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)
1527 unsigned long rq_weight;
1528 unsigned long shares;
1534 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1537 rq_weight = NICE_0_LOAD;
1541 * \Sum shares * rq_weight
1542 * shares = -----------------------
1546 shares = (sd_shares * rq_weight) / sd_rq_weight;
1547 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1549 if (abs(shares - tg->se[cpu]->load.weight) >
1550 sysctl_sched_shares_thresh) {
1551 struct rq *rq = cpu_rq(cpu);
1552 unsigned long flags;
1554 spin_lock_irqsave(&rq->lock, flags);
1555 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1556 __set_se_shares(tg->se[cpu], shares);
1557 spin_unlock_irqrestore(&rq->lock, flags);
1562 * Re-compute the task group their per cpu shares over the given domain.
1563 * This needs to be done in a bottom-up fashion because the rq weight of a
1564 * parent group depends on the shares of its child groups.
1566 static int tg_shares_up(struct task_group *tg, void *data)
1568 unsigned long weight, rq_weight = 0, eff_weight = 0;
1569 unsigned long shares = 0;
1570 struct sched_domain *sd = data;
1573 for_each_cpu(i, sched_domain_span(sd)) {
1575 * If there are currently no tasks on the cpu pretend there
1576 * is one of average load so that when a new task gets to
1577 * run here it will not get delayed by group starvation.
1579 weight = tg->cfs_rq[i]->load.weight;
1580 tg->cfs_rq[i]->rq_weight = weight;
1581 rq_weight += weight;
1584 weight = NICE_0_LOAD;
1586 eff_weight += weight;
1587 shares += tg->cfs_rq[i]->shares;
1590 if ((!shares && rq_weight) || shares > tg->shares)
1591 shares = tg->shares;
1593 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1594 shares = tg->shares;
1596 for_each_cpu(i, sched_domain_span(sd)) {
1597 unsigned long sd_rq_weight = rq_weight;
1599 if (!tg->cfs_rq[i]->rq_weight)
1600 sd_rq_weight = eff_weight;
1602 update_group_shares_cpu(tg, i, shares, sd_rq_weight);
1609 * Compute the cpu's hierarchical load factor for each task group.
1610 * This needs to be done in a top-down fashion because the load of a child
1611 * group is a fraction of its parents load.
1613 static int tg_load_down(struct task_group *tg, void *data)
1616 long cpu = (long)data;
1619 load = cpu_rq(cpu)->load.weight;
1621 load = tg->parent->cfs_rq[cpu]->h_load;
1622 load *= tg->cfs_rq[cpu]->shares;
1623 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1626 tg->cfs_rq[cpu]->h_load = load;
1631 static void update_shares(struct sched_domain *sd)
1636 if (root_task_group_empty())
1639 now = cpu_clock(raw_smp_processor_id());
1640 elapsed = now - sd->last_update;
1642 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1643 sd->last_update = now;
1644 walk_tg_tree(tg_nop, tg_shares_up, sd);
1648 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1650 if (root_task_group_empty())
1653 spin_unlock(&rq->lock);
1655 spin_lock(&rq->lock);
1658 static void update_h_load(long cpu)
1660 if (root_task_group_empty())
1663 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1668 static inline void update_shares(struct sched_domain *sd)
1672 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1678 #ifdef CONFIG_PREEMPT
1681 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1682 * way at the expense of forcing extra atomic operations in all
1683 * invocations. This assures that the double_lock is acquired using the
1684 * same underlying policy as the spinlock_t on this architecture, which
1685 * reduces latency compared to the unfair variant below. However, it
1686 * also adds more overhead and therefore may reduce throughput.
1688 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1689 __releases(this_rq->lock)
1690 __acquires(busiest->lock)
1691 __acquires(this_rq->lock)
1693 spin_unlock(&this_rq->lock);
1694 double_rq_lock(this_rq, busiest);
1701 * Unfair double_lock_balance: Optimizes throughput at the expense of
1702 * latency by eliminating extra atomic operations when the locks are
1703 * already in proper order on entry. This favors lower cpu-ids and will
1704 * grant the double lock to lower cpus over higher ids under contention,
1705 * regardless of entry order into the function.
1707 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1708 __releases(this_rq->lock)
1709 __acquires(busiest->lock)
1710 __acquires(this_rq->lock)
1714 if (unlikely(!spin_trylock(&busiest->lock))) {
1715 if (busiest < this_rq) {
1716 spin_unlock(&this_rq->lock);
1717 spin_lock(&busiest->lock);
1718 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1721 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1726 #endif /* CONFIG_PREEMPT */
1729 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1731 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1733 if (unlikely(!irqs_disabled())) {
1734 /* printk() doesn't work good under rq->lock */
1735 spin_unlock(&this_rq->lock);
1739 return _double_lock_balance(this_rq, busiest);
1742 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1743 __releases(busiest->lock)
1745 spin_unlock(&busiest->lock);
1746 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1750 #ifdef CONFIG_FAIR_GROUP_SCHED
1751 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1754 cfs_rq->shares = shares;
1759 static void calc_load_account_active(struct rq *this_rq);
1761 #include "sched_stats.h"
1762 #include "sched_idletask.c"
1763 #include "sched_fair.c"
1764 #include "sched_rt.c"
1765 #ifdef CONFIG_SCHED_DEBUG
1766 # include "sched_debug.c"
1769 #define sched_class_highest (&rt_sched_class)
1770 #define for_each_class(class) \
1771 for (class = sched_class_highest; class; class = class->next)
1773 static void inc_nr_running(struct rq *rq)
1778 static void dec_nr_running(struct rq *rq)
1783 static void set_load_weight(struct task_struct *p)
1785 if (task_has_rt_policy(p)) {
1786 p->se.load.weight = prio_to_weight[0] * 2;
1787 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1792 * SCHED_IDLE tasks get minimal weight:
1794 if (p->policy == SCHED_IDLE) {
1795 p->se.load.weight = WEIGHT_IDLEPRIO;
1796 p->se.load.inv_weight = WMULT_IDLEPRIO;
1800 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1801 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1804 static void update_avg(u64 *avg, u64 sample)
1806 s64 diff = sample - *avg;
1810 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1813 p->se.start_runtime = p->se.sum_exec_runtime;
1815 sched_info_queued(p);
1816 p->sched_class->enqueue_task(rq, p, wakeup);
1820 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1823 if (p->se.last_wakeup) {
1824 update_avg(&p->se.avg_overlap,
1825 p->se.sum_exec_runtime - p->se.last_wakeup);
1826 p->se.last_wakeup = 0;
1828 update_avg(&p->se.avg_wakeup,
1829 sysctl_sched_wakeup_granularity);
1833 sched_info_dequeued(p);
1834 p->sched_class->dequeue_task(rq, p, sleep);
1839 * __normal_prio - return the priority that is based on the static prio
1841 static inline int __normal_prio(struct task_struct *p)
1843 return p->static_prio;
1847 * Calculate the expected normal priority: i.e. priority
1848 * without taking RT-inheritance into account. Might be
1849 * boosted by interactivity modifiers. Changes upon fork,
1850 * setprio syscalls, and whenever the interactivity
1851 * estimator recalculates.
1853 static inline int normal_prio(struct task_struct *p)
1857 if (task_has_rt_policy(p))
1858 prio = MAX_RT_PRIO-1 - p->rt_priority;
1860 prio = __normal_prio(p);
1865 * Calculate the current priority, i.e. the priority
1866 * taken into account by the scheduler. This value might
1867 * be boosted by RT tasks, or might be boosted by
1868 * interactivity modifiers. Will be RT if the task got
1869 * RT-boosted. If not then it returns p->normal_prio.
1871 static int effective_prio(struct task_struct *p)
1873 p->normal_prio = normal_prio(p);
1875 * If we are RT tasks or we were boosted to RT priority,
1876 * keep the priority unchanged. Otherwise, update priority
1877 * to the normal priority:
1879 if (!rt_prio(p->prio))
1880 return p->normal_prio;
1885 * activate_task - move a task to the runqueue.
1887 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1889 if (task_contributes_to_load(p))
1890 rq->nr_uninterruptible--;
1892 enqueue_task(rq, p, wakeup);
1897 * deactivate_task - remove a task from the runqueue.
1899 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1901 if (task_contributes_to_load(p))
1902 rq->nr_uninterruptible++;
1904 dequeue_task(rq, p, sleep);
1909 * task_curr - is this task currently executing on a CPU?
1910 * @p: the task in question.
1912 inline int task_curr(const struct task_struct *p)
1914 return cpu_curr(task_cpu(p)) == p;
1917 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1919 set_task_rq(p, cpu);
1922 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1923 * successfuly executed on another CPU. We must ensure that updates of
1924 * per-task data have been completed by this moment.
1927 task_thread_info(p)->cpu = cpu;
1931 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1932 const struct sched_class *prev_class,
1933 int oldprio, int running)
1935 if (prev_class != p->sched_class) {
1936 if (prev_class->switched_from)
1937 prev_class->switched_from(rq, p, running);
1938 p->sched_class->switched_to(rq, p, running);
1940 p->sched_class->prio_changed(rq, p, oldprio, running);
1945 /* Used instead of source_load when we know the type == 0 */
1946 static unsigned long weighted_cpuload(const int cpu)
1948 return cpu_rq(cpu)->load.weight;
1952 * Is this task likely cache-hot:
1955 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1960 * Buddy candidates are cache hot:
1962 if (sched_feat(CACHE_HOT_BUDDY) &&
1963 (&p->se == cfs_rq_of(&p->se)->next ||
1964 &p->se == cfs_rq_of(&p->se)->last))
1967 if (p->sched_class != &fair_sched_class)
1970 if (sysctl_sched_migration_cost == -1)
1972 if (sysctl_sched_migration_cost == 0)
1975 delta = now - p->se.exec_start;
1977 return delta < (s64)sysctl_sched_migration_cost;
1981 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1983 int old_cpu = task_cpu(p);
1984 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1985 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1986 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1989 clock_offset = old_rq->clock - new_rq->clock;
1991 trace_sched_migrate_task(p, new_cpu);
1993 #ifdef CONFIG_SCHEDSTATS
1994 if (p->se.wait_start)
1995 p->se.wait_start -= clock_offset;
1996 if (p->se.sleep_start)
1997 p->se.sleep_start -= clock_offset;
1998 if (p->se.block_start)
1999 p->se.block_start -= clock_offset;
2001 if (old_cpu != new_cpu) {
2002 p->se.nr_migrations++;
2003 new_rq->nr_migrations_in++;
2004 #ifdef CONFIG_SCHEDSTATS
2005 if (task_hot(p, old_rq->clock, NULL))
2006 schedstat_inc(p, se.nr_forced2_migrations);
2008 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2011 p->se.vruntime -= old_cfsrq->min_vruntime -
2012 new_cfsrq->min_vruntime;
2014 __set_task_cpu(p, new_cpu);
2017 struct migration_req {
2018 struct list_head list;
2020 struct task_struct *task;
2023 struct completion done;
2027 * The task's runqueue lock must be held.
2028 * Returns true if you have to wait for migration thread.
2031 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2033 struct rq *rq = task_rq(p);
2036 * If the task is not on a runqueue (and not running), then
2037 * it is sufficient to simply update the task's cpu field.
2039 if (!p->se.on_rq && !task_running(rq, p)) {
2040 set_task_cpu(p, dest_cpu);
2044 init_completion(&req->done);
2046 req->dest_cpu = dest_cpu;
2047 list_add(&req->list, &rq->migration_queue);
2053 * wait_task_context_switch - wait for a thread to complete at least one
2056 * @p must not be current.
2058 void wait_task_context_switch(struct task_struct *p)
2060 unsigned long nvcsw, nivcsw, flags;
2068 * The runqueue is assigned before the actual context
2069 * switch. We need to take the runqueue lock.
2071 * We could check initially without the lock but it is
2072 * very likely that we need to take the lock in every
2075 rq = task_rq_lock(p, &flags);
2076 running = task_running(rq, p);
2077 task_rq_unlock(rq, &flags);
2079 if (likely(!running))
2082 * The switch count is incremented before the actual
2083 * context switch. We thus wait for two switches to be
2084 * sure at least one completed.
2086 if ((p->nvcsw - nvcsw) > 1)
2088 if ((p->nivcsw - nivcsw) > 1)
2096 * wait_task_inactive - wait for a thread to unschedule.
2098 * If @match_state is nonzero, it's the @p->state value just checked and
2099 * not expected to change. If it changes, i.e. @p might have woken up,
2100 * then return zero. When we succeed in waiting for @p to be off its CPU,
2101 * we return a positive number (its total switch count). If a second call
2102 * a short while later returns the same number, the caller can be sure that
2103 * @p has remained unscheduled the whole time.
2105 * The caller must ensure that the task *will* unschedule sometime soon,
2106 * else this function might spin for a *long* time. This function can't
2107 * be called with interrupts off, or it may introduce deadlock with
2108 * smp_call_function() if an IPI is sent by the same process we are
2109 * waiting to become inactive.
2111 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2113 unsigned long flags;
2120 * We do the initial early heuristics without holding
2121 * any task-queue locks at all. We'll only try to get
2122 * the runqueue lock when things look like they will
2128 * If the task is actively running on another CPU
2129 * still, just relax and busy-wait without holding
2132 * NOTE! Since we don't hold any locks, it's not
2133 * even sure that "rq" stays as the right runqueue!
2134 * But we don't care, since "task_running()" will
2135 * return false if the runqueue has changed and p
2136 * is actually now running somewhere else!
2138 while (task_running(rq, p)) {
2139 if (match_state && unlikely(p->state != match_state))
2145 * Ok, time to look more closely! We need the rq
2146 * lock now, to be *sure*. If we're wrong, we'll
2147 * just go back and repeat.
2149 rq = task_rq_lock(p, &flags);
2150 trace_sched_wait_task(rq, p);
2151 running = task_running(rq, p);
2152 on_rq = p->se.on_rq;
2154 if (!match_state || p->state == match_state)
2155 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2156 task_rq_unlock(rq, &flags);
2159 * If it changed from the expected state, bail out now.
2161 if (unlikely(!ncsw))
2165 * Was it really running after all now that we
2166 * checked with the proper locks actually held?
2168 * Oops. Go back and try again..
2170 if (unlikely(running)) {
2176 * It's not enough that it's not actively running,
2177 * it must be off the runqueue _entirely_, and not
2180 * So if it was still runnable (but just not actively
2181 * running right now), it's preempted, and we should
2182 * yield - it could be a while.
2184 if (unlikely(on_rq)) {
2185 schedule_timeout_uninterruptible(1);
2190 * Ahh, all good. It wasn't running, and it wasn't
2191 * runnable, which means that it will never become
2192 * running in the future either. We're all done!
2201 * kick_process - kick a running thread to enter/exit the kernel
2202 * @p: the to-be-kicked thread
2204 * Cause a process which is running on another CPU to enter
2205 * kernel-mode, without any delay. (to get signals handled.)
2207 * NOTE: this function doesnt have to take the runqueue lock,
2208 * because all it wants to ensure is that the remote task enters
2209 * the kernel. If the IPI races and the task has been migrated
2210 * to another CPU then no harm is done and the purpose has been
2213 void kick_process(struct task_struct *p)
2219 if ((cpu != smp_processor_id()) && task_curr(p))
2220 smp_send_reschedule(cpu);
2223 EXPORT_SYMBOL_GPL(kick_process);
2226 * Return a low guess at the load of a migration-source cpu weighted
2227 * according to the scheduling class and "nice" value.
2229 * We want to under-estimate the load of migration sources, to
2230 * balance conservatively.
2232 static unsigned long source_load(int cpu, int type)
2234 struct rq *rq = cpu_rq(cpu);
2235 unsigned long total = weighted_cpuload(cpu);
2237 if (type == 0 || !sched_feat(LB_BIAS))
2240 return min(rq->cpu_load[type-1], total);
2244 * Return a high guess at the load of a migration-target cpu weighted
2245 * according to the scheduling class and "nice" value.
2247 static unsigned long target_load(int cpu, int type)
2249 struct rq *rq = cpu_rq(cpu);
2250 unsigned long total = weighted_cpuload(cpu);
2252 if (type == 0 || !sched_feat(LB_BIAS))
2255 return max(rq->cpu_load[type-1], total);
2259 * find_idlest_group finds and returns the least busy CPU group within the
2262 static struct sched_group *
2263 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2265 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2266 unsigned long min_load = ULONG_MAX, this_load = 0;
2267 int load_idx = sd->forkexec_idx;
2268 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2271 unsigned long load, avg_load;
2275 /* Skip over this group if it has no CPUs allowed */
2276 if (!cpumask_intersects(sched_group_cpus(group),
2280 local_group = cpumask_test_cpu(this_cpu,
2281 sched_group_cpus(group));
2283 /* Tally up the load of all CPUs in the group */
2286 for_each_cpu(i, sched_group_cpus(group)) {
2287 /* Bias balancing toward cpus of our domain */
2289 load = source_load(i, load_idx);
2291 load = target_load(i, load_idx);
2296 /* Adjust by relative CPU power of the group */
2297 avg_load = sg_div_cpu_power(group,
2298 avg_load * SCHED_LOAD_SCALE);
2301 this_load = avg_load;
2303 } else if (avg_load < min_load) {
2304 min_load = avg_load;
2307 } while (group = group->next, group != sd->groups);
2309 if (!idlest || 100*this_load < imbalance*min_load)
2315 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2318 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2320 unsigned long load, min_load = ULONG_MAX;
2324 /* Traverse only the allowed CPUs */
2325 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2326 load = weighted_cpuload(i);
2328 if (load < min_load || (load == min_load && i == this_cpu)) {
2338 * sched_balance_self: balance the current task (running on cpu) in domains
2339 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2342 * Balance, ie. select the least loaded group.
2344 * Returns the target CPU number, or the same CPU if no balancing is needed.
2346 * preempt must be disabled.
2348 static int sched_balance_self(int cpu, int flag)
2350 struct task_struct *t = current;
2351 struct sched_domain *tmp, *sd = NULL;
2353 for_each_domain(cpu, tmp) {
2355 * If power savings logic is enabled for a domain, stop there.
2357 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2359 if (tmp->flags & flag)
2367 struct sched_group *group;
2368 int new_cpu, weight;
2370 if (!(sd->flags & flag)) {
2375 group = find_idlest_group(sd, t, cpu);
2381 new_cpu = find_idlest_cpu(group, t, cpu);
2382 if (new_cpu == -1 || new_cpu == cpu) {
2383 /* Now try balancing at a lower domain level of cpu */
2388 /* Now try balancing at a lower domain level of new_cpu */
2390 weight = cpumask_weight(sched_domain_span(sd));
2392 for_each_domain(cpu, tmp) {
2393 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2395 if (tmp->flags & flag)
2398 /* while loop will break here if sd == NULL */
2404 #endif /* CONFIG_SMP */
2407 * task_oncpu_function_call - call a function on the cpu on which a task runs
2408 * @p: the task to evaluate
2409 * @func: the function to be called
2410 * @info: the function call argument
2412 * Calls the function @func when the task is currently running. This might
2413 * be on the current CPU, which just calls the function directly
2415 void task_oncpu_function_call(struct task_struct *p,
2416 void (*func) (void *info), void *info)
2423 smp_call_function_single(cpu, func, info, 1);
2428 * try_to_wake_up - wake up a thread
2429 * @p: the to-be-woken-up thread
2430 * @state: the mask of task states that can be woken
2431 * @sync: do a synchronous wakeup?
2433 * Put it on the run-queue if it's not already there. The "current"
2434 * thread is always on the run-queue (except when the actual
2435 * re-schedule is in progress), and as such you're allowed to do
2436 * the simpler "current->state = TASK_RUNNING" to mark yourself
2437 * runnable without the overhead of this.
2439 * returns failure only if the task is already active.
2441 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2443 int cpu, orig_cpu, this_cpu, success = 0;
2444 unsigned long flags;
2448 if (!sched_feat(SYNC_WAKEUPS))
2452 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2453 struct sched_domain *sd;
2455 this_cpu = raw_smp_processor_id();
2458 for_each_domain(this_cpu, sd) {
2459 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2468 rq = task_rq_lock(p, &flags);
2469 update_rq_clock(rq);
2470 old_state = p->state;
2471 if (!(old_state & state))
2479 this_cpu = smp_processor_id();
2482 if (unlikely(task_running(rq, p)))
2485 cpu = p->sched_class->select_task_rq(p, sync);
2486 if (cpu != orig_cpu) {
2487 set_task_cpu(p, cpu);
2488 task_rq_unlock(rq, &flags);
2489 /* might preempt at this point */
2490 rq = task_rq_lock(p, &flags);
2491 old_state = p->state;
2492 if (!(old_state & state))
2497 this_cpu = smp_processor_id();
2501 #ifdef CONFIG_SCHEDSTATS
2502 schedstat_inc(rq, ttwu_count);
2503 if (cpu == this_cpu)
2504 schedstat_inc(rq, ttwu_local);
2506 struct sched_domain *sd;
2507 for_each_domain(this_cpu, sd) {
2508 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2509 schedstat_inc(sd, ttwu_wake_remote);
2514 #endif /* CONFIG_SCHEDSTATS */
2517 #endif /* CONFIG_SMP */
2518 schedstat_inc(p, se.nr_wakeups);
2520 schedstat_inc(p, se.nr_wakeups_sync);
2521 if (orig_cpu != cpu)
2522 schedstat_inc(p, se.nr_wakeups_migrate);
2523 if (cpu == this_cpu)
2524 schedstat_inc(p, se.nr_wakeups_local);
2526 schedstat_inc(p, se.nr_wakeups_remote);
2527 activate_task(rq, p, 1);
2531 * Only attribute actual wakeups done by this task.
2533 if (!in_interrupt()) {
2534 struct sched_entity *se = ¤t->se;
2535 u64 sample = se->sum_exec_runtime;
2537 if (se->last_wakeup)
2538 sample -= se->last_wakeup;
2540 sample -= se->start_runtime;
2541 update_avg(&se->avg_wakeup, sample);
2543 se->last_wakeup = se->sum_exec_runtime;
2547 trace_sched_wakeup(rq, p, success);
2548 check_preempt_curr(rq, p, sync);
2550 p->state = TASK_RUNNING;
2552 if (p->sched_class->task_wake_up)
2553 p->sched_class->task_wake_up(rq, p);
2556 task_rq_unlock(rq, &flags);
2562 * wake_up_process - Wake up a specific process
2563 * @p: The process to be woken up.
2565 * Attempt to wake up the nominated process and move it to the set of runnable
2566 * processes. Returns 1 if the process was woken up, 0 if it was already
2569 * It may be assumed that this function implies a write memory barrier before
2570 * changing the task state if and only if any tasks are woken up.
2572 int wake_up_process(struct task_struct *p)
2574 return try_to_wake_up(p, TASK_ALL, 0);
2576 EXPORT_SYMBOL(wake_up_process);
2578 int wake_up_state(struct task_struct *p, unsigned int state)
2580 return try_to_wake_up(p, state, 0);
2584 * Perform scheduler related setup for a newly forked process p.
2585 * p is forked by current.
2587 * __sched_fork() is basic setup used by init_idle() too:
2589 static void __sched_fork(struct task_struct *p)
2591 p->se.exec_start = 0;
2592 p->se.sum_exec_runtime = 0;
2593 p->se.prev_sum_exec_runtime = 0;
2594 p->se.nr_migrations = 0;
2595 p->se.last_wakeup = 0;
2596 p->se.avg_overlap = 0;
2597 p->se.start_runtime = 0;
2598 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2600 #ifdef CONFIG_SCHEDSTATS
2601 p->se.wait_start = 0;
2603 p->se.wait_count = 0;
2606 p->se.sleep_start = 0;
2607 p->se.sleep_max = 0;
2608 p->se.sum_sleep_runtime = 0;
2610 p->se.block_start = 0;
2611 p->se.block_max = 0;
2613 p->se.slice_max = 0;
2615 p->se.nr_migrations_cold = 0;
2616 p->se.nr_failed_migrations_affine = 0;
2617 p->se.nr_failed_migrations_running = 0;
2618 p->se.nr_failed_migrations_hot = 0;
2619 p->se.nr_forced_migrations = 0;
2620 p->se.nr_forced2_migrations = 0;
2622 p->se.nr_wakeups = 0;
2623 p->se.nr_wakeups_sync = 0;
2624 p->se.nr_wakeups_migrate = 0;
2625 p->se.nr_wakeups_local = 0;
2626 p->se.nr_wakeups_remote = 0;
2627 p->se.nr_wakeups_affine = 0;
2628 p->se.nr_wakeups_affine_attempts = 0;
2629 p->se.nr_wakeups_passive = 0;
2630 p->se.nr_wakeups_idle = 0;
2634 INIT_LIST_HEAD(&p->rt.run_list);
2636 INIT_LIST_HEAD(&p->se.group_node);
2638 #ifdef CONFIG_PREEMPT_NOTIFIERS
2639 INIT_HLIST_HEAD(&p->preempt_notifiers);
2643 * We mark the process as running here, but have not actually
2644 * inserted it onto the runqueue yet. This guarantees that
2645 * nobody will actually run it, and a signal or other external
2646 * event cannot wake it up and insert it on the runqueue either.
2648 p->state = TASK_RUNNING;
2652 * fork()/clone()-time setup:
2654 void sched_fork(struct task_struct *p, int clone_flags)
2656 int cpu = get_cpu();
2661 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2663 set_task_cpu(p, cpu);
2666 * Make sure we do not leak PI boosting priority to the child.
2668 p->prio = current->normal_prio;
2671 * Revert to default priority/policy on fork if requested.
2673 if (unlikely(p->sched_reset_on_fork)) {
2674 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR)
2675 p->policy = SCHED_NORMAL;
2677 if (p->normal_prio < DEFAULT_PRIO)
2678 p->prio = DEFAULT_PRIO;
2680 if (PRIO_TO_NICE(p->static_prio) < 0) {
2681 p->static_prio = NICE_TO_PRIO(0);
2686 * We don't need the reset flag anymore after the fork. It has
2687 * fulfilled its duty:
2689 p->sched_reset_on_fork = 0;
2692 if (!rt_prio(p->prio))
2693 p->sched_class = &fair_sched_class;
2695 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2696 if (likely(sched_info_on()))
2697 memset(&p->sched_info, 0, sizeof(p->sched_info));
2699 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2702 #ifdef CONFIG_PREEMPT
2703 /* Want to start with kernel preemption disabled. */
2704 task_thread_info(p)->preempt_count = 1;
2706 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2712 * wake_up_new_task - wake up a newly created task for the first time.
2714 * This function will do some initial scheduler statistics housekeeping
2715 * that must be done for every newly created context, then puts the task
2716 * on the runqueue and wakes it.
2718 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2720 unsigned long flags;
2723 rq = task_rq_lock(p, &flags);
2724 BUG_ON(p->state != TASK_RUNNING);
2725 update_rq_clock(rq);
2727 p->prio = effective_prio(p);
2729 if (!p->sched_class->task_new || !current->se.on_rq) {
2730 activate_task(rq, p, 0);
2733 * Let the scheduling class do new task startup
2734 * management (if any):
2736 p->sched_class->task_new(rq, p);
2739 trace_sched_wakeup_new(rq, p, 1);
2740 check_preempt_curr(rq, p, 0);
2742 if (p->sched_class->task_wake_up)
2743 p->sched_class->task_wake_up(rq, p);
2745 task_rq_unlock(rq, &flags);
2748 #ifdef CONFIG_PREEMPT_NOTIFIERS
2751 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2752 * @notifier: notifier struct to register
2754 void preempt_notifier_register(struct preempt_notifier *notifier)
2756 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2758 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2761 * preempt_notifier_unregister - no longer interested in preemption notifications
2762 * @notifier: notifier struct to unregister
2764 * This is safe to call from within a preemption notifier.
2766 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2768 hlist_del(¬ifier->link);
2770 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2772 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2774 struct preempt_notifier *notifier;
2775 struct hlist_node *node;
2777 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2778 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2782 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2783 struct task_struct *next)
2785 struct preempt_notifier *notifier;
2786 struct hlist_node *node;
2788 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2789 notifier->ops->sched_out(notifier, next);
2792 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2794 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2799 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2800 struct task_struct *next)
2804 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2807 * prepare_task_switch - prepare to switch tasks
2808 * @rq: the runqueue preparing to switch
2809 * @prev: the current task that is being switched out
2810 * @next: the task we are going to switch to.
2812 * This is called with the rq lock held and interrupts off. It must
2813 * be paired with a subsequent finish_task_switch after the context
2816 * prepare_task_switch sets up locking and calls architecture specific
2820 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2821 struct task_struct *next)
2823 fire_sched_out_preempt_notifiers(prev, next);
2824 prepare_lock_switch(rq, next);
2825 prepare_arch_switch(next);
2829 * finish_task_switch - clean up after a task-switch
2830 * @rq: runqueue associated with task-switch
2831 * @prev: the thread we just switched away from.
2833 * finish_task_switch must be called after the context switch, paired
2834 * with a prepare_task_switch call before the context switch.
2835 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2836 * and do any other architecture-specific cleanup actions.
2838 * Note that we may have delayed dropping an mm in context_switch(). If
2839 * so, we finish that here outside of the runqueue lock. (Doing it
2840 * with the lock held can cause deadlocks; see schedule() for
2843 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2844 __releases(rq->lock)
2846 struct mm_struct *mm = rq->prev_mm;
2852 * A task struct has one reference for the use as "current".
2853 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2854 * schedule one last time. The schedule call will never return, and
2855 * the scheduled task must drop that reference.
2856 * The test for TASK_DEAD must occur while the runqueue locks are
2857 * still held, otherwise prev could be scheduled on another cpu, die
2858 * there before we look at prev->state, and then the reference would
2860 * Manfred Spraul <manfred@colorfullife.com>
2862 prev_state = prev->state;
2863 finish_arch_switch(prev);
2864 perf_counter_task_sched_in(current, cpu_of(rq));
2865 finish_lock_switch(rq, prev);
2867 fire_sched_in_preempt_notifiers(current);
2870 if (unlikely(prev_state == TASK_DEAD)) {
2872 * Remove function-return probe instances associated with this
2873 * task and put them back on the free list.
2875 kprobe_flush_task(prev);
2876 put_task_struct(prev);
2882 /* assumes rq->lock is held */
2883 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2885 if (prev->sched_class->pre_schedule)
2886 prev->sched_class->pre_schedule(rq, prev);
2889 /* rq->lock is NOT held, but preemption is disabled */
2890 static inline void post_schedule(struct rq *rq)
2892 if (rq->post_schedule) {
2893 unsigned long flags;
2895 spin_lock_irqsave(&rq->lock, flags);
2896 if (rq->curr->sched_class->post_schedule)
2897 rq->curr->sched_class->post_schedule(rq);
2898 spin_unlock_irqrestore(&rq->lock, flags);
2900 rq->post_schedule = 0;
2906 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2910 static inline void post_schedule(struct rq *rq)
2917 * schedule_tail - first thing a freshly forked thread must call.
2918 * @prev: the thread we just switched away from.
2920 asmlinkage void schedule_tail(struct task_struct *prev)
2921 __releases(rq->lock)
2923 struct rq *rq = this_rq();
2925 finish_task_switch(rq, prev);
2928 * FIXME: do we need to worry about rq being invalidated by the
2933 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2934 /* In this case, finish_task_switch does not reenable preemption */
2937 if (current->set_child_tid)
2938 put_user(task_pid_vnr(current), current->set_child_tid);
2942 * context_switch - switch to the new MM and the new
2943 * thread's register state.
2946 context_switch(struct rq *rq, struct task_struct *prev,
2947 struct task_struct *next)
2949 struct mm_struct *mm, *oldmm;
2951 prepare_task_switch(rq, prev, next);
2952 trace_sched_switch(rq, prev, next);
2954 oldmm = prev->active_mm;
2956 * For paravirt, this is coupled with an exit in switch_to to
2957 * combine the page table reload and the switch backend into
2960 arch_start_context_switch(prev);
2962 if (unlikely(!mm)) {
2963 next->active_mm = oldmm;
2964 atomic_inc(&oldmm->mm_count);
2965 enter_lazy_tlb(oldmm, next);
2967 switch_mm(oldmm, mm, next);
2969 if (unlikely(!prev->mm)) {
2970 prev->active_mm = NULL;
2971 rq->prev_mm = oldmm;
2974 * Since the runqueue lock will be released by the next
2975 * task (which is an invalid locking op but in the case
2976 * of the scheduler it's an obvious special-case), so we
2977 * do an early lockdep release here:
2979 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2980 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2983 /* Here we just switch the register state and the stack. */
2984 switch_to(prev, next, prev);
2988 * this_rq must be evaluated again because prev may have moved
2989 * CPUs since it called schedule(), thus the 'rq' on its stack
2990 * frame will be invalid.
2992 finish_task_switch(this_rq(), prev);
2996 * nr_running, nr_uninterruptible and nr_context_switches:
2998 * externally visible scheduler statistics: current number of runnable
2999 * threads, current number of uninterruptible-sleeping threads, total
3000 * number of context switches performed since bootup.
3002 unsigned long nr_running(void)
3004 unsigned long i, sum = 0;
3006 for_each_online_cpu(i)
3007 sum += cpu_rq(i)->nr_running;
3012 unsigned long nr_uninterruptible(void)
3014 unsigned long i, sum = 0;
3016 for_each_possible_cpu(i)
3017 sum += cpu_rq(i)->nr_uninterruptible;
3020 * Since we read the counters lockless, it might be slightly
3021 * inaccurate. Do not allow it to go below zero though:
3023 if (unlikely((long)sum < 0))
3029 unsigned long long nr_context_switches(void)
3032 unsigned long long sum = 0;
3034 for_each_possible_cpu(i)
3035 sum += cpu_rq(i)->nr_switches;
3040 unsigned long nr_iowait(void)
3042 unsigned long i, sum = 0;
3044 for_each_possible_cpu(i)
3045 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3050 /* Variables and functions for calc_load */
3051 static atomic_long_t calc_load_tasks;
3052 static unsigned long calc_load_update;
3053 unsigned long avenrun[3];
3054 EXPORT_SYMBOL(avenrun);
3057 * get_avenrun - get the load average array
3058 * @loads: pointer to dest load array
3059 * @offset: offset to add
3060 * @shift: shift count to shift the result left
3062 * These values are estimates at best, so no need for locking.
3064 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3066 loads[0] = (avenrun[0] + offset) << shift;
3067 loads[1] = (avenrun[1] + offset) << shift;
3068 loads[2] = (avenrun[2] + offset) << shift;
3071 static unsigned long
3072 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3075 load += active * (FIXED_1 - exp);
3076 return load >> FSHIFT;
3080 * calc_load - update the avenrun load estimates 10 ticks after the
3081 * CPUs have updated calc_load_tasks.
3083 void calc_global_load(void)
3085 unsigned long upd = calc_load_update + 10;
3088 if (time_before(jiffies, upd))
3091 active = atomic_long_read(&calc_load_tasks);
3092 active = active > 0 ? active * FIXED_1 : 0;
3094 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3095 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3096 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3098 calc_load_update += LOAD_FREQ;
3102 * Either called from update_cpu_load() or from a cpu going idle
3104 static void calc_load_account_active(struct rq *this_rq)
3106 long nr_active, delta;
3108 nr_active = this_rq->nr_running;
3109 nr_active += (long) this_rq->nr_uninterruptible;
3111 if (nr_active != this_rq->calc_load_active) {
3112 delta = nr_active - this_rq->calc_load_active;
3113 this_rq->calc_load_active = nr_active;
3114 atomic_long_add(delta, &calc_load_tasks);
3119 * Externally visible per-cpu scheduler statistics:
3120 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3122 u64 cpu_nr_migrations(int cpu)
3124 return cpu_rq(cpu)->nr_migrations_in;
3128 * Update rq->cpu_load[] statistics. This function is usually called every
3129 * scheduler tick (TICK_NSEC).
3131 static void update_cpu_load(struct rq *this_rq)
3133 unsigned long this_load = this_rq->load.weight;
3136 this_rq->nr_load_updates++;
3138 /* Update our load: */
3139 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3140 unsigned long old_load, new_load;
3142 /* scale is effectively 1 << i now, and >> i divides by scale */
3144 old_load = this_rq->cpu_load[i];
3145 new_load = this_load;
3147 * Round up the averaging division if load is increasing. This
3148 * prevents us from getting stuck on 9 if the load is 10, for
3151 if (new_load > old_load)
3152 new_load += scale-1;
3153 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3156 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3157 this_rq->calc_load_update += LOAD_FREQ;
3158 calc_load_account_active(this_rq);
3165 * double_rq_lock - safely lock two runqueues
3167 * Note this does not disable interrupts like task_rq_lock,
3168 * you need to do so manually before calling.
3170 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3171 __acquires(rq1->lock)
3172 __acquires(rq2->lock)
3174 BUG_ON(!irqs_disabled());
3176 spin_lock(&rq1->lock);
3177 __acquire(rq2->lock); /* Fake it out ;) */
3180 spin_lock(&rq1->lock);
3181 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3183 spin_lock(&rq2->lock);
3184 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3187 update_rq_clock(rq1);
3188 update_rq_clock(rq2);
3192 * double_rq_unlock - safely unlock two runqueues
3194 * Note this does not restore interrupts like task_rq_unlock,
3195 * you need to do so manually after calling.
3197 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3198 __releases(rq1->lock)
3199 __releases(rq2->lock)
3201 spin_unlock(&rq1->lock);
3203 spin_unlock(&rq2->lock);
3205 __release(rq2->lock);
3209 * If dest_cpu is allowed for this process, migrate the task to it.
3210 * This is accomplished by forcing the cpu_allowed mask to only
3211 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3212 * the cpu_allowed mask is restored.
3214 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3216 struct migration_req req;
3217 unsigned long flags;
3220 rq = task_rq_lock(p, &flags);
3221 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3222 || unlikely(!cpu_active(dest_cpu)))
3225 /* force the process onto the specified CPU */
3226 if (migrate_task(p, dest_cpu, &req)) {
3227 /* Need to wait for migration thread (might exit: take ref). */
3228 struct task_struct *mt = rq->migration_thread;
3230 get_task_struct(mt);
3231 task_rq_unlock(rq, &flags);
3232 wake_up_process(mt);
3233 put_task_struct(mt);
3234 wait_for_completion(&req.done);
3239 task_rq_unlock(rq, &flags);
3243 * sched_exec - execve() is a valuable balancing opportunity, because at
3244 * this point the task has the smallest effective memory and cache footprint.
3246 void sched_exec(void)
3248 int new_cpu, this_cpu = get_cpu();
3249 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3251 if (new_cpu != this_cpu)
3252 sched_migrate_task(current, new_cpu);
3256 * pull_task - move a task from a remote runqueue to the local runqueue.
3257 * Both runqueues must be locked.
3259 static void pull_task(struct rq *src_rq, struct task_struct *p,
3260 struct rq *this_rq, int this_cpu)
3262 deactivate_task(src_rq, p, 0);
3263 set_task_cpu(p, this_cpu);
3264 activate_task(this_rq, p, 0);
3266 * Note that idle threads have a prio of MAX_PRIO, for this test
3267 * to be always true for them.
3269 check_preempt_curr(this_rq, p, 0);
3273 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3276 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3277 struct sched_domain *sd, enum cpu_idle_type idle,
3280 int tsk_cache_hot = 0;
3282 * We do not migrate tasks that are:
3283 * 1) running (obviously), or
3284 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3285 * 3) are cache-hot on their current CPU.
3287 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3288 schedstat_inc(p, se.nr_failed_migrations_affine);
3293 if (task_running(rq, p)) {
3294 schedstat_inc(p, se.nr_failed_migrations_running);
3299 * Aggressive migration if:
3300 * 1) task is cache cold, or
3301 * 2) too many balance attempts have failed.
3304 tsk_cache_hot = task_hot(p, rq->clock, sd);
3305 if (!tsk_cache_hot ||
3306 sd->nr_balance_failed > sd->cache_nice_tries) {
3307 #ifdef CONFIG_SCHEDSTATS
3308 if (tsk_cache_hot) {
3309 schedstat_inc(sd, lb_hot_gained[idle]);
3310 schedstat_inc(p, se.nr_forced_migrations);
3316 if (tsk_cache_hot) {
3317 schedstat_inc(p, se.nr_failed_migrations_hot);
3323 static unsigned long
3324 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3325 unsigned long max_load_move, struct sched_domain *sd,
3326 enum cpu_idle_type idle, int *all_pinned,
3327 int *this_best_prio, struct rq_iterator *iterator)
3329 int loops = 0, pulled = 0, pinned = 0;
3330 struct task_struct *p;
3331 long rem_load_move = max_load_move;
3333 if (max_load_move == 0)
3339 * Start the load-balancing iterator:
3341 p = iterator->start(iterator->arg);
3343 if (!p || loops++ > sysctl_sched_nr_migrate)
3346 if ((p->se.load.weight >> 1) > rem_load_move ||
3347 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3348 p = iterator->next(iterator->arg);
3352 pull_task(busiest, p, this_rq, this_cpu);
3354 rem_load_move -= p->se.load.weight;
3356 #ifdef CONFIG_PREEMPT
3358 * NEWIDLE balancing is a source of latency, so preemptible kernels
3359 * will stop after the first task is pulled to minimize the critical
3362 if (idle == CPU_NEWLY_IDLE)
3367 * We only want to steal up to the prescribed amount of weighted load.
3369 if (rem_load_move > 0) {
3370 if (p->prio < *this_best_prio)
3371 *this_best_prio = p->prio;
3372 p = iterator->next(iterator->arg);
3377 * Right now, this is one of only two places pull_task() is called,
3378 * so we can safely collect pull_task() stats here rather than
3379 * inside pull_task().
3381 schedstat_add(sd, lb_gained[idle], pulled);
3384 *all_pinned = pinned;
3386 return max_load_move - rem_load_move;
3390 * move_tasks tries to move up to max_load_move weighted load from busiest to
3391 * this_rq, as part of a balancing operation within domain "sd".
3392 * Returns 1 if successful and 0 otherwise.
3394 * Called with both runqueues locked.
3396 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3397 unsigned long max_load_move,
3398 struct sched_domain *sd, enum cpu_idle_type idle,
3401 const struct sched_class *class = sched_class_highest;
3402 unsigned long total_load_moved = 0;
3403 int this_best_prio = this_rq->curr->prio;
3407 class->load_balance(this_rq, this_cpu, busiest,
3408 max_load_move - total_load_moved,
3409 sd, idle, all_pinned, &this_best_prio);
3410 class = class->next;
3412 #ifdef CONFIG_PREEMPT
3414 * NEWIDLE balancing is a source of latency, so preemptible
3415 * kernels will stop after the first task is pulled to minimize
3416 * the critical section.
3418 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3421 } while (class && max_load_move > total_load_moved);
3423 return total_load_moved > 0;
3427 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3428 struct sched_domain *sd, enum cpu_idle_type idle,
3429 struct rq_iterator *iterator)
3431 struct task_struct *p = iterator->start(iterator->arg);
3435 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3436 pull_task(busiest, p, this_rq, this_cpu);
3438 * Right now, this is only the second place pull_task()
3439 * is called, so we can safely collect pull_task()
3440 * stats here rather than inside pull_task().
3442 schedstat_inc(sd, lb_gained[idle]);
3446 p = iterator->next(iterator->arg);
3453 * move_one_task tries to move exactly one task from busiest to this_rq, as
3454 * part of active balancing operations within "domain".
3455 * Returns 1 if successful and 0 otherwise.
3457 * Called with both runqueues locked.
3459 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3460 struct sched_domain *sd, enum cpu_idle_type idle)
3462 const struct sched_class *class;
3464 for (class = sched_class_highest; class; class = class->next)
3465 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3470 /********** Helpers for find_busiest_group ************************/
3472 * sd_lb_stats - Structure to store the statistics of a sched_domain
3473 * during load balancing.
3475 struct sd_lb_stats {
3476 struct sched_group *busiest; /* Busiest group in this sd */
3477 struct sched_group *this; /* Local group in this sd */
3478 unsigned long total_load; /* Total load of all groups in sd */
3479 unsigned long total_pwr; /* Total power of all groups in sd */
3480 unsigned long avg_load; /* Average load across all groups in sd */
3482 /** Statistics of this group */
3483 unsigned long this_load;
3484 unsigned long this_load_per_task;
3485 unsigned long this_nr_running;
3487 /* Statistics of the busiest group */
3488 unsigned long max_load;
3489 unsigned long busiest_load_per_task;
3490 unsigned long busiest_nr_running;
3492 int group_imb; /* Is there imbalance in this sd */
3493 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3494 int power_savings_balance; /* Is powersave balance needed for this sd */
3495 struct sched_group *group_min; /* Least loaded group in sd */
3496 struct sched_group *group_leader; /* Group which relieves group_min */
3497 unsigned long min_load_per_task; /* load_per_task in group_min */
3498 unsigned long leader_nr_running; /* Nr running of group_leader */
3499 unsigned long min_nr_running; /* Nr running of group_min */
3504 * sg_lb_stats - stats of a sched_group required for load_balancing
3506 struct sg_lb_stats {
3507 unsigned long avg_load; /*Avg load across the CPUs of the group */
3508 unsigned long group_load; /* Total load over the CPUs of the group */
3509 unsigned long sum_nr_running; /* Nr tasks running in the group */
3510 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3511 unsigned long group_capacity;
3512 int group_imb; /* Is there an imbalance in the group ? */
3516 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3517 * @group: The group whose first cpu is to be returned.
3519 static inline unsigned int group_first_cpu(struct sched_group *group)
3521 return cpumask_first(sched_group_cpus(group));
3525 * get_sd_load_idx - Obtain the load index for a given sched domain.
3526 * @sd: The sched_domain whose load_idx is to be obtained.
3527 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3529 static inline int get_sd_load_idx(struct sched_domain *sd,
3530 enum cpu_idle_type idle)
3536 load_idx = sd->busy_idx;
3539 case CPU_NEWLY_IDLE:
3540 load_idx = sd->newidle_idx;
3543 load_idx = sd->idle_idx;
3551 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3553 * init_sd_power_savings_stats - Initialize power savings statistics for
3554 * the given sched_domain, during load balancing.
3556 * @sd: Sched domain whose power-savings statistics are to be initialized.
3557 * @sds: Variable containing the statistics for sd.
3558 * @idle: Idle status of the CPU at which we're performing load-balancing.
3560 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3561 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3564 * Busy processors will not participate in power savings
3567 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3568 sds->power_savings_balance = 0;
3570 sds->power_savings_balance = 1;
3571 sds->min_nr_running = ULONG_MAX;
3572 sds->leader_nr_running = 0;
3577 * update_sd_power_savings_stats - Update the power saving stats for a
3578 * sched_domain while performing load balancing.
3580 * @group: sched_group belonging to the sched_domain under consideration.
3581 * @sds: Variable containing the statistics of the sched_domain
3582 * @local_group: Does group contain the CPU for which we're performing
3584 * @sgs: Variable containing the statistics of the group.
3586 static inline void update_sd_power_savings_stats(struct sched_group *group,
3587 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3590 if (!sds->power_savings_balance)
3594 * If the local group is idle or completely loaded
3595 * no need to do power savings balance at this domain
3597 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3598 !sds->this_nr_running))
3599 sds->power_savings_balance = 0;
3602 * If a group is already running at full capacity or idle,
3603 * don't include that group in power savings calculations
3605 if (!sds->power_savings_balance ||
3606 sgs->sum_nr_running >= sgs->group_capacity ||
3607 !sgs->sum_nr_running)
3611 * Calculate the group which has the least non-idle load.
3612 * This is the group from where we need to pick up the load
3615 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3616 (sgs->sum_nr_running == sds->min_nr_running &&
3617 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3618 sds->group_min = group;
3619 sds->min_nr_running = sgs->sum_nr_running;
3620 sds->min_load_per_task = sgs->sum_weighted_load /
3621 sgs->sum_nr_running;
3625 * Calculate the group which is almost near its
3626 * capacity but still has some space to pick up some load
3627 * from other group and save more power
3629 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3632 if (sgs->sum_nr_running > sds->leader_nr_running ||
3633 (sgs->sum_nr_running == sds->leader_nr_running &&
3634 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3635 sds->group_leader = group;
3636 sds->leader_nr_running = sgs->sum_nr_running;
3641 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3642 * @sds: Variable containing the statistics of the sched_domain
3643 * under consideration.
3644 * @this_cpu: Cpu at which we're currently performing load-balancing.
3645 * @imbalance: Variable to store the imbalance.
3648 * Check if we have potential to perform some power-savings balance.
3649 * If yes, set the busiest group to be the least loaded group in the
3650 * sched_domain, so that it's CPUs can be put to idle.
3652 * Returns 1 if there is potential to perform power-savings balance.
3655 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3656 int this_cpu, unsigned long *imbalance)
3658 if (!sds->power_savings_balance)
3661 if (sds->this != sds->group_leader ||
3662 sds->group_leader == sds->group_min)
3665 *imbalance = sds->min_load_per_task;
3666 sds->busiest = sds->group_min;
3668 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3669 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3670 group_first_cpu(sds->group_leader);
3676 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3677 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3678 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3683 static inline void update_sd_power_savings_stats(struct sched_group *group,
3684 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3689 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3690 int this_cpu, unsigned long *imbalance)
3694 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3698 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3699 * @group: sched_group whose statistics are to be updated.
3700 * @this_cpu: Cpu for which load balance is currently performed.
3701 * @idle: Idle status of this_cpu
3702 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3703 * @sd_idle: Idle status of the sched_domain containing group.
3704 * @local_group: Does group contain this_cpu.
3705 * @cpus: Set of cpus considered for load balancing.
3706 * @balance: Should we balance.
3707 * @sgs: variable to hold the statistics for this group.
3709 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3710 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3711 int local_group, const struct cpumask *cpus,
3712 int *balance, struct sg_lb_stats *sgs)
3714 unsigned long load, max_cpu_load, min_cpu_load;
3716 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3717 unsigned long sum_avg_load_per_task;
3718 unsigned long avg_load_per_task;
3721 balance_cpu = group_first_cpu(group);
3723 /* Tally up the load of all CPUs in the group */
3724 sum_avg_load_per_task = avg_load_per_task = 0;
3726 min_cpu_load = ~0UL;
3728 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3729 struct rq *rq = cpu_rq(i);
3731 if (*sd_idle && rq->nr_running)
3734 /* Bias balancing toward cpus of our domain */
3736 if (idle_cpu(i) && !first_idle_cpu) {
3741 load = target_load(i, load_idx);
3743 load = source_load(i, load_idx);
3744 if (load > max_cpu_load)
3745 max_cpu_load = load;
3746 if (min_cpu_load > load)
3747 min_cpu_load = load;
3750 sgs->group_load += load;
3751 sgs->sum_nr_running += rq->nr_running;
3752 sgs->sum_weighted_load += weighted_cpuload(i);
3754 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3758 * First idle cpu or the first cpu(busiest) in this sched group
3759 * is eligible for doing load balancing at this and above
3760 * domains. In the newly idle case, we will allow all the cpu's
3761 * to do the newly idle load balance.
3763 if (idle != CPU_NEWLY_IDLE && local_group &&
3764 balance_cpu != this_cpu && balance) {
3769 /* Adjust by relative CPU power of the group */
3770 sgs->avg_load = sg_div_cpu_power(group,
3771 sgs->group_load * SCHED_LOAD_SCALE);
3775 * Consider the group unbalanced when the imbalance is larger
3776 * than the average weight of two tasks.
3778 * APZ: with cgroup the avg task weight can vary wildly and
3779 * might not be a suitable number - should we keep a
3780 * normalized nr_running number somewhere that negates
3783 avg_load_per_task = sg_div_cpu_power(group,
3784 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3786 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3789 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3794 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3795 * @sd: sched_domain whose statistics are to be updated.
3796 * @this_cpu: Cpu for which load balance is currently performed.
3797 * @idle: Idle status of this_cpu
3798 * @sd_idle: Idle status of the sched_domain containing group.
3799 * @cpus: Set of cpus considered for load balancing.
3800 * @balance: Should we balance.
3801 * @sds: variable to hold the statistics for this sched_domain.
3803 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3804 enum cpu_idle_type idle, int *sd_idle,
3805 const struct cpumask *cpus, int *balance,
3806 struct sd_lb_stats *sds)
3808 struct sched_group *group = sd->groups;
3809 struct sg_lb_stats sgs;
3812 init_sd_power_savings_stats(sd, sds, idle);
3813 load_idx = get_sd_load_idx(sd, idle);
3818 local_group = cpumask_test_cpu(this_cpu,
3819 sched_group_cpus(group));
3820 memset(&sgs, 0, sizeof(sgs));
3821 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3822 local_group, cpus, balance, &sgs);
3824 if (local_group && balance && !(*balance))
3827 sds->total_load += sgs.group_load;
3828 sds->total_pwr += group->__cpu_power;
3831 sds->this_load = sgs.avg_load;
3833 sds->this_nr_running = sgs.sum_nr_running;
3834 sds->this_load_per_task = sgs.sum_weighted_load;
3835 } else if (sgs.avg_load > sds->max_load &&
3836 (sgs.sum_nr_running > sgs.group_capacity ||
3838 sds->max_load = sgs.avg_load;
3839 sds->busiest = group;
3840 sds->busiest_nr_running = sgs.sum_nr_running;
3841 sds->busiest_load_per_task = sgs.sum_weighted_load;
3842 sds->group_imb = sgs.group_imb;
3845 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3846 group = group->next;
3847 } while (group != sd->groups);
3852 * fix_small_imbalance - Calculate the minor imbalance that exists
3853 * amongst the groups of a sched_domain, during
3855 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3856 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3857 * @imbalance: Variable to store the imbalance.
3859 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3860 int this_cpu, unsigned long *imbalance)
3862 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3863 unsigned int imbn = 2;
3865 if (sds->this_nr_running) {
3866 sds->this_load_per_task /= sds->this_nr_running;
3867 if (sds->busiest_load_per_task >
3868 sds->this_load_per_task)
3871 sds->this_load_per_task =
3872 cpu_avg_load_per_task(this_cpu);
3874 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3875 sds->busiest_load_per_task * imbn) {
3876 *imbalance = sds->busiest_load_per_task;
3881 * OK, we don't have enough imbalance to justify moving tasks,
3882 * however we may be able to increase total CPU power used by
3886 pwr_now += sds->busiest->__cpu_power *
3887 min(sds->busiest_load_per_task, sds->max_load);
3888 pwr_now += sds->this->__cpu_power *
3889 min(sds->this_load_per_task, sds->this_load);
3890 pwr_now /= SCHED_LOAD_SCALE;
3892 /* Amount of load we'd subtract */
3893 tmp = sg_div_cpu_power(sds->busiest,
3894 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3895 if (sds->max_load > tmp)
3896 pwr_move += sds->busiest->__cpu_power *
3897 min(sds->busiest_load_per_task, sds->max_load - tmp);
3899 /* Amount of load we'd add */
3900 if (sds->max_load * sds->busiest->__cpu_power <
3901 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3902 tmp = sg_div_cpu_power(sds->this,
3903 sds->max_load * sds->busiest->__cpu_power);
3905 tmp = sg_div_cpu_power(sds->this,
3906 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3907 pwr_move += sds->this->__cpu_power *
3908 min(sds->this_load_per_task, sds->this_load + tmp);
3909 pwr_move /= SCHED_LOAD_SCALE;
3911 /* Move if we gain throughput */
3912 if (pwr_move > pwr_now)
3913 *imbalance = sds->busiest_load_per_task;
3917 * calculate_imbalance - Calculate the amount of imbalance present within the
3918 * groups of a given sched_domain during load balance.
3919 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3920 * @this_cpu: Cpu for which currently load balance is being performed.
3921 * @imbalance: The variable to store the imbalance.
3923 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3924 unsigned long *imbalance)
3926 unsigned long max_pull;
3928 * In the presence of smp nice balancing, certain scenarios can have
3929 * max load less than avg load(as we skip the groups at or below
3930 * its cpu_power, while calculating max_load..)
3932 if (sds->max_load < sds->avg_load) {
3934 return fix_small_imbalance(sds, this_cpu, imbalance);
3937 /* Don't want to pull so many tasks that a group would go idle */
3938 max_pull = min(sds->max_load - sds->avg_load,
3939 sds->max_load - sds->busiest_load_per_task);
3941 /* How much load to actually move to equalise the imbalance */
3942 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3943 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3947 * if *imbalance is less than the average load per runnable task
3948 * there is no gaurantee that any tasks will be moved so we'll have
3949 * a think about bumping its value to force at least one task to be
3952 if (*imbalance < sds->busiest_load_per_task)
3953 return fix_small_imbalance(sds, this_cpu, imbalance);
3956 /******* find_busiest_group() helpers end here *********************/
3959 * find_busiest_group - Returns the busiest group within the sched_domain
3960 * if there is an imbalance. If there isn't an imbalance, and
3961 * the user has opted for power-savings, it returns a group whose
3962 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3963 * such a group exists.
3965 * Also calculates the amount of weighted load which should be moved
3966 * to restore balance.
3968 * @sd: The sched_domain whose busiest group is to be returned.
3969 * @this_cpu: The cpu for which load balancing is currently being performed.
3970 * @imbalance: Variable which stores amount of weighted load which should
3971 * be moved to restore balance/put a group to idle.
3972 * @idle: The idle status of this_cpu.
3973 * @sd_idle: The idleness of sd
3974 * @cpus: The set of CPUs under consideration for load-balancing.
3975 * @balance: Pointer to a variable indicating if this_cpu
3976 * is the appropriate cpu to perform load balancing at this_level.
3978 * Returns: - the busiest group if imbalance exists.
3979 * - If no imbalance and user has opted for power-savings balance,
3980 * return the least loaded group whose CPUs can be
3981 * put to idle by rebalancing its tasks onto our group.
3983 static struct sched_group *
3984 find_busiest_group(struct sched_domain *sd, int this_cpu,
3985 unsigned long *imbalance, enum cpu_idle_type idle,
3986 int *sd_idle, const struct cpumask *cpus, int *balance)
3988 struct sd_lb_stats sds;
3990 memset(&sds, 0, sizeof(sds));
3993 * Compute the various statistics relavent for load balancing at
3996 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3999 /* Cases where imbalance does not exist from POV of this_cpu */
4000 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4002 * 2) There is no busy sibling group to pull from.
4003 * 3) This group is the busiest group.
4004 * 4) This group is more busy than the avg busieness at this
4006 * 5) The imbalance is within the specified limit.
4007 * 6) Any rebalance would lead to ping-pong
4009 if (balance && !(*balance))
4012 if (!sds.busiest || sds.busiest_nr_running == 0)
4015 if (sds.this_load >= sds.max_load)
4018 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4020 if (sds.this_load >= sds.avg_load)
4023 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4026 sds.busiest_load_per_task /= sds.busiest_nr_running;
4028 sds.busiest_load_per_task =
4029 min(sds.busiest_load_per_task, sds.avg_load);
4032 * We're trying to get all the cpus to the average_load, so we don't
4033 * want to push ourselves above the average load, nor do we wish to
4034 * reduce the max loaded cpu below the average load, as either of these
4035 * actions would just result in more rebalancing later, and ping-pong
4036 * tasks around. Thus we look for the minimum possible imbalance.
4037 * Negative imbalances (*we* are more loaded than anyone else) will
4038 * be counted as no imbalance for these purposes -- we can't fix that
4039 * by pulling tasks to us. Be careful of negative numbers as they'll
4040 * appear as very large values with unsigned longs.
4042 if (sds.max_load <= sds.busiest_load_per_task)
4045 /* Looks like there is an imbalance. Compute it */
4046 calculate_imbalance(&sds, this_cpu, imbalance);
4051 * There is no obvious imbalance. But check if we can do some balancing
4054 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4062 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4065 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4066 unsigned long imbalance, const struct cpumask *cpus)
4068 struct rq *busiest = NULL, *rq;
4069 unsigned long max_load = 0;
4072 for_each_cpu(i, sched_group_cpus(group)) {
4075 if (!cpumask_test_cpu(i, cpus))
4079 wl = weighted_cpuload(i);
4081 if (rq->nr_running == 1 && wl > imbalance)
4084 if (wl > max_load) {
4094 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4095 * so long as it is large enough.
4097 #define MAX_PINNED_INTERVAL 512
4099 /* Working cpumask for load_balance and load_balance_newidle. */
4100 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4103 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4104 * tasks if there is an imbalance.
4106 static int load_balance(int this_cpu, struct rq *this_rq,
4107 struct sched_domain *sd, enum cpu_idle_type idle,
4110 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4111 struct sched_group *group;
4112 unsigned long imbalance;
4114 unsigned long flags;
4115 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4117 cpumask_setall(cpus);
4120 * When power savings policy is enabled for the parent domain, idle
4121 * sibling can pick up load irrespective of busy siblings. In this case,
4122 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4123 * portraying it as CPU_NOT_IDLE.
4125 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4126 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4129 schedstat_inc(sd, lb_count[idle]);
4133 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4140 schedstat_inc(sd, lb_nobusyg[idle]);
4144 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4146 schedstat_inc(sd, lb_nobusyq[idle]);
4150 BUG_ON(busiest == this_rq);
4152 schedstat_add(sd, lb_imbalance[idle], imbalance);
4155 if (busiest->nr_running > 1) {
4157 * Attempt to move tasks. If find_busiest_group has found
4158 * an imbalance but busiest->nr_running <= 1, the group is
4159 * still unbalanced. ld_moved simply stays zero, so it is
4160 * correctly treated as an imbalance.
4162 local_irq_save(flags);
4163 double_rq_lock(this_rq, busiest);
4164 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4165 imbalance, sd, idle, &all_pinned);
4166 double_rq_unlock(this_rq, busiest);
4167 local_irq_restore(flags);
4170 * some other cpu did the load balance for us.
4172 if (ld_moved && this_cpu != smp_processor_id())
4173 resched_cpu(this_cpu);
4175 /* All tasks on this runqueue were pinned by CPU affinity */
4176 if (unlikely(all_pinned)) {
4177 cpumask_clear_cpu(cpu_of(busiest), cpus);
4178 if (!cpumask_empty(cpus))
4185 schedstat_inc(sd, lb_failed[idle]);
4186 sd->nr_balance_failed++;
4188 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4190 spin_lock_irqsave(&busiest->lock, flags);
4192 /* don't kick the migration_thread, if the curr
4193 * task on busiest cpu can't be moved to this_cpu
4195 if (!cpumask_test_cpu(this_cpu,
4196 &busiest->curr->cpus_allowed)) {
4197 spin_unlock_irqrestore(&busiest->lock, flags);
4199 goto out_one_pinned;
4202 if (!busiest->active_balance) {
4203 busiest->active_balance = 1;
4204 busiest->push_cpu = this_cpu;
4207 spin_unlock_irqrestore(&busiest->lock, flags);
4209 wake_up_process(busiest->migration_thread);
4212 * We've kicked active balancing, reset the failure
4215 sd->nr_balance_failed = sd->cache_nice_tries+1;
4218 sd->nr_balance_failed = 0;
4220 if (likely(!active_balance)) {
4221 /* We were unbalanced, so reset the balancing interval */
4222 sd->balance_interval = sd->min_interval;
4225 * If we've begun active balancing, start to back off. This
4226 * case may not be covered by the all_pinned logic if there
4227 * is only 1 task on the busy runqueue (because we don't call
4230 if (sd->balance_interval < sd->max_interval)
4231 sd->balance_interval *= 2;
4234 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4235 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4241 schedstat_inc(sd, lb_balanced[idle]);
4243 sd->nr_balance_failed = 0;
4246 /* tune up the balancing interval */
4247 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4248 (sd->balance_interval < sd->max_interval))
4249 sd->balance_interval *= 2;
4251 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4252 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4263 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4264 * tasks if there is an imbalance.
4266 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4267 * this_rq is locked.
4270 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4272 struct sched_group *group;
4273 struct rq *busiest = NULL;
4274 unsigned long imbalance;
4278 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4280 cpumask_setall(cpus);
4283 * When power savings policy is enabled for the parent domain, idle
4284 * sibling can pick up load irrespective of busy siblings. In this case,
4285 * let the state of idle sibling percolate up as IDLE, instead of
4286 * portraying it as CPU_NOT_IDLE.
4288 if (sd->flags & SD_SHARE_CPUPOWER &&
4289 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4292 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4294 update_shares_locked(this_rq, sd);
4295 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4296 &sd_idle, cpus, NULL);
4298 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4302 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4304 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4308 BUG_ON(busiest == this_rq);
4310 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4313 if (busiest->nr_running > 1) {
4314 /* Attempt to move tasks */
4315 double_lock_balance(this_rq, busiest);
4316 /* this_rq->clock is already updated */
4317 update_rq_clock(busiest);
4318 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4319 imbalance, sd, CPU_NEWLY_IDLE,
4321 double_unlock_balance(this_rq, busiest);
4323 if (unlikely(all_pinned)) {
4324 cpumask_clear_cpu(cpu_of(busiest), cpus);
4325 if (!cpumask_empty(cpus))
4331 int active_balance = 0;
4333 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4334 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4335 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4338 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4341 if (sd->nr_balance_failed++ < 2)
4345 * The only task running in a non-idle cpu can be moved to this
4346 * cpu in an attempt to completely freeup the other CPU
4347 * package. The same method used to move task in load_balance()
4348 * have been extended for load_balance_newidle() to speedup
4349 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4351 * The package power saving logic comes from
4352 * find_busiest_group(). If there are no imbalance, then
4353 * f_b_g() will return NULL. However when sched_mc={1,2} then
4354 * f_b_g() will select a group from which a running task may be
4355 * pulled to this cpu in order to make the other package idle.
4356 * If there is no opportunity to make a package idle and if
4357 * there are no imbalance, then f_b_g() will return NULL and no
4358 * action will be taken in load_balance_newidle().
4360 * Under normal task pull operation due to imbalance, there
4361 * will be more than one task in the source run queue and
4362 * move_tasks() will succeed. ld_moved will be true and this
4363 * active balance code will not be triggered.
4366 /* Lock busiest in correct order while this_rq is held */
4367 double_lock_balance(this_rq, busiest);
4370 * don't kick the migration_thread, if the curr
4371 * task on busiest cpu can't be moved to this_cpu
4373 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4374 double_unlock_balance(this_rq, busiest);
4379 if (!busiest->active_balance) {
4380 busiest->active_balance = 1;
4381 busiest->push_cpu = this_cpu;
4385 double_unlock_balance(this_rq, busiest);
4387 * Should not call ttwu while holding a rq->lock
4389 spin_unlock(&this_rq->lock);
4391 wake_up_process(busiest->migration_thread);
4392 spin_lock(&this_rq->lock);
4395 sd->nr_balance_failed = 0;
4397 update_shares_locked(this_rq, sd);
4401 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4402 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4403 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4405 sd->nr_balance_failed = 0;
4411 * idle_balance is called by schedule() if this_cpu is about to become
4412 * idle. Attempts to pull tasks from other CPUs.
4414 static void idle_balance(int this_cpu, struct rq *this_rq)
4416 struct sched_domain *sd;
4417 int pulled_task = 0;
4418 unsigned long next_balance = jiffies + HZ;
4420 for_each_domain(this_cpu, sd) {
4421 unsigned long interval;
4423 if (!(sd->flags & SD_LOAD_BALANCE))
4426 if (sd->flags & SD_BALANCE_NEWIDLE)
4427 /* If we've pulled tasks over stop searching: */
4428 pulled_task = load_balance_newidle(this_cpu, this_rq,
4431 interval = msecs_to_jiffies(sd->balance_interval);
4432 if (time_after(next_balance, sd->last_balance + interval))
4433 next_balance = sd->last_balance + interval;
4437 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4439 * We are going idle. next_balance may be set based on
4440 * a busy processor. So reset next_balance.
4442 this_rq->next_balance = next_balance;
4447 * active_load_balance is run by migration threads. It pushes running tasks
4448 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4449 * running on each physical CPU where possible, and avoids physical /
4450 * logical imbalances.
4452 * Called with busiest_rq locked.
4454 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4456 int target_cpu = busiest_rq->push_cpu;
4457 struct sched_domain *sd;
4458 struct rq *target_rq;
4460 /* Is there any task to move? */
4461 if (busiest_rq->nr_running <= 1)
4464 target_rq = cpu_rq(target_cpu);
4467 * This condition is "impossible", if it occurs
4468 * we need to fix it. Originally reported by
4469 * Bjorn Helgaas on a 128-cpu setup.
4471 BUG_ON(busiest_rq == target_rq);
4473 /* move a task from busiest_rq to target_rq */
4474 double_lock_balance(busiest_rq, target_rq);
4475 update_rq_clock(busiest_rq);
4476 update_rq_clock(target_rq);
4478 /* Search for an sd spanning us and the target CPU. */
4479 for_each_domain(target_cpu, sd) {
4480 if ((sd->flags & SD_LOAD_BALANCE) &&
4481 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4486 schedstat_inc(sd, alb_count);
4488 if (move_one_task(target_rq, target_cpu, busiest_rq,
4490 schedstat_inc(sd, alb_pushed);
4492 schedstat_inc(sd, alb_failed);
4494 double_unlock_balance(busiest_rq, target_rq);
4499 atomic_t load_balancer;
4500 cpumask_var_t cpu_mask;
4501 cpumask_var_t ilb_grp_nohz_mask;
4502 } nohz ____cacheline_aligned = {
4503 .load_balancer = ATOMIC_INIT(-1),
4506 int get_nohz_load_balancer(void)
4508 return atomic_read(&nohz.load_balancer);
4511 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4513 * lowest_flag_domain - Return lowest sched_domain containing flag.
4514 * @cpu: The cpu whose lowest level of sched domain is to
4516 * @flag: The flag to check for the lowest sched_domain
4517 * for the given cpu.
4519 * Returns the lowest sched_domain of a cpu which contains the given flag.
4521 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4523 struct sched_domain *sd;
4525 for_each_domain(cpu, sd)
4526 if (sd && (sd->flags & flag))
4533 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4534 * @cpu: The cpu whose domains we're iterating over.
4535 * @sd: variable holding the value of the power_savings_sd
4537 * @flag: The flag to filter the sched_domains to be iterated.
4539 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4540 * set, starting from the lowest sched_domain to the highest.
4542 #define for_each_flag_domain(cpu, sd, flag) \
4543 for (sd = lowest_flag_domain(cpu, flag); \
4544 (sd && (sd->flags & flag)); sd = sd->parent)
4547 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4548 * @ilb_group: group to be checked for semi-idleness
4550 * Returns: 1 if the group is semi-idle. 0 otherwise.
4552 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4553 * and atleast one non-idle CPU. This helper function checks if the given
4554 * sched_group is semi-idle or not.
4556 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4558 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4559 sched_group_cpus(ilb_group));
4562 * A sched_group is semi-idle when it has atleast one busy cpu
4563 * and atleast one idle cpu.
4565 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4568 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4574 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4575 * @cpu: The cpu which is nominating a new idle_load_balancer.
4577 * Returns: Returns the id of the idle load balancer if it exists,
4578 * Else, returns >= nr_cpu_ids.
4580 * This algorithm picks the idle load balancer such that it belongs to a
4581 * semi-idle powersavings sched_domain. The idea is to try and avoid
4582 * completely idle packages/cores just for the purpose of idle load balancing
4583 * when there are other idle cpu's which are better suited for that job.
4585 static int find_new_ilb(int cpu)
4587 struct sched_domain *sd;
4588 struct sched_group *ilb_group;
4591 * Have idle load balancer selection from semi-idle packages only
4592 * when power-aware load balancing is enabled
4594 if (!(sched_smt_power_savings || sched_mc_power_savings))
4598 * Optimize for the case when we have no idle CPUs or only one
4599 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4601 if (cpumask_weight(nohz.cpu_mask) < 2)
4604 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4605 ilb_group = sd->groups;
4608 if (is_semi_idle_group(ilb_group))
4609 return cpumask_first(nohz.ilb_grp_nohz_mask);
4611 ilb_group = ilb_group->next;
4613 } while (ilb_group != sd->groups);
4617 return cpumask_first(nohz.cpu_mask);
4619 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4620 static inline int find_new_ilb(int call_cpu)
4622 return cpumask_first(nohz.cpu_mask);
4627 * This routine will try to nominate the ilb (idle load balancing)
4628 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4629 * load balancing on behalf of all those cpus. If all the cpus in the system
4630 * go into this tickless mode, then there will be no ilb owner (as there is
4631 * no need for one) and all the cpus will sleep till the next wakeup event
4634 * For the ilb owner, tick is not stopped. And this tick will be used
4635 * for idle load balancing. ilb owner will still be part of
4638 * While stopping the tick, this cpu will become the ilb owner if there
4639 * is no other owner. And will be the owner till that cpu becomes busy
4640 * or if all cpus in the system stop their ticks at which point
4641 * there is no need for ilb owner.
4643 * When the ilb owner becomes busy, it nominates another owner, during the
4644 * next busy scheduler_tick()
4646 int select_nohz_load_balancer(int stop_tick)
4648 int cpu = smp_processor_id();
4651 cpu_rq(cpu)->in_nohz_recently = 1;
4653 if (!cpu_active(cpu)) {
4654 if (atomic_read(&nohz.load_balancer) != cpu)
4658 * If we are going offline and still the leader,
4661 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4667 cpumask_set_cpu(cpu, nohz.cpu_mask);
4669 /* time for ilb owner also to sleep */
4670 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4671 if (atomic_read(&nohz.load_balancer) == cpu)
4672 atomic_set(&nohz.load_balancer, -1);
4676 if (atomic_read(&nohz.load_balancer) == -1) {
4677 /* make me the ilb owner */
4678 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4680 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4683 if (!(sched_smt_power_savings ||
4684 sched_mc_power_savings))
4687 * Check to see if there is a more power-efficient
4690 new_ilb = find_new_ilb(cpu);
4691 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4692 atomic_set(&nohz.load_balancer, -1);
4693 resched_cpu(new_ilb);
4699 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4702 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4704 if (atomic_read(&nohz.load_balancer) == cpu)
4705 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4712 static DEFINE_SPINLOCK(balancing);
4715 * It checks each scheduling domain to see if it is due to be balanced,
4716 * and initiates a balancing operation if so.
4718 * Balancing parameters are set up in arch_init_sched_domains.
4720 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4723 struct rq *rq = cpu_rq(cpu);
4724 unsigned long interval;
4725 struct sched_domain *sd;
4726 /* Earliest time when we have to do rebalance again */
4727 unsigned long next_balance = jiffies + 60*HZ;
4728 int update_next_balance = 0;
4731 for_each_domain(cpu, sd) {
4732 if (!(sd->flags & SD_LOAD_BALANCE))
4735 interval = sd->balance_interval;
4736 if (idle != CPU_IDLE)
4737 interval *= sd->busy_factor;
4739 /* scale ms to jiffies */
4740 interval = msecs_to_jiffies(interval);
4741 if (unlikely(!interval))
4743 if (interval > HZ*NR_CPUS/10)
4744 interval = HZ*NR_CPUS/10;
4746 need_serialize = sd->flags & SD_SERIALIZE;
4748 if (need_serialize) {
4749 if (!spin_trylock(&balancing))
4753 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4754 if (load_balance(cpu, rq, sd, idle, &balance)) {
4756 * We've pulled tasks over so either we're no
4757 * longer idle, or one of our SMT siblings is
4760 idle = CPU_NOT_IDLE;
4762 sd->last_balance = jiffies;
4765 spin_unlock(&balancing);
4767 if (time_after(next_balance, sd->last_balance + interval)) {
4768 next_balance = sd->last_balance + interval;
4769 update_next_balance = 1;
4773 * Stop the load balance at this level. There is another
4774 * CPU in our sched group which is doing load balancing more
4782 * next_balance will be updated only when there is a need.
4783 * When the cpu is attached to null domain for ex, it will not be
4786 if (likely(update_next_balance))
4787 rq->next_balance = next_balance;
4791 * run_rebalance_domains is triggered when needed from the scheduler tick.
4792 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4793 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4795 static void run_rebalance_domains(struct softirq_action *h)
4797 int this_cpu = smp_processor_id();
4798 struct rq *this_rq = cpu_rq(this_cpu);
4799 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4800 CPU_IDLE : CPU_NOT_IDLE;
4802 rebalance_domains(this_cpu, idle);
4806 * If this cpu is the owner for idle load balancing, then do the
4807 * balancing on behalf of the other idle cpus whose ticks are
4810 if (this_rq->idle_at_tick &&
4811 atomic_read(&nohz.load_balancer) == this_cpu) {
4815 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4816 if (balance_cpu == this_cpu)
4820 * If this cpu gets work to do, stop the load balancing
4821 * work being done for other cpus. Next load
4822 * balancing owner will pick it up.
4827 rebalance_domains(balance_cpu, CPU_IDLE);
4829 rq = cpu_rq(balance_cpu);
4830 if (time_after(this_rq->next_balance, rq->next_balance))
4831 this_rq->next_balance = rq->next_balance;
4837 static inline int on_null_domain(int cpu)
4839 return !rcu_dereference(cpu_rq(cpu)->sd);
4843 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4845 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4846 * idle load balancing owner or decide to stop the periodic load balancing,
4847 * if the whole system is idle.
4849 static inline void trigger_load_balance(struct rq *rq, int cpu)
4853 * If we were in the nohz mode recently and busy at the current
4854 * scheduler tick, then check if we need to nominate new idle
4857 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4858 rq->in_nohz_recently = 0;
4860 if (atomic_read(&nohz.load_balancer) == cpu) {
4861 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4862 atomic_set(&nohz.load_balancer, -1);
4865 if (atomic_read(&nohz.load_balancer) == -1) {
4866 int ilb = find_new_ilb(cpu);
4868 if (ilb < nr_cpu_ids)
4874 * If this cpu is idle and doing idle load balancing for all the
4875 * cpus with ticks stopped, is it time for that to stop?
4877 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4878 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4884 * If this cpu is idle and the idle load balancing is done by
4885 * someone else, then no need raise the SCHED_SOFTIRQ
4887 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4888 cpumask_test_cpu(cpu, nohz.cpu_mask))
4891 /* Don't need to rebalance while attached to NULL domain */
4892 if (time_after_eq(jiffies, rq->next_balance) &&
4893 likely(!on_null_domain(cpu)))
4894 raise_softirq(SCHED_SOFTIRQ);
4897 #else /* CONFIG_SMP */
4900 * on UP we do not need to balance between CPUs:
4902 static inline void idle_balance(int cpu, struct rq *rq)
4908 DEFINE_PER_CPU(struct kernel_stat, kstat);
4910 EXPORT_PER_CPU_SYMBOL(kstat);
4913 * Return any ns on the sched_clock that have not yet been accounted in
4914 * @p in case that task is currently running.
4916 * Called with task_rq_lock() held on @rq.
4918 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4922 if (task_current(rq, p)) {
4923 update_rq_clock(rq);
4924 ns = rq->clock - p->se.exec_start;
4932 unsigned long long task_delta_exec(struct task_struct *p)
4934 unsigned long flags;
4938 rq = task_rq_lock(p, &flags);
4939 ns = do_task_delta_exec(p, rq);
4940 task_rq_unlock(rq, &flags);
4946 * Return accounted runtime for the task.
4947 * In case the task is currently running, return the runtime plus current's
4948 * pending runtime that have not been accounted yet.
4950 unsigned long long task_sched_runtime(struct task_struct *p)
4952 unsigned long flags;
4956 rq = task_rq_lock(p, &flags);
4957 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4958 task_rq_unlock(rq, &flags);
4964 * Return sum_exec_runtime for the thread group.
4965 * In case the task is currently running, return the sum plus current's
4966 * pending runtime that have not been accounted yet.
4968 * Note that the thread group might have other running tasks as well,
4969 * so the return value not includes other pending runtime that other
4970 * running tasks might have.
4972 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4974 struct task_cputime totals;
4975 unsigned long flags;
4979 rq = task_rq_lock(p, &flags);
4980 thread_group_cputime(p, &totals);
4981 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4982 task_rq_unlock(rq, &flags);
4988 * Account user cpu time to a process.
4989 * @p: the process that the cpu time gets accounted to
4990 * @cputime: the cpu time spent in user space since the last update
4991 * @cputime_scaled: cputime scaled by cpu frequency
4993 void account_user_time(struct task_struct *p, cputime_t cputime,
4994 cputime_t cputime_scaled)
4996 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4999 /* Add user time to process. */
5000 p->utime = cputime_add(p->utime, cputime);
5001 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5002 account_group_user_time(p, cputime);
5004 /* Add user time to cpustat. */
5005 tmp = cputime_to_cputime64(cputime);
5006 if (TASK_NICE(p) > 0)
5007 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5009 cpustat->user = cputime64_add(cpustat->user, tmp);
5011 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5012 /* Account for user time used */
5013 acct_update_integrals(p);
5017 * Account guest cpu time to a process.
5018 * @p: the process that the cpu time gets accounted to
5019 * @cputime: the cpu time spent in virtual machine since the last update
5020 * @cputime_scaled: cputime scaled by cpu frequency
5022 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5023 cputime_t cputime_scaled)
5026 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5028 tmp = cputime_to_cputime64(cputime);
5030 /* Add guest time to process. */
5031 p->utime = cputime_add(p->utime, cputime);
5032 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5033 account_group_user_time(p, cputime);
5034 p->gtime = cputime_add(p->gtime, cputime);
5036 /* Add guest time to cpustat. */
5037 cpustat->user = cputime64_add(cpustat->user, tmp);
5038 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5042 * Account system cpu time to a process.
5043 * @p: the process that the cpu time gets accounted to
5044 * @hardirq_offset: the offset to subtract from hardirq_count()
5045 * @cputime: the cpu time spent in kernel space since the last update
5046 * @cputime_scaled: cputime scaled by cpu frequency
5048 void account_system_time(struct task_struct *p, int hardirq_offset,
5049 cputime_t cputime, cputime_t cputime_scaled)
5051 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5054 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5055 account_guest_time(p, cputime, cputime_scaled);
5059 /* Add system time to process. */
5060 p->stime = cputime_add(p->stime, cputime);
5061 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5062 account_group_system_time(p, cputime);
5064 /* Add system time to cpustat. */
5065 tmp = cputime_to_cputime64(cputime);
5066 if (hardirq_count() - hardirq_offset)
5067 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5068 else if (softirq_count())
5069 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5071 cpustat->system = cputime64_add(cpustat->system, tmp);
5073 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5075 /* Account for system time used */
5076 acct_update_integrals(p);
5080 * Account for involuntary wait time.
5081 * @steal: the cpu time spent in involuntary wait
5083 void account_steal_time(cputime_t cputime)
5085 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5086 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5088 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5092 * Account for idle time.
5093 * @cputime: the cpu time spent in idle wait
5095 void account_idle_time(cputime_t cputime)
5097 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5098 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5099 struct rq *rq = this_rq();
5101 if (atomic_read(&rq->nr_iowait) > 0)
5102 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5104 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5107 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5110 * Account a single tick of cpu time.
5111 * @p: the process that the cpu time gets accounted to
5112 * @user_tick: indicates if the tick is a user or a system tick
5114 void account_process_tick(struct task_struct *p, int user_tick)
5116 cputime_t one_jiffy = jiffies_to_cputime(1);
5117 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5118 struct rq *rq = this_rq();
5121 account_user_time(p, one_jiffy, one_jiffy_scaled);
5122 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5123 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5126 account_idle_time(one_jiffy);
5130 * Account multiple ticks of steal time.
5131 * @p: the process from which the cpu time has been stolen
5132 * @ticks: number of stolen ticks
5134 void account_steal_ticks(unsigned long ticks)
5136 account_steal_time(jiffies_to_cputime(ticks));
5140 * Account multiple ticks of idle time.
5141 * @ticks: number of stolen ticks
5143 void account_idle_ticks(unsigned long ticks)
5145 account_idle_time(jiffies_to_cputime(ticks));
5151 * Use precise platform statistics if available:
5153 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5154 cputime_t task_utime(struct task_struct *p)
5159 cputime_t task_stime(struct task_struct *p)
5164 cputime_t task_utime(struct task_struct *p)
5166 clock_t utime = cputime_to_clock_t(p->utime),
5167 total = utime + cputime_to_clock_t(p->stime);
5171 * Use CFS's precise accounting:
5173 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5177 do_div(temp, total);
5179 utime = (clock_t)temp;
5181 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5182 return p->prev_utime;
5185 cputime_t task_stime(struct task_struct *p)
5190 * Use CFS's precise accounting. (we subtract utime from
5191 * the total, to make sure the total observed by userspace
5192 * grows monotonically - apps rely on that):
5194 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5195 cputime_to_clock_t(task_utime(p));
5198 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5200 return p->prev_stime;
5204 inline cputime_t task_gtime(struct task_struct *p)
5210 * This function gets called by the timer code, with HZ frequency.
5211 * We call it with interrupts disabled.
5213 * It also gets called by the fork code, when changing the parent's
5216 void scheduler_tick(void)
5218 int cpu = smp_processor_id();
5219 struct rq *rq = cpu_rq(cpu);
5220 struct task_struct *curr = rq->curr;
5224 spin_lock(&rq->lock);
5225 update_rq_clock(rq);
5226 update_cpu_load(rq);
5227 curr->sched_class->task_tick(rq, curr, 0);
5228 spin_unlock(&rq->lock);
5230 perf_counter_task_tick(curr, cpu);
5233 rq->idle_at_tick = idle_cpu(cpu);
5234 trigger_load_balance(rq, cpu);
5238 notrace unsigned long get_parent_ip(unsigned long addr)
5240 if (in_lock_functions(addr)) {
5241 addr = CALLER_ADDR2;
5242 if (in_lock_functions(addr))
5243 addr = CALLER_ADDR3;
5248 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5249 defined(CONFIG_PREEMPT_TRACER))
5251 void __kprobes add_preempt_count(int val)
5253 #ifdef CONFIG_DEBUG_PREEMPT
5257 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5260 preempt_count() += val;
5261 #ifdef CONFIG_DEBUG_PREEMPT
5263 * Spinlock count overflowing soon?
5265 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5268 if (preempt_count() == val)
5269 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5271 EXPORT_SYMBOL(add_preempt_count);
5273 void __kprobes sub_preempt_count(int val)
5275 #ifdef CONFIG_DEBUG_PREEMPT
5279 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5282 * Is the spinlock portion underflowing?
5284 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5285 !(preempt_count() & PREEMPT_MASK)))
5289 if (preempt_count() == val)
5290 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5291 preempt_count() -= val;
5293 EXPORT_SYMBOL(sub_preempt_count);
5298 * Print scheduling while atomic bug:
5300 static noinline void __schedule_bug(struct task_struct *prev)
5302 struct pt_regs *regs = get_irq_regs();
5304 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5305 prev->comm, prev->pid, preempt_count());
5307 debug_show_held_locks(prev);
5309 if (irqs_disabled())
5310 print_irqtrace_events(prev);
5319 * Various schedule()-time debugging checks and statistics:
5321 static inline void schedule_debug(struct task_struct *prev)
5324 * Test if we are atomic. Since do_exit() needs to call into
5325 * schedule() atomically, we ignore that path for now.
5326 * Otherwise, whine if we are scheduling when we should not be.
5328 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5329 __schedule_bug(prev);
5331 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5333 schedstat_inc(this_rq(), sched_count);
5334 #ifdef CONFIG_SCHEDSTATS
5335 if (unlikely(prev->lock_depth >= 0)) {
5336 schedstat_inc(this_rq(), bkl_count);
5337 schedstat_inc(prev, sched_info.bkl_count);
5342 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5344 if (prev->state == TASK_RUNNING) {
5345 u64 runtime = prev->se.sum_exec_runtime;
5347 runtime -= prev->se.prev_sum_exec_runtime;
5348 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5351 * In order to avoid avg_overlap growing stale when we are
5352 * indeed overlapping and hence not getting put to sleep, grow
5353 * the avg_overlap on preemption.
5355 * We use the average preemption runtime because that
5356 * correlates to the amount of cache footprint a task can
5359 update_avg(&prev->se.avg_overlap, runtime);
5361 prev->sched_class->put_prev_task(rq, prev);
5365 * Pick up the highest-prio task:
5367 static inline struct task_struct *
5368 pick_next_task(struct rq *rq)
5370 const struct sched_class *class;
5371 struct task_struct *p;
5374 * Optimization: we know that if all tasks are in
5375 * the fair class we can call that function directly:
5377 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5378 p = fair_sched_class.pick_next_task(rq);
5383 class = sched_class_highest;
5385 p = class->pick_next_task(rq);
5389 * Will never be NULL as the idle class always
5390 * returns a non-NULL p:
5392 class = class->next;
5397 * schedule() is the main scheduler function.
5399 asmlinkage void __sched schedule(void)
5401 struct task_struct *prev, *next;
5402 unsigned long *switch_count;
5408 cpu = smp_processor_id();
5412 switch_count = &prev->nivcsw;
5414 release_kernel_lock(prev);
5415 need_resched_nonpreemptible:
5417 schedule_debug(prev);
5419 if (sched_feat(HRTICK))
5422 spin_lock_irq(&rq->lock);
5423 update_rq_clock(rq);
5424 clear_tsk_need_resched(prev);
5426 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5427 if (unlikely(signal_pending_state(prev->state, prev)))
5428 prev->state = TASK_RUNNING;
5430 deactivate_task(rq, prev, 1);
5431 switch_count = &prev->nvcsw;
5434 pre_schedule(rq, prev);
5436 if (unlikely(!rq->nr_running))
5437 idle_balance(cpu, rq);
5439 put_prev_task(rq, prev);
5440 next = pick_next_task(rq);
5442 if (likely(prev != next)) {
5443 sched_info_switch(prev, next);
5444 perf_counter_task_sched_out(prev, next, cpu);
5450 context_switch(rq, prev, next); /* unlocks the rq */
5452 * the context switch might have flipped the stack from under
5453 * us, hence refresh the local variables.
5455 cpu = smp_processor_id();
5458 spin_unlock_irq(&rq->lock);
5462 if (unlikely(reacquire_kernel_lock(current) < 0))
5463 goto need_resched_nonpreemptible;
5465 preempt_enable_no_resched();
5469 EXPORT_SYMBOL(schedule);
5473 * Look out! "owner" is an entirely speculative pointer
5474 * access and not reliable.
5476 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5481 if (!sched_feat(OWNER_SPIN))
5484 #ifdef CONFIG_DEBUG_PAGEALLOC
5486 * Need to access the cpu field knowing that
5487 * DEBUG_PAGEALLOC could have unmapped it if
5488 * the mutex owner just released it and exited.
5490 if (probe_kernel_address(&owner->cpu, cpu))
5497 * Even if the access succeeded (likely case),
5498 * the cpu field may no longer be valid.
5500 if (cpu >= nr_cpumask_bits)
5504 * We need to validate that we can do a
5505 * get_cpu() and that we have the percpu area.
5507 if (!cpu_online(cpu))
5514 * Owner changed, break to re-assess state.
5516 if (lock->owner != owner)
5520 * Is that owner really running on that cpu?
5522 if (task_thread_info(rq->curr) != owner || need_resched())
5532 #ifdef CONFIG_PREEMPT
5534 * this is the entry point to schedule() from in-kernel preemption
5535 * off of preempt_enable. Kernel preemptions off return from interrupt
5536 * occur there and call schedule directly.
5538 asmlinkage void __sched preempt_schedule(void)
5540 struct thread_info *ti = current_thread_info();
5543 * If there is a non-zero preempt_count or interrupts are disabled,
5544 * we do not want to preempt the current task. Just return..
5546 if (likely(ti->preempt_count || irqs_disabled()))
5550 add_preempt_count(PREEMPT_ACTIVE);
5552 sub_preempt_count(PREEMPT_ACTIVE);
5555 * Check again in case we missed a preemption opportunity
5556 * between schedule and now.
5559 } while (need_resched());
5561 EXPORT_SYMBOL(preempt_schedule);
5564 * this is the entry point to schedule() from kernel preemption
5565 * off of irq context.
5566 * Note, that this is called and return with irqs disabled. This will
5567 * protect us against recursive calling from irq.
5569 asmlinkage void __sched preempt_schedule_irq(void)
5571 struct thread_info *ti = current_thread_info();
5573 /* Catch callers which need to be fixed */
5574 BUG_ON(ti->preempt_count || !irqs_disabled());
5577 add_preempt_count(PREEMPT_ACTIVE);
5580 local_irq_disable();
5581 sub_preempt_count(PREEMPT_ACTIVE);
5584 * Check again in case we missed a preemption opportunity
5585 * between schedule and now.
5588 } while (need_resched());
5591 #endif /* CONFIG_PREEMPT */
5593 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5596 return try_to_wake_up(curr->private, mode, sync);
5598 EXPORT_SYMBOL(default_wake_function);
5601 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5602 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5603 * number) then we wake all the non-exclusive tasks and one exclusive task.
5605 * There are circumstances in which we can try to wake a task which has already
5606 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5607 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5609 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5610 int nr_exclusive, int sync, void *key)
5612 wait_queue_t *curr, *next;
5614 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5615 unsigned flags = curr->flags;
5617 if (curr->func(curr, mode, sync, key) &&
5618 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5624 * __wake_up - wake up threads blocked on a waitqueue.
5626 * @mode: which threads
5627 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5628 * @key: is directly passed to the wakeup function
5630 * It may be assumed that this function implies a write memory barrier before
5631 * changing the task state if and only if any tasks are woken up.
5633 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5634 int nr_exclusive, void *key)
5636 unsigned long flags;
5638 spin_lock_irqsave(&q->lock, flags);
5639 __wake_up_common(q, mode, nr_exclusive, 0, key);
5640 spin_unlock_irqrestore(&q->lock, flags);
5642 EXPORT_SYMBOL(__wake_up);
5645 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5647 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5649 __wake_up_common(q, mode, 1, 0, NULL);
5652 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5654 __wake_up_common(q, mode, 1, 0, key);
5658 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5660 * @mode: which threads
5661 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5662 * @key: opaque value to be passed to wakeup targets
5664 * The sync wakeup differs that the waker knows that it will schedule
5665 * away soon, so while the target thread will be woken up, it will not
5666 * be migrated to another CPU - ie. the two threads are 'synchronized'
5667 * with each other. This can prevent needless bouncing between CPUs.
5669 * On UP it can prevent extra preemption.
5671 * It may be assumed that this function implies a write memory barrier before
5672 * changing the task state if and only if any tasks are woken up.
5674 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5675 int nr_exclusive, void *key)
5677 unsigned long flags;
5683 if (unlikely(!nr_exclusive))
5686 spin_lock_irqsave(&q->lock, flags);
5687 __wake_up_common(q, mode, nr_exclusive, sync, key);
5688 spin_unlock_irqrestore(&q->lock, flags);
5690 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5693 * __wake_up_sync - see __wake_up_sync_key()
5695 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5697 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5699 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5702 * complete: - signals a single thread waiting on this completion
5703 * @x: holds the state of this particular completion
5705 * This will wake up a single thread waiting on this completion. Threads will be
5706 * awakened in the same order in which they were queued.
5708 * See also complete_all(), wait_for_completion() and related routines.
5710 * It may be assumed that this function implies a write memory barrier before
5711 * changing the task state if and only if any tasks are woken up.
5713 void complete(struct completion *x)
5715 unsigned long flags;
5717 spin_lock_irqsave(&x->wait.lock, flags);
5719 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5720 spin_unlock_irqrestore(&x->wait.lock, flags);
5722 EXPORT_SYMBOL(complete);
5725 * complete_all: - signals all threads waiting on this completion
5726 * @x: holds the state of this particular completion
5728 * This will wake up all threads waiting on this particular completion event.
5730 * It may be assumed that this function implies a write memory barrier before
5731 * changing the task state if and only if any tasks are woken up.
5733 void complete_all(struct completion *x)
5735 unsigned long flags;
5737 spin_lock_irqsave(&x->wait.lock, flags);
5738 x->done += UINT_MAX/2;
5739 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5740 spin_unlock_irqrestore(&x->wait.lock, flags);
5742 EXPORT_SYMBOL(complete_all);
5744 static inline long __sched
5745 do_wait_for_common(struct completion *x, long timeout, int state)
5748 DECLARE_WAITQUEUE(wait, current);
5750 wait.flags |= WQ_FLAG_EXCLUSIVE;
5751 __add_wait_queue_tail(&x->wait, &wait);
5753 if (signal_pending_state(state, current)) {
5754 timeout = -ERESTARTSYS;
5757 __set_current_state(state);
5758 spin_unlock_irq(&x->wait.lock);
5759 timeout = schedule_timeout(timeout);
5760 spin_lock_irq(&x->wait.lock);
5761 } while (!x->done && timeout);
5762 __remove_wait_queue(&x->wait, &wait);
5767 return timeout ?: 1;
5771 wait_for_common(struct completion *x, long timeout, int state)
5775 spin_lock_irq(&x->wait.lock);
5776 timeout = do_wait_for_common(x, timeout, state);
5777 spin_unlock_irq(&x->wait.lock);
5782 * wait_for_completion: - waits for completion of a task
5783 * @x: holds the state of this particular completion
5785 * This waits to be signaled for completion of a specific task. It is NOT
5786 * interruptible and there is no timeout.
5788 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5789 * and interrupt capability. Also see complete().
5791 void __sched wait_for_completion(struct completion *x)
5793 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5795 EXPORT_SYMBOL(wait_for_completion);
5798 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5799 * @x: holds the state of this particular completion
5800 * @timeout: timeout value in jiffies
5802 * This waits for either a completion of a specific task to be signaled or for a
5803 * specified timeout to expire. The timeout is in jiffies. It is not
5806 unsigned long __sched
5807 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5809 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5811 EXPORT_SYMBOL(wait_for_completion_timeout);
5814 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5815 * @x: holds the state of this particular completion
5817 * This waits for completion of a specific task to be signaled. It is
5820 int __sched wait_for_completion_interruptible(struct completion *x)
5822 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5823 if (t == -ERESTARTSYS)
5827 EXPORT_SYMBOL(wait_for_completion_interruptible);
5830 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5831 * @x: holds the state of this particular completion
5832 * @timeout: timeout value in jiffies
5834 * This waits for either a completion of a specific task to be signaled or for a
5835 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5837 unsigned long __sched
5838 wait_for_completion_interruptible_timeout(struct completion *x,
5839 unsigned long timeout)
5841 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5843 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5846 * wait_for_completion_killable: - waits for completion of a task (killable)
5847 * @x: holds the state of this particular completion
5849 * This waits to be signaled for completion of a specific task. It can be
5850 * interrupted by a kill signal.
5852 int __sched wait_for_completion_killable(struct completion *x)
5854 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5855 if (t == -ERESTARTSYS)
5859 EXPORT_SYMBOL(wait_for_completion_killable);
5862 * try_wait_for_completion - try to decrement a completion without blocking
5863 * @x: completion structure
5865 * Returns: 0 if a decrement cannot be done without blocking
5866 * 1 if a decrement succeeded.
5868 * If a completion is being used as a counting completion,
5869 * attempt to decrement the counter without blocking. This
5870 * enables us to avoid waiting if the resource the completion
5871 * is protecting is not available.
5873 bool try_wait_for_completion(struct completion *x)
5877 spin_lock_irq(&x->wait.lock);
5882 spin_unlock_irq(&x->wait.lock);
5885 EXPORT_SYMBOL(try_wait_for_completion);
5888 * completion_done - Test to see if a completion has any waiters
5889 * @x: completion structure
5891 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5892 * 1 if there are no waiters.
5895 bool completion_done(struct completion *x)
5899 spin_lock_irq(&x->wait.lock);
5902 spin_unlock_irq(&x->wait.lock);
5905 EXPORT_SYMBOL(completion_done);
5908 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5910 unsigned long flags;
5913 init_waitqueue_entry(&wait, current);
5915 __set_current_state(state);
5917 spin_lock_irqsave(&q->lock, flags);
5918 __add_wait_queue(q, &wait);
5919 spin_unlock(&q->lock);
5920 timeout = schedule_timeout(timeout);
5921 spin_lock_irq(&q->lock);
5922 __remove_wait_queue(q, &wait);
5923 spin_unlock_irqrestore(&q->lock, flags);
5928 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5930 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5932 EXPORT_SYMBOL(interruptible_sleep_on);
5935 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5937 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5939 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5941 void __sched sleep_on(wait_queue_head_t *q)
5943 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5945 EXPORT_SYMBOL(sleep_on);
5947 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5949 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5951 EXPORT_SYMBOL(sleep_on_timeout);
5953 #ifdef CONFIG_RT_MUTEXES
5956 * rt_mutex_setprio - set the current priority of a task
5958 * @prio: prio value (kernel-internal form)
5960 * This function changes the 'effective' priority of a task. It does
5961 * not touch ->normal_prio like __setscheduler().
5963 * Used by the rt_mutex code to implement priority inheritance logic.
5965 void rt_mutex_setprio(struct task_struct *p, int prio)
5967 unsigned long flags;
5968 int oldprio, on_rq, running;
5970 const struct sched_class *prev_class = p->sched_class;
5972 BUG_ON(prio < 0 || prio > MAX_PRIO);
5974 rq = task_rq_lock(p, &flags);
5975 update_rq_clock(rq);
5978 on_rq = p->se.on_rq;
5979 running = task_current(rq, p);
5981 dequeue_task(rq, p, 0);
5983 p->sched_class->put_prev_task(rq, p);
5986 p->sched_class = &rt_sched_class;
5988 p->sched_class = &fair_sched_class;
5993 p->sched_class->set_curr_task(rq);
5995 enqueue_task(rq, p, 0);
5997 check_class_changed(rq, p, prev_class, oldprio, running);
5999 task_rq_unlock(rq, &flags);
6004 void set_user_nice(struct task_struct *p, long nice)
6006 int old_prio, delta, on_rq;
6007 unsigned long flags;
6010 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6013 * We have to be careful, if called from sys_setpriority(),
6014 * the task might be in the middle of scheduling on another CPU.
6016 rq = task_rq_lock(p, &flags);
6017 update_rq_clock(rq);
6019 * The RT priorities are set via sched_setscheduler(), but we still
6020 * allow the 'normal' nice value to be set - but as expected
6021 * it wont have any effect on scheduling until the task is
6022 * SCHED_FIFO/SCHED_RR:
6024 if (task_has_rt_policy(p)) {
6025 p->static_prio = NICE_TO_PRIO(nice);
6028 on_rq = p->se.on_rq;
6030 dequeue_task(rq, p, 0);
6032 p->static_prio = NICE_TO_PRIO(nice);
6035 p->prio = effective_prio(p);
6036 delta = p->prio - old_prio;
6039 enqueue_task(rq, p, 0);
6041 * If the task increased its priority or is running and
6042 * lowered its priority, then reschedule its CPU:
6044 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6045 resched_task(rq->curr);
6048 task_rq_unlock(rq, &flags);
6050 EXPORT_SYMBOL(set_user_nice);
6053 * can_nice - check if a task can reduce its nice value
6057 int can_nice(const struct task_struct *p, const int nice)
6059 /* convert nice value [19,-20] to rlimit style value [1,40] */
6060 int nice_rlim = 20 - nice;
6062 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6063 capable(CAP_SYS_NICE));
6066 #ifdef __ARCH_WANT_SYS_NICE
6069 * sys_nice - change the priority of the current process.
6070 * @increment: priority increment
6072 * sys_setpriority is a more generic, but much slower function that
6073 * does similar things.
6075 SYSCALL_DEFINE1(nice, int, increment)
6080 * Setpriority might change our priority at the same moment.
6081 * We don't have to worry. Conceptually one call occurs first
6082 * and we have a single winner.
6084 if (increment < -40)
6089 nice = TASK_NICE(current) + increment;
6095 if (increment < 0 && !can_nice(current, nice))
6098 retval = security_task_setnice(current, nice);
6102 set_user_nice(current, nice);
6109 * task_prio - return the priority value of a given task.
6110 * @p: the task in question.
6112 * This is the priority value as seen by users in /proc.
6113 * RT tasks are offset by -200. Normal tasks are centered
6114 * around 0, value goes from -16 to +15.
6116 int task_prio(const struct task_struct *p)
6118 return p->prio - MAX_RT_PRIO;
6122 * task_nice - return the nice value of a given task.
6123 * @p: the task in question.
6125 int task_nice(const struct task_struct *p)
6127 return TASK_NICE(p);
6129 EXPORT_SYMBOL(task_nice);
6132 * idle_cpu - is a given cpu idle currently?
6133 * @cpu: the processor in question.
6135 int idle_cpu(int cpu)
6137 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6141 * idle_task - return the idle task for a given cpu.
6142 * @cpu: the processor in question.
6144 struct task_struct *idle_task(int cpu)
6146 return cpu_rq(cpu)->idle;
6150 * find_process_by_pid - find a process with a matching PID value.
6151 * @pid: the pid in question.
6153 static struct task_struct *find_process_by_pid(pid_t pid)
6155 return pid ? find_task_by_vpid(pid) : current;
6158 /* Actually do priority change: must hold rq lock. */
6160 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6162 BUG_ON(p->se.on_rq);
6165 switch (p->policy) {
6169 p->sched_class = &fair_sched_class;
6173 p->sched_class = &rt_sched_class;
6177 p->rt_priority = prio;
6178 p->normal_prio = normal_prio(p);
6179 /* we are holding p->pi_lock already */
6180 p->prio = rt_mutex_getprio(p);
6185 * check the target process has a UID that matches the current process's
6187 static bool check_same_owner(struct task_struct *p)
6189 const struct cred *cred = current_cred(), *pcred;
6193 pcred = __task_cred(p);
6194 match = (cred->euid == pcred->euid ||
6195 cred->euid == pcred->uid);
6200 static int __sched_setscheduler(struct task_struct *p, int policy,
6201 struct sched_param *param, bool user)
6203 int retval, oldprio, oldpolicy = -1, on_rq, running;
6204 unsigned long flags;
6205 const struct sched_class *prev_class = p->sched_class;
6209 /* may grab non-irq protected spin_locks */
6210 BUG_ON(in_interrupt());
6212 /* double check policy once rq lock held */
6214 reset_on_fork = p->sched_reset_on_fork;
6215 policy = oldpolicy = p->policy;
6217 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6218 policy &= ~SCHED_RESET_ON_FORK;
6220 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6221 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6222 policy != SCHED_IDLE)
6227 * Valid priorities for SCHED_FIFO and SCHED_RR are
6228 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6229 * SCHED_BATCH and SCHED_IDLE is 0.
6231 if (param->sched_priority < 0 ||
6232 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6233 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6235 if (rt_policy(policy) != (param->sched_priority != 0))
6239 * Allow unprivileged RT tasks to decrease priority:
6241 if (user && !capable(CAP_SYS_NICE)) {
6242 if (rt_policy(policy)) {
6243 unsigned long rlim_rtprio;
6245 if (!lock_task_sighand(p, &flags))
6247 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6248 unlock_task_sighand(p, &flags);
6250 /* can't set/change the rt policy */
6251 if (policy != p->policy && !rlim_rtprio)
6254 /* can't increase priority */
6255 if (param->sched_priority > p->rt_priority &&
6256 param->sched_priority > rlim_rtprio)
6260 * Like positive nice levels, dont allow tasks to
6261 * move out of SCHED_IDLE either:
6263 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6266 /* can't change other user's priorities */
6267 if (!check_same_owner(p))
6270 /* Normal users shall not reset the sched_reset_on_fork flag */
6271 if (p->sched_reset_on_fork && !reset_on_fork)
6276 #ifdef CONFIG_RT_GROUP_SCHED
6278 * Do not allow realtime tasks into groups that have no runtime
6281 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6282 task_group(p)->rt_bandwidth.rt_runtime == 0)
6286 retval = security_task_setscheduler(p, policy, param);
6292 * make sure no PI-waiters arrive (or leave) while we are
6293 * changing the priority of the task:
6295 spin_lock_irqsave(&p->pi_lock, flags);
6297 * To be able to change p->policy safely, the apropriate
6298 * runqueue lock must be held.
6300 rq = __task_rq_lock(p);
6301 /* recheck policy now with rq lock held */
6302 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6303 policy = oldpolicy = -1;
6304 __task_rq_unlock(rq);
6305 spin_unlock_irqrestore(&p->pi_lock, flags);
6308 update_rq_clock(rq);
6309 on_rq = p->se.on_rq;
6310 running = task_current(rq, p);
6312 deactivate_task(rq, p, 0);
6314 p->sched_class->put_prev_task(rq, p);
6316 p->sched_reset_on_fork = reset_on_fork;
6319 __setscheduler(rq, p, policy, param->sched_priority);
6322 p->sched_class->set_curr_task(rq);
6324 activate_task(rq, p, 0);
6326 check_class_changed(rq, p, prev_class, oldprio, running);
6328 __task_rq_unlock(rq);
6329 spin_unlock_irqrestore(&p->pi_lock, flags);
6331 rt_mutex_adjust_pi(p);
6337 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6338 * @p: the task in question.
6339 * @policy: new policy.
6340 * @param: structure containing the new RT priority.
6342 * NOTE that the task may be already dead.
6344 int sched_setscheduler(struct task_struct *p, int policy,
6345 struct sched_param *param)
6347 return __sched_setscheduler(p, policy, param, true);
6349 EXPORT_SYMBOL_GPL(sched_setscheduler);
6352 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6353 * @p: the task in question.
6354 * @policy: new policy.
6355 * @param: structure containing the new RT priority.
6357 * Just like sched_setscheduler, only don't bother checking if the
6358 * current context has permission. For example, this is needed in
6359 * stop_machine(): we create temporary high priority worker threads,
6360 * but our caller might not have that capability.
6362 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6363 struct sched_param *param)
6365 return __sched_setscheduler(p, policy, param, false);
6369 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6371 struct sched_param lparam;
6372 struct task_struct *p;
6375 if (!param || pid < 0)
6377 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6382 p = find_process_by_pid(pid);
6384 retval = sched_setscheduler(p, policy, &lparam);
6391 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6392 * @pid: the pid in question.
6393 * @policy: new policy.
6394 * @param: structure containing the new RT priority.
6396 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6397 struct sched_param __user *, param)
6399 /* negative values for policy are not valid */
6403 return do_sched_setscheduler(pid, policy, param);
6407 * sys_sched_setparam - set/change the RT priority of a thread
6408 * @pid: the pid in question.
6409 * @param: structure containing the new RT priority.
6411 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6413 return do_sched_setscheduler(pid, -1, param);
6417 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6418 * @pid: the pid in question.
6420 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6422 struct task_struct *p;
6429 read_lock(&tasklist_lock);
6430 p = find_process_by_pid(pid);
6432 retval = security_task_getscheduler(p);
6435 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6437 read_unlock(&tasklist_lock);
6442 * sys_sched_getparam - get the RT priority of a thread
6443 * @pid: the pid in question.
6444 * @param: structure containing the RT priority.
6446 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6448 struct sched_param lp;
6449 struct task_struct *p;
6452 if (!param || pid < 0)
6455 read_lock(&tasklist_lock);
6456 p = find_process_by_pid(pid);
6461 retval = security_task_getscheduler(p);
6465 lp.sched_priority = p->rt_priority;
6466 read_unlock(&tasklist_lock);
6469 * This one might sleep, we cannot do it with a spinlock held ...
6471 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6476 read_unlock(&tasklist_lock);
6480 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6482 cpumask_var_t cpus_allowed, new_mask;
6483 struct task_struct *p;
6487 read_lock(&tasklist_lock);
6489 p = find_process_by_pid(pid);
6491 read_unlock(&tasklist_lock);
6497 * It is not safe to call set_cpus_allowed with the
6498 * tasklist_lock held. We will bump the task_struct's
6499 * usage count and then drop tasklist_lock.
6502 read_unlock(&tasklist_lock);
6504 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6508 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6510 goto out_free_cpus_allowed;
6513 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6516 retval = security_task_setscheduler(p, 0, NULL);
6520 cpuset_cpus_allowed(p, cpus_allowed);
6521 cpumask_and(new_mask, in_mask, cpus_allowed);
6523 retval = set_cpus_allowed_ptr(p, new_mask);
6526 cpuset_cpus_allowed(p, cpus_allowed);
6527 if (!cpumask_subset(new_mask, cpus_allowed)) {
6529 * We must have raced with a concurrent cpuset
6530 * update. Just reset the cpus_allowed to the
6531 * cpuset's cpus_allowed
6533 cpumask_copy(new_mask, cpus_allowed);
6538 free_cpumask_var(new_mask);
6539 out_free_cpus_allowed:
6540 free_cpumask_var(cpus_allowed);
6547 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6548 struct cpumask *new_mask)
6550 if (len < cpumask_size())
6551 cpumask_clear(new_mask);
6552 else if (len > cpumask_size())
6553 len = cpumask_size();
6555 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6559 * sys_sched_setaffinity - set the cpu affinity of a process
6560 * @pid: pid of the process
6561 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6562 * @user_mask_ptr: user-space pointer to the new cpu mask
6564 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6565 unsigned long __user *, user_mask_ptr)
6567 cpumask_var_t new_mask;
6570 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6573 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6575 retval = sched_setaffinity(pid, new_mask);
6576 free_cpumask_var(new_mask);
6580 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6582 struct task_struct *p;
6586 read_lock(&tasklist_lock);
6589 p = find_process_by_pid(pid);
6593 retval = security_task_getscheduler(p);
6597 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6600 read_unlock(&tasklist_lock);
6607 * sys_sched_getaffinity - get the cpu affinity of a process
6608 * @pid: pid of the process
6609 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6610 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6612 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6613 unsigned long __user *, user_mask_ptr)
6618 if (len < cpumask_size())
6621 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6624 ret = sched_getaffinity(pid, mask);
6626 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6629 ret = cpumask_size();
6631 free_cpumask_var(mask);
6637 * sys_sched_yield - yield the current processor to other threads.
6639 * This function yields the current CPU to other tasks. If there are no
6640 * other threads running on this CPU then this function will return.
6642 SYSCALL_DEFINE0(sched_yield)
6644 struct rq *rq = this_rq_lock();
6646 schedstat_inc(rq, yld_count);
6647 current->sched_class->yield_task(rq);
6650 * Since we are going to call schedule() anyway, there's
6651 * no need to preempt or enable interrupts:
6653 __release(rq->lock);
6654 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6655 _raw_spin_unlock(&rq->lock);
6656 preempt_enable_no_resched();
6663 static inline int should_resched(void)
6665 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6668 static void __cond_resched(void)
6670 add_preempt_count(PREEMPT_ACTIVE);
6672 sub_preempt_count(PREEMPT_ACTIVE);
6675 int __sched _cond_resched(void)
6677 if (should_resched()) {
6683 EXPORT_SYMBOL(_cond_resched);
6686 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6687 * call schedule, and on return reacquire the lock.
6689 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6690 * operations here to prevent schedule() from being called twice (once via
6691 * spin_unlock(), once by hand).
6693 int __cond_resched_lock(spinlock_t *lock)
6695 int resched = should_resched();
6698 if (spin_needbreak(lock) || resched) {
6709 EXPORT_SYMBOL(__cond_resched_lock);
6711 int __sched __cond_resched_softirq(void)
6713 BUG_ON(!in_softirq());
6715 if (should_resched()) {
6723 EXPORT_SYMBOL(__cond_resched_softirq);
6726 * yield - yield the current processor to other threads.
6728 * This is a shortcut for kernel-space yielding - it marks the
6729 * thread runnable and calls sys_sched_yield().
6731 void __sched yield(void)
6733 set_current_state(TASK_RUNNING);
6736 EXPORT_SYMBOL(yield);
6739 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6740 * that process accounting knows that this is a task in IO wait state.
6742 * But don't do that if it is a deliberate, throttling IO wait (this task
6743 * has set its backing_dev_info: the queue against which it should throttle)
6745 void __sched io_schedule(void)
6747 struct rq *rq = raw_rq();
6749 delayacct_blkio_start();
6750 atomic_inc(&rq->nr_iowait);
6752 atomic_dec(&rq->nr_iowait);
6753 delayacct_blkio_end();
6755 EXPORT_SYMBOL(io_schedule);
6757 long __sched io_schedule_timeout(long timeout)
6759 struct rq *rq = raw_rq();
6762 delayacct_blkio_start();
6763 atomic_inc(&rq->nr_iowait);
6764 ret = schedule_timeout(timeout);
6765 atomic_dec(&rq->nr_iowait);
6766 delayacct_blkio_end();
6771 * sys_sched_get_priority_max - return maximum RT priority.
6772 * @policy: scheduling class.
6774 * this syscall returns the maximum rt_priority that can be used
6775 * by a given scheduling class.
6777 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6784 ret = MAX_USER_RT_PRIO-1;
6796 * sys_sched_get_priority_min - return minimum RT priority.
6797 * @policy: scheduling class.
6799 * this syscall returns the minimum rt_priority that can be used
6800 * by a given scheduling class.
6802 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6820 * sys_sched_rr_get_interval - return the default timeslice of a process.
6821 * @pid: pid of the process.
6822 * @interval: userspace pointer to the timeslice value.
6824 * this syscall writes the default timeslice value of a given process
6825 * into the user-space timespec buffer. A value of '0' means infinity.
6827 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6828 struct timespec __user *, interval)
6830 struct task_struct *p;
6831 unsigned int time_slice;
6839 read_lock(&tasklist_lock);
6840 p = find_process_by_pid(pid);
6844 retval = security_task_getscheduler(p);
6849 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6850 * tasks that are on an otherwise idle runqueue:
6853 if (p->policy == SCHED_RR) {
6854 time_slice = DEF_TIMESLICE;
6855 } else if (p->policy != SCHED_FIFO) {
6856 struct sched_entity *se = &p->se;
6857 unsigned long flags;
6860 rq = task_rq_lock(p, &flags);
6861 if (rq->cfs.load.weight)
6862 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6863 task_rq_unlock(rq, &flags);
6865 read_unlock(&tasklist_lock);
6866 jiffies_to_timespec(time_slice, &t);
6867 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6871 read_unlock(&tasklist_lock);
6875 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6877 void sched_show_task(struct task_struct *p)
6879 unsigned long free = 0;
6882 state = p->state ? __ffs(p->state) + 1 : 0;
6883 printk(KERN_INFO "%-13.13s %c", p->comm,
6884 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6885 #if BITS_PER_LONG == 32
6886 if (state == TASK_RUNNING)
6887 printk(KERN_CONT " running ");
6889 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6891 if (state == TASK_RUNNING)
6892 printk(KERN_CONT " running task ");
6894 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6896 #ifdef CONFIG_DEBUG_STACK_USAGE
6897 free = stack_not_used(p);
6899 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6900 task_pid_nr(p), task_pid_nr(p->real_parent),
6901 (unsigned long)task_thread_info(p)->flags);
6903 show_stack(p, NULL);
6906 void show_state_filter(unsigned long state_filter)
6908 struct task_struct *g, *p;
6910 #if BITS_PER_LONG == 32
6912 " task PC stack pid father\n");
6915 " task PC stack pid father\n");
6917 read_lock(&tasklist_lock);
6918 do_each_thread(g, p) {
6920 * reset the NMI-timeout, listing all files on a slow
6921 * console might take alot of time:
6923 touch_nmi_watchdog();
6924 if (!state_filter || (p->state & state_filter))
6926 } while_each_thread(g, p);
6928 touch_all_softlockup_watchdogs();
6930 #ifdef CONFIG_SCHED_DEBUG
6931 sysrq_sched_debug_show();
6933 read_unlock(&tasklist_lock);
6935 * Only show locks if all tasks are dumped:
6937 if (state_filter == -1)
6938 debug_show_all_locks();
6941 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6943 idle->sched_class = &idle_sched_class;
6947 * init_idle - set up an idle thread for a given CPU
6948 * @idle: task in question
6949 * @cpu: cpu the idle task belongs to
6951 * NOTE: this function does not set the idle thread's NEED_RESCHED
6952 * flag, to make booting more robust.
6954 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6956 struct rq *rq = cpu_rq(cpu);
6957 unsigned long flags;
6959 spin_lock_irqsave(&rq->lock, flags);
6962 idle->se.exec_start = sched_clock();
6964 idle->prio = idle->normal_prio = MAX_PRIO;
6965 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6966 __set_task_cpu(idle, cpu);
6968 rq->curr = rq->idle = idle;
6969 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6972 spin_unlock_irqrestore(&rq->lock, flags);
6974 /* Set the preempt count _outside_ the spinlocks! */
6975 #if defined(CONFIG_PREEMPT)
6976 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6978 task_thread_info(idle)->preempt_count = 0;
6981 * The idle tasks have their own, simple scheduling class:
6983 idle->sched_class = &idle_sched_class;
6984 ftrace_graph_init_task(idle);
6988 * In a system that switches off the HZ timer nohz_cpu_mask
6989 * indicates which cpus entered this state. This is used
6990 * in the rcu update to wait only for active cpus. For system
6991 * which do not switch off the HZ timer nohz_cpu_mask should
6992 * always be CPU_BITS_NONE.
6994 cpumask_var_t nohz_cpu_mask;
6997 * Increase the granularity value when there are more CPUs,
6998 * because with more CPUs the 'effective latency' as visible
6999 * to users decreases. But the relationship is not linear,
7000 * so pick a second-best guess by going with the log2 of the
7003 * This idea comes from the SD scheduler of Con Kolivas:
7005 static inline void sched_init_granularity(void)
7007 unsigned int factor = 1 + ilog2(num_online_cpus());
7008 const unsigned long limit = 200000000;
7010 sysctl_sched_min_granularity *= factor;
7011 if (sysctl_sched_min_granularity > limit)
7012 sysctl_sched_min_granularity = limit;
7014 sysctl_sched_latency *= factor;
7015 if (sysctl_sched_latency > limit)
7016 sysctl_sched_latency = limit;
7018 sysctl_sched_wakeup_granularity *= factor;
7020 sysctl_sched_shares_ratelimit *= factor;
7025 * This is how migration works:
7027 * 1) we queue a struct migration_req structure in the source CPU's
7028 * runqueue and wake up that CPU's migration thread.
7029 * 2) we down() the locked semaphore => thread blocks.
7030 * 3) migration thread wakes up (implicitly it forces the migrated
7031 * thread off the CPU)
7032 * 4) it gets the migration request and checks whether the migrated
7033 * task is still in the wrong runqueue.
7034 * 5) if it's in the wrong runqueue then the migration thread removes
7035 * it and puts it into the right queue.
7036 * 6) migration thread up()s the semaphore.
7037 * 7) we wake up and the migration is done.
7041 * Change a given task's CPU affinity. Migrate the thread to a
7042 * proper CPU and schedule it away if the CPU it's executing on
7043 * is removed from the allowed bitmask.
7045 * NOTE: the caller must have a valid reference to the task, the
7046 * task must not exit() & deallocate itself prematurely. The
7047 * call is not atomic; no spinlocks may be held.
7049 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7051 struct migration_req req;
7052 unsigned long flags;
7056 rq = task_rq_lock(p, &flags);
7057 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7062 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7063 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7068 if (p->sched_class->set_cpus_allowed)
7069 p->sched_class->set_cpus_allowed(p, new_mask);
7071 cpumask_copy(&p->cpus_allowed, new_mask);
7072 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7075 /* Can the task run on the task's current CPU? If so, we're done */
7076 if (cpumask_test_cpu(task_cpu(p), new_mask))
7079 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7080 /* Need help from migration thread: drop lock and wait. */
7081 struct task_struct *mt = rq->migration_thread;
7083 get_task_struct(mt);
7084 task_rq_unlock(rq, &flags);
7085 wake_up_process(rq->migration_thread);
7086 put_task_struct(mt);
7087 wait_for_completion(&req.done);
7088 tlb_migrate_finish(p->mm);
7092 task_rq_unlock(rq, &flags);
7096 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7099 * Move (not current) task off this cpu, onto dest cpu. We're doing
7100 * this because either it can't run here any more (set_cpus_allowed()
7101 * away from this CPU, or CPU going down), or because we're
7102 * attempting to rebalance this task on exec (sched_exec).
7104 * So we race with normal scheduler movements, but that's OK, as long
7105 * as the task is no longer on this CPU.
7107 * Returns non-zero if task was successfully migrated.
7109 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7111 struct rq *rq_dest, *rq_src;
7114 if (unlikely(!cpu_active(dest_cpu)))
7117 rq_src = cpu_rq(src_cpu);
7118 rq_dest = cpu_rq(dest_cpu);
7120 double_rq_lock(rq_src, rq_dest);
7121 /* Already moved. */
7122 if (task_cpu(p) != src_cpu)
7124 /* Affinity changed (again). */
7125 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7128 on_rq = p->se.on_rq;
7130 deactivate_task(rq_src, p, 0);
7132 set_task_cpu(p, dest_cpu);
7134 activate_task(rq_dest, p, 0);
7135 check_preempt_curr(rq_dest, p, 0);
7140 double_rq_unlock(rq_src, rq_dest);
7145 * migration_thread - this is a highprio system thread that performs
7146 * thread migration by bumping thread off CPU then 'pushing' onto
7149 static int migration_thread(void *data)
7151 int cpu = (long)data;
7155 BUG_ON(rq->migration_thread != current);
7157 set_current_state(TASK_INTERRUPTIBLE);
7158 while (!kthread_should_stop()) {
7159 struct migration_req *req;
7160 struct list_head *head;
7162 spin_lock_irq(&rq->lock);
7164 if (cpu_is_offline(cpu)) {
7165 spin_unlock_irq(&rq->lock);
7169 if (rq->active_balance) {
7170 active_load_balance(rq, cpu);
7171 rq->active_balance = 0;
7174 head = &rq->migration_queue;
7176 if (list_empty(head)) {
7177 spin_unlock_irq(&rq->lock);
7179 set_current_state(TASK_INTERRUPTIBLE);
7182 req = list_entry(head->next, struct migration_req, list);
7183 list_del_init(head->next);
7185 spin_unlock(&rq->lock);
7186 __migrate_task(req->task, cpu, req->dest_cpu);
7189 complete(&req->done);
7191 __set_current_state(TASK_RUNNING);
7196 #ifdef CONFIG_HOTPLUG_CPU
7198 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7202 local_irq_disable();
7203 ret = __migrate_task(p, src_cpu, dest_cpu);
7209 * Figure out where task on dead CPU should go, use force if necessary.
7211 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7214 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7217 /* Look for allowed, online CPU in same node. */
7218 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7219 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7222 /* Any allowed, online CPU? */
7223 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7224 if (dest_cpu < nr_cpu_ids)
7227 /* No more Mr. Nice Guy. */
7228 if (dest_cpu >= nr_cpu_ids) {
7229 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7230 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7233 * Don't tell them about moving exiting tasks or
7234 * kernel threads (both mm NULL), since they never
7237 if (p->mm && printk_ratelimit()) {
7238 printk(KERN_INFO "process %d (%s) no "
7239 "longer affine to cpu%d\n",
7240 task_pid_nr(p), p->comm, dead_cpu);
7245 /* It can have affinity changed while we were choosing. */
7246 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7251 * While a dead CPU has no uninterruptible tasks queued at this point,
7252 * it might still have a nonzero ->nr_uninterruptible counter, because
7253 * for performance reasons the counter is not stricly tracking tasks to
7254 * their home CPUs. So we just add the counter to another CPU's counter,
7255 * to keep the global sum constant after CPU-down:
7257 static void migrate_nr_uninterruptible(struct rq *rq_src)
7259 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7260 unsigned long flags;
7262 local_irq_save(flags);
7263 double_rq_lock(rq_src, rq_dest);
7264 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7265 rq_src->nr_uninterruptible = 0;
7266 double_rq_unlock(rq_src, rq_dest);
7267 local_irq_restore(flags);
7270 /* Run through task list and migrate tasks from the dead cpu. */
7271 static void migrate_live_tasks(int src_cpu)
7273 struct task_struct *p, *t;
7275 read_lock(&tasklist_lock);
7277 do_each_thread(t, p) {
7281 if (task_cpu(p) == src_cpu)
7282 move_task_off_dead_cpu(src_cpu, p);
7283 } while_each_thread(t, p);
7285 read_unlock(&tasklist_lock);
7289 * Schedules idle task to be the next runnable task on current CPU.
7290 * It does so by boosting its priority to highest possible.
7291 * Used by CPU offline code.
7293 void sched_idle_next(void)
7295 int this_cpu = smp_processor_id();
7296 struct rq *rq = cpu_rq(this_cpu);
7297 struct task_struct *p = rq->idle;
7298 unsigned long flags;
7300 /* cpu has to be offline */
7301 BUG_ON(cpu_online(this_cpu));
7304 * Strictly not necessary since rest of the CPUs are stopped by now
7305 * and interrupts disabled on the current cpu.
7307 spin_lock_irqsave(&rq->lock, flags);
7309 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7311 update_rq_clock(rq);
7312 activate_task(rq, p, 0);
7314 spin_unlock_irqrestore(&rq->lock, flags);
7318 * Ensures that the idle task is using init_mm right before its cpu goes
7321 void idle_task_exit(void)
7323 struct mm_struct *mm = current->active_mm;
7325 BUG_ON(cpu_online(smp_processor_id()));
7328 switch_mm(mm, &init_mm, current);
7332 /* called under rq->lock with disabled interrupts */
7333 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7335 struct rq *rq = cpu_rq(dead_cpu);
7337 /* Must be exiting, otherwise would be on tasklist. */
7338 BUG_ON(!p->exit_state);
7340 /* Cannot have done final schedule yet: would have vanished. */
7341 BUG_ON(p->state == TASK_DEAD);
7346 * Drop lock around migration; if someone else moves it,
7347 * that's OK. No task can be added to this CPU, so iteration is
7350 spin_unlock_irq(&rq->lock);
7351 move_task_off_dead_cpu(dead_cpu, p);
7352 spin_lock_irq(&rq->lock);
7357 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7358 static void migrate_dead_tasks(unsigned int dead_cpu)
7360 struct rq *rq = cpu_rq(dead_cpu);
7361 struct task_struct *next;
7364 if (!rq->nr_running)
7366 update_rq_clock(rq);
7367 next = pick_next_task(rq);
7370 next->sched_class->put_prev_task(rq, next);
7371 migrate_dead(dead_cpu, next);
7377 * remove the tasks which were accounted by rq from calc_load_tasks.
7379 static void calc_global_load_remove(struct rq *rq)
7381 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7382 rq->calc_load_active = 0;
7384 #endif /* CONFIG_HOTPLUG_CPU */
7386 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7388 static struct ctl_table sd_ctl_dir[] = {
7390 .procname = "sched_domain",
7396 static struct ctl_table sd_ctl_root[] = {
7398 .ctl_name = CTL_KERN,
7399 .procname = "kernel",
7401 .child = sd_ctl_dir,
7406 static struct ctl_table *sd_alloc_ctl_entry(int n)
7408 struct ctl_table *entry =
7409 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7414 static void sd_free_ctl_entry(struct ctl_table **tablep)
7416 struct ctl_table *entry;
7419 * In the intermediate directories, both the child directory and
7420 * procname are dynamically allocated and could fail but the mode
7421 * will always be set. In the lowest directory the names are
7422 * static strings and all have proc handlers.
7424 for (entry = *tablep; entry->mode; entry++) {
7426 sd_free_ctl_entry(&entry->child);
7427 if (entry->proc_handler == NULL)
7428 kfree(entry->procname);
7436 set_table_entry(struct ctl_table *entry,
7437 const char *procname, void *data, int maxlen,
7438 mode_t mode, proc_handler *proc_handler)
7440 entry->procname = procname;
7442 entry->maxlen = maxlen;
7444 entry->proc_handler = proc_handler;
7447 static struct ctl_table *
7448 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7450 struct ctl_table *table = sd_alloc_ctl_entry(13);
7455 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7456 sizeof(long), 0644, proc_doulongvec_minmax);
7457 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7458 sizeof(long), 0644, proc_doulongvec_minmax);
7459 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7460 sizeof(int), 0644, proc_dointvec_minmax);
7461 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7462 sizeof(int), 0644, proc_dointvec_minmax);
7463 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7464 sizeof(int), 0644, proc_dointvec_minmax);
7465 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7466 sizeof(int), 0644, proc_dointvec_minmax);
7467 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7468 sizeof(int), 0644, proc_dointvec_minmax);
7469 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7470 sizeof(int), 0644, proc_dointvec_minmax);
7471 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7472 sizeof(int), 0644, proc_dointvec_minmax);
7473 set_table_entry(&table[9], "cache_nice_tries",
7474 &sd->cache_nice_tries,
7475 sizeof(int), 0644, proc_dointvec_minmax);
7476 set_table_entry(&table[10], "flags", &sd->flags,
7477 sizeof(int), 0644, proc_dointvec_minmax);
7478 set_table_entry(&table[11], "name", sd->name,
7479 CORENAME_MAX_SIZE, 0444, proc_dostring);
7480 /* &table[12] is terminator */
7485 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7487 struct ctl_table *entry, *table;
7488 struct sched_domain *sd;
7489 int domain_num = 0, i;
7492 for_each_domain(cpu, sd)
7494 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7499 for_each_domain(cpu, sd) {
7500 snprintf(buf, 32, "domain%d", i);
7501 entry->procname = kstrdup(buf, GFP_KERNEL);
7503 entry->child = sd_alloc_ctl_domain_table(sd);
7510 static struct ctl_table_header *sd_sysctl_header;
7511 static void register_sched_domain_sysctl(void)
7513 int i, cpu_num = num_online_cpus();
7514 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7517 WARN_ON(sd_ctl_dir[0].child);
7518 sd_ctl_dir[0].child = entry;
7523 for_each_online_cpu(i) {
7524 snprintf(buf, 32, "cpu%d", i);
7525 entry->procname = kstrdup(buf, GFP_KERNEL);
7527 entry->child = sd_alloc_ctl_cpu_table(i);
7531 WARN_ON(sd_sysctl_header);
7532 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7535 /* may be called multiple times per register */
7536 static void unregister_sched_domain_sysctl(void)
7538 if (sd_sysctl_header)
7539 unregister_sysctl_table(sd_sysctl_header);
7540 sd_sysctl_header = NULL;
7541 if (sd_ctl_dir[0].child)
7542 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7545 static void register_sched_domain_sysctl(void)
7548 static void unregister_sched_domain_sysctl(void)
7553 static void set_rq_online(struct rq *rq)
7556 const struct sched_class *class;
7558 cpumask_set_cpu(rq->cpu, rq->rd->online);
7561 for_each_class(class) {
7562 if (class->rq_online)
7563 class->rq_online(rq);
7568 static void set_rq_offline(struct rq *rq)
7571 const struct sched_class *class;
7573 for_each_class(class) {
7574 if (class->rq_offline)
7575 class->rq_offline(rq);
7578 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7584 * migration_call - callback that gets triggered when a CPU is added.
7585 * Here we can start up the necessary migration thread for the new CPU.
7587 static int __cpuinit
7588 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7590 struct task_struct *p;
7591 int cpu = (long)hcpu;
7592 unsigned long flags;
7597 case CPU_UP_PREPARE:
7598 case CPU_UP_PREPARE_FROZEN:
7599 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7602 kthread_bind(p, cpu);
7603 /* Must be high prio: stop_machine expects to yield to it. */
7604 rq = task_rq_lock(p, &flags);
7605 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7606 task_rq_unlock(rq, &flags);
7608 cpu_rq(cpu)->migration_thread = p;
7609 rq->calc_load_update = calc_load_update;
7613 case CPU_ONLINE_FROZEN:
7614 /* Strictly unnecessary, as first user will wake it. */
7615 wake_up_process(cpu_rq(cpu)->migration_thread);
7617 /* Update our root-domain */
7619 spin_lock_irqsave(&rq->lock, flags);
7621 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7625 spin_unlock_irqrestore(&rq->lock, flags);
7628 #ifdef CONFIG_HOTPLUG_CPU
7629 case CPU_UP_CANCELED:
7630 case CPU_UP_CANCELED_FROZEN:
7631 if (!cpu_rq(cpu)->migration_thread)
7633 /* Unbind it from offline cpu so it can run. Fall thru. */
7634 kthread_bind(cpu_rq(cpu)->migration_thread,
7635 cpumask_any(cpu_online_mask));
7636 kthread_stop(cpu_rq(cpu)->migration_thread);
7637 put_task_struct(cpu_rq(cpu)->migration_thread);
7638 cpu_rq(cpu)->migration_thread = NULL;
7642 case CPU_DEAD_FROZEN:
7643 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7644 migrate_live_tasks(cpu);
7646 kthread_stop(rq->migration_thread);
7647 put_task_struct(rq->migration_thread);
7648 rq->migration_thread = NULL;
7649 /* Idle task back to normal (off runqueue, low prio) */
7650 spin_lock_irq(&rq->lock);
7651 update_rq_clock(rq);
7652 deactivate_task(rq, rq->idle, 0);
7653 rq->idle->static_prio = MAX_PRIO;
7654 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7655 rq->idle->sched_class = &idle_sched_class;
7656 migrate_dead_tasks(cpu);
7657 spin_unlock_irq(&rq->lock);
7659 migrate_nr_uninterruptible(rq);
7660 BUG_ON(rq->nr_running != 0);
7661 calc_global_load_remove(rq);
7663 * No need to migrate the tasks: it was best-effort if
7664 * they didn't take sched_hotcpu_mutex. Just wake up
7667 spin_lock_irq(&rq->lock);
7668 while (!list_empty(&rq->migration_queue)) {
7669 struct migration_req *req;
7671 req = list_entry(rq->migration_queue.next,
7672 struct migration_req, list);
7673 list_del_init(&req->list);
7674 spin_unlock_irq(&rq->lock);
7675 complete(&req->done);
7676 spin_lock_irq(&rq->lock);
7678 spin_unlock_irq(&rq->lock);
7682 case CPU_DYING_FROZEN:
7683 /* Update our root-domain */
7685 spin_lock_irqsave(&rq->lock, flags);
7687 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7690 spin_unlock_irqrestore(&rq->lock, flags);
7698 * Register at high priority so that task migration (migrate_all_tasks)
7699 * happens before everything else. This has to be lower priority than
7700 * the notifier in the perf_counter subsystem, though.
7702 static struct notifier_block __cpuinitdata migration_notifier = {
7703 .notifier_call = migration_call,
7707 static int __init migration_init(void)
7709 void *cpu = (void *)(long)smp_processor_id();
7712 /* Start one for the boot CPU: */
7713 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7714 BUG_ON(err == NOTIFY_BAD);
7715 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7716 register_cpu_notifier(&migration_notifier);
7720 early_initcall(migration_init);
7725 #ifdef CONFIG_SCHED_DEBUG
7727 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7728 struct cpumask *groupmask)
7730 struct sched_group *group = sd->groups;
7733 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7734 cpumask_clear(groupmask);
7736 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7738 if (!(sd->flags & SD_LOAD_BALANCE)) {
7739 printk("does not load-balance\n");
7741 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7746 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7748 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7749 printk(KERN_ERR "ERROR: domain->span does not contain "
7752 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7753 printk(KERN_ERR "ERROR: domain->groups does not contain"
7757 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7761 printk(KERN_ERR "ERROR: group is NULL\n");
7765 if (!group->__cpu_power) {
7766 printk(KERN_CONT "\n");
7767 printk(KERN_ERR "ERROR: domain->cpu_power not "
7772 if (!cpumask_weight(sched_group_cpus(group))) {
7773 printk(KERN_CONT "\n");
7774 printk(KERN_ERR "ERROR: empty group\n");
7778 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7779 printk(KERN_CONT "\n");
7780 printk(KERN_ERR "ERROR: repeated CPUs\n");
7784 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7786 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7788 printk(KERN_CONT " %s", str);
7789 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7790 printk(KERN_CONT " (__cpu_power = %d)",
7791 group->__cpu_power);
7794 group = group->next;
7795 } while (group != sd->groups);
7796 printk(KERN_CONT "\n");
7798 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7799 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7802 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7803 printk(KERN_ERR "ERROR: parent span is not a superset "
7804 "of domain->span\n");
7808 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7810 cpumask_var_t groupmask;
7814 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7818 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7820 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7821 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7826 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7833 free_cpumask_var(groupmask);
7835 #else /* !CONFIG_SCHED_DEBUG */
7836 # define sched_domain_debug(sd, cpu) do { } while (0)
7837 #endif /* CONFIG_SCHED_DEBUG */
7839 static int sd_degenerate(struct sched_domain *sd)
7841 if (cpumask_weight(sched_domain_span(sd)) == 1)
7844 /* Following flags need at least 2 groups */
7845 if (sd->flags & (SD_LOAD_BALANCE |
7846 SD_BALANCE_NEWIDLE |
7850 SD_SHARE_PKG_RESOURCES)) {
7851 if (sd->groups != sd->groups->next)
7855 /* Following flags don't use groups */
7856 if (sd->flags & (SD_WAKE_IDLE |
7865 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7867 unsigned long cflags = sd->flags, pflags = parent->flags;
7869 if (sd_degenerate(parent))
7872 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7875 /* Does parent contain flags not in child? */
7876 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7877 if (cflags & SD_WAKE_AFFINE)
7878 pflags &= ~SD_WAKE_BALANCE;
7879 /* Flags needing groups don't count if only 1 group in parent */
7880 if (parent->groups == parent->groups->next) {
7881 pflags &= ~(SD_LOAD_BALANCE |
7882 SD_BALANCE_NEWIDLE |
7886 SD_SHARE_PKG_RESOURCES);
7887 if (nr_node_ids == 1)
7888 pflags &= ~SD_SERIALIZE;
7890 if (~cflags & pflags)
7896 static void free_rootdomain(struct root_domain *rd)
7898 cpupri_cleanup(&rd->cpupri);
7900 free_cpumask_var(rd->rto_mask);
7901 free_cpumask_var(rd->online);
7902 free_cpumask_var(rd->span);
7906 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7908 struct root_domain *old_rd = NULL;
7909 unsigned long flags;
7911 spin_lock_irqsave(&rq->lock, flags);
7916 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7919 cpumask_clear_cpu(rq->cpu, old_rd->span);
7922 * If we dont want to free the old_rt yet then
7923 * set old_rd to NULL to skip the freeing later
7926 if (!atomic_dec_and_test(&old_rd->refcount))
7930 atomic_inc(&rd->refcount);
7933 cpumask_set_cpu(rq->cpu, rd->span);
7934 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7937 spin_unlock_irqrestore(&rq->lock, flags);
7940 free_rootdomain(old_rd);
7943 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7945 gfp_t gfp = GFP_KERNEL;
7947 memset(rd, 0, sizeof(*rd));
7952 if (!alloc_cpumask_var(&rd->span, gfp))
7954 if (!alloc_cpumask_var(&rd->online, gfp))
7956 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7959 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7964 free_cpumask_var(rd->rto_mask);
7966 free_cpumask_var(rd->online);
7968 free_cpumask_var(rd->span);
7973 static void init_defrootdomain(void)
7975 init_rootdomain(&def_root_domain, true);
7977 atomic_set(&def_root_domain.refcount, 1);
7980 static struct root_domain *alloc_rootdomain(void)
7982 struct root_domain *rd;
7984 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7988 if (init_rootdomain(rd, false) != 0) {
7997 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7998 * hold the hotplug lock.
8001 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8003 struct rq *rq = cpu_rq(cpu);
8004 struct sched_domain *tmp;
8006 /* Remove the sched domains which do not contribute to scheduling. */
8007 for (tmp = sd; tmp; ) {
8008 struct sched_domain *parent = tmp->parent;
8012 if (sd_parent_degenerate(tmp, parent)) {
8013 tmp->parent = parent->parent;
8015 parent->parent->child = tmp;
8020 if (sd && sd_degenerate(sd)) {
8026 sched_domain_debug(sd, cpu);
8028 rq_attach_root(rq, rd);
8029 rcu_assign_pointer(rq->sd, sd);
8032 /* cpus with isolated domains */
8033 static cpumask_var_t cpu_isolated_map;
8035 /* Setup the mask of cpus configured for isolated domains */
8036 static int __init isolated_cpu_setup(char *str)
8038 cpulist_parse(str, cpu_isolated_map);
8042 __setup("isolcpus=", isolated_cpu_setup);
8045 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8046 * to a function which identifies what group(along with sched group) a CPU
8047 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8048 * (due to the fact that we keep track of groups covered with a struct cpumask).
8050 * init_sched_build_groups will build a circular linked list of the groups
8051 * covered by the given span, and will set each group's ->cpumask correctly,
8052 * and ->cpu_power to 0.
8055 init_sched_build_groups(const struct cpumask *span,
8056 const struct cpumask *cpu_map,
8057 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8058 struct sched_group **sg,
8059 struct cpumask *tmpmask),
8060 struct cpumask *covered, struct cpumask *tmpmask)
8062 struct sched_group *first = NULL, *last = NULL;
8065 cpumask_clear(covered);
8067 for_each_cpu(i, span) {
8068 struct sched_group *sg;
8069 int group = group_fn(i, cpu_map, &sg, tmpmask);
8072 if (cpumask_test_cpu(i, covered))
8075 cpumask_clear(sched_group_cpus(sg));
8076 sg->__cpu_power = 0;
8078 for_each_cpu(j, span) {
8079 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8082 cpumask_set_cpu(j, covered);
8083 cpumask_set_cpu(j, sched_group_cpus(sg));
8094 #define SD_NODES_PER_DOMAIN 16
8099 * find_next_best_node - find the next node to include in a sched_domain
8100 * @node: node whose sched_domain we're building
8101 * @used_nodes: nodes already in the sched_domain
8103 * Find the next node to include in a given scheduling domain. Simply
8104 * finds the closest node not already in the @used_nodes map.
8106 * Should use nodemask_t.
8108 static int find_next_best_node(int node, nodemask_t *used_nodes)
8110 int i, n, val, min_val, best_node = 0;
8114 for (i = 0; i < nr_node_ids; i++) {
8115 /* Start at @node */
8116 n = (node + i) % nr_node_ids;
8118 if (!nr_cpus_node(n))
8121 /* Skip already used nodes */
8122 if (node_isset(n, *used_nodes))
8125 /* Simple min distance search */
8126 val = node_distance(node, n);
8128 if (val < min_val) {
8134 node_set(best_node, *used_nodes);
8139 * sched_domain_node_span - get a cpumask for a node's sched_domain
8140 * @node: node whose cpumask we're constructing
8141 * @span: resulting cpumask
8143 * Given a node, construct a good cpumask for its sched_domain to span. It
8144 * should be one that prevents unnecessary balancing, but also spreads tasks
8147 static void sched_domain_node_span(int node, struct cpumask *span)
8149 nodemask_t used_nodes;
8152 cpumask_clear(span);
8153 nodes_clear(used_nodes);
8155 cpumask_or(span, span, cpumask_of_node(node));
8156 node_set(node, used_nodes);
8158 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8159 int next_node = find_next_best_node(node, &used_nodes);
8161 cpumask_or(span, span, cpumask_of_node(next_node));
8164 #endif /* CONFIG_NUMA */
8166 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8169 * The cpus mask in sched_group and sched_domain hangs off the end.
8171 * ( See the the comments in include/linux/sched.h:struct sched_group
8172 * and struct sched_domain. )
8174 struct static_sched_group {
8175 struct sched_group sg;
8176 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8179 struct static_sched_domain {
8180 struct sched_domain sd;
8181 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8185 * SMT sched-domains:
8187 #ifdef CONFIG_SCHED_SMT
8188 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8189 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8192 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8193 struct sched_group **sg, struct cpumask *unused)
8196 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8199 #endif /* CONFIG_SCHED_SMT */
8202 * multi-core sched-domains:
8204 #ifdef CONFIG_SCHED_MC
8205 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8206 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8207 #endif /* CONFIG_SCHED_MC */
8209 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8211 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8212 struct sched_group **sg, struct cpumask *mask)
8216 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8217 group = cpumask_first(mask);
8219 *sg = &per_cpu(sched_group_core, group).sg;
8222 #elif defined(CONFIG_SCHED_MC)
8224 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8225 struct sched_group **sg, struct cpumask *unused)
8228 *sg = &per_cpu(sched_group_core, cpu).sg;
8233 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8234 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8237 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8238 struct sched_group **sg, struct cpumask *mask)
8241 #ifdef CONFIG_SCHED_MC
8242 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8243 group = cpumask_first(mask);
8244 #elif defined(CONFIG_SCHED_SMT)
8245 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8246 group = cpumask_first(mask);
8251 *sg = &per_cpu(sched_group_phys, group).sg;
8257 * The init_sched_build_groups can't handle what we want to do with node
8258 * groups, so roll our own. Now each node has its own list of groups which
8259 * gets dynamically allocated.
8261 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8262 static struct sched_group ***sched_group_nodes_bycpu;
8264 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8265 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8267 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8268 struct sched_group **sg,
8269 struct cpumask *nodemask)
8273 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8274 group = cpumask_first(nodemask);
8277 *sg = &per_cpu(sched_group_allnodes, group).sg;
8281 static void init_numa_sched_groups_power(struct sched_group *group_head)
8283 struct sched_group *sg = group_head;
8289 for_each_cpu(j, sched_group_cpus(sg)) {
8290 struct sched_domain *sd;
8292 sd = &per_cpu(phys_domains, j).sd;
8293 if (j != group_first_cpu(sd->groups)) {
8295 * Only add "power" once for each
8301 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8304 } while (sg != group_head);
8306 #endif /* CONFIG_NUMA */
8309 /* Free memory allocated for various sched_group structures */
8310 static void free_sched_groups(const struct cpumask *cpu_map,
8311 struct cpumask *nodemask)
8315 for_each_cpu(cpu, cpu_map) {
8316 struct sched_group **sched_group_nodes
8317 = sched_group_nodes_bycpu[cpu];
8319 if (!sched_group_nodes)
8322 for (i = 0; i < nr_node_ids; i++) {
8323 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8325 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8326 if (cpumask_empty(nodemask))
8336 if (oldsg != sched_group_nodes[i])
8339 kfree(sched_group_nodes);
8340 sched_group_nodes_bycpu[cpu] = NULL;
8343 #else /* !CONFIG_NUMA */
8344 static void free_sched_groups(const struct cpumask *cpu_map,
8345 struct cpumask *nodemask)
8348 #endif /* CONFIG_NUMA */
8351 * Initialize sched groups cpu_power.
8353 * cpu_power indicates the capacity of sched group, which is used while
8354 * distributing the load between different sched groups in a sched domain.
8355 * Typically cpu_power for all the groups in a sched domain will be same unless
8356 * there are asymmetries in the topology. If there are asymmetries, group
8357 * having more cpu_power will pickup more load compared to the group having
8360 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8361 * the maximum number of tasks a group can handle in the presence of other idle
8362 * or lightly loaded groups in the same sched domain.
8364 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8366 struct sched_domain *child;
8367 struct sched_group *group;
8369 WARN_ON(!sd || !sd->groups);
8371 if (cpu != group_first_cpu(sd->groups))
8376 sd->groups->__cpu_power = 0;
8379 * For perf policy, if the groups in child domain share resources
8380 * (for example cores sharing some portions of the cache hierarchy
8381 * or SMT), then set this domain groups cpu_power such that each group
8382 * can handle only one task, when there are other idle groups in the
8383 * same sched domain.
8385 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8387 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
8388 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
8393 * add cpu_power of each child group to this groups cpu_power
8395 group = child->groups;
8397 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8398 group = group->next;
8399 } while (group != child->groups);
8403 * Initializers for schedule domains
8404 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8407 #ifdef CONFIG_SCHED_DEBUG
8408 # define SD_INIT_NAME(sd, type) sd->name = #type
8410 # define SD_INIT_NAME(sd, type) do { } while (0)
8413 #define SD_INIT(sd, type) sd_init_##type(sd)
8415 #define SD_INIT_FUNC(type) \
8416 static noinline void sd_init_##type(struct sched_domain *sd) \
8418 memset(sd, 0, sizeof(*sd)); \
8419 *sd = SD_##type##_INIT; \
8420 sd->level = SD_LV_##type; \
8421 SD_INIT_NAME(sd, type); \
8426 SD_INIT_FUNC(ALLNODES)
8429 #ifdef CONFIG_SCHED_SMT
8430 SD_INIT_FUNC(SIBLING)
8432 #ifdef CONFIG_SCHED_MC
8436 static int default_relax_domain_level = -1;
8438 static int __init setup_relax_domain_level(char *str)
8442 val = simple_strtoul(str, NULL, 0);
8443 if (val < SD_LV_MAX)
8444 default_relax_domain_level = val;
8448 __setup("relax_domain_level=", setup_relax_domain_level);
8450 static void set_domain_attribute(struct sched_domain *sd,
8451 struct sched_domain_attr *attr)
8455 if (!attr || attr->relax_domain_level < 0) {
8456 if (default_relax_domain_level < 0)
8459 request = default_relax_domain_level;
8461 request = attr->relax_domain_level;
8462 if (request < sd->level) {
8463 /* turn off idle balance on this domain */
8464 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8466 /* turn on idle balance on this domain */
8467 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8472 * Build sched domains for a given set of cpus and attach the sched domains
8473 * to the individual cpus
8475 static int __build_sched_domains(const struct cpumask *cpu_map,
8476 struct sched_domain_attr *attr)
8478 int i, err = -ENOMEM;
8479 struct root_domain *rd;
8480 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
8483 cpumask_var_t domainspan, covered, notcovered;
8484 struct sched_group **sched_group_nodes = NULL;
8485 int sd_allnodes = 0;
8487 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
8489 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
8490 goto free_domainspan;
8491 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
8495 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
8496 goto free_notcovered;
8497 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
8499 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
8500 goto free_this_sibling_map;
8501 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
8502 goto free_this_core_map;
8503 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
8504 goto free_send_covered;
8508 * Allocate the per-node list of sched groups
8510 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
8512 if (!sched_group_nodes) {
8513 printk(KERN_WARNING "Can not alloc sched group node list\n");
8518 rd = alloc_rootdomain();
8520 printk(KERN_WARNING "Cannot alloc root domain\n");
8521 goto free_sched_groups;
8525 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
8529 * Set up domains for cpus specified by the cpu_map.
8531 for_each_cpu(i, cpu_map) {
8532 struct sched_domain *sd = NULL, *p;
8534 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
8537 if (cpumask_weight(cpu_map) >
8538 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
8539 sd = &per_cpu(allnodes_domains, i).sd;
8540 SD_INIT(sd, ALLNODES);
8541 set_domain_attribute(sd, attr);
8542 cpumask_copy(sched_domain_span(sd), cpu_map);
8543 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
8549 sd = &per_cpu(node_domains, i).sd;
8551 set_domain_attribute(sd, attr);
8552 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8556 cpumask_and(sched_domain_span(sd),
8557 sched_domain_span(sd), cpu_map);
8561 sd = &per_cpu(phys_domains, i).sd;
8563 set_domain_attribute(sd, attr);
8564 cpumask_copy(sched_domain_span(sd), nodemask);
8568 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
8570 #ifdef CONFIG_SCHED_MC
8572 sd = &per_cpu(core_domains, i).sd;
8574 set_domain_attribute(sd, attr);
8575 cpumask_and(sched_domain_span(sd), cpu_map,
8576 cpu_coregroup_mask(i));
8579 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
8582 #ifdef CONFIG_SCHED_SMT
8584 sd = &per_cpu(cpu_domains, i).sd;
8585 SD_INIT(sd, SIBLING);
8586 set_domain_attribute(sd, attr);
8587 cpumask_and(sched_domain_span(sd),
8588 topology_thread_cpumask(i), cpu_map);
8591 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8595 #ifdef CONFIG_SCHED_SMT
8596 /* Set up CPU (sibling) groups */
8597 for_each_cpu(i, cpu_map) {
8598 cpumask_and(this_sibling_map,
8599 topology_thread_cpumask(i), cpu_map);
8600 if (i != cpumask_first(this_sibling_map))
8603 init_sched_build_groups(this_sibling_map, cpu_map,
8605 send_covered, tmpmask);
8609 #ifdef CONFIG_SCHED_MC
8610 /* Set up multi-core groups */
8611 for_each_cpu(i, cpu_map) {
8612 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8613 if (i != cpumask_first(this_core_map))
8616 init_sched_build_groups(this_core_map, cpu_map,
8618 send_covered, tmpmask);
8622 /* Set up physical groups */
8623 for (i = 0; i < nr_node_ids; i++) {
8624 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8625 if (cpumask_empty(nodemask))
8628 init_sched_build_groups(nodemask, cpu_map,
8630 send_covered, tmpmask);
8634 /* Set up node groups */
8636 init_sched_build_groups(cpu_map, cpu_map,
8637 &cpu_to_allnodes_group,
8638 send_covered, tmpmask);
8641 for (i = 0; i < nr_node_ids; i++) {
8642 /* Set up node groups */
8643 struct sched_group *sg, *prev;
8646 cpumask_clear(covered);
8647 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8648 if (cpumask_empty(nodemask)) {
8649 sched_group_nodes[i] = NULL;
8653 sched_domain_node_span(i, domainspan);
8654 cpumask_and(domainspan, domainspan, cpu_map);
8656 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8659 printk(KERN_WARNING "Can not alloc domain group for "
8663 sched_group_nodes[i] = sg;
8664 for_each_cpu(j, nodemask) {
8665 struct sched_domain *sd;
8667 sd = &per_cpu(node_domains, j).sd;
8670 sg->__cpu_power = 0;
8671 cpumask_copy(sched_group_cpus(sg), nodemask);
8673 cpumask_or(covered, covered, nodemask);
8676 for (j = 0; j < nr_node_ids; j++) {
8677 int n = (i + j) % nr_node_ids;
8679 cpumask_complement(notcovered, covered);
8680 cpumask_and(tmpmask, notcovered, cpu_map);
8681 cpumask_and(tmpmask, tmpmask, domainspan);
8682 if (cpumask_empty(tmpmask))
8685 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8686 if (cpumask_empty(tmpmask))
8689 sg = kmalloc_node(sizeof(struct sched_group) +
8694 "Can not alloc domain group for node %d\n", j);
8697 sg->__cpu_power = 0;
8698 cpumask_copy(sched_group_cpus(sg), tmpmask);
8699 sg->next = prev->next;
8700 cpumask_or(covered, covered, tmpmask);
8707 /* Calculate CPU power for physical packages and nodes */
8708 #ifdef CONFIG_SCHED_SMT
8709 for_each_cpu(i, cpu_map) {
8710 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8712 init_sched_groups_power(i, sd);
8715 #ifdef CONFIG_SCHED_MC
8716 for_each_cpu(i, cpu_map) {
8717 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8719 init_sched_groups_power(i, sd);
8723 for_each_cpu(i, cpu_map) {
8724 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8726 init_sched_groups_power(i, sd);
8730 for (i = 0; i < nr_node_ids; i++)
8731 init_numa_sched_groups_power(sched_group_nodes[i]);
8734 struct sched_group *sg;
8736 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8738 init_numa_sched_groups_power(sg);
8742 /* Attach the domains */
8743 for_each_cpu(i, cpu_map) {
8744 struct sched_domain *sd;
8745 #ifdef CONFIG_SCHED_SMT
8746 sd = &per_cpu(cpu_domains, i).sd;
8747 #elif defined(CONFIG_SCHED_MC)
8748 sd = &per_cpu(core_domains, i).sd;
8750 sd = &per_cpu(phys_domains, i).sd;
8752 cpu_attach_domain(sd, rd, i);
8758 free_cpumask_var(tmpmask);
8760 free_cpumask_var(send_covered);
8762 free_cpumask_var(this_core_map);
8763 free_this_sibling_map:
8764 free_cpumask_var(this_sibling_map);
8766 free_cpumask_var(nodemask);
8769 free_cpumask_var(notcovered);
8771 free_cpumask_var(covered);
8773 free_cpumask_var(domainspan);
8780 kfree(sched_group_nodes);
8786 free_sched_groups(cpu_map, tmpmask);
8787 free_rootdomain(rd);
8792 static int build_sched_domains(const struct cpumask *cpu_map)
8794 return __build_sched_domains(cpu_map, NULL);
8797 static struct cpumask *doms_cur; /* current sched domains */
8798 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8799 static struct sched_domain_attr *dattr_cur;
8800 /* attribues of custom domains in 'doms_cur' */
8803 * Special case: If a kmalloc of a doms_cur partition (array of
8804 * cpumask) fails, then fallback to a single sched domain,
8805 * as determined by the single cpumask fallback_doms.
8807 static cpumask_var_t fallback_doms;
8810 * arch_update_cpu_topology lets virtualized architectures update the
8811 * cpu core maps. It is supposed to return 1 if the topology changed
8812 * or 0 if it stayed the same.
8814 int __attribute__((weak)) arch_update_cpu_topology(void)
8820 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8821 * For now this just excludes isolated cpus, but could be used to
8822 * exclude other special cases in the future.
8824 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8828 arch_update_cpu_topology();
8830 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8832 doms_cur = fallback_doms;
8833 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8835 err = build_sched_domains(doms_cur);
8836 register_sched_domain_sysctl();
8841 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8842 struct cpumask *tmpmask)
8844 free_sched_groups(cpu_map, tmpmask);
8848 * Detach sched domains from a group of cpus specified in cpu_map
8849 * These cpus will now be attached to the NULL domain
8851 static void detach_destroy_domains(const struct cpumask *cpu_map)
8853 /* Save because hotplug lock held. */
8854 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8857 for_each_cpu(i, cpu_map)
8858 cpu_attach_domain(NULL, &def_root_domain, i);
8859 synchronize_sched();
8860 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8863 /* handle null as "default" */
8864 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8865 struct sched_domain_attr *new, int idx_new)
8867 struct sched_domain_attr tmp;
8874 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8875 new ? (new + idx_new) : &tmp,
8876 sizeof(struct sched_domain_attr));
8880 * Partition sched domains as specified by the 'ndoms_new'
8881 * cpumasks in the array doms_new[] of cpumasks. This compares
8882 * doms_new[] to the current sched domain partitioning, doms_cur[].
8883 * It destroys each deleted domain and builds each new domain.
8885 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8886 * The masks don't intersect (don't overlap.) We should setup one
8887 * sched domain for each mask. CPUs not in any of the cpumasks will
8888 * not be load balanced. If the same cpumask appears both in the
8889 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8892 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8893 * ownership of it and will kfree it when done with it. If the caller
8894 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8895 * ndoms_new == 1, and partition_sched_domains() will fallback to
8896 * the single partition 'fallback_doms', it also forces the domains
8899 * If doms_new == NULL it will be replaced with cpu_online_mask.
8900 * ndoms_new == 0 is a special case for destroying existing domains,
8901 * and it will not create the default domain.
8903 * Call with hotplug lock held
8905 /* FIXME: Change to struct cpumask *doms_new[] */
8906 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8907 struct sched_domain_attr *dattr_new)
8912 mutex_lock(&sched_domains_mutex);
8914 /* always unregister in case we don't destroy any domains */
8915 unregister_sched_domain_sysctl();
8917 /* Let architecture update cpu core mappings. */
8918 new_topology = arch_update_cpu_topology();
8920 n = doms_new ? ndoms_new : 0;
8922 /* Destroy deleted domains */
8923 for (i = 0; i < ndoms_cur; i++) {
8924 for (j = 0; j < n && !new_topology; j++) {
8925 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8926 && dattrs_equal(dattr_cur, i, dattr_new, j))
8929 /* no match - a current sched domain not in new doms_new[] */
8930 detach_destroy_domains(doms_cur + i);
8935 if (doms_new == NULL) {
8937 doms_new = fallback_doms;
8938 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8939 WARN_ON_ONCE(dattr_new);
8942 /* Build new domains */
8943 for (i = 0; i < ndoms_new; i++) {
8944 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8945 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8946 && dattrs_equal(dattr_new, i, dattr_cur, j))
8949 /* no match - add a new doms_new */
8950 __build_sched_domains(doms_new + i,
8951 dattr_new ? dattr_new + i : NULL);
8956 /* Remember the new sched domains */
8957 if (doms_cur != fallback_doms)
8959 kfree(dattr_cur); /* kfree(NULL) is safe */
8960 doms_cur = doms_new;
8961 dattr_cur = dattr_new;
8962 ndoms_cur = ndoms_new;
8964 register_sched_domain_sysctl();
8966 mutex_unlock(&sched_domains_mutex);
8969 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8970 static void arch_reinit_sched_domains(void)
8974 /* Destroy domains first to force the rebuild */
8975 partition_sched_domains(0, NULL, NULL);
8977 rebuild_sched_domains();
8981 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8983 unsigned int level = 0;
8985 if (sscanf(buf, "%u", &level) != 1)
8989 * level is always be positive so don't check for
8990 * level < POWERSAVINGS_BALANCE_NONE which is 0
8991 * What happens on 0 or 1 byte write,
8992 * need to check for count as well?
8995 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8999 sched_smt_power_savings = level;
9001 sched_mc_power_savings = level;
9003 arch_reinit_sched_domains();
9008 #ifdef CONFIG_SCHED_MC
9009 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9012 return sprintf(page, "%u\n", sched_mc_power_savings);
9014 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9015 const char *buf, size_t count)
9017 return sched_power_savings_store(buf, count, 0);
9019 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9020 sched_mc_power_savings_show,
9021 sched_mc_power_savings_store);
9024 #ifdef CONFIG_SCHED_SMT
9025 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9028 return sprintf(page, "%u\n", sched_smt_power_savings);
9030 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9031 const char *buf, size_t count)
9033 return sched_power_savings_store(buf, count, 1);
9035 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9036 sched_smt_power_savings_show,
9037 sched_smt_power_savings_store);
9040 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9044 #ifdef CONFIG_SCHED_SMT
9046 err = sysfs_create_file(&cls->kset.kobj,
9047 &attr_sched_smt_power_savings.attr);
9049 #ifdef CONFIG_SCHED_MC
9050 if (!err && mc_capable())
9051 err = sysfs_create_file(&cls->kset.kobj,
9052 &attr_sched_mc_power_savings.attr);
9056 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9058 #ifndef CONFIG_CPUSETS
9060 * Add online and remove offline CPUs from the scheduler domains.
9061 * When cpusets are enabled they take over this function.
9063 static int update_sched_domains(struct notifier_block *nfb,
9064 unsigned long action, void *hcpu)
9068 case CPU_ONLINE_FROZEN:
9070 case CPU_DEAD_FROZEN:
9071 partition_sched_domains(1, NULL, NULL);
9080 static int update_runtime(struct notifier_block *nfb,
9081 unsigned long action, void *hcpu)
9083 int cpu = (int)(long)hcpu;
9086 case CPU_DOWN_PREPARE:
9087 case CPU_DOWN_PREPARE_FROZEN:
9088 disable_runtime(cpu_rq(cpu));
9091 case CPU_DOWN_FAILED:
9092 case CPU_DOWN_FAILED_FROZEN:
9094 case CPU_ONLINE_FROZEN:
9095 enable_runtime(cpu_rq(cpu));
9103 void __init sched_init_smp(void)
9105 cpumask_var_t non_isolated_cpus;
9107 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9109 #if defined(CONFIG_NUMA)
9110 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9112 BUG_ON(sched_group_nodes_bycpu == NULL);
9115 mutex_lock(&sched_domains_mutex);
9116 arch_init_sched_domains(cpu_online_mask);
9117 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9118 if (cpumask_empty(non_isolated_cpus))
9119 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9120 mutex_unlock(&sched_domains_mutex);
9123 #ifndef CONFIG_CPUSETS
9124 /* XXX: Theoretical race here - CPU may be hotplugged now */
9125 hotcpu_notifier(update_sched_domains, 0);
9128 /* RT runtime code needs to handle some hotplug events */
9129 hotcpu_notifier(update_runtime, 0);
9133 /* Move init over to a non-isolated CPU */
9134 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9136 sched_init_granularity();
9137 free_cpumask_var(non_isolated_cpus);
9139 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9140 init_sched_rt_class();
9143 void __init sched_init_smp(void)
9145 sched_init_granularity();
9147 #endif /* CONFIG_SMP */
9149 const_debug unsigned int sysctl_timer_migration = 1;
9151 int in_sched_functions(unsigned long addr)
9153 return in_lock_functions(addr) ||
9154 (addr >= (unsigned long)__sched_text_start
9155 && addr < (unsigned long)__sched_text_end);
9158 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9160 cfs_rq->tasks_timeline = RB_ROOT;
9161 INIT_LIST_HEAD(&cfs_rq->tasks);
9162 #ifdef CONFIG_FAIR_GROUP_SCHED
9165 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9168 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9170 struct rt_prio_array *array;
9173 array = &rt_rq->active;
9174 for (i = 0; i < MAX_RT_PRIO; i++) {
9175 INIT_LIST_HEAD(array->queue + i);
9176 __clear_bit(i, array->bitmap);
9178 /* delimiter for bitsearch: */
9179 __set_bit(MAX_RT_PRIO, array->bitmap);
9181 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9182 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9184 rt_rq->highest_prio.next = MAX_RT_PRIO;
9188 rt_rq->rt_nr_migratory = 0;
9189 rt_rq->overloaded = 0;
9190 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9194 rt_rq->rt_throttled = 0;
9195 rt_rq->rt_runtime = 0;
9196 spin_lock_init(&rt_rq->rt_runtime_lock);
9198 #ifdef CONFIG_RT_GROUP_SCHED
9199 rt_rq->rt_nr_boosted = 0;
9204 #ifdef CONFIG_FAIR_GROUP_SCHED
9205 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9206 struct sched_entity *se, int cpu, int add,
9207 struct sched_entity *parent)
9209 struct rq *rq = cpu_rq(cpu);
9210 tg->cfs_rq[cpu] = cfs_rq;
9211 init_cfs_rq(cfs_rq, rq);
9214 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9217 /* se could be NULL for init_task_group */
9222 se->cfs_rq = &rq->cfs;
9224 se->cfs_rq = parent->my_q;
9227 se->load.weight = tg->shares;
9228 se->load.inv_weight = 0;
9229 se->parent = parent;
9233 #ifdef CONFIG_RT_GROUP_SCHED
9234 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9235 struct sched_rt_entity *rt_se, int cpu, int add,
9236 struct sched_rt_entity *parent)
9238 struct rq *rq = cpu_rq(cpu);
9240 tg->rt_rq[cpu] = rt_rq;
9241 init_rt_rq(rt_rq, rq);
9243 rt_rq->rt_se = rt_se;
9244 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9246 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9248 tg->rt_se[cpu] = rt_se;
9253 rt_se->rt_rq = &rq->rt;
9255 rt_se->rt_rq = parent->my_q;
9257 rt_se->my_q = rt_rq;
9258 rt_se->parent = parent;
9259 INIT_LIST_HEAD(&rt_se->run_list);
9263 void __init sched_init(void)
9266 unsigned long alloc_size = 0, ptr;
9268 #ifdef CONFIG_FAIR_GROUP_SCHED
9269 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9271 #ifdef CONFIG_RT_GROUP_SCHED
9272 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9274 #ifdef CONFIG_USER_SCHED
9277 #ifdef CONFIG_CPUMASK_OFFSTACK
9278 alloc_size += num_possible_cpus() * cpumask_size();
9281 * As sched_init() is called before page_alloc is setup,
9282 * we use alloc_bootmem().
9285 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9287 #ifdef CONFIG_FAIR_GROUP_SCHED
9288 init_task_group.se = (struct sched_entity **)ptr;
9289 ptr += nr_cpu_ids * sizeof(void **);
9291 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9292 ptr += nr_cpu_ids * sizeof(void **);
9294 #ifdef CONFIG_USER_SCHED
9295 root_task_group.se = (struct sched_entity **)ptr;
9296 ptr += nr_cpu_ids * sizeof(void **);
9298 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9299 ptr += nr_cpu_ids * sizeof(void **);
9300 #endif /* CONFIG_USER_SCHED */
9301 #endif /* CONFIG_FAIR_GROUP_SCHED */
9302 #ifdef CONFIG_RT_GROUP_SCHED
9303 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9304 ptr += nr_cpu_ids * sizeof(void **);
9306 init_task_group.rt_rq = (struct rt_rq **)ptr;
9307 ptr += nr_cpu_ids * sizeof(void **);
9309 #ifdef CONFIG_USER_SCHED
9310 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9311 ptr += nr_cpu_ids * sizeof(void **);
9313 root_task_group.rt_rq = (struct rt_rq **)ptr;
9314 ptr += nr_cpu_ids * sizeof(void **);
9315 #endif /* CONFIG_USER_SCHED */
9316 #endif /* CONFIG_RT_GROUP_SCHED */
9317 #ifdef CONFIG_CPUMASK_OFFSTACK
9318 for_each_possible_cpu(i) {
9319 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9320 ptr += cpumask_size();
9322 #endif /* CONFIG_CPUMASK_OFFSTACK */
9326 init_defrootdomain();
9329 init_rt_bandwidth(&def_rt_bandwidth,
9330 global_rt_period(), global_rt_runtime());
9332 #ifdef CONFIG_RT_GROUP_SCHED
9333 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9334 global_rt_period(), global_rt_runtime());
9335 #ifdef CONFIG_USER_SCHED
9336 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9337 global_rt_period(), RUNTIME_INF);
9338 #endif /* CONFIG_USER_SCHED */
9339 #endif /* CONFIG_RT_GROUP_SCHED */
9341 #ifdef CONFIG_GROUP_SCHED
9342 list_add(&init_task_group.list, &task_groups);
9343 INIT_LIST_HEAD(&init_task_group.children);
9345 #ifdef CONFIG_USER_SCHED
9346 INIT_LIST_HEAD(&root_task_group.children);
9347 init_task_group.parent = &root_task_group;
9348 list_add(&init_task_group.siblings, &root_task_group.children);
9349 #endif /* CONFIG_USER_SCHED */
9350 #endif /* CONFIG_GROUP_SCHED */
9352 for_each_possible_cpu(i) {
9356 spin_lock_init(&rq->lock);
9358 rq->calc_load_active = 0;
9359 rq->calc_load_update = jiffies + LOAD_FREQ;
9360 init_cfs_rq(&rq->cfs, rq);
9361 init_rt_rq(&rq->rt, rq);
9362 #ifdef CONFIG_FAIR_GROUP_SCHED
9363 init_task_group.shares = init_task_group_load;
9364 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9365 #ifdef CONFIG_CGROUP_SCHED
9367 * How much cpu bandwidth does init_task_group get?
9369 * In case of task-groups formed thr' the cgroup filesystem, it
9370 * gets 100% of the cpu resources in the system. This overall
9371 * system cpu resource is divided among the tasks of
9372 * init_task_group and its child task-groups in a fair manner,
9373 * based on each entity's (task or task-group's) weight
9374 * (se->load.weight).
9376 * In other words, if init_task_group has 10 tasks of weight
9377 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9378 * then A0's share of the cpu resource is:
9380 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9382 * We achieve this by letting init_task_group's tasks sit
9383 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9385 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9386 #elif defined CONFIG_USER_SCHED
9387 root_task_group.shares = NICE_0_LOAD;
9388 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9390 * In case of task-groups formed thr' the user id of tasks,
9391 * init_task_group represents tasks belonging to root user.
9392 * Hence it forms a sibling of all subsequent groups formed.
9393 * In this case, init_task_group gets only a fraction of overall
9394 * system cpu resource, based on the weight assigned to root
9395 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9396 * by letting tasks of init_task_group sit in a separate cfs_rq
9397 * (init_cfs_rq) and having one entity represent this group of
9398 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9400 init_tg_cfs_entry(&init_task_group,
9401 &per_cpu(init_cfs_rq, i),
9402 &per_cpu(init_sched_entity, i), i, 1,
9403 root_task_group.se[i]);
9406 #endif /* CONFIG_FAIR_GROUP_SCHED */
9408 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9409 #ifdef CONFIG_RT_GROUP_SCHED
9410 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9411 #ifdef CONFIG_CGROUP_SCHED
9412 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9413 #elif defined CONFIG_USER_SCHED
9414 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9415 init_tg_rt_entry(&init_task_group,
9416 &per_cpu(init_rt_rq, i),
9417 &per_cpu(init_sched_rt_entity, i), i, 1,
9418 root_task_group.rt_se[i]);
9422 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9423 rq->cpu_load[j] = 0;
9427 rq->post_schedule = 0;
9428 rq->active_balance = 0;
9429 rq->next_balance = jiffies;
9433 rq->migration_thread = NULL;
9434 INIT_LIST_HEAD(&rq->migration_queue);
9435 rq_attach_root(rq, &def_root_domain);
9438 atomic_set(&rq->nr_iowait, 0);
9441 set_load_weight(&init_task);
9443 #ifdef CONFIG_PREEMPT_NOTIFIERS
9444 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9448 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9451 #ifdef CONFIG_RT_MUTEXES
9452 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9456 * The boot idle thread does lazy MMU switching as well:
9458 atomic_inc(&init_mm.mm_count);
9459 enter_lazy_tlb(&init_mm, current);
9462 * Make us the idle thread. Technically, schedule() should not be
9463 * called from this thread, however somewhere below it might be,
9464 * but because we are the idle thread, we just pick up running again
9465 * when this runqueue becomes "idle".
9467 init_idle(current, smp_processor_id());
9469 calc_load_update = jiffies + LOAD_FREQ;
9472 * During early bootup we pretend to be a normal task:
9474 current->sched_class = &fair_sched_class;
9476 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9477 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9480 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9481 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9483 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9486 perf_counter_init();
9488 scheduler_running = 1;
9491 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9492 static inline int preempt_count_equals(int preempt_offset)
9494 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9496 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9499 void __might_sleep(char *file, int line, int preempt_offset)
9502 static unsigned long prev_jiffy; /* ratelimiting */
9504 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9505 system_state != SYSTEM_RUNNING || oops_in_progress)
9507 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9509 prev_jiffy = jiffies;
9512 "BUG: sleeping function called from invalid context at %s:%d\n",
9515 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9516 in_atomic(), irqs_disabled(),
9517 current->pid, current->comm);
9519 debug_show_held_locks(current);
9520 if (irqs_disabled())
9521 print_irqtrace_events(current);
9525 EXPORT_SYMBOL(__might_sleep);
9528 #ifdef CONFIG_MAGIC_SYSRQ
9529 static void normalize_task(struct rq *rq, struct task_struct *p)
9533 update_rq_clock(rq);
9534 on_rq = p->se.on_rq;
9536 deactivate_task(rq, p, 0);
9537 __setscheduler(rq, p, SCHED_NORMAL, 0);
9539 activate_task(rq, p, 0);
9540 resched_task(rq->curr);
9544 void normalize_rt_tasks(void)
9546 struct task_struct *g, *p;
9547 unsigned long flags;
9550 read_lock_irqsave(&tasklist_lock, flags);
9551 do_each_thread(g, p) {
9553 * Only normalize user tasks:
9558 p->se.exec_start = 0;
9559 #ifdef CONFIG_SCHEDSTATS
9560 p->se.wait_start = 0;
9561 p->se.sleep_start = 0;
9562 p->se.block_start = 0;
9567 * Renice negative nice level userspace
9570 if (TASK_NICE(p) < 0 && p->mm)
9571 set_user_nice(p, 0);
9575 spin_lock(&p->pi_lock);
9576 rq = __task_rq_lock(p);
9578 normalize_task(rq, p);
9580 __task_rq_unlock(rq);
9581 spin_unlock(&p->pi_lock);
9582 } while_each_thread(g, p);
9584 read_unlock_irqrestore(&tasklist_lock, flags);
9587 #endif /* CONFIG_MAGIC_SYSRQ */
9591 * These functions are only useful for the IA64 MCA handling.
9593 * They can only be called when the whole system has been
9594 * stopped - every CPU needs to be quiescent, and no scheduling
9595 * activity can take place. Using them for anything else would
9596 * be a serious bug, and as a result, they aren't even visible
9597 * under any other configuration.
9601 * curr_task - return the current task for a given cpu.
9602 * @cpu: the processor in question.
9604 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9606 struct task_struct *curr_task(int cpu)
9608 return cpu_curr(cpu);
9612 * set_curr_task - set the current task for a given cpu.
9613 * @cpu: the processor in question.
9614 * @p: the task pointer to set.
9616 * Description: This function must only be used when non-maskable interrupts
9617 * are serviced on a separate stack. It allows the architecture to switch the
9618 * notion of the current task on a cpu in a non-blocking manner. This function
9619 * must be called with all CPU's synchronized, and interrupts disabled, the
9620 * and caller must save the original value of the current task (see
9621 * curr_task() above) and restore that value before reenabling interrupts and
9622 * re-starting the system.
9624 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9626 void set_curr_task(int cpu, struct task_struct *p)
9633 #ifdef CONFIG_FAIR_GROUP_SCHED
9634 static void free_fair_sched_group(struct task_group *tg)
9638 for_each_possible_cpu(i) {
9640 kfree(tg->cfs_rq[i]);
9650 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9652 struct cfs_rq *cfs_rq;
9653 struct sched_entity *se;
9657 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9660 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9664 tg->shares = NICE_0_LOAD;
9666 for_each_possible_cpu(i) {
9669 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9670 GFP_KERNEL, cpu_to_node(i));
9674 se = kzalloc_node(sizeof(struct sched_entity),
9675 GFP_KERNEL, cpu_to_node(i));
9679 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9688 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9690 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9691 &cpu_rq(cpu)->leaf_cfs_rq_list);
9694 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9696 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9698 #else /* !CONFG_FAIR_GROUP_SCHED */
9699 static inline void free_fair_sched_group(struct task_group *tg)
9704 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9709 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9713 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9716 #endif /* CONFIG_FAIR_GROUP_SCHED */
9718 #ifdef CONFIG_RT_GROUP_SCHED
9719 static void free_rt_sched_group(struct task_group *tg)
9723 destroy_rt_bandwidth(&tg->rt_bandwidth);
9725 for_each_possible_cpu(i) {
9727 kfree(tg->rt_rq[i]);
9729 kfree(tg->rt_se[i]);
9737 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9739 struct rt_rq *rt_rq;
9740 struct sched_rt_entity *rt_se;
9744 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9747 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9751 init_rt_bandwidth(&tg->rt_bandwidth,
9752 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9754 for_each_possible_cpu(i) {
9757 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9758 GFP_KERNEL, cpu_to_node(i));
9762 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9763 GFP_KERNEL, cpu_to_node(i));
9767 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9776 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9778 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9779 &cpu_rq(cpu)->leaf_rt_rq_list);
9782 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9784 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9786 #else /* !CONFIG_RT_GROUP_SCHED */
9787 static inline void free_rt_sched_group(struct task_group *tg)
9792 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9797 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9801 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9804 #endif /* CONFIG_RT_GROUP_SCHED */
9806 #ifdef CONFIG_GROUP_SCHED
9807 static void free_sched_group(struct task_group *tg)
9809 free_fair_sched_group(tg);
9810 free_rt_sched_group(tg);
9814 /* allocate runqueue etc for a new task group */
9815 struct task_group *sched_create_group(struct task_group *parent)
9817 struct task_group *tg;
9818 unsigned long flags;
9821 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9823 return ERR_PTR(-ENOMEM);
9825 if (!alloc_fair_sched_group(tg, parent))
9828 if (!alloc_rt_sched_group(tg, parent))
9831 spin_lock_irqsave(&task_group_lock, flags);
9832 for_each_possible_cpu(i) {
9833 register_fair_sched_group(tg, i);
9834 register_rt_sched_group(tg, i);
9836 list_add_rcu(&tg->list, &task_groups);
9838 WARN_ON(!parent); /* root should already exist */
9840 tg->parent = parent;
9841 INIT_LIST_HEAD(&tg->children);
9842 list_add_rcu(&tg->siblings, &parent->children);
9843 spin_unlock_irqrestore(&task_group_lock, flags);
9848 free_sched_group(tg);
9849 return ERR_PTR(-ENOMEM);
9852 /* rcu callback to free various structures associated with a task group */
9853 static void free_sched_group_rcu(struct rcu_head *rhp)
9855 /* now it should be safe to free those cfs_rqs */
9856 free_sched_group(container_of(rhp, struct task_group, rcu));
9859 /* Destroy runqueue etc associated with a task group */
9860 void sched_destroy_group(struct task_group *tg)
9862 unsigned long flags;
9865 spin_lock_irqsave(&task_group_lock, flags);
9866 for_each_possible_cpu(i) {
9867 unregister_fair_sched_group(tg, i);
9868 unregister_rt_sched_group(tg, i);
9870 list_del_rcu(&tg->list);
9871 list_del_rcu(&tg->siblings);
9872 spin_unlock_irqrestore(&task_group_lock, flags);
9874 /* wait for possible concurrent references to cfs_rqs complete */
9875 call_rcu(&tg->rcu, free_sched_group_rcu);
9878 /* change task's runqueue when it moves between groups.
9879 * The caller of this function should have put the task in its new group
9880 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9881 * reflect its new group.
9883 void sched_move_task(struct task_struct *tsk)
9886 unsigned long flags;
9889 rq = task_rq_lock(tsk, &flags);
9891 update_rq_clock(rq);
9893 running = task_current(rq, tsk);
9894 on_rq = tsk->se.on_rq;
9897 dequeue_task(rq, tsk, 0);
9898 if (unlikely(running))
9899 tsk->sched_class->put_prev_task(rq, tsk);
9901 set_task_rq(tsk, task_cpu(tsk));
9903 #ifdef CONFIG_FAIR_GROUP_SCHED
9904 if (tsk->sched_class->moved_group)
9905 tsk->sched_class->moved_group(tsk);
9908 if (unlikely(running))
9909 tsk->sched_class->set_curr_task(rq);
9911 enqueue_task(rq, tsk, 0);
9913 task_rq_unlock(rq, &flags);
9915 #endif /* CONFIG_GROUP_SCHED */
9917 #ifdef CONFIG_FAIR_GROUP_SCHED
9918 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9920 struct cfs_rq *cfs_rq = se->cfs_rq;
9925 dequeue_entity(cfs_rq, se, 0);
9927 se->load.weight = shares;
9928 se->load.inv_weight = 0;
9931 enqueue_entity(cfs_rq, se, 0);
9934 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9936 struct cfs_rq *cfs_rq = se->cfs_rq;
9937 struct rq *rq = cfs_rq->rq;
9938 unsigned long flags;
9940 spin_lock_irqsave(&rq->lock, flags);
9941 __set_se_shares(se, shares);
9942 spin_unlock_irqrestore(&rq->lock, flags);
9945 static DEFINE_MUTEX(shares_mutex);
9947 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9950 unsigned long flags;
9953 * We can't change the weight of the root cgroup.
9958 if (shares < MIN_SHARES)
9959 shares = MIN_SHARES;
9960 else if (shares > MAX_SHARES)
9961 shares = MAX_SHARES;
9963 mutex_lock(&shares_mutex);
9964 if (tg->shares == shares)
9967 spin_lock_irqsave(&task_group_lock, flags);
9968 for_each_possible_cpu(i)
9969 unregister_fair_sched_group(tg, i);
9970 list_del_rcu(&tg->siblings);
9971 spin_unlock_irqrestore(&task_group_lock, flags);
9973 /* wait for any ongoing reference to this group to finish */
9974 synchronize_sched();
9977 * Now we are free to modify the group's share on each cpu
9978 * w/o tripping rebalance_share or load_balance_fair.
9980 tg->shares = shares;
9981 for_each_possible_cpu(i) {
9985 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9986 set_se_shares(tg->se[i], shares);
9990 * Enable load balance activity on this group, by inserting it back on
9991 * each cpu's rq->leaf_cfs_rq_list.
9993 spin_lock_irqsave(&task_group_lock, flags);
9994 for_each_possible_cpu(i)
9995 register_fair_sched_group(tg, i);
9996 list_add_rcu(&tg->siblings, &tg->parent->children);
9997 spin_unlock_irqrestore(&task_group_lock, flags);
9999 mutex_unlock(&shares_mutex);
10003 unsigned long sched_group_shares(struct task_group *tg)
10009 #ifdef CONFIG_RT_GROUP_SCHED
10011 * Ensure that the real time constraints are schedulable.
10013 static DEFINE_MUTEX(rt_constraints_mutex);
10015 static unsigned long to_ratio(u64 period, u64 runtime)
10017 if (runtime == RUNTIME_INF)
10020 return div64_u64(runtime << 20, period);
10023 /* Must be called with tasklist_lock held */
10024 static inline int tg_has_rt_tasks(struct task_group *tg)
10026 struct task_struct *g, *p;
10028 do_each_thread(g, p) {
10029 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10031 } while_each_thread(g, p);
10036 struct rt_schedulable_data {
10037 struct task_group *tg;
10042 static int tg_schedulable(struct task_group *tg, void *data)
10044 struct rt_schedulable_data *d = data;
10045 struct task_group *child;
10046 unsigned long total, sum = 0;
10047 u64 period, runtime;
10049 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10050 runtime = tg->rt_bandwidth.rt_runtime;
10053 period = d->rt_period;
10054 runtime = d->rt_runtime;
10057 #ifdef CONFIG_USER_SCHED
10058 if (tg == &root_task_group) {
10059 period = global_rt_period();
10060 runtime = global_rt_runtime();
10065 * Cannot have more runtime than the period.
10067 if (runtime > period && runtime != RUNTIME_INF)
10071 * Ensure we don't starve existing RT tasks.
10073 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10076 total = to_ratio(period, runtime);
10079 * Nobody can have more than the global setting allows.
10081 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10085 * The sum of our children's runtime should not exceed our own.
10087 list_for_each_entry_rcu(child, &tg->children, siblings) {
10088 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10089 runtime = child->rt_bandwidth.rt_runtime;
10091 if (child == d->tg) {
10092 period = d->rt_period;
10093 runtime = d->rt_runtime;
10096 sum += to_ratio(period, runtime);
10105 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10107 struct rt_schedulable_data data = {
10109 .rt_period = period,
10110 .rt_runtime = runtime,
10113 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10116 static int tg_set_bandwidth(struct task_group *tg,
10117 u64 rt_period, u64 rt_runtime)
10121 mutex_lock(&rt_constraints_mutex);
10122 read_lock(&tasklist_lock);
10123 err = __rt_schedulable(tg, rt_period, rt_runtime);
10127 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10128 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10129 tg->rt_bandwidth.rt_runtime = rt_runtime;
10131 for_each_possible_cpu(i) {
10132 struct rt_rq *rt_rq = tg->rt_rq[i];
10134 spin_lock(&rt_rq->rt_runtime_lock);
10135 rt_rq->rt_runtime = rt_runtime;
10136 spin_unlock(&rt_rq->rt_runtime_lock);
10138 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10140 read_unlock(&tasklist_lock);
10141 mutex_unlock(&rt_constraints_mutex);
10146 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10148 u64 rt_runtime, rt_period;
10150 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10151 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10152 if (rt_runtime_us < 0)
10153 rt_runtime = RUNTIME_INF;
10155 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10158 long sched_group_rt_runtime(struct task_group *tg)
10162 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10165 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10166 do_div(rt_runtime_us, NSEC_PER_USEC);
10167 return rt_runtime_us;
10170 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10172 u64 rt_runtime, rt_period;
10174 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10175 rt_runtime = tg->rt_bandwidth.rt_runtime;
10177 if (rt_period == 0)
10180 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10183 long sched_group_rt_period(struct task_group *tg)
10187 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10188 do_div(rt_period_us, NSEC_PER_USEC);
10189 return rt_period_us;
10192 static int sched_rt_global_constraints(void)
10194 u64 runtime, period;
10197 if (sysctl_sched_rt_period <= 0)
10200 runtime = global_rt_runtime();
10201 period = global_rt_period();
10204 * Sanity check on the sysctl variables.
10206 if (runtime > period && runtime != RUNTIME_INF)
10209 mutex_lock(&rt_constraints_mutex);
10210 read_lock(&tasklist_lock);
10211 ret = __rt_schedulable(NULL, 0, 0);
10212 read_unlock(&tasklist_lock);
10213 mutex_unlock(&rt_constraints_mutex);
10218 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10220 /* Don't accept realtime tasks when there is no way for them to run */
10221 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10227 #else /* !CONFIG_RT_GROUP_SCHED */
10228 static int sched_rt_global_constraints(void)
10230 unsigned long flags;
10233 if (sysctl_sched_rt_period <= 0)
10237 * There's always some RT tasks in the root group
10238 * -- migration, kstopmachine etc..
10240 if (sysctl_sched_rt_runtime == 0)
10243 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10244 for_each_possible_cpu(i) {
10245 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10247 spin_lock(&rt_rq->rt_runtime_lock);
10248 rt_rq->rt_runtime = global_rt_runtime();
10249 spin_unlock(&rt_rq->rt_runtime_lock);
10251 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10255 #endif /* CONFIG_RT_GROUP_SCHED */
10257 int sched_rt_handler(struct ctl_table *table, int write,
10258 struct file *filp, void __user *buffer, size_t *lenp,
10262 int old_period, old_runtime;
10263 static DEFINE_MUTEX(mutex);
10265 mutex_lock(&mutex);
10266 old_period = sysctl_sched_rt_period;
10267 old_runtime = sysctl_sched_rt_runtime;
10269 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10271 if (!ret && write) {
10272 ret = sched_rt_global_constraints();
10274 sysctl_sched_rt_period = old_period;
10275 sysctl_sched_rt_runtime = old_runtime;
10277 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10278 def_rt_bandwidth.rt_period =
10279 ns_to_ktime(global_rt_period());
10282 mutex_unlock(&mutex);
10287 #ifdef CONFIG_CGROUP_SCHED
10289 /* return corresponding task_group object of a cgroup */
10290 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10292 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10293 struct task_group, css);
10296 static struct cgroup_subsys_state *
10297 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10299 struct task_group *tg, *parent;
10301 if (!cgrp->parent) {
10302 /* This is early initialization for the top cgroup */
10303 return &init_task_group.css;
10306 parent = cgroup_tg(cgrp->parent);
10307 tg = sched_create_group(parent);
10309 return ERR_PTR(-ENOMEM);
10315 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10317 struct task_group *tg = cgroup_tg(cgrp);
10319 sched_destroy_group(tg);
10323 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10324 struct task_struct *tsk)
10326 #ifdef CONFIG_RT_GROUP_SCHED
10327 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10330 /* We don't support RT-tasks being in separate groups */
10331 if (tsk->sched_class != &fair_sched_class)
10339 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10340 struct cgroup *old_cont, struct task_struct *tsk)
10342 sched_move_task(tsk);
10345 #ifdef CONFIG_FAIR_GROUP_SCHED
10346 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10349 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10352 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10354 struct task_group *tg = cgroup_tg(cgrp);
10356 return (u64) tg->shares;
10358 #endif /* CONFIG_FAIR_GROUP_SCHED */
10360 #ifdef CONFIG_RT_GROUP_SCHED
10361 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10364 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10367 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10369 return sched_group_rt_runtime(cgroup_tg(cgrp));
10372 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10375 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10378 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10380 return sched_group_rt_period(cgroup_tg(cgrp));
10382 #endif /* CONFIG_RT_GROUP_SCHED */
10384 static struct cftype cpu_files[] = {
10385 #ifdef CONFIG_FAIR_GROUP_SCHED
10388 .read_u64 = cpu_shares_read_u64,
10389 .write_u64 = cpu_shares_write_u64,
10392 #ifdef CONFIG_RT_GROUP_SCHED
10394 .name = "rt_runtime_us",
10395 .read_s64 = cpu_rt_runtime_read,
10396 .write_s64 = cpu_rt_runtime_write,
10399 .name = "rt_period_us",
10400 .read_u64 = cpu_rt_period_read_uint,
10401 .write_u64 = cpu_rt_period_write_uint,
10406 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10408 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10411 struct cgroup_subsys cpu_cgroup_subsys = {
10413 .create = cpu_cgroup_create,
10414 .destroy = cpu_cgroup_destroy,
10415 .can_attach = cpu_cgroup_can_attach,
10416 .attach = cpu_cgroup_attach,
10417 .populate = cpu_cgroup_populate,
10418 .subsys_id = cpu_cgroup_subsys_id,
10422 #endif /* CONFIG_CGROUP_SCHED */
10424 #ifdef CONFIG_CGROUP_CPUACCT
10427 * CPU accounting code for task groups.
10429 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10430 * (balbir@in.ibm.com).
10433 /* track cpu usage of a group of tasks and its child groups */
10435 struct cgroup_subsys_state css;
10436 /* cpuusage holds pointer to a u64-type object on every cpu */
10438 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10439 struct cpuacct *parent;
10442 struct cgroup_subsys cpuacct_subsys;
10444 /* return cpu accounting group corresponding to this container */
10445 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10447 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10448 struct cpuacct, css);
10451 /* return cpu accounting group to which this task belongs */
10452 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10454 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10455 struct cpuacct, css);
10458 /* create a new cpu accounting group */
10459 static struct cgroup_subsys_state *cpuacct_create(
10460 struct cgroup_subsys *ss, struct cgroup *cgrp)
10462 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10468 ca->cpuusage = alloc_percpu(u64);
10472 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10473 if (percpu_counter_init(&ca->cpustat[i], 0))
10474 goto out_free_counters;
10477 ca->parent = cgroup_ca(cgrp->parent);
10483 percpu_counter_destroy(&ca->cpustat[i]);
10484 free_percpu(ca->cpuusage);
10488 return ERR_PTR(-ENOMEM);
10491 /* destroy an existing cpu accounting group */
10493 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10495 struct cpuacct *ca = cgroup_ca(cgrp);
10498 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10499 percpu_counter_destroy(&ca->cpustat[i]);
10500 free_percpu(ca->cpuusage);
10504 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10506 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10509 #ifndef CONFIG_64BIT
10511 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10513 spin_lock_irq(&cpu_rq(cpu)->lock);
10515 spin_unlock_irq(&cpu_rq(cpu)->lock);
10523 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10525 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10527 #ifndef CONFIG_64BIT
10529 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10531 spin_lock_irq(&cpu_rq(cpu)->lock);
10533 spin_unlock_irq(&cpu_rq(cpu)->lock);
10539 /* return total cpu usage (in nanoseconds) of a group */
10540 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10542 struct cpuacct *ca = cgroup_ca(cgrp);
10543 u64 totalcpuusage = 0;
10546 for_each_present_cpu(i)
10547 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10549 return totalcpuusage;
10552 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10555 struct cpuacct *ca = cgroup_ca(cgrp);
10564 for_each_present_cpu(i)
10565 cpuacct_cpuusage_write(ca, i, 0);
10571 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10572 struct seq_file *m)
10574 struct cpuacct *ca = cgroup_ca(cgroup);
10578 for_each_present_cpu(i) {
10579 percpu = cpuacct_cpuusage_read(ca, i);
10580 seq_printf(m, "%llu ", (unsigned long long) percpu);
10582 seq_printf(m, "\n");
10586 static const char *cpuacct_stat_desc[] = {
10587 [CPUACCT_STAT_USER] = "user",
10588 [CPUACCT_STAT_SYSTEM] = "system",
10591 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10592 struct cgroup_map_cb *cb)
10594 struct cpuacct *ca = cgroup_ca(cgrp);
10597 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10598 s64 val = percpu_counter_read(&ca->cpustat[i]);
10599 val = cputime64_to_clock_t(val);
10600 cb->fill(cb, cpuacct_stat_desc[i], val);
10605 static struct cftype files[] = {
10608 .read_u64 = cpuusage_read,
10609 .write_u64 = cpuusage_write,
10612 .name = "usage_percpu",
10613 .read_seq_string = cpuacct_percpu_seq_read,
10617 .read_map = cpuacct_stats_show,
10621 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10623 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10627 * charge this task's execution time to its accounting group.
10629 * called with rq->lock held.
10631 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10633 struct cpuacct *ca;
10636 if (unlikely(!cpuacct_subsys.active))
10639 cpu = task_cpu(tsk);
10645 for (; ca; ca = ca->parent) {
10646 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10647 *cpuusage += cputime;
10654 * Charge the system/user time to the task's accounting group.
10656 static void cpuacct_update_stats(struct task_struct *tsk,
10657 enum cpuacct_stat_index idx, cputime_t val)
10659 struct cpuacct *ca;
10661 if (unlikely(!cpuacct_subsys.active))
10668 percpu_counter_add(&ca->cpustat[idx], val);
10674 struct cgroup_subsys cpuacct_subsys = {
10676 .create = cpuacct_create,
10677 .destroy = cpuacct_destroy,
10678 .populate = cpuacct_populate,
10679 .subsys_id = cpuacct_subsys_id,
10681 #endif /* CONFIG_CGROUP_CPUACCT */