications
2485 * @notifier: notifier struct to unregister
2487 * This is safe to call from within a preemption notifier.
2489 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2491 hlist_del(¬ifier->link);
2493 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2495 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2497 struct preempt_notifier *notifier;
2498 struct hlist_node *node;
2500 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2501 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2505 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2506 struct task_struct *next)
2508 struct preempt_notifier *notifier;
2509 struct hlist_node *node;
2511 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2512 notifier->ops->sched_out(notifier, next);
2515 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2517 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2522 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2523 struct task_struct *next)
2527 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2530 * prepare_task_switch - prepare to switch tasks
2531 * @rq: the runqueue preparing to switch
2532 * @prev: the current task that is being switched out
2533 * @next: the task we are going to switch to.
2535 * This is called with the rq lock held and interrupts off. It must
2536 * be paired with a subsequent finish_task_switch after the context
2539 * prepare_task_switch sets up locking and calls architecture specific
2543 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2544 struct task_struct *next)
2546 fire_sched_out_preempt_notifiers(prev, next);
2547 prepare_lock_switch(rq, next);
2548 prepare_arch_switch(next);
2552 * finish_task_switch - clean up after a task-switch
2553 * @rq: runqueue associated with task-switch
2554 * @prev: the thread we just switched away from.
2556 * finish_task_switch must be called after the context switch, paired
2557 * with a prepare_task_switch call before the context switch.
2558 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2559 * and do any other architecture-specific cleanup actions.
2561 * Note that we may have delayed dropping an mm in context_switch(). If
2562 * so, we finish that here outside of the runqueue lock. (Doing it
2563 * with the lock held can cause deadlocks; see schedule() for
2566 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2567 __releases(rq->lock)
2569 struct mm_struct *mm = rq->prev_mm;
2575 * A task struct has one reference for the use as "current".
2576 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2577 * schedule one last time. The schedule call will never return, and
2578 * the scheduled task must drop that reference.
2579 * The test for TASK_DEAD must occur while the runqueue locks are
2580 * still held, otherwise prev could be scheduled on another cpu, die
2581 * there before we look at prev->state, and then the reference would
2583 * Manfred Spraul <manfred@colorfullife.com>
2585 prev_state = prev->state;
2586 finish_arch_switch(prev);
2587 finish_lock_switch(rq, prev);
2589 if (current->sched_class->post_schedule)
2590 current->sched_class->post_schedule(rq);
2593 fire_sched_in_preempt_notifiers(current);
2596 if (unlikely(prev_state == TASK_DEAD)) {
2598 * Remove function-return probe instances associated with this
2599 * task and put them back on the free list.
2601 kprobe_flush_task(prev);
2602 put_task_struct(prev);
2607 * schedule_tail - first thing a freshly forked thread must call.
2608 * @prev: the thread we just switched away from.
2610 asmlinkage void schedule_tail(struct task_struct *prev)
2611 __releases(rq->lock)
2613 struct rq *rq = this_rq();
2615 finish_task_switch(rq, prev);
2616 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2617 /* In this case, finish_task_switch does not reenable preemption */
2620 if (current->set_child_tid)
2621 put_user(task_pid_vnr(current), current->set_child_tid);
2625 * context_switch - switch to the new MM and the new
2626 * thread's register state.
2629 context_switch(struct rq *rq, struct task_struct *prev,
2630 struct task_struct *next)
2632 struct mm_struct *mm, *oldmm;
2634 prepare_task_switch(rq, prev, next);
2635 trace_sched_switch(rq, prev, next);
2637 oldmm = prev->active_mm;
2639 * For paravirt, this is coupled with an exit in switch_to to
2640 * combine the page table reload and the switch backend into
2643 arch_enter_lazy_cpu_mode();
2645 if (unlikely(!mm)) {
2646 next->active_mm = oldmm;
2647 atomic_inc(&oldmm->mm_count);
2648 enter_lazy_tlb(oldmm, next);
2650 switch_mm(oldmm, mm, next);
2652 if (unlikely(!prev->mm)) {
2653 prev->active_mm = NULL;
2654 rq->prev_mm = oldmm;
2657 * Since the runqueue lock will be released by the next
2658 * task (which is an invalid locking op but in the case
2659 * of the scheduler it's an obvious special-case), so we
2660 * do an early lockdep release here:
2662 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2663 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2666 /* Here we just switch the register state and the stack. */
2667 switch_to(prev, next, prev);
2671 * this_rq must be evaluated again because prev may have moved
2672 * CPUs since it called schedule(), thus the 'rq' on its stack
2673 * frame will be invalid.
2675 finish_task_switch(this_rq(), prev);
2679 * nr_running, nr_uninterruptible and nr_context_switches:
2681 * externally visible scheduler statistics: current number of runnable
2682 * threads, current number of uninterruptible-sleeping threads, total
2683 * number of context switches performed since bootup.
2685 unsigned long nr_running(void)
2687 unsigned long i, sum = 0;
2689 for_each_online_cpu(i)
2690 sum += cpu_rq(i)->nr_running;
2695 unsigned long nr_uninterruptible(void)
2697 unsigned long i, sum = 0;
2699 for_each_possible_cpu(i)
2700 sum += cpu_rq(i)->nr_uninterruptible;
2703 * Since we read the counters lockless, it might be slightly
2704 * inaccurate. Do not allow it to go below zero though:
2706 if (unlikely((long)sum < 0))
2712 unsigned long long nr_context_switches(void)
2715 unsigned long long sum = 0;
2717 for_each_possible_cpu(i)
2718 sum += cpu_rq(i)->nr_switches;
2723 unsigned long nr_iowait(void)
2725 unsigned long i, sum = 0;
2727 for_each_possible_cpu(i)
2728 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2733 unsigned long nr_active(void)
2735 unsigned long i, running = 0, uninterruptible = 0;
2737 for_each_online_cpu(i) {
2738 running += cpu_rq(i)->nr_running;
2739 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2742 if (unlikely((long)uninterruptible < 0))
2743 uninterruptible = 0;
2745 return running + uninterruptible;
2749 * Update rq->cpu_load[] statistics. This function is usually called every
2750 * scheduler tick (TICK_NSEC).
2752 static void update_cpu_load(struct rq *this_rq)
2754 unsigned long this_load = this_rq->load.weight;
2757 this_rq->nr_load_updates++;
2759 /* Update our load: */
2760 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2761 unsigned long old_load, new_load;
2763 /* scale is effectively 1 << i now, and >> i divides by scale */
2765 old_load = this_rq->cpu_load[i];
2766 new_load = this_load;
2768 * Round up the averaging division if load is increasing. This
2769 * prevents us from getting stuck on 9 if the load is 10, for
2772 if (new_load > old_load)
2773 new_load += scale-1;
2774 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2781 * double_rq_lock - safely lock two runqueues
2783 * Note this does not disable interrupts like task_rq_lock,
2784 * you need to do so manually before calling.
2786 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2787 __acquires(rq1->lock)
2788 __acquires(rq2->lock)
2790 BUG_ON(!irqs_disabled());
2792 spin_lock(&rq1->lock);
2793 __acquire(rq2->lock); /* Fake it out ;) */
2796 spin_lock(&rq1->lock);
2797 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2799 spin_lock(&rq2->lock);
2800 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2803 update_rq_clock(rq1);
2804 update_rq_clock(rq2);
2808 * double_rq_unlock - safely unlock two runqueues
2810 * Note this does not restore interrupts like task_rq_unlock,
2811 * you need to do so manually after calling.
2813 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2814 __releases(rq1->lock)
2815 __releases(rq2->lock)
2817 spin_unlock(&rq1->lock);
2819 spin_unlock(&rq2->lock);
2821 __release(rq2->lock);
2825 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2827 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2828 __releases(this_rq->lock)
2829 __acquires(busiest->lock)
2830 __acquires(this_rq->lock)
2834 if (unlikely(!irqs_disabled())) {
2835 /* printk() doesn't work good under rq->lock */
2836 spin_unlock(&this_rq->lock);
2839 if (unlikely(!spin_trylock(&busiest->lock))) {
2840 if (busiest < this_rq) {
2841 spin_unlock(&this_rq->lock);
2842 spin_lock(&busiest->lock);
2843 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2846 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2851 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2852 __releases(busiest->lock)
2854 spin_unlock(&busiest->lock);
2855 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2859 * If dest_cpu is allowed for this process, migrate the task to it.
2860 * This is accomplished by forcing the cpu_allowed mask to only
2861 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2862 * the cpu_allowed mask is restored.
2864 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2866 struct migration_req req;
2867 unsigned long flags;
2870 rq = task_rq_lock(p, &flags);
2871 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2872 || unlikely(!cpu_active(dest_cpu)))
2875 /* force the process onto the specified CPU */
2876 if (migrate_task(p, dest_cpu, &req)) {
2877 /* Need to wait for migration thread (might exit: take ref). */
2878 struct task_struct *mt = rq->migration_thread;
2880 get_task_struct(mt);
2881 task_rq_unlock(rq, &flags);
2882 wake_up_process(mt);
2883 put_task_struct(mt);
2884 wait_for_completion(&req.done);
2889 task_rq_unlock(rq, &flags);
2893 * sched_exec - execve() is a valuable balancing opportunity, because at
2894 * this point the task has the smallest effective memory and cache footprint.
2896 void sched_exec(void)
2898 int new_cpu, this_cpu = get_cpu();
2899 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2901 if (new_cpu != this_cpu)
2902 sched_migrate_task(current, new_cpu);
2906 * pull_task - move a task from a remote runqueue to the local runqueue.
2907 * Both runqueues must be locked.
2909 static void pull_task(struct rq *src_rq, struct task_struct *p,
2910 struct rq *this_rq, int this_cpu)
2912 deactivate_task(src_rq, p, 0);
2913 set_task_cpu(p, this_cpu);
2914 activate_task(this_rq, p, 0);
2916 * Note that idle threads have a prio of MAX_PRIO, for this test
2917 * to be always true for them.
2919 check_preempt_curr(this_rq, p, 0);
2923 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2926 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2927 struct sched_domain *sd, enum cpu_idle_type idle,
2931 * We do not migrate tasks that are:
2932 * 1) running (obviously), or
2933 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2934 * 3) are cache-hot on their current CPU.
2936 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2937 schedstat_inc(p, se.nr_failed_migrations_affine);
2942 if (task_running(rq, p)) {
2943 schedstat_inc(p, se.nr_failed_migrations_running);
2948 * Aggressive migration if:
2949 * 1) task is cache cold, or
2950 * 2) too many balance attempts have failed.
2953 if (!task_hot(p, rq->clock, sd) ||
2954 sd->nr_balance_failed > sd->cache_nice_tries) {
2955 #ifdef CONFIG_SCHEDSTATS
2956 if (task_hot(p, rq->clock, sd)) {
2957 schedstat_inc(sd, lb_hot_gained[idle]);
2958 schedstat_inc(p, se.nr_forced_migrations);
2964 if (task_hot(p, rq->clock, sd)) {
2965 schedstat_inc(p, se.nr_failed_migrations_hot);
2971 static unsigned long
2972 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2973 unsigned long max_load_move, struct sched_domain *sd,
2974 enum cpu_idle_type idle, int *all_pinned,
2975 int *this_best_prio, struct rq_iterator *iterator)
2977 int loops = 0, pulled = 0, pinned = 0;
2978 struct task_struct *p;
2979 long rem_load_move = max_load_move;
2981 if (max_load_move == 0)
2987 * Start the load-balancing iterator:
2989 p = iterator->start(iterator->arg);
2991 if (!p || loops++ > sysctl_sched_nr_migrate)
2994 if ((p->se.load.weight >> 1) > rem_load_move ||
2995 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2996 p = iterator->next(iterator->arg);
3000 pull_task(busiest, p, this_rq, this_cpu);
3002 rem_load_move -= p->se.load.weight;
3005 * We only want to steal up to the prescribed amount of weighted load.
3007 if (rem_load_move > 0) {
3008 if (p->prio < *this_best_prio)
3009 *this_best_prio = p->prio;
3010 p = iterator->next(iterator->arg);
3015 * Right now, this is one of only two places pull_task() is called,
3016 * so we can safely collect pull_task() stats here rather than
3017 * inside pull_task().
3019 schedstat_add(sd, lb_gained[idle], pulled);
3022 *all_pinned = pinned;
3024 return max_load_move - rem_load_move;
3028 * move_tasks tries to move up to max_load_move weighted load from busiest to
3029 * this_rq, as part of a balancing operation within domain "sd".
3030 * Returns 1 if successful and 0 otherwise.
3032 * Called with both runqueues locked.
3034 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3035 unsigned long max_load_move,
3036 struct sched_domain *sd, enum cpu_idle_type idle,
3039 const struct sched_class *class = sched_class_highest;
3040 unsigned long total_load_moved = 0;
3041 int this_best_prio = this_rq->curr->prio;
3045 class->load_balance(this_rq, this_cpu, busiest,
3046 max_load_move - total_load_moved,
3047 sd, idle, all_pinned, &this_best_prio);
3048 class = class->next;
3050 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3053 } while (class && max_load_move > total_load_moved);
3055 return total_load_moved > 0;
3059 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3060 struct sched_domain *sd, enum cpu_idle_type idle,
3061 struct rq_iterator *iterator)
3063 struct task_struct *p = iterator->start(iterator->arg);
3067 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3068 pull_task(busiest, p, this_rq, this_cpu);
3070 * Right now, this is only the second place pull_task()
3071 * is called, so we can safely collect pull_task()
3072 * stats here rather than inside pull_task().
3074 schedstat_inc(sd, lb_gained[idle]);
3078 p = iterator->next(iterator->arg);
3085 * move_one_task tries to move exactly one task from busiest to this_rq, as
3086 * part of active balancing operations within "domain".
3087 * Returns 1 if successful and 0 otherwise.
3089 * Called with both runqueues locked.
3091 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3092 struct sched_domain *sd, enum cpu_idle_type idle)
3094 const struct sched_class *class;
3096 for (class = sched_class_highest; class; class = class->next)
3097 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3104 * find_busiest_group finds and returns the busiest CPU group within the
3105 * domain. It calculates and returns the amount of weighted load which
3106 * should be moved to restore balance via the imbalance parameter.
3108 static struct sched_group *
3109 find_busiest_group(struct sched_domain *sd, int this_cpu,
3110 unsigned long *imbalance, enum cpu_idle_type idle,
3111 int *sd_idle, const cpumask_t *cpus, int *balance)
3113 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3114 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3115 unsigned long max_pull;
3116 unsigned long busiest_load_per_task, busiest_nr_running;
3117 unsigned long this_load_per_task, this_nr_running;
3118 int load_idx, group_imb = 0;
3119 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3120 int power_savings_balance = 1;
3121 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3122 unsigned long min_nr_running = ULONG_MAX;
3123 struct sched_group *group_min = NULL, *group_leader = NULL;
3126 max_load = this_load = total_load = total_pwr = 0;
3127 busiest_load_per_task = busiest_nr_running = 0;
3128 this_load_per_task = this_nr_running = 0;
3130 if (idle == CPU_NOT_IDLE)
3131 load_idx = sd->busy_idx;
3132 else if (idle == CPU_NEWLY_IDLE)
3133 load_idx = sd->newidle_idx;
3135 load_idx = sd->idle_idx;
3138 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3141 int __group_imb = 0;
3142 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3143 unsigned long sum_nr_running, sum_weighted_load;
3144 unsigned long sum_avg_load_per_task;
3145 unsigned long avg_load_per_task;
3147 local_group = cpu_isset(this_cpu, group->cpumask);
3150 balance_cpu = first_cpu(group->cpumask);
3152 /* Tally up the load of all CPUs in the group */
3153 sum_weighted_load = sum_nr_running = avg_load = 0;
3154 sum_avg_load_per_task = avg_load_per_task = 0;
3157 min_cpu_load = ~0UL;
3159 for_each_cpu_mask_nr(i, group->cpumask) {
3162 if (!cpu_isset(i, *cpus))
3167 if (*sd_idle && rq->nr_running)
3170 /* Bias balancing toward cpus of our domain */
3172 if (idle_cpu(i) && !first_idle_cpu) {
3177 load = target_load(i, load_idx);
3179 load = source_load(i, load_idx);
3180 if (load > max_cpu_load)
3181 max_cpu_load = load;
3182 if (min_cpu_load > load)
3183 min_cpu_load = load;
3187 sum_nr_running += rq->nr_running;
3188 sum_weighted_load += weighted_cpuload(i);
3190 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3194 * First idle cpu or the first cpu(busiest) in this sched group
3195 * is eligible for doing load balancing at this and above
3196 * domains. In the newly idle case, we will allow all the cpu's
3197 * to do the newly idle load balance.
3199 if (idle != CPU_NEWLY_IDLE && local_group &&
3200 balance_cpu != this_cpu && balance) {
3205 total_load += avg_load;
3206 total_pwr += group->__cpu_power;
3208 /* Adjust by relative CPU power of the group */
3209 avg_load = sg_div_cpu_power(group,
3210 avg_load * SCHED_LOAD_SCALE);
3214 * Consider the group unbalanced when the imbalance is larger
3215 * than the average weight of two tasks.
3217 * APZ: with cgroup the avg task weight can vary wildly and
3218 * might not be a suitable number - should we keep a
3219 * normalized nr_running number somewhere that negates
3222 avg_load_per_task = sg_div_cpu_power(group,
3223 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3225 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3228 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3231 this_load = avg_load;
3233 this_nr_running = sum_nr_running;
3234 this_load_per_task = sum_weighted_load;
3235 } else if (avg_load > max_load &&
3236 (sum_nr_running > group_capacity || __group_imb)) {
3237 max_load = avg_load;
3239 busiest_nr_running = sum_nr_running;
3240 busiest_load_per_task = sum_weighted_load;
3241 group_imb = __group_imb;
3244 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3246 * Busy processors will not participate in power savings
3249 if (idle == CPU_NOT_IDLE ||
3250 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3254 * If the local group is idle or completely loaded
3255 * no need to do power savings balance at this domain
3257 if (local_group && (this_nr_running >= group_capacity ||
3259 power_savings_balance = 0;
3262 * If a group is already running at full capacity or idle,
3263 * don't include that group in power savings calculations
3265 if (!power_savings_balance || sum_nr_running >= group_capacity
3270 * Calculate the group which has the least non-idle load.
3271 * This is the group from where we need to pick up the load
3274 if ((sum_nr_running < min_nr_running) ||
3275 (sum_nr_running == min_nr_running &&
3276 first_cpu(group->cpumask) <
3277 first_cpu(group_min->cpumask))) {
3279 min_nr_running = sum_nr_running;
3280 min_load_per_task = sum_weighted_load /
3285 * Calculate the group which is almost near its
3286 * capacity but still has some space to pick up some load
3287 * from other group and save more power
3289 if (sum_nr_running <= group_capacity - 1) {
3290 if (sum_nr_running > leader_nr_running ||
3291 (sum_nr_running == leader_nr_running &&
3292 first_cpu(group->cpumask) >
3293 first_cpu(group_leader->cpumask))) {
3294 group_leader = group;
3295 leader_nr_running = sum_nr_running;
3300 group = group->next;
3301 } while (group != sd->groups);
3303 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3306 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3308 if (this_load >= avg_load ||
3309 100*max_load <= sd->imbalance_pct*this_load)
3312 busiest_load_per_task /= busiest_nr_running;
3314 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3317 * We're trying to get all the cpus to the average_load, so we don't
3318 * want to push ourselves above the average load, nor do we wish to
3319 * reduce the max loaded cpu below the average load, as either of these
3320 * actions would just result in more rebalancing later, and ping-pong
3321 * tasks around. Thus we look for the minimum possible imbalance.
3322 * Negative imbalances (*we* are more loaded than anyone else) will
3323 * be counted as no imbalance for these purposes -- we can't fix that
3324 * by pulling tasks to us. Be careful of negative numbers as they'll
3325 * appear as very large values with unsigned longs.
3327 if (max_load <= busiest_load_per_task)
3331 * In the presence of smp nice balancing, certain scenarios can have
3332 * max load less than avg load(as we skip the groups at or below
3333 * its cpu_power, while calculating max_load..)
3335 if (max_load < avg_load) {
3337 goto small_imbalance;
3340 /* Don't want to pull so many tasks that a group would go idle */
3341 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3343 /* How much load to actually move to equalise the imbalance */
3344 *imbalance = min(max_pull * busiest->__cpu_power,
3345 (avg_load - this_load) * this->__cpu_power)
3349 * if *imbalance is less than the average load per runnable task
3350 * there is no gaurantee that any tasks will be moved so we'll have
3351 * a think about bumping its value to force at least one task to be
3354 if (*imbalance < busiest_load_per_task) {
3355 unsigned long tmp, pwr_now, pwr_move;
3359 pwr_move = pwr_now = 0;
3361 if (this_nr_running) {
3362 this_load_per_task /= this_nr_running;
3363 if (busiest_load_per_task > this_load_per_task)
3366 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3368 if (max_load - this_load + busiest_load_per_task >=
3369 busiest_load_per_task * imbn) {
3370 *imbalance = busiest_load_per_task;
3375 * OK, we don't have enough imbalance to justify moving tasks,
3376 * however we may be able to increase total CPU power used by
3380 pwr_now += busiest->__cpu_power *
3381 min(busiest_load_per_task, max_load);
3382 pwr_now += this->__cpu_power *
3383 min(this_load_per_task, this_load);
3384 pwr_now /= SCHED_LOAD_SCALE;
3386 /* Amount of load we'd subtract */
3387 tmp = sg_div_cpu_power(busiest,
3388 busiest_load_per_task * SCHED_LOAD_SCALE);
3390 pwr_move += busiest->__cpu_power *
3391 min(busiest_load_per_task, max_load - tmp);
3393 /* Amount of load we'd add */
3394 if (max_load * busiest->__cpu_power <
3395 busiest_load_per_task * SCHED_LOAD_SCALE)
3396 tmp = sg_div_cpu_power(this,
3397 max_load * busiest->__cpu_power);
3399 tmp = sg_div_cpu_power(this,
3400 busiest_load_per_task * SCHED_LOAD_SCALE);
3401 pwr_move += this->__cpu_power *
3402 min(this_load_per_task, this_load + tmp);
3403 pwr_move /= SCHED_LOAD_SCALE;
3405 /* Move if we gain throughput */
3406 if (pwr_move > pwr_now)
3407 *imbalance = busiest_load_per_task;
3413 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3414 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3417 if (this == group_leader && group_leader != group_min) {
3418 *imbalance = min_load_per_task;
3428 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3431 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3432 unsigned long imbalance, const cpumask_t *cpus)
3434 struct rq *busiest = NULL, *rq;
3435 unsigned long max_load = 0;
3438 for_each_cpu_mask_nr(i, group->cpumask) {
3441 if (!cpu_isset(i, *cpus))
3445 wl = weighted_cpuload(i);
3447 if (rq->nr_running == 1 && wl > imbalance)
3450 if (wl > max_load) {
3460 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3461 * so long as it is large enough.
3463 #define MAX_PINNED_INTERVAL 512
3466 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3467 * tasks if there is an imbalance.
3469 static int load_balance(int this_cpu, struct rq *this_rq,
3470 struct sched_domain *sd, enum cpu_idle_type idle,
3471 int *balance, cpumask_t *cpus)
3473 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3474 struct sched_group *group;
3475 unsigned long imbalance;
3477 unsigned long flags;
3482 * When power savings policy is enabled for the parent domain, idle
3483 * sibling can pick up load irrespective of busy siblings. In this case,
3484 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3485 * portraying it as CPU_NOT_IDLE.
3487 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3488 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3491 schedstat_inc(sd, lb_count[idle]);
3495 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3502 schedstat_inc(sd, lb_nobusyg[idle]);
3506 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3508 schedstat_inc(sd, lb_nobusyq[idle]);
3512 BUG_ON(busiest == this_rq);
3514 schedstat_add(sd, lb_imbalance[idle], imbalance);
3517 if (busiest->nr_running > 1) {
3519 * Attempt to move tasks. If find_busiest_group has found
3520 * an imbalance but busiest->nr_running <= 1, the group is
3521 * still unbalanced. ld_moved simply stays zero, so it is
3522 * correctly treated as an imbalance.
3524 local_irq_save(flags);
3525 double_rq_lock(this_rq, busiest);
3526 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3527 imbalance, sd, idle, &all_pinned);
3528 double_rq_unlock(this_rq, busiest);
3529 local_irq_restore(flags);
3532 * some other cpu did the load balance for us.
3534 if (ld_moved && this_cpu != smp_processor_id())
3535 resched_cpu(this_cpu);
3537 /* All tasks on this runqueue were pinned by CPU affinity */
3538 if (unlikely(all_pinned)) {
3539 cpu_clear(cpu_of(busiest), *cpus);
3540 if (!cpus_empty(*cpus))
3547 schedstat_inc(sd, lb_failed[idle]);
3548 sd->nr_balance_failed++;
3550 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3552 spin_lock_irqsave(&busiest->lock, flags);
3554 /* don't kick the migration_thread, if the curr
3555 * task on busiest cpu can't be moved to this_cpu
3557 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3558 spin_unlock_irqrestore(&busiest->lock, flags);
3560 goto out_one_pinned;
3563 if (!busiest->active_balance) {
3564 busiest->active_balance = 1;
3565 busiest->push_cpu = this_cpu;
3568 spin_unlock_irqrestore(&busiest->lock, flags);
3570 wake_up_process(busiest->migration_thread);
3573 * We've kicked active balancing, reset the failure
3576 sd->nr_balance_failed = sd->cache_nice_tries+1;
3579 sd->nr_balance_failed = 0;
3581 if (likely(!active_balance)) {
3582 /* We were unbalanced, so reset the balancing interval */
3583 sd->balance_interval = sd->min_interval;
3586 * If we've begun active balancing, start to back off. This
3587 * case may not be covered by the all_pinned logic if there
3588 * is only 1 task on the busy runqueue (because we don't call
3591 if (sd->balance_interval < sd->max_interval)
3592 sd->balance_interval *= 2;
3595 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3596 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3602 schedstat_inc(sd, lb_balanced[idle]);
3604 sd->nr_balance_failed = 0;
3607 /* tune up the balancing interval */
3608 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3609 (sd->balance_interval < sd->max_interval))
3610 sd->balance_interval *= 2;
3612 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3613 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3624 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3625 * tasks if there is an imbalance.
3627 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3628 * this_rq is locked.
3631 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3634 struct sched_group *group;
3635 struct rq *busiest = NULL;
3636 unsigned long imbalance;
3644 * When power savings policy is enabled for the parent domain, idle
3645 * sibling can pick up load irrespective of busy siblings. In this case,
3646 * let the state of idle sibling percolate up as IDLE, instead of
3647 * portraying it as CPU_NOT_IDLE.
3649 if (sd->flags & SD_SHARE_CPUPOWER &&
3650 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3653 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3655 update_shares_locked(this_rq, sd);
3656 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3657 &sd_idle, cpus, NULL);
3659 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3663 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3665 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3669 BUG_ON(busiest == this_rq);
3671 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3674 if (busiest->nr_running > 1) {
3675 /* Attempt to move tasks */
3676 double_lock_balance(this_rq, busiest);
3677 /* this_rq->clock is already updated */
3678 update_rq_clock(busiest);
3679 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3680 imbalance, sd, CPU_NEWLY_IDLE,
3682 double_unlock_balance(this_rq, busiest);
3684 if (unlikely(all_pinned)) {
3685 cpu_clear(cpu_of(busiest), *cpus);
3686 if (!cpus_empty(*cpus))
3692 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3693 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3694 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3697 sd->nr_balance_failed = 0;
3699 update_shares_locked(this_rq, sd);
3703 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3704 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3705 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3707 sd->nr_balance_failed = 0;
3713 * idle_balance is called by schedule() if this_cpu is about to become
3714 * idle. Attempts to pull tasks from other CPUs.
3716 static void idle_balance(int this_cpu, struct rq *this_rq)
3718 struct sched_domain *sd;
3719 int pulled_task = -1;
3720 unsigned long next_balance = jiffies + HZ;
3723 for_each_domain(this_cpu, sd) {
3724 unsigned long interval;
3726 if (!(sd->flags & SD_LOAD_BALANCE))
3729 if (sd->flags & SD_BALANCE_NEWIDLE)
3730 /* If we've pulled tasks over stop searching: */
3731 pulled_task = load_balance_newidle(this_cpu, this_rq,
3734 interval = msecs_to_jiffies(sd->balance_interval);
3735 if (time_after(next_balance, sd->last_balance + interval))
3736 next_balance = sd->last_balance + interval;
3740 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3742 * We are going idle. next_balance may be set based on
3743 * a busy processor. So reset next_balance.
3745 this_rq->next_balance = next_balance;
3750 * active_load_balance is run by migration threads. It pushes running tasks
3751 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3752 * running on each physical CPU where possible, and avoids physical /
3753 * logical imbalances.
3755 * Called with busiest_rq locked.
3757 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3759 int target_cpu = busiest_rq->push_cpu;
3760 struct sched_domain *sd;
3761 struct rq *target_rq;
3763 /* Is there any task to move? */
3764 if (busiest_rq->nr_running <= 1)
3767 target_rq = cpu_rq(target_cpu);
3770 * This condition is "impossible", if it occurs
3771 * we need to fix it. Originally reported by
3772 * Bjorn Helgaas on a 128-cpu setup.
3774 BUG_ON(busiest_rq == target_rq);
3776 /* move a task from busiest_rq to target_rq */
3777 double_lock_balance(busiest_rq, target_rq);
3778 update_rq_clock(busiest_rq);
3779 update_rq_clock(target_rq);
3781 /* Search for an sd spanning us and the target CPU. */
3782 for_each_domain(target_cpu, sd) {
3783 if ((sd->flags & SD_LOAD_BALANCE) &&
3784 cpu_isset(busiest_cpu, sd->span))
3789 schedstat_inc(sd, alb_count);
3791 if (move_one_task(target_rq, target_cpu, busiest_rq,
3793 schedstat_inc(sd, alb_pushed);
3795 schedstat_inc(sd, alb_failed);
3797 double_unlock_balance(busiest_rq, target_rq);
3802 atomic_t load_balancer;
3804 } nohz ____cacheline_aligned = {
3805 .load_balancer = ATOMIC_INIT(-1),
3806 .cpu_mask = CPU_MASK_NONE,
3810 * This routine will try to nominate the ilb (idle load balancing)
3811 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3812 * load balancing on behalf of all those cpus. If all the cpus in the system
3813 * go into this tickless mode, then there will be no ilb owner (as there is
3814 * no need for one) and all the cpus will sleep till the next wakeup event
3817 * For the ilb owner, tick is not stopped. And this tick will be used
3818 * for idle load balancing. ilb owner will still be part of
3821 * While stopping the tick, this cpu will become the ilb owner if there
3822 * is no other owner. And will be the owner till that cpu becomes busy
3823 * or if all cpus in the system stop their ticks at which point
3824 * there is no need for ilb owner.
3826 * When the ilb owner becomes busy, it nominates another owner, during the
3827 * next busy scheduler_tick()
3829 int select_nohz_load_balancer(int stop_tick)
3831 int cpu = smp_processor_id();
3834 cpu_set(cpu, nohz.cpu_mask);
3835 cpu_rq(cpu)->in_nohz_recently = 1;
3838 * If we are going offline and still the leader, give up!
3840 if (!cpu_active(cpu) &&
3841 atomic_read(&nohz.load_balancer) == cpu) {
3842 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3847 /* time for ilb owner also to sleep */
3848 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3849 if (atomic_read(&nohz.load_balancer) == cpu)
3850 atomic_set(&nohz.load_balancer, -1);
3854 if (atomic_read(&nohz.load_balancer) == -1) {
3855 /* make me the ilb owner */
3856 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3858 } else if (atomic_read(&nohz.load_balancer) == cpu)
3861 if (!cpu_isset(cpu, nohz.cpu_mask))
3864 cpu_clear(cpu, nohz.cpu_mask);
3866 if (atomic_read(&nohz.load_balancer) == cpu)
3867 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3874 static DEFINE_SPINLOCK(balancing);
3877 * It checks each scheduling domain to see if it is due to be balanced,
3878 * and initiates a balancing operation if so.
3880 * Balancing parameters are set up in arch_init_sched_domains.
3882 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3885 struct rq *rq = cpu_rq(cpu);
3886 unsigned long interval;
3887 struct sched_domain *sd;
3888 /* Earliest time when we have to do rebalance again */
3889 unsigned long next_balance = jiffies + 60*HZ;
3890 int update_next_balance = 0;
3894 for_each_domain(cpu, sd) {
3895 if (!(sd->flags & SD_LOAD_BALANCE))
3898 interval = sd->balance_interval;
3899 if (idle != CPU_IDLE)
3900 interval *= sd->busy_factor;
3902 /* scale ms to jiffies */
3903 interval = msecs_to_jiffies(interval);
3904 if (unlikely(!interval))
3906 if (interval > HZ*NR_CPUS/10)
3907 interval = HZ*NR_CPUS/10;
3909 need_serialize = sd->flags & SD_SERIALIZE;
3911 if (need_serialize) {
3912 if (!spin_trylock(&balancing))
3916 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3917 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3919 * We've pulled tasks over so either we're no
3920 * longer idle, or one of our SMT siblings is
3923 idle = CPU_NOT_IDLE;
3925 sd->last_balance = jiffies;
3928 spin_unlock(&balancing);
3930 if (time_after(next_balance, sd->last_balance + interval)) {
3931 next_balance = sd->last_balance + interval;
3932 update_next_balance = 1;
3936 * Stop the load balance at this level. There is another
3937 * CPU in our sched group which is doing load balancing more
3945 * next_balance will be updated only when there is a need.
3946 * When the cpu is attached to null domain for ex, it will not be
3949 if (likely(update_next_balance))
3950 rq->next_balance = next_balance;
3954 * run_rebalance_domains is triggered when needed from the scheduler tick.
3955 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3956 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3958 static void run_rebalance_domains(struct softirq_action *h)
3960 int this_cpu = smp_processor_id();
3961 struct rq *this_rq = cpu_rq(this_cpu);
3962 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3963 CPU_IDLE : CPU_NOT_IDLE;
3965 rebalance_domains(this_cpu, idle);
3969 * If this cpu is the owner for idle load balancing, then do the
3970 * balancing on behalf of the other idle cpus whose ticks are
3973 if (this_rq->idle_at_tick &&
3974 atomic_read(&nohz.load_balancer) == this_cpu) {
3975 cpumask_t cpus = nohz.cpu_mask;
3979 cpu_clear(this_cpu, cpus);
3980 for_each_cpu_mask_nr(balance_cpu, cpus) {
3982 * If this cpu gets work to do, stop the load balancing
3983 * work being done for other cpus. Next load
3984 * balancing owner will pick it up.
3989 rebalance_domains(balance_cpu, CPU_IDLE);
3991 rq = cpu_rq(balance_cpu);
3992 if (time_after(this_rq->next_balance, rq->next_balance))
3993 this_rq->next_balance = rq->next_balance;
4000 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4002 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4003 * idle load balancing owner or decide to stop the periodic load balancing,
4004 * if the whole system is idle.
4006 static inline void trigger_load_balance(struct rq *rq, int cpu)
4010 * If we were in the nohz mode recently and busy at the current
4011 * scheduler tick, then check if we need to nominate new idle
4014 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4015 rq->in_nohz_recently = 0;
4017 if (atomic_read(&nohz.load_balancer) == cpu) {
4018 cpu_clear(cpu, nohz.cpu_mask);
4019 atomic_set(&nohz.load_balancer, -1);
4022 if (atomic_read(&nohz.load_balancer) == -1) {
4024 * simple selection for now: Nominate the
4025 * first cpu in the nohz list to be the next
4028 * TBD: Traverse the sched domains and nominate
4029 * the nearest cpu in the nohz.cpu_mask.
4031 int ilb = first_cpu(nohz.cpu_mask);
4033 if (ilb < nr_cpu_ids)
4039 * If this cpu is idle and doing idle load balancing for all the
4040 * cpus with ticks stopped, is it time for that to stop?
4042 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4043 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4049 * If this cpu is idle and the idle load balancing is done by
4050 * someone else, then no need raise the SCHED_SOFTIRQ
4052 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4053 cpu_isset(cpu, nohz.cpu_mask))
4056 if (time_after_eq(jiffies, rq->next_balance))
4057 raise_softirq(SCHED_SOFTIRQ);
4060 #else /* CONFIG_SMP */
4063 * on UP we do not need to balance between CPUs:
4065 static inline void idle_balance(int cpu, struct rq *rq)
4071 DEFINE_PER_CPU(struct kernel_stat, kstat);
4073 EXPORT_PER_CPU_SYMBOL(kstat);
4076 * Return any ns on the sched_clock that have not yet been banked in
4077 * @p in case that task is currently running.
4079 unsigned long long task_delta_exec(struct task_struct *p)
4081 unsigned long flags;
4085 rq = task_rq_lock(p, &flags);
4087 if (task_current(rq, p)) {
4090 update_rq_clock(rq);
4091 delta_exec = rq->clock - p->se.exec_start;
4092 if ((s64)delta_exec > 0)
4096 task_rq_unlock(rq, &flags);
4102 * Account user cpu time to a process.
4103 * @p: the process that the cpu time gets accounted to
4104 * @cputime: the cpu time spent in user space since the last update
4106 void account_user_time(struct task_struct *p, cputime_t cputime)
4108 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4111 p->utime = cputime_add(p->utime, cputime);
4112 account_group_user_time(p, cputime);
4114 /* Add user time to cpustat. */
4115 tmp = cputime_to_cputime64(cputime);
4116 if (TASK_NICE(p) > 0)
4117 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4119 cpustat->user = cputime64_add(cpustat->user, tmp);
4120 /* Account for user time used */
4121 acct_update_integrals(p);
4125 * Account guest cpu time to a process.
4126 * @p: the process that the cpu time gets accounted to
4127 * @cputime: the cpu time spent in virtual machine since the last update
4129 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4132 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4134 tmp = cputime_to_cputime64(cputime);
4136 p->utime = cputime_add(p->utime, cputime);
4137 account_group_user_time(p, cputime);
4138 p->gtime = cputime_add(p->gtime, cputime);
4140 cpustat->user = cputime64_add(cpustat->user, tmp);
4141 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4145 * Account scaled user cpu time to a process.
4146 * @p: the process that the cpu time gets accounted to
4147 * @cputime: the cpu time spent in user space since the last update
4149 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4151 p->utimescaled = cputime_add(p->utimescaled, cputime);
4155 * Account system cpu time to a process.
4156 * @p: the process that the cpu time gets accounted to
4157 * @hardirq_offset: the offset to subtract from hardirq_count()
4158 * @cputime: the cpu time spent in kernel space since the last update
4160 void account_system_time(struct task_struct *p, int hardirq_offset,
4163 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4164 struct rq *rq = this_rq();
4167 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4168 account_guest_time(p, cputime);
4172 p->stime = cputime_add(p->stime, cputime);
4173 account_group_system_time(p, cputime);
4175 /* Add system time to cpustat. */
4176 tmp = cputime_to_cputime64(cputime);
4177 if (hardirq_count() - hardirq_offset)
4178 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4179 else if (softirq_count())
4180 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4181 else if (p != rq->idle)
4182 cpustat->system = cputime64_add(cpustat->system, tmp);
4183 else if (atomic_read(&rq->nr_iowait) > 0)
4184 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4186 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4187 /* Account for system time used */
4188 acct_update_integrals(p);
4192 * Account scaled system cpu time to a process.
4193 * @p: the process that the cpu time gets accounted to
4194 * @hardirq_offset: the offset to subtract from hardirq_count()
4195 * @cputime: the cpu time spent in kernel space since the last update
4197 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4199 p->stimescaled = cputime_add(p->stimescaled, cputime);
4203 * Account for involuntary wait time.
4204 * @p: the process from which the cpu time has been stolen
4205 * @steal: the cpu time spent in involuntary wait
4207 void account_steal_time(struct task_struct *p, cputime_t steal)
4209 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4210 cputime64_t tmp = cputime_to_cputime64(steal);
4211 struct rq *rq = this_rq();
4213 if (p == rq->idle) {
4214 p->stime = cputime_add(p->stime, steal);
4215 account_group_system_time(p, steal);
4216 if (atomic_read(&rq->nr_iowait) > 0)
4217 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4219 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4221 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4225 * Use precise platform statistics if available:
4227 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4228 cputime_t task_utime(struct task_struct *p)
4233 cputime_t task_stime(struct task_struct *p)
4238 cputime_t task_utime(struct task_struct *p)
4240 clock_t utime = cputime_to_clock_t(p->utime),
4241 total = utime + cputime_to_clock_t(p->stime);
4245 * Use CFS's precise accounting:
4247 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4251 do_div(temp, total);
4253 utime = (clock_t)temp;
4255 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4256 return p->prev_utime;
4259 cputime_t task_stime(struct task_struct *p)
4264 * Use CFS's precise accounting. (we subtract utime from
4265 * the total, to make sure the total observed by userspace
4266 * grows monotonically - apps rely on that):
4268 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4269 cputime_to_clock_t(task_utime(p));
4272 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4274 return p->prev_stime;
4278 inline cputime_t task_gtime(struct task_struct *p)
4284 * This function gets called by the timer code, with HZ frequency.
4285 * We call it with interrupts disabled.
4287 * It also gets called by the fork code, when changing the parent's
4290 void scheduler_tick(void)
4292 int cpu = smp_processor_id();
4293 struct rq *rq = cpu_rq(cpu);
4294 struct task_struct *curr = rq->curr;
4298 spin_lock(&rq->lock);
4299 update_rq_clock(rq);
4300 update_cpu_load(rq);
4301 curr->sched_class->task_tick(rq, curr, 0);
4302 spin_unlock(&rq->lock);
4305 rq->idle_at_tick = idle_cpu(cpu);
4306 trigger_load_balance(rq, cpu);
4310 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4311 defined(CONFIG_PREEMPT_TRACER))
4313 static inline unsigned long get_parent_ip(unsigned long addr)
4315 if (in_lock_functions(addr)) {
4316 addr = CALLER_ADDR2;
4317 if (in_lock_functions(addr))
4318 addr = CALLER_ADDR3;
4323 void __kprobes add_preempt_count(int val)
4325 #ifdef CONFIG_DEBUG_PREEMPT
4329 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4332 preempt_count() += val;
4333 #ifdef CONFIG_DEBUG_PREEMPT
4335 * Spinlock count overflowing soon?
4337 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4340 if (preempt_count() == val)
4341 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4343 EXPORT_SYMBOL(add_preempt_count);
4345 void __kprobes sub_preempt_count(int val)
4347 #ifdef CONFIG_DEBUG_PREEMPT
4351 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4354 * Is the spinlock portion underflowing?
4356 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4357 !(preempt_count() & PREEMPT_MASK)))
4361 if (preempt_count() == val)
4362 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4363 preempt_count() -= val;
4365 EXPORT_SYMBOL(sub_preempt_count);
4370 * Print scheduling while atomic bug:
4372 static noinline void __schedule_bug(struct task_struct *prev)
4374 struct pt_regs *regs = get_irq_regs();
4376 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4377 prev->comm, prev->pid, preempt_count());
4379 debug_show_held_locks(prev);
4381 if (irqs_disabled())
4382 print_irqtrace_events(prev);
4391 * Various schedule()-time debugging checks and statistics:
4393 static inline void schedule_debug(struct task_struct *prev)
4396 * Test if we are atomic. Since do_exit() needs to call into
4397 * schedule() atomically, we ignore that path for now.
4398 * Otherwise, whine if we are scheduling when we should not be.
4400 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4401 __schedule_bug(prev);
4403 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4405 schedstat_inc(this_rq(), sched_count);
4406 #ifdef CONFIG_SCHEDSTATS
4407 if (unlikely(prev->lock_depth >= 0)) {
4408 schedstat_inc(this_rq(), bkl_count);
4409 schedstat_inc(prev, sched_info.bkl_count);
4415 * Pick up the highest-prio task:
4417 static inline struct task_struct *
4418 pick_next_task(struct rq *rq, struct task_struct *prev)
4420 const struct sched_class *class;
4421 struct task_struct *p;
4424 * Optimization: we know that if all tasks are in
4425 * the fair class we can call that function directly:
4427 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4428 p = fair_sched_class.pick_next_task(rq);
4433 class = sched_class_highest;
4435 p = class->pick_next_task(rq);
4439 * Will never be NULL as the idle class always
4440 * returns a non-NULL p:
4442 class = class->next;
4447 * schedule() is the main scheduler function.
4449 asmlinkage void __sched schedule(void)
4451 struct task_struct *prev, *next;
4452 unsigned long *switch_count;
4458 cpu = smp_processor_id();
4462 switch_count = &prev->nivcsw;
4464 release_kernel_lock(prev);
4465 need_resched_nonpreemptible:
4467 schedule_debug(prev);
4469 if (sched_feat(HRTICK))
4472 spin_lock_irq(&rq->lock);
4473 update_rq_clock(rq);
4474 clear_tsk_need_resched(prev);
4476 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4477 if (unlikely(signal_pending_state(prev->state, prev)))
4478 prev->state = TASK_RUNNING;
4480 deactivate_task(rq, prev, 1);
4481 switch_count = &prev->nvcsw;
4485 if (prev->sched_class->pre_schedule)
4486 prev->sched_class->pre_schedule(rq, prev);
4489 if (unlikely(!rq->nr_running))
4490 idle_balance(cpu, rq);
4492 prev->sched_class->put_prev_task(rq, prev);
4493 next = pick_next_task(rq, prev);
4495 if (likely(prev != next)) {
4496 sched_info_switch(prev, next);
4502 context_switch(rq, prev, next); /* unlocks the rq */
4504 * the context switch might have flipped the stack from under
4505 * us, hence refresh the local variables.
4507 cpu = smp_processor_id();
4510 spin_unlock_irq(&rq->lock);
4512 if (unlikely(reacquire_kernel_lock(current) < 0))
4513 goto need_resched_nonpreemptible;
4515 preempt_enable_no_resched();
4516 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4519 EXPORT_SYMBOL(schedule);
4521 #ifdef CONFIG_PREEMPT
4523 * this is the entry point to schedule() from in-kernel preemption
4524 * off of preempt_enable. Kernel preemptions off return from interrupt
4525 * occur there and call schedule directly.
4527 asmlinkage void __sched preempt_schedule(void)
4529 struct thread_info *ti = current_thread_info();
4532 * If there is a non-zero preempt_count or interrupts are disabled,
4533 * we do not want to preempt the current task. Just return..
4535 if (likely(ti->preempt_count || irqs_disabled()))
4539 add_preempt_count(PREEMPT_ACTIVE);
4541 sub_preempt_count(PREEMPT_ACTIVE);
4544 * Check again in case we missed a preemption opportunity
4545 * between schedule and now.
4548 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4550 EXPORT_SYMBOL(preempt_schedule);
4553 * this is the entry point to schedule() from kernel preemption
4554 * off of irq context.
4555 * Note, that this is called and return with irqs disabled. This will
4556 * protect us against recursive calling from irq.
4558 asmlinkage void __sched preempt_schedule_irq(void)
4560 struct thread_info *ti = current_thread_info();
4562 /* Catch callers which need to be fixed */
4563 BUG_ON(ti->preempt_count || !irqs_disabled());
4566 add_preempt_count(PREEMPT_ACTIVE);
4569 local_irq_disable();
4570 sub_preempt_count(PREEMPT_ACTIVE);
4573 * Check again in case we missed a preemption opportunity
4574 * between schedule and now.
4577 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4580 #endif /* CONFIG_PREEMPT */
4582 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4585 return try_to_wake_up(curr->private, mode, sync);
4587 EXPORT_SYMBOL(default_wake_function);
4590 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4591 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4592 * number) then we wake all the non-exclusive tasks and one exclusive task.
4594 * There are circumstances in which we can try to wake a task which has already
4595 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4596 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4598 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4599 int nr_exclusive, int sync, void *key)
4601 wait_queue_t *curr, *next;
4603 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4604 unsigned flags = curr->flags;
4606 if (curr->func(curr, mode, sync, key) &&
4607 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4613 * __wake_up - wake up threads blocked on a waitqueue.
4615 * @mode: which threads
4616 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4617 * @key: is directly passed to the wakeup function
4619 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4620 int nr_exclusive, void *key)
4622 unsigned long flags;
4624 spin_lock_irqsave(&q->lock, flags);
4625 __wake_up_common(q, mode, nr_exclusive, 0, key);
4626 spin_unlock_irqrestore(&q->lock, flags);
4628 EXPORT_SYMBOL(__wake_up);
4631 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4633 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4635 __wake_up_common(q, mode, 1, 0, NULL);
4639 * __wake_up_sync - wake up threads blocked on a waitqueue.
4641 * @mode: which threads
4642 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4644 * The sync wakeup differs that the waker knows that it will schedule
4645 * away soon, so while the target thread will be woken up, it will not
4646 * be migrated to another CPU - ie. the two threads are 'synchronized'
4647 * with each other. This can prevent needless bouncing between CPUs.
4649 * On UP it can prevent extra preemption.
4652 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4654 unsigned long flags;
4660 if (unlikely(!nr_exclusive))
4663 spin_lock_irqsave(&q->lock, flags);
4664 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4665 spin_unlock_irqrestore(&q->lock, flags);
4667 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4670 * complete: - signals a single thread waiting on this completion
4671 * @x: holds the state of this particular completion
4673 * This will wake up a single thread waiting on this completion. Threads will be
4674 * awakened in the same order in which they were queued.
4676 * See also complete_all(), wait_for_completion() and related routines.
4678 void complete(struct completion *x)
4680 unsigned long flags;
4682 spin_lock_irqsave(&x->wait.lock, flags);
4684 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4685 spin_unlock_irqrestore(&x->wait.lock, flags);
4687 EXPORT_SYMBOL(complete);
4690 * complete_all: - signals all threads waiting on this completion
4691 * @x: holds the state of this particular completion
4693 * This will wake up all threads waiting on this particular completion event.
4695 void complete_all(struct completion *x)
4697 unsigned long flags;
4699 spin_lock_irqsave(&x->wait.lock, flags);
4700 x->done += UINT_MAX/2;
4701 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4702 spin_unlock_irqrestore(&x->wait.lock, flags);
4704 EXPORT_SYMBOL(complete_all);
4706 static inline long __sched
4707 do_wait_for_common(struct completion *x, long timeout, int state)
4710 DECLARE_WAITQUEUE(wait, current);
4712 wait.flags |= WQ_FLAG_EXCLUSIVE;
4713 __add_wait_queue_tail(&x->wait, &wait);
4715 if (signal_pending_state(state, current)) {
4716 timeout = -ERESTARTSYS;
4719 __set_current_state(state);
4720 spin_unlock_irq(&x->wait.lock);
4721 timeout = schedule_timeout(timeout);
4722 spin_lock_irq(&x->wait.lock);
4723 } while (!x->done && timeout);
4724 __remove_wait_queue(&x->wait, &wait);
4729 return timeout ?: 1;
4733 wait_for_common(struct completion *x, long timeout, int state)
4737 spin_lock_irq(&x->wait.lock);
4738 timeout = do_wait_for_common(x, timeout, state);
4739 spin_unlock_irq(&x->wait.lock);
4744 * wait_for_completion: - waits for completion of a task
4745 * @x: holds the state of this particular completion
4747 * This waits to be signaled for completion of a specific task. It is NOT
4748 * interruptible and there is no timeout.
4750 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4751 * and interrupt capability. Also see complete().
4753 void __sched wait_for_completion(struct completion *x)
4755 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4757 EXPORT_SYMBOL(wait_for_completion);
4760 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4761 * @x: holds the state of this particular completion
4762 * @timeout: timeout value in jiffies
4764 * This waits for either a completion of a specific task to be signaled or for a
4765 * specified timeout to expire. The timeout is in jiffies. It is not
4768 unsigned long __sched
4769 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4771 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4773 EXPORT_SYMBOL(wait_for_completion_timeout);
4776 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4777 * @x: holds the state of this particular completion
4779 * This waits for completion of a specific task to be signaled. It is
4782 int __sched wait_for_completion_interruptible(struct completion *x)
4784 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4785 if (t == -ERESTARTSYS)
4789 EXPORT_SYMBOL(wait_for_completion_interruptible);
4792 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4793 * @x: holds the state of this particular completion
4794 * @timeout: timeout value in jiffies
4796 * This waits for either a completion of a specific task to be signaled or for a
4797 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4799 unsigned long __sched
4800 wait_for_completion_interruptible_timeout(struct completion *x,
4801 unsigned long timeout)
4803 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4805 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4808 * wait_for_completion_killable: - waits for completion of a task (killable)
4809 * @x: holds the state of this particular completion
4811 * This waits to be signaled for completion of a specific task. It can be
4812 * interrupted by a kill signal.
4814 int __sched wait_for_completion_killable(struct completion *x)
4816 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4817 if (t == -ERESTARTSYS)
4821 EXPORT_SYMBOL(wait_for_completion_killable);
4824 * try_wait_for_completion - try to decrement a completion without blocking
4825 * @x: completion structure
4827 * Returns: 0 if a decrement cannot be done without blocking
4828 * 1 if a decrement succeeded.
4830 * If a completion is being used as a counting completion,
4831 * attempt to decrement the counter without blocking. This
4832 * enables us to avoid waiting if the resource the completion
4833 * is protecting is not available.
4835 bool try_wait_for_completion(struct completion *x)
4839 spin_lock_irq(&x->wait.lock);
4844 spin_unlock_irq(&x->wait.lock);
4847 EXPORT_SYMBOL(try_wait_for_completion);
4850 * completion_done - Test to see if a completion has any waiters
4851 * @x: completion structure
4853 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4854 * 1 if there are no waiters.
4857 bool completion_done(struct completion *x)
4861 spin_lock_irq(&x->wait.lock);
4864 spin_unlock_irq(&x->wait.lock);
4867 EXPORT_SYMBOL(completion_done);
4870 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4872 unsigned long flags;
4875 init_waitqueue_entry(&wait, current);
4877 __set_current_state(state);
4879 spin_lock_irqsave(&q->lock, flags);
4880 __add_wait_queue(q, &wait);
4881 spin_unlock(&q->lock);
4882 timeout = schedule_timeout(timeout);
4883 spin_lock_irq(&q->lock);
4884 __remove_wait_queue(q, &wait);
4885 spin_unlock_irqrestore(&q->lock, flags);
4890 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4892 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4894 EXPORT_SYMBOL(interruptible_sleep_on);
4897 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4899 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4901 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4903 void __sched sleep_on(wait_queue_head_t *q)
4905 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4907 EXPORT_SYMBOL(sleep_on);
4909 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4911 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4913 EXPORT_SYMBOL(sleep_on_timeout);
4915 #ifdef CONFIG_RT_MUTEXES
4918 * rt_mutex_setprio - set the current priority of a task
4920 * @prio: prio value (kernel-internal form)
4922 * This function changes the 'effective' priority of a task. It does
4923 * not touch ->normal_prio like __setscheduler().
4925 * Used by the rt_mutex code to implement priority inheritance logic.
4927 void rt_mutex_setprio(struct task_struct *p, int prio)
4929 unsigned long flags;
4930 int oldprio, on_rq, running;
4932 const struct sched_class *prev_class = p->sched_class;
4934 BUG_ON(prio < 0 || prio > MAX_PRIO);
4936 rq = task_rq_lock(p, &flags);
4937 update_rq_clock(rq);
4940 on_rq = p->se.on_rq;
4941 running = task_current(rq, p);
4943 dequeue_task(rq, p, 0);
4945 p->sched_class->put_prev_task(rq, p);
4948 p->sched_class = &rt_sched_class;
4950 p->sched_class = &fair_sched_class;
4955 p->sched_class->set_curr_task(rq);
4957 enqueue_task(rq, p, 0);
4959 check_class_changed(rq, p, prev_class, oldprio, running);
4961 task_rq_unlock(rq, &flags);
4966 void set_user_nice(struct task_struct *p, long nice)
4968 int old_prio, delta, on_rq;
4969 unsigned long flags;
4972 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4975 * We have to be careful, if called from sys_setpriority(),
4976 * the task might be in the middle of scheduling on another CPU.
4978 rq = task_rq_lock(p, &flags);
4979 update_rq_clock(rq);
4981 * The RT priorities are set via sched_setscheduler(), but we still
4982 * allow the 'normal' nice value to be set - but as expected
4983 * it wont have any effect on scheduling until the task is
4984 * SCHED_FIFO/SCHED_RR:
4986 if (task_has_rt_policy(p)) {
4987 p->static_prio = NICE_TO_PRIO(nice);
4990 on_rq = p->se.on_rq;
4992 dequeue_task(rq, p, 0);
4994 p->static_prio = NICE_TO_PRIO(nice);
4997 p->prio = effective_prio(p);
4998 delta = p->prio - old_prio;
5001 enqueue_task(rq, p, 0);
5003 * If the task increased its priority or is running and
5004 * lowered its priority, then reschedule its CPU:
5006 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5007 resched_task(rq->curr);
5010 task_rq_unlock(rq, &flags);
5012 EXPORT_SYMBOL(set_user_nice);
5015 * can_nice - check if a task can reduce its nice value
5019 int can_nice(const struct task_struct *p, const int nice)
5021 /* convert nice value [19,-20] to rlimit style value [1,40] */
5022 int nice_rlim = 20 - nice;
5024 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5025 capable(CAP_SYS_NICE));
5028 #ifdef __ARCH_WANT_SYS_NICE
5031 * sys_nice - change the priority of the current process.
5032 * @increment: priority increment
5034 * sys_setpriority is a more generic, but much slower function that
5035 * does similar things.
5037 asmlinkage long sys_nice(int increment)
5042 * Setpriority might change our priority at the same moment.
5043 * We don't have to worry. Conceptually one call occurs first
5044 * and we have a single winner.
5046 if (increment < -40)
5051 nice = PRIO_TO_NICE(current->static_prio) + increment;
5057 if (increment < 0 && !can_nice(current, nice))
5060 retval = security_task_setnice(current, nice);
5064 set_user_nice(current, nice);
5071 * task_prio - return the priority value of a given task.
5072 * @p: the task in question.
5074 * This is the priority value as seen by users in /proc.
5075 * RT tasks are offset by -200. Normal tasks are centered
5076 * around 0, value goes from -16 to +15.
5078 int task_prio(const struct task_struct *p)
5080 return p->prio - MAX_RT_PRIO;
5084 * task_nice - return the nice value of a given task.
5085 * @p: the task in question.
5087 int task_nice(const struct task_struct *p)
5089 return TASK_NICE(p);
5091 EXPORT_SYMBOL(task_nice);
5094 * idle_cpu - is a given cpu idle currently?
5095 * @cpu: the processor in question.
5097 int idle_cpu(int cpu)
5099 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5103 * idle_task - return the idle task for a given cpu.
5104 * @cpu: the processor in question.
5106 struct task_struct *idle_task(int cpu)
5108 return cpu_rq(cpu)->idle;
5112 * find_process_by_pid - find a process with a matching PID value.
5113 * @pid: the pid in question.
5115 static struct task_struct *find_process_by_pid(pid_t pid)
5117 return pid ? find_task_by_vpid(pid) : current;
5120 /* Actually do priority change: must hold rq lock. */
5122 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5124 BUG_ON(p->se.on_rq);
5127 switch (p->policy) {
5131 p->sched_class = &fair_sched_class;
5135 p->sched_class = &rt_sched_class;
5139 p->rt_priority = prio;
5140 p->normal_prio = normal_prio(p);
5141 /* we are holding p->pi_lock already */
5142 p->prio = rt_mutex_getprio(p);
5147 * check the target process has a UID that matches the current process's
5149 static bool check_same_owner(struct task_struct *p)
5151 const struct cred *cred = current_cred(), *pcred;
5155 pcred = __task_cred(p);
5156 match = (cred->euid == pcred->euid ||
5157 cred->euid == pcred->uid);
5162 static int __sched_setscheduler(struct task_struct *p, int policy,
5163 struct sched_param *param, bool user)
5165 int retval, oldprio, oldpolicy = -1, on_rq, running;
5166 unsigned long flags;
5167 const struct sched_class *prev_class = p->sched_class;
5170 /* may grab non-irq protected spin_locks */
5171 BUG_ON(in_interrupt());
5173 /* double check policy once rq lock held */
5175 policy = oldpolicy = p->policy;
5176 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5177 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5178 policy != SCHED_IDLE)
5181 * Valid priorities for SCHED_FIFO and SCHED_RR are
5182 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5183 * SCHED_BATCH and SCHED_IDLE is 0.
5185 if (param->sched_priority < 0 ||
5186 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5187 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5189 if (rt_policy(policy) != (param->sched_priority != 0))
5193 * Allow unprivileged RT tasks to decrease priority:
5195 if (user && !capable(CAP_SYS_NICE)) {
5196 if (rt_policy(policy)) {
5197 unsigned long rlim_rtprio;
5199 if (!lock_task_sighand(p, &flags))
5201 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5202 unlock_task_sighand(p, &flags);
5204 /* can't set/change the rt policy */
5205 if (policy != p->policy && !rlim_rtprio)
5208 /* can't increase priority */
5209 if (param->sched_priority > p->rt_priority &&
5210 param->sched_priority > rlim_rtprio)
5214 * Like positive nice levels, dont allow tasks to
5215 * move out of SCHED_IDLE either:
5217 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5220 /* can't change other user's priorities */
5221 if (!check_same_owner(p))
5226 #ifdef CONFIG_RT_GROUP_SCHED
5228 * Do not allow realtime tasks into groups that have no runtime
5231 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5232 task_group(p)->rt_bandwidth.rt_runtime == 0)
5236 retval = security_task_setscheduler(p, policy, param);
5242 * make sure no PI-waiters arrive (or leave) while we are
5243 * changing the priority of the task:
5245 spin_lock_irqsave(&p->pi_lock, flags);
5247 * To be able to change p->policy safely, the apropriate
5248 * runqueue lock must be held.
5250 rq = __task_rq_lock(p);
5251 /* recheck policy now with rq lock held */
5252 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5253 policy = oldpolicy = -1;
5254 __task_rq_unlock(rq);
5255 spin_unlock_irqrestore(&p->pi_lock, flags);
5258 update_rq_clock(rq);
5259 on_rq = p->se.on_rq;
5260 running = task_current(rq, p);
5262 deactivate_task(rq, p, 0);
5264 p->sched_class->put_prev_task(rq, p);
5267 __setscheduler(rq, p, policy, param->sched_priority);
5270 p->sched_class->set_curr_task(rq);
5272 activate_task(rq, p, 0);
5274 check_class_changed(rq, p, prev_class, oldprio, running);
5276 __task_rq_unlock(rq);
5277 spin_unlock_irqrestore(&p->pi_lock, flags);
5279 rt_mutex_adjust_pi(p);
5285 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5286 * @p: the task in question.
5287 * @policy: new policy.
5288 * @param: structure containing the new RT priority.
5290 * NOTE that the task may be already dead.
5292 int sched_setscheduler(struct task_struct *p, int policy,
5293 struct sched_param *param)
5295 return __sched_setscheduler(p, policy, param, true);
5297 EXPORT_SYMBOL_GPL(sched_setscheduler);
5300 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5301 * @p: the task in question.
5302 * @policy: new policy.
5303 * @param: structure containing the new RT priority.
5305 * Just like sched_setscheduler, only don't bother checking if the
5306 * current context has permission. For example, this is needed in
5307 * stop_machine(): we create temporary high priority worker threads,
5308 * but our caller might not have that capability.
5310 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5311 struct sched_param *param)
5313 return __sched_setscheduler(p, policy, param, false);
5317 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5319 struct sched_param lparam;
5320 struct task_struct *p;
5323 if (!param || pid < 0)
5325 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5330 p = find_process_by_pid(pid);
5332 retval = sched_setscheduler(p, policy, &lparam);
5339 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5340 * @pid: the pid in question.
5341 * @policy: new policy.
5342 * @param: structure containing the new RT priority.
5345 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5347 /* negative values for policy are not valid */
5351 return do_sched_setscheduler(pid, policy, param);
5355 * sys_sched_setparam - set/change the RT priority of a thread
5356 * @pid: the pid in question.
5357 * @param: structure containing the new RT priority.
5359 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5361 return do_sched_setscheduler(pid, -1, param);
5365 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5366 * @pid: the pid in question.
5368 asmlinkage long sys_sched_getscheduler(pid_t pid)
5370 struct task_struct *p;
5377 read_lock(&tasklist_lock);
5378 p = find_process_by_pid(pid);
5380 retval = security_task_getscheduler(p);
5384 read_unlock(&tasklist_lock);
5389 * sys_sched_getscheduler - get the RT priority of a thread
5390 * @pid: the pid in question.
5391 * @param: structure containing the RT priority.
5393 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5395 struct sched_param lp;
5396 struct task_struct *p;
5399 if (!param || pid < 0)
5402 read_lock(&tasklist_lock);
5403 p = find_process_by_pid(pid);
5408 retval = security_task_getscheduler(p);
5412 lp.sched_priority = p->rt_priority;
5413 read_unlock(&tasklist_lock);
5416 * This one might sleep, we cannot do it with a spinlock held ...
5418 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5423 read_unlock(&tasklist_lock);
5427 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5429 cpumask_t cpus_allowed;
5430 cpumask_t new_mask = *in_mask;
5431 struct task_struct *p;
5435 read_lock(&tasklist_lock);
5437 p = find_process_by_pid(pid);
5439 read_unlock(&tasklist_lock);
5445 * It is not safe to call set_cpus_allowed with the
5446 * tasklist_lock held. We will bump the task_struct's
5447 * usage count and then drop tasklist_lock.
5450 read_unlock(&tasklist_lock);
5453 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5456 retval = security_task_setscheduler(p, 0, NULL);
5460 cpuset_cpus_allowed(p, &cpus_allowed);
5461 cpus_and(new_mask, new_mask, cpus_allowed);
5463 retval = set_cpus_allowed_ptr(p, &new_mask);
5466 cpuset_cpus_allowed(p, &cpus_allowed);
5467 if (!cpus_subset(new_mask, cpus_allowed)) {
5469 * We must have raced with a concurrent cpuset
5470 * update. Just reset the cpus_allowed to the
5471 * cpuset's cpus_allowed
5473 new_mask = cpus_allowed;
5483 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5484 cpumask_t *new_mask)
5486 if (len < sizeof(cpumask_t)) {
5487 memset(new_mask, 0, sizeof(cpumask_t));
5488 } else if (len > sizeof(cpumask_t)) {
5489 len = sizeof(cpumask_t);
5491 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5495 * sys_sched_setaffinity - set the cpu affinity of a process
5496 * @pid: pid of the process
5497 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5498 * @user_mask_ptr: user-space pointer to the new cpu mask
5500 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5501 unsigned long __user *user_mask_ptr)
5506 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5510 return sched_setaffinity(pid, &new_mask);
5513 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5515 struct task_struct *p;
5519 read_lock(&tasklist_lock);
5522 p = find_process_by_pid(pid);
5526 retval = security_task_getscheduler(p);
5530 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5533 read_unlock(&tasklist_lock);
5540 * sys_sched_getaffinity - get the cpu affinity of a process
5541 * @pid: pid of the process
5542 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5543 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5545 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5546 unsigned long __user *user_mask_ptr)
5551 if (len < sizeof(cpumask_t))
5554 ret = sched_getaffinity(pid, &mask);
5558 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5561 return sizeof(cpumask_t);
5565 * sys_sched_yield - yield the current processor to other threads.
5567 * This function yields the current CPU to other tasks. If there are no
5568 * other threads running on this CPU then this function will return.
5570 asmlinkage long sys_sched_yield(void)
5572 struct rq *rq = this_rq_lock();
5574 schedstat_inc(rq, yld_count);
5575 current->sched_class->yield_task(rq);
5578 * Since we are going to call schedule() anyway, there's
5579 * no need to preempt or enable interrupts:
5581 __release(rq->lock);
5582 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5583 _raw_spin_unlock(&rq->lock);
5584 preempt_enable_no_resched();
5591 static void __cond_resched(void)
5593 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5594 __might_sleep(__FILE__, __LINE__);
5597 * The BKS might be reacquired before we have dropped
5598 * PREEMPT_ACTIVE, which could trigger a second
5599 * cond_resched() call.
5602 add_preempt_count(PREEMPT_ACTIVE);
5604 sub_preempt_count(PREEMPT_ACTIVE);
5605 } while (need_resched());
5608 int __sched _cond_resched(void)
5610 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5611 system_state == SYSTEM_RUNNING) {
5617 EXPORT_SYMBOL(_cond_resched);
5620 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5621 * call schedule, and on return reacquire the lock.
5623 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5624 * operations here to prevent schedule() from being called twice (once via
5625 * spin_unlock(), once by hand).
5627 int cond_resched_lock(spinlock_t *lock)
5629 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5632 if (spin_needbreak(lock) || resched) {
5634 if (resched && need_resched())
5643 EXPORT_SYMBOL(cond_resched_lock);
5645 int __sched cond_resched_softirq(void)
5647 BUG_ON(!in_softirq());
5649 if (need_resched() && system_state == SYSTEM_RUNNING) {
5657 EXPORT_SYMBOL(cond_resched_softirq);
5660 * yield - yield the current processor to other threads.
5662 * This is a shortcut for kernel-space yielding - it marks the
5663 * thread runnable and calls sys_sched_yield().
5665 void __sched yield(void)
5667 set_current_state(TASK_RUNNING);
5670 EXPORT_SYMBOL(yield);
5673 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5674 * that process accounting knows that this is a task in IO wait state.
5676 * But don't do that if it is a deliberate, throttling IO wait (this task
5677 * has set its backing_dev_info: the queue against which it should throttle)
5679 void __sched io_schedule(void)
5681 struct rq *rq = &__raw_get_cpu_var(runqueues);
5683 delayacct_blkio_start();
5684 atomic_inc(&rq->nr_iowait);
5686 atomic_dec(&rq->nr_iowait);
5687 delayacct_blkio_end();
5689 EXPORT_SYMBOL(io_schedule);
5691 long __sched io_schedule_timeout(long timeout)
5693 struct rq *rq = &__raw_get_cpu_var(runqueues);
5696 delayacct_blkio_start();
5697 atomic_inc(&rq->nr_iowait);
5698 ret = schedule_timeout(timeout);
5699 atomic_dec(&rq->nr_iowait);
5700 delayacct_blkio_end();
5705 * sys_sched_get_priority_max - return maximum RT priority.
5706 * @policy: scheduling class.
5708 * this syscall returns the maximum rt_priority that can be used
5709 * by a given scheduling class.
5711 asmlinkage long sys_sched_get_priority_max(int policy)
5718 ret = MAX_USER_RT_PRIO-1;
5730 * sys_sched_get_priority_min - return minimum RT priority.
5731 * @policy: scheduling class.
5733 * this syscall returns the minimum rt_priority that can be used
5734 * by a given scheduling class.
5736 asmlinkage long sys_sched_get_priority_min(int policy)
5754 * sys_sched_rr_get_interval - return the default timeslice of a process.
5755 * @pid: pid of the process.
5756 * @interval: userspace pointer to the timeslice value.
5758 * this syscall writes the default timeslice value of a given process
5759 * into the user-space timespec buffer. A value of '0' means infinity.
5762 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5764 struct task_struct *p;
5765 unsigned int time_slice;
5773 read_lock(&tasklist_lock);
5774 p = find_process_by_pid(pid);
5778 retval = security_task_getscheduler(p);
5783 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5784 * tasks that are on an otherwise idle runqueue:
5787 if (p->policy == SCHED_RR) {
5788 time_slice = DEF_TIMESLICE;
5789 } else if (p->policy != SCHED_FIFO) {
5790 struct sched_entity *se = &p->se;
5791 unsigned long flags;
5794 rq = task_rq_lock(p, &flags);
5795 if (rq->cfs.load.weight)
5796 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5797 task_rq_unlock(rq, &flags);
5799 read_unlock(&tasklist_lock);
5800 jiffies_to_timespec(time_slice, &t);
5801 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5805 read_unlock(&tasklist_lock);
5809 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5811 void sched_show_task(struct task_struct *p)
5813 unsigned long free = 0;
5816 state = p->state ? __ffs(p->state) + 1 : 0;
5817 printk(KERN_INFO "%-13.13s %c", p->comm,
5818 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5819 #if BITS_PER_LONG == 32
5820 if (state == TASK_RUNNING)
5821 printk(KERN_CONT " running ");
5823 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5825 if (state == TASK_RUNNING)
5826 printk(KERN_CONT " running task ");
5828 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5830 #ifdef CONFIG_DEBUG_STACK_USAGE
5832 unsigned long *n = end_of_stack(p);
5835 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5838 printk(KERN_CONT "%5lu %5d %6d\n", free,
5839 task_pid_nr(p), task_pid_nr(p->real_parent));
5841 show_stack(p, NULL);
5844 void show_state_filter(unsigned long state_filter)
5846 struct task_struct *g, *p;
5848 #if BITS_PER_LONG == 32
5850 " task PC stack pid father\n");
5853 " task PC stack pid father\n");
5855 read_lock(&tasklist_lock);
5856 do_each_thread(g, p) {
5858 * reset the NMI-timeout, listing all files on a slow
5859 * console might take alot of time:
5861 touch_nmi_watchdog();
5862 if (!state_filter || (p->state & state_filter))
5864 } while_each_thread(g, p);
5866 touch_all_softlockup_watchdogs();
5868 #ifdef CONFIG_SCHED_DEBUG
5869 sysrq_sched_debug_show();
5871 read_unlock(&tasklist_lock);
5873 * Only show locks if all tasks are dumped:
5875 if (state_filter == -1)
5876 debug_show_all_locks();
5879 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5881 idle->sched_class = &idle_sched_class;
5885 * init_idle - set up an idle thread for a given CPU
5886 * @idle: task in question
5887 * @cpu: cpu the idle task belongs to
5889 * NOTE: this function does not set the idle thread's NEED_RESCHED
5890 * flag, to make booting more robust.
5892 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5894 struct rq *rq = cpu_rq(cpu);
5895 unsigned long flags;
5897 spin_lock_irqsave(&rq->lock, flags);
5900 idle->se.exec_start = sched_clock();
5902 idle->prio = idle->normal_prio = MAX_PRIO;
5903 idle->cpus_allowed = cpumask_of_cpu(cpu);
5904 __set_task_cpu(idle, cpu);
5906 rq->curr = rq->idle = idle;
5907 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5910 spin_unlock_irqrestore(&rq->lock, flags);
5912 /* Set the preempt count _outside_ the spinlocks! */
5913 #if defined(CONFIG_PREEMPT)
5914 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5916 task_thread_info(idle)->preempt_count = 0;
5919 * The idle tasks have their own, simple scheduling class:
5921 idle->sched_class = &idle_sched_class;
5922 ftrace_graph_init_task(idle);
5926 * In a system that switches off the HZ timer nohz_cpu_mask
5927 * indicates which cpus entered this state. This is used
5928 * in the rcu update to wait only for active cpus. For system
5929 * which do not switch off the HZ timer nohz_cpu_mask should
5930 * always be CPU_MASK_NONE.
5932 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5935 * Increase the granularity value when there are more CPUs,
5936 * because with more CPUs the 'effective latency' as visible
5937 * to users decreases. But the relationship is not linear,
5938 * so pick a second-best guess by going with the log2 of the
5941 * This idea comes from the SD scheduler of Con Kolivas:
5943 static inline void sched_init_granularity(void)
5945 unsigned int factor = 1 + ilog2(num_online_cpus());
5946 const unsigned long limit = 200000000;
5948 sysctl_sched_min_granularity *= factor;
5949 if (sysctl_sched_min_granularity > limit)
5950 sysctl_sched_min_granularity = limit;
5952 sysctl_sched_latency *= factor;
5953 if (sysctl_sched_latency > limit)
5954 sysctl_sched_latency = limit;
5956 sysctl_sched_wakeup_granularity *= factor;
5958 sysctl_sched_shares_ratelimit *= factor;
5963 * This is how migration works:
5965 * 1) we queue a struct migration_req structure in the source CPU's
5966 * runqueue and wake up that CPU's migration thread.
5967 * 2) we down() the locked semaphore => thread blocks.
5968 * 3) migration thread wakes up (implicitly it forces the migrated
5969 * thread off the CPU)
5970 * 4) it gets the migration request and checks whether the migrated
5971 * task is still in the wrong runqueue.
5972 * 5) if it's in the wrong runqueue then the migration thread removes
5973 * it and puts it into the right queue.
5974 * 6) migration thread up()s the semaphore.
5975 * 7) we wake up and the migration is done.
5979 * Change a given task's CPU affinity. Migrate the thread to a
5980 * proper CPU and schedule it away if the CPU it's executing on
5981 * is removed from the allowed bitmask.
5983 * NOTE: the caller must have a valid reference to the task, the
5984 * task must not exit() & deallocate itself prematurely. The
5985 * call is not atomic; no spinlocks may be held.
5987 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5989 struct migration_req req;
5990 unsigned long flags;
5994 rq = task_rq_lock(p, &flags);
5995 if (!cpus_intersects(*new_mask, cpu_online_map)) {
6000 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6001 !cpus_equal(p->cpus_allowed, *new_mask))) {
6006 if (p->sched_class->set_cpus_allowed)
6007 p->sched_class->set_cpus_allowed(p, new_mask);
6009 p->cpus_allowed = *new_mask;
6010 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
6013 /* Can the task run on the task's current CPU? If so, we're done */
6014 if (cpu_isset(task_cpu(p), *new_mask))
6017 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
6018 /* Need help from migration thread: drop lock and wait. */
6019 task_rq_unlock(rq, &flags);
6020 wake_up_process(rq->migration_thread);
6021 wait_for_completion(&req.done);
6022 tlb_migrate_finish(p->mm);
6026 task_rq_unlock(rq, &flags);
6030 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6033 * Move (not current) task off this cpu, onto dest cpu. We're doing
6034 * this because either it can't run here any more (set_cpus_allowed()
6035 * away from this CPU, or CPU going down), or because we're
6036 * attempting to rebalance this task on exec (sched_exec).
6038 * So we race with normal scheduler movements, but that's OK, as long
6039 * as the task is no longer on this CPU.
6041 * Returns non-zero if task was successfully migrated.
6043 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6045 struct rq *rq_dest, *rq_src;
6048 if (unlikely(!cpu_active(dest_cpu)))
6051 rq_src = cpu_rq(src_cpu);
6052 rq_dest = cpu_rq(dest_cpu);
6054 double_rq_lock(rq_src, rq_dest);
6055 /* Already moved. */
6056 if (task_cpu(p) != src_cpu)
6058 /* Affinity changed (again). */
6059 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6062 on_rq = p->se.on_rq;
6064 deactivate_task(rq_src, p, 0);
6066 set_task_cpu(p, dest_cpu);
6068 activate_task(rq_dest, p, 0);
6069 check_preempt_curr(rq_dest, p, 0);
6074 double_rq_unlock(rq_src, rq_dest);
6079 * migration_thread - this is a highprio system thread that performs
6080 * thread migration by bumping thread off CPU then 'pushing' onto
6083 static int migration_thread(void *data)
6085 int cpu = (long)data;
6089 BUG_ON(rq->migration_thread != current);
6091 set_current_state(TASK_INTERRUPTIBLE);
6092 while (!kthread_should_stop()) {
6093 struct migration_req *req;
6094 struct list_head *head;
6096 spin_lock_irq(&rq->lock);
6098 if (cpu_is_offline(cpu)) {
6099 spin_unlock_irq(&rq->lock);
6103 if (rq->active_balance) {
6104 active_load_balance(rq, cpu);
6105 rq->active_balance = 0;
6108 head = &rq->migration_queue;
6110 if (list_empty(head)) {
6111 spin_unlock_irq(&rq->lock);
6113 set_current_state(TASK_INTERRUPTIBLE);
6116 req = list_entry(head->next, struct migration_req, list);
6117 list_del_init(head->next);
6119 spin_unlock(&rq->lock);
6120 __migrate_task(req->task, cpu, req->dest_cpu);
6123 complete(&req->done);
6125 __set_current_state(TASK_RUNNING);
6129 /* Wait for kthread_stop */
6130 set_current_state(TASK_INTERRUPTIBLE);
6131 while (!kthread_should_stop()) {
6133 set_current_state(TASK_INTERRUPTIBLE);
6135 __set_current_state(TASK_RUNNING);
6139 #ifdef CONFIG_HOTPLUG_CPU
6141 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6145 local_irq_disable();
6146 ret = __migrate_task(p, src_cpu, dest_cpu);
6152 * Figure out where task on dead CPU should go, use force if necessary.
6153 * NOTE: interrupts should be disabled by the caller
6155 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6157 unsigned long flags;
6164 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6165 cpus_and(mask, mask, p->cpus_allowed);
6166 dest_cpu = any_online_cpu(mask);
6168 /* On any allowed CPU? */
6169 if (dest_cpu >= nr_cpu_ids)
6170 dest_cpu = any_online_cpu(p->cpus_allowed);
6172 /* No more Mr. Nice Guy. */
6173 if (dest_cpu >= nr_cpu_ids) {
6174 cpumask_t cpus_allowed;
6176 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6178 * Try to stay on the same cpuset, where the
6179 * current cpuset may be a subset of all cpus.
6180 * The cpuset_cpus_allowed_locked() variant of
6181 * cpuset_cpus_allowed() will not block. It must be
6182 * called within calls to cpuset_lock/cpuset_unlock.
6184 rq = task_rq_lock(p, &flags);
6185 p->cpus_allowed = cpus_allowed;
6186 dest_cpu = any_online_cpu(p->cpus_allowed);
6187 task_rq_unlock(rq, &flags);
6190 * Don't tell them about moving exiting tasks or
6191 * kernel threads (both mm NULL), since they never
6194 if (p->mm && printk_ratelimit()) {
6195 printk(KERN_INFO "process %d (%s) no "
6196 "longer affine to cpu%d\n",
6197 task_pid_nr(p), p->comm, dead_cpu);
6200 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6204 * While a dead CPU has no uninterruptible tasks queued at this point,
6205 * it might still have a nonzero ->nr_uninterruptible counter, because
6206 * for performance reasons the counter is not stricly tracking tasks to
6207 * their home CPUs. So we just add the counter to another CPU's counter,
6208 * to keep the global sum constant after CPU-down:
6210 static void migrate_nr_uninterruptible(struct rq *rq_src)
6212 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6213 unsigned long flags;
6215 local_irq_save(flags);
6216 double_rq_lock(rq_src, rq_dest);
6217 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6218 rq_src->nr_uninterruptible = 0;
6219 double_rq_unlock(rq_src, rq_dest);
6220 local_irq_restore(flags);
6223 /* Run through task list and migrate tasks from the dead cpu. */
6224 static void migrate_live_tasks(int src_cpu)
6226 struct task_struct *p, *t;
6228 read_lock(&tasklist_lock);
6230 do_each_thread(t, p) {
6234 if (task_cpu(p) == src_cpu)
6235 move_task_off_dead_cpu(src_cpu, p);
6236 } while_each_thread(t, p);
6238 read_unlock(&tasklist_lock);
6242 * Schedules idle task to be the next runnable task on current CPU.
6243 * It does so by boosting its priority to highest possible.
6244 * Used by CPU offline code.
6246 void sched_idle_next(void)
6248 int this_cpu = smp_processor_id();
6249 struct rq *rq = cpu_rq(this_cpu);
6250 struct task_struct *p = rq->idle;
6251 unsigned long flags;
6253 /* cpu has to be offline */
6254 BUG_ON(cpu_online(this_cpu));
6257 * Strictly not necessary since rest of the CPUs are stopped by now
6258 * and interrupts disabled on the current cpu.
6260 spin_lock_irqsave(&rq->lock, flags);
6262 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6264 update_rq_clock(rq);
6265 activate_task(rq, p, 0);
6267 spin_unlock_irqrestore(&rq->lock, flags);
6271 * Ensures that the idle task is using init_mm right before its cpu goes
6274 void idle_task_exit(void)
6276 struct mm_struct *mm = current->active_mm;
6278 BUG_ON(cpu_online(smp_processor_id()));
6281 switch_mm(mm, &init_mm, current);
6285 /* called under rq->lock with disabled interrupts */
6286 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6288 struct rq *rq = cpu_rq(dead_cpu);
6290 /* Must be exiting, otherwise would be on tasklist. */
6291 BUG_ON(!p->exit_state);
6293 /* Cannot have done final schedule yet: would have vanished. */
6294 BUG_ON(p->state == TASK_DEAD);
6299 * Drop lock around migration; if someone else moves it,
6300 * that's OK. No task can be added to this CPU, so iteration is
6303 spin_unlock_irq(&rq->lock);
6304 move_task_off_dead_cpu(dead_cpu, p);
6305 spin_lock_irq(&rq->lock);
6310 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6311 static void migrate_dead_tasks(unsigned int dead_cpu)
6313 struct rq *rq = cpu_rq(dead_cpu);
6314 struct task_struct *next;
6317 if (!rq->nr_running)
6319 update_rq_clock(rq);
6320 next = pick_next_task(rq, rq->curr);
6323 next->sched_class->put_prev_task(rq, next);
6324 migrate_dead(dead_cpu, next);
6328 #endif /* CONFIG_HOTPLUG_CPU */
6330 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6332 static struct ctl_table sd_ctl_dir[] = {
6334 .procname = "sched_domain",
6340 static struct ctl_table sd_ctl_root[] = {
6342 .ctl_name = CTL_KERN,
6343 .procname = "kernel",
6345 .child = sd_ctl_dir,
6350 static struct ctl_table *sd_alloc_ctl_entry(int n)
6352 struct ctl_table *entry =
6353 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6358 static void sd_free_ctl_entry(struct ctl_table **tablep)
6360 struct ctl_table *entry;
6363 * In the intermediate directories, both the child directory and
6364 * procname are dynamically allocated and could fail but the mode
6365 * will always be set. In the lowest directory the names are
6366 * static strings and all have proc handlers.
6368 for (entry = *tablep; entry->mode; entry++) {
6370 sd_free_ctl_entry(&entry->child);
6371 if (entry->proc_handler == NULL)
6372 kfree(entry->procname);
6380 set_table_entry(struct ctl_table *entry,
6381 const char *procname, void *data, int maxlen,
6382 mode_t mode, proc_handler *proc_handler)
6384 entry->procname = procname;
6386 entry->maxlen = maxlen;
6388 entry->proc_handler = proc_handler;
6391 static struct ctl_table *
6392 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6394 struct ctl_table *table = sd_alloc_ctl_entry(13);
6399 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6400 sizeof(long), 0644, proc_doulongvec_minmax);
6401 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6402 sizeof(long), 0644, proc_doulongvec_minmax);
6403 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6404 sizeof(int), 0644, proc_dointvec_minmax);
6405 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6406 sizeof(int), 0644, proc_dointvec_minmax);
6407 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6408 sizeof(int), 0644, proc_dointvec_minmax);
6409 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6410 sizeof(int), 0644, proc_dointvec_minmax);
6411 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6412 sizeof(int), 0644, proc_dointvec_minmax);
6413 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6414 sizeof(int), 0644, proc_dointvec_minmax);
6415 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6416 sizeof(int), 0644, proc_dointvec_minmax);
6417 set_table_entry(&table[9], "cache_nice_tries",
6418 &sd->cache_nice_tries,
6419 sizeof(int), 0644, proc_dointvec_minmax);
6420 set_table_entry(&table[10], "flags", &sd->flags,
6421 sizeof(int), 0644, proc_dointvec_minmax);
6422 set_table_entry(&table[11], "name", sd->name,
6423 CORENAME_MAX_SIZE, 0444, proc_dostring);
6424 /* &table[12] is terminator */
6429 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6431 struct ctl_table *entry, *table;
6432 struct sched_domain *sd;
6433 int domain_num = 0, i;
6436 for_each_domain(cpu, sd)
6438 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6443 for_each_domain(cpu, sd) {
6444 snprintf(buf, 32, "domain%d", i);
6445 entry->procname = kstrdup(buf, GFP_KERNEL);
6447 entry->child = sd_alloc_ctl_domain_table(sd);
6454 static struct ctl_table_header *sd_sysctl_header;
6455 static void register_sched_domain_sysctl(void)
6457 int i, cpu_num = num_online_cpus();
6458 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6461 WARN_ON(sd_ctl_dir[0].child);
6462 sd_ctl_dir[0].child = entry;
6467 for_each_online_cpu(i) {
6468 snprintf(buf, 32, "cpu%d", i);
6469 entry->procname = kstrdup(buf, GFP_KERNEL);
6471 entry->child = sd_alloc_ctl_cpu_table(i);
6475 WARN_ON(sd_sysctl_header);
6476 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6479 /* may be called multiple times per register */
6480 static void unregister_sched_domain_sysctl(void)
6482 if (sd_sysctl_header)
6483 unregister_sysctl_table(sd_sysctl_header);
6484 sd_sysctl_header = NULL;
6485 if (sd_ctl_dir[0].child)
6486 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6489 static void register_sched_domain_sysctl(void)
6492 static void unregister_sched_domain_sysctl(void)
6497 static void set_rq_online(struct rq *rq)
6500 const struct sched_class *class;
6502 cpu_set(rq->cpu, rq->rd->online);
6505 for_each_class(class) {
6506 if (class->rq_online)
6507 class->rq_online(rq);
6512 static void set_rq_offline(struct rq *rq)
6515 const struct sched_class *class;
6517 for_each_class(class) {
6518 if (class->rq_offline)
6519 class->rq_offline(rq);
6522 cpu_clear(rq->cpu, rq->rd->online);
6528 * migration_call - callback that gets triggered when a CPU is added.
6529 * Here we can start up the necessary migration thread for the new CPU.
6531 static int __cpuinit
6532 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6534 struct task_struct *p;
6535 int cpu = (long)hcpu;
6536 unsigned long flags;
6541 case CPU_UP_PREPARE:
6542 case CPU_UP_PREPARE_FROZEN:
6543 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6546 kthread_bind(p, cpu);
6547 /* Must be high prio: stop_machine expects to yield to it. */
6548 rq = task_rq_lock(p, &flags);
6549 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6550 task_rq_unlock(rq, &flags);
6551 cpu_rq(cpu)->migration_thread = p;
6555 case CPU_ONLINE_FROZEN:
6556 /* Strictly unnecessary, as first user will wake it. */
6557 wake_up_process(cpu_rq(cpu)->migration_thread);
6559 /* Update our root-domain */
6561 spin_lock_irqsave(&rq->lock, flags);
6563 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6567 spin_unlock_irqrestore(&rq->lock, flags);
6570 #ifdef CONFIG_HOTPLUG_CPU
6571 case CPU_UP_CANCELED:
6572 case CPU_UP_CANCELED_FROZEN:
6573 if (!cpu_rq(cpu)->migration_thread)
6575 /* Unbind it from offline cpu so it can run. Fall thru. */
6576 kthread_bind(cpu_rq(cpu)->migration_thread,
6577 any_online_cpu(cpu_online_map));
6578 kthread_stop(cpu_rq(cpu)->migration_thread);
6579 cpu_rq(cpu)->migration_thread = NULL;
6583 case CPU_DEAD_FROZEN:
6584 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6585 migrate_live_tasks(cpu);
6587 kthread_stop(rq->migration_thread);
6588 rq->migration_thread = NULL;
6589 /* Idle task back to normal (off runqueue, low prio) */
6590 spin_lock_irq(&rq->lock);
6591 update_rq_clock(rq);
6592 deactivate_task(rq, rq->idle, 0);
6593 rq->idle->static_prio = MAX_PRIO;
6594 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6595 rq->idle->sched_class = &idle_sched_class;
6596 migrate_dead_tasks(cpu);
6597 spin_unlock_irq(&rq->lock);
6599 migrate_nr_uninterruptible(rq);
6600 BUG_ON(rq->nr_running != 0);
6603 * No need to migrate the tasks: it was best-effort if
6604 * they didn't take sched_hotcpu_mutex. Just wake up
6607 spin_lock_irq(&rq->lock);
6608 while (!list_empty(&rq->migration_queue)) {
6609 struct migration_req *req;
6611 req = list_entry(rq->migration_queue.next,
6612 struct migration_req, list);
6613 list_del_init(&req->list);
6614 spin_unlock_irq(&rq->lock);
6615 complete(&req->done);
6616 spin_lock_irq(&rq->lock);
6618 spin_unlock_irq(&rq->lock);
6622 case CPU_DYING_FROZEN:
6623 /* Update our root-domain */
6625 spin_lock_irqsave(&rq->lock, flags);
6627 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6630 spin_unlock_irqrestore(&rq->lock, flags);
6637 /* Register at highest priority so that task migration (migrate_all_tasks)
6638 * happens before everything else.
6640 static struct notifier_block __cpuinitdata migration_notifier = {
6641 .notifier_call = migration_call,
6645 static int __init migration_init(void)
6647 void *cpu = (void *)(long)smp_processor_id();
6650 /* Start one for the boot CPU: */
6651 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6652 BUG_ON(err == NOTIFY_BAD);
6653 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6654 register_cpu_notifier(&migration_notifier);
6658 early_initcall(migration_init);
6663 #ifdef CONFIG_SCHED_DEBUG
6665 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6678 case SD_LV_ALLNODES:
6687 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6688 cpumask_t *groupmask)
6690 struct sched_group *group = sd->groups;
6693 cpulist_scnprintf(str, sizeof(str), sd->span);
6694 cpus_clear(*groupmask);
6696 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6698 if (!(sd->flags & SD_LOAD_BALANCE)) {
6699 printk("does not load-balance\n");
6701 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6706 printk(KERN_CONT "span %s level %s\n",
6707 str, sd_level_to_string(sd->level));
6709 if (!cpu_isset(cpu, sd->span)) {
6710 printk(KERN_ERR "ERROR: domain->span does not contain "
6713 if (!cpu_isset(cpu, group->cpumask)) {
6714 printk(KERN_ERR "ERROR: domain->groups does not contain"
6718 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6722 printk(KERN_ERR "ERROR: group is NULL\n");
6726 if (!group->__cpu_power) {
6727 printk(KERN_CONT "\n");
6728 printk(KERN_ERR "ERROR: domain->cpu_power not "
6733 if (!cpus_weight(group->cpumask)) {
6734 printk(KERN_CONT "\n");
6735 printk(KERN_ERR "ERROR: empty group\n");
6739 if (cpus_intersects(*groupmask, group->cpumask)) {
6740 printk(KERN_CONT "\n");
6741 printk(KERN_ERR "ERROR: repeated CPUs\n");
6745 cpus_or(*groupmask, *groupmask, group->cpumask);
6747 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6748 printk(KERN_CONT " %s", str);
6750 group = group->next;
6751 } while (group != sd->groups);
6752 printk(KERN_CONT "\n");
6754 if (!cpus_equal(sd->span, *groupmask))
6755 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6757 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6758 printk(KERN_ERR "ERROR: parent span is not a superset "
6759 "of domain->span\n");
6763 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6765 cpumask_t *groupmask;
6769 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6773 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6775 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6777 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6782 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6791 #else /* !CONFIG_SCHED_DEBUG */
6792 # define sched_domain_debug(sd, cpu) do { } while (0)
6793 #endif /* CONFIG_SCHED_DEBUG */
6795 static int sd_degenerate(struct sched_domain *sd)
6797 if (cpus_weight(sd->span) == 1)
6800 /* Following flags need at least 2 groups */
6801 if (sd->flags & (SD_LOAD_BALANCE |
6802 SD_BALANCE_NEWIDLE |
6806 SD_SHARE_PKG_RESOURCES)) {
6807 if (sd->groups != sd->groups->next)
6811 /* Following flags don't use groups */
6812 if (sd->flags & (SD_WAKE_IDLE |
6821 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6823 unsigned long cflags = sd->flags, pflags = parent->flags;
6825 if (sd_degenerate(parent))
6828 if (!cpus_equal(sd->span, parent->span))
6831 /* Does parent contain flags not in child? */
6832 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6833 if (cflags & SD_WAKE_AFFINE)
6834 pflags &= ~SD_WAKE_BALANCE;
6835 /* Flags needing groups don't count if only 1 group in parent */
6836 if (parent->groups == parent->groups->next) {
6837 pflags &= ~(SD_LOAD_BALANCE |
6838 SD_BALANCE_NEWIDLE |
6842 SD_SHARE_PKG_RESOURCES);
6844 if (~cflags & pflags)
6850 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6852 unsigned long flags;
6854 spin_lock_irqsave(&rq->lock, flags);
6857 struct root_domain *old_rd = rq->rd;
6859 if (cpu_isset(rq->cpu, old_rd->online))
6862 cpu_clear(rq->cpu, old_rd->span);
6864 if (atomic_dec_and_test(&old_rd->refcount))
6868 atomic_inc(&rd->refcount);
6871 cpu_set(rq->cpu, rd->span);
6872 if (cpu_isset(rq->cpu, cpu_online_map))
6875 spin_unlock_irqrestore(&rq->lock, flags);
6878 static void init_rootdomain(struct root_domain *rd)
6880 memset(rd, 0, sizeof(*rd));
6882 cpus_clear(rd->span);
6883 cpus_clear(rd->online);
6885 cpupri_init(&rd->cpupri);
6888 static void init_defrootdomain(void)
6890 init_rootdomain(&def_root_domain);
6891 atomic_set(&def_root_domain.refcount, 1);
6894 static struct root_domain *alloc_rootdomain(void)
6896 struct root_domain *rd;
6898 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6902 init_rootdomain(rd);
6908 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6909 * hold the hotplug lock.
6912 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6914 struct rq *rq = cpu_rq(cpu);
6915 struct sched_domain *tmp;
6917 /* Remove the sched domains which do not contribute to scheduling. */
6918 for (tmp = sd; tmp; ) {
6919 struct sched_domain *parent = tmp->parent;
6923 if (sd_parent_degenerate(tmp, parent)) {
6924 tmp->parent = parent->parent;
6926 parent->parent->child = tmp;
6931 if (sd && sd_degenerate(sd)) {
6937 sched_domain_debug(sd, cpu);
6939 rq_attach_root(rq, rd);
6940 rcu_assign_pointer(rq->sd, sd);
6943 /* cpus with isolated domains */
6944 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6946 /* Setup the mask of cpus configured for isolated domains */
6947 static int __init isolated_cpu_setup(char *str)
6949 static int __initdata ints[NR_CPUS];
6952 str = get_options(str, ARRAY_SIZE(ints), ints);
6953 cpus_clear(cpu_isolated_map);
6954 for (i = 1; i <= ints[0]; i++)
6955 if (ints[i] < NR_CPUS)
6956 cpu_set(ints[i], cpu_isolated_map);
6960 __setup("isolcpus=", isolated_cpu_setup);
6963 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6964 * to a function which identifies what group(along with sched group) a CPU
6965 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6966 * (due to the fact that we keep track of groups covered with a cpumask_t).
6968 * init_sched_build_groups will build a circular linked list of the groups
6969 * covered by the given span, and will set each group's ->cpumask correctly,
6970 * and ->cpu_power to 0.
6973 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6974 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6975 struct sched_group **sg,
6976 cpumask_t *tmpmask),
6977 cpumask_t *covered, cpumask_t *tmpmask)
6979 struct sched_group *first = NULL, *last = NULL;
6982 cpus_clear(*covered);
6984 for_each_cpu_mask_nr(i, *span) {
6985 struct sched_group *sg;
6986 int group = group_fn(i, cpu_map, &sg, tmpmask);
6989 if (cpu_isset(i, *covered))
6992 cpus_clear(sg->cpumask);
6993 sg->__cpu_power = 0;
6995 for_each_cpu_mask_nr(j, *span) {
6996 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6999 cpu_set(j, *covered);
7000 cpu_set(j, sg->cpumask);
7011 #define SD_NODES_PER_DOMAIN 16
7016 * find_next_best_node - find the next node to include in a sched_domain
7017 * @node: node whose sched_domain we're building
7018 * @used_nodes: nodes already in the sched_domain
7020 * Find the next node to include in a given scheduling domain. Simply
7021 * finds the closest node not already in the @used_nodes map.
7023 * Should use nodemask_t.
7025 static int find_next_best_node(int node, nodemask_t *used_nodes)
7027 int i, n, val, min_val, best_node = 0;
7031 for (i = 0; i < nr_node_ids; i++) {
7032 /* Start at @node */
7033 n = (node + i) % nr_node_ids;
7035 if (!nr_cpus_node(n))
7038 /* Skip already used nodes */
7039 if (node_isset(n, *used_nodes))
7042 /* Simple min distance search */
7043 val = node_distance(node, n);
7045 if (val < min_val) {
7051 node_set(best_node, *used_nodes);
7056 * sched_domain_node_span - get a cpumask for a node's sched_domain
7057 * @node: node whose cpumask we're constructing
7058 * @span: resulting cpumask
7060 * Given a node, construct a good cpumask for its sched_domain to span. It
7061 * should be one that prevents unnecessary balancing, but also spreads tasks
7064 static void sched_domain_node_span(int node, cpumask_t *span)
7066 nodemask_t used_nodes;
7067 node_to_cpumask_ptr(nodemask, node);
7071 nodes_clear(used_nodes);
7073 cpus_or(*span, *span, *nodemask);
7074 node_set(node, used_nodes);
7076 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7077 int next_node = find_next_best_node(node, &used_nodes);
7079 node_to_cpumask_ptr_next(nodemask, next_node);
7080 cpus_or(*span, *span, *nodemask);
7083 #endif /* CONFIG_NUMA */
7085 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7088 * SMT sched-domains:
7090 #ifdef CONFIG_SCHED_SMT
7091 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7092 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7095 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7099 *sg = &per_cpu(sched_group_cpus, cpu);
7102 #endif /* CONFIG_SCHED_SMT */
7105 * multi-core sched-domains:
7107 #ifdef CONFIG_SCHED_MC
7108 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7109 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7110 #endif /* CONFIG_SCHED_MC */
7112 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7114 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7119 *mask = per_cpu(cpu_sibling_map, cpu);
7120 cpus_and(*mask, *mask, *cpu_map);
7121 group = first_cpu(*mask);
7123 *sg = &per_cpu(sched_group_core, group);
7126 #elif defined(CONFIG_SCHED_MC)
7128 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7132 *sg = &per_cpu(sched_group_core, cpu);
7137 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7138 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7141 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7145 #ifdef CONFIG_SCHED_MC
7146 *mask = cpu_coregroup_map(cpu);
7147 cpus_and(*mask, *mask, *cpu_map);
7148 group = first_cpu(*mask);
7149 #elif defined(CONFIG_SCHED_SMT)
7150 *mask = per_cpu(cpu_sibling_map, cpu);
7151 cpus_and(*mask, *mask, *cpu_map);
7152 group = first_cpu(*mask);
7157 *sg = &per_cpu(sched_group_phys, group);
7163 * The init_sched_build_groups can't handle what we want to do with node
7164 * groups, so roll our own. Now each node has its own list of groups which
7165 * gets dynamically allocated.
7167 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7168 static struct sched_group ***sched_group_nodes_bycpu;
7170 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7171 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7173 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7174 struct sched_group **sg, cpumask_t *nodemask)
7178 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7179 cpus_and(*nodemask, *nodemask, *cpu_map);
7180 group = first_cpu(*nodemask);
7183 *sg = &per_cpu(sched_group_allnodes, group);
7187 static void init_numa_sched_groups_power(struct sched_group *group_head)
7189 struct sched_group *sg = group_head;
7195 for_each_cpu_mask_nr(j, sg->cpumask) {
7196 struct sched_domain *sd;
7198 sd = &per_cpu(phys_domains, j);
7199 if (j != first_cpu(sd->groups->cpumask)) {
7201 * Only add "power" once for each
7207 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7210 } while (sg != group_head);
7212 #endif /* CONFIG_NUMA */
7215 /* Free memory allocated for various sched_group structures */
7216 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7220 for_each_cpu_mask_nr(cpu, *cpu_map) {
7221 struct sched_group **sched_group_nodes
7222 = sched_group_nodes_bycpu[cpu];
7224 if (!sched_group_nodes)
7227 for (i = 0; i < nr_node_ids; i++) {
7228 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7230 *nodemask = node_to_cpumask(i);
7231 cpus_and(*nodemask, *nodemask, *cpu_map);
7232 if (cpus_empty(*nodemask))
7242 if (oldsg != sched_group_nodes[i])
7245 kfree(sched_group_nodes);
7246 sched_group_nodes_bycpu[cpu] = NULL;
7249 #else /* !CONFIG_NUMA */
7250 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7253 #endif /* CONFIG_NUMA */
7256 * Initialize sched groups cpu_power.
7258 * cpu_power indicates the capacity of sched group, which is used while
7259 * distributing the load between different sched groups in a sched domain.
7260 * Typically cpu_power for all the groups in a sched domain will be same unless
7261 * there are asymmetries in the topology. If there are asymmetries, group
7262 * having more cpu_power will pickup more load compared to the group having
7265 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7266 * the maximum number of tasks a group can handle in the presence of other idle
7267 * or lightly loaded groups in the same sched domain.
7269 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7271 struct sched_domain *child;
7272 struct sched_group *group;
7274 WARN_ON(!sd || !sd->groups);
7276 if (cpu != first_cpu(sd->groups->cpumask))
7281 sd->groups->__cpu_power = 0;
7284 * For perf policy, if the groups in child domain share resources
7285 * (for example cores sharing some portions of the cache hierarchy
7286 * or SMT), then set this domain groups cpu_power such that each group
7287 * can handle only one task, when there are other idle groups in the
7288 * same sched domain.
7290 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7292 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7293 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7298 * add cpu_power of each child group to this groups cpu_power
7300 group = child->groups;
7302 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7303 group = group->next;
7304 } while (group != child->groups);
7308 * Initializers for schedule domains
7309 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7312 #ifdef CONFIG_SCHED_DEBUG
7313 # define SD_INIT_NAME(sd, type) sd->name = #type
7315 # define SD_INIT_NAME(sd, type) do { } while (0)
7318 #define SD_INIT(sd, type) sd_init_##type(sd)
7320 #define SD_INIT_FUNC(type) \
7321 static noinline void sd_init_##type(struct sched_domain *sd) \
7323 memset(sd, 0, sizeof(*sd)); \
7324 *sd = SD_##type##_INIT; \
7325 sd->level = SD_LV_##type; \
7326 SD_INIT_NAME(sd, type); \
7331 SD_INIT_FUNC(ALLNODES)
7334 #ifdef CONFIG_SCHED_SMT
7335 SD_INIT_FUNC(SIBLING)
7337 #ifdef CONFIG_SCHED_MC
7342 * To minimize stack usage kmalloc room for cpumasks and share the
7343 * space as the usage in build_sched_domains() dictates. Used only
7344 * if the amount of space is significant.
7347 cpumask_t tmpmask; /* make this one first */
7350 cpumask_t this_sibling_map;
7351 cpumask_t this_core_map;
7353 cpumask_t send_covered;
7356 cpumask_t domainspan;
7358 cpumask_t notcovered;
7363 #define SCHED_CPUMASK_ALLOC 1
7364 #define SCHED_CPUMASK_FREE(v) kfree(v)
7365 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7367 #define SCHED_CPUMASK_ALLOC 0
7368 #define SCHED_CPUMASK_FREE(v)
7369 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7372 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7373 ((unsigned long)(a) + offsetof(struct allmasks, v))
7375 static int default_relax_domain_level = -1;
7377 static int __init setup_relax_domain_level(char *str)
7381 val = simple_strtoul(str, NULL, 0);
7382 if (val < SD_LV_MAX)
7383 default_relax_domain_level = val;
7387 __setup("relax_domain_level=", setup_relax_domain_level);
7389 static void set_domain_attribute(struct sched_domain *sd,
7390 struct sched_domain_attr *attr)
7394 if (!attr || attr->relax_domain_level < 0) {
7395 if (default_relax_domain_level < 0)
7398 request = default_relax_domain_level;
7400 request = attr->relax_domain_level;
7401 if (request < sd->level) {
7402 /* turn off idle balance on this domain */
7403 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7405 /* turn on idle balance on this domain */
7406 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7411 * Build sched domains for a given set of cpus and attach the sched domains
7412 * to the individual cpus
7414 static int __build_sched_domains(const cpumask_t *cpu_map,
7415 struct sched_domain_attr *attr)
7418 struct root_domain *rd;
7419 SCHED_CPUMASK_DECLARE(allmasks);
7422 struct sched_group **sched_group_nodes = NULL;
7423 int sd_allnodes = 0;
7426 * Allocate the per-node list of sched groups
7428 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7430 if (!sched_group_nodes) {
7431 printk(KERN_WARNING "Can not alloc sched group node list\n");
7436 rd = alloc_rootdomain();
7438 printk(KERN_WARNING "Cannot alloc root domain\n");
7440 kfree(sched_group_nodes);
7445 #if SCHED_CPUMASK_ALLOC
7446 /* get space for all scratch cpumask variables */
7447 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7449 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7452 kfree(sched_group_nodes);
7457 tmpmask = (cpumask_t *)allmasks;
7461 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7465 * Set up domains for cpus specified by the cpu_map.
7467 for_each_cpu_mask_nr(i, *cpu_map) {
7468 struct sched_domain *sd = NULL, *p;
7469 SCHED_CPUMASK_VAR(nodemask, allmasks);
7471 *nodemask = node_to_cpumask(cpu_to_node(i));
7472 cpus_and(*nodemask, *nodemask, *cpu_map);
7475 if (cpus_weight(*cpu_map) >
7476 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7477 sd = &per_cpu(allnodes_domains, i);
7478 SD_INIT(sd, ALLNODES);
7479 set_domain_attribute(sd, attr);
7480 sd->span = *cpu_map;
7481 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7487 sd = &per_cpu(node_domains, i);
7489 set_domain_attribute(sd, attr);
7490 sched_domain_node_span(cpu_to_node(i), &sd->span);
7494 cpus_and(sd->span, sd->span, *cpu_map);
7498 sd = &per_cpu(phys_domains, i);
7500 set_domain_attribute(sd, attr);
7501 sd->span = *nodemask;
7505 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7507 #ifdef CONFIG_SCHED_MC
7509 sd = &per_cpu(core_domains, i);
7511 set_domain_attribute(sd, attr);
7512 sd->span = cpu_coregroup_map(i);
7513 cpus_and(sd->span, sd->span, *cpu_map);
7516 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7519 #ifdef CONFIG_SCHED_SMT
7521 sd = &per_cpu(cpu_domains, i);
7522 SD_INIT(sd, SIBLING);
7523 set_domain_attribute(sd, attr);
7524 sd->span = per_cpu(cpu_sibling_map, i);
7525 cpus_and(sd->span, sd->span, *cpu_map);
7528 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7532 #ifdef CONFIG_SCHED_SMT
7533 /* Set up CPU (sibling) groups */
7534 for_each_cpu_mask_nr(i, *cpu_map) {
7535 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7536 SCHED_CPUMASK_VAR(send_covered, allmasks);
7538 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7539 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7540 if (i != first_cpu(*this_sibling_map))
7543 init_sched_build_groups(this_sibling_map, cpu_map,
7545 send_covered, tmpmask);
7549 #ifdef CONFIG_SCHED_MC
7550 /* Set up multi-core groups */
7551 for_each_cpu_mask_nr(i, *cpu_map) {
7552 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7553 SCHED_CPUMASK_VAR(send_covered, allmasks);
7555 *this_core_map = cpu_coregroup_map(i);
7556 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7557 if (i != first_cpu(*this_core_map))
7560 init_sched_build_groups(this_core_map, cpu_map,
7562 send_covered, tmpmask);
7566 /* Set up physical groups */
7567 for (i = 0; i < nr_node_ids; i++) {
7568 SCHED_CPUMASK_VAR(nodemask, allmasks);
7569 SCHED_CPUMASK_VAR(send_covered, allmasks);
7571 *nodemask = node_to_cpumask(i);
7572 cpus_and(*nodemask, *nodemask, *cpu_map);
7573 if (cpus_empty(*nodemask))
7576 init_sched_build_groups(nodemask, cpu_map,
7578 send_covered, tmpmask);
7582 /* Set up node groups */
7584 SCHED_CPUMASK_VAR(send_covered, allmasks);
7586 init_sched_build_groups(cpu_map, cpu_map,
7587 &cpu_to_allnodes_group,
7588 send_covered, tmpmask);
7591 for (i = 0; i < nr_node_ids; i++) {
7592 /* Set up node groups */
7593 struct sched_group *sg, *prev;
7594 SCHED_CPUMASK_VAR(nodemask, allmasks);
7595 SCHED_CPUMASK_VAR(domainspan, allmasks);
7596 SCHED_CPUMASK_VAR(covered, allmasks);
7599 *nodemask = node_to_cpumask(i);
7600 cpus_clear(*covered);
7602 cpus_and(*nodemask, *nodemask, *cpu_map);
7603 if (cpus_empty(*nodemask)) {
7604 sched_group_nodes[i] = NULL;
7608 sched_domain_node_span(i, domainspan);
7609 cpus_and(*domainspan, *domainspan, *cpu_map);
7611 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7613 printk(KERN_WARNING "Can not alloc domain group for "
7617 sched_group_nodes[i] = sg;
7618 for_each_cpu_mask_nr(j, *nodemask) {
7619 struct sched_domain *sd;
7621 sd = &per_cpu(node_domains, j);
7624 sg->__cpu_power = 0;
7625 sg->cpumask = *nodemask;
7627 cpus_or(*covered, *covered, *nodemask);
7630 for (j = 0; j < nr_node_ids; j++) {
7631 SCHED_CPUMASK_VAR(notcovered, allmasks);
7632 int n = (i + j) % nr_node_ids;
7633 node_to_cpumask_ptr(pnodemask, n);
7635 cpus_complement(*notcovered, *covered);
7636 cpus_and(*tmpmask, *notcovered, *cpu_map);
7637 cpus_and(*tmpmask, *tmpmask, *domainspan);
7638 if (cpus_empty(*tmpmask))
7641 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7642 if (cpus_empty(*tmpmask))
7645 sg = kmalloc_node(sizeof(struct sched_group),
7649 "Can not alloc domain group for node %d\n", j);
7652 sg->__cpu_power = 0;
7653 sg->cpumask = *tmpmask;
7654 sg->next = prev->next;
7655 cpus_or(*covered, *covered, *tmpmask);
7662 /* Calculate CPU power for physical packages and nodes */
7663 #ifdef CONFIG_SCHED_SMT
7664 for_each_cpu_mask_nr(i, *cpu_map) {
7665 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7667 init_sched_groups_power(i, sd);
7670 #ifdef CONFIG_SCHED_MC
7671 for_each_cpu_mask_nr(i, *cpu_map) {
7672 struct sched_domain *sd = &per_cpu(core_domains, i);
7674 init_sched_groups_power(i, sd);
7678 for_each_cpu_mask_nr(i, *cpu_map) {
7679 struct sched_domain *sd = &per_cpu(phys_domains, i);
7681 init_sched_groups_power(i, sd);
7685 for (i = 0; i < nr_node_ids; i++)
7686 init_numa_sched_groups_power(sched_group_nodes[i]);
7689 struct sched_group *sg;
7691 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7693 init_numa_sched_groups_power(sg);
7697 /* Attach the domains */
7698 for_each_cpu_mask_nr(i, *cpu_map) {
7699 struct sched_domain *sd;
7700 #ifdef CONFIG_SCHED_SMT
7701 sd = &per_cpu(cpu_domains, i);
7702 #elif defined(CONFIG_SCHED_MC)
7703 sd = &per_cpu(core_domains, i);
7705 sd = &per_cpu(phys_domains, i);
7707 cpu_attach_domain(sd, rd, i);
7710 SCHED_CPUMASK_FREE((void *)allmasks);
7715 free_sched_groups(cpu_map, tmpmask);
7716 SCHED_CPUMASK_FREE((void *)allmasks);
7722 static int build_sched_domains(const cpumask_t *cpu_map)
7724 return __build_sched_domains(cpu_map, NULL);
7727 static cpumask_t *doms_cur; /* current sched domains */
7728 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7729 static struct sched_domain_attr *dattr_cur;
7730 /* attribues of custom domains in 'doms_cur' */
7733 * Special case: If a kmalloc of a doms_cur partition (array of
7734 * cpumask_t) fails, then fallback to a single sched domain,
7735 * as determined by the single cpumask_t fallback_doms.
7737 static cpumask_t fallback_doms;
7739 void __attribute__((weak)) arch_update_cpu_topology(void)
7744 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7745 * For now this just excludes isolated cpus, but could be used to
7746 * exclude other special cases in the future.
7748 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7752 arch_update_cpu_topology();
7754 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7756 doms_cur = &fallback_doms;
7757 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7759 err = build_sched_domains(doms_cur);
7760 register_sched_domain_sysctl();
7765 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7768 free_sched_groups(cpu_map, tmpmask);
7772 * Detach sched domains from a group of cpus specified in cpu_map
7773 * These cpus will now be attached to the NULL domain
7775 static void detach_destroy_domains(const cpumask_t *cpu_map)
7780 unregister_sched_domain_sysctl();
7782 for_each_cpu_mask_nr(i, *cpu_map)
7783 cpu_attach_domain(NULL, &def_root_domain, i);
7784 synchronize_sched();
7785 arch_destroy_sched_domains(cpu_map, &tmpmask);
7788 /* handle null as "default" */
7789 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7790 struct sched_domain_attr *new, int idx_new)
7792 struct sched_domain_attr tmp;
7799 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7800 new ? (new + idx_new) : &tmp,
7801 sizeof(struct sched_domain_attr));
7805 * Partition sched domains as specified by the 'ndoms_new'
7806 * cpumasks in the array doms_new[] of cpumasks. This compares
7807 * doms_new[] to the current sched domain partitioning, doms_cur[].
7808 * It destroys each deleted domain and builds each new domain.
7810 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7811 * The masks don't intersect (don't overlap.) We should setup one
7812 * sched domain for each mask. CPUs not in any of the cpumasks will
7813 * not be load balanced. If the same cpumask appears both in the
7814 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7817 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7818 * ownership of it and will kfree it when done with it. If the caller
7819 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7820 * ndoms_new == 1, and partition_sched_domains() will fallback to
7821 * the single partition 'fallback_doms', it also forces the domains
7824 * If doms_new == NULL it will be replaced with cpu_online_map.
7825 * ndoms_new == 0 is a special case for destroying existing domains,
7826 * and it will not create the default domain.
7828 * Call with hotplug lock held
7830 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7831 struct sched_domain_attr *dattr_new)
7835 mutex_lock(&sched_domains_mutex);
7837 /* always unregister in case we don't destroy any domains */
7838 unregister_sched_domain_sysctl();
7840 n = doms_new ? ndoms_new : 0;
7842 /* Destroy deleted domains */
7843 for (i = 0; i < ndoms_cur; i++) {
7844 for (j = 0; j < n; j++) {
7845 if (cpus_equal(doms_cur[i], doms_new[j])
7846 && dattrs_equal(dattr_cur, i, dattr_new, j))
7849 /* no match - a current sched domain not in new doms_new[] */
7850 detach_destroy_domains(doms_cur + i);
7855 if (doms_new == NULL) {
7857 doms_new = &fallback_doms;
7858 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7862 /* Build new domains */
7863 for (i = 0; i < ndoms_new; i++) {
7864 for (j = 0; j < ndoms_cur; j++) {
7865 if (cpus_equal(doms_new[i], doms_cur[j])
7866 && dattrs_equal(dattr_new, i, dattr_cur, j))
7869 /* no match - add a new doms_new */
7870 __build_sched_domains(doms_new + i,
7871 dattr_new ? dattr_new + i : NULL);
7876 /* Remember the new sched domains */
7877 if (doms_cur != &fallback_doms)
7879 kfree(dattr_cur); /* kfree(NULL) is safe */
7880 doms_cur = doms_new;
7881 dattr_cur = dattr_new;
7882 ndoms_cur = ndoms_new;
7884 register_sched_domain_sysctl();
7886 mutex_unlock(&sched_domains_mutex);
7889 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7890 int arch_reinit_sched_domains(void)
7894 /* Destroy domains first to force the rebuild */
7895 partition_sched_domains(0, NULL, NULL);
7897 rebuild_sched_domains();
7903 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7907 if (buf[0] != '0' && buf[0] != '1')
7911 sched_smt_power_savings = (buf[0] == '1');
7913 sched_mc_power_savings = (buf[0] == '1');
7915 ret = arch_reinit_sched_domains();
7917 return ret ? ret : count;
7920 #ifdef CONFIG_SCHED_MC
7921 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7924 return sprintf(page, "%u\n", sched_mc_power_savings);
7926 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7927 const char *buf, size_t count)
7929 return sched_power_savings_store(buf, count, 0);
7931 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7932 sched_mc_power_savings_show,
7933 sched_mc_power_savings_store);
7936 #ifdef CONFIG_SCHED_SMT
7937 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7940 return sprintf(page, "%u\n", sched_smt_power_savings);
7942 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7943 const char *buf, size_t count)
7945 return sched_power_savings_store(buf, count, 1);
7947 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7948 sched_smt_power_savings_show,
7949 sched_smt_power_savings_store);
7952 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7956 #ifdef CONFIG_SCHED_SMT
7958 err = sysfs_create_file(&cls->kset.kobj,
7959 &attr_sched_smt_power_savings.attr);
7961 #ifdef CONFIG_SCHED_MC
7962 if (!err && mc_capable())
7963 err = sysfs_create_file(&cls->kset.kobj,
7964 &attr_sched_mc_power_savings.attr);
7968 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7970 #ifndef CONFIG_CPUSETS
7972 * Add online and remove offline CPUs from the scheduler domains.
7973 * When cpusets are enabled they take over this function.
7975 static int update_sched_domains(struct notifier_block *nfb,
7976 unsigned long action, void *hcpu)
7980 case CPU_ONLINE_FROZEN:
7982 case CPU_DEAD_FROZEN:
7983 partition_sched_domains(1, NULL, NULL);
7992 static int update_runtime(struct notifier_block *nfb,
7993 unsigned long action, void *hcpu)
7995 int cpu = (int)(long)hcpu;
7998 case CPU_DOWN_PREPARE:
7999 case CPU_DOWN_PREPARE_FROZEN:
8000 disable_runtime(cpu_rq(cpu));
8003 case CPU_DOWN_FAILED:
8004 case CPU_DOWN_FAILED_FROZEN:
8006 case CPU_ONLINE_FROZEN:
8007 enable_runtime(cpu_rq(cpu));
8015 void __init sched_init_smp(void)
8017 cpumask_t non_isolated_cpus;
8019 #if defined(CONFIG_NUMA)
8020 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8022 BUG_ON(sched_group_nodes_bycpu == NULL);
8025 mutex_lock(&sched_domains_mutex);
8026 arch_init_sched_domains(&cpu_online_map);
8027 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
8028 if (cpus_empty(non_isolated_cpus))
8029 cpu_set(smp_processor_id(), non_isolated_cpus);
8030 mutex_unlock(&sched_domains_mutex);
8033 #ifndef CONFIG_CPUSETS
8034 /* XXX: Theoretical race here - CPU may be hotplugged now */
8035 hotcpu_notifier(update_sched_domains, 0);
8038 /* RT runtime code needs to handle some hotplug events */
8039 hotcpu_notifier(update_runtime, 0);
8043 /* Move init over to a non-isolated CPU */
8044 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
8046 sched_init_granularity();
8049 void __init sched_init_smp(void)
8051 sched_init_granularity();
8053 #endif /* CONFIG_SMP */
8055 int in_sched_functions(unsigned long addr)
8057 return in_lock_functions(addr) ||
8058 (addr >= (unsigned long)__sched_text_start
8059 && addr < (unsigned long)__sched_text_end);
8062 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8064 cfs_rq->tasks_timeline = RB_ROOT;
8065 INIT_LIST_HEAD(&cfs_rq->tasks);
8066 #ifdef CONFIG_FAIR_GROUP_SCHED
8069 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8072 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8074 struct rt_prio_array *array;
8077 array = &rt_rq->active;
8078 for (i = 0; i < MAX_RT_PRIO; i++) {
8079 INIT_LIST_HEAD(array->queue + i);
8080 __clear_bit(i, array->bitmap);
8082 /* delimiter for bitsearch: */
8083 __set_bit(MAX_RT_PRIO, array->bitmap);
8085 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8086 rt_rq->highest_prio = MAX_RT_PRIO;
8089 rt_rq->rt_nr_migratory = 0;
8090 rt_rq->overloaded = 0;
8094 rt_rq->rt_throttled = 0;
8095 rt_rq->rt_runtime = 0;
8096 spin_lock_init(&rt_rq->rt_runtime_lock);
8098 #ifdef CONFIG_RT_GROUP_SCHED
8099 rt_rq->rt_nr_boosted = 0;
8104 #ifdef CONFIG_FAIR_GROUP_SCHED
8105 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8106 struct sched_entity *se, int cpu, int add,
8107 struct sched_entity *parent)
8109 struct rq *rq = cpu_rq(cpu);
8110 tg->cfs_rq[cpu] = cfs_rq;
8111 init_cfs_rq(cfs_rq, rq);
8114 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8117 /* se could be NULL for init_task_group */
8122 se->cfs_rq = &rq->cfs;
8124 se->cfs_rq = parent->my_q;
8127 se->load.weight = tg->shares;
8128 se->load.inv_weight = 0;
8129 se->parent = parent;
8133 #ifdef CONFIG_RT_GROUP_SCHED
8134 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8135 struct sched_rt_entity *rt_se, int cpu, int add,
8136 struct sched_rt_entity *parent)
8138 struct rq *rq = cpu_rq(cpu);
8140 tg->rt_rq[cpu] = rt_rq;
8141 init_rt_rq(rt_rq, rq);
8143 rt_rq->rt_se = rt_se;
8144 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8146 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8148 tg->rt_se[cpu] = rt_se;
8153 rt_se->rt_rq = &rq->rt;
8155 rt_se->rt_rq = parent->my_q;
8157 rt_se->my_q = rt_rq;
8158 rt_se->parent = parent;
8159 INIT_LIST_HEAD(&rt_se->run_list);
8163 void __init sched_init(void)
8166 unsigned long alloc_size = 0, ptr;
8168 #ifdef CONFIG_FAIR_GROUP_SCHED
8169 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8171 #ifdef CONFIG_RT_GROUP_SCHED
8172 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8174 #ifdef CONFIG_USER_SCHED
8178 * As sched_init() is called before page_alloc is setup,
8179 * we use alloc_bootmem().
8182 ptr = (unsigned long)alloc_bootmem(alloc_size);
8184 #ifdef CONFIG_FAIR_GROUP_SCHED
8185 init_task_group.se = (struct sched_entity **)ptr;
8186 ptr += nr_cpu_ids * sizeof(void **);
8188 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8189 ptr += nr_cpu_ids * sizeof(void **);
8191 #ifdef CONFIG_USER_SCHED
8192 root_task_group.se = (struct sched_entity **)ptr;
8193 ptr += nr_cpu_ids * sizeof(void **);
8195 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8196 ptr += nr_cpu_ids * sizeof(void **);
8197 #endif /* CONFIG_USER_SCHED */
8198 #endif /* CONFIG_FAIR_GROUP_SCHED */
8199 #ifdef CONFIG_RT_GROUP_SCHED
8200 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8201 ptr += nr_cpu_ids * sizeof(void **);
8203 init_task_group.rt_rq = (struct rt_rq **)ptr;
8204 ptr += nr_cpu_ids * sizeof(void **);
8206 #ifdef CONFIG_USER_SCHED
8207 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8208 ptr += nr_cpu_ids * sizeof(void **);
8210 root_task_group.rt_rq = (struct rt_rq **)ptr;
8211 ptr += nr_cpu_ids * sizeof(void **);
8212 #endif /* CONFIG_USER_SCHED */
8213 #endif /* CONFIG_RT_GROUP_SCHED */
8217 init_defrootdomain();
8220 init_rt_bandwidth(&def_rt_bandwidth,
8221 global_rt_period(), global_rt_runtime());
8223 #ifdef CONFIG_RT_GROUP_SCHED
8224 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8225 global_rt_period(), global_rt_runtime());
8226 #ifdef CONFIG_USER_SCHED
8227 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8228 global_rt_period(), RUNTIME_INF);
8229 #endif /* CONFIG_USER_SCHED */
8230 #endif /* CONFIG_RT_GROUP_SCHED */
8232 #ifdef CONFIG_GROUP_SCHED
8233 list_add(&init_task_group.list, &task_groups);
8234 INIT_LIST_HEAD(&init_task_group.children);
8236 #ifdef CONFIG_USER_SCHED
8237 INIT_LIST_HEAD(&root_task_group.children);
8238 init_task_group.parent = &root_task_group;
8239 list_add(&init_task_group.siblings, &root_task_group.children);
8240 #endif /* CONFIG_USER_SCHED */
8241 #endif /* CONFIG_GROUP_SCHED */
8243 for_each_possible_cpu(i) {
8247 spin_lock_init(&rq->lock);
8249 init_cfs_rq(&rq->cfs, rq);
8250 init_rt_rq(&rq->rt, rq);
8251 #ifdef CONFIG_FAIR_GROUP_SCHED
8252 init_task_group.shares = init_task_group_load;
8253 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8254 #ifdef CONFIG_CGROUP_SCHED
8256 * How much cpu bandwidth does init_task_group get?
8258 * In case of task-groups formed thr' the cgroup filesystem, it
8259 * gets 100% of the cpu resources in the system. This overall
8260 * system cpu resource is divided among the tasks of
8261 * init_task_group and its child task-groups in a fair manner,
8262 * based on each entity's (task or task-group's) weight
8263 * (se->load.weight).
8265 * In other words, if init_task_group has 10 tasks of weight
8266 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8267 * then A0's share of the cpu resource is:
8269 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8271 * We achieve this by letting init_task_group's tasks sit
8272 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8274 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8275 #elif defined CONFIG_USER_SCHED
8276 root_task_group.shares = NICE_0_LOAD;
8277 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8279 * In case of task-groups formed thr' the user id of tasks,
8280 * init_task_group represents tasks belonging to root user.
8281 * Hence it forms a sibling of all subsequent groups formed.
8282 * In this case, init_task_group gets only a fraction of overall
8283 * system cpu resource, based on the weight assigned to root
8284 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8285 * by letting tasks of init_task_group sit in a separate cfs_rq
8286 * (init_cfs_rq) and having one entity represent this group of
8287 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8289 init_tg_cfs_entry(&init_task_group,
8290 &per_cpu(init_cfs_rq, i),
8291 &per_cpu(init_sched_entity, i), i, 1,
8292 root_task_group.se[i]);
8295 #endif /* CONFIG_FAIR_GROUP_SCHED */
8297 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8298 #ifdef CONFIG_RT_GROUP_SCHED
8299 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8300 #ifdef CONFIG_CGROUP_SCHED
8301 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8302 #elif defined CONFIG_USER_SCHED
8303 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8304 init_tg_rt_entry(&init_task_group,
8305 &per_cpu(init_rt_rq, i),
8306 &per_cpu(init_sched_rt_entity, i), i, 1,
8307 root_task_group.rt_se[i]);
8311 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8312 rq->cpu_load[j] = 0;
8316 rq->active_balance = 0;
8317 rq->next_balance = jiffies;
8321 rq->migration_thread = NULL;
8322 INIT_LIST_HEAD(&rq->migration_queue);
8323 rq_attach_root(rq, &def_root_domain);
8326 atomic_set(&rq->nr_iowait, 0);
8329 set_load_weight(&init_task);
8331 #ifdef CONFIG_PREEMPT_NOTIFIERS
8332 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8336 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8339 #ifdef CONFIG_RT_MUTEXES
8340 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8344 * The boot idle thread does lazy MMU switching as well:
8346 atomic_inc(&init_mm.mm_count);
8347 enter_lazy_tlb(&init_mm, current);
8350 * Make us the idle thread. Technically, schedule() should not be
8351 * called from this thread, however somewhere below it might be,
8352 * but because we are the idle thread, we just pick up running again
8353 * when this runqueue becomes "idle".
8355 init_idle(current, smp_processor_id());
8357 * During early bootup we pretend to be a normal task:
8359 current->sched_class = &fair_sched_class;
8361 scheduler_running = 1;
8364 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8365 void __might_sleep(char *file, int line)
8368 static unsigned long prev_jiffy; /* ratelimiting */
8370 if ((!in_atomic() && !irqs_disabled()) ||
8371 system_state != SYSTEM_RUNNING || oops_in_progress)
8373 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8375 prev_jiffy = jiffies;
8378 "BUG: sleeping function called from invalid context at %s:%d\n",
8381 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8382 in_atomic(), irqs_disabled(),
8383 current->pid, current->comm);
8385 debug_show_held_locks(current);
8386 if (irqs_disabled())
8387 print_irqtrace_events(current);
8391 EXPORT_SYMBOL(__might_sleep);
8394 #ifdef CONFIG_MAGIC_SYSRQ
8395 static void normalize_task(struct rq *rq, struct task_struct *p)
8399 update_rq_clock(rq);
8400 on_rq = p->se.on_rq;
8402 deactivate_task(rq, p, 0);
8403 __setscheduler(rq, p, SCHED_NORMAL, 0);
8405 activate_task(rq, p, 0);
8406 resched_task(rq->curr);
8410 void normalize_rt_tasks(void)
8412 struct task_struct *g, *p;
8413 unsigned long flags;
8416 read_lock_irqsave(&tasklist_lock, flags);
8417 do_each_thread(g, p) {
8419 * Only normalize user tasks:
8424 p->se.exec_start = 0;
8425 #ifdef CONFIG_SCHEDSTATS
8426 p->se.wait_start = 0;
8427 p->se.sleep_start = 0;
8428 p->se.block_start = 0;
8433 * Renice negative nice level userspace
8436 if (TASK_NICE(p) < 0 && p->mm)
8437 set_user_nice(p, 0);
8441 spin_lock(&p->pi_lock);
8442 rq = __task_rq_lock(p);
8444 normalize_task(rq, p);
8446 __task_rq_unlock(rq);
8447 spin_unlock(&p->pi_lock);
8448 } while_each_thread(g, p);
8450 read_unlock_irqrestore(&tasklist_lock, flags);
8453 #endif /* CONFIG_MAGIC_SYSRQ */
8457 * These functions are only useful for the IA64 MCA handling.
8459 * They can only be called when the whole system has been
8460 * stopped - every CPU needs to be quiescent, and no scheduling
8461 * activity can take place. Using them for anything else would
8462 * be a serious bug, and as a result, they aren't even visible
8463 * under any other configuration.
8467 * curr_task - return the current task for a given cpu.
8468 * @cpu: the processor in question.
8470 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8472 struct task_struct *curr_task(int cpu)
8474 return cpu_curr(cpu);
8478 * set_curr_task - set the current task for a given cpu.
8479 * @cpu: the processor in question.
8480 * @p: the task pointer to set.
8482 * Description: This function must only be used when non-maskable interrupts
8483 * are serviced on a separate stack. It allows the architecture to switch the
8484 * notion of the current task on a cpu in a non-blocking manner. This function
8485 * must be called with all CPU's synchronized, and interrupts disabled, the
8486 * and caller must save the original value of the current task (see
8487 * curr_task() above) and restore that value before reenabling interrupts and
8488 * re-starting the system.
8490 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8492 void set_curr_task(int cpu, struct task_struct *p)
8499 #ifdef CONFIG_FAIR_GROUP_SCHED
8500 static void free_fair_sched_group(struct task_group *tg)
8504 for_each_possible_cpu(i) {
8506 kfree(tg->cfs_rq[i]);
8516 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8518 struct cfs_rq *cfs_rq;
8519 struct sched_entity *se, *parent_se;
8523 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8526 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8530 tg->shares = NICE_0_LOAD;
8532 for_each_possible_cpu(i) {
8535 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8536 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8540 se = kmalloc_node(sizeof(struct sched_entity),
8541 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8545 parent_se = parent ? parent->se[i] : NULL;
8546 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8555 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8557 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8558 &cpu_rq(cpu)->leaf_cfs_rq_list);
8561 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8563 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8565 #else /* !CONFG_FAIR_GROUP_SCHED */
8566 static inline void free_fair_sched_group(struct task_group *tg)
8571 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8576 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8580 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8583 #endif /* CONFIG_FAIR_GROUP_SCHED */
8585 #ifdef CONFIG_RT_GROUP_SCHED
8586 static void free_rt_sched_group(struct task_group *tg)
8590 destroy_rt_bandwidth(&tg->rt_bandwidth);
8592 for_each_possible_cpu(i) {
8594 kfree(tg->rt_rq[i]);
8596 kfree(tg->rt_se[i]);
8604 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8606 struct rt_rq *rt_rq;
8607 struct sched_rt_entity *rt_se, *parent_se;
8611 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8614 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8618 init_rt_bandwidth(&tg->rt_bandwidth,
8619 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8621 for_each_possible_cpu(i) {
8624 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8625 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8629 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8630 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8634 parent_se = parent ? parent->rt_se[i] : NULL;
8635 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8644 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8646 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8647 &cpu_rq(cpu)->leaf_rt_rq_list);
8650 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8652 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8654 #else /* !CONFIG_RT_GROUP_SCHED */
8655 static inline void free_rt_sched_group(struct task_group *tg)
8660 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8665 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8669 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8672 #endif /* CONFIG_RT_GROUP_SCHED */
8674 #ifdef CONFIG_GROUP_SCHED
8675 static void free_sched_group(struct task_group *tg)
8677 free_fair_sched_group(tg);
8678 free_rt_sched_group(tg);
8682 /* allocate runqueue etc for a new task group */
8683 struct task_group *sched_create_group(struct task_group *parent)
8685 struct task_group *tg;
8686 unsigned long flags;
8689 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8691 return ERR_PTR(-ENOMEM);
8693 if (!alloc_fair_sched_group(tg, parent))
8696 if (!alloc_rt_sched_group(tg, parent))
8699 spin_lock_irqsave(&task_group_lock, flags);
8700 for_each_possible_cpu(i) {
8701 register_fair_sched_group(tg, i);
8702 register_rt_sched_group(tg, i);
8704 list_add_rcu(&tg->list, &task_groups);
8706 WARN_ON(!parent); /* root should already exist */
8708 tg->parent = parent;
8709 INIT_LIST_HEAD(&tg->children);
8710 list_add_rcu(&tg->siblings, &parent->children);
8711 spin_unlock_irqrestore(&task_group_lock, flags);
8716 free_sched_group(tg);
8717 return ERR_PTR(-ENOMEM);
8720 /* rcu callback to free various structures associated with a task group */
8721 static void free_sched_group_rcu(struct rcu_head *rhp)
8723 /* now it should be safe to free those cfs_rqs */
8724 free_sched_group(container_of(rhp, struct task_group, rcu));
8727 /* Destroy runqueue etc associated with a task group */
8728 void sched_destroy_group(struct task_group *tg)
8730 unsigned long flags;
8733 spin_lock_irqsave(&task_group_lock, flags);
8734 for_each_possible_cpu(i) {
8735 unregister_fair_sched_group(tg, i);
8736 unregister_rt_sched_group(tg, i);
8738 list_del_rcu(&tg->list);
8739 list_del_rcu(&tg->siblings);
8740 spin_unlock_irqrestore(&task_group_lock, flags);
8742 /* wait for possible concurrent references to cfs_rqs complete */
8743 call_rcu(&tg->rcu, free_sched_group_rcu);
8746 /* change task's runqueue when it moves between groups.
8747 * The caller of this function should have put the task in its new group
8748 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8749 * reflect its new group.
8751 void sched_move_task(struct task_struct *tsk)
8754 unsigned long flags;
8757 rq = task_rq_lock(tsk, &flags);
8759 update_rq_clock(rq);
8761 running = task_current(rq, tsk);
8762 on_rq = tsk->se.on_rq;
8765 dequeue_task(rq, tsk, 0);
8766 if (unlikely(running))
8767 tsk->sched_class->put_prev_task(rq, tsk);
8769 set_task_rq(tsk, task_cpu(tsk));
8771 #ifdef CONFIG_FAIR_GROUP_SCHED
8772 if (tsk->sched_class->moved_group)
8773 tsk->sched_class->moved_group(tsk);
8776 if (unlikely(running))
8777 tsk->sched_class->set_curr_task(rq);
8779 enqueue_task(rq, tsk, 0);
8781 task_rq_unlock(rq, &flags);
8783 #endif /* CONFIG_GROUP_SCHED */
8785 #ifdef CONFIG_FAIR_GROUP_SCHED
8786 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8788 struct cfs_rq *cfs_rq = se->cfs_rq;
8793 dequeue_entity(cfs_rq, se, 0);
8795 se->load.weight = shares;
8796 se->load.inv_weight = 0;
8799 enqueue_entity(cfs_rq, se, 0);
8802 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8804 struct cfs_rq *cfs_rq = se->cfs_rq;
8805 struct rq *rq = cfs_rq->rq;
8806 unsigned long flags;
8808 spin_lock_irqsave(&rq->lock, flags);
8809 __set_se_shares(se, shares);
8810 spin_unlock_irqrestore(&rq->lock, flags);
8813 static DEFINE_MUTEX(shares_mutex);
8815 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8818 unsigned long flags;
8821 * We can't change the weight of the root cgroup.
8826 if (shares < MIN_SHARES)
8827 shares = MIN_SHARES;
8828 else if (shares > MAX_SHARES)
8829 shares = MAX_SHARES;
8831 mutex_lock(&shares_mutex);
8832 if (tg->shares == shares)
8835 spin_lock_irqsave(&task_group_lock, flags);
8836 for_each_possible_cpu(i)
8837 unregister_fair_sched_group(tg, i);
8838 list_del_rcu(&tg->siblings);
8839 spin_unlock_irqrestore(&task_group_lock, flags);
8841 /* wait for any ongoing reference to this group to finish */
8842 synchronize_sched();
8845 * Now we are free to modify the group's share on each cpu
8846 * w/o tripping rebalance_share or load_balance_fair.
8848 tg->shares = shares;
8849 for_each_possible_cpu(i) {
8853 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8854 set_se_shares(tg->se[i], shares);
8858 * Enable load balance activity on this group, by inserting it back on
8859 * each cpu's rq->leaf_cfs_rq_list.
8861 spin_lock_irqsave(&task_group_lock, flags);
8862 for_each_possible_cpu(i)
8863 register_fair_sched_group(tg, i);
8864 list_add_rcu(&tg->siblings, &tg->parent->children);
8865 spin_unlock_irqrestore(&task_group_lock, flags);
8867 mutex_unlock(&shares_mutex);
8871 unsigned long sched_group_shares(struct task_group *tg)
8877 #ifdef CONFIG_RT_GROUP_SCHED
8879 * Ensure that the real time constraints are schedulable.
8881 static DEFINE_MUTEX(rt_constraints_mutex);
8883 static unsigned long to_ratio(u64 period, u64 runtime)
8885 if (runtime == RUNTIME_INF)
8888 return div64_u64(runtime << 20, period);
8891 /* Must be called with tasklist_lock held */
8892 static inline int tg_has_rt_tasks(struct task_group *tg)
8894 struct task_struct *g, *p;
8896 do_each_thread(g, p) {
8897 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8899 } while_each_thread(g, p);
8904 struct rt_schedulable_data {
8905 struct task_group *tg;
8910 static int tg_schedulable(struct task_group *tg, void *data)
8912 struct rt_schedulable_data *d = data;
8913 struct task_group *child;
8914 unsigned long total, sum = 0;
8915 u64 period, runtime;
8917 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8918 runtime = tg->rt_bandwidth.rt_runtime;
8921 period = d->rt_period;
8922 runtime = d->rt_runtime;
8926 * Cannot have more runtime than the period.
8928 if (runtime > period && runtime != RUNTIME_INF)
8932 * Ensure we don't starve existing RT tasks.
8934 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8937 total = to_ratio(period, runtime);
8940 * Nobody can have more than the global setting allows.
8942 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8946 * The sum of our children's runtime should not exceed our own.
8948 list_for_each_entry_rcu(child, &tg->children, siblings) {
8949 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8950 runtime = child->rt_bandwidth.rt_runtime;
8952 if (child == d->tg) {
8953 period = d->rt_period;
8954 runtime = d->rt_runtime;
8957 sum += to_ratio(period, runtime);
8966 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8968 struct rt_schedulable_data data = {
8970 .rt_period = period,
8971 .rt_runtime = runtime,
8974 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8977 static int tg_set_bandwidth(struct task_group *tg,
8978 u64 rt_period, u64 rt_runtime)
8982 mutex_lock(&rt_constraints_mutex);
8983 read_lock(&tasklist_lock);
8984 err = __rt_schedulable(tg, rt_period, rt_runtime);
8988 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8989 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8990 tg->rt_bandwidth.rt_runtime = rt_runtime;
8992 for_each_possible_cpu(i) {
8993 struct rt_rq *rt_rq = tg->rt_rq[i];
8995 spin_lock(&rt_rq->rt_runtime_lock);
8996 rt_rq->rt_runtime = rt_runtime;
8997 spin_unlock(&rt_rq->rt_runtime_lock);
8999 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9001 read_unlock(&tasklist_lock);
9002 mutex_unlock(&rt_constraints_mutex);
9007 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9009 u64 rt_runtime, rt_period;
9011 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9012 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9013 if (rt_runtime_us < 0)
9014 rt_runtime = RUNTIME_INF;
9016 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9019 long sched_group_rt_runtime(struct task_group *tg)
9023 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9026 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9027 do_div(rt_runtime_us, NSEC_PER_USEC);
9028 return rt_runtime_us;
9031 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9033 u64 rt_runtime, rt_period;
9035 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9036 rt_runtime = tg->rt_bandwidth.rt_runtime;
9041 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9044 long sched_group_rt_period(struct task_group *tg)
9048 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9049 do_div(rt_period_us, NSEC_PER_USEC);
9050 return rt_period_us;
9053 static int sched_rt_global_constraints(void)
9055 u64 runtime, period;
9058 if (sysctl_sched_rt_period <= 0)
9061 runtime = global_rt_runtime();
9062 period = global_rt_period();
9065 * Sanity check on the sysctl variables.
9067 if (runtime > period && runtime != RUNTIME_INF)
9070 mutex_lock(&rt_constraints_mutex);
9071 read_lock(&tasklist_lock);
9072 ret = __rt_schedulable(NULL, 0, 0);
9073 read_unlock(&tasklist_lock);
9074 mutex_unlock(&rt_constraints_mutex);
9078 #else /* !CONFIG_RT_GROUP_SCHED */
9079 static int sched_rt_global_constraints(void)
9081 unsigned long flags;
9084 if (sysctl_sched_rt_period <= 0)
9087 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9088 for_each_possible_cpu(i) {
9089 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9091 spin_lock(&rt_rq->rt_runtime_lock);
9092 rt_rq->rt_runtime = global_rt_runtime();
9093 spin_unlock(&rt_rq->rt_runtime_lock);
9095 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9099 #endif /* CONFIG_RT_GROUP_SCHED */
9101 int sched_rt_handler(struct ctl_table *table, int write,
9102 struct file *filp, void __user *buffer, size_t *lenp,
9106 int old_period, old_runtime;
9107 static DEFINE_MUTEX(mutex);
9110 old_period = sysctl_sched_rt_period;
9111 old_runtime = sysctl_sched_rt_runtime;
9113 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9115 if (!ret && write) {
9116 ret = sched_rt_global_constraints();
9118 sysctl_sched_rt_period = old_period;
9119 sysctl_sched_rt_runtime = old_runtime;
9121 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9122 def_rt_bandwidth.rt_period =
9123 ns_to_ktime(global_rt_period());
9126 mutex_unlock(&mutex);
9131 #ifdef CONFIG_CGROUP_SCHED
9133 /* return corresponding task_group object of a cgroup */
9134 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9136 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9137 struct task_group, css);
9140 static struct cgroup_subsys_state *
9141 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9143 struct task_group *tg, *parent;
9145 if (!cgrp->parent) {
9146 /* This is early initialization for the top cgroup */
9147 return &init_task_group.css;
9150 parent = cgroup_tg(cgrp->parent);
9151 tg = sched_create_group(parent);
9153 return ERR_PTR(-ENOMEM);
9159 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9161 struct task_group *tg = cgroup_tg(cgrp);
9163 sched_destroy_group(tg);
9167 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9168 struct task_struct *tsk)
9170 #ifdef CONFIG_RT_GROUP_SCHED
9171 /* Don't accept realtime tasks when there is no way for them to run */
9172 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9175 /* We don't support RT-tasks being in separate groups */
9176 if (tsk->sched_class != &fair_sched_class)
9184 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9185 struct cgroup *old_cont, struct task_struct *tsk)
9187 sched_move_task(tsk);
9190 #ifdef CONFIG_FAIR_GROUP_SCHED
9191 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9194 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9197 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9199 struct task_group *tg = cgroup_tg(cgrp);
9201 return (u64) tg->shares;
9203 #endif /* CONFIG_FAIR_GROUP_SCHED */
9205 #ifdef CONFIG_RT_GROUP_SCHED
9206 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9209 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9212 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9214 return sched_group_rt_runtime(cgroup_tg(cgrp));
9217 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9220 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9223 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9225 return sched_group_rt_period(cgroup_tg(cgrp));
9227 #endif /* CONFIG_RT_GROUP_SCHED */
9229 static struct cftype cpu_files[] = {
9230 #ifdef CONFIG_FAIR_GROUP_SCHED
9233 .read_u64 = cpu_shares_read_u64,
9234 .write_u64 = cpu_shares_write_u64,
9237 #ifdef CONFIG_RT_GROUP_SCHED
9239 .name = "rt_runtime_us",
9240 .read_s64 = cpu_rt_runtime_read,
9241 .write_s64 = cpu_rt_runtime_write,
9244 .name = "rt_period_us",
9245 .read_u64 = cpu_rt_period_read_uint,
9246 .write_u64 = cpu_rt_period_write_uint,
9251 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9253 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9256 struct cgroup_subsys cpu_cgroup_subsys = {
9258 .create = cpu_cgroup_create,
9259 .destroy = cpu_cgroup_destroy,
9260 .can_attach = cpu_cgroup_can_attach,
9261 .attach = cpu_cgroup_attach,
9262 .populate = cpu_cgroup_populate,
9263 .subsys_id = cpu_cgroup_subsys_id,
9267 #endif /* CONFIG_CGROUP_SCHED */
9269 #ifdef CONFIG_CGROUP_CPUACCT
9272 * CPU accounting code for task groups.
9274 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9275 * (balbir@in.ibm.com).
9278 /* track cpu usage of a group of tasks */
9280 struct cgroup_subsys_state css;
9281 /* cpuusage holds pointer to a u64-type object on every cpu */
9285 struct cgroup_subsys cpuacct_subsys;
9287 /* return cpu accounting group corresponding to this container */
9288 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9290 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9291 struct cpuacct, css);
9294 /* return cpu accounting group to which this task belongs */
9295 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9297 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9298 struct cpuacct, css);
9301 /* create a new cpu accounting group */
9302 static struct cgroup_subsys_state *cpuacct_create(
9303 struct cgroup_subsys *ss, struct cgroup *cgrp)
9305 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9308 return ERR_PTR(-ENOMEM);
9310 ca->cpuusage = alloc_percpu(u64);
9311 if (!ca->cpuusage) {
9313 return ERR_PTR(-ENOMEM);
9319 /* destroy an existing cpu accounting group */
9321 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9323 struct cpuacct *ca = cgroup_ca(cgrp);
9325 free_percpu(ca->cpuusage);
9329 /* return total cpu usage (in nanoseconds) of a group */
9330 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9332 struct cpuacct *ca = cgroup_ca(cgrp);
9333 u64 totalcpuusage = 0;
9336 for_each_possible_cpu(i) {
9337 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9340 * Take rq->lock to make 64-bit addition safe on 32-bit
9343 spin_lock_irq(&cpu_rq(i)->lock);
9344 totalcpuusage += *cpuusage;
9345 spin_unlock_irq(&cpu_rq(i)->lock);
9348 return totalcpuusage;
9351 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9354 struct cpuacct *ca = cgroup_ca(cgrp);
9363 for_each_possible_cpu(i) {
9364 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9366 spin_lock_irq(&cpu_rq(i)->lock);
9368 spin_unlock_irq(&cpu_rq(i)->lock);
9374 static struct cftype files[] = {
9377 .read_u64 = cpuusage_read,
9378 .write_u64 = cpuusage_write,
9382 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9384 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9388 * charge this task's execution time to its accounting group.
9390 * called with rq->lock held.
9392 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9396 if (!cpuacct_subsys.active)
9401 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9403 *cpuusage += cputime;
9407 struct cgroup_subsys cpuacct_subsys = {
9409 .create = cpuacct_create,
9410 .destroy = cpuacct_destroy,
9411 .populate = cpuacct_populate,
9412 .subsys_id = cpuacct_subsys_id,
9414 #endif /* CONFIG_CGROUP_CPUACCT */