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/perf_counter.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/reciprocal_div.h>
68 #include <linux/unistd.h>
69 #include <linux/pagemap.h>
70 #include <linux/hrtimer.h>
71 #include <linux/tick.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
125 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
128 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
129 * Since cpu_power is a 'constant', we can use a reciprocal divide.
131 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
133 return reciprocal_divide(load, sg->reciprocal_cpu_power);
137 * Each time a sched group cpu_power is changed,
138 * we must compute its reciprocal value
140 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
142 sg->__cpu_power += val;
143 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
147 static inline int rt_policy(int policy)
149 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
154 static inline int task_has_rt_policy(struct task_struct *p)
156 return rt_policy(p->policy);
160 * This is the priority-queue data structure of the RT scheduling class:
162 struct rt_prio_array {
163 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
164 struct list_head queue[MAX_RT_PRIO];
167 struct rt_bandwidth {
168 /* nests inside the rq lock: */
169 spinlock_t rt_runtime_lock;
172 struct hrtimer rt_period_timer;
175 static struct rt_bandwidth def_rt_bandwidth;
177 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
179 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
181 struct rt_bandwidth *rt_b =
182 container_of(timer, struct rt_bandwidth, rt_period_timer);
188 now = hrtimer_cb_get_time(timer);
189 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
194 idle = do_sched_rt_period_timer(rt_b, overrun);
197 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
201 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
203 rt_b->rt_period = ns_to_ktime(period);
204 rt_b->rt_runtime = runtime;
206 spin_lock_init(&rt_b->rt_runtime_lock);
208 hrtimer_init(&rt_b->rt_period_timer,
209 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
210 rt_b->rt_period_timer.function = sched_rt_period_timer;
213 static inline int rt_bandwidth_enabled(void)
215 return sysctl_sched_rt_runtime >= 0;
218 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
222 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
225 if (hrtimer_active(&rt_b->rt_period_timer))
228 spin_lock(&rt_b->rt_runtime_lock);
233 if (hrtimer_active(&rt_b->rt_period_timer))
236 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
237 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
239 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
240 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
241 delta = ktime_to_ns(ktime_sub(hard, soft));
242 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
243 HRTIMER_MODE_ABS_PINNED, 0);
245 spin_unlock(&rt_b->rt_runtime_lock);
248 #ifdef CONFIG_RT_GROUP_SCHED
249 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
251 hrtimer_cancel(&rt_b->rt_period_timer);
256 * sched_domains_mutex serializes calls to arch_init_sched_domains,
257 * detach_destroy_domains and partition_sched_domains.
259 static DEFINE_MUTEX(sched_domains_mutex);
261 #ifdef CONFIG_GROUP_SCHED
263 #include <linux/cgroup.h>
267 static LIST_HEAD(task_groups);
269 /* task group related information */
271 #ifdef CONFIG_CGROUP_SCHED
272 struct cgroup_subsys_state css;
275 #ifdef CONFIG_USER_SCHED
279 #ifdef CONFIG_FAIR_GROUP_SCHED
280 /* schedulable entities of this group on each cpu */
281 struct sched_entity **se;
282 /* runqueue "owned" by this group on each cpu */
283 struct cfs_rq **cfs_rq;
284 unsigned long shares;
287 #ifdef CONFIG_RT_GROUP_SCHED
288 struct sched_rt_entity **rt_se;
289 struct rt_rq **rt_rq;
291 struct rt_bandwidth rt_bandwidth;
295 struct list_head list;
297 struct task_group *parent;
298 struct list_head siblings;
299 struct list_head children;
302 #ifdef CONFIG_USER_SCHED
304 /* Helper function to pass uid information to create_sched_user() */
305 void set_tg_uid(struct user_struct *user)
307 user->tg->uid = user->uid;
312 * Every UID task group (including init_task_group aka UID-0) will
313 * be a child to this group.
315 struct task_group root_task_group;
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 /* Default task group's sched entity on each cpu */
319 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
320 /* Default task group's cfs_rq on each cpu */
321 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
322 #endif /* CONFIG_FAIR_GROUP_SCHED */
324 #ifdef CONFIG_RT_GROUP_SCHED
325 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
326 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
327 #endif /* CONFIG_RT_GROUP_SCHED */
328 #else /* !CONFIG_USER_SCHED */
329 #define root_task_group init_task_group
330 #endif /* CONFIG_USER_SCHED */
332 /* task_group_lock serializes add/remove of task groups and also changes to
333 * a task group's cpu shares.
335 static DEFINE_SPINLOCK(task_group_lock);
338 static int root_task_group_empty(void)
340 return list_empty(&root_task_group.children);
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 #ifdef CONFIG_USER_SCHED
346 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
347 #else /* !CONFIG_USER_SCHED */
348 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
349 #endif /* CONFIG_USER_SCHED */
352 * A weight of 0 or 1 can cause arithmetics problems.
353 * A weight of a cfs_rq is the sum of weights of which entities
354 * are queued on this cfs_rq, so a weight of a entity should not be
355 * too large, so as the shares value of a task group.
356 * (The default weight is 1024 - so there's no practical
357 * limitation from this.)
360 #define MAX_SHARES (1UL << 18)
362 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
365 /* Default task group.
366 * Every task in system belong to this group at bootup.
368 struct task_group init_task_group;
370 /* return group to which a task belongs */
371 static inline struct task_group *task_group(struct task_struct *p)
373 struct task_group *tg;
375 #ifdef CONFIG_USER_SCHED
377 tg = __task_cred(p)->user->tg;
379 #elif defined(CONFIG_CGROUP_SCHED)
380 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
381 struct task_group, css);
383 tg = &init_task_group;
388 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
389 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
393 p->se.parent = task_group(p)->se[cpu];
396 #ifdef CONFIG_RT_GROUP_SCHED
397 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
398 p->rt.parent = task_group(p)->rt_se[cpu];
405 static int root_task_group_empty(void)
411 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
412 static inline struct task_group *task_group(struct task_struct *p)
417 #endif /* CONFIG_GROUP_SCHED */
419 /* CFS-related fields in a runqueue */
421 struct load_weight load;
422 unsigned long nr_running;
427 struct rb_root tasks_timeline;
428 struct rb_node *rb_leftmost;
430 struct list_head tasks;
431 struct list_head *balance_iterator;
434 * 'curr' points to currently running entity on this cfs_rq.
435 * It is set to NULL otherwise (i.e when none are currently running).
437 struct sched_entity *curr, *next, *last;
439 unsigned int nr_spread_over;
441 #ifdef CONFIG_FAIR_GROUP_SCHED
442 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
445 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
446 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
447 * (like users, containers etc.)
449 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
450 * list is used during load balance.
452 struct list_head leaf_cfs_rq_list;
453 struct task_group *tg; /* group that "owns" this runqueue */
457 * the part of load.weight contributed by tasks
459 unsigned long task_weight;
462 * h_load = weight * f(tg)
464 * Where f(tg) is the recursive weight fraction assigned to
467 unsigned long h_load;
470 * this cpu's part of tg->shares
472 unsigned long shares;
475 * load.weight at the time we set shares
477 unsigned long rq_weight;
482 /* Real-Time classes' related field in a runqueue: */
484 struct rt_prio_array active;
485 unsigned long rt_nr_running;
486 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
488 int curr; /* highest queued rt task prio */
490 int next; /* next highest */
495 unsigned long rt_nr_migratory;
496 unsigned long rt_nr_total;
498 struct plist_head pushable_tasks;
503 /* Nests inside the rq lock: */
504 spinlock_t rt_runtime_lock;
506 #ifdef CONFIG_RT_GROUP_SCHED
507 unsigned long rt_nr_boosted;
510 struct list_head leaf_rt_rq_list;
511 struct task_group *tg;
512 struct sched_rt_entity *rt_se;
519 * We add the notion of a root-domain which will be used to define per-domain
520 * variables. Each exclusive cpuset essentially defines an island domain by
521 * fully partitioning the member cpus from any other cpuset. Whenever a new
522 * exclusive cpuset is created, we also create and attach a new root-domain
529 cpumask_var_t online;
532 * The "RT overload" flag: it gets set if a CPU has more than
533 * one runnable RT task.
535 cpumask_var_t rto_mask;
538 struct cpupri cpupri;
540 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
542 * Preferred wake up cpu nominated by sched_mc balance that will be
543 * used when most cpus are idle in the system indicating overall very
544 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
546 unsigned int sched_mc_preferred_wakeup_cpu;
551 * By default the system creates a single root-domain with all cpus as
552 * members (mimicking the global state we have today).
554 static struct root_domain def_root_domain;
559 * This is the main, per-CPU runqueue data structure.
561 * Locking rule: those places that want to lock multiple runqueues
562 * (such as the load balancing or the thread migration code), lock
563 * acquire operations must be ordered by ascending &runqueue.
570 * nr_running and cpu_load should be in the same cacheline because
571 * remote CPUs use both these fields when doing load calculation.
573 unsigned long nr_running;
574 #define CPU_LOAD_IDX_MAX 5
575 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
577 unsigned long last_tick_seen;
578 unsigned char in_nohz_recently;
580 /* capture load from *all* tasks on this cpu: */
581 struct load_weight load;
582 unsigned long nr_load_updates;
584 u64 nr_migrations_in;
589 #ifdef CONFIG_FAIR_GROUP_SCHED
590 /* list of leaf cfs_rq on this cpu: */
591 struct list_head leaf_cfs_rq_list;
593 #ifdef CONFIG_RT_GROUP_SCHED
594 struct list_head leaf_rt_rq_list;
598 * This is part of a global counter where only the total sum
599 * over all CPUs matters. A task can increase this counter on
600 * one CPU and if it got migrated afterwards it may decrease
601 * it on another CPU. Always updated under the runqueue lock:
603 unsigned long nr_uninterruptible;
605 struct task_struct *curr, *idle;
606 unsigned long next_balance;
607 struct mm_struct *prev_mm;
614 struct root_domain *rd;
615 struct sched_domain *sd;
617 unsigned char idle_at_tick;
618 /* For active balancing */
622 /* cpu of this runqueue: */
626 unsigned long avg_load_per_task;
628 struct task_struct *migration_thread;
629 struct list_head migration_queue;
632 /* calc_load related fields */
633 unsigned long calc_load_update;
634 long calc_load_active;
636 #ifdef CONFIG_SCHED_HRTICK
638 int hrtick_csd_pending;
639 struct call_single_data hrtick_csd;
641 struct hrtimer hrtick_timer;
644 #ifdef CONFIG_SCHEDSTATS
646 struct sched_info rq_sched_info;
647 unsigned long long rq_cpu_time;
648 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
650 /* sys_sched_yield() stats */
651 unsigned int yld_count;
653 /* schedule() stats */
654 unsigned int sched_switch;
655 unsigned int sched_count;
656 unsigned int sched_goidle;
658 /* try_to_wake_up() stats */
659 unsigned int ttwu_count;
660 unsigned int ttwu_local;
663 unsigned int bkl_count;
667 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
669 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
671 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
674 static inline int cpu_of(struct rq *rq)
684 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
685 * See detach_destroy_domains: synchronize_sched for details.
687 * The domain tree of any CPU may only be accessed from within
688 * preempt-disabled sections.
690 #define for_each_domain(cpu, __sd) \
691 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
693 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
694 #define this_rq() (&__get_cpu_var(runqueues))
695 #define task_rq(p) cpu_rq(task_cpu(p))
696 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
697 #define raw_rq() (&__raw_get_cpu_var(runqueues))
699 inline void update_rq_clock(struct rq *rq)
701 rq->clock = sched_clock_cpu(cpu_of(rq));
705 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
707 #ifdef CONFIG_SCHED_DEBUG
708 # define const_debug __read_mostly
710 # define const_debug static const
716 * Returns true if the current cpu runqueue is locked.
717 * This interface allows printk to be called with the runqueue lock
718 * held and know whether or not it is OK to wake up the klogd.
720 int runqueue_is_locked(void)
723 struct rq *rq = cpu_rq(cpu);
726 ret = spin_is_locked(&rq->lock);
732 * Debugging: various feature bits
735 #define SCHED_FEAT(name, enabled) \
736 __SCHED_FEAT_##name ,
739 #include "sched_features.h"
744 #define SCHED_FEAT(name, enabled) \
745 (1UL << __SCHED_FEAT_##name) * enabled |
747 const_debug unsigned int sysctl_sched_features =
748 #include "sched_features.h"
753 #ifdef CONFIG_SCHED_DEBUG
754 #define SCHED_FEAT(name, enabled) \
757 static __read_mostly char *sched_feat_names[] = {
758 #include "sched_features.h"
764 static int sched_feat_show(struct seq_file *m, void *v)
768 for (i = 0; sched_feat_names[i]; i++) {
769 if (!(sysctl_sched_features & (1UL << i)))
771 seq_printf(m, "%s ", sched_feat_names[i]);
779 sched_feat_write(struct file *filp, const char __user *ubuf,
780 size_t cnt, loff_t *ppos)
790 if (copy_from_user(&buf, ubuf, cnt))
795 if (strncmp(buf, "NO_", 3) == 0) {
800 for (i = 0; sched_feat_names[i]; i++) {
801 int len = strlen(sched_feat_names[i]);
803 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
805 sysctl_sched_features &= ~(1UL << i);
807 sysctl_sched_features |= (1UL << i);
812 if (!sched_feat_names[i])
820 static int sched_feat_open(struct inode *inode, struct file *filp)
822 return single_open(filp, sched_feat_show, NULL);
825 static struct file_operations sched_feat_fops = {
826 .open = sched_feat_open,
827 .write = sched_feat_write,
830 .release = single_release,
833 static __init int sched_init_debug(void)
835 debugfs_create_file("sched_features", 0644, NULL, NULL,
840 late_initcall(sched_init_debug);
844 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
847 * Number of tasks to iterate in a single balance run.
848 * Limited because this is done with IRQs disabled.
850 const_debug unsigned int sysctl_sched_nr_migrate = 32;
853 * ratelimit for updating the group shares.
856 unsigned int sysctl_sched_shares_ratelimit = 250000;
859 * Inject some fuzzyness into changing the per-cpu group shares
860 * this avoids remote rq-locks at the expense of fairness.
863 unsigned int sysctl_sched_shares_thresh = 4;
866 * period over which we measure -rt task cpu usage in us.
869 unsigned int sysctl_sched_rt_period = 1000000;
871 static __read_mostly int scheduler_running;
874 * part of the period that we allow rt tasks to run in us.
877 int sysctl_sched_rt_runtime = 950000;
879 static inline u64 global_rt_period(void)
881 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
884 static inline u64 global_rt_runtime(void)
886 if (sysctl_sched_rt_runtime < 0)
889 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
892 #ifndef prepare_arch_switch
893 # define prepare_arch_switch(next) do { } while (0)
895 #ifndef finish_arch_switch
896 # define finish_arch_switch(prev) do { } while (0)
899 static inline int task_current(struct rq *rq, struct task_struct *p)
901 return rq->curr == p;
904 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
905 static inline int task_running(struct rq *rq, struct task_struct *p)
907 return task_current(rq, p);
910 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
914 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
916 #ifdef CONFIG_DEBUG_SPINLOCK
917 /* this is a valid case when another task releases the spinlock */
918 rq->lock.owner = current;
921 * If we are tracking spinlock dependencies then we have to
922 * fix up the runqueue lock - which gets 'carried over' from
925 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
927 spin_unlock_irq(&rq->lock);
930 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
931 static inline int task_running(struct rq *rq, struct task_struct *p)
936 return task_current(rq, p);
940 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
944 * We can optimise this out completely for !SMP, because the
945 * SMP rebalancing from interrupt is the only thing that cares
950 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
951 spin_unlock_irq(&rq->lock);
953 spin_unlock(&rq->lock);
957 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
961 * After ->oncpu is cleared, the task can be moved to a different CPU.
962 * We must ensure this doesn't happen until the switch is completely
968 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
972 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
975 * __task_rq_lock - lock the runqueue a given task resides on.
976 * Must be called interrupts disabled.
978 static inline struct rq *__task_rq_lock(struct task_struct *p)
982 struct rq *rq = task_rq(p);
983 spin_lock(&rq->lock);
984 if (likely(rq == task_rq(p)))
986 spin_unlock(&rq->lock);
991 * task_rq_lock - lock the runqueue a given task resides on and disable
992 * interrupts. Note the ordering: we can safely lookup the task_rq without
993 * explicitly disabling preemption.
995 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1001 local_irq_save(*flags);
1003 spin_lock(&rq->lock);
1004 if (likely(rq == task_rq(p)))
1006 spin_unlock_irqrestore(&rq->lock, *flags);
1010 void task_rq_unlock_wait(struct task_struct *p)
1012 struct rq *rq = task_rq(p);
1014 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1015 spin_unlock_wait(&rq->lock);
1018 static void __task_rq_unlock(struct rq *rq)
1019 __releases(rq->lock)
1021 spin_unlock(&rq->lock);
1024 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1025 __releases(rq->lock)
1027 spin_unlock_irqrestore(&rq->lock, *flags);
1031 * this_rq_lock - lock this runqueue and disable interrupts.
1033 static struct rq *this_rq_lock(void)
1034 __acquires(rq->lock)
1038 local_irq_disable();
1040 spin_lock(&rq->lock);
1045 #ifdef CONFIG_SCHED_HRTICK
1047 * Use HR-timers to deliver accurate preemption points.
1049 * Its all a bit involved since we cannot program an hrt while holding the
1050 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1053 * When we get rescheduled we reprogram the hrtick_timer outside of the
1059 * - enabled by features
1060 * - hrtimer is actually high res
1062 static inline int hrtick_enabled(struct rq *rq)
1064 if (!sched_feat(HRTICK))
1066 if (!cpu_active(cpu_of(rq)))
1068 return hrtimer_is_hres_active(&rq->hrtick_timer);
1071 static void hrtick_clear(struct rq *rq)
1073 if (hrtimer_active(&rq->hrtick_timer))
1074 hrtimer_cancel(&rq->hrtick_timer);
1078 * High-resolution timer tick.
1079 * Runs from hardirq context with interrupts disabled.
1081 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1083 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1085 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1087 spin_lock(&rq->lock);
1088 update_rq_clock(rq);
1089 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1090 spin_unlock(&rq->lock);
1092 return HRTIMER_NORESTART;
1097 * called from hardirq (IPI) context
1099 static void __hrtick_start(void *arg)
1101 struct rq *rq = arg;
1103 spin_lock(&rq->lock);
1104 hrtimer_restart(&rq->hrtick_timer);
1105 rq->hrtick_csd_pending = 0;
1106 spin_unlock(&rq->lock);
1110 * Called to set the hrtick timer state.
1112 * called with rq->lock held and irqs disabled
1114 static void hrtick_start(struct rq *rq, u64 delay)
1116 struct hrtimer *timer = &rq->hrtick_timer;
1117 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1119 hrtimer_set_expires(timer, time);
1121 if (rq == this_rq()) {
1122 hrtimer_restart(timer);
1123 } else if (!rq->hrtick_csd_pending) {
1124 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1125 rq->hrtick_csd_pending = 1;
1130 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1132 int cpu = (int)(long)hcpu;
1135 case CPU_UP_CANCELED:
1136 case CPU_UP_CANCELED_FROZEN:
1137 case CPU_DOWN_PREPARE:
1138 case CPU_DOWN_PREPARE_FROZEN:
1140 case CPU_DEAD_FROZEN:
1141 hrtick_clear(cpu_rq(cpu));
1148 static __init void init_hrtick(void)
1150 hotcpu_notifier(hotplug_hrtick, 0);
1154 * Called to set the hrtick timer state.
1156 * called with rq->lock held and irqs disabled
1158 static void hrtick_start(struct rq *rq, u64 delay)
1160 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1161 HRTIMER_MODE_REL_PINNED, 0);
1164 static inline void init_hrtick(void)
1167 #endif /* CONFIG_SMP */
1169 static void init_rq_hrtick(struct rq *rq)
1172 rq->hrtick_csd_pending = 0;
1174 rq->hrtick_csd.flags = 0;
1175 rq->hrtick_csd.func = __hrtick_start;
1176 rq->hrtick_csd.info = rq;
1179 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1180 rq->hrtick_timer.function = hrtick;
1182 #else /* CONFIG_SCHED_HRTICK */
1183 static inline void hrtick_clear(struct rq *rq)
1187 static inline void init_rq_hrtick(struct rq *rq)
1191 static inline void init_hrtick(void)
1194 #endif /* CONFIG_SCHED_HRTICK */
1197 * resched_task - mark a task 'to be rescheduled now'.
1199 * On UP this means the setting of the need_resched flag, on SMP it
1200 * might also involve a cross-CPU call to trigger the scheduler on
1205 #ifndef tsk_is_polling
1206 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1209 static void resched_task(struct task_struct *p)
1213 assert_spin_locked(&task_rq(p)->lock);
1215 if (test_tsk_need_resched(p))
1218 set_tsk_need_resched(p);
1221 if (cpu == smp_processor_id())
1224 /* NEED_RESCHED must be visible before we test polling */
1226 if (!tsk_is_polling(p))
1227 smp_send_reschedule(cpu);
1230 static void resched_cpu(int cpu)
1232 struct rq *rq = cpu_rq(cpu);
1233 unsigned long flags;
1235 if (!spin_trylock_irqsave(&rq->lock, flags))
1237 resched_task(cpu_curr(cpu));
1238 spin_unlock_irqrestore(&rq->lock, flags);
1243 * When add_timer_on() enqueues a timer into the timer wheel of an
1244 * idle CPU then this timer might expire before the next timer event
1245 * which is scheduled to wake up that CPU. In case of a completely
1246 * idle system the next event might even be infinite time into the
1247 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1248 * leaves the inner idle loop so the newly added timer is taken into
1249 * account when the CPU goes back to idle and evaluates the timer
1250 * wheel for the next timer event.
1252 void wake_up_idle_cpu(int cpu)
1254 struct rq *rq = cpu_rq(cpu);
1256 if (cpu == smp_processor_id())
1260 * This is safe, as this function is called with the timer
1261 * wheel base lock of (cpu) held. When the CPU is on the way
1262 * to idle and has not yet set rq->curr to idle then it will
1263 * be serialized on the timer wheel base lock and take the new
1264 * timer into account automatically.
1266 if (rq->curr != rq->idle)
1270 * We can set TIF_RESCHED on the idle task of the other CPU
1271 * lockless. The worst case is that the other CPU runs the
1272 * idle task through an additional NOOP schedule()
1274 set_tsk_need_resched(rq->idle);
1276 /* NEED_RESCHED must be visible before we test polling */
1278 if (!tsk_is_polling(rq->idle))
1279 smp_send_reschedule(cpu);
1281 #endif /* CONFIG_NO_HZ */
1283 #else /* !CONFIG_SMP */
1284 static void resched_task(struct task_struct *p)
1286 assert_spin_locked(&task_rq(p)->lock);
1287 set_tsk_need_resched(p);
1289 #endif /* CONFIG_SMP */
1291 #if BITS_PER_LONG == 32
1292 # define WMULT_CONST (~0UL)
1294 # define WMULT_CONST (1UL << 32)
1297 #define WMULT_SHIFT 32
1300 * Shift right and round:
1302 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1305 * delta *= weight / lw
1307 static unsigned long
1308 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1309 struct load_weight *lw)
1313 if (!lw->inv_weight) {
1314 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1317 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1321 tmp = (u64)delta_exec * weight;
1323 * Check whether we'd overflow the 64-bit multiplication:
1325 if (unlikely(tmp > WMULT_CONST))
1326 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1329 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1331 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1334 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1340 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1347 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1348 * of tasks with abnormal "nice" values across CPUs the contribution that
1349 * each task makes to its run queue's load is weighted according to its
1350 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1351 * scaled version of the new time slice allocation that they receive on time
1355 #define WEIGHT_IDLEPRIO 3
1356 #define WMULT_IDLEPRIO 1431655765
1359 * Nice levels are multiplicative, with a gentle 10% change for every
1360 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1361 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1362 * that remained on nice 0.
1364 * The "10% effect" is relative and cumulative: from _any_ nice level,
1365 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1366 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1367 * If a task goes up by ~10% and another task goes down by ~10% then
1368 * the relative distance between them is ~25%.)
1370 static const int prio_to_weight[40] = {
1371 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1372 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1373 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1374 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1375 /* 0 */ 1024, 820, 655, 526, 423,
1376 /* 5 */ 335, 272, 215, 172, 137,
1377 /* 10 */ 110, 87, 70, 56, 45,
1378 /* 15 */ 36, 29, 23, 18, 15,
1382 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1384 * In cases where the weight does not change often, we can use the
1385 * precalculated inverse to speed up arithmetics by turning divisions
1386 * into multiplications:
1388 static const u32 prio_to_wmult[40] = {
1389 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1390 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1391 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1392 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1393 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1394 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1395 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1396 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1399 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1402 * runqueue iterator, to support SMP load-balancing between different
1403 * scheduling classes, without having to expose their internal data
1404 * structures to the load-balancing proper:
1406 struct rq_iterator {
1408 struct task_struct *(*start)(void *);
1409 struct task_struct *(*next)(void *);
1413 static unsigned long
1414 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1415 unsigned long max_load_move, struct sched_domain *sd,
1416 enum cpu_idle_type idle, int *all_pinned,
1417 int *this_best_prio, struct rq_iterator *iterator);
1420 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1421 struct sched_domain *sd, enum cpu_idle_type idle,
1422 struct rq_iterator *iterator);
1425 /* Time spent by the tasks of the cpu accounting group executing in ... */
1426 enum cpuacct_stat_index {
1427 CPUACCT_STAT_USER, /* ... user mode */
1428 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1430 CPUACCT_STAT_NSTATS,
1433 #ifdef CONFIG_CGROUP_CPUACCT
1434 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1435 static void cpuacct_update_stats(struct task_struct *tsk,
1436 enum cpuacct_stat_index idx, cputime_t val);
1438 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1439 static inline void cpuacct_update_stats(struct task_struct *tsk,
1440 enum cpuacct_stat_index idx, cputime_t val) {}
1443 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1445 update_load_add(&rq->load, load);
1448 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1450 update_load_sub(&rq->load, load);
1453 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1454 typedef int (*tg_visitor)(struct task_group *, void *);
1457 * Iterate the full tree, calling @down when first entering a node and @up when
1458 * leaving it for the final time.
1460 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1462 struct task_group *parent, *child;
1466 parent = &root_task_group;
1468 ret = (*down)(parent, data);
1471 list_for_each_entry_rcu(child, &parent->children, siblings) {
1478 ret = (*up)(parent, data);
1483 parent = parent->parent;
1492 static int tg_nop(struct task_group *tg, void *data)
1499 static unsigned long source_load(int cpu, int type);
1500 static unsigned long target_load(int cpu, int type);
1501 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1503 static unsigned long cpu_avg_load_per_task(int cpu)
1505 struct rq *rq = cpu_rq(cpu);
1506 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1509 rq->avg_load_per_task = rq->load.weight / nr_running;
1511 rq->avg_load_per_task = 0;
1513 return rq->avg_load_per_task;
1516 #ifdef CONFIG_FAIR_GROUP_SCHED
1518 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1521 * Calculate and set the cpu's group shares.
1524 update_group_shares_cpu(struct task_group *tg, int cpu,
1525 unsigned long sd_shares, unsigned long sd_rq_weight,
1526 unsigned long sd_eff_weight)
1528 unsigned long rq_weight;
1529 unsigned long shares;
1535 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1538 rq_weight = NICE_0_LOAD;
1539 if (sd_rq_weight == sd_eff_weight)
1540 sd_eff_weight += NICE_0_LOAD;
1541 sd_rq_weight = sd_eff_weight;
1545 * \Sum_j shares_j * rq_weight_i
1546 * shares_i = -----------------------------
1547 * \Sum_j rq_weight_j
1549 shares = (sd_shares * rq_weight) / sd_rq_weight;
1550 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1552 if (abs(shares - tg->se[cpu]->load.weight) >
1553 sysctl_sched_shares_thresh) {
1554 struct rq *rq = cpu_rq(cpu);
1555 unsigned long flags;
1557 spin_lock_irqsave(&rq->lock, flags);
1558 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1559 __set_se_shares(tg->se[cpu], shares);
1560 spin_unlock_irqrestore(&rq->lock, flags);
1565 * Re-compute the task group their per cpu shares over the given domain.
1566 * This needs to be done in a bottom-up fashion because the rq weight of a
1567 * parent group depends on the shares of its child groups.
1569 static int tg_shares_up(struct task_group *tg, void *data)
1571 unsigned long weight, rq_weight = 0, eff_weight = 0;
1572 unsigned long shares = 0;
1573 struct sched_domain *sd = data;
1576 for_each_cpu(i, sched_domain_span(sd)) {
1578 * If there are currently no tasks on the cpu pretend there
1579 * is one of average load so that when a new task gets to
1580 * run here it will not get delayed by group starvation.
1582 weight = tg->cfs_rq[i]->load.weight;
1583 tg->cfs_rq[i]->rq_weight = weight;
1584 rq_weight += weight;
1587 weight = NICE_0_LOAD;
1589 eff_weight += weight;
1590 shares += tg->cfs_rq[i]->shares;
1593 if ((!shares && rq_weight) || shares > tg->shares)
1594 shares = tg->shares;
1596 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1597 shares = tg->shares;
1599 for_each_cpu(i, sched_domain_span(sd))
1600 update_group_shares_cpu(tg, i, shares, rq_weight, eff_weight);
1606 * Compute the cpu's hierarchical load factor for each task group.
1607 * This needs to be done in a top-down fashion because the load of a child
1608 * group is a fraction of its parents load.
1610 static int tg_load_down(struct task_group *tg, void *data)
1613 long cpu = (long)data;
1616 load = cpu_rq(cpu)->load.weight;
1618 load = tg->parent->cfs_rq[cpu]->h_load;
1619 load *= tg->cfs_rq[cpu]->shares;
1620 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1623 tg->cfs_rq[cpu]->h_load = load;
1628 static void update_shares(struct sched_domain *sd)
1633 if (root_task_group_empty())
1636 now = cpu_clock(raw_smp_processor_id());
1637 elapsed = now - sd->last_update;
1639 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1640 sd->last_update = now;
1641 walk_tg_tree(tg_nop, tg_shares_up, sd);
1645 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1647 if (root_task_group_empty())
1650 spin_unlock(&rq->lock);
1652 spin_lock(&rq->lock);
1655 static void update_h_load(long cpu)
1657 if (root_task_group_empty())
1660 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1665 static inline void update_shares(struct sched_domain *sd)
1669 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1675 #ifdef CONFIG_PREEMPT
1678 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1679 * way at the expense of forcing extra atomic operations in all
1680 * invocations. This assures that the double_lock is acquired using the
1681 * same underlying policy as the spinlock_t on this architecture, which
1682 * reduces latency compared to the unfair variant below. However, it
1683 * also adds more overhead and therefore may reduce throughput.
1685 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1686 __releases(this_rq->lock)
1687 __acquires(busiest->lock)
1688 __acquires(this_rq->lock)
1690 spin_unlock(&this_rq->lock);
1691 double_rq_lock(this_rq, busiest);
1698 * Unfair double_lock_balance: Optimizes throughput at the expense of
1699 * latency by eliminating extra atomic operations when the locks are
1700 * already in proper order on entry. This favors lower cpu-ids and will
1701 * grant the double lock to lower cpus over higher ids under contention,
1702 * regardless of entry order into the function.
1704 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1705 __releases(this_rq->lock)
1706 __acquires(busiest->lock)
1707 __acquires(this_rq->lock)
1711 if (unlikely(!spin_trylock(&busiest->lock))) {
1712 if (busiest < this_rq) {
1713 spin_unlock(&this_rq->lock);
1714 spin_lock(&busiest->lock);
1715 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1718 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1723 #endif /* CONFIG_PREEMPT */
1726 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1728 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1730 if (unlikely(!irqs_disabled())) {
1731 /* printk() doesn't work good under rq->lock */
1732 spin_unlock(&this_rq->lock);
1736 return _double_lock_balance(this_rq, busiest);
1739 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1740 __releases(busiest->lock)
1742 spin_unlock(&busiest->lock);
1743 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1747 #ifdef CONFIG_FAIR_GROUP_SCHED
1748 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1751 cfs_rq->shares = shares;
1756 static void calc_load_account_active(struct rq *this_rq);
1758 #include "sched_stats.h"
1759 #include "sched_idletask.c"
1760 #include "sched_fair.c"
1761 #include "sched_rt.c"
1762 #ifdef CONFIG_SCHED_DEBUG
1763 # include "sched_debug.c"
1766 #define sched_class_highest (&rt_sched_class)
1767 #define for_each_class(class) \
1768 for (class = sched_class_highest; class; class = class->next)
1770 static void inc_nr_running(struct rq *rq)
1775 static void dec_nr_running(struct rq *rq)
1780 static void set_load_weight(struct task_struct *p)
1782 if (task_has_rt_policy(p)) {
1783 p->se.load.weight = prio_to_weight[0] * 2;
1784 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1789 * SCHED_IDLE tasks get minimal weight:
1791 if (p->policy == SCHED_IDLE) {
1792 p->se.load.weight = WEIGHT_IDLEPRIO;
1793 p->se.load.inv_weight = WMULT_IDLEPRIO;
1797 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1798 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1801 static void update_avg(u64 *avg, u64 sample)
1803 s64 diff = sample - *avg;
1807 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1810 p->se.start_runtime = p->se.sum_exec_runtime;
1812 sched_info_queued(p);
1813 p->sched_class->enqueue_task(rq, p, wakeup);
1817 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1820 if (p->se.last_wakeup) {
1821 update_avg(&p->se.avg_overlap,
1822 p->se.sum_exec_runtime - p->se.last_wakeup);
1823 p->se.last_wakeup = 0;
1825 update_avg(&p->se.avg_wakeup,
1826 sysctl_sched_wakeup_granularity);
1830 sched_info_dequeued(p);
1831 p->sched_class->dequeue_task(rq, p, sleep);
1836 * __normal_prio - return the priority that is based on the static prio
1838 static inline int __normal_prio(struct task_struct *p)
1840 return p->static_prio;
1844 * Calculate the expected normal priority: i.e. priority
1845 * without taking RT-inheritance into account. Might be
1846 * boosted by interactivity modifiers. Changes upon fork,
1847 * setprio syscalls, and whenever the interactivity
1848 * estimator recalculates.
1850 static inline int normal_prio(struct task_struct *p)
1854 if (task_has_rt_policy(p))
1855 prio = MAX_RT_PRIO-1 - p->rt_priority;
1857 prio = __normal_prio(p);
1862 * Calculate the current priority, i.e. the priority
1863 * taken into account by the scheduler. This value might
1864 * be boosted by RT tasks, or might be boosted by
1865 * interactivity modifiers. Will be RT if the task got
1866 * RT-boosted. If not then it returns p->normal_prio.
1868 static int effective_prio(struct task_struct *p)
1870 p->normal_prio = normal_prio(p);
1872 * If we are RT tasks or we were boosted to RT priority,
1873 * keep the priority unchanged. Otherwise, update priority
1874 * to the normal priority:
1876 if (!rt_prio(p->prio))
1877 return p->normal_prio;
1882 * activate_task - move a task to the runqueue.
1884 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1886 if (task_contributes_to_load(p))
1887 rq->nr_uninterruptible--;
1889 enqueue_task(rq, p, wakeup);
1894 * deactivate_task - remove a task from the runqueue.
1896 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1898 if (task_contributes_to_load(p))
1899 rq->nr_uninterruptible++;
1901 dequeue_task(rq, p, sleep);
1906 * task_curr - is this task currently executing on a CPU?
1907 * @p: the task in question.
1909 inline int task_curr(const struct task_struct *p)
1911 return cpu_curr(task_cpu(p)) == p;
1914 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1916 set_task_rq(p, cpu);
1919 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1920 * successfuly executed on another CPU. We must ensure that updates of
1921 * per-task data have been completed by this moment.
1924 task_thread_info(p)->cpu = cpu;
1928 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1929 const struct sched_class *prev_class,
1930 int oldprio, int running)
1932 if (prev_class != p->sched_class) {
1933 if (prev_class->switched_from)
1934 prev_class->switched_from(rq, p, running);
1935 p->sched_class->switched_to(rq, p, running);
1937 p->sched_class->prio_changed(rq, p, oldprio, running);
1942 /* Used instead of source_load when we know the type == 0 */
1943 static unsigned long weighted_cpuload(const int cpu)
1945 return cpu_rq(cpu)->load.weight;
1949 * Is this task likely cache-hot:
1952 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1957 * Buddy candidates are cache hot:
1959 if (sched_feat(CACHE_HOT_BUDDY) &&
1960 (&p->se == cfs_rq_of(&p->se)->next ||
1961 &p->se == cfs_rq_of(&p->se)->last))
1964 if (p->sched_class != &fair_sched_class)
1967 if (sysctl_sched_migration_cost == -1)
1969 if (sysctl_sched_migration_cost == 0)
1972 delta = now - p->se.exec_start;
1974 return delta < (s64)sysctl_sched_migration_cost;
1978 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1980 int old_cpu = task_cpu(p);
1981 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1982 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1983 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1986 clock_offset = old_rq->clock - new_rq->clock;
1988 trace_sched_migrate_task(p, new_cpu);
1990 #ifdef CONFIG_SCHEDSTATS
1991 if (p->se.wait_start)
1992 p->se.wait_start -= clock_offset;
1993 if (p->se.sleep_start)
1994 p->se.sleep_start -= clock_offset;
1995 if (p->se.block_start)
1996 p->se.block_start -= clock_offset;
1998 if (old_cpu != new_cpu) {
1999 p->se.nr_migrations++;
2000 new_rq->nr_migrations_in++;
2001 #ifdef CONFIG_SCHEDSTATS
2002 if (task_hot(p, old_rq->clock, NULL))
2003 schedstat_inc(p, se.nr_forced2_migrations);
2005 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2008 p->se.vruntime -= old_cfsrq->min_vruntime -
2009 new_cfsrq->min_vruntime;
2011 __set_task_cpu(p, new_cpu);
2014 struct migration_req {
2015 struct list_head list;
2017 struct task_struct *task;
2020 struct completion done;
2024 * The task's runqueue lock must be held.
2025 * Returns true if you have to wait for migration thread.
2028 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2030 struct rq *rq = task_rq(p);
2033 * If the task is not on a runqueue (and not running), then
2034 * it is sufficient to simply update the task's cpu field.
2036 if (!p->se.on_rq && !task_running(rq, p)) {
2037 set_task_cpu(p, dest_cpu);
2041 init_completion(&req->done);
2043 req->dest_cpu = dest_cpu;
2044 list_add(&req->list, &rq->migration_queue);
2050 * wait_task_context_switch - wait for a thread to complete at least one
2053 * @p must not be current.
2055 void wait_task_context_switch(struct task_struct *p)
2057 unsigned long nvcsw, nivcsw, flags;
2065 * The runqueue is assigned before the actual context
2066 * switch. We need to take the runqueue lock.
2068 * We could check initially without the lock but it is
2069 * very likely that we need to take the lock in every
2072 rq = task_rq_lock(p, &flags);
2073 running = task_running(rq, p);
2074 task_rq_unlock(rq, &flags);
2076 if (likely(!running))
2079 * The switch count is incremented before the actual
2080 * context switch. We thus wait for two switches to be
2081 * sure at least one completed.
2083 if ((p->nvcsw - nvcsw) > 1)
2085 if ((p->nivcsw - nivcsw) > 1)
2093 * wait_task_inactive - wait for a thread to unschedule.
2095 * If @match_state is nonzero, it's the @p->state value just checked and
2096 * not expected to change. If it changes, i.e. @p might have woken up,
2097 * then return zero. When we succeed in waiting for @p to be off its CPU,
2098 * we return a positive number (its total switch count). If a second call
2099 * a short while later returns the same number, the caller can be sure that
2100 * @p has remained unscheduled the whole time.
2102 * The caller must ensure that the task *will* unschedule sometime soon,
2103 * else this function might spin for a *long* time. This function can't
2104 * be called with interrupts off, or it may introduce deadlock with
2105 * smp_call_function() if an IPI is sent by the same process we are
2106 * waiting to become inactive.
2108 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2110 unsigned long flags;
2117 * We do the initial early heuristics without holding
2118 * any task-queue locks at all. We'll only try to get
2119 * the runqueue lock when things look like they will
2125 * If the task is actively running on another CPU
2126 * still, just relax and busy-wait without holding
2129 * NOTE! Since we don't hold any locks, it's not
2130 * even sure that "rq" stays as the right runqueue!
2131 * But we don't care, since "task_running()" will
2132 * return false if the runqueue has changed and p
2133 * is actually now running somewhere else!
2135 while (task_running(rq, p)) {
2136 if (match_state && unlikely(p->state != match_state))
2142 * Ok, time to look more closely! We need the rq
2143 * lock now, to be *sure*. If we're wrong, we'll
2144 * just go back and repeat.
2146 rq = task_rq_lock(p, &flags);
2147 trace_sched_wait_task(rq, p);
2148 running = task_running(rq, p);
2149 on_rq = p->se.on_rq;
2151 if (!match_state || p->state == match_state)
2152 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2153 task_rq_unlock(rq, &flags);
2156 * If it changed from the expected state, bail out now.
2158 if (unlikely(!ncsw))
2162 * Was it really running after all now that we
2163 * checked with the proper locks actually held?
2165 * Oops. Go back and try again..
2167 if (unlikely(running)) {
2173 * It's not enough that it's not actively running,
2174 * it must be off the runqueue _entirely_, and not
2177 * So if it was still runnable (but just not actively
2178 * running right now), it's preempted, and we should
2179 * yield - it could be a while.
2181 if (unlikely(on_rq)) {
2182 schedule_timeout_uninterruptible(1);
2187 * Ahh, all good. It wasn't running, and it wasn't
2188 * runnable, which means that it will never become
2189 * running in the future either. We're all done!
2198 * kick_process - kick a running thread to enter/exit the kernel
2199 * @p: the to-be-kicked thread
2201 * Cause a process which is running on another CPU to enter
2202 * kernel-mode, without any delay. (to get signals handled.)
2204 * NOTE: this function doesnt have to take the runqueue lock,
2205 * because all it wants to ensure is that the remote task enters
2206 * the kernel. If the IPI races and the task has been migrated
2207 * to another CPU then no harm is done and the purpose has been
2210 void kick_process(struct task_struct *p)
2216 if ((cpu != smp_processor_id()) && task_curr(p))
2217 smp_send_reschedule(cpu);
2220 EXPORT_SYMBOL_GPL(kick_process);
2223 * Return a low guess at the load of a migration-source cpu weighted
2224 * according to the scheduling class and "nice" value.
2226 * We want to under-estimate the load of migration sources, to
2227 * balance conservatively.
2229 static unsigned long source_load(int cpu, int type)
2231 struct rq *rq = cpu_rq(cpu);
2232 unsigned long total = weighted_cpuload(cpu);
2234 if (type == 0 || !sched_feat(LB_BIAS))
2237 return min(rq->cpu_load[type-1], total);
2241 * Return a high guess at the load of a migration-target cpu weighted
2242 * according to the scheduling class and "nice" value.
2244 static unsigned long target_load(int cpu, int type)
2246 struct rq *rq = cpu_rq(cpu);
2247 unsigned long total = weighted_cpuload(cpu);
2249 if (type == 0 || !sched_feat(LB_BIAS))
2252 return max(rq->cpu_load[type-1], total);
2256 * find_idlest_group finds and returns the least busy CPU group within the
2259 static struct sched_group *
2260 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2262 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2263 unsigned long min_load = ULONG_MAX, this_load = 0;
2264 int load_idx = sd->forkexec_idx;
2265 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2268 unsigned long load, avg_load;
2272 /* Skip over this group if it has no CPUs allowed */
2273 if (!cpumask_intersects(sched_group_cpus(group),
2277 local_group = cpumask_test_cpu(this_cpu,
2278 sched_group_cpus(group));
2280 /* Tally up the load of all CPUs in the group */
2283 for_each_cpu(i, sched_group_cpus(group)) {
2284 /* Bias balancing toward cpus of our domain */
2286 load = source_load(i, load_idx);
2288 load = target_load(i, load_idx);
2293 /* Adjust by relative CPU power of the group */
2294 avg_load = sg_div_cpu_power(group,
2295 avg_load * SCHED_LOAD_SCALE);
2298 this_load = avg_load;
2300 } else if (avg_load < min_load) {
2301 min_load = avg_load;
2304 } while (group = group->next, group != sd->groups);
2306 if (!idlest || 100*this_load < imbalance*min_load)
2312 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2315 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2317 unsigned long load, min_load = ULONG_MAX;
2321 /* Traverse only the allowed CPUs */
2322 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2323 load = weighted_cpuload(i);
2325 if (load < min_load || (load == min_load && i == this_cpu)) {
2335 * sched_balance_self: balance the current task (running on cpu) in domains
2336 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2339 * Balance, ie. select the least loaded group.
2341 * Returns the target CPU number, or the same CPU if no balancing is needed.
2343 * preempt must be disabled.
2345 static int sched_balance_self(int cpu, int flag)
2347 struct task_struct *t = current;
2348 struct sched_domain *tmp, *sd = NULL;
2350 for_each_domain(cpu, tmp) {
2352 * If power savings logic is enabled for a domain, stop there.
2354 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2356 if (tmp->flags & flag)
2364 struct sched_group *group;
2365 int new_cpu, weight;
2367 if (!(sd->flags & flag)) {
2372 group = find_idlest_group(sd, t, cpu);
2378 new_cpu = find_idlest_cpu(group, t, cpu);
2379 if (new_cpu == -1 || new_cpu == cpu) {
2380 /* Now try balancing at a lower domain level of cpu */
2385 /* Now try balancing at a lower domain level of new_cpu */
2387 weight = cpumask_weight(sched_domain_span(sd));
2389 for_each_domain(cpu, tmp) {
2390 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2392 if (tmp->flags & flag)
2395 /* while loop will break here if sd == NULL */
2401 #endif /* CONFIG_SMP */
2404 * task_oncpu_function_call - call a function on the cpu on which a task runs
2405 * @p: the task to evaluate
2406 * @func: the function to be called
2407 * @info: the function call argument
2409 * Calls the function @func when the task is currently running. This might
2410 * be on the current CPU, which just calls the function directly
2412 void task_oncpu_function_call(struct task_struct *p,
2413 void (*func) (void *info), void *info)
2420 smp_call_function_single(cpu, func, info, 1);
2425 * try_to_wake_up - wake up a thread
2426 * @p: the to-be-woken-up thread
2427 * @state: the mask of task states that can be woken
2428 * @sync: do a synchronous wakeup?
2430 * Put it on the run-queue if it's not already there. The "current"
2431 * thread is always on the run-queue (except when the actual
2432 * re-schedule is in progress), and as such you're allowed to do
2433 * the simpler "current->state = TASK_RUNNING" to mark yourself
2434 * runnable without the overhead of this.
2436 * returns failure only if the task is already active.
2438 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2440 int cpu, orig_cpu, this_cpu, success = 0;
2441 unsigned long flags;
2445 if (!sched_feat(SYNC_WAKEUPS))
2449 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2450 struct sched_domain *sd;
2452 this_cpu = raw_smp_processor_id();
2455 for_each_domain(this_cpu, sd) {
2456 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2465 rq = task_rq_lock(p, &flags);
2466 update_rq_clock(rq);
2467 old_state = p->state;
2468 if (!(old_state & state))
2476 this_cpu = smp_processor_id();
2479 if (unlikely(task_running(rq, p)))
2482 cpu = p->sched_class->select_task_rq(p, sync);
2483 if (cpu != orig_cpu) {
2484 set_task_cpu(p, cpu);
2485 task_rq_unlock(rq, &flags);
2486 /* might preempt at this point */
2487 rq = task_rq_lock(p, &flags);
2488 old_state = p->state;
2489 if (!(old_state & state))
2494 this_cpu = smp_processor_id();
2498 #ifdef CONFIG_SCHEDSTATS
2499 schedstat_inc(rq, ttwu_count);
2500 if (cpu == this_cpu)
2501 schedstat_inc(rq, ttwu_local);
2503 struct sched_domain *sd;
2504 for_each_domain(this_cpu, sd) {
2505 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2506 schedstat_inc(sd, ttwu_wake_remote);
2511 #endif /* CONFIG_SCHEDSTATS */
2514 #endif /* CONFIG_SMP */
2515 schedstat_inc(p, se.nr_wakeups);
2517 schedstat_inc(p, se.nr_wakeups_sync);
2518 if (orig_cpu != cpu)
2519 schedstat_inc(p, se.nr_wakeups_migrate);
2520 if (cpu == this_cpu)
2521 schedstat_inc(p, se.nr_wakeups_local);
2523 schedstat_inc(p, se.nr_wakeups_remote);
2524 activate_task(rq, p, 1);
2528 * Only attribute actual wakeups done by this task.
2530 if (!in_interrupt()) {
2531 struct sched_entity *se = ¤t->se;
2532 u64 sample = se->sum_exec_runtime;
2534 if (se->last_wakeup)
2535 sample -= se->last_wakeup;
2537 sample -= se->start_runtime;
2538 update_avg(&se->avg_wakeup, sample);
2540 se->last_wakeup = se->sum_exec_runtime;
2544 trace_sched_wakeup(rq, p, success);
2545 check_preempt_curr(rq, p, sync);
2547 p->state = TASK_RUNNING;
2549 if (p->sched_class->task_wake_up)
2550 p->sched_class->task_wake_up(rq, p);
2553 task_rq_unlock(rq, &flags);
2559 * wake_up_process - Wake up a specific process
2560 * @p: The process to be woken up.
2562 * Attempt to wake up the nominated process and move it to the set of runnable
2563 * processes. Returns 1 if the process was woken up, 0 if it was already
2566 * It may be assumed that this function implies a write memory barrier before
2567 * changing the task state if and only if any tasks are woken up.
2569 int wake_up_process(struct task_struct *p)
2571 return try_to_wake_up(p, TASK_ALL, 0);
2573 EXPORT_SYMBOL(wake_up_process);
2575 int wake_up_state(struct task_struct *p, unsigned int state)
2577 return try_to_wake_up(p, state, 0);
2581 * Perform scheduler related setup for a newly forked process p.
2582 * p is forked by current.
2584 * __sched_fork() is basic setup used by init_idle() too:
2586 static void __sched_fork(struct task_struct *p)
2588 p->se.exec_start = 0;
2589 p->se.sum_exec_runtime = 0;
2590 p->se.prev_sum_exec_runtime = 0;
2591 p->se.nr_migrations = 0;
2592 p->se.last_wakeup = 0;
2593 p->se.avg_overlap = 0;
2594 p->se.start_runtime = 0;
2595 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2597 #ifdef CONFIG_SCHEDSTATS
2598 p->se.wait_start = 0;
2600 p->se.wait_count = 0;
2603 p->se.sleep_start = 0;
2604 p->se.sleep_max = 0;
2605 p->se.sum_sleep_runtime = 0;
2607 p->se.block_start = 0;
2608 p->se.block_max = 0;
2610 p->se.slice_max = 0;
2612 p->se.nr_migrations_cold = 0;
2613 p->se.nr_failed_migrations_affine = 0;
2614 p->se.nr_failed_migrations_running = 0;
2615 p->se.nr_failed_migrations_hot = 0;
2616 p->se.nr_forced_migrations = 0;
2617 p->se.nr_forced2_migrations = 0;
2619 p->se.nr_wakeups = 0;
2620 p->se.nr_wakeups_sync = 0;
2621 p->se.nr_wakeups_migrate = 0;
2622 p->se.nr_wakeups_local = 0;
2623 p->se.nr_wakeups_remote = 0;
2624 p->se.nr_wakeups_affine = 0;
2625 p->se.nr_wakeups_affine_attempts = 0;
2626 p->se.nr_wakeups_passive = 0;
2627 p->se.nr_wakeups_idle = 0;
2631 INIT_LIST_HEAD(&p->rt.run_list);
2633 INIT_LIST_HEAD(&p->se.group_node);
2635 #ifdef CONFIG_PREEMPT_NOTIFIERS
2636 INIT_HLIST_HEAD(&p->preempt_notifiers);
2640 * We mark the process as running here, but have not actually
2641 * inserted it onto the runqueue yet. This guarantees that
2642 * nobody will actually run it, and a signal or other external
2643 * event cannot wake it up and insert it on the runqueue either.
2645 p->state = TASK_RUNNING;
2649 * fork()/clone()-time setup:
2651 void sched_fork(struct task_struct *p, int clone_flags)
2653 int cpu = get_cpu();
2658 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2660 set_task_cpu(p, cpu);
2663 * Make sure we do not leak PI boosting priority to the child.
2665 p->prio = current->normal_prio;
2668 * Revert to default priority/policy on fork if requested.
2670 if (unlikely(p->sched_reset_on_fork)) {
2671 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR)
2672 p->policy = SCHED_NORMAL;
2674 if (p->normal_prio < DEFAULT_PRIO)
2675 p->prio = DEFAULT_PRIO;
2677 if (PRIO_TO_NICE(p->static_prio) < 0) {
2678 p->static_prio = NICE_TO_PRIO(0);
2683 * We don't need the reset flag anymore after the fork. It has
2684 * fulfilled its duty:
2686 p->sched_reset_on_fork = 0;
2689 if (!rt_prio(p->prio))
2690 p->sched_class = &fair_sched_class;
2692 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2693 if (likely(sched_info_on()))
2694 memset(&p->sched_info, 0, sizeof(p->sched_info));
2696 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2699 #ifdef CONFIG_PREEMPT
2700 /* Want to start with kernel preemption disabled. */
2701 task_thread_info(p)->preempt_count = 1;
2703 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2709 * wake_up_new_task - wake up a newly created task for the first time.
2711 * This function will do some initial scheduler statistics housekeeping
2712 * that must be done for every newly created context, then puts the task
2713 * on the runqueue and wakes it.
2715 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2717 unsigned long flags;
2720 rq = task_rq_lock(p, &flags);
2721 BUG_ON(p->state != TASK_RUNNING);
2722 update_rq_clock(rq);
2724 p->prio = effective_prio(p);
2726 if (!p->sched_class->task_new || !current->se.on_rq) {
2727 activate_task(rq, p, 0);
2730 * Let the scheduling class do new task startup
2731 * management (if any):
2733 p->sched_class->task_new(rq, p);
2736 trace_sched_wakeup_new(rq, p, 1);
2737 check_preempt_curr(rq, p, 0);
2739 if (p->sched_class->task_wake_up)
2740 p->sched_class->task_wake_up(rq, p);
2742 task_rq_unlock(rq, &flags);
2745 #ifdef CONFIG_PREEMPT_NOTIFIERS
2748 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2749 * @notifier: notifier struct to register
2751 void preempt_notifier_register(struct preempt_notifier *notifier)
2753 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2755 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2758 * preempt_notifier_unregister - no longer interested in preemption notifications
2759 * @notifier: notifier struct to unregister
2761 * This is safe to call from within a preemption notifier.
2763 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2765 hlist_del(¬ifier->link);
2767 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2769 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2771 struct preempt_notifier *notifier;
2772 struct hlist_node *node;
2774 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2775 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2779 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2780 struct task_struct *next)
2782 struct preempt_notifier *notifier;
2783 struct hlist_node *node;
2785 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2786 notifier->ops->sched_out(notifier, next);
2789 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2791 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2796 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2797 struct task_struct *next)
2801 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2804 * prepare_task_switch - prepare to switch tasks
2805 * @rq: the runqueue preparing to switch
2806 * @prev: the current task that is being switched out
2807 * @next: the task we are going to switch to.
2809 * This is called with the rq lock held and interrupts off. It must
2810 * be paired with a subsequent finish_task_switch after the context
2813 * prepare_task_switch sets up locking and calls architecture specific
2817 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2818 struct task_struct *next)
2820 fire_sched_out_preempt_notifiers(prev, next);
2821 prepare_lock_switch(rq, next);
2822 prepare_arch_switch(next);
2826 * finish_task_switch - clean up after a task-switch
2827 * @rq: runqueue associated with task-switch
2828 * @prev: the thread we just switched away from.
2830 * finish_task_switch must be called after the context switch, paired
2831 * with a prepare_task_switch call before the context switch.
2832 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2833 * and do any other architecture-specific cleanup actions.
2835 * Note that we may have delayed dropping an mm in context_switch(). If
2836 * so, we finish that here outside of the runqueue lock. (Doing it
2837 * with the lock held can cause deadlocks; see schedule() for
2840 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2841 __releases(rq->lock)
2843 struct mm_struct *mm = rq->prev_mm;
2849 * A task struct has one reference for the use as "current".
2850 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2851 * schedule one last time. The schedule call will never return, and
2852 * the scheduled task must drop that reference.
2853 * The test for TASK_DEAD must occur while the runqueue locks are
2854 * still held, otherwise prev could be scheduled on another cpu, die
2855 * there before we look at prev->state, and then the reference would
2857 * Manfred Spraul <manfred@colorfullife.com>
2859 prev_state = prev->state;
2860 finish_arch_switch(prev);
2861 perf_counter_task_sched_in(current, cpu_of(rq));
2862 finish_lock_switch(rq, prev);
2864 fire_sched_in_preempt_notifiers(current);
2867 if (unlikely(prev_state == TASK_DEAD)) {
2869 * Remove function-return probe instances associated with this
2870 * task and put them back on the free list.
2872 kprobe_flush_task(prev);
2873 put_task_struct(prev);
2879 /* assumes rq->lock is held */
2880 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2882 if (prev->sched_class->pre_schedule)
2883 prev->sched_class->pre_schedule(rq, prev);
2886 /* rq->lock is NOT held, but preemption is disabled */
2887 static inline void post_schedule(struct rq *rq)
2889 if (rq->post_schedule) {
2890 unsigned long flags;
2892 spin_lock_irqsave(&rq->lock, flags);
2893 if (rq->curr->sched_class->post_schedule)
2894 rq->curr->sched_class->post_schedule(rq);
2895 spin_unlock_irqrestore(&rq->lock, flags);
2897 rq->post_schedule = 0;
2903 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2907 static inline void post_schedule(struct rq *rq)
2914 * schedule_tail - first thing a freshly forked thread must call.
2915 * @prev: the thread we just switched away from.
2917 asmlinkage void schedule_tail(struct task_struct *prev)
2918 __releases(rq->lock)
2920 struct rq *rq = this_rq();
2922 finish_task_switch(rq, prev);
2925 * FIXME: do we need to worry about rq being invalidated by the
2930 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2931 /* In this case, finish_task_switch does not reenable preemption */
2934 if (current->set_child_tid)
2935 put_user(task_pid_vnr(current), current->set_child_tid);
2939 * context_switch - switch to the new MM and the new
2940 * thread's register state.
2943 context_switch(struct rq *rq, struct task_struct *prev,
2944 struct task_struct *next)
2946 struct mm_struct *mm, *oldmm;
2948 prepare_task_switch(rq, prev, next);
2949 trace_sched_switch(rq, prev, next);
2951 oldmm = prev->active_mm;
2953 * For paravirt, this is coupled with an exit in switch_to to
2954 * combine the page table reload and the switch backend into
2957 arch_start_context_switch(prev);
2959 if (unlikely(!mm)) {
2960 next->active_mm = oldmm;
2961 atomic_inc(&oldmm->mm_count);
2962 enter_lazy_tlb(oldmm, next);
2964 switch_mm(oldmm, mm, next);
2966 if (unlikely(!prev->mm)) {
2967 prev->active_mm = NULL;
2968 rq->prev_mm = oldmm;
2971 * Since the runqueue lock will be released by the next
2972 * task (which is an invalid locking op but in the case
2973 * of the scheduler it's an obvious special-case), so we
2974 * do an early lockdep release here:
2976 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2977 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2980 /* Here we just switch the register state and the stack. */
2981 switch_to(prev, next, prev);
2985 * this_rq must be evaluated again because prev may have moved
2986 * CPUs since it called schedule(), thus the 'rq' on its stack
2987 * frame will be invalid.
2989 finish_task_switch(this_rq(), prev);
2993 * nr_running, nr_uninterruptible and nr_context_switches:
2995 * externally visible scheduler statistics: current number of runnable
2996 * threads, current number of uninterruptible-sleeping threads, total
2997 * number of context switches performed since bootup.
2999 unsigned long nr_running(void)
3001 unsigned long i, sum = 0;
3003 for_each_online_cpu(i)
3004 sum += cpu_rq(i)->nr_running;
3009 unsigned long nr_uninterruptible(void)
3011 unsigned long i, sum = 0;
3013 for_each_possible_cpu(i)
3014 sum += cpu_rq(i)->nr_uninterruptible;
3017 * Since we read the counters lockless, it might be slightly
3018 * inaccurate. Do not allow it to go below zero though:
3020 if (unlikely((long)sum < 0))
3026 unsigned long long nr_context_switches(void)
3029 unsigned long long sum = 0;
3031 for_each_possible_cpu(i)
3032 sum += cpu_rq(i)->nr_switches;
3037 unsigned long nr_iowait(void)
3039 unsigned long i, sum = 0;
3041 for_each_possible_cpu(i)
3042 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3047 /* Variables and functions for calc_load */
3048 static atomic_long_t calc_load_tasks;
3049 static unsigned long calc_load_update;
3050 unsigned long avenrun[3];
3051 EXPORT_SYMBOL(avenrun);
3054 * get_avenrun - get the load average array
3055 * @loads: pointer to dest load array
3056 * @offset: offset to add
3057 * @shift: shift count to shift the result left
3059 * These values are estimates at best, so no need for locking.
3061 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3063 loads[0] = (avenrun[0] + offset) << shift;
3064 loads[1] = (avenrun[1] + offset) << shift;
3065 loads[2] = (avenrun[2] + offset) << shift;
3068 static unsigned long
3069 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3072 load += active * (FIXED_1 - exp);
3073 return load >> FSHIFT;
3077 * calc_load - update the avenrun load estimates 10 ticks after the
3078 * CPUs have updated calc_load_tasks.
3080 void calc_global_load(void)
3082 unsigned long upd = calc_load_update + 10;
3085 if (time_before(jiffies, upd))
3088 active = atomic_long_read(&calc_load_tasks);
3089 active = active > 0 ? active * FIXED_1 : 0;
3091 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3092 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3093 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3095 calc_load_update += LOAD_FREQ;
3099 * Either called from update_cpu_load() or from a cpu going idle
3101 static void calc_load_account_active(struct rq *this_rq)
3103 long nr_active, delta;
3105 nr_active = this_rq->nr_running;
3106 nr_active += (long) this_rq->nr_uninterruptible;
3108 if (nr_active != this_rq->calc_load_active) {
3109 delta = nr_active - this_rq->calc_load_active;
3110 this_rq->calc_load_active = nr_active;
3111 atomic_long_add(delta, &calc_load_tasks);
3116 * Externally visible per-cpu scheduler statistics:
3117 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3119 u64 cpu_nr_migrations(int cpu)
3121 return cpu_rq(cpu)->nr_migrations_in;
3125 * Update rq->cpu_load[] statistics. This function is usually called every
3126 * scheduler tick (TICK_NSEC).
3128 static void update_cpu_load(struct rq *this_rq)
3130 unsigned long this_load = this_rq->load.weight;
3133 this_rq->nr_load_updates++;
3135 /* Update our load: */
3136 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3137 unsigned long old_load, new_load;
3139 /* scale is effectively 1 << i now, and >> i divides by scale */
3141 old_load = this_rq->cpu_load[i];
3142 new_load = this_load;
3144 * Round up the averaging division if load is increasing. This
3145 * prevents us from getting stuck on 9 if the load is 10, for
3148 if (new_load > old_load)
3149 new_load += scale-1;
3150 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3153 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3154 this_rq->calc_load_update += LOAD_FREQ;
3155 calc_load_account_active(this_rq);
3162 * double_rq_lock - safely lock two runqueues
3164 * Note this does not disable interrupts like task_rq_lock,
3165 * you need to do so manually before calling.
3167 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3168 __acquires(rq1->lock)
3169 __acquires(rq2->lock)
3171 BUG_ON(!irqs_disabled());
3173 spin_lock(&rq1->lock);
3174 __acquire(rq2->lock); /* Fake it out ;) */
3177 spin_lock(&rq1->lock);
3178 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3180 spin_lock(&rq2->lock);
3181 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3184 update_rq_clock(rq1);
3185 update_rq_clock(rq2);
3189 * double_rq_unlock - safely unlock two runqueues
3191 * Note this does not restore interrupts like task_rq_unlock,
3192 * you need to do so manually after calling.
3194 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3195 __releases(rq1->lock)
3196 __releases(rq2->lock)
3198 spin_unlock(&rq1->lock);
3200 spin_unlock(&rq2->lock);
3202 __release(rq2->lock);
3206 * If dest_cpu is allowed for this process, migrate the task to it.
3207 * This is accomplished by forcing the cpu_allowed mask to only
3208 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3209 * the cpu_allowed mask is restored.
3211 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3213 struct migration_req req;
3214 unsigned long flags;
3217 rq = task_rq_lock(p, &flags);
3218 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3219 || unlikely(!cpu_active(dest_cpu)))
3222 /* force the process onto the specified CPU */
3223 if (migrate_task(p, dest_cpu, &req)) {
3224 /* Need to wait for migration thread (might exit: take ref). */
3225 struct task_struct *mt = rq->migration_thread;
3227 get_task_struct(mt);
3228 task_rq_unlock(rq, &flags);
3229 wake_up_process(mt);
3230 put_task_struct(mt);
3231 wait_for_completion(&req.done);
3236 task_rq_unlock(rq, &flags);
3240 * sched_exec - execve() is a valuable balancing opportunity, because at
3241 * this point the task has the smallest effective memory and cache footprint.
3243 void sched_exec(void)
3245 int new_cpu, this_cpu = get_cpu();
3246 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3248 if (new_cpu != this_cpu)
3249 sched_migrate_task(current, new_cpu);
3253 * pull_task - move a task from a remote runqueue to the local runqueue.
3254 * Both runqueues must be locked.
3256 static void pull_task(struct rq *src_rq, struct task_struct *p,
3257 struct rq *this_rq, int this_cpu)
3259 deactivate_task(src_rq, p, 0);
3260 set_task_cpu(p, this_cpu);
3261 activate_task(this_rq, p, 0);
3263 * Note that idle threads have a prio of MAX_PRIO, for this test
3264 * to be always true for them.
3266 check_preempt_curr(this_rq, p, 0);
3270 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3273 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3274 struct sched_domain *sd, enum cpu_idle_type idle,
3277 int tsk_cache_hot = 0;
3279 * We do not migrate tasks that are:
3280 * 1) running (obviously), or
3281 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3282 * 3) are cache-hot on their current CPU.
3284 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3285 schedstat_inc(p, se.nr_failed_migrations_affine);
3290 if (task_running(rq, p)) {
3291 schedstat_inc(p, se.nr_failed_migrations_running);
3296 * Aggressive migration if:
3297 * 1) task is cache cold, or
3298 * 2) too many balance attempts have failed.
3301 tsk_cache_hot = task_hot(p, rq->clock, sd);
3302 if (!tsk_cache_hot ||
3303 sd->nr_balance_failed > sd->cache_nice_tries) {
3304 #ifdef CONFIG_SCHEDSTATS
3305 if (tsk_cache_hot) {
3306 schedstat_inc(sd, lb_hot_gained[idle]);
3307 schedstat_inc(p, se.nr_forced_migrations);
3313 if (tsk_cache_hot) {
3314 schedstat_inc(p, se.nr_failed_migrations_hot);
3320 static unsigned long
3321 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3322 unsigned long max_load_move, struct sched_domain *sd,
3323 enum cpu_idle_type idle, int *all_pinned,
3324 int *this_best_prio, struct rq_iterator *iterator)
3326 int loops = 0, pulled = 0, pinned = 0;
3327 struct task_struct *p;
3328 long rem_load_move = max_load_move;
3330 if (max_load_move == 0)
3336 * Start the load-balancing iterator:
3338 p = iterator->start(iterator->arg);
3340 if (!p || loops++ > sysctl_sched_nr_migrate)
3343 if ((p->se.load.weight >> 1) > rem_load_move ||
3344 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3345 p = iterator->next(iterator->arg);
3349 pull_task(busiest, p, this_rq, this_cpu);
3351 rem_load_move -= p->se.load.weight;
3353 #ifdef CONFIG_PREEMPT
3355 * NEWIDLE balancing is a source of latency, so preemptible kernels
3356 * will stop after the first task is pulled to minimize the critical
3359 if (idle == CPU_NEWLY_IDLE)
3364 * We only want to steal up to the prescribed amount of weighted load.
3366 if (rem_load_move > 0) {
3367 if (p->prio < *this_best_prio)
3368 *this_best_prio = p->prio;
3369 p = iterator->next(iterator->arg);
3374 * Right now, this is one of only two places pull_task() is called,
3375 * so we can safely collect pull_task() stats here rather than
3376 * inside pull_task().
3378 schedstat_add(sd, lb_gained[idle], pulled);
3381 *all_pinned = pinned;
3383 return max_load_move - rem_load_move;
3387 * move_tasks tries to move up to max_load_move weighted load from busiest to
3388 * this_rq, as part of a balancing operation within domain "sd".
3389 * Returns 1 if successful and 0 otherwise.
3391 * Called with both runqueues locked.
3393 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3394 unsigned long max_load_move,
3395 struct sched_domain *sd, enum cpu_idle_type idle,
3398 const struct sched_class *class = sched_class_highest;
3399 unsigned long total_load_moved = 0;
3400 int this_best_prio = this_rq->curr->prio;
3404 class->load_balance(this_rq, this_cpu, busiest,
3405 max_load_move - total_load_moved,
3406 sd, idle, all_pinned, &this_best_prio);
3407 class = class->next;
3409 #ifdef CONFIG_PREEMPT
3411 * NEWIDLE balancing is a source of latency, so preemptible
3412 * kernels will stop after the first task is pulled to minimize
3413 * the critical section.
3415 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3418 } while (class && max_load_move > total_load_moved);
3420 return total_load_moved > 0;
3424 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3425 struct sched_domain *sd, enum cpu_idle_type idle,
3426 struct rq_iterator *iterator)
3428 struct task_struct *p = iterator->start(iterator->arg);
3432 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3433 pull_task(busiest, p, this_rq, this_cpu);
3435 * Right now, this is only the second place pull_task()
3436 * is called, so we can safely collect pull_task()
3437 * stats here rather than inside pull_task().
3439 schedstat_inc(sd, lb_gained[idle]);
3443 p = iterator->next(iterator->arg);
3450 * move_one_task tries to move exactly one task from busiest to this_rq, as
3451 * part of active balancing operations within "domain".
3452 * Returns 1 if successful and 0 otherwise.
3454 * Called with both runqueues locked.
3456 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3457 struct sched_domain *sd, enum cpu_idle_type idle)
3459 const struct sched_class *class;
3461 for_each_class(class) {
3462 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3468 /********** Helpers for find_busiest_group ************************/
3470 * sd_lb_stats - Structure to store the statistics of a sched_domain
3471 * during load balancing.
3473 struct sd_lb_stats {
3474 struct sched_group *busiest; /* Busiest group in this sd */
3475 struct sched_group *this; /* Local group in this sd */
3476 unsigned long total_load; /* Total load of all groups in sd */
3477 unsigned long total_pwr; /* Total power of all groups in sd */
3478 unsigned long avg_load; /* Average load across all groups in sd */
3480 /** Statistics of this group */
3481 unsigned long this_load;
3482 unsigned long this_load_per_task;
3483 unsigned long this_nr_running;
3485 /* Statistics of the busiest group */
3486 unsigned long max_load;
3487 unsigned long busiest_load_per_task;
3488 unsigned long busiest_nr_running;
3490 int group_imb; /* Is there imbalance in this sd */
3491 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3492 int power_savings_balance; /* Is powersave balance needed for this sd */
3493 struct sched_group *group_min; /* Least loaded group in sd */
3494 struct sched_group *group_leader; /* Group which relieves group_min */
3495 unsigned long min_load_per_task; /* load_per_task in group_min */
3496 unsigned long leader_nr_running; /* Nr running of group_leader */
3497 unsigned long min_nr_running; /* Nr running of group_min */
3502 * sg_lb_stats - stats of a sched_group required for load_balancing
3504 struct sg_lb_stats {
3505 unsigned long avg_load; /*Avg load across the CPUs of the group */
3506 unsigned long group_load; /* Total load over the CPUs of the group */
3507 unsigned long sum_nr_running; /* Nr tasks running in the group */
3508 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3509 unsigned long group_capacity;
3510 int group_imb; /* Is there an imbalance in the group ? */
3514 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3515 * @group: The group whose first cpu is to be returned.
3517 static inline unsigned int group_first_cpu(struct sched_group *group)
3519 return cpumask_first(sched_group_cpus(group));
3523 * get_sd_load_idx - Obtain the load index for a given sched domain.
3524 * @sd: The sched_domain whose load_idx is to be obtained.
3525 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3527 static inline int get_sd_load_idx(struct sched_domain *sd,
3528 enum cpu_idle_type idle)
3534 load_idx = sd->busy_idx;
3537 case CPU_NEWLY_IDLE:
3538 load_idx = sd->newidle_idx;
3541 load_idx = sd->idle_idx;
3549 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3551 * init_sd_power_savings_stats - Initialize power savings statistics for
3552 * the given sched_domain, during load balancing.
3554 * @sd: Sched domain whose power-savings statistics are to be initialized.
3555 * @sds: Variable containing the statistics for sd.
3556 * @idle: Idle status of the CPU at which we're performing load-balancing.
3558 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3559 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3562 * Busy processors will not participate in power savings
3565 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3566 sds->power_savings_balance = 0;
3568 sds->power_savings_balance = 1;
3569 sds->min_nr_running = ULONG_MAX;
3570 sds->leader_nr_running = 0;
3575 * update_sd_power_savings_stats - Update the power saving stats for a
3576 * sched_domain while performing load balancing.
3578 * @group: sched_group belonging to the sched_domain under consideration.
3579 * @sds: Variable containing the statistics of the sched_domain
3580 * @local_group: Does group contain the CPU for which we're performing
3582 * @sgs: Variable containing the statistics of the group.
3584 static inline void update_sd_power_savings_stats(struct sched_group *group,
3585 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3588 if (!sds->power_savings_balance)
3592 * If the local group is idle or completely loaded
3593 * no need to do power savings balance at this domain
3595 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3596 !sds->this_nr_running))
3597 sds->power_savings_balance = 0;
3600 * If a group is already running at full capacity or idle,
3601 * don't include that group in power savings calculations
3603 if (!sds->power_savings_balance ||
3604 sgs->sum_nr_running >= sgs->group_capacity ||
3605 !sgs->sum_nr_running)
3609 * Calculate the group which has the least non-idle load.
3610 * This is the group from where we need to pick up the load
3613 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3614 (sgs->sum_nr_running == sds->min_nr_running &&
3615 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3616 sds->group_min = group;
3617 sds->min_nr_running = sgs->sum_nr_running;
3618 sds->min_load_per_task = sgs->sum_weighted_load /
3619 sgs->sum_nr_running;
3623 * Calculate the group which is almost near its
3624 * capacity but still has some space to pick up some load
3625 * from other group and save more power
3627 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3630 if (sgs->sum_nr_running > sds->leader_nr_running ||
3631 (sgs->sum_nr_running == sds->leader_nr_running &&
3632 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3633 sds->group_leader = group;
3634 sds->leader_nr_running = sgs->sum_nr_running;
3639 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3640 * @sds: Variable containing the statistics of the sched_domain
3641 * under consideration.
3642 * @this_cpu: Cpu at which we're currently performing load-balancing.
3643 * @imbalance: Variable to store the imbalance.
3646 * Check if we have potential to perform some power-savings balance.
3647 * If yes, set the busiest group to be the least loaded group in the
3648 * sched_domain, so that it's CPUs can be put to idle.
3650 * Returns 1 if there is potential to perform power-savings balance.
3653 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3654 int this_cpu, unsigned long *imbalance)
3656 if (!sds->power_savings_balance)
3659 if (sds->this != sds->group_leader ||
3660 sds->group_leader == sds->group_min)
3663 *imbalance = sds->min_load_per_task;
3664 sds->busiest = sds->group_min;
3666 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3667 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3668 group_first_cpu(sds->group_leader);
3674 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3675 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3676 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3681 static inline void update_sd_power_savings_stats(struct sched_group *group,
3682 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3687 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3688 int this_cpu, unsigned long *imbalance)
3692 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3696 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3697 * @group: sched_group whose statistics are to be updated.
3698 * @this_cpu: Cpu for which load balance is currently performed.
3699 * @idle: Idle status of this_cpu
3700 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3701 * @sd_idle: Idle status of the sched_domain containing group.
3702 * @local_group: Does group contain this_cpu.
3703 * @cpus: Set of cpus considered for load balancing.
3704 * @balance: Should we balance.
3705 * @sgs: variable to hold the statistics for this group.
3707 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3708 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3709 int local_group, const struct cpumask *cpus,
3710 int *balance, struct sg_lb_stats *sgs)
3712 unsigned long load, max_cpu_load, min_cpu_load;
3714 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3715 unsigned long sum_avg_load_per_task;
3716 unsigned long avg_load_per_task;
3719 balance_cpu = group_first_cpu(group);
3721 /* Tally up the load of all CPUs in the group */
3722 sum_avg_load_per_task = avg_load_per_task = 0;
3724 min_cpu_load = ~0UL;
3726 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3727 struct rq *rq = cpu_rq(i);
3729 if (*sd_idle && rq->nr_running)
3732 /* Bias balancing toward cpus of our domain */
3734 if (idle_cpu(i) && !first_idle_cpu) {
3739 load = target_load(i, load_idx);
3741 load = source_load(i, load_idx);
3742 if (load > max_cpu_load)
3743 max_cpu_load = load;
3744 if (min_cpu_load > load)
3745 min_cpu_load = load;
3748 sgs->group_load += load;
3749 sgs->sum_nr_running += rq->nr_running;
3750 sgs->sum_weighted_load += weighted_cpuload(i);
3752 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3756 * First idle cpu or the first cpu(busiest) in this sched group
3757 * is eligible for doing load balancing at this and above
3758 * domains. In the newly idle case, we will allow all the cpu's
3759 * to do the newly idle load balance.
3761 if (idle != CPU_NEWLY_IDLE && local_group &&
3762 balance_cpu != this_cpu && balance) {
3767 /* Adjust by relative CPU power of the group */
3768 sgs->avg_load = sg_div_cpu_power(group,
3769 sgs->group_load * SCHED_LOAD_SCALE);
3773 * Consider the group unbalanced when the imbalance is larger
3774 * than the average weight of two tasks.
3776 * APZ: with cgroup the avg task weight can vary wildly and
3777 * might not be a suitable number - should we keep a
3778 * normalized nr_running number somewhere that negates
3781 avg_load_per_task = sg_div_cpu_power(group,
3782 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3784 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3787 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3792 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3793 * @sd: sched_domain whose statistics are to be updated.
3794 * @this_cpu: Cpu for which load balance is currently performed.
3795 * @idle: Idle status of this_cpu
3796 * @sd_idle: Idle status of the sched_domain containing group.
3797 * @cpus: Set of cpus considered for load balancing.
3798 * @balance: Should we balance.
3799 * @sds: variable to hold the statistics for this sched_domain.
3801 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3802 enum cpu_idle_type idle, int *sd_idle,
3803 const struct cpumask *cpus, int *balance,
3804 struct sd_lb_stats *sds)
3806 struct sched_group *group = sd->groups;
3807 struct sg_lb_stats sgs;
3810 init_sd_power_savings_stats(sd, sds, idle);
3811 load_idx = get_sd_load_idx(sd, idle);
3816 local_group = cpumask_test_cpu(this_cpu,
3817 sched_group_cpus(group));
3818 memset(&sgs, 0, sizeof(sgs));
3819 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3820 local_group, cpus, balance, &sgs);
3822 if (local_group && balance && !(*balance))
3825 sds->total_load += sgs.group_load;
3826 sds->total_pwr += group->__cpu_power;
3829 sds->this_load = sgs.avg_load;
3831 sds->this_nr_running = sgs.sum_nr_running;
3832 sds->this_load_per_task = sgs.sum_weighted_load;
3833 } else if (sgs.avg_load > sds->max_load &&
3834 (sgs.sum_nr_running > sgs.group_capacity ||
3836 sds->max_load = sgs.avg_load;
3837 sds->busiest = group;
3838 sds->busiest_nr_running = sgs.sum_nr_running;
3839 sds->busiest_load_per_task = sgs.sum_weighted_load;
3840 sds->group_imb = sgs.group_imb;
3843 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3844 group = group->next;
3845 } while (group != sd->groups);
3850 * fix_small_imbalance - Calculate the minor imbalance that exists
3851 * amongst the groups of a sched_domain, during
3853 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3854 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3855 * @imbalance: Variable to store the imbalance.
3857 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3858 int this_cpu, unsigned long *imbalance)
3860 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3861 unsigned int imbn = 2;
3863 if (sds->this_nr_running) {
3864 sds->this_load_per_task /= sds->this_nr_running;
3865 if (sds->busiest_load_per_task >
3866 sds->this_load_per_task)
3869 sds->this_load_per_task =
3870 cpu_avg_load_per_task(this_cpu);
3872 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3873 sds->busiest_load_per_task * imbn) {
3874 *imbalance = sds->busiest_load_per_task;
3879 * OK, we don't have enough imbalance to justify moving tasks,
3880 * however we may be able to increase total CPU power used by
3884 pwr_now += sds->busiest->__cpu_power *
3885 min(sds->busiest_load_per_task, sds->max_load);
3886 pwr_now += sds->this->__cpu_power *
3887 min(sds->this_load_per_task, sds->this_load);
3888 pwr_now /= SCHED_LOAD_SCALE;
3890 /* Amount of load we'd subtract */
3891 tmp = sg_div_cpu_power(sds->busiest,
3892 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3893 if (sds->max_load > tmp)
3894 pwr_move += sds->busiest->__cpu_power *
3895 min(sds->busiest_load_per_task, sds->max_load - tmp);
3897 /* Amount of load we'd add */
3898 if (sds->max_load * sds->busiest->__cpu_power <
3899 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3900 tmp = sg_div_cpu_power(sds->this,
3901 sds->max_load * sds->busiest->__cpu_power);
3903 tmp = sg_div_cpu_power(sds->this,
3904 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3905 pwr_move += sds->this->__cpu_power *
3906 min(sds->this_load_per_task, sds->this_load + tmp);
3907 pwr_move /= SCHED_LOAD_SCALE;
3909 /* Move if we gain throughput */
3910 if (pwr_move > pwr_now)
3911 *imbalance = sds->busiest_load_per_task;
3915 * calculate_imbalance - Calculate the amount of imbalance present within the
3916 * groups of a given sched_domain during load balance.
3917 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3918 * @this_cpu: Cpu for which currently load balance is being performed.
3919 * @imbalance: The variable to store the imbalance.
3921 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3922 unsigned long *imbalance)
3924 unsigned long max_pull;
3926 * In the presence of smp nice balancing, certain scenarios can have
3927 * max load less than avg load(as we skip the groups at or below
3928 * its cpu_power, while calculating max_load..)
3930 if (sds->max_load < sds->avg_load) {
3932 return fix_small_imbalance(sds, this_cpu, imbalance);
3935 /* Don't want to pull so many tasks that a group would go idle */
3936 max_pull = min(sds->max_load - sds->avg_load,
3937 sds->max_load - sds->busiest_load_per_task);
3939 /* How much load to actually move to equalise the imbalance */
3940 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3941 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3945 * if *imbalance is less than the average load per runnable task
3946 * there is no gaurantee that any tasks will be moved so we'll have
3947 * a think about bumping its value to force at least one task to be
3950 if (*imbalance < sds->busiest_load_per_task)
3951 return fix_small_imbalance(sds, this_cpu, imbalance);
3954 /******* find_busiest_group() helpers end here *********************/
3957 * find_busiest_group - Returns the busiest group within the sched_domain
3958 * if there is an imbalance. If there isn't an imbalance, and
3959 * the user has opted for power-savings, it returns a group whose
3960 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3961 * such a group exists.
3963 * Also calculates the amount of weighted load which should be moved
3964 * to restore balance.
3966 * @sd: The sched_domain whose busiest group is to be returned.
3967 * @this_cpu: The cpu for which load balancing is currently being performed.
3968 * @imbalance: Variable which stores amount of weighted load which should
3969 * be moved to restore balance/put a group to idle.
3970 * @idle: The idle status of this_cpu.
3971 * @sd_idle: The idleness of sd
3972 * @cpus: The set of CPUs under consideration for load-balancing.
3973 * @balance: Pointer to a variable indicating if this_cpu
3974 * is the appropriate cpu to perform load balancing at this_level.
3976 * Returns: - the busiest group if imbalance exists.
3977 * - If no imbalance and user has opted for power-savings balance,
3978 * return the least loaded group whose CPUs can be
3979 * put to idle by rebalancing its tasks onto our group.
3981 static struct sched_group *
3982 find_busiest_group(struct sched_domain *sd, int this_cpu,
3983 unsigned long *imbalance, enum cpu_idle_type idle,
3984 int *sd_idle, const struct cpumask *cpus, int *balance)
3986 struct sd_lb_stats sds;
3988 memset(&sds, 0, sizeof(sds));
3991 * Compute the various statistics relavent for load balancing at
3994 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3997 /* Cases where imbalance does not exist from POV of this_cpu */
3998 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4000 * 2) There is no busy sibling group to pull from.
4001 * 3) This group is the busiest group.
4002 * 4) This group is more busy than the avg busieness at this
4004 * 5) The imbalance is within the specified limit.
4005 * 6) Any rebalance would lead to ping-pong
4007 if (balance && !(*balance))
4010 if (!sds.busiest || sds.busiest_nr_running == 0)
4013 if (sds.this_load >= sds.max_load)
4016 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4018 if (sds.this_load >= sds.avg_load)
4021 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4024 sds.busiest_load_per_task /= sds.busiest_nr_running;
4026 sds.busiest_load_per_task =
4027 min(sds.busiest_load_per_task, sds.avg_load);
4030 * We're trying to get all the cpus to the average_load, so we don't
4031 * want to push ourselves above the average load, nor do we wish to
4032 * reduce the max loaded cpu below the average load, as either of these
4033 * actions would just result in more rebalancing later, and ping-pong
4034 * tasks around. Thus we look for the minimum possible imbalance.
4035 * Negative imbalances (*we* are more loaded than anyone else) will
4036 * be counted as no imbalance for these purposes -- we can't fix that
4037 * by pulling tasks to us. Be careful of negative numbers as they'll
4038 * appear as very large values with unsigned longs.
4040 if (sds.max_load <= sds.busiest_load_per_task)
4043 /* Looks like there is an imbalance. Compute it */
4044 calculate_imbalance(&sds, this_cpu, imbalance);
4049 * There is no obvious imbalance. But check if we can do some balancing
4052 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4060 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4063 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4064 unsigned long imbalance, const struct cpumask *cpus)
4066 struct rq *busiest = NULL, *rq;
4067 unsigned long max_load = 0;
4070 for_each_cpu(i, sched_group_cpus(group)) {
4073 if (!cpumask_test_cpu(i, cpus))
4077 wl = weighted_cpuload(i);
4079 if (rq->nr_running == 1 && wl > imbalance)
4082 if (wl > max_load) {
4092 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4093 * so long as it is large enough.
4095 #define MAX_PINNED_INTERVAL 512
4097 /* Working cpumask for load_balance and load_balance_newidle. */
4098 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4101 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4102 * tasks if there is an imbalance.
4104 static int load_balance(int this_cpu, struct rq *this_rq,
4105 struct sched_domain *sd, enum cpu_idle_type idle,
4108 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4109 struct sched_group *group;
4110 unsigned long imbalance;
4112 unsigned long flags;
4113 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4115 cpumask_setall(cpus);
4118 * When power savings policy is enabled for the parent domain, idle
4119 * sibling can pick up load irrespective of busy siblings. In this case,
4120 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4121 * portraying it as CPU_NOT_IDLE.
4123 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4124 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4127 schedstat_inc(sd, lb_count[idle]);
4131 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4138 schedstat_inc(sd, lb_nobusyg[idle]);
4142 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4144 schedstat_inc(sd, lb_nobusyq[idle]);
4148 BUG_ON(busiest == this_rq);
4150 schedstat_add(sd, lb_imbalance[idle], imbalance);
4153 if (busiest->nr_running > 1) {
4155 * Attempt to move tasks. If find_busiest_group has found
4156 * an imbalance but busiest->nr_running <= 1, the group is
4157 * still unbalanced. ld_moved simply stays zero, so it is
4158 * correctly treated as an imbalance.
4160 local_irq_save(flags);
4161 double_rq_lock(this_rq, busiest);
4162 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4163 imbalance, sd, idle, &all_pinned);
4164 double_rq_unlock(this_rq, busiest);
4165 local_irq_restore(flags);
4168 * some other cpu did the load balance for us.
4170 if (ld_moved && this_cpu != smp_processor_id())
4171 resched_cpu(this_cpu);
4173 /* All tasks on this runqueue were pinned by CPU affinity */
4174 if (unlikely(all_pinned)) {
4175 cpumask_clear_cpu(cpu_of(busiest), cpus);
4176 if (!cpumask_empty(cpus))
4183 schedstat_inc(sd, lb_failed[idle]);
4184 sd->nr_balance_failed++;
4186 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4188 spin_lock_irqsave(&busiest->lock, flags);
4190 /* don't kick the migration_thread, if the curr
4191 * task on busiest cpu can't be moved to this_cpu
4193 if (!cpumask_test_cpu(this_cpu,
4194 &busiest->curr->cpus_allowed)) {
4195 spin_unlock_irqrestore(&busiest->lock, flags);
4197 goto out_one_pinned;
4200 if (!busiest->active_balance) {
4201 busiest->active_balance = 1;
4202 busiest->push_cpu = this_cpu;
4205 spin_unlock_irqrestore(&busiest->lock, flags);
4207 wake_up_process(busiest->migration_thread);
4210 * We've kicked active balancing, reset the failure
4213 sd->nr_balance_failed = sd->cache_nice_tries+1;
4216 sd->nr_balance_failed = 0;
4218 if (likely(!active_balance)) {
4219 /* We were unbalanced, so reset the balancing interval */
4220 sd->balance_interval = sd->min_interval;
4223 * If we've begun active balancing, start to back off. This
4224 * case may not be covered by the all_pinned logic if there
4225 * is only 1 task on the busy runqueue (because we don't call
4228 if (sd->balance_interval < sd->max_interval)
4229 sd->balance_interval *= 2;
4232 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4233 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4239 schedstat_inc(sd, lb_balanced[idle]);
4241 sd->nr_balance_failed = 0;
4244 /* tune up the balancing interval */
4245 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4246 (sd->balance_interval < sd->max_interval))
4247 sd->balance_interval *= 2;
4249 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4250 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4261 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4262 * tasks if there is an imbalance.
4264 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4265 * this_rq is locked.
4268 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4270 struct sched_group *group;
4271 struct rq *busiest = NULL;
4272 unsigned long imbalance;
4276 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4278 cpumask_setall(cpus);
4281 * When power savings policy is enabled for the parent domain, idle
4282 * sibling can pick up load irrespective of busy siblings. In this case,
4283 * let the state of idle sibling percolate up as IDLE, instead of
4284 * portraying it as CPU_NOT_IDLE.
4286 if (sd->flags & SD_SHARE_CPUPOWER &&
4287 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4290 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4292 update_shares_locked(this_rq, sd);
4293 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4294 &sd_idle, cpus, NULL);
4296 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4300 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4302 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4306 BUG_ON(busiest == this_rq);
4308 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4311 if (busiest->nr_running > 1) {
4312 /* Attempt to move tasks */
4313 double_lock_balance(this_rq, busiest);
4314 /* this_rq->clock is already updated */
4315 update_rq_clock(busiest);
4316 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4317 imbalance, sd, CPU_NEWLY_IDLE,
4319 double_unlock_balance(this_rq, busiest);
4321 if (unlikely(all_pinned)) {
4322 cpumask_clear_cpu(cpu_of(busiest), cpus);
4323 if (!cpumask_empty(cpus))
4329 int active_balance = 0;
4331 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4332 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4333 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4336 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4339 if (sd->nr_balance_failed++ < 2)
4343 * The only task running in a non-idle cpu can be moved to this
4344 * cpu in an attempt to completely freeup the other CPU
4345 * package. The same method used to move task in load_balance()
4346 * have been extended for load_balance_newidle() to speedup
4347 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4349 * The package power saving logic comes from
4350 * find_busiest_group(). If there are no imbalance, then
4351 * f_b_g() will return NULL. However when sched_mc={1,2} then
4352 * f_b_g() will select a group from which a running task may be
4353 * pulled to this cpu in order to make the other package idle.
4354 * If there is no opportunity to make a package idle and if
4355 * there are no imbalance, then f_b_g() will return NULL and no
4356 * action will be taken in load_balance_newidle().
4358 * Under normal task pull operation due to imbalance, there
4359 * will be more than one task in the source run queue and
4360 * move_tasks() will succeed. ld_moved will be true and this
4361 * active balance code will not be triggered.
4364 /* Lock busiest in correct order while this_rq is held */
4365 double_lock_balance(this_rq, busiest);
4368 * don't kick the migration_thread, if the curr
4369 * task on busiest cpu can't be moved to this_cpu
4371 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4372 double_unlock_balance(this_rq, busiest);
4377 if (!busiest->active_balance) {
4378 busiest->active_balance = 1;
4379 busiest->push_cpu = this_cpu;
4383 double_unlock_balance(this_rq, busiest);
4385 * Should not call ttwu while holding a rq->lock
4387 spin_unlock(&this_rq->lock);
4389 wake_up_process(busiest->migration_thread);
4390 spin_lock(&this_rq->lock);
4393 sd->nr_balance_failed = 0;
4395 update_shares_locked(this_rq, sd);
4399 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4400 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4401 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4403 sd->nr_balance_failed = 0;
4409 * idle_balance is called by schedule() if this_cpu is about to become
4410 * idle. Attempts to pull tasks from other CPUs.
4412 static void idle_balance(int this_cpu, struct rq *this_rq)
4414 struct sched_domain *sd;
4415 int pulled_task = 0;
4416 unsigned long next_balance = jiffies + HZ;
4418 for_each_domain(this_cpu, sd) {
4419 unsigned long interval;
4421 if (!(sd->flags & SD_LOAD_BALANCE))
4424 if (sd->flags & SD_BALANCE_NEWIDLE)
4425 /* If we've pulled tasks over stop searching: */
4426 pulled_task = load_balance_newidle(this_cpu, this_rq,
4429 interval = msecs_to_jiffies(sd->balance_interval);
4430 if (time_after(next_balance, sd->last_balance + interval))
4431 next_balance = sd->last_balance + interval;
4435 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4437 * We are going idle. next_balance may be set based on
4438 * a busy processor. So reset next_balance.
4440 this_rq->next_balance = next_balance;
4445 * active_load_balance is run by migration threads. It pushes running tasks
4446 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4447 * running on each physical CPU where possible, and avoids physical /
4448 * logical imbalances.
4450 * Called with busiest_rq locked.
4452 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4454 int target_cpu = busiest_rq->push_cpu;
4455 struct sched_domain *sd;
4456 struct rq *target_rq;
4458 /* Is there any task to move? */
4459 if (busiest_rq->nr_running <= 1)
4462 target_rq = cpu_rq(target_cpu);
4465 * This condition is "impossible", if it occurs
4466 * we need to fix it. Originally reported by
4467 * Bjorn Helgaas on a 128-cpu setup.
4469 BUG_ON(busiest_rq == target_rq);
4471 /* move a task from busiest_rq to target_rq */
4472 double_lock_balance(busiest_rq, target_rq);
4473 update_rq_clock(busiest_rq);
4474 update_rq_clock(target_rq);
4476 /* Search for an sd spanning us and the target CPU. */
4477 for_each_domain(target_cpu, sd) {
4478 if ((sd->flags & SD_LOAD_BALANCE) &&
4479 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4484 schedstat_inc(sd, alb_count);
4486 if (move_one_task(target_rq, target_cpu, busiest_rq,
4488 schedstat_inc(sd, alb_pushed);
4490 schedstat_inc(sd, alb_failed);
4492 double_unlock_balance(busiest_rq, target_rq);
4497 atomic_t load_balancer;
4498 cpumask_var_t cpu_mask;
4499 cpumask_var_t ilb_grp_nohz_mask;
4500 } nohz ____cacheline_aligned = {
4501 .load_balancer = ATOMIC_INIT(-1),
4504 int get_nohz_load_balancer(void)
4506 return atomic_read(&nohz.load_balancer);
4509 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4511 * lowest_flag_domain - Return lowest sched_domain containing flag.
4512 * @cpu: The cpu whose lowest level of sched domain is to
4514 * @flag: The flag to check for the lowest sched_domain
4515 * for the given cpu.
4517 * Returns the lowest sched_domain of a cpu which contains the given flag.
4519 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4521 struct sched_domain *sd;
4523 for_each_domain(cpu, sd)
4524 if (sd && (sd->flags & flag))
4531 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4532 * @cpu: The cpu whose domains we're iterating over.
4533 * @sd: variable holding the value of the power_savings_sd
4535 * @flag: The flag to filter the sched_domains to be iterated.
4537 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4538 * set, starting from the lowest sched_domain to the highest.
4540 #define for_each_flag_domain(cpu, sd, flag) \
4541 for (sd = lowest_flag_domain(cpu, flag); \
4542 (sd && (sd->flags & flag)); sd = sd->parent)
4545 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4546 * @ilb_group: group to be checked for semi-idleness
4548 * Returns: 1 if the group is semi-idle. 0 otherwise.
4550 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4551 * and atleast one non-idle CPU. This helper function checks if the given
4552 * sched_group is semi-idle or not.
4554 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4556 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4557 sched_group_cpus(ilb_group));
4560 * A sched_group is semi-idle when it has atleast one busy cpu
4561 * and atleast one idle cpu.
4563 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4566 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4572 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4573 * @cpu: The cpu which is nominating a new idle_load_balancer.
4575 * Returns: Returns the id of the idle load balancer if it exists,
4576 * Else, returns >= nr_cpu_ids.
4578 * This algorithm picks the idle load balancer such that it belongs to a
4579 * semi-idle powersavings sched_domain. The idea is to try and avoid
4580 * completely idle packages/cores just for the purpose of idle load balancing
4581 * when there are other idle cpu's which are better suited for that job.
4583 static int find_new_ilb(int cpu)
4585 struct sched_domain *sd;
4586 struct sched_group *ilb_group;
4589 * Have idle load balancer selection from semi-idle packages only
4590 * when power-aware load balancing is enabled
4592 if (!(sched_smt_power_savings || sched_mc_power_savings))
4596 * Optimize for the case when we have no idle CPUs or only one
4597 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4599 if (cpumask_weight(nohz.cpu_mask) < 2)
4602 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4603 ilb_group = sd->groups;
4606 if (is_semi_idle_group(ilb_group))
4607 return cpumask_first(nohz.ilb_grp_nohz_mask);
4609 ilb_group = ilb_group->next;
4611 } while (ilb_group != sd->groups);
4615 return cpumask_first(nohz.cpu_mask);
4617 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4618 static inline int find_new_ilb(int call_cpu)
4620 return cpumask_first(nohz.cpu_mask);
4625 * This routine will try to nominate the ilb (idle load balancing)
4626 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4627 * load balancing on behalf of all those cpus. If all the cpus in the system
4628 * go into this tickless mode, then there will be no ilb owner (as there is
4629 * no need for one) and all the cpus will sleep till the next wakeup event
4632 * For the ilb owner, tick is not stopped. And this tick will be used
4633 * for idle load balancing. ilb owner will still be part of
4636 * While stopping the tick, this cpu will become the ilb owner if there
4637 * is no other owner. And will be the owner till that cpu becomes busy
4638 * or if all cpus in the system stop their ticks at which point
4639 * there is no need for ilb owner.
4641 * When the ilb owner becomes busy, it nominates another owner, during the
4642 * next busy scheduler_tick()
4644 int select_nohz_load_balancer(int stop_tick)
4646 int cpu = smp_processor_id();
4649 cpu_rq(cpu)->in_nohz_recently = 1;
4651 if (!cpu_active(cpu)) {
4652 if (atomic_read(&nohz.load_balancer) != cpu)
4656 * If we are going offline and still the leader,
4659 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4665 cpumask_set_cpu(cpu, nohz.cpu_mask);
4667 /* time for ilb owner also to sleep */
4668 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4669 if (atomic_read(&nohz.load_balancer) == cpu)
4670 atomic_set(&nohz.load_balancer, -1);
4674 if (atomic_read(&nohz.load_balancer) == -1) {
4675 /* make me the ilb owner */
4676 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4678 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4681 if (!(sched_smt_power_savings ||
4682 sched_mc_power_savings))
4685 * Check to see if there is a more power-efficient
4688 new_ilb = find_new_ilb(cpu);
4689 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4690 atomic_set(&nohz.load_balancer, -1);
4691 resched_cpu(new_ilb);
4697 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4700 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4702 if (atomic_read(&nohz.load_balancer) == cpu)
4703 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4710 static DEFINE_SPINLOCK(balancing);
4713 * It checks each scheduling domain to see if it is due to be balanced,
4714 * and initiates a balancing operation if so.
4716 * Balancing parameters are set up in arch_init_sched_domains.
4718 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4721 struct rq *rq = cpu_rq(cpu);
4722 unsigned long interval;
4723 struct sched_domain *sd;
4724 /* Earliest time when we have to do rebalance again */
4725 unsigned long next_balance = jiffies + 60*HZ;
4726 int update_next_balance = 0;
4729 for_each_domain(cpu, sd) {
4730 if (!(sd->flags & SD_LOAD_BALANCE))
4733 interval = sd->balance_interval;
4734 if (idle != CPU_IDLE)
4735 interval *= sd->busy_factor;
4737 /* scale ms to jiffies */
4738 interval = msecs_to_jiffies(interval);
4739 if (unlikely(!interval))
4741 if (interval > HZ*NR_CPUS/10)
4742 interval = HZ*NR_CPUS/10;
4744 need_serialize = sd->flags & SD_SERIALIZE;
4746 if (need_serialize) {
4747 if (!spin_trylock(&balancing))
4751 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4752 if (load_balance(cpu, rq, sd, idle, &balance)) {
4754 * We've pulled tasks over so either we're no
4755 * longer idle, or one of our SMT siblings is
4758 idle = CPU_NOT_IDLE;
4760 sd->last_balance = jiffies;
4763 spin_unlock(&balancing);
4765 if (time_after(next_balance, sd->last_balance + interval)) {
4766 next_balance = sd->last_balance + interval;
4767 update_next_balance = 1;
4771 * Stop the load balance at this level. There is another
4772 * CPU in our sched group which is doing load balancing more
4780 * next_balance will be updated only when there is a need.
4781 * When the cpu is attached to null domain for ex, it will not be
4784 if (likely(update_next_balance))
4785 rq->next_balance = next_balance;
4789 * run_rebalance_domains is triggered when needed from the scheduler tick.
4790 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4791 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4793 static void run_rebalance_domains(struct softirq_action *h)
4795 int this_cpu = smp_processor_id();
4796 struct rq *this_rq = cpu_rq(this_cpu);
4797 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4798 CPU_IDLE : CPU_NOT_IDLE;
4800 rebalance_domains(this_cpu, idle);
4804 * If this cpu is the owner for idle load balancing, then do the
4805 * balancing on behalf of the other idle cpus whose ticks are
4808 if (this_rq->idle_at_tick &&
4809 atomic_read(&nohz.load_balancer) == this_cpu) {
4813 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4814 if (balance_cpu == this_cpu)
4818 * If this cpu gets work to do, stop the load balancing
4819 * work being done for other cpus. Next load
4820 * balancing owner will pick it up.
4825 rebalance_domains(balance_cpu, CPU_IDLE);
4827 rq = cpu_rq(balance_cpu);
4828 if (time_after(this_rq->next_balance, rq->next_balance))
4829 this_rq->next_balance = rq->next_balance;
4835 static inline int on_null_domain(int cpu)
4837 return !rcu_dereference(cpu_rq(cpu)->sd);
4841 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4843 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4844 * idle load balancing owner or decide to stop the periodic load balancing,
4845 * if the whole system is idle.
4847 static inline void trigger_load_balance(struct rq *rq, int cpu)
4851 * If we were in the nohz mode recently and busy at the current
4852 * scheduler tick, then check if we need to nominate new idle
4855 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4856 rq->in_nohz_recently = 0;
4858 if (atomic_read(&nohz.load_balancer) == cpu) {
4859 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4860 atomic_set(&nohz.load_balancer, -1);
4863 if (atomic_read(&nohz.load_balancer) == -1) {
4864 int ilb = find_new_ilb(cpu);
4866 if (ilb < nr_cpu_ids)
4872 * If this cpu is idle and doing idle load balancing for all the
4873 * cpus with ticks stopped, is it time for that to stop?
4875 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4876 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4882 * If this cpu is idle and the idle load balancing is done by
4883 * someone else, then no need raise the SCHED_SOFTIRQ
4885 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4886 cpumask_test_cpu(cpu, nohz.cpu_mask))
4889 /* Don't need to rebalance while attached to NULL domain */
4890 if (time_after_eq(jiffies, rq->next_balance) &&
4891 likely(!on_null_domain(cpu)))
4892 raise_softirq(SCHED_SOFTIRQ);
4895 #else /* CONFIG_SMP */
4898 * on UP we do not need to balance between CPUs:
4900 static inline void idle_balance(int cpu, struct rq *rq)
4906 DEFINE_PER_CPU(struct kernel_stat, kstat);
4908 EXPORT_PER_CPU_SYMBOL(kstat);
4911 * Return any ns on the sched_clock that have not yet been accounted in
4912 * @p in case that task is currently running.
4914 * Called with task_rq_lock() held on @rq.
4916 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4920 if (task_current(rq, p)) {
4921 update_rq_clock(rq);
4922 ns = rq->clock - p->se.exec_start;
4930 unsigned long long task_delta_exec(struct task_struct *p)
4932 unsigned long flags;
4936 rq = task_rq_lock(p, &flags);
4937 ns = do_task_delta_exec(p, rq);
4938 task_rq_unlock(rq, &flags);
4944 * Return accounted runtime for the task.
4945 * In case the task is currently running, return the runtime plus current's
4946 * pending runtime that have not been accounted yet.
4948 unsigned long long task_sched_runtime(struct task_struct *p)
4950 unsigned long flags;
4954 rq = task_rq_lock(p, &flags);
4955 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4956 task_rq_unlock(rq, &flags);
4962 * Return sum_exec_runtime for the thread group.
4963 * In case the task is currently running, return the sum plus current's
4964 * pending runtime that have not been accounted yet.
4966 * Note that the thread group might have other running tasks as well,
4967 * so the return value not includes other pending runtime that other
4968 * running tasks might have.
4970 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4972 struct task_cputime totals;
4973 unsigned long flags;
4977 rq = task_rq_lock(p, &flags);
4978 thread_group_cputime(p, &totals);
4979 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4980 task_rq_unlock(rq, &flags);
4986 * Account user cpu time to a process.
4987 * @p: the process that the cpu time gets accounted to
4988 * @cputime: the cpu time spent in user space since the last update
4989 * @cputime_scaled: cputime scaled by cpu frequency
4991 void account_user_time(struct task_struct *p, cputime_t cputime,
4992 cputime_t cputime_scaled)
4994 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4997 /* Add user time to process. */
4998 p->utime = cputime_add(p->utime, cputime);
4999 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5000 account_group_user_time(p, cputime);
5002 /* Add user time to cpustat. */
5003 tmp = cputime_to_cputime64(cputime);
5004 if (TASK_NICE(p) > 0)
5005 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5007 cpustat->user = cputime64_add(cpustat->user, tmp);
5009 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5010 /* Account for user time used */
5011 acct_update_integrals(p);
5015 * Account guest cpu time to a process.
5016 * @p: the process that the cpu time gets accounted to
5017 * @cputime: the cpu time spent in virtual machine since the last update
5018 * @cputime_scaled: cputime scaled by cpu frequency
5020 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5021 cputime_t cputime_scaled)
5024 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5026 tmp = cputime_to_cputime64(cputime);
5028 /* Add guest time to process. */
5029 p->utime = cputime_add(p->utime, cputime);
5030 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5031 account_group_user_time(p, cputime);
5032 p->gtime = cputime_add(p->gtime, cputime);
5034 /* Add guest time to cpustat. */
5035 cpustat->user = cputime64_add(cpustat->user, tmp);
5036 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5040 * Account system cpu time to a process.
5041 * @p: the process that the cpu time gets accounted to
5042 * @hardirq_offset: the offset to subtract from hardirq_count()
5043 * @cputime: the cpu time spent in kernel space since the last update
5044 * @cputime_scaled: cputime scaled by cpu frequency
5046 void account_system_time(struct task_struct *p, int hardirq_offset,
5047 cputime_t cputime, cputime_t cputime_scaled)
5049 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5052 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5053 account_guest_time(p, cputime, cputime_scaled);
5057 /* Add system time to process. */
5058 p->stime = cputime_add(p->stime, cputime);
5059 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5060 account_group_system_time(p, cputime);
5062 /* Add system time to cpustat. */
5063 tmp = cputime_to_cputime64(cputime);
5064 if (hardirq_count() - hardirq_offset)
5065 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5066 else if (softirq_count())
5067 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5069 cpustat->system = cputime64_add(cpustat->system, tmp);
5071 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5073 /* Account for system time used */
5074 acct_update_integrals(p);
5078 * Account for involuntary wait time.
5079 * @steal: the cpu time spent in involuntary wait
5081 void account_steal_time(cputime_t cputime)
5083 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5084 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5086 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5090 * Account for idle time.
5091 * @cputime: the cpu time spent in idle wait
5093 void account_idle_time(cputime_t cputime)
5095 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5096 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5097 struct rq *rq = this_rq();
5099 if (atomic_read(&rq->nr_iowait) > 0)
5100 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5102 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5105 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5108 * Account a single tick of cpu time.
5109 * @p: the process that the cpu time gets accounted to
5110 * @user_tick: indicates if the tick is a user or a system tick
5112 void account_process_tick(struct task_struct *p, int user_tick)
5114 cputime_t one_jiffy = jiffies_to_cputime(1);
5115 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5116 struct rq *rq = this_rq();
5119 account_user_time(p, one_jiffy, one_jiffy_scaled);
5120 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5121 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5124 account_idle_time(one_jiffy);
5128 * Account multiple ticks of steal time.
5129 * @p: the process from which the cpu time has been stolen
5130 * @ticks: number of stolen ticks
5132 void account_steal_ticks(unsigned long ticks)
5134 account_steal_time(jiffies_to_cputime(ticks));
5138 * Account multiple ticks of idle time.
5139 * @ticks: number of stolen ticks
5141 void account_idle_ticks(unsigned long ticks)
5143 account_idle_time(jiffies_to_cputime(ticks));
5149 * Use precise platform statistics if available:
5151 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5152 cputime_t task_utime(struct task_struct *p)
5157 cputime_t task_stime(struct task_struct *p)
5162 cputime_t task_utime(struct task_struct *p)
5164 clock_t utime = cputime_to_clock_t(p->utime),
5165 total = utime + cputime_to_clock_t(p->stime);
5169 * Use CFS's precise accounting:
5171 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5175 do_div(temp, total);
5177 utime = (clock_t)temp;
5179 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5180 return p->prev_utime;
5183 cputime_t task_stime(struct task_struct *p)
5188 * Use CFS's precise accounting. (we subtract utime from
5189 * the total, to make sure the total observed by userspace
5190 * grows monotonically - apps rely on that):
5192 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5193 cputime_to_clock_t(task_utime(p));
5196 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5198 return p->prev_stime;
5202 inline cputime_t task_gtime(struct task_struct *p)
5208 * This function gets called by the timer code, with HZ frequency.
5209 * We call it with interrupts disabled.
5211 * It also gets called by the fork code, when changing the parent's
5214 void scheduler_tick(void)
5216 int cpu = smp_processor_id();
5217 struct rq *rq = cpu_rq(cpu);
5218 struct task_struct *curr = rq->curr;
5222 spin_lock(&rq->lock);
5223 update_rq_clock(rq);
5224 update_cpu_load(rq);
5225 curr->sched_class->task_tick(rq, curr, 0);
5226 spin_unlock(&rq->lock);
5228 perf_counter_task_tick(curr, cpu);
5231 rq->idle_at_tick = idle_cpu(cpu);
5232 trigger_load_balance(rq, cpu);
5236 notrace unsigned long get_parent_ip(unsigned long addr)
5238 if (in_lock_functions(addr)) {
5239 addr = CALLER_ADDR2;
5240 if (in_lock_functions(addr))
5241 addr = CALLER_ADDR3;
5246 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5247 defined(CONFIG_PREEMPT_TRACER))
5249 void __kprobes add_preempt_count(int val)
5251 #ifdef CONFIG_DEBUG_PREEMPT
5255 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5258 preempt_count() += val;
5259 #ifdef CONFIG_DEBUG_PREEMPT
5261 * Spinlock count overflowing soon?
5263 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5266 if (preempt_count() == val)
5267 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5269 EXPORT_SYMBOL(add_preempt_count);
5271 void __kprobes sub_preempt_count(int val)
5273 #ifdef CONFIG_DEBUG_PREEMPT
5277 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5280 * Is the spinlock portion underflowing?
5282 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5283 !(preempt_count() & PREEMPT_MASK)))
5287 if (preempt_count() == val)
5288 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5289 preempt_count() -= val;
5291 EXPORT_SYMBOL(sub_preempt_count);
5296 * Print scheduling while atomic bug:
5298 static noinline void __schedule_bug(struct task_struct *prev)
5300 struct pt_regs *regs = get_irq_regs();
5302 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5303 prev->comm, prev->pid, preempt_count());
5305 debug_show_held_locks(prev);
5307 if (irqs_disabled())
5308 print_irqtrace_events(prev);
5317 * Various schedule()-time debugging checks and statistics:
5319 static inline void schedule_debug(struct task_struct *prev)
5322 * Test if we are atomic. Since do_exit() needs to call into
5323 * schedule() atomically, we ignore that path for now.
5324 * Otherwise, whine if we are scheduling when we should not be.
5326 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5327 __schedule_bug(prev);
5329 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5331 schedstat_inc(this_rq(), sched_count);
5332 #ifdef CONFIG_SCHEDSTATS
5333 if (unlikely(prev->lock_depth >= 0)) {
5334 schedstat_inc(this_rq(), bkl_count);
5335 schedstat_inc(prev, sched_info.bkl_count);
5340 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5342 if (prev->state == TASK_RUNNING) {
5343 u64 runtime = prev->se.sum_exec_runtime;
5345 runtime -= prev->se.prev_sum_exec_runtime;
5346 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5349 * In order to avoid avg_overlap growing stale when we are
5350 * indeed overlapping and hence not getting put to sleep, grow
5351 * the avg_overlap on preemption.
5353 * We use the average preemption runtime because that
5354 * correlates to the amount of cache footprint a task can
5357 update_avg(&prev->se.avg_overlap, runtime);
5359 prev->sched_class->put_prev_task(rq, prev);
5363 * Pick up the highest-prio task:
5365 static inline struct task_struct *
5366 pick_next_task(struct rq *rq)
5368 const struct sched_class *class;
5369 struct task_struct *p;
5372 * Optimization: we know that if all tasks are in
5373 * the fair class we can call that function directly:
5375 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5376 p = fair_sched_class.pick_next_task(rq);
5381 class = sched_class_highest;
5383 p = class->pick_next_task(rq);
5387 * Will never be NULL as the idle class always
5388 * returns a non-NULL p:
5390 class = class->next;
5395 * schedule() is the main scheduler function.
5397 asmlinkage void __sched schedule(void)
5399 struct task_struct *prev, *next;
5400 unsigned long *switch_count;
5406 cpu = smp_processor_id();
5410 switch_count = &prev->nivcsw;
5412 release_kernel_lock(prev);
5413 need_resched_nonpreemptible:
5415 schedule_debug(prev);
5417 if (sched_feat(HRTICK))
5420 spin_lock_irq(&rq->lock);
5421 update_rq_clock(rq);
5422 clear_tsk_need_resched(prev);
5424 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5425 if (unlikely(signal_pending_state(prev->state, prev)))
5426 prev->state = TASK_RUNNING;
5428 deactivate_task(rq, prev, 1);
5429 switch_count = &prev->nvcsw;
5432 pre_schedule(rq, prev);
5434 if (unlikely(!rq->nr_running))
5435 idle_balance(cpu, rq);
5437 put_prev_task(rq, prev);
5438 next = pick_next_task(rq);
5440 if (likely(prev != next)) {
5441 sched_info_switch(prev, next);
5442 perf_counter_task_sched_out(prev, next, cpu);
5448 context_switch(rq, prev, next); /* unlocks the rq */
5450 * the context switch might have flipped the stack from under
5451 * us, hence refresh the local variables.
5453 cpu = smp_processor_id();
5456 spin_unlock_irq(&rq->lock);
5460 if (unlikely(reacquire_kernel_lock(current) < 0))
5461 goto need_resched_nonpreemptible;
5463 preempt_enable_no_resched();
5467 EXPORT_SYMBOL(schedule);
5471 * Look out! "owner" is an entirely speculative pointer
5472 * access and not reliable.
5474 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5479 if (!sched_feat(OWNER_SPIN))
5482 #ifdef CONFIG_DEBUG_PAGEALLOC
5484 * Need to access the cpu field knowing that
5485 * DEBUG_PAGEALLOC could have unmapped it if
5486 * the mutex owner just released it and exited.
5488 if (probe_kernel_address(&owner->cpu, cpu))
5495 * Even if the access succeeded (likely case),
5496 * the cpu field may no longer be valid.
5498 if (cpu >= nr_cpumask_bits)
5502 * We need to validate that we can do a
5503 * get_cpu() and that we have the percpu area.
5505 if (!cpu_online(cpu))
5512 * Owner changed, break to re-assess state.
5514 if (lock->owner != owner)
5518 * Is that owner really running on that cpu?
5520 if (task_thread_info(rq->curr) != owner || need_resched())
5530 #ifdef CONFIG_PREEMPT
5532 * this is the entry point to schedule() from in-kernel preemption
5533 * off of preempt_enable. Kernel preemptions off return from interrupt
5534 * occur there and call schedule directly.
5536 asmlinkage void __sched preempt_schedule(void)
5538 struct thread_info *ti = current_thread_info();
5541 * If there is a non-zero preempt_count or interrupts are disabled,
5542 * we do not want to preempt the current task. Just return..
5544 if (likely(ti->preempt_count || irqs_disabled()))
5548 add_preempt_count(PREEMPT_ACTIVE);
5550 sub_preempt_count(PREEMPT_ACTIVE);
5553 * Check again in case we missed a preemption opportunity
5554 * between schedule and now.
5557 } while (need_resched());
5559 EXPORT_SYMBOL(preempt_schedule);
5562 * this is the entry point to schedule() from kernel preemption
5563 * off of irq context.
5564 * Note, that this is called and return with irqs disabled. This will
5565 * protect us against recursive calling from irq.
5567 asmlinkage void __sched preempt_schedule_irq(void)
5569 struct thread_info *ti = current_thread_info();
5571 /* Catch callers which need to be fixed */
5572 BUG_ON(ti->preempt_count || !irqs_disabled());
5575 add_preempt_count(PREEMPT_ACTIVE);
5578 local_irq_disable();
5579 sub_preempt_count(PREEMPT_ACTIVE);
5582 * Check again in case we missed a preemption opportunity
5583 * between schedule and now.
5586 } while (need_resched());
5589 #endif /* CONFIG_PREEMPT */
5591 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5594 return try_to_wake_up(curr->private, mode, sync);
5596 EXPORT_SYMBOL(default_wake_function);
5599 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5600 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5601 * number) then we wake all the non-exclusive tasks and one exclusive task.
5603 * There are circumstances in which we can try to wake a task which has already
5604 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5605 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5607 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5608 int nr_exclusive, int sync, void *key)
5610 wait_queue_t *curr, *next;
5612 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5613 unsigned flags = curr->flags;
5615 if (curr->func(curr, mode, sync, key) &&
5616 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5622 * __wake_up - wake up threads blocked on a waitqueue.
5624 * @mode: which threads
5625 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5626 * @key: is directly passed to the wakeup function
5628 * It may be assumed that this function implies a write memory barrier before
5629 * changing the task state if and only if any tasks are woken up.
5631 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5632 int nr_exclusive, void *key)
5634 unsigned long flags;
5636 spin_lock_irqsave(&q->lock, flags);
5637 __wake_up_common(q, mode, nr_exclusive, 0, key);
5638 spin_unlock_irqrestore(&q->lock, flags);
5640 EXPORT_SYMBOL(__wake_up);
5643 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5645 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5647 __wake_up_common(q, mode, 1, 0, NULL);
5650 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5652 __wake_up_common(q, mode, 1, 0, key);
5656 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5658 * @mode: which threads
5659 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5660 * @key: opaque value to be passed to wakeup targets
5662 * The sync wakeup differs that the waker knows that it will schedule
5663 * away soon, so while the target thread will be woken up, it will not
5664 * be migrated to another CPU - ie. the two threads are 'synchronized'
5665 * with each other. This can prevent needless bouncing between CPUs.
5667 * On UP it can prevent extra preemption.
5669 * It may be assumed that this function implies a write memory barrier before
5670 * changing the task state if and only if any tasks are woken up.
5672 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5673 int nr_exclusive, void *key)
5675 unsigned long flags;
5681 if (unlikely(!nr_exclusive))
5684 spin_lock_irqsave(&q->lock, flags);
5685 __wake_up_common(q, mode, nr_exclusive, sync, key);
5686 spin_unlock_irqrestore(&q->lock, flags);
5688 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5691 * __wake_up_sync - see __wake_up_sync_key()
5693 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5695 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5697 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5700 * complete: - signals a single thread waiting on this completion
5701 * @x: holds the state of this particular completion
5703 * This will wake up a single thread waiting on this completion. Threads will be
5704 * awakened in the same order in which they were queued.
5706 * See also complete_all(), wait_for_completion() and related routines.
5708 * It may be assumed that this function implies a write memory barrier before
5709 * changing the task state if and only if any tasks are woken up.
5711 void complete(struct completion *x)
5713 unsigned long flags;
5715 spin_lock_irqsave(&x->wait.lock, flags);
5717 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5718 spin_unlock_irqrestore(&x->wait.lock, flags);
5720 EXPORT_SYMBOL(complete);
5723 * complete_all: - signals all threads waiting on this completion
5724 * @x: holds the state of this particular completion
5726 * This will wake up all threads waiting on this particular completion event.
5728 * It may be assumed that this function implies a write memory barrier before
5729 * changing the task state if and only if any tasks are woken up.
5731 void complete_all(struct completion *x)
5733 unsigned long flags;
5735 spin_lock_irqsave(&x->wait.lock, flags);
5736 x->done += UINT_MAX/2;
5737 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5738 spin_unlock_irqrestore(&x->wait.lock, flags);
5740 EXPORT_SYMBOL(complete_all);
5742 static inline long __sched
5743 do_wait_for_common(struct completion *x, long timeout, int state)
5746 DECLARE_WAITQUEUE(wait, current);
5748 wait.flags |= WQ_FLAG_EXCLUSIVE;
5749 __add_wait_queue_tail(&x->wait, &wait);
5751 if (signal_pending_state(state, current)) {
5752 timeout = -ERESTARTSYS;
5755 __set_current_state(state);
5756 spin_unlock_irq(&x->wait.lock);
5757 timeout = schedule_timeout(timeout);
5758 spin_lock_irq(&x->wait.lock);
5759 } while (!x->done && timeout);
5760 __remove_wait_queue(&x->wait, &wait);
5765 return timeout ?: 1;
5769 wait_for_common(struct completion *x, long timeout, int state)
5773 spin_lock_irq(&x->wait.lock);
5774 timeout = do_wait_for_common(x, timeout, state);
5775 spin_unlock_irq(&x->wait.lock);
5780 * wait_for_completion: - waits for completion of a task
5781 * @x: holds the state of this particular completion
5783 * This waits to be signaled for completion of a specific task. It is NOT
5784 * interruptible and there is no timeout.
5786 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5787 * and interrupt capability. Also see complete().
5789 void __sched wait_for_completion(struct completion *x)
5791 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5793 EXPORT_SYMBOL(wait_for_completion);
5796 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5797 * @x: holds the state of this particular completion
5798 * @timeout: timeout value in jiffies
5800 * This waits for either a completion of a specific task to be signaled or for a
5801 * specified timeout to expire. The timeout is in jiffies. It is not
5804 unsigned long __sched
5805 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5807 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5809 EXPORT_SYMBOL(wait_for_completion_timeout);
5812 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5813 * @x: holds the state of this particular completion
5815 * This waits for completion of a specific task to be signaled. It is
5818 int __sched wait_for_completion_interruptible(struct completion *x)
5820 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5821 if (t == -ERESTARTSYS)
5825 EXPORT_SYMBOL(wait_for_completion_interruptible);
5828 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5829 * @x: holds the state of this particular completion
5830 * @timeout: timeout value in jiffies
5832 * This waits for either a completion of a specific task to be signaled or for a
5833 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5835 unsigned long __sched
5836 wait_for_completion_interruptible_timeout(struct completion *x,
5837 unsigned long timeout)
5839 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5841 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5844 * wait_for_completion_killable: - waits for completion of a task (killable)
5845 * @x: holds the state of this particular completion
5847 * This waits to be signaled for completion of a specific task. It can be
5848 * interrupted by a kill signal.
5850 int __sched wait_for_completion_killable(struct completion *x)
5852 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5853 if (t == -ERESTARTSYS)
5857 EXPORT_SYMBOL(wait_for_completion_killable);
5860 * try_wait_for_completion - try to decrement a completion without blocking
5861 * @x: completion structure
5863 * Returns: 0 if a decrement cannot be done without blocking
5864 * 1 if a decrement succeeded.
5866 * If a completion is being used as a counting completion,
5867 * attempt to decrement the counter without blocking. This
5868 * enables us to avoid waiting if the resource the completion
5869 * is protecting is not available.
5871 bool try_wait_for_completion(struct completion *x)
5875 spin_lock_irq(&x->wait.lock);
5880 spin_unlock_irq(&x->wait.lock);
5883 EXPORT_SYMBOL(try_wait_for_completion);
5886 * completion_done - Test to see if a completion has any waiters
5887 * @x: completion structure
5889 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5890 * 1 if there are no waiters.
5893 bool completion_done(struct completion *x)
5897 spin_lock_irq(&x->wait.lock);
5900 spin_unlock_irq(&x->wait.lock);
5903 EXPORT_SYMBOL(completion_done);
5906 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5908 unsigned long flags;
5911 init_waitqueue_entry(&wait, current);
5913 __set_current_state(state);
5915 spin_lock_irqsave(&q->lock, flags);
5916 __add_wait_queue(q, &wait);
5917 spin_unlock(&q->lock);
5918 timeout = schedule_timeout(timeout);
5919 spin_lock_irq(&q->lock);
5920 __remove_wait_queue(q, &wait);
5921 spin_unlock_irqrestore(&q->lock, flags);
5926 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5928 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5930 EXPORT_SYMBOL(interruptible_sleep_on);
5933 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5935 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5937 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5939 void __sched sleep_on(wait_queue_head_t *q)
5941 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5943 EXPORT_SYMBOL(sleep_on);
5945 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5947 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5949 EXPORT_SYMBOL(sleep_on_timeout);
5951 #ifdef CONFIG_RT_MUTEXES
5954 * rt_mutex_setprio - set the current priority of a task
5956 * @prio: prio value (kernel-internal form)
5958 * This function changes the 'effective' priority of a task. It does
5959 * not touch ->normal_prio like __setscheduler().
5961 * Used by the rt_mutex code to implement priority inheritance logic.
5963 void rt_mutex_setprio(struct task_struct *p, int prio)
5965 unsigned long flags;
5966 int oldprio, on_rq, running;
5968 const struct sched_class *prev_class = p->sched_class;
5970 BUG_ON(prio < 0 || prio > MAX_PRIO);
5972 rq = task_rq_lock(p, &flags);
5973 update_rq_clock(rq);
5976 on_rq = p->se.on_rq;
5977 running = task_current(rq, p);
5979 dequeue_task(rq, p, 0);
5981 p->sched_class->put_prev_task(rq, p);
5984 p->sched_class = &rt_sched_class;
5986 p->sched_class = &fair_sched_class;
5991 p->sched_class->set_curr_task(rq);
5993 enqueue_task(rq, p, 0);
5995 check_class_changed(rq, p, prev_class, oldprio, running);
5997 task_rq_unlock(rq, &flags);
6002 void set_user_nice(struct task_struct *p, long nice)
6004 int old_prio, delta, on_rq;
6005 unsigned long flags;
6008 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6011 * We have to be careful, if called from sys_setpriority(),
6012 * the task might be in the middle of scheduling on another CPU.
6014 rq = task_rq_lock(p, &flags);
6015 update_rq_clock(rq);
6017 * The RT priorities are set via sched_setscheduler(), but we still
6018 * allow the 'normal' nice value to be set - but as expected
6019 * it wont have any effect on scheduling until the task is
6020 * SCHED_FIFO/SCHED_RR:
6022 if (task_has_rt_policy(p)) {
6023 p->static_prio = NICE_TO_PRIO(nice);
6026 on_rq = p->se.on_rq;
6028 dequeue_task(rq, p, 0);
6030 p->static_prio = NICE_TO_PRIO(nice);
6033 p->prio = effective_prio(p);
6034 delta = p->prio - old_prio;
6037 enqueue_task(rq, p, 0);
6039 * If the task increased its priority or is running and
6040 * lowered its priority, then reschedule its CPU:
6042 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6043 resched_task(rq->curr);
6046 task_rq_unlock(rq, &flags);
6048 EXPORT_SYMBOL(set_user_nice);
6051 * can_nice - check if a task can reduce its nice value
6055 int can_nice(const struct task_struct *p, const int nice)
6057 /* convert nice value [19,-20] to rlimit style value [1,40] */
6058 int nice_rlim = 20 - nice;
6060 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6061 capable(CAP_SYS_NICE));
6064 #ifdef __ARCH_WANT_SYS_NICE
6067 * sys_nice - change the priority of the current process.
6068 * @increment: priority increment
6070 * sys_setpriority is a more generic, but much slower function that
6071 * does similar things.
6073 SYSCALL_DEFINE1(nice, int, increment)
6078 * Setpriority might change our priority at the same moment.
6079 * We don't have to worry. Conceptually one call occurs first
6080 * and we have a single winner.
6082 if (increment < -40)
6087 nice = TASK_NICE(current) + increment;
6093 if (increment < 0 && !can_nice(current, nice))
6096 retval = security_task_setnice(current, nice);
6100 set_user_nice(current, nice);
6107 * task_prio - return the priority value of a given task.
6108 * @p: the task in question.
6110 * This is the priority value as seen by users in /proc.
6111 * RT tasks are offset by -200. Normal tasks are centered
6112 * around 0, value goes from -16 to +15.
6114 int task_prio(const struct task_struct *p)
6116 return p->prio - MAX_RT_PRIO;
6120 * task_nice - return the nice value of a given task.
6121 * @p: the task in question.
6123 int task_nice(const struct task_struct *p)
6125 return TASK_NICE(p);
6127 EXPORT_SYMBOL(task_nice);
6130 * idle_cpu - is a given cpu idle currently?
6131 * @cpu: the processor in question.
6133 int idle_cpu(int cpu)
6135 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6139 * idle_task - return the idle task for a given cpu.
6140 * @cpu: the processor in question.
6142 struct task_struct *idle_task(int cpu)
6144 return cpu_rq(cpu)->idle;
6148 * find_process_by_pid - find a process with a matching PID value.
6149 * @pid: the pid in question.
6151 static struct task_struct *find_process_by_pid(pid_t pid)
6153 return pid ? find_task_by_vpid(pid) : current;
6156 /* Actually do priority change: must hold rq lock. */
6158 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6160 BUG_ON(p->se.on_rq);
6163 switch (p->policy) {
6167 p->sched_class = &fair_sched_class;
6171 p->sched_class = &rt_sched_class;
6175 p->rt_priority = prio;
6176 p->normal_prio = normal_prio(p);
6177 /* we are holding p->pi_lock already */
6178 p->prio = rt_mutex_getprio(p);
6183 * check the target process has a UID that matches the current process's
6185 static bool check_same_owner(struct task_struct *p)
6187 const struct cred *cred = current_cred(), *pcred;
6191 pcred = __task_cred(p);
6192 match = (cred->euid == pcred->euid ||
6193 cred->euid == pcred->uid);
6198 static int __sched_setscheduler(struct task_struct *p, int policy,
6199 struct sched_param *param, bool user)
6201 int retval, oldprio, oldpolicy = -1, on_rq, running;
6202 unsigned long flags;
6203 const struct sched_class *prev_class = p->sched_class;
6207 /* may grab non-irq protected spin_locks */
6208 BUG_ON(in_interrupt());
6210 /* double check policy once rq lock held */
6212 reset_on_fork = p->sched_reset_on_fork;
6213 policy = oldpolicy = p->policy;
6215 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6216 policy &= ~SCHED_RESET_ON_FORK;
6218 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6219 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6220 policy != SCHED_IDLE)
6225 * Valid priorities for SCHED_FIFO and SCHED_RR are
6226 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6227 * SCHED_BATCH and SCHED_IDLE is 0.
6229 if (param->sched_priority < 0 ||
6230 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6231 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6233 if (rt_policy(policy) != (param->sched_priority != 0))
6237 * Allow unprivileged RT tasks to decrease priority:
6239 if (user && !capable(CAP_SYS_NICE)) {
6240 if (rt_policy(policy)) {
6241 unsigned long rlim_rtprio;
6243 if (!lock_task_sighand(p, &flags))
6245 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6246 unlock_task_sighand(p, &flags);
6248 /* can't set/change the rt policy */
6249 if (policy != p->policy && !rlim_rtprio)
6252 /* can't increase priority */
6253 if (param->sched_priority > p->rt_priority &&
6254 param->sched_priority > rlim_rtprio)
6258 * Like positive nice levels, dont allow tasks to
6259 * move out of SCHED_IDLE either:
6261 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6264 /* can't change other user's priorities */
6265 if (!check_same_owner(p))
6268 /* Normal users shall not reset the sched_reset_on_fork flag */
6269 if (p->sched_reset_on_fork && !reset_on_fork)
6274 #ifdef CONFIG_RT_GROUP_SCHED
6276 * Do not allow realtime tasks into groups that have no runtime
6279 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6280 task_group(p)->rt_bandwidth.rt_runtime == 0)
6284 retval = security_task_setscheduler(p, policy, param);
6290 * make sure no PI-waiters arrive (or leave) while we are
6291 * changing the priority of the task:
6293 spin_lock_irqsave(&p->pi_lock, flags);
6295 * To be able to change p->policy safely, the apropriate
6296 * runqueue lock must be held.
6298 rq = __task_rq_lock(p);
6299 /* recheck policy now with rq lock held */
6300 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6301 policy = oldpolicy = -1;
6302 __task_rq_unlock(rq);
6303 spin_unlock_irqrestore(&p->pi_lock, flags);
6306 update_rq_clock(rq);
6307 on_rq = p->se.on_rq;
6308 running = task_current(rq, p);
6310 deactivate_task(rq, p, 0);
6312 p->sched_class->put_prev_task(rq, p);
6314 p->sched_reset_on_fork = reset_on_fork;
6317 __setscheduler(rq, p, policy, param->sched_priority);
6320 p->sched_class->set_curr_task(rq);
6322 activate_task(rq, p, 0);
6324 check_class_changed(rq, p, prev_class, oldprio, running);
6326 __task_rq_unlock(rq);
6327 spin_unlock_irqrestore(&p->pi_lock, flags);
6329 rt_mutex_adjust_pi(p);
6335 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6336 * @p: the task in question.
6337 * @policy: new policy.
6338 * @param: structure containing the new RT priority.
6340 * NOTE that the task may be already dead.
6342 int sched_setscheduler(struct task_struct *p, int policy,
6343 struct sched_param *param)
6345 return __sched_setscheduler(p, policy, param, true);
6347 EXPORT_SYMBOL_GPL(sched_setscheduler);
6350 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6351 * @p: the task in question.
6352 * @policy: new policy.
6353 * @param: structure containing the new RT priority.
6355 * Just like sched_setscheduler, only don't bother checking if the
6356 * current context has permission. For example, this is needed in
6357 * stop_machine(): we create temporary high priority worker threads,
6358 * but our caller might not have that capability.
6360 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6361 struct sched_param *param)
6363 return __sched_setscheduler(p, policy, param, false);
6367 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6369 struct sched_param lparam;
6370 struct task_struct *p;
6373 if (!param || pid < 0)
6375 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6380 p = find_process_by_pid(pid);
6382 retval = sched_setscheduler(p, policy, &lparam);
6389 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6390 * @pid: the pid in question.
6391 * @policy: new policy.
6392 * @param: structure containing the new RT priority.
6394 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6395 struct sched_param __user *, param)
6397 /* negative values for policy are not valid */
6401 return do_sched_setscheduler(pid, policy, param);
6405 * sys_sched_setparam - set/change the RT priority of a thread
6406 * @pid: the pid in question.
6407 * @param: structure containing the new RT priority.
6409 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6411 return do_sched_setscheduler(pid, -1, param);
6415 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6416 * @pid: the pid in question.
6418 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6420 struct task_struct *p;
6427 read_lock(&tasklist_lock);
6428 p = find_process_by_pid(pid);
6430 retval = security_task_getscheduler(p);
6433 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6435 read_unlock(&tasklist_lock);
6440 * sys_sched_getparam - get the RT priority of a thread
6441 * @pid: the pid in question.
6442 * @param: structure containing the RT priority.
6444 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6446 struct sched_param lp;
6447 struct task_struct *p;
6450 if (!param || pid < 0)
6453 read_lock(&tasklist_lock);
6454 p = find_process_by_pid(pid);
6459 retval = security_task_getscheduler(p);
6463 lp.sched_priority = p->rt_priority;
6464 read_unlock(&tasklist_lock);
6467 * This one might sleep, we cannot do it with a spinlock held ...
6469 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6474 read_unlock(&tasklist_lock);
6478 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6480 cpumask_var_t cpus_allowed, new_mask;
6481 struct task_struct *p;
6485 read_lock(&tasklist_lock);
6487 p = find_process_by_pid(pid);
6489 read_unlock(&tasklist_lock);
6495 * It is not safe to call set_cpus_allowed with the
6496 * tasklist_lock held. We will bump the task_struct's
6497 * usage count and then drop tasklist_lock.
6500 read_unlock(&tasklist_lock);
6502 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6506 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6508 goto out_free_cpus_allowed;
6511 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6514 retval = security_task_setscheduler(p, 0, NULL);
6518 cpuset_cpus_allowed(p, cpus_allowed);
6519 cpumask_and(new_mask, in_mask, cpus_allowed);
6521 retval = set_cpus_allowed_ptr(p, new_mask);
6524 cpuset_cpus_allowed(p, cpus_allowed);
6525 if (!cpumask_subset(new_mask, cpus_allowed)) {
6527 * We must have raced with a concurrent cpuset
6528 * update. Just reset the cpus_allowed to the
6529 * cpuset's cpus_allowed
6531 cpumask_copy(new_mask, cpus_allowed);
6536 free_cpumask_var(new_mask);
6537 out_free_cpus_allowed:
6538 free_cpumask_var(cpus_allowed);
6545 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6546 struct cpumask *new_mask)
6548 if (len < cpumask_size())
6549 cpumask_clear(new_mask);
6550 else if (len > cpumask_size())
6551 len = cpumask_size();
6553 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6557 * sys_sched_setaffinity - set the cpu affinity of a process
6558 * @pid: pid of the process
6559 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6560 * @user_mask_ptr: user-space pointer to the new cpu mask
6562 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6563 unsigned long __user *, user_mask_ptr)
6565 cpumask_var_t new_mask;
6568 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6571 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6573 retval = sched_setaffinity(pid, new_mask);
6574 free_cpumask_var(new_mask);
6578 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6580 struct task_struct *p;
6584 read_lock(&tasklist_lock);
6587 p = find_process_by_pid(pid);
6591 retval = security_task_getscheduler(p);
6595 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6598 read_unlock(&tasklist_lock);
6605 * sys_sched_getaffinity - get the cpu affinity of a process
6606 * @pid: pid of the process
6607 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6608 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6610 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6611 unsigned long __user *, user_mask_ptr)
6616 if (len < cpumask_size())
6619 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6622 ret = sched_getaffinity(pid, mask);
6624 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6627 ret = cpumask_size();
6629 free_cpumask_var(mask);
6635 * sys_sched_yield - yield the current processor to other threads.
6637 * This function yields the current CPU to other tasks. If there are no
6638 * other threads running on this CPU then this function will return.
6640 SYSCALL_DEFINE0(sched_yield)
6642 struct rq *rq = this_rq_lock();
6644 schedstat_inc(rq, yld_count);
6645 current->sched_class->yield_task(rq);
6648 * Since we are going to call schedule() anyway, there's
6649 * no need to preempt or enable interrupts:
6651 __release(rq->lock);
6652 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6653 _raw_spin_unlock(&rq->lock);
6654 preempt_enable_no_resched();
6661 static inline int should_resched(void)
6663 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6666 static void __cond_resched(void)
6668 add_preempt_count(PREEMPT_ACTIVE);
6670 sub_preempt_count(PREEMPT_ACTIVE);
6673 int __sched _cond_resched(void)
6675 if (should_resched()) {
6681 EXPORT_SYMBOL(_cond_resched);
6684 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6685 * call schedule, and on return reacquire the lock.
6687 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6688 * operations here to prevent schedule() from being called twice (once via
6689 * spin_unlock(), once by hand).
6691 int __cond_resched_lock(spinlock_t *lock)
6693 int resched = should_resched();
6696 if (spin_needbreak(lock) || resched) {
6707 EXPORT_SYMBOL(__cond_resched_lock);
6709 int __sched __cond_resched_softirq(void)
6711 BUG_ON(!in_softirq());
6713 if (should_resched()) {
6721 EXPORT_SYMBOL(__cond_resched_softirq);
6724 * yield - yield the current processor to other threads.
6726 * This is a shortcut for kernel-space yielding - it marks the
6727 * thread runnable and calls sys_sched_yield().
6729 void __sched yield(void)
6731 set_current_state(TASK_RUNNING);
6734 EXPORT_SYMBOL(yield);
6737 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6738 * that process accounting knows that this is a task in IO wait state.
6740 * But don't do that if it is a deliberate, throttling IO wait (this task
6741 * has set its backing_dev_info: the queue against which it should throttle)
6743 void __sched io_schedule(void)
6745 struct rq *rq = raw_rq();
6747 delayacct_blkio_start();
6748 atomic_inc(&rq->nr_iowait);
6750 atomic_dec(&rq->nr_iowait);
6751 delayacct_blkio_end();
6753 EXPORT_SYMBOL(io_schedule);
6755 long __sched io_schedule_timeout(long timeout)
6757 struct rq *rq = raw_rq();
6760 delayacct_blkio_start();
6761 atomic_inc(&rq->nr_iowait);
6762 ret = schedule_timeout(timeout);
6763 atomic_dec(&rq->nr_iowait);
6764 delayacct_blkio_end();
6769 * sys_sched_get_priority_max - return maximum RT priority.
6770 * @policy: scheduling class.
6772 * this syscall returns the maximum rt_priority that can be used
6773 * by a given scheduling class.
6775 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6782 ret = MAX_USER_RT_PRIO-1;
6794 * sys_sched_get_priority_min - return minimum RT priority.
6795 * @policy: scheduling class.
6797 * this syscall returns the minimum rt_priority that can be used
6798 * by a given scheduling class.
6800 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6818 * sys_sched_rr_get_interval - return the default timeslice of a process.
6819 * @pid: pid of the process.
6820 * @interval: userspace pointer to the timeslice value.
6822 * this syscall writes the default timeslice value of a given process
6823 * into the user-space timespec buffer. A value of '0' means infinity.
6825 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6826 struct timespec __user *, interval)
6828 struct task_struct *p;
6829 unsigned int time_slice;
6837 read_lock(&tasklist_lock);
6838 p = find_process_by_pid(pid);
6842 retval = security_task_getscheduler(p);
6847 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6848 * tasks that are on an otherwise idle runqueue:
6851 if (p->policy == SCHED_RR) {
6852 time_slice = DEF_TIMESLICE;
6853 } else if (p->policy != SCHED_FIFO) {
6854 struct sched_entity *se = &p->se;
6855 unsigned long flags;
6858 rq = task_rq_lock(p, &flags);
6859 if (rq->cfs.load.weight)
6860 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6861 task_rq_unlock(rq, &flags);
6863 read_unlock(&tasklist_lock);
6864 jiffies_to_timespec(time_slice, &t);
6865 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6869 read_unlock(&tasklist_lock);
6873 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6875 void sched_show_task(struct task_struct *p)
6877 unsigned long free = 0;
6880 state = p->state ? __ffs(p->state) + 1 : 0;
6881 printk(KERN_INFO "%-13.13s %c", p->comm,
6882 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6883 #if BITS_PER_LONG == 32
6884 if (state == TASK_RUNNING)
6885 printk(KERN_CONT " running ");
6887 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6889 if (state == TASK_RUNNING)
6890 printk(KERN_CONT " running task ");
6892 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6894 #ifdef CONFIG_DEBUG_STACK_USAGE
6895 free = stack_not_used(p);
6897 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6898 task_pid_nr(p), task_pid_nr(p->real_parent),
6899 (unsigned long)task_thread_info(p)->flags);
6901 show_stack(p, NULL);
6904 void show_state_filter(unsigned long state_filter)
6906 struct task_struct *g, *p;
6908 #if BITS_PER_LONG == 32
6910 " task PC stack pid father\n");
6913 " task PC stack pid father\n");
6915 read_lock(&tasklist_lock);
6916 do_each_thread(g, p) {
6918 * reset the NMI-timeout, listing all files on a slow
6919 * console might take alot of time:
6921 touch_nmi_watchdog();
6922 if (!state_filter || (p->state & state_filter))
6924 } while_each_thread(g, p);
6926 touch_all_softlockup_watchdogs();
6928 #ifdef CONFIG_SCHED_DEBUG
6929 sysrq_sched_debug_show();
6931 read_unlock(&tasklist_lock);
6933 * Only show locks if all tasks are dumped:
6935 if (state_filter == -1)
6936 debug_show_all_locks();
6939 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6941 idle->sched_class = &idle_sched_class;
6945 * init_idle - set up an idle thread for a given CPU
6946 * @idle: task in question
6947 * @cpu: cpu the idle task belongs to
6949 * NOTE: this function does not set the idle thread's NEED_RESCHED
6950 * flag, to make booting more robust.
6952 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6954 struct rq *rq = cpu_rq(cpu);
6955 unsigned long flags;
6957 spin_lock_irqsave(&rq->lock, flags);
6960 idle->se.exec_start = sched_clock();
6962 idle->prio = idle->normal_prio = MAX_PRIO;
6963 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6964 __set_task_cpu(idle, cpu);
6966 rq->curr = rq->idle = idle;
6967 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6970 spin_unlock_irqrestore(&rq->lock, flags);
6972 /* Set the preempt count _outside_ the spinlocks! */
6973 #if defined(CONFIG_PREEMPT)
6974 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6976 task_thread_info(idle)->preempt_count = 0;
6979 * The idle tasks have their own, simple scheduling class:
6981 idle->sched_class = &idle_sched_class;
6982 ftrace_graph_init_task(idle);
6986 * In a system that switches off the HZ timer nohz_cpu_mask
6987 * indicates which cpus entered this state. This is used
6988 * in the rcu update to wait only for active cpus. For system
6989 * which do not switch off the HZ timer nohz_cpu_mask should
6990 * always be CPU_BITS_NONE.
6992 cpumask_var_t nohz_cpu_mask;
6995 * Increase the granularity value when there are more CPUs,
6996 * because with more CPUs the 'effective latency' as visible
6997 * to users decreases. But the relationship is not linear,
6998 * so pick a second-best guess by going with the log2 of the
7001 * This idea comes from the SD scheduler of Con Kolivas:
7003 static inline void sched_init_granularity(void)
7005 unsigned int factor = 1 + ilog2(num_online_cpus());
7006 const unsigned long limit = 200000000;
7008 sysctl_sched_min_granularity *= factor;
7009 if (sysctl_sched_min_granularity > limit)
7010 sysctl_sched_min_granularity = limit;
7012 sysctl_sched_latency *= factor;
7013 if (sysctl_sched_latency > limit)
7014 sysctl_sched_latency = limit;
7016 sysctl_sched_wakeup_granularity *= factor;
7018 sysctl_sched_shares_ratelimit *= factor;
7023 * This is how migration works:
7025 * 1) we queue a struct migration_req structure in the source CPU's
7026 * runqueue and wake up that CPU's migration thread.
7027 * 2) we down() the locked semaphore => thread blocks.
7028 * 3) migration thread wakes up (implicitly it forces the migrated
7029 * thread off the CPU)
7030 * 4) it gets the migration request and checks whether the migrated
7031 * task is still in the wrong runqueue.
7032 * 5) if it's in the wrong runqueue then the migration thread removes
7033 * it and puts it into the right queue.
7034 * 6) migration thread up()s the semaphore.
7035 * 7) we wake up and the migration is done.
7039 * Change a given task's CPU affinity. Migrate the thread to a
7040 * proper CPU and schedule it away if the CPU it's executing on
7041 * is removed from the allowed bitmask.
7043 * NOTE: the caller must have a valid reference to the task, the
7044 * task must not exit() & deallocate itself prematurely. The
7045 * call is not atomic; no spinlocks may be held.
7047 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7049 struct migration_req req;
7050 unsigned long flags;
7054 rq = task_rq_lock(p, &flags);
7055 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7060 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7061 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7066 if (p->sched_class->set_cpus_allowed)
7067 p->sched_class->set_cpus_allowed(p, new_mask);
7069 cpumask_copy(&p->cpus_allowed, new_mask);
7070 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7073 /* Can the task run on the task's current CPU? If so, we're done */
7074 if (cpumask_test_cpu(task_cpu(p), new_mask))
7077 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7078 /* Need help from migration thread: drop lock and wait. */
7079 struct task_struct *mt = rq->migration_thread;
7081 get_task_struct(mt);
7082 task_rq_unlock(rq, &flags);
7083 wake_up_process(rq->migration_thread);
7084 put_task_struct(mt);
7085 wait_for_completion(&req.done);
7086 tlb_migrate_finish(p->mm);
7090 task_rq_unlock(rq, &flags);
7094 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7097 * Move (not current) task off this cpu, onto dest cpu. We're doing
7098 * this because either it can't run here any more (set_cpus_allowed()
7099 * away from this CPU, or CPU going down), or because we're
7100 * attempting to rebalance this task on exec (sched_exec).
7102 * So we race with normal scheduler movements, but that's OK, as long
7103 * as the task is no longer on this CPU.
7105 * Returns non-zero if task was successfully migrated.
7107 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7109 struct rq *rq_dest, *rq_src;
7112 if (unlikely(!cpu_active(dest_cpu)))
7115 rq_src = cpu_rq(src_cpu);
7116 rq_dest = cpu_rq(dest_cpu);
7118 double_rq_lock(rq_src, rq_dest);
7119 /* Already moved. */
7120 if (task_cpu(p) != src_cpu)
7122 /* Affinity changed (again). */
7123 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7126 on_rq = p->se.on_rq;
7128 deactivate_task(rq_src, p, 0);
7130 set_task_cpu(p, dest_cpu);
7132 activate_task(rq_dest, p, 0);
7133 check_preempt_curr(rq_dest, p, 0);
7138 double_rq_unlock(rq_src, rq_dest);
7143 * migration_thread - this is a highprio system thread that performs
7144 * thread migration by bumping thread off CPU then 'pushing' onto
7147 static int migration_thread(void *data)
7149 int cpu = (long)data;
7153 BUG_ON(rq->migration_thread != current);
7155 set_current_state(TASK_INTERRUPTIBLE);
7156 while (!kthread_should_stop()) {
7157 struct migration_req *req;
7158 struct list_head *head;
7160 spin_lock_irq(&rq->lock);
7162 if (cpu_is_offline(cpu)) {
7163 spin_unlock_irq(&rq->lock);
7167 if (rq->active_balance) {
7168 active_load_balance(rq, cpu);
7169 rq->active_balance = 0;
7172 head = &rq->migration_queue;
7174 if (list_empty(head)) {
7175 spin_unlock_irq(&rq->lock);
7177 set_current_state(TASK_INTERRUPTIBLE);
7180 req = list_entry(head->next, struct migration_req, list);
7181 list_del_init(head->next);
7183 spin_unlock(&rq->lock);
7184 __migrate_task(req->task, cpu, req->dest_cpu);
7187 complete(&req->done);
7189 __set_current_state(TASK_RUNNING);
7194 #ifdef CONFIG_HOTPLUG_CPU
7196 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7200 local_irq_disable();
7201 ret = __migrate_task(p, src_cpu, dest_cpu);
7207 * Figure out where task on dead CPU should go, use force if necessary.
7209 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7212 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7215 /* Look for allowed, online CPU in same node. */
7216 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7217 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7220 /* Any allowed, online CPU? */
7221 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7222 if (dest_cpu < nr_cpu_ids)
7225 /* No more Mr. Nice Guy. */
7226 if (dest_cpu >= nr_cpu_ids) {
7227 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7228 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7231 * Don't tell them about moving exiting tasks or
7232 * kernel threads (both mm NULL), since they never
7235 if (p->mm && printk_ratelimit()) {
7236 printk(KERN_INFO "process %d (%s) no "
7237 "longer affine to cpu%d\n",
7238 task_pid_nr(p), p->comm, dead_cpu);
7243 /* It can have affinity changed while we were choosing. */
7244 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7249 * While a dead CPU has no uninterruptible tasks queued at this point,
7250 * it might still have a nonzero ->nr_uninterruptible counter, because
7251 * for performance reasons the counter is not stricly tracking tasks to
7252 * their home CPUs. So we just add the counter to another CPU's counter,
7253 * to keep the global sum constant after CPU-down:
7255 static void migrate_nr_uninterruptible(struct rq *rq_src)
7257 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7258 unsigned long flags;
7260 local_irq_save(flags);
7261 double_rq_lock(rq_src, rq_dest);
7262 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7263 rq_src->nr_uninterruptible = 0;
7264 double_rq_unlock(rq_src, rq_dest);
7265 local_irq_restore(flags);
7268 /* Run through task list and migrate tasks from the dead cpu. */
7269 static void migrate_live_tasks(int src_cpu)
7271 struct task_struct *p, *t;
7273 read_lock(&tasklist_lock);
7275 do_each_thread(t, p) {
7279 if (task_cpu(p) == src_cpu)
7280 move_task_off_dead_cpu(src_cpu, p);
7281 } while_each_thread(t, p);
7283 read_unlock(&tasklist_lock);
7287 * Schedules idle task to be the next runnable task on current CPU.
7288 * It does so by boosting its priority to highest possible.
7289 * Used by CPU offline code.
7291 void sched_idle_next(void)
7293 int this_cpu = smp_processor_id();
7294 struct rq *rq = cpu_rq(this_cpu);
7295 struct task_struct *p = rq->idle;
7296 unsigned long flags;
7298 /* cpu has to be offline */
7299 BUG_ON(cpu_online(this_cpu));
7302 * Strictly not necessary since rest of the CPUs are stopped by now
7303 * and interrupts disabled on the current cpu.
7305 spin_lock_irqsave(&rq->lock, flags);
7307 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7309 update_rq_clock(rq);
7310 activate_task(rq, p, 0);
7312 spin_unlock_irqrestore(&rq->lock, flags);
7316 * Ensures that the idle task is using init_mm right before its cpu goes
7319 void idle_task_exit(void)
7321 struct mm_struct *mm = current->active_mm;
7323 BUG_ON(cpu_online(smp_processor_id()));
7326 switch_mm(mm, &init_mm, current);
7330 /* called under rq->lock with disabled interrupts */
7331 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7333 struct rq *rq = cpu_rq(dead_cpu);
7335 /* Must be exiting, otherwise would be on tasklist. */
7336 BUG_ON(!p->exit_state);
7338 /* Cannot have done final schedule yet: would have vanished. */
7339 BUG_ON(p->state == TASK_DEAD);
7344 * Drop lock around migration; if someone else moves it,
7345 * that's OK. No task can be added to this CPU, so iteration is
7348 spin_unlock_irq(&rq->lock);
7349 move_task_off_dead_cpu(dead_cpu, p);
7350 spin_lock_irq(&rq->lock);
7355 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7356 static void migrate_dead_tasks(unsigned int dead_cpu)
7358 struct rq *rq = cpu_rq(dead_cpu);
7359 struct task_struct *next;
7362 if (!rq->nr_running)
7364 update_rq_clock(rq);
7365 next = pick_next_task(rq);
7368 next->sched_class->put_prev_task(rq, next);
7369 migrate_dead(dead_cpu, next);
7375 * remove the tasks which were accounted by rq from calc_load_tasks.
7377 static void calc_global_load_remove(struct rq *rq)
7379 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7380 rq->calc_load_active = 0;
7382 #endif /* CONFIG_HOTPLUG_CPU */
7384 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7386 static struct ctl_table sd_ctl_dir[] = {
7388 .procname = "sched_domain",
7394 static struct ctl_table sd_ctl_root[] = {
7396 .ctl_name = CTL_KERN,
7397 .procname = "kernel",
7399 .child = sd_ctl_dir,
7404 static struct ctl_table *sd_alloc_ctl_entry(int n)
7406 struct ctl_table *entry =
7407 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7412 static void sd_free_ctl_entry(struct ctl_table **tablep)
7414 struct ctl_table *entry;
7417 * In the intermediate directories, both the child directory and
7418 * procname are dynamically allocated and could fail but the mode
7419 * will always be set. In the lowest directory the names are
7420 * static strings and all have proc handlers.
7422 for (entry = *tablep; entry->mode; entry++) {
7424 sd_free_ctl_entry(&entry->child);
7425 if (entry->proc_handler == NULL)
7426 kfree(entry->procname);
7434 set_table_entry(struct ctl_table *entry,
7435 const char *procname, void *data, int maxlen,
7436 mode_t mode, proc_handler *proc_handler)
7438 entry->procname = procname;
7440 entry->maxlen = maxlen;
7442 entry->proc_handler = proc_handler;
7445 static struct ctl_table *
7446 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7448 struct ctl_table *table = sd_alloc_ctl_entry(13);
7453 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7454 sizeof(long), 0644, proc_doulongvec_minmax);
7455 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7456 sizeof(long), 0644, proc_doulongvec_minmax);
7457 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7458 sizeof(int), 0644, proc_dointvec_minmax);
7459 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7460 sizeof(int), 0644, proc_dointvec_minmax);
7461 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7462 sizeof(int), 0644, proc_dointvec_minmax);
7463 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7464 sizeof(int), 0644, proc_dointvec_minmax);
7465 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7466 sizeof(int), 0644, proc_dointvec_minmax);
7467 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7468 sizeof(int), 0644, proc_dointvec_minmax);
7469 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7470 sizeof(int), 0644, proc_dointvec_minmax);
7471 set_table_entry(&table[9], "cache_nice_tries",
7472 &sd->cache_nice_tries,
7473 sizeof(int), 0644, proc_dointvec_minmax);
7474 set_table_entry(&table[10], "flags", &sd->flags,
7475 sizeof(int), 0644, proc_dointvec_minmax);
7476 set_table_entry(&table[11], "name", sd->name,
7477 CORENAME_MAX_SIZE, 0444, proc_dostring);
7478 /* &table[12] is terminator */
7483 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7485 struct ctl_table *entry, *table;
7486 struct sched_domain *sd;
7487 int domain_num = 0, i;
7490 for_each_domain(cpu, sd)
7492 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7497 for_each_domain(cpu, sd) {
7498 snprintf(buf, 32, "domain%d", i);
7499 entry->procname = kstrdup(buf, GFP_KERNEL);
7501 entry->child = sd_alloc_ctl_domain_table(sd);
7508 static struct ctl_table_header *sd_sysctl_header;
7509 static void register_sched_domain_sysctl(void)
7511 int i, cpu_num = num_online_cpus();
7512 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7515 WARN_ON(sd_ctl_dir[0].child);
7516 sd_ctl_dir[0].child = entry;
7521 for_each_online_cpu(i) {
7522 snprintf(buf, 32, "cpu%d", i);
7523 entry->procname = kstrdup(buf, GFP_KERNEL);
7525 entry->child = sd_alloc_ctl_cpu_table(i);
7529 WARN_ON(sd_sysctl_header);
7530 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7533 /* may be called multiple times per register */
7534 static void unregister_sched_domain_sysctl(void)
7536 if (sd_sysctl_header)
7537 unregister_sysctl_table(sd_sysctl_header);
7538 sd_sysctl_header = NULL;
7539 if (sd_ctl_dir[0].child)
7540 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7543 static void register_sched_domain_sysctl(void)
7546 static void unregister_sched_domain_sysctl(void)
7551 static void set_rq_online(struct rq *rq)
7554 const struct sched_class *class;
7556 cpumask_set_cpu(rq->cpu, rq->rd->online);
7559 for_each_class(class) {
7560 if (class->rq_online)
7561 class->rq_online(rq);
7566 static void set_rq_offline(struct rq *rq)
7569 const struct sched_class *class;
7571 for_each_class(class) {
7572 if (class->rq_offline)
7573 class->rq_offline(rq);
7576 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7582 * migration_call - callback that gets triggered when a CPU is added.
7583 * Here we can start up the necessary migration thread for the new CPU.
7585 static int __cpuinit
7586 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7588 struct task_struct *p;
7589 int cpu = (long)hcpu;
7590 unsigned long flags;
7595 case CPU_UP_PREPARE:
7596 case CPU_UP_PREPARE_FROZEN:
7597 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7600 kthread_bind(p, cpu);
7601 /* Must be high prio: stop_machine expects to yield to it. */
7602 rq = task_rq_lock(p, &flags);
7603 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7604 task_rq_unlock(rq, &flags);
7606 cpu_rq(cpu)->migration_thread = p;
7607 rq->calc_load_update = calc_load_update;
7611 case CPU_ONLINE_FROZEN:
7612 /* Strictly unnecessary, as first user will wake it. */
7613 wake_up_process(cpu_rq(cpu)->migration_thread);
7615 /* Update our root-domain */
7617 spin_lock_irqsave(&rq->lock, flags);
7619 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7623 spin_unlock_irqrestore(&rq->lock, flags);
7626 #ifdef CONFIG_HOTPLUG_CPU
7627 case CPU_UP_CANCELED:
7628 case CPU_UP_CANCELED_FROZEN:
7629 if (!cpu_rq(cpu)->migration_thread)
7631 /* Unbind it from offline cpu so it can run. Fall thru. */
7632 kthread_bind(cpu_rq(cpu)->migration_thread,
7633 cpumask_any(cpu_online_mask));
7634 kthread_stop(cpu_rq(cpu)->migration_thread);
7635 put_task_struct(cpu_rq(cpu)->migration_thread);
7636 cpu_rq(cpu)->migration_thread = NULL;
7640 case CPU_DEAD_FROZEN:
7641 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7642 migrate_live_tasks(cpu);
7644 kthread_stop(rq->migration_thread);
7645 put_task_struct(rq->migration_thread);
7646 rq->migration_thread = NULL;
7647 /* Idle task back to normal (off runqueue, low prio) */
7648 spin_lock_irq(&rq->lock);
7649 update_rq_clock(rq);
7650 deactivate_task(rq, rq->idle, 0);
7651 rq->idle->static_prio = MAX_PRIO;
7652 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7653 rq->idle->sched_class = &idle_sched_class;
7654 migrate_dead_tasks(cpu);
7655 spin_unlock_irq(&rq->lock);
7657 migrate_nr_uninterruptible(rq);
7658 BUG_ON(rq->nr_running != 0);
7659 calc_global_load_remove(rq);
7661 * No need to migrate the tasks: it was best-effort if
7662 * they didn't take sched_hotcpu_mutex. Just wake up
7665 spin_lock_irq(&rq->lock);
7666 while (!list_empty(&rq->migration_queue)) {
7667 struct migration_req *req;
7669 req = list_entry(rq->migration_queue.next,
7670 struct migration_req, list);
7671 list_del_init(&req->list);
7672 spin_unlock_irq(&rq->lock);
7673 complete(&req->done);
7674 spin_lock_irq(&rq->lock);
7676 spin_unlock_irq(&rq->lock);
7680 case CPU_DYING_FROZEN:
7681 /* Update our root-domain */
7683 spin_lock_irqsave(&rq->lock, flags);
7685 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7688 spin_unlock_irqrestore(&rq->lock, flags);
7696 * Register at high priority so that task migration (migrate_all_tasks)
7697 * happens before everything else. This has to be lower priority than
7698 * the notifier in the perf_counter subsystem, though.
7700 static struct notifier_block __cpuinitdata migration_notifier = {
7701 .notifier_call = migration_call,
7705 static int __init migration_init(void)
7707 void *cpu = (void *)(long)smp_processor_id();
7710 /* Start one for the boot CPU: */
7711 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7712 BUG_ON(err == NOTIFY_BAD);
7713 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7714 register_cpu_notifier(&migration_notifier);
7718 early_initcall(migration_init);
7723 #ifdef CONFIG_SCHED_DEBUG
7725 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7726 struct cpumask *groupmask)
7728 struct sched_group *group = sd->groups;
7731 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7732 cpumask_clear(groupmask);
7734 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7736 if (!(sd->flags & SD_LOAD_BALANCE)) {
7737 printk("does not load-balance\n");
7739 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7744 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7746 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7747 printk(KERN_ERR "ERROR: domain->span does not contain "
7750 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7751 printk(KERN_ERR "ERROR: domain->groups does not contain"
7755 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7759 printk(KERN_ERR "ERROR: group is NULL\n");
7763 if (!group->__cpu_power) {
7764 printk(KERN_CONT "\n");
7765 printk(KERN_ERR "ERROR: domain->cpu_power not "
7770 if (!cpumask_weight(sched_group_cpus(group))) {
7771 printk(KERN_CONT "\n");
7772 printk(KERN_ERR "ERROR: empty group\n");
7776 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7777 printk(KERN_CONT "\n");
7778 printk(KERN_ERR "ERROR: repeated CPUs\n");
7782 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7784 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7786 printk(KERN_CONT " %s", str);
7787 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7788 printk(KERN_CONT " (__cpu_power = %d)",
7789 group->__cpu_power);
7792 group = group->next;
7793 } while (group != sd->groups);
7794 printk(KERN_CONT "\n");
7796 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7797 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7800 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7801 printk(KERN_ERR "ERROR: parent span is not a superset "
7802 "of domain->span\n");
7806 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7808 cpumask_var_t groupmask;
7812 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7816 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7818 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7819 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7824 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7831 free_cpumask_var(groupmask);
7833 #else /* !CONFIG_SCHED_DEBUG */
7834 # define sched_domain_debug(sd, cpu) do { } while (0)
7835 #endif /* CONFIG_SCHED_DEBUG */
7837 static int sd_degenerate(struct sched_domain *sd)
7839 if (cpumask_weight(sched_domain_span(sd)) == 1)
7842 /* Following flags need at least 2 groups */
7843 if (sd->flags & (SD_LOAD_BALANCE |
7844 SD_BALANCE_NEWIDLE |
7848 SD_SHARE_PKG_RESOURCES)) {
7849 if (sd->groups != sd->groups->next)
7853 /* Following flags don't use groups */
7854 if (sd->flags & (SD_WAKE_IDLE |
7863 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7865 unsigned long cflags = sd->flags, pflags = parent->flags;
7867 if (sd_degenerate(parent))
7870 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7873 /* Does parent contain flags not in child? */
7874 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7875 if (cflags & SD_WAKE_AFFINE)
7876 pflags &= ~SD_WAKE_BALANCE;
7877 /* Flags needing groups don't count if only 1 group in parent */
7878 if (parent->groups == parent->groups->next) {
7879 pflags &= ~(SD_LOAD_BALANCE |
7880 SD_BALANCE_NEWIDLE |
7884 SD_SHARE_PKG_RESOURCES);
7885 if (nr_node_ids == 1)
7886 pflags &= ~SD_SERIALIZE;
7888 if (~cflags & pflags)
7894 static void free_rootdomain(struct root_domain *rd)
7896 cpupri_cleanup(&rd->cpupri);
7898 free_cpumask_var(rd->rto_mask);
7899 free_cpumask_var(rd->online);
7900 free_cpumask_var(rd->span);
7904 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7906 struct root_domain *old_rd = NULL;
7907 unsigned long flags;
7909 spin_lock_irqsave(&rq->lock, flags);
7914 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7917 cpumask_clear_cpu(rq->cpu, old_rd->span);
7920 * If we dont want to free the old_rt yet then
7921 * set old_rd to NULL to skip the freeing later
7924 if (!atomic_dec_and_test(&old_rd->refcount))
7928 atomic_inc(&rd->refcount);
7931 cpumask_set_cpu(rq->cpu, rd->span);
7932 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7935 spin_unlock_irqrestore(&rq->lock, flags);
7938 free_rootdomain(old_rd);
7941 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7943 gfp_t gfp = GFP_KERNEL;
7945 memset(rd, 0, sizeof(*rd));
7950 if (!alloc_cpumask_var(&rd->span, gfp))
7952 if (!alloc_cpumask_var(&rd->online, gfp))
7954 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7957 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7962 free_cpumask_var(rd->rto_mask);
7964 free_cpumask_var(rd->online);
7966 free_cpumask_var(rd->span);
7971 static void init_defrootdomain(void)
7973 init_rootdomain(&def_root_domain, true);
7975 atomic_set(&def_root_domain.refcount, 1);
7978 static struct root_domain *alloc_rootdomain(void)
7980 struct root_domain *rd;
7982 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7986 if (init_rootdomain(rd, false) != 0) {
7995 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7996 * hold the hotplug lock.
7999 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8001 struct rq *rq = cpu_rq(cpu);
8002 struct sched_domain *tmp;
8004 /* Remove the sched domains which do not contribute to scheduling. */
8005 for (tmp = sd; tmp; ) {
8006 struct sched_domain *parent = tmp->parent;
8010 if (sd_parent_degenerate(tmp, parent)) {
8011 tmp->parent = parent->parent;
8013 parent->parent->child = tmp;
8018 if (sd && sd_degenerate(sd)) {
8024 sched_domain_debug(sd, cpu);
8026 rq_attach_root(rq, rd);
8027 rcu_assign_pointer(rq->sd, sd);
8030 /* cpus with isolated domains */
8031 static cpumask_var_t cpu_isolated_map;
8033 /* Setup the mask of cpus configured for isolated domains */
8034 static int __init isolated_cpu_setup(char *str)
8036 cpulist_parse(str, cpu_isolated_map);
8040 __setup("isolcpus=", isolated_cpu_setup);
8043 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8044 * to a function which identifies what group(along with sched group) a CPU
8045 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8046 * (due to the fact that we keep track of groups covered with a struct cpumask).
8048 * init_sched_build_groups will build a circular linked list of the groups
8049 * covered by the given span, and will set each group's ->cpumask correctly,
8050 * and ->cpu_power to 0.
8053 init_sched_build_groups(const struct cpumask *span,
8054 const struct cpumask *cpu_map,
8055 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8056 struct sched_group **sg,
8057 struct cpumask *tmpmask),
8058 struct cpumask *covered, struct cpumask *tmpmask)
8060 struct sched_group *first = NULL, *last = NULL;
8063 cpumask_clear(covered);
8065 for_each_cpu(i, span) {
8066 struct sched_group *sg;
8067 int group = group_fn(i, cpu_map, &sg, tmpmask);
8070 if (cpumask_test_cpu(i, covered))
8073 cpumask_clear(sched_group_cpus(sg));
8074 sg->__cpu_power = 0;
8076 for_each_cpu(j, span) {
8077 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8080 cpumask_set_cpu(j, covered);
8081 cpumask_set_cpu(j, sched_group_cpus(sg));
8092 #define SD_NODES_PER_DOMAIN 16
8097 * find_next_best_node - find the next node to include in a sched_domain
8098 * @node: node whose sched_domain we're building
8099 * @used_nodes: nodes already in the sched_domain
8101 * Find the next node to include in a given scheduling domain. Simply
8102 * finds the closest node not already in the @used_nodes map.
8104 * Should use nodemask_t.
8106 static int find_next_best_node(int node, nodemask_t *used_nodes)
8108 int i, n, val, min_val, best_node = 0;
8112 for (i = 0; i < nr_node_ids; i++) {
8113 /* Start at @node */
8114 n = (node + i) % nr_node_ids;
8116 if (!nr_cpus_node(n))
8119 /* Skip already used nodes */
8120 if (node_isset(n, *used_nodes))
8123 /* Simple min distance search */
8124 val = node_distance(node, n);
8126 if (val < min_val) {
8132 node_set(best_node, *used_nodes);
8137 * sched_domain_node_span - get a cpumask for a node's sched_domain
8138 * @node: node whose cpumask we're constructing
8139 * @span: resulting cpumask
8141 * Given a node, construct a good cpumask for its sched_domain to span. It
8142 * should be one that prevents unnecessary balancing, but also spreads tasks
8145 static void sched_domain_node_span(int node, struct cpumask *span)
8147 nodemask_t used_nodes;
8150 cpumask_clear(span);
8151 nodes_clear(used_nodes);
8153 cpumask_or(span, span, cpumask_of_node(node));
8154 node_set(node, used_nodes);
8156 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8157 int next_node = find_next_best_node(node, &used_nodes);
8159 cpumask_or(span, span, cpumask_of_node(next_node));
8162 #endif /* CONFIG_NUMA */
8164 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8167 * The cpus mask in sched_group and sched_domain hangs off the end.
8169 * ( See the the comments in include/linux/sched.h:struct sched_group
8170 * and struct sched_domain. )
8172 struct static_sched_group {
8173 struct sched_group sg;
8174 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8177 struct static_sched_domain {
8178 struct sched_domain sd;
8179 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8183 * SMT sched-domains:
8185 #ifdef CONFIG_SCHED_SMT
8186 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8187 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8190 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8191 struct sched_group **sg, struct cpumask *unused)
8194 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8197 #endif /* CONFIG_SCHED_SMT */
8200 * multi-core sched-domains:
8202 #ifdef CONFIG_SCHED_MC
8203 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8204 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8205 #endif /* CONFIG_SCHED_MC */
8207 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8209 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8210 struct sched_group **sg, struct cpumask *mask)
8214 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8215 group = cpumask_first(mask);
8217 *sg = &per_cpu(sched_group_core, group).sg;
8220 #elif defined(CONFIG_SCHED_MC)
8222 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8223 struct sched_group **sg, struct cpumask *unused)
8226 *sg = &per_cpu(sched_group_core, cpu).sg;
8231 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8232 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8235 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8236 struct sched_group **sg, struct cpumask *mask)
8239 #ifdef CONFIG_SCHED_MC
8240 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8241 group = cpumask_first(mask);
8242 #elif defined(CONFIG_SCHED_SMT)
8243 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8244 group = cpumask_first(mask);
8249 *sg = &per_cpu(sched_group_phys, group).sg;
8255 * The init_sched_build_groups can't handle what we want to do with node
8256 * groups, so roll our own. Now each node has its own list of groups which
8257 * gets dynamically allocated.
8259 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8260 static struct sched_group ***sched_group_nodes_bycpu;
8262 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8263 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8265 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8266 struct sched_group **sg,
8267 struct cpumask *nodemask)
8271 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8272 group = cpumask_first(nodemask);
8275 *sg = &per_cpu(sched_group_allnodes, group).sg;
8279 static void init_numa_sched_groups_power(struct sched_group *group_head)
8281 struct sched_group *sg = group_head;
8287 for_each_cpu(j, sched_group_cpus(sg)) {
8288 struct sched_domain *sd;
8290 sd = &per_cpu(phys_domains, j).sd;
8291 if (j != group_first_cpu(sd->groups)) {
8293 * Only add "power" once for each
8299 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8302 } while (sg != group_head);
8304 #endif /* CONFIG_NUMA */
8307 /* Free memory allocated for various sched_group structures */
8308 static void free_sched_groups(const struct cpumask *cpu_map,
8309 struct cpumask *nodemask)
8313 for_each_cpu(cpu, cpu_map) {
8314 struct sched_group **sched_group_nodes
8315 = sched_group_nodes_bycpu[cpu];
8317 if (!sched_group_nodes)
8320 for (i = 0; i < nr_node_ids; i++) {
8321 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8323 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8324 if (cpumask_empty(nodemask))
8334 if (oldsg != sched_group_nodes[i])
8337 kfree(sched_group_nodes);
8338 sched_group_nodes_bycpu[cpu] = NULL;
8341 #else /* !CONFIG_NUMA */
8342 static void free_sched_groups(const struct cpumask *cpu_map,
8343 struct cpumask *nodemask)
8346 #endif /* CONFIG_NUMA */
8349 * Initialize sched groups cpu_power.
8351 * cpu_power indicates the capacity of sched group, which is used while
8352 * distributing the load between different sched groups in a sched domain.
8353 * Typically cpu_power for all the groups in a sched domain will be same unless
8354 * there are asymmetries in the topology. If there are asymmetries, group
8355 * having more cpu_power will pickup more load compared to the group having
8358 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8359 * the maximum number of tasks a group can handle in the presence of other idle
8360 * or lightly loaded groups in the same sched domain.
8362 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8364 struct sched_domain *child;
8365 struct sched_group *group;
8367 WARN_ON(!sd || !sd->groups);
8369 if (cpu != group_first_cpu(sd->groups))
8374 sd->groups->__cpu_power = 0;
8377 * For perf policy, if the groups in child domain share resources
8378 * (for example cores sharing some portions of the cache hierarchy
8379 * or SMT), then set this domain groups cpu_power such that each group
8380 * can handle only one task, when there are other idle groups in the
8381 * same sched domain.
8383 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8385 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
8386 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
8391 * add cpu_power of each child group to this groups cpu_power
8393 group = child->groups;
8395 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8396 group = group->next;
8397 } while (group != child->groups);
8401 * Initializers for schedule domains
8402 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8405 #ifdef CONFIG_SCHED_DEBUG
8406 # define SD_INIT_NAME(sd, type) sd->name = #type
8408 # define SD_INIT_NAME(sd, type) do { } while (0)
8411 #define SD_INIT(sd, type) sd_init_##type(sd)
8413 #define SD_INIT_FUNC(type) \
8414 static noinline void sd_init_##type(struct sched_domain *sd) \
8416 memset(sd, 0, sizeof(*sd)); \
8417 *sd = SD_##type##_INIT; \
8418 sd->level = SD_LV_##type; \
8419 SD_INIT_NAME(sd, type); \
8424 SD_INIT_FUNC(ALLNODES)
8427 #ifdef CONFIG_SCHED_SMT
8428 SD_INIT_FUNC(SIBLING)
8430 #ifdef CONFIG_SCHED_MC
8434 static int default_relax_domain_level = -1;
8436 static int __init setup_relax_domain_level(char *str)
8440 val = simple_strtoul(str, NULL, 0);
8441 if (val < SD_LV_MAX)
8442 default_relax_domain_level = val;
8446 __setup("relax_domain_level=", setup_relax_domain_level);
8448 static void set_domain_attribute(struct sched_domain *sd,
8449 struct sched_domain_attr *attr)
8453 if (!attr || attr->relax_domain_level < 0) {
8454 if (default_relax_domain_level < 0)
8457 request = default_relax_domain_level;
8459 request = attr->relax_domain_level;
8460 if (request < sd->level) {
8461 /* turn off idle balance on this domain */
8462 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8464 /* turn on idle balance on this domain */
8465 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8470 * Build sched domains for a given set of cpus and attach the sched domains
8471 * to the individual cpus
8473 static int __build_sched_domains(const struct cpumask *cpu_map,
8474 struct sched_domain_attr *attr)
8476 int i, err = -ENOMEM;
8477 struct root_domain *rd;
8478 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
8481 cpumask_var_t domainspan, covered, notcovered;
8482 struct sched_group **sched_group_nodes = NULL;
8483 int sd_allnodes = 0;
8485 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
8487 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
8488 goto free_domainspan;
8489 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
8493 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
8494 goto free_notcovered;
8495 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
8497 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
8498 goto free_this_sibling_map;
8499 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
8500 goto free_this_core_map;
8501 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
8502 goto free_send_covered;
8506 * Allocate the per-node list of sched groups
8508 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
8510 if (!sched_group_nodes) {
8511 printk(KERN_WARNING "Can not alloc sched group node list\n");
8516 rd = alloc_rootdomain();
8518 printk(KERN_WARNING "Cannot alloc root domain\n");
8519 goto free_sched_groups;
8523 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
8527 * Set up domains for cpus specified by the cpu_map.
8529 for_each_cpu(i, cpu_map) {
8530 struct sched_domain *sd = NULL, *p;
8532 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
8535 if (cpumask_weight(cpu_map) >
8536 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
8537 sd = &per_cpu(allnodes_domains, i).sd;
8538 SD_INIT(sd, ALLNODES);
8539 set_domain_attribute(sd, attr);
8540 cpumask_copy(sched_domain_span(sd), cpu_map);
8541 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
8547 sd = &per_cpu(node_domains, i).sd;
8549 set_domain_attribute(sd, attr);
8550 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8554 cpumask_and(sched_domain_span(sd),
8555 sched_domain_span(sd), cpu_map);
8559 sd = &per_cpu(phys_domains, i).sd;
8561 set_domain_attribute(sd, attr);
8562 cpumask_copy(sched_domain_span(sd), nodemask);
8566 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
8568 #ifdef CONFIG_SCHED_MC
8570 sd = &per_cpu(core_domains, i).sd;
8572 set_domain_attribute(sd, attr);
8573 cpumask_and(sched_domain_span(sd), cpu_map,
8574 cpu_coregroup_mask(i));
8577 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
8580 #ifdef CONFIG_SCHED_SMT
8582 sd = &per_cpu(cpu_domains, i).sd;
8583 SD_INIT(sd, SIBLING);
8584 set_domain_attribute(sd, attr);
8585 cpumask_and(sched_domain_span(sd),
8586 topology_thread_cpumask(i), cpu_map);
8589 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8593 #ifdef CONFIG_SCHED_SMT
8594 /* Set up CPU (sibling) groups */
8595 for_each_cpu(i, cpu_map) {
8596 cpumask_and(this_sibling_map,
8597 topology_thread_cpumask(i), cpu_map);
8598 if (i != cpumask_first(this_sibling_map))
8601 init_sched_build_groups(this_sibling_map, cpu_map,
8603 send_covered, tmpmask);
8607 #ifdef CONFIG_SCHED_MC
8608 /* Set up multi-core groups */
8609 for_each_cpu(i, cpu_map) {
8610 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8611 if (i != cpumask_first(this_core_map))
8614 init_sched_build_groups(this_core_map, cpu_map,
8616 send_covered, tmpmask);
8620 /* Set up physical groups */
8621 for (i = 0; i < nr_node_ids; i++) {
8622 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8623 if (cpumask_empty(nodemask))
8626 init_sched_build_groups(nodemask, cpu_map,
8628 send_covered, tmpmask);
8632 /* Set up node groups */
8634 init_sched_build_groups(cpu_map, cpu_map,
8635 &cpu_to_allnodes_group,
8636 send_covered, tmpmask);
8639 for (i = 0; i < nr_node_ids; i++) {
8640 /* Set up node groups */
8641 struct sched_group *sg, *prev;
8644 cpumask_clear(covered);
8645 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8646 if (cpumask_empty(nodemask)) {
8647 sched_group_nodes[i] = NULL;
8651 sched_domain_node_span(i, domainspan);
8652 cpumask_and(domainspan, domainspan, cpu_map);
8654 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8657 printk(KERN_WARNING "Can not alloc domain group for "
8661 sched_group_nodes[i] = sg;
8662 for_each_cpu(j, nodemask) {
8663 struct sched_domain *sd;
8665 sd = &per_cpu(node_domains, j).sd;
8668 sg->__cpu_power = 0;
8669 cpumask_copy(sched_group_cpus(sg), nodemask);
8671 cpumask_or(covered, covered, nodemask);
8674 for (j = 0; j < nr_node_ids; j++) {
8675 int n = (i + j) % nr_node_ids;
8677 cpumask_complement(notcovered, covered);
8678 cpumask_and(tmpmask, notcovered, cpu_map);
8679 cpumask_and(tmpmask, tmpmask, domainspan);
8680 if (cpumask_empty(tmpmask))
8683 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8684 if (cpumask_empty(tmpmask))
8687 sg = kmalloc_node(sizeof(struct sched_group) +
8692 "Can not alloc domain group for node %d\n", j);
8695 sg->__cpu_power = 0;
8696 cpumask_copy(sched_group_cpus(sg), tmpmask);
8697 sg->next = prev->next;
8698 cpumask_or(covered, covered, tmpmask);
8705 /* Calculate CPU power for physical packages and nodes */
8706 #ifdef CONFIG_SCHED_SMT
8707 for_each_cpu(i, cpu_map) {
8708 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8710 init_sched_groups_power(i, sd);
8713 #ifdef CONFIG_SCHED_MC
8714 for_each_cpu(i, cpu_map) {
8715 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8717 init_sched_groups_power(i, sd);
8721 for_each_cpu(i, cpu_map) {
8722 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8724 init_sched_groups_power(i, sd);
8728 for (i = 0; i < nr_node_ids; i++)
8729 init_numa_sched_groups_power(sched_group_nodes[i]);
8732 struct sched_group *sg;
8734 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8736 init_numa_sched_groups_power(sg);
8740 /* Attach the domains */
8741 for_each_cpu(i, cpu_map) {
8742 struct sched_domain *sd;
8743 #ifdef CONFIG_SCHED_SMT
8744 sd = &per_cpu(cpu_domains, i).sd;
8745 #elif defined(CONFIG_SCHED_MC)
8746 sd = &per_cpu(core_domains, i).sd;
8748 sd = &per_cpu(phys_domains, i).sd;
8750 cpu_attach_domain(sd, rd, i);
8756 free_cpumask_var(tmpmask);
8758 free_cpumask_var(send_covered);
8760 free_cpumask_var(this_core_map);
8761 free_this_sibling_map:
8762 free_cpumask_var(this_sibling_map);
8764 free_cpumask_var(nodemask);
8767 free_cpumask_var(notcovered);
8769 free_cpumask_var(covered);
8771 free_cpumask_var(domainspan);
8778 kfree(sched_group_nodes);
8784 free_sched_groups(cpu_map, tmpmask);
8785 free_rootdomain(rd);
8790 static int build_sched_domains(const struct cpumask *cpu_map)
8792 return __build_sched_domains(cpu_map, NULL);
8795 static struct cpumask *doms_cur; /* current sched domains */
8796 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8797 static struct sched_domain_attr *dattr_cur;
8798 /* attribues of custom domains in 'doms_cur' */
8801 * Special case: If a kmalloc of a doms_cur partition (array of
8802 * cpumask) fails, then fallback to a single sched domain,
8803 * as determined by the single cpumask fallback_doms.
8805 static cpumask_var_t fallback_doms;
8808 * arch_update_cpu_topology lets virtualized architectures update the
8809 * cpu core maps. It is supposed to return 1 if the topology changed
8810 * or 0 if it stayed the same.
8812 int __attribute__((weak)) arch_update_cpu_topology(void)
8818 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8819 * For now this just excludes isolated cpus, but could be used to
8820 * exclude other special cases in the future.
8822 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8826 arch_update_cpu_topology();
8828 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8830 doms_cur = fallback_doms;
8831 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8833 err = build_sched_domains(doms_cur);
8834 register_sched_domain_sysctl();
8839 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8840 struct cpumask *tmpmask)
8842 free_sched_groups(cpu_map, tmpmask);
8846 * Detach sched domains from a group of cpus specified in cpu_map
8847 * These cpus will now be attached to the NULL domain
8849 static void detach_destroy_domains(const struct cpumask *cpu_map)
8851 /* Save because hotplug lock held. */
8852 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8855 for_each_cpu(i, cpu_map)
8856 cpu_attach_domain(NULL, &def_root_domain, i);
8857 synchronize_sched();
8858 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8861 /* handle null as "default" */
8862 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8863 struct sched_domain_attr *new, int idx_new)
8865 struct sched_domain_attr tmp;
8872 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8873 new ? (new + idx_new) : &tmp,
8874 sizeof(struct sched_domain_attr));
8878 * Partition sched domains as specified by the 'ndoms_new'
8879 * cpumasks in the array doms_new[] of cpumasks. This compares
8880 * doms_new[] to the current sched domain partitioning, doms_cur[].
8881 * It destroys each deleted domain and builds each new domain.
8883 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8884 * The masks don't intersect (don't overlap.) We should setup one
8885 * sched domain for each mask. CPUs not in any of the cpumasks will
8886 * not be load balanced. If the same cpumask appears both in the
8887 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8890 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8891 * ownership of it and will kfree it when done with it. If the caller
8892 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8893 * ndoms_new == 1, and partition_sched_domains() will fallback to
8894 * the single partition 'fallback_doms', it also forces the domains
8897 * If doms_new == NULL it will be replaced with cpu_online_mask.
8898 * ndoms_new == 0 is a special case for destroying existing domains,
8899 * and it will not create the default domain.
8901 * Call with hotplug lock held
8903 /* FIXME: Change to struct cpumask *doms_new[] */
8904 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8905 struct sched_domain_attr *dattr_new)
8910 mutex_lock(&sched_domains_mutex);
8912 /* always unregister in case we don't destroy any domains */
8913 unregister_sched_domain_sysctl();
8915 /* Let architecture update cpu core mappings. */
8916 new_topology = arch_update_cpu_topology();
8918 n = doms_new ? ndoms_new : 0;
8920 /* Destroy deleted domains */
8921 for (i = 0; i < ndoms_cur; i++) {
8922 for (j = 0; j < n && !new_topology; j++) {
8923 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8924 && dattrs_equal(dattr_cur, i, dattr_new, j))
8927 /* no match - a current sched domain not in new doms_new[] */
8928 detach_destroy_domains(doms_cur + i);
8933 if (doms_new == NULL) {
8935 doms_new = fallback_doms;
8936 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8937 WARN_ON_ONCE(dattr_new);
8940 /* Build new domains */
8941 for (i = 0; i < ndoms_new; i++) {
8942 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8943 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8944 && dattrs_equal(dattr_new, i, dattr_cur, j))
8947 /* no match - add a new doms_new */
8948 __build_sched_domains(doms_new + i,
8949 dattr_new ? dattr_new + i : NULL);
8954 /* Remember the new sched domains */
8955 if (doms_cur != fallback_doms)
8957 kfree(dattr_cur); /* kfree(NULL) is safe */
8958 doms_cur = doms_new;
8959 dattr_cur = dattr_new;
8960 ndoms_cur = ndoms_new;
8962 register_sched_domain_sysctl();
8964 mutex_unlock(&sched_domains_mutex);
8967 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8968 static void arch_reinit_sched_domains(void)
8972 /* Destroy domains first to force the rebuild */
8973 partition_sched_domains(0, NULL, NULL);
8975 rebuild_sched_domains();
8979 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8981 unsigned int level = 0;
8983 if (sscanf(buf, "%u", &level) != 1)
8987 * level is always be positive so don't check for
8988 * level < POWERSAVINGS_BALANCE_NONE which is 0
8989 * What happens on 0 or 1 byte write,
8990 * need to check for count as well?
8993 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8997 sched_smt_power_savings = level;
8999 sched_mc_power_savings = level;
9001 arch_reinit_sched_domains();
9006 #ifdef CONFIG_SCHED_MC
9007 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9010 return sprintf(page, "%u\n", sched_mc_power_savings);
9012 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9013 const char *buf, size_t count)
9015 return sched_power_savings_store(buf, count, 0);
9017 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9018 sched_mc_power_savings_show,
9019 sched_mc_power_savings_store);
9022 #ifdef CONFIG_SCHED_SMT
9023 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9026 return sprintf(page, "%u\n", sched_smt_power_savings);
9028 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9029 const char *buf, size_t count)
9031 return sched_power_savings_store(buf, count, 1);
9033 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9034 sched_smt_power_savings_show,
9035 sched_smt_power_savings_store);
9038 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9042 #ifdef CONFIG_SCHED_SMT
9044 err = sysfs_create_file(&cls->kset.kobj,
9045 &attr_sched_smt_power_savings.attr);
9047 #ifdef CONFIG_SCHED_MC
9048 if (!err && mc_capable())
9049 err = sysfs_create_file(&cls->kset.kobj,
9050 &attr_sched_mc_power_savings.attr);
9054 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9056 #ifndef CONFIG_CPUSETS
9058 * Add online and remove offline CPUs from the scheduler domains.
9059 * When cpusets are enabled they take over this function.
9061 static int update_sched_domains(struct notifier_block *nfb,
9062 unsigned long action, void *hcpu)
9066 case CPU_ONLINE_FROZEN:
9068 case CPU_DEAD_FROZEN:
9069 partition_sched_domains(1, NULL, NULL);
9078 static int update_runtime(struct notifier_block *nfb,
9079 unsigned long action, void *hcpu)
9081 int cpu = (int)(long)hcpu;
9084 case CPU_DOWN_PREPARE:
9085 case CPU_DOWN_PREPARE_FROZEN:
9086 disable_runtime(cpu_rq(cpu));
9089 case CPU_DOWN_FAILED:
9090 case CPU_DOWN_FAILED_FROZEN:
9092 case CPU_ONLINE_FROZEN:
9093 enable_runtime(cpu_rq(cpu));
9101 void __init sched_init_smp(void)
9103 cpumask_var_t non_isolated_cpus;
9105 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9107 #if defined(CONFIG_NUMA)
9108 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9110 BUG_ON(sched_group_nodes_bycpu == NULL);
9113 mutex_lock(&sched_domains_mutex);
9114 arch_init_sched_domains(cpu_online_mask);
9115 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9116 if (cpumask_empty(non_isolated_cpus))
9117 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9118 mutex_unlock(&sched_domains_mutex);
9121 #ifndef CONFIG_CPUSETS
9122 /* XXX: Theoretical race here - CPU may be hotplugged now */
9123 hotcpu_notifier(update_sched_domains, 0);
9126 /* RT runtime code needs to handle some hotplug events */
9127 hotcpu_notifier(update_runtime, 0);
9131 /* Move init over to a non-isolated CPU */
9132 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9134 sched_init_granularity();
9135 free_cpumask_var(non_isolated_cpus);
9137 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9138 init_sched_rt_class();
9141 void __init sched_init_smp(void)
9143 sched_init_granularity();
9145 #endif /* CONFIG_SMP */
9147 const_debug unsigned int sysctl_timer_migration = 1;
9149 int in_sched_functions(unsigned long addr)
9151 return in_lock_functions(addr) ||
9152 (addr >= (unsigned long)__sched_text_start
9153 && addr < (unsigned long)__sched_text_end);
9156 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9158 cfs_rq->tasks_timeline = RB_ROOT;
9159 INIT_LIST_HEAD(&cfs_rq->tasks);
9160 #ifdef CONFIG_FAIR_GROUP_SCHED
9163 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9166 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9168 struct rt_prio_array *array;
9171 array = &rt_rq->active;
9172 for (i = 0; i < MAX_RT_PRIO; i++) {
9173 INIT_LIST_HEAD(array->queue + i);
9174 __clear_bit(i, array->bitmap);
9176 /* delimiter for bitsearch: */
9177 __set_bit(MAX_RT_PRIO, array->bitmap);
9179 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9180 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9182 rt_rq->highest_prio.next = MAX_RT_PRIO;
9186 rt_rq->rt_nr_migratory = 0;
9187 rt_rq->overloaded = 0;
9188 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9192 rt_rq->rt_throttled = 0;
9193 rt_rq->rt_runtime = 0;
9194 spin_lock_init(&rt_rq->rt_runtime_lock);
9196 #ifdef CONFIG_RT_GROUP_SCHED
9197 rt_rq->rt_nr_boosted = 0;
9202 #ifdef CONFIG_FAIR_GROUP_SCHED
9203 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9204 struct sched_entity *se, int cpu, int add,
9205 struct sched_entity *parent)
9207 struct rq *rq = cpu_rq(cpu);
9208 tg->cfs_rq[cpu] = cfs_rq;
9209 init_cfs_rq(cfs_rq, rq);
9212 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9215 /* se could be NULL for init_task_group */
9220 se->cfs_rq = &rq->cfs;
9222 se->cfs_rq = parent->my_q;
9225 se->load.weight = tg->shares;
9226 se->load.inv_weight = 0;
9227 se->parent = parent;
9231 #ifdef CONFIG_RT_GROUP_SCHED
9232 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9233 struct sched_rt_entity *rt_se, int cpu, int add,
9234 struct sched_rt_entity *parent)
9236 struct rq *rq = cpu_rq(cpu);
9238 tg->rt_rq[cpu] = rt_rq;
9239 init_rt_rq(rt_rq, rq);
9241 rt_rq->rt_se = rt_se;
9242 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9244 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9246 tg->rt_se[cpu] = rt_se;
9251 rt_se->rt_rq = &rq->rt;
9253 rt_se->rt_rq = parent->my_q;
9255 rt_se->my_q = rt_rq;
9256 rt_se->parent = parent;
9257 INIT_LIST_HEAD(&rt_se->run_list);
9261 void __init sched_init(void)
9264 unsigned long alloc_size = 0, ptr;
9266 #ifdef CONFIG_FAIR_GROUP_SCHED
9267 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9269 #ifdef CONFIG_RT_GROUP_SCHED
9270 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9272 #ifdef CONFIG_USER_SCHED
9275 #ifdef CONFIG_CPUMASK_OFFSTACK
9276 alloc_size += num_possible_cpus() * cpumask_size();
9279 * As sched_init() is called before page_alloc is setup,
9280 * we use alloc_bootmem().
9283 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9285 #ifdef CONFIG_FAIR_GROUP_SCHED
9286 init_task_group.se = (struct sched_entity **)ptr;
9287 ptr += nr_cpu_ids * sizeof(void **);
9289 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9290 ptr += nr_cpu_ids * sizeof(void **);
9292 #ifdef CONFIG_USER_SCHED
9293 root_task_group.se = (struct sched_entity **)ptr;
9294 ptr += nr_cpu_ids * sizeof(void **);
9296 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9297 ptr += nr_cpu_ids * sizeof(void **);
9298 #endif /* CONFIG_USER_SCHED */
9299 #endif /* CONFIG_FAIR_GROUP_SCHED */
9300 #ifdef CONFIG_RT_GROUP_SCHED
9301 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9302 ptr += nr_cpu_ids * sizeof(void **);
9304 init_task_group.rt_rq = (struct rt_rq **)ptr;
9305 ptr += nr_cpu_ids * sizeof(void **);
9307 #ifdef CONFIG_USER_SCHED
9308 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9309 ptr += nr_cpu_ids * sizeof(void **);
9311 root_task_group.rt_rq = (struct rt_rq **)ptr;
9312 ptr += nr_cpu_ids * sizeof(void **);
9313 #endif /* CONFIG_USER_SCHED */
9314 #endif /* CONFIG_RT_GROUP_SCHED */
9315 #ifdef CONFIG_CPUMASK_OFFSTACK
9316 for_each_possible_cpu(i) {
9317 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9318 ptr += cpumask_size();
9320 #endif /* CONFIG_CPUMASK_OFFSTACK */
9324 init_defrootdomain();
9327 init_rt_bandwidth(&def_rt_bandwidth,
9328 global_rt_period(), global_rt_runtime());
9330 #ifdef CONFIG_RT_GROUP_SCHED
9331 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9332 global_rt_period(), global_rt_runtime());
9333 #ifdef CONFIG_USER_SCHED
9334 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9335 global_rt_period(), RUNTIME_INF);
9336 #endif /* CONFIG_USER_SCHED */
9337 #endif /* CONFIG_RT_GROUP_SCHED */
9339 #ifdef CONFIG_GROUP_SCHED
9340 list_add(&init_task_group.list, &task_groups);
9341 INIT_LIST_HEAD(&init_task_group.children);
9343 #ifdef CONFIG_USER_SCHED
9344 INIT_LIST_HEAD(&root_task_group.children);
9345 init_task_group.parent = &root_task_group;
9346 list_add(&init_task_group.siblings, &root_task_group.children);
9347 #endif /* CONFIG_USER_SCHED */
9348 #endif /* CONFIG_GROUP_SCHED */
9350 for_each_possible_cpu(i) {
9354 spin_lock_init(&rq->lock);
9356 rq->calc_load_active = 0;
9357 rq->calc_load_update = jiffies + LOAD_FREQ;
9358 init_cfs_rq(&rq->cfs, rq);
9359 init_rt_rq(&rq->rt, rq);
9360 #ifdef CONFIG_FAIR_GROUP_SCHED
9361 init_task_group.shares = init_task_group_load;
9362 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9363 #ifdef CONFIG_CGROUP_SCHED
9365 * How much cpu bandwidth does init_task_group get?
9367 * In case of task-groups formed thr' the cgroup filesystem, it
9368 * gets 100% of the cpu resources in the system. This overall
9369 * system cpu resource is divided among the tasks of
9370 * init_task_group and its child task-groups in a fair manner,
9371 * based on each entity's (task or task-group's) weight
9372 * (se->load.weight).
9374 * In other words, if init_task_group has 10 tasks of weight
9375 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9376 * then A0's share of the cpu resource is:
9378 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9380 * We achieve this by letting init_task_group's tasks sit
9381 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9383 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9384 #elif defined CONFIG_USER_SCHED
9385 root_task_group.shares = NICE_0_LOAD;
9386 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9388 * In case of task-groups formed thr' the user id of tasks,
9389 * init_task_group represents tasks belonging to root user.
9390 * Hence it forms a sibling of all subsequent groups formed.
9391 * In this case, init_task_group gets only a fraction of overall
9392 * system cpu resource, based on the weight assigned to root
9393 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9394 * by letting tasks of init_task_group sit in a separate cfs_rq
9395 * (init_cfs_rq) and having one entity represent this group of
9396 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9398 init_tg_cfs_entry(&init_task_group,
9399 &per_cpu(init_cfs_rq, i),
9400 &per_cpu(init_sched_entity, i), i, 1,
9401 root_task_group.se[i]);
9404 #endif /* CONFIG_FAIR_GROUP_SCHED */
9406 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9407 #ifdef CONFIG_RT_GROUP_SCHED
9408 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9409 #ifdef CONFIG_CGROUP_SCHED
9410 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9411 #elif defined CONFIG_USER_SCHED
9412 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9413 init_tg_rt_entry(&init_task_group,
9414 &per_cpu(init_rt_rq, i),
9415 &per_cpu(init_sched_rt_entity, i), i, 1,
9416 root_task_group.rt_se[i]);
9420 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9421 rq->cpu_load[j] = 0;
9425 rq->post_schedule = 0;
9426 rq->active_balance = 0;
9427 rq->next_balance = jiffies;
9431 rq->migration_thread = NULL;
9432 INIT_LIST_HEAD(&rq->migration_queue);
9433 rq_attach_root(rq, &def_root_domain);
9436 atomic_set(&rq->nr_iowait, 0);
9439 set_load_weight(&init_task);
9441 #ifdef CONFIG_PREEMPT_NOTIFIERS
9442 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9446 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9449 #ifdef CONFIG_RT_MUTEXES
9450 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9454 * The boot idle thread does lazy MMU switching as well:
9456 atomic_inc(&init_mm.mm_count);
9457 enter_lazy_tlb(&init_mm, current);
9460 * Make us the idle thread. Technically, schedule() should not be
9461 * called from this thread, however somewhere below it might be,
9462 * but because we are the idle thread, we just pick up running again
9463 * when this runqueue becomes "idle".
9465 init_idle(current, smp_processor_id());
9467 calc_load_update = jiffies + LOAD_FREQ;
9470 * During early bootup we pretend to be a normal task:
9472 current->sched_class = &fair_sched_class;
9474 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9475 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9478 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9479 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9481 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9484 perf_counter_init();
9486 scheduler_running = 1;
9489 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9490 static inline int preempt_count_equals(int preempt_offset)
9492 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9494 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9497 void __might_sleep(char *file, int line, int preempt_offset)
9500 static unsigned long prev_jiffy; /* ratelimiting */
9502 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9503 system_state != SYSTEM_RUNNING || oops_in_progress)
9505 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9507 prev_jiffy = jiffies;
9510 "BUG: sleeping function called from invalid context at %s:%d\n",
9513 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9514 in_atomic(), irqs_disabled(),
9515 current->pid, current->comm);
9517 debug_show_held_locks(current);
9518 if (irqs_disabled())
9519 print_irqtrace_events(current);
9523 EXPORT_SYMBOL(__might_sleep);
9526 #ifdef CONFIG_MAGIC_SYSRQ
9527 static void normalize_task(struct rq *rq, struct task_struct *p)
9531 update_rq_clock(rq);
9532 on_rq = p->se.on_rq;
9534 deactivate_task(rq, p, 0);
9535 __setscheduler(rq, p, SCHED_NORMAL, 0);
9537 activate_task(rq, p, 0);
9538 resched_task(rq->curr);
9542 void normalize_rt_tasks(void)
9544 struct task_struct *g, *p;
9545 unsigned long flags;
9548 read_lock_irqsave(&tasklist_lock, flags);
9549 do_each_thread(g, p) {
9551 * Only normalize user tasks:
9556 p->se.exec_start = 0;
9557 #ifdef CONFIG_SCHEDSTATS
9558 p->se.wait_start = 0;
9559 p->se.sleep_start = 0;
9560 p->se.block_start = 0;
9565 * Renice negative nice level userspace
9568 if (TASK_NICE(p) < 0 && p->mm)
9569 set_user_nice(p, 0);
9573 spin_lock(&p->pi_lock);
9574 rq = __task_rq_lock(p);
9576 normalize_task(rq, p);
9578 __task_rq_unlock(rq);
9579 spin_unlock(&p->pi_lock);
9580 } while_each_thread(g, p);
9582 read_unlock_irqrestore(&tasklist_lock, flags);
9585 #endif /* CONFIG_MAGIC_SYSRQ */
9589 * These functions are only useful for the IA64 MCA handling.
9591 * They can only be called when the whole system has been
9592 * stopped - every CPU needs to be quiescent, and no scheduling
9593 * activity can take place. Using them for anything else would
9594 * be a serious bug, and as a result, they aren't even visible
9595 * under any other configuration.
9599 * curr_task - return the current task for a given cpu.
9600 * @cpu: the processor in question.
9602 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9604 struct task_struct *curr_task(int cpu)
9606 return cpu_curr(cpu);
9610 * set_curr_task - set the current task for a given cpu.
9611 * @cpu: the processor in question.
9612 * @p: the task pointer to set.
9614 * Description: This function must only be used when non-maskable interrupts
9615 * are serviced on a separate stack. It allows the architecture to switch the
9616 * notion of the current task on a cpu in a non-blocking manner. This function
9617 * must be called with all CPU's synchronized, and interrupts disabled, the
9618 * and caller must save the original value of the current task (see
9619 * curr_task() above) and restore that value before reenabling interrupts and
9620 * re-starting the system.
9622 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9624 void set_curr_task(int cpu, struct task_struct *p)
9631 #ifdef CONFIG_FAIR_GROUP_SCHED
9632 static void free_fair_sched_group(struct task_group *tg)
9636 for_each_possible_cpu(i) {
9638 kfree(tg->cfs_rq[i]);
9648 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9650 struct cfs_rq *cfs_rq;
9651 struct sched_entity *se;
9655 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9658 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9662 tg->shares = NICE_0_LOAD;
9664 for_each_possible_cpu(i) {
9667 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9668 GFP_KERNEL, cpu_to_node(i));
9672 se = kzalloc_node(sizeof(struct sched_entity),
9673 GFP_KERNEL, cpu_to_node(i));
9677 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9686 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9688 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9689 &cpu_rq(cpu)->leaf_cfs_rq_list);
9692 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9694 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9696 #else /* !CONFG_FAIR_GROUP_SCHED */
9697 static inline void free_fair_sched_group(struct task_group *tg)
9702 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9707 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9711 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9714 #endif /* CONFIG_FAIR_GROUP_SCHED */
9716 #ifdef CONFIG_RT_GROUP_SCHED
9717 static void free_rt_sched_group(struct task_group *tg)
9721 destroy_rt_bandwidth(&tg->rt_bandwidth);
9723 for_each_possible_cpu(i) {
9725 kfree(tg->rt_rq[i]);
9727 kfree(tg->rt_se[i]);
9735 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9737 struct rt_rq *rt_rq;
9738 struct sched_rt_entity *rt_se;
9742 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9745 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9749 init_rt_bandwidth(&tg->rt_bandwidth,
9750 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9752 for_each_possible_cpu(i) {
9755 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9756 GFP_KERNEL, cpu_to_node(i));
9760 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9761 GFP_KERNEL, cpu_to_node(i));
9765 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9774 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9776 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9777 &cpu_rq(cpu)->leaf_rt_rq_list);
9780 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9782 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9784 #else /* !CONFIG_RT_GROUP_SCHED */
9785 static inline void free_rt_sched_group(struct task_group *tg)
9790 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9795 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9799 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9802 #endif /* CONFIG_RT_GROUP_SCHED */
9804 #ifdef CONFIG_GROUP_SCHED
9805 static void free_sched_group(struct task_group *tg)
9807 free_fair_sched_group(tg);
9808 free_rt_sched_group(tg);
9812 /* allocate runqueue etc for a new task group */
9813 struct task_group *sched_create_group(struct task_group *parent)
9815 struct task_group *tg;
9816 unsigned long flags;
9819 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9821 return ERR_PTR(-ENOMEM);
9823 if (!alloc_fair_sched_group(tg, parent))
9826 if (!alloc_rt_sched_group(tg, parent))
9829 spin_lock_irqsave(&task_group_lock, flags);
9830 for_each_possible_cpu(i) {
9831 register_fair_sched_group(tg, i);
9832 register_rt_sched_group(tg, i);
9834 list_add_rcu(&tg->list, &task_groups);
9836 WARN_ON(!parent); /* root should already exist */
9838 tg->parent = parent;
9839 INIT_LIST_HEAD(&tg->children);
9840 list_add_rcu(&tg->siblings, &parent->children);
9841 spin_unlock_irqrestore(&task_group_lock, flags);
9846 free_sched_group(tg);
9847 return ERR_PTR(-ENOMEM);
9850 /* rcu callback to free various structures associated with a task group */
9851 static void free_sched_group_rcu(struct rcu_head *rhp)
9853 /* now it should be safe to free those cfs_rqs */
9854 free_sched_group(container_of(rhp, struct task_group, rcu));
9857 /* Destroy runqueue etc associated with a task group */
9858 void sched_destroy_group(struct task_group *tg)
9860 unsigned long flags;
9863 spin_lock_irqsave(&task_group_lock, flags);
9864 for_each_possible_cpu(i) {
9865 unregister_fair_sched_group(tg, i);
9866 unregister_rt_sched_group(tg, i);
9868 list_del_rcu(&tg->list);
9869 list_del_rcu(&tg->siblings);
9870 spin_unlock_irqrestore(&task_group_lock, flags);
9872 /* wait for possible concurrent references to cfs_rqs complete */
9873 call_rcu(&tg->rcu, free_sched_group_rcu);
9876 /* change task's runqueue when it moves between groups.
9877 * The caller of this function should have put the task in its new group
9878 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9879 * reflect its new group.
9881 void sched_move_task(struct task_struct *tsk)
9884 unsigned long flags;
9887 rq = task_rq_lock(tsk, &flags);
9889 update_rq_clock(rq);
9891 running = task_current(rq, tsk);
9892 on_rq = tsk->se.on_rq;
9895 dequeue_task(rq, tsk, 0);
9896 if (unlikely(running))
9897 tsk->sched_class->put_prev_task(rq, tsk);
9899 set_task_rq(tsk, task_cpu(tsk));
9901 #ifdef CONFIG_FAIR_GROUP_SCHED
9902 if (tsk->sched_class->moved_group)
9903 tsk->sched_class->moved_group(tsk);
9906 if (unlikely(running))
9907 tsk->sched_class->set_curr_task(rq);
9909 enqueue_task(rq, tsk, 0);
9911 task_rq_unlock(rq, &flags);
9913 #endif /* CONFIG_GROUP_SCHED */
9915 #ifdef CONFIG_FAIR_GROUP_SCHED
9916 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9918 struct cfs_rq *cfs_rq = se->cfs_rq;
9923 dequeue_entity(cfs_rq, se, 0);
9925 se->load.weight = shares;
9926 se->load.inv_weight = 0;
9929 enqueue_entity(cfs_rq, se, 0);
9932 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9934 struct cfs_rq *cfs_rq = se->cfs_rq;
9935 struct rq *rq = cfs_rq->rq;
9936 unsigned long flags;
9938 spin_lock_irqsave(&rq->lock, flags);
9939 __set_se_shares(se, shares);
9940 spin_unlock_irqrestore(&rq->lock, flags);
9943 static DEFINE_MUTEX(shares_mutex);
9945 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9948 unsigned long flags;
9951 * We can't change the weight of the root cgroup.
9956 if (shares < MIN_SHARES)
9957 shares = MIN_SHARES;
9958 else if (shares > MAX_SHARES)
9959 shares = MAX_SHARES;
9961 mutex_lock(&shares_mutex);
9962 if (tg->shares == shares)
9965 spin_lock_irqsave(&task_group_lock, flags);
9966 for_each_possible_cpu(i)
9967 unregister_fair_sched_group(tg, i);
9968 list_del_rcu(&tg->siblings);
9969 spin_unlock_irqrestore(&task_group_lock, flags);
9971 /* wait for any ongoing reference to this group to finish */
9972 synchronize_sched();
9975 * Now we are free to modify the group's share on each cpu
9976 * w/o tripping rebalance_share or load_balance_fair.
9978 tg->shares = shares;
9979 for_each_possible_cpu(i) {
9983 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9984 set_se_shares(tg->se[i], shares);
9988 * Enable load balance activity on this group, by inserting it back on
9989 * each cpu's rq->leaf_cfs_rq_list.
9991 spin_lock_irqsave(&task_group_lock, flags);
9992 for_each_possible_cpu(i)
9993 register_fair_sched_group(tg, i);
9994 list_add_rcu(&tg->siblings, &tg->parent->children);
9995 spin_unlock_irqrestore(&task_group_lock, flags);
9997 mutex_unlock(&shares_mutex);
10001 unsigned long sched_group_shares(struct task_group *tg)
10007 #ifdef CONFIG_RT_GROUP_SCHED
10009 * Ensure that the real time constraints are schedulable.
10011 static DEFINE_MUTEX(rt_constraints_mutex);
10013 static unsigned long to_ratio(u64 period, u64 runtime)
10015 if (runtime == RUNTIME_INF)
10018 return div64_u64(runtime << 20, period);
10021 /* Must be called with tasklist_lock held */
10022 static inline int tg_has_rt_tasks(struct task_group *tg)
10024 struct task_struct *g, *p;
10026 do_each_thread(g, p) {
10027 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10029 } while_each_thread(g, p);
10034 struct rt_schedulable_data {
10035 struct task_group *tg;
10040 static int tg_schedulable(struct task_group *tg, void *data)
10042 struct rt_schedulable_data *d = data;
10043 struct task_group *child;
10044 unsigned long total, sum = 0;
10045 u64 period, runtime;
10047 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10048 runtime = tg->rt_bandwidth.rt_runtime;
10051 period = d->rt_period;
10052 runtime = d->rt_runtime;
10055 #ifdef CONFIG_USER_SCHED
10056 if (tg == &root_task_group) {
10057 period = global_rt_period();
10058 runtime = global_rt_runtime();
10063 * Cannot have more runtime than the period.
10065 if (runtime > period && runtime != RUNTIME_INF)
10069 * Ensure we don't starve existing RT tasks.
10071 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10074 total = to_ratio(period, runtime);
10077 * Nobody can have more than the global setting allows.
10079 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10083 * The sum of our children's runtime should not exceed our own.
10085 list_for_each_entry_rcu(child, &tg->children, siblings) {
10086 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10087 runtime = child->rt_bandwidth.rt_runtime;
10089 if (child == d->tg) {
10090 period = d->rt_period;
10091 runtime = d->rt_runtime;
10094 sum += to_ratio(period, runtime);
10103 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10105 struct rt_schedulable_data data = {
10107 .rt_period = period,
10108 .rt_runtime = runtime,
10111 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10114 static int tg_set_bandwidth(struct task_group *tg,
10115 u64 rt_period, u64 rt_runtime)
10119 mutex_lock(&rt_constraints_mutex);
10120 read_lock(&tasklist_lock);
10121 err = __rt_schedulable(tg, rt_period, rt_runtime);
10125 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10126 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10127 tg->rt_bandwidth.rt_runtime = rt_runtime;
10129 for_each_possible_cpu(i) {
10130 struct rt_rq *rt_rq = tg->rt_rq[i];
10132 spin_lock(&rt_rq->rt_runtime_lock);
10133 rt_rq->rt_runtime = rt_runtime;
10134 spin_unlock(&rt_rq->rt_runtime_lock);
10136 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10138 read_unlock(&tasklist_lock);
10139 mutex_unlock(&rt_constraints_mutex);
10144 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10146 u64 rt_runtime, rt_period;
10148 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10149 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10150 if (rt_runtime_us < 0)
10151 rt_runtime = RUNTIME_INF;
10153 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10156 long sched_group_rt_runtime(struct task_group *tg)
10160 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10163 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10164 do_div(rt_runtime_us, NSEC_PER_USEC);
10165 return rt_runtime_us;
10168 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10170 u64 rt_runtime, rt_period;
10172 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10173 rt_runtime = tg->rt_bandwidth.rt_runtime;
10175 if (rt_period == 0)
10178 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10181 long sched_group_rt_period(struct task_group *tg)
10185 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10186 do_div(rt_period_us, NSEC_PER_USEC);
10187 return rt_period_us;
10190 static int sched_rt_global_constraints(void)
10192 u64 runtime, period;
10195 if (sysctl_sched_rt_period <= 0)
10198 runtime = global_rt_runtime();
10199 period = global_rt_period();
10202 * Sanity check on the sysctl variables.
10204 if (runtime > period && runtime != RUNTIME_INF)
10207 mutex_lock(&rt_constraints_mutex);
10208 read_lock(&tasklist_lock);
10209 ret = __rt_schedulable(NULL, 0, 0);
10210 read_unlock(&tasklist_lock);
10211 mutex_unlock(&rt_constraints_mutex);
10216 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10218 /* Don't accept realtime tasks when there is no way for them to run */
10219 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10225 #else /* !CONFIG_RT_GROUP_SCHED */
10226 static int sched_rt_global_constraints(void)
10228 unsigned long flags;
10231 if (sysctl_sched_rt_period <= 0)
10235 * There's always some RT tasks in the root group
10236 * -- migration, kstopmachine etc..
10238 if (sysctl_sched_rt_runtime == 0)
10241 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10242 for_each_possible_cpu(i) {
10243 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10245 spin_lock(&rt_rq->rt_runtime_lock);
10246 rt_rq->rt_runtime = global_rt_runtime();
10247 spin_unlock(&rt_rq->rt_runtime_lock);
10249 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10253 #endif /* CONFIG_RT_GROUP_SCHED */
10255 int sched_rt_handler(struct ctl_table *table, int write,
10256 struct file *filp, void __user *buffer, size_t *lenp,
10260 int old_period, old_runtime;
10261 static DEFINE_MUTEX(mutex);
10263 mutex_lock(&mutex);
10264 old_period = sysctl_sched_rt_period;
10265 old_runtime = sysctl_sched_rt_runtime;
10267 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10269 if (!ret && write) {
10270 ret = sched_rt_global_constraints();
10272 sysctl_sched_rt_period = old_period;
10273 sysctl_sched_rt_runtime = old_runtime;
10275 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10276 def_rt_bandwidth.rt_period =
10277 ns_to_ktime(global_rt_period());
10280 mutex_unlock(&mutex);
10285 #ifdef CONFIG_CGROUP_SCHED
10287 /* return corresponding task_group object of a cgroup */
10288 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10290 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10291 struct task_group, css);
10294 static struct cgroup_subsys_state *
10295 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10297 struct task_group *tg, *parent;
10299 if (!cgrp->parent) {
10300 /* This is early initialization for the top cgroup */
10301 return &init_task_group.css;
10304 parent = cgroup_tg(cgrp->parent);
10305 tg = sched_create_group(parent);
10307 return ERR_PTR(-ENOMEM);
10313 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10315 struct task_group *tg = cgroup_tg(cgrp);
10317 sched_destroy_group(tg);
10321 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10322 struct task_struct *tsk)
10324 #ifdef CONFIG_RT_GROUP_SCHED
10325 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10328 /* We don't support RT-tasks being in separate groups */
10329 if (tsk->sched_class != &fair_sched_class)
10337 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10338 struct cgroup *old_cont, struct task_struct *tsk)
10340 sched_move_task(tsk);
10343 #ifdef CONFIG_FAIR_GROUP_SCHED
10344 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10347 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10350 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10352 struct task_group *tg = cgroup_tg(cgrp);
10354 return (u64) tg->shares;
10356 #endif /* CONFIG_FAIR_GROUP_SCHED */
10358 #ifdef CONFIG_RT_GROUP_SCHED
10359 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10362 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10365 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10367 return sched_group_rt_runtime(cgroup_tg(cgrp));
10370 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10373 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10376 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10378 return sched_group_rt_period(cgroup_tg(cgrp));
10380 #endif /* CONFIG_RT_GROUP_SCHED */
10382 static struct cftype cpu_files[] = {
10383 #ifdef CONFIG_FAIR_GROUP_SCHED
10386 .read_u64 = cpu_shares_read_u64,
10387 .write_u64 = cpu_shares_write_u64,
10390 #ifdef CONFIG_RT_GROUP_SCHED
10392 .name = "rt_runtime_us",
10393 .read_s64 = cpu_rt_runtime_read,
10394 .write_s64 = cpu_rt_runtime_write,
10397 .name = "rt_period_us",
10398 .read_u64 = cpu_rt_period_read_uint,
10399 .write_u64 = cpu_rt_period_write_uint,
10404 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10406 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10409 struct cgroup_subsys cpu_cgroup_subsys = {
10411 .create = cpu_cgroup_create,
10412 .destroy = cpu_cgroup_destroy,
10413 .can_attach = cpu_cgroup_can_attach,
10414 .attach = cpu_cgroup_attach,
10415 .populate = cpu_cgroup_populate,
10416 .subsys_id = cpu_cgroup_subsys_id,
10420 #endif /* CONFIG_CGROUP_SCHED */
10422 #ifdef CONFIG_CGROUP_CPUACCT
10425 * CPU accounting code for task groups.
10427 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10428 * (balbir@in.ibm.com).
10431 /* track cpu usage of a group of tasks and its child groups */
10433 struct cgroup_subsys_state css;
10434 /* cpuusage holds pointer to a u64-type object on every cpu */
10436 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10437 struct cpuacct *parent;
10440 struct cgroup_subsys cpuacct_subsys;
10442 /* return cpu accounting group corresponding to this container */
10443 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10445 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10446 struct cpuacct, css);
10449 /* return cpu accounting group to which this task belongs */
10450 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10452 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10453 struct cpuacct, css);
10456 /* create a new cpu accounting group */
10457 static struct cgroup_subsys_state *cpuacct_create(
10458 struct cgroup_subsys *ss, struct cgroup *cgrp)
10460 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10466 ca->cpuusage = alloc_percpu(u64);
10470 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10471 if (percpu_counter_init(&ca->cpustat[i], 0))
10472 goto out_free_counters;
10475 ca->parent = cgroup_ca(cgrp->parent);
10481 percpu_counter_destroy(&ca->cpustat[i]);
10482 free_percpu(ca->cpuusage);
10486 return ERR_PTR(-ENOMEM);
10489 /* destroy an existing cpu accounting group */
10491 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10493 struct cpuacct *ca = cgroup_ca(cgrp);
10496 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10497 percpu_counter_destroy(&ca->cpustat[i]);
10498 free_percpu(ca->cpuusage);
10502 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10504 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10507 #ifndef CONFIG_64BIT
10509 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10511 spin_lock_irq(&cpu_rq(cpu)->lock);
10513 spin_unlock_irq(&cpu_rq(cpu)->lock);
10521 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10523 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10525 #ifndef CONFIG_64BIT
10527 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10529 spin_lock_irq(&cpu_rq(cpu)->lock);
10531 spin_unlock_irq(&cpu_rq(cpu)->lock);
10537 /* return total cpu usage (in nanoseconds) of a group */
10538 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10540 struct cpuacct *ca = cgroup_ca(cgrp);
10541 u64 totalcpuusage = 0;
10544 for_each_present_cpu(i)
10545 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10547 return totalcpuusage;
10550 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10553 struct cpuacct *ca = cgroup_ca(cgrp);
10562 for_each_present_cpu(i)
10563 cpuacct_cpuusage_write(ca, i, 0);
10569 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10570 struct seq_file *m)
10572 struct cpuacct *ca = cgroup_ca(cgroup);
10576 for_each_present_cpu(i) {
10577 percpu = cpuacct_cpuusage_read(ca, i);
10578 seq_printf(m, "%llu ", (unsigned long long) percpu);
10580 seq_printf(m, "\n");
10584 static const char *cpuacct_stat_desc[] = {
10585 [CPUACCT_STAT_USER] = "user",
10586 [CPUACCT_STAT_SYSTEM] = "system",
10589 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10590 struct cgroup_map_cb *cb)
10592 struct cpuacct *ca = cgroup_ca(cgrp);
10595 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10596 s64 val = percpu_counter_read(&ca->cpustat[i]);
10597 val = cputime64_to_clock_t(val);
10598 cb->fill(cb, cpuacct_stat_desc[i], val);
10603 static struct cftype files[] = {
10606 .read_u64 = cpuusage_read,
10607 .write_u64 = cpuusage_write,
10610 .name = "usage_percpu",
10611 .read_seq_string = cpuacct_percpu_seq_read,
10615 .read_map = cpuacct_stats_show,
10619 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10621 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10625 * charge this task's execution time to its accounting group.
10627 * called with rq->lock held.
10629 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10631 struct cpuacct *ca;
10634 if (unlikely(!cpuacct_subsys.active))
10637 cpu = task_cpu(tsk);
10643 for (; ca; ca = ca->parent) {
10644 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10645 *cpuusage += cputime;
10652 * Charge the system/user time to the task's accounting group.
10654 static void cpuacct_update_stats(struct task_struct *tsk,
10655 enum cpuacct_stat_index idx, cputime_t val)
10657 struct cpuacct *ca;
10659 if (unlikely(!cpuacct_subsys.active))
10666 percpu_counter_add(&ca->cpustat[idx], val);
10672 struct cgroup_subsys cpuacct_subsys = {
10674 .create = cpuacct_create,
10675 .destroy = cpuacct_destroy,
10676 .populate = cpuacct_populate,
10677 .subsys_id = cpuacct_subsys_id,
10679 #endif /* CONFIG_CGROUP_CPUACCT */