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_event.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/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_CGROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 struct cgroup_subsys_state css;
248 #ifdef CONFIG_FAIR_GROUP_SCHED
249 /* schedulable entities of this group on each cpu */
250 struct sched_entity **se;
251 /* runqueue "owned" by this group on each cpu */
252 struct cfs_rq **cfs_rq;
253 unsigned long shares;
256 #ifdef CONFIG_RT_GROUP_SCHED
257 struct sched_rt_entity **rt_se;
258 struct rt_rq **rt_rq;
260 struct rt_bandwidth rt_bandwidth;
264 struct list_head list;
266 struct task_group *parent;
267 struct list_head siblings;
268 struct list_head children;
271 #define root_task_group init_task_group
273 /* task_group_lock serializes add/remove of task groups and also changes to
274 * a task group's cpu shares.
276 static DEFINE_SPINLOCK(task_group_lock);
278 #ifdef CONFIG_FAIR_GROUP_SCHED
281 static int root_task_group_empty(void)
283 return list_empty(&root_task_group.children);
287 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
290 * A weight of 0 or 1 can cause arithmetics problems.
291 * A weight of a cfs_rq is the sum of weights of which entities
292 * are queued on this cfs_rq, so a weight of a entity should not be
293 * too large, so as the shares value of a task group.
294 * (The default weight is 1024 - so there's no practical
295 * limitation from this.)
298 #define MAX_SHARES (1UL << 18)
300 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
303 /* Default task group.
304 * Every task in system belong to this group at bootup.
306 struct task_group init_task_group;
308 /* return group to which a task belongs */
309 static inline struct task_group *task_group(struct task_struct *p)
311 struct task_group *tg;
313 #ifdef CONFIG_CGROUP_SCHED
314 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
315 struct task_group, css);
317 tg = &init_task_group;
322 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
323 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
327 p->se.parent = task_group(p)->se[cpu];
330 #ifdef CONFIG_RT_GROUP_SCHED
331 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
332 p->rt.parent = task_group(p)->rt_se[cpu];
338 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
339 static inline struct task_group *task_group(struct task_struct *p)
344 #endif /* CONFIG_CGROUP_SCHED */
346 /* CFS-related fields in a runqueue */
348 struct load_weight load;
349 unsigned long nr_running;
354 struct rb_root tasks_timeline;
355 struct rb_node *rb_leftmost;
357 struct list_head tasks;
358 struct list_head *balance_iterator;
361 * 'curr' points to currently running entity on this cfs_rq.
362 * It is set to NULL otherwise (i.e when none are currently running).
364 struct sched_entity *curr, *next, *last;
366 unsigned int nr_spread_over;
368 #ifdef CONFIG_FAIR_GROUP_SCHED
369 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
372 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
373 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
374 * (like users, containers etc.)
376 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
377 * list is used during load balance.
379 struct list_head leaf_cfs_rq_list;
380 struct task_group *tg; /* group that "owns" this runqueue */
384 * the part of load.weight contributed by tasks
386 unsigned long task_weight;
389 * h_load = weight * f(tg)
391 * Where f(tg) is the recursive weight fraction assigned to
394 unsigned long h_load;
397 * this cpu's part of tg->shares
399 unsigned long shares;
402 * load.weight at the time we set shares
404 unsigned long rq_weight;
409 /* Real-Time classes' related field in a runqueue: */
411 struct rt_prio_array active;
412 unsigned long rt_nr_running;
413 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
415 int curr; /* highest queued rt task prio */
417 int next; /* next highest */
422 unsigned long rt_nr_migratory;
423 unsigned long rt_nr_total;
425 struct plist_head pushable_tasks;
430 /* Nests inside the rq lock: */
431 spinlock_t rt_runtime_lock;
433 #ifdef CONFIG_RT_GROUP_SCHED
434 unsigned long rt_nr_boosted;
437 struct list_head leaf_rt_rq_list;
438 struct task_group *tg;
439 struct sched_rt_entity *rt_se;
446 * We add the notion of a root-domain which will be used to define per-domain
447 * variables. Each exclusive cpuset essentially defines an island domain by
448 * fully partitioning the member cpus from any other cpuset. Whenever a new
449 * exclusive cpuset is created, we also create and attach a new root-domain
456 cpumask_var_t online;
459 * The "RT overload" flag: it gets set if a CPU has more than
460 * one runnable RT task.
462 cpumask_var_t rto_mask;
465 struct cpupri cpupri;
470 * By default the system creates a single root-domain with all cpus as
471 * members (mimicking the global state we have today).
473 static struct root_domain def_root_domain;
478 * This is the main, per-CPU runqueue data structure.
480 * Locking rule: those places that want to lock multiple runqueues
481 * (such as the load balancing or the thread migration code), lock
482 * acquire operations must be ordered by ascending &runqueue.
489 * nr_running and cpu_load should be in the same cacheline because
490 * remote CPUs use both these fields when doing load calculation.
492 unsigned long nr_running;
493 #define CPU_LOAD_IDX_MAX 5
494 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
496 unsigned long last_tick_seen;
497 unsigned char in_nohz_recently;
499 /* capture load from *all* tasks on this cpu: */
500 struct load_weight load;
501 unsigned long nr_load_updates;
507 #ifdef CONFIG_FAIR_GROUP_SCHED
508 /* list of leaf cfs_rq on this cpu: */
509 struct list_head leaf_cfs_rq_list;
511 #ifdef CONFIG_RT_GROUP_SCHED
512 struct list_head leaf_rt_rq_list;
516 * This is part of a global counter where only the total sum
517 * over all CPUs matters. A task can increase this counter on
518 * one CPU and if it got migrated afterwards it may decrease
519 * it on another CPU. Always updated under the runqueue lock:
521 unsigned long nr_uninterruptible;
523 struct task_struct *curr, *idle;
524 unsigned long next_balance;
525 struct mm_struct *prev_mm;
532 struct root_domain *rd;
533 struct sched_domain *sd;
535 unsigned char idle_at_tick;
536 /* For active balancing */
540 /* cpu of this runqueue: */
544 unsigned long avg_load_per_task;
546 struct task_struct *migration_thread;
547 struct list_head migration_queue;
555 /* calc_load related fields */
556 unsigned long calc_load_update;
557 long calc_load_active;
559 #ifdef CONFIG_SCHED_HRTICK
561 int hrtick_csd_pending;
562 struct call_single_data hrtick_csd;
564 struct hrtimer hrtick_timer;
567 #ifdef CONFIG_SCHEDSTATS
569 struct sched_info rq_sched_info;
570 unsigned long long rq_cpu_time;
571 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
573 /* sys_sched_yield() stats */
574 unsigned int yld_count;
576 /* schedule() stats */
577 unsigned int sched_switch;
578 unsigned int sched_count;
579 unsigned int sched_goidle;
581 /* try_to_wake_up() stats */
582 unsigned int ttwu_count;
583 unsigned int ttwu_local;
586 unsigned int bkl_count;
590 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
593 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
595 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
598 static inline int cpu_of(struct rq *rq)
608 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
609 * See detach_destroy_domains: synchronize_sched for details.
611 * The domain tree of any CPU may only be accessed from within
612 * preempt-disabled sections.
614 #define for_each_domain(cpu, __sd) \
615 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
617 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
618 #define this_rq() (&__get_cpu_var(runqueues))
619 #define task_rq(p) cpu_rq(task_cpu(p))
620 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
621 #define raw_rq() (&__raw_get_cpu_var(runqueues))
623 inline void update_rq_clock(struct rq *rq)
625 rq->clock = sched_clock_cpu(cpu_of(rq));
629 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
631 #ifdef CONFIG_SCHED_DEBUG
632 # define const_debug __read_mostly
634 # define const_debug static const
639 * @cpu: the processor in question.
641 * Returns true if the current cpu runqueue is locked.
642 * This interface allows printk to be called with the runqueue lock
643 * held and know whether or not it is OK to wake up the klogd.
645 int runqueue_is_locked(int cpu)
647 return spin_is_locked(&cpu_rq(cpu)->lock);
651 * Debugging: various feature bits
654 #define SCHED_FEAT(name, enabled) \
655 __SCHED_FEAT_##name ,
658 #include "sched_features.h"
663 #define SCHED_FEAT(name, enabled) \
664 (1UL << __SCHED_FEAT_##name) * enabled |
666 const_debug unsigned int sysctl_sched_features =
667 #include "sched_features.h"
672 #ifdef CONFIG_SCHED_DEBUG
673 #define SCHED_FEAT(name, enabled) \
676 static __read_mostly char *sched_feat_names[] = {
677 #include "sched_features.h"
683 static int sched_feat_show(struct seq_file *m, void *v)
687 for (i = 0; sched_feat_names[i]; i++) {
688 if (!(sysctl_sched_features & (1UL << i)))
690 seq_printf(m, "%s ", sched_feat_names[i]);
698 sched_feat_write(struct file *filp, const char __user *ubuf,
699 size_t cnt, loff_t *ppos)
709 if (copy_from_user(&buf, ubuf, cnt))
715 if (strncmp(buf, "NO_", 3) == 0) {
720 for (i = 0; sched_feat_names[i]; i++) {
721 if (strcmp(cmp, sched_feat_names[i]) == 0) {
723 sysctl_sched_features &= ~(1UL << i);
725 sysctl_sched_features |= (1UL << i);
730 if (!sched_feat_names[i])
738 static int sched_feat_open(struct inode *inode, struct file *filp)
740 return single_open(filp, sched_feat_show, NULL);
743 static const struct file_operations sched_feat_fops = {
744 .open = sched_feat_open,
745 .write = sched_feat_write,
748 .release = single_release,
751 static __init int sched_init_debug(void)
753 debugfs_create_file("sched_features", 0644, NULL, NULL,
758 late_initcall(sched_init_debug);
762 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
765 * Number of tasks to iterate in a single balance run.
766 * Limited because this is done with IRQs disabled.
768 const_debug unsigned int sysctl_sched_nr_migrate = 32;
771 * ratelimit for updating the group shares.
774 unsigned int sysctl_sched_shares_ratelimit = 250000;
775 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
778 * Inject some fuzzyness into changing the per-cpu group shares
779 * this avoids remote rq-locks at the expense of fairness.
782 unsigned int sysctl_sched_shares_thresh = 4;
785 * period over which we average the RT time consumption, measured
790 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
793 * period over which we measure -rt task cpu usage in us.
796 unsigned int sysctl_sched_rt_period = 1000000;
798 static __read_mostly int scheduler_running;
801 * part of the period that we allow rt tasks to run in us.
804 int sysctl_sched_rt_runtime = 950000;
806 static inline u64 global_rt_period(void)
808 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
811 static inline u64 global_rt_runtime(void)
813 if (sysctl_sched_rt_runtime < 0)
816 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
819 #ifndef prepare_arch_switch
820 # define prepare_arch_switch(next) do { } while (0)
822 #ifndef finish_arch_switch
823 # define finish_arch_switch(prev) do { } while (0)
826 static inline int task_current(struct rq *rq, struct task_struct *p)
828 return rq->curr == p;
831 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
832 static inline int task_running(struct rq *rq, struct task_struct *p)
834 return task_current(rq, p);
837 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
841 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
843 #ifdef CONFIG_DEBUG_SPINLOCK
844 /* this is a valid case when another task releases the spinlock */
845 rq->lock.owner = current;
848 * If we are tracking spinlock dependencies then we have to
849 * fix up the runqueue lock - which gets 'carried over' from
852 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
854 spin_unlock_irq(&rq->lock);
857 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
858 static inline int task_running(struct rq *rq, struct task_struct *p)
863 return task_current(rq, p);
867 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
871 * We can optimise this out completely for !SMP, because the
872 * SMP rebalancing from interrupt is the only thing that cares
877 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
878 spin_unlock_irq(&rq->lock);
880 spin_unlock(&rq->lock);
884 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
888 * After ->oncpu is cleared, the task can be moved to a different CPU.
889 * We must ensure this doesn't happen until the switch is completely
895 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
899 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
902 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
905 static inline int task_is_waking(struct task_struct *p)
907 return unlikely(p->state == TASK_WAKING);
911 * __task_rq_lock - lock the runqueue a given task resides on.
912 * Must be called interrupts disabled.
914 static inline struct rq *__task_rq_lock(struct task_struct *p)
921 spin_lock(&rq->lock);
922 if (likely(rq == task_rq(p)))
924 spin_unlock(&rq->lock);
929 * task_rq_lock - lock the runqueue a given task resides on and disable
930 * interrupts. Note the ordering: we can safely lookup the task_rq without
931 * explicitly disabling preemption.
933 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
939 local_irq_save(*flags);
941 spin_lock(&rq->lock);
942 if (likely(rq == task_rq(p)))
944 spin_unlock_irqrestore(&rq->lock, *flags);
948 void task_rq_unlock_wait(struct task_struct *p)
950 struct rq *rq = task_rq(p);
952 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
953 spin_unlock_wait(&rq->lock);
956 static void __task_rq_unlock(struct rq *rq)
959 spin_unlock(&rq->lock);
962 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
965 spin_unlock_irqrestore(&rq->lock, *flags);
969 * this_rq_lock - lock this runqueue and disable interrupts.
971 static struct rq *this_rq_lock(void)
978 spin_lock(&rq->lock);
983 #ifdef CONFIG_SCHED_HRTICK
985 * Use HR-timers to deliver accurate preemption points.
987 * Its all a bit involved since we cannot program an hrt while holding the
988 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
991 * When we get rescheduled we reprogram the hrtick_timer outside of the
997 * - enabled by features
998 * - hrtimer is actually high res
1000 static inline int hrtick_enabled(struct rq *rq)
1002 if (!sched_feat(HRTICK))
1004 if (!cpu_active(cpu_of(rq)))
1006 return hrtimer_is_hres_active(&rq->hrtick_timer);
1009 static void hrtick_clear(struct rq *rq)
1011 if (hrtimer_active(&rq->hrtick_timer))
1012 hrtimer_cancel(&rq->hrtick_timer);
1016 * High-resolution timer tick.
1017 * Runs from hardirq context with interrupts disabled.
1019 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1021 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1023 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1025 spin_lock(&rq->lock);
1026 update_rq_clock(rq);
1027 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1028 spin_unlock(&rq->lock);
1030 return HRTIMER_NORESTART;
1035 * called from hardirq (IPI) context
1037 static void __hrtick_start(void *arg)
1039 struct rq *rq = arg;
1041 spin_lock(&rq->lock);
1042 hrtimer_restart(&rq->hrtick_timer);
1043 rq->hrtick_csd_pending = 0;
1044 spin_unlock(&rq->lock);
1048 * Called to set the hrtick timer state.
1050 * called with rq->lock held and irqs disabled
1052 static void hrtick_start(struct rq *rq, u64 delay)
1054 struct hrtimer *timer = &rq->hrtick_timer;
1055 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1057 hrtimer_set_expires(timer, time);
1059 if (rq == this_rq()) {
1060 hrtimer_restart(timer);
1061 } else if (!rq->hrtick_csd_pending) {
1062 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1063 rq->hrtick_csd_pending = 1;
1068 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1070 int cpu = (int)(long)hcpu;
1073 case CPU_UP_CANCELED:
1074 case CPU_UP_CANCELED_FROZEN:
1075 case CPU_DOWN_PREPARE:
1076 case CPU_DOWN_PREPARE_FROZEN:
1078 case CPU_DEAD_FROZEN:
1079 hrtick_clear(cpu_rq(cpu));
1086 static __init void init_hrtick(void)
1088 hotcpu_notifier(hotplug_hrtick, 0);
1092 * Called to set the hrtick timer state.
1094 * called with rq->lock held and irqs disabled
1096 static void hrtick_start(struct rq *rq, u64 delay)
1098 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1099 HRTIMER_MODE_REL_PINNED, 0);
1102 static inline void init_hrtick(void)
1105 #endif /* CONFIG_SMP */
1107 static void init_rq_hrtick(struct rq *rq)
1110 rq->hrtick_csd_pending = 0;
1112 rq->hrtick_csd.flags = 0;
1113 rq->hrtick_csd.func = __hrtick_start;
1114 rq->hrtick_csd.info = rq;
1117 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1118 rq->hrtick_timer.function = hrtick;
1120 #else /* CONFIG_SCHED_HRTICK */
1121 static inline void hrtick_clear(struct rq *rq)
1125 static inline void init_rq_hrtick(struct rq *rq)
1129 static inline void init_hrtick(void)
1132 #endif /* CONFIG_SCHED_HRTICK */
1135 * resched_task - mark a task 'to be rescheduled now'.
1137 * On UP this means the setting of the need_resched flag, on SMP it
1138 * might also involve a cross-CPU call to trigger the scheduler on
1143 #ifndef tsk_is_polling
1144 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1147 static void resched_task(struct task_struct *p)
1151 assert_spin_locked(&task_rq(p)->lock);
1153 if (test_tsk_need_resched(p))
1156 set_tsk_need_resched(p);
1159 if (cpu == smp_processor_id())
1162 /* NEED_RESCHED must be visible before we test polling */
1164 if (!tsk_is_polling(p))
1165 smp_send_reschedule(cpu);
1168 static void resched_cpu(int cpu)
1170 struct rq *rq = cpu_rq(cpu);
1171 unsigned long flags;
1173 if (!spin_trylock_irqsave(&rq->lock, flags))
1175 resched_task(cpu_curr(cpu));
1176 spin_unlock_irqrestore(&rq->lock, flags);
1181 * When add_timer_on() enqueues a timer into the timer wheel of an
1182 * idle CPU then this timer might expire before the next timer event
1183 * which is scheduled to wake up that CPU. In case of a completely
1184 * idle system the next event might even be infinite time into the
1185 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1186 * leaves the inner idle loop so the newly added timer is taken into
1187 * account when the CPU goes back to idle and evaluates the timer
1188 * wheel for the next timer event.
1190 void wake_up_idle_cpu(int cpu)
1192 struct rq *rq = cpu_rq(cpu);
1194 if (cpu == smp_processor_id())
1198 * This is safe, as this function is called with the timer
1199 * wheel base lock of (cpu) held. When the CPU is on the way
1200 * to idle and has not yet set rq->curr to idle then it will
1201 * be serialized on the timer wheel base lock and take the new
1202 * timer into account automatically.
1204 if (rq->curr != rq->idle)
1208 * We can set TIF_RESCHED on the idle task of the other CPU
1209 * lockless. The worst case is that the other CPU runs the
1210 * idle task through an additional NOOP schedule()
1212 set_tsk_need_resched(rq->idle);
1214 /* NEED_RESCHED must be visible before we test polling */
1216 if (!tsk_is_polling(rq->idle))
1217 smp_send_reschedule(cpu);
1219 #endif /* CONFIG_NO_HZ */
1221 static u64 sched_avg_period(void)
1223 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1226 static void sched_avg_update(struct rq *rq)
1228 s64 period = sched_avg_period();
1230 while ((s64)(rq->clock - rq->age_stamp) > period) {
1232 * Inline assembly required to prevent the compiler
1233 * optimising this loop into a divmod call.
1234 * See __iter_div_u64_rem() for another example of this.
1236 asm("" : "+rm" (rq->age_stamp));
1237 rq->age_stamp += period;
1242 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1244 rq->rt_avg += rt_delta;
1245 sched_avg_update(rq);
1248 #else /* !CONFIG_SMP */
1249 static void resched_task(struct task_struct *p)
1251 assert_spin_locked(&task_rq(p)->lock);
1252 set_tsk_need_resched(p);
1255 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1259 static void sched_avg_update(struct rq *rq)
1262 #endif /* CONFIG_SMP */
1264 #if BITS_PER_LONG == 32
1265 # define WMULT_CONST (~0UL)
1267 # define WMULT_CONST (1UL << 32)
1270 #define WMULT_SHIFT 32
1273 * Shift right and round:
1275 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1278 * delta *= weight / lw
1280 static unsigned long
1281 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1282 struct load_weight *lw)
1286 if (!lw->inv_weight) {
1287 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1290 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1294 tmp = (u64)delta_exec * weight;
1296 * Check whether we'd overflow the 64-bit multiplication:
1298 if (unlikely(tmp > WMULT_CONST))
1299 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1302 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1304 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1307 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1313 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1320 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1321 * of tasks with abnormal "nice" values across CPUs the contribution that
1322 * each task makes to its run queue's load is weighted according to its
1323 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1324 * scaled version of the new time slice allocation that they receive on time
1328 #define WEIGHT_IDLEPRIO 3
1329 #define WMULT_IDLEPRIO 1431655765
1332 * Nice levels are multiplicative, with a gentle 10% change for every
1333 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1334 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1335 * that remained on nice 0.
1337 * The "10% effect" is relative and cumulative: from _any_ nice level,
1338 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1339 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1340 * If a task goes up by ~10% and another task goes down by ~10% then
1341 * the relative distance between them is ~25%.)
1343 static const int prio_to_weight[40] = {
1344 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1345 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1346 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1347 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1348 /* 0 */ 1024, 820, 655, 526, 423,
1349 /* 5 */ 335, 272, 215, 172, 137,
1350 /* 10 */ 110, 87, 70, 56, 45,
1351 /* 15 */ 36, 29, 23, 18, 15,
1355 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1357 * In cases where the weight does not change often, we can use the
1358 * precalculated inverse to speed up arithmetics by turning divisions
1359 * into multiplications:
1361 static const u32 prio_to_wmult[40] = {
1362 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1363 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1364 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1365 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1366 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1367 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1368 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1369 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1372 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1375 * runqueue iterator, to support SMP load-balancing between different
1376 * scheduling classes, without having to expose their internal data
1377 * structures to the load-balancing proper:
1379 struct rq_iterator {
1381 struct task_struct *(*start)(void *);
1382 struct task_struct *(*next)(void *);
1386 static unsigned long
1387 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1388 unsigned long max_load_move, struct sched_domain *sd,
1389 enum cpu_idle_type idle, int *all_pinned,
1390 int *this_best_prio, struct rq_iterator *iterator);
1393 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1394 struct sched_domain *sd, enum cpu_idle_type idle,
1395 struct rq_iterator *iterator);
1398 /* Time spent by the tasks of the cpu accounting group executing in ... */
1399 enum cpuacct_stat_index {
1400 CPUACCT_STAT_USER, /* ... user mode */
1401 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1403 CPUACCT_STAT_NSTATS,
1406 #ifdef CONFIG_CGROUP_CPUACCT
1407 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1408 static void cpuacct_update_stats(struct task_struct *tsk,
1409 enum cpuacct_stat_index idx, cputime_t val);
1411 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1412 static inline void cpuacct_update_stats(struct task_struct *tsk,
1413 enum cpuacct_stat_index idx, cputime_t val) {}
1416 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1418 update_load_add(&rq->load, load);
1421 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1423 update_load_sub(&rq->load, load);
1426 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1427 typedef int (*tg_visitor)(struct task_group *, void *);
1430 * Iterate the full tree, calling @down when first entering a node and @up when
1431 * leaving it for the final time.
1433 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1435 struct task_group *parent, *child;
1439 parent = &root_task_group;
1441 ret = (*down)(parent, data);
1444 list_for_each_entry_rcu(child, &parent->children, siblings) {
1451 ret = (*up)(parent, data);
1456 parent = parent->parent;
1465 static int tg_nop(struct task_group *tg, void *data)
1472 /* Used instead of source_load when we know the type == 0 */
1473 static unsigned long weighted_cpuload(const int cpu)
1475 return cpu_rq(cpu)->load.weight;
1479 * Return a low guess at the load of a migration-source cpu weighted
1480 * according to the scheduling class and "nice" value.
1482 * We want to under-estimate the load of migration sources, to
1483 * balance conservatively.
1485 static unsigned long source_load(int cpu, int type)
1487 struct rq *rq = cpu_rq(cpu);
1488 unsigned long total = weighted_cpuload(cpu);
1490 if (type == 0 || !sched_feat(LB_BIAS))
1493 return min(rq->cpu_load[type-1], total);
1497 * Return a high guess at the load of a migration-target cpu weighted
1498 * according to the scheduling class and "nice" value.
1500 static unsigned long target_load(int cpu, int type)
1502 struct rq *rq = cpu_rq(cpu);
1503 unsigned long total = weighted_cpuload(cpu);
1505 if (type == 0 || !sched_feat(LB_BIAS))
1508 return max(rq->cpu_load[type-1], total);
1511 static struct sched_group *group_of(int cpu)
1513 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1521 static unsigned long power_of(int cpu)
1523 struct sched_group *group = group_of(cpu);
1526 return SCHED_LOAD_SCALE;
1528 return group->cpu_power;
1531 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1533 static unsigned long cpu_avg_load_per_task(int cpu)
1535 struct rq *rq = cpu_rq(cpu);
1536 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1539 rq->avg_load_per_task = rq->load.weight / nr_running;
1541 rq->avg_load_per_task = 0;
1543 return rq->avg_load_per_task;
1546 #ifdef CONFIG_FAIR_GROUP_SCHED
1548 static __read_mostly unsigned long *update_shares_data;
1550 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1553 * Calculate and set the cpu's group shares.
1555 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1556 unsigned long sd_shares,
1557 unsigned long sd_rq_weight,
1558 unsigned long *usd_rq_weight)
1560 unsigned long shares, rq_weight;
1563 rq_weight = usd_rq_weight[cpu];
1566 rq_weight = NICE_0_LOAD;
1570 * \Sum_j shares_j * rq_weight_i
1571 * shares_i = -----------------------------
1572 * \Sum_j rq_weight_j
1574 shares = (sd_shares * rq_weight) / sd_rq_weight;
1575 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1577 if (abs(shares - tg->se[cpu]->load.weight) >
1578 sysctl_sched_shares_thresh) {
1579 struct rq *rq = cpu_rq(cpu);
1580 unsigned long flags;
1582 spin_lock_irqsave(&rq->lock, flags);
1583 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1584 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1585 __set_se_shares(tg->se[cpu], shares);
1586 spin_unlock_irqrestore(&rq->lock, flags);
1591 * Re-compute the task group their per cpu shares over the given domain.
1592 * This needs to be done in a bottom-up fashion because the rq weight of a
1593 * parent group depends on the shares of its child groups.
1595 static int tg_shares_up(struct task_group *tg, void *data)
1597 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1598 unsigned long *usd_rq_weight;
1599 struct sched_domain *sd = data;
1600 unsigned long flags;
1606 local_irq_save(flags);
1607 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1609 for_each_cpu(i, sched_domain_span(sd)) {
1610 weight = tg->cfs_rq[i]->load.weight;
1611 usd_rq_weight[i] = weight;
1613 rq_weight += weight;
1615 * If there are currently no tasks on the cpu pretend there
1616 * is one of average load so that when a new task gets to
1617 * run here it will not get delayed by group starvation.
1620 weight = NICE_0_LOAD;
1622 sum_weight += weight;
1623 shares += tg->cfs_rq[i]->shares;
1627 rq_weight = sum_weight;
1629 if ((!shares && rq_weight) || shares > tg->shares)
1630 shares = tg->shares;
1632 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1633 shares = tg->shares;
1635 for_each_cpu(i, sched_domain_span(sd))
1636 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1638 local_irq_restore(flags);
1644 * Compute the cpu's hierarchical load factor for each task group.
1645 * This needs to be done in a top-down fashion because the load of a child
1646 * group is a fraction of its parents load.
1648 static int tg_load_down(struct task_group *tg, void *data)
1651 long cpu = (long)data;
1654 load = cpu_rq(cpu)->load.weight;
1656 load = tg->parent->cfs_rq[cpu]->h_load;
1657 load *= tg->cfs_rq[cpu]->shares;
1658 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1661 tg->cfs_rq[cpu]->h_load = load;
1666 static void update_shares(struct sched_domain *sd)
1671 if (root_task_group_empty())
1674 now = cpu_clock(raw_smp_processor_id());
1675 elapsed = now - sd->last_update;
1677 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1678 sd->last_update = now;
1679 walk_tg_tree(tg_nop, tg_shares_up, sd);
1683 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1685 if (root_task_group_empty())
1688 spin_unlock(&rq->lock);
1690 spin_lock(&rq->lock);
1693 static void update_h_load(long cpu)
1695 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1700 static inline void update_shares(struct sched_domain *sd)
1704 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1710 #ifdef CONFIG_PREEMPT
1712 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1715 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1716 * way at the expense of forcing extra atomic operations in all
1717 * invocations. This assures that the double_lock is acquired using the
1718 * same underlying policy as the spinlock_t on this architecture, which
1719 * reduces latency compared to the unfair variant below. However, it
1720 * also adds more overhead and therefore may reduce throughput.
1722 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1723 __releases(this_rq->lock)
1724 __acquires(busiest->lock)
1725 __acquires(this_rq->lock)
1727 spin_unlock(&this_rq->lock);
1728 double_rq_lock(this_rq, busiest);
1735 * Unfair double_lock_balance: Optimizes throughput at the expense of
1736 * latency by eliminating extra atomic operations when the locks are
1737 * already in proper order on entry. This favors lower cpu-ids and will
1738 * grant the double lock to lower cpus over higher ids under contention,
1739 * regardless of entry order into the function.
1741 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1742 __releases(this_rq->lock)
1743 __acquires(busiest->lock)
1744 __acquires(this_rq->lock)
1748 if (unlikely(!spin_trylock(&busiest->lock))) {
1749 if (busiest < this_rq) {
1750 spin_unlock(&this_rq->lock);
1751 spin_lock(&busiest->lock);
1752 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1755 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1760 #endif /* CONFIG_PREEMPT */
1763 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1765 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1767 if (unlikely(!irqs_disabled())) {
1768 /* printk() doesn't work good under rq->lock */
1769 spin_unlock(&this_rq->lock);
1773 return _double_lock_balance(this_rq, busiest);
1776 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1777 __releases(busiest->lock)
1779 spin_unlock(&busiest->lock);
1780 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1784 #ifdef CONFIG_FAIR_GROUP_SCHED
1785 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1788 cfs_rq->shares = shares;
1793 static void calc_load_account_active(struct rq *this_rq);
1794 static void update_sysctl(void);
1796 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1798 set_task_rq(p, cpu);
1801 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1802 * successfuly executed on another CPU. We must ensure that updates of
1803 * per-task data have been completed by this moment.
1806 task_thread_info(p)->cpu = cpu;
1810 #include "sched_stats.h"
1811 #include "sched_idletask.c"
1812 #include "sched_fair.c"
1813 #include "sched_rt.c"
1814 #ifdef CONFIG_SCHED_DEBUG
1815 # include "sched_debug.c"
1818 #define sched_class_highest (&rt_sched_class)
1819 #define for_each_class(class) \
1820 for (class = sched_class_highest; class; class = class->next)
1822 static void inc_nr_running(struct rq *rq)
1827 static void dec_nr_running(struct rq *rq)
1832 static void set_load_weight(struct task_struct *p)
1834 if (task_has_rt_policy(p)) {
1835 p->se.load.weight = prio_to_weight[0] * 2;
1836 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1841 * SCHED_IDLE tasks get minimal weight:
1843 if (p->policy == SCHED_IDLE) {
1844 p->se.load.weight = WEIGHT_IDLEPRIO;
1845 p->se.load.inv_weight = WMULT_IDLEPRIO;
1849 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1850 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1853 static void update_avg(u64 *avg, u64 sample)
1855 s64 diff = sample - *avg;
1860 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1863 p->se.start_runtime = p->se.sum_exec_runtime;
1865 sched_info_queued(p);
1866 p->sched_class->enqueue_task(rq, p, wakeup, head);
1870 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1873 if (p->se.last_wakeup) {
1874 update_avg(&p->se.avg_overlap,
1875 p->se.sum_exec_runtime - p->se.last_wakeup);
1876 p->se.last_wakeup = 0;
1878 update_avg(&p->se.avg_wakeup,
1879 sysctl_sched_wakeup_granularity);
1883 sched_info_dequeued(p);
1884 p->sched_class->dequeue_task(rq, p, sleep);
1889 * __normal_prio - return the priority that is based on the static prio
1891 static inline int __normal_prio(struct task_struct *p)
1893 return p->static_prio;
1897 * Calculate the expected normal priority: i.e. priority
1898 * without taking RT-inheritance into account. Might be
1899 * boosted by interactivity modifiers. Changes upon fork,
1900 * setprio syscalls, and whenever the interactivity
1901 * estimator recalculates.
1903 static inline int normal_prio(struct task_struct *p)
1907 if (task_has_rt_policy(p))
1908 prio = MAX_RT_PRIO-1 - p->rt_priority;
1910 prio = __normal_prio(p);
1915 * Calculate the current priority, i.e. the priority
1916 * taken into account by the scheduler. This value might
1917 * be boosted by RT tasks, or might be boosted by
1918 * interactivity modifiers. Will be RT if the task got
1919 * RT-boosted. If not then it returns p->normal_prio.
1921 static int effective_prio(struct task_struct *p)
1923 p->normal_prio = normal_prio(p);
1925 * If we are RT tasks or we were boosted to RT priority,
1926 * keep the priority unchanged. Otherwise, update priority
1927 * to the normal priority:
1929 if (!rt_prio(p->prio))
1930 return p->normal_prio;
1935 * activate_task - move a task to the runqueue.
1937 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1939 if (task_contributes_to_load(p))
1940 rq->nr_uninterruptible--;
1942 enqueue_task(rq, p, wakeup, false);
1947 * deactivate_task - remove a task from the runqueue.
1949 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1951 if (task_contributes_to_load(p))
1952 rq->nr_uninterruptible++;
1954 dequeue_task(rq, p, sleep);
1959 * task_curr - is this task currently executing on a CPU?
1960 * @p: the task in question.
1962 inline int task_curr(const struct task_struct *p)
1964 return cpu_curr(task_cpu(p)) == p;
1967 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1968 const struct sched_class *prev_class,
1969 int oldprio, int running)
1971 if (prev_class != p->sched_class) {
1972 if (prev_class->switched_from)
1973 prev_class->switched_from(rq, p, running);
1974 p->sched_class->switched_to(rq, p, running);
1976 p->sched_class->prio_changed(rq, p, oldprio, running);
1980 * kthread_bind - bind a just-created kthread to a cpu.
1981 * @p: thread created by kthread_create().
1982 * @cpu: cpu (might not be online, must be possible) for @k to run on.
1984 * Description: This function is equivalent to set_cpus_allowed(),
1985 * except that @cpu doesn't need to be online, and the thread must be
1986 * stopped (i.e., just returned from kthread_create()).
1988 * Function lives here instead of kthread.c because it messes with
1989 * scheduler internals which require locking.
1991 void kthread_bind(struct task_struct *p, unsigned int cpu)
1993 /* Must have done schedule() in kthread() before we set_task_cpu */
1994 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
1999 p->cpus_allowed = cpumask_of_cpu(cpu);
2000 p->rt.nr_cpus_allowed = 1;
2001 p->flags |= PF_THREAD_BOUND;
2003 EXPORT_SYMBOL(kthread_bind);
2007 * Is this task likely cache-hot:
2010 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2014 if (p->sched_class != &fair_sched_class)
2018 * Buddy candidates are cache hot:
2020 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2021 (&p->se == cfs_rq_of(&p->se)->next ||
2022 &p->se == cfs_rq_of(&p->se)->last))
2025 if (sysctl_sched_migration_cost == -1)
2027 if (sysctl_sched_migration_cost == 0)
2030 delta = now - p->se.exec_start;
2032 return delta < (s64)sysctl_sched_migration_cost;
2036 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2038 int old_cpu = task_cpu(p);
2040 #ifdef CONFIG_SCHED_DEBUG
2042 * We should never call set_task_cpu() on a blocked task,
2043 * ttwu() will sort out the placement.
2045 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2046 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2049 trace_sched_migrate_task(p, new_cpu);
2051 if (old_cpu != new_cpu) {
2052 p->se.nr_migrations++;
2053 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2057 __set_task_cpu(p, new_cpu);
2060 struct migration_req {
2061 struct list_head list;
2063 struct task_struct *task;
2066 struct completion done;
2070 * The task's runqueue lock must be held.
2071 * Returns true if you have to wait for migration thread.
2074 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2076 struct rq *rq = task_rq(p);
2079 * If the task is not on a runqueue (and not running), then
2080 * the next wake-up will properly place the task.
2082 if (!p->se.on_rq && !task_running(rq, p))
2085 init_completion(&req->done);
2087 req->dest_cpu = dest_cpu;
2088 list_add(&req->list, &rq->migration_queue);
2094 * wait_task_context_switch - wait for a thread to complete at least one
2097 * @p must not be current.
2099 void wait_task_context_switch(struct task_struct *p)
2101 unsigned long nvcsw, nivcsw, flags;
2109 * The runqueue is assigned before the actual context
2110 * switch. We need to take the runqueue lock.
2112 * We could check initially without the lock but it is
2113 * very likely that we need to take the lock in every
2116 rq = task_rq_lock(p, &flags);
2117 running = task_running(rq, p);
2118 task_rq_unlock(rq, &flags);
2120 if (likely(!running))
2123 * The switch count is incremented before the actual
2124 * context switch. We thus wait for two switches to be
2125 * sure at least one completed.
2127 if ((p->nvcsw - nvcsw) > 1)
2129 if ((p->nivcsw - nivcsw) > 1)
2137 * wait_task_inactive - wait for a thread to unschedule.
2139 * If @match_state is nonzero, it's the @p->state value just checked and
2140 * not expected to change. If it changes, i.e. @p might have woken up,
2141 * then return zero. When we succeed in waiting for @p to be off its CPU,
2142 * we return a positive number (its total switch count). If a second call
2143 * a short while later returns the same number, the caller can be sure that
2144 * @p has remained unscheduled the whole time.
2146 * The caller must ensure that the task *will* unschedule sometime soon,
2147 * else this function might spin for a *long* time. This function can't
2148 * be called with interrupts off, or it may introduce deadlock with
2149 * smp_call_function() if an IPI is sent by the same process we are
2150 * waiting to become inactive.
2152 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2154 unsigned long flags;
2161 * We do the initial early heuristics without holding
2162 * any task-queue locks at all. We'll only try to get
2163 * the runqueue lock when things look like they will
2169 * If the task is actively running on another CPU
2170 * still, just relax and busy-wait without holding
2173 * NOTE! Since we don't hold any locks, it's not
2174 * even sure that "rq" stays as the right runqueue!
2175 * But we don't care, since "task_running()" will
2176 * return false if the runqueue has changed and p
2177 * is actually now running somewhere else!
2179 while (task_running(rq, p)) {
2180 if (match_state && unlikely(p->state != match_state))
2186 * Ok, time to look more closely! We need the rq
2187 * lock now, to be *sure*. If we're wrong, we'll
2188 * just go back and repeat.
2190 rq = task_rq_lock(p, &flags);
2191 trace_sched_wait_task(rq, p);
2192 running = task_running(rq, p);
2193 on_rq = p->se.on_rq;
2195 if (!match_state || p->state == match_state)
2196 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2197 task_rq_unlock(rq, &flags);
2200 * If it changed from the expected state, bail out now.
2202 if (unlikely(!ncsw))
2206 * Was it really running after all now that we
2207 * checked with the proper locks actually held?
2209 * Oops. Go back and try again..
2211 if (unlikely(running)) {
2217 * It's not enough that it's not actively running,
2218 * it must be off the runqueue _entirely_, and not
2221 * So if it was still runnable (but just not actively
2222 * running right now), it's preempted, and we should
2223 * yield - it could be a while.
2225 if (unlikely(on_rq)) {
2226 schedule_timeout_uninterruptible(1);
2231 * Ahh, all good. It wasn't running, and it wasn't
2232 * runnable, which means that it will never become
2233 * running in the future either. We're all done!
2242 * kick_process - kick a running thread to enter/exit the kernel
2243 * @p: the to-be-kicked thread
2245 * Cause a process which is running on another CPU to enter
2246 * kernel-mode, without any delay. (to get signals handled.)
2248 * NOTE: this function doesnt have to take the runqueue lock,
2249 * because all it wants to ensure is that the remote task enters
2250 * the kernel. If the IPI races and the task has been migrated
2251 * to another CPU then no harm is done and the purpose has been
2254 void kick_process(struct task_struct *p)
2260 if ((cpu != smp_processor_id()) && task_curr(p))
2261 smp_send_reschedule(cpu);
2264 EXPORT_SYMBOL_GPL(kick_process);
2265 #endif /* CONFIG_SMP */
2268 * task_oncpu_function_call - call a function on the cpu on which a task runs
2269 * @p: the task to evaluate
2270 * @func: the function to be called
2271 * @info: the function call argument
2273 * Calls the function @func when the task is currently running. This might
2274 * be on the current CPU, which just calls the function directly
2276 void task_oncpu_function_call(struct task_struct *p,
2277 void (*func) (void *info), void *info)
2284 smp_call_function_single(cpu, func, info, 1);
2290 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2292 static int select_fallback_rq(int cpu, struct task_struct *p)
2295 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2297 /* Look for allowed, online CPU in same node. */
2298 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2299 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2302 /* Any allowed, online CPU? */
2303 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2304 if (dest_cpu < nr_cpu_ids)
2307 /* No more Mr. Nice Guy. */
2308 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2309 dest_cpu = cpuset_cpus_allowed_fallback(p);
2311 * Don't tell them about moving exiting tasks or
2312 * kernel threads (both mm NULL), since they never
2315 if (p->mm && printk_ratelimit()) {
2316 printk(KERN_INFO "process %d (%s) no "
2317 "longer affine to cpu%d\n",
2318 task_pid_nr(p), p->comm, cpu);
2326 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2329 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2331 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2334 * In order not to call set_task_cpu() on a blocking task we need
2335 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2338 * Since this is common to all placement strategies, this lives here.
2340 * [ this allows ->select_task() to simply return task_cpu(p) and
2341 * not worry about this generic constraint ]
2343 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2345 cpu = select_fallback_rq(task_cpu(p), p);
2352 * try_to_wake_up - wake up a thread
2353 * @p: the to-be-woken-up thread
2354 * @state: the mask of task states that can be woken
2355 * @sync: do a synchronous wakeup?
2357 * Put it on the run-queue if it's not already there. The "current"
2358 * thread is always on the run-queue (except when the actual
2359 * re-schedule is in progress), and as such you're allowed to do
2360 * the simpler "current->state = TASK_RUNNING" to mark yourself
2361 * runnable without the overhead of this.
2363 * returns failure only if the task is already active.
2365 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2368 int cpu, orig_cpu, this_cpu, success = 0;
2369 unsigned long flags;
2370 struct rq *rq, *orig_rq;
2372 if (!sched_feat(SYNC_WAKEUPS))
2373 wake_flags &= ~WF_SYNC;
2375 this_cpu = get_cpu();
2378 rq = orig_rq = task_rq_lock(p, &flags);
2379 update_rq_clock(rq);
2380 if (!(p->state & state))
2390 if (unlikely(task_running(rq, p)))
2394 * In order to handle concurrent wakeups and release the rq->lock
2395 * we put the task in TASK_WAKING state.
2397 * First fix up the nr_uninterruptible count:
2399 if (task_contributes_to_load(p)) {
2400 if (likely(cpu_online(orig_cpu)))
2401 rq->nr_uninterruptible--;
2403 this_rq()->nr_uninterruptible--;
2405 p->state = TASK_WAKING;
2407 if (p->sched_class->task_waking)
2408 p->sched_class->task_waking(rq, p);
2410 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2411 if (cpu != orig_cpu)
2412 set_task_cpu(p, cpu);
2413 __task_rq_unlock(rq);
2416 spin_lock(&rq->lock);
2417 update_rq_clock(rq);
2420 * We migrated the task without holding either rq->lock, however
2421 * since the task is not on the task list itself, nobody else
2422 * will try and migrate the task, hence the rq should match the
2423 * cpu we just moved it to.
2425 WARN_ON(task_cpu(p) != cpu);
2426 WARN_ON(p->state != TASK_WAKING);
2428 #ifdef CONFIG_SCHEDSTATS
2429 schedstat_inc(rq, ttwu_count);
2430 if (cpu == this_cpu)
2431 schedstat_inc(rq, ttwu_local);
2433 struct sched_domain *sd;
2434 for_each_domain(this_cpu, sd) {
2435 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2436 schedstat_inc(sd, ttwu_wake_remote);
2441 #endif /* CONFIG_SCHEDSTATS */
2444 #endif /* CONFIG_SMP */
2445 schedstat_inc(p, se.nr_wakeups);
2446 if (wake_flags & WF_SYNC)
2447 schedstat_inc(p, se.nr_wakeups_sync);
2448 if (orig_cpu != cpu)
2449 schedstat_inc(p, se.nr_wakeups_migrate);
2450 if (cpu == this_cpu)
2451 schedstat_inc(p, se.nr_wakeups_local);
2453 schedstat_inc(p, se.nr_wakeups_remote);
2454 activate_task(rq, p, 1);
2458 * Only attribute actual wakeups done by this task.
2460 if (!in_interrupt()) {
2461 struct sched_entity *se = ¤t->se;
2462 u64 sample = se->sum_exec_runtime;
2464 if (se->last_wakeup)
2465 sample -= se->last_wakeup;
2467 sample -= se->start_runtime;
2468 update_avg(&se->avg_wakeup, sample);
2470 se->last_wakeup = se->sum_exec_runtime;
2474 trace_sched_wakeup(rq, p, success);
2475 check_preempt_curr(rq, p, wake_flags);
2477 p->state = TASK_RUNNING;
2479 if (p->sched_class->task_woken)
2480 p->sched_class->task_woken(rq, p);
2482 if (unlikely(rq->idle_stamp)) {
2483 u64 delta = rq->clock - rq->idle_stamp;
2484 u64 max = 2*sysctl_sched_migration_cost;
2489 update_avg(&rq->avg_idle, delta);
2494 task_rq_unlock(rq, &flags);
2501 * wake_up_process - Wake up a specific process
2502 * @p: The process to be woken up.
2504 * Attempt to wake up the nominated process and move it to the set of runnable
2505 * processes. Returns 1 if the process was woken up, 0 if it was already
2508 * It may be assumed that this function implies a write memory barrier before
2509 * changing the task state if and only if any tasks are woken up.
2511 int wake_up_process(struct task_struct *p)
2513 return try_to_wake_up(p, TASK_ALL, 0);
2515 EXPORT_SYMBOL(wake_up_process);
2517 int wake_up_state(struct task_struct *p, unsigned int state)
2519 return try_to_wake_up(p, state, 0);
2523 * Perform scheduler related setup for a newly forked process p.
2524 * p is forked by current.
2526 * __sched_fork() is basic setup used by init_idle() too:
2528 static void __sched_fork(struct task_struct *p)
2530 p->se.exec_start = 0;
2531 p->se.sum_exec_runtime = 0;
2532 p->se.prev_sum_exec_runtime = 0;
2533 p->se.nr_migrations = 0;
2534 p->se.last_wakeup = 0;
2535 p->se.avg_overlap = 0;
2536 p->se.start_runtime = 0;
2537 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2538 p->se.avg_running = 0;
2540 #ifdef CONFIG_SCHEDSTATS
2541 p->se.wait_start = 0;
2543 p->se.wait_count = 0;
2546 p->se.sleep_start = 0;
2547 p->se.sleep_max = 0;
2548 p->se.sum_sleep_runtime = 0;
2550 p->se.block_start = 0;
2551 p->se.block_max = 0;
2553 p->se.slice_max = 0;
2555 p->se.nr_migrations_cold = 0;
2556 p->se.nr_failed_migrations_affine = 0;
2557 p->se.nr_failed_migrations_running = 0;
2558 p->se.nr_failed_migrations_hot = 0;
2559 p->se.nr_forced_migrations = 0;
2561 p->se.nr_wakeups = 0;
2562 p->se.nr_wakeups_sync = 0;
2563 p->se.nr_wakeups_migrate = 0;
2564 p->se.nr_wakeups_local = 0;
2565 p->se.nr_wakeups_remote = 0;
2566 p->se.nr_wakeups_affine = 0;
2567 p->se.nr_wakeups_affine_attempts = 0;
2568 p->se.nr_wakeups_passive = 0;
2569 p->se.nr_wakeups_idle = 0;
2573 INIT_LIST_HEAD(&p->rt.run_list);
2575 INIT_LIST_HEAD(&p->se.group_node);
2577 #ifdef CONFIG_PREEMPT_NOTIFIERS
2578 INIT_HLIST_HEAD(&p->preempt_notifiers);
2583 * fork()/clone()-time setup:
2585 void sched_fork(struct task_struct *p, int clone_flags)
2587 int cpu = get_cpu();
2591 * We mark the process as running here. This guarantees that
2592 * nobody will actually run it, and a signal or other external
2593 * event cannot wake it up and insert it on the runqueue either.
2595 p->state = TASK_RUNNING;
2598 * Revert to default priority/policy on fork if requested.
2600 if (unlikely(p->sched_reset_on_fork)) {
2601 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2602 p->policy = SCHED_NORMAL;
2603 p->normal_prio = p->static_prio;
2606 if (PRIO_TO_NICE(p->static_prio) < 0) {
2607 p->static_prio = NICE_TO_PRIO(0);
2608 p->normal_prio = p->static_prio;
2613 * We don't need the reset flag anymore after the fork. It has
2614 * fulfilled its duty:
2616 p->sched_reset_on_fork = 0;
2620 * Make sure we do not leak PI boosting priority to the child.
2622 p->prio = current->normal_prio;
2624 if (!rt_prio(p->prio))
2625 p->sched_class = &fair_sched_class;
2627 if (p->sched_class->task_fork)
2628 p->sched_class->task_fork(p);
2630 set_task_cpu(p, cpu);
2632 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2633 if (likely(sched_info_on()))
2634 memset(&p->sched_info, 0, sizeof(p->sched_info));
2636 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2639 #ifdef CONFIG_PREEMPT
2640 /* Want to start with kernel preemption disabled. */
2641 task_thread_info(p)->preempt_count = 1;
2643 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2649 * wake_up_new_task - wake up a newly created task for the first time.
2651 * This function will do some initial scheduler statistics housekeeping
2652 * that must be done for every newly created context, then puts the task
2653 * on the runqueue and wakes it.
2655 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2657 unsigned long flags;
2659 int cpu = get_cpu();
2662 rq = task_rq_lock(p, &flags);
2663 p->state = TASK_WAKING;
2666 * Fork balancing, do it here and not earlier because:
2667 * - cpus_allowed can change in the fork path
2668 * - any previously selected cpu might disappear through hotplug
2670 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2671 * without people poking at ->cpus_allowed.
2673 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2674 set_task_cpu(p, cpu);
2676 p->state = TASK_RUNNING;
2677 task_rq_unlock(rq, &flags);
2680 rq = task_rq_lock(p, &flags);
2681 update_rq_clock(rq);
2682 activate_task(rq, p, 0);
2683 trace_sched_wakeup_new(rq, p, 1);
2684 check_preempt_curr(rq, p, WF_FORK);
2686 if (p->sched_class->task_woken)
2687 p->sched_class->task_woken(rq, p);
2689 task_rq_unlock(rq, &flags);
2693 #ifdef CONFIG_PREEMPT_NOTIFIERS
2696 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2697 * @notifier: notifier struct to register
2699 void preempt_notifier_register(struct preempt_notifier *notifier)
2701 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2703 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2706 * preempt_notifier_unregister - no longer interested in preemption notifications
2707 * @notifier: notifier struct to unregister
2709 * This is safe to call from within a preemption notifier.
2711 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2713 hlist_del(¬ifier->link);
2715 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2717 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2719 struct preempt_notifier *notifier;
2720 struct hlist_node *node;
2722 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2723 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2727 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2728 struct task_struct *next)
2730 struct preempt_notifier *notifier;
2731 struct hlist_node *node;
2733 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2734 notifier->ops->sched_out(notifier, next);
2737 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2739 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2744 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2745 struct task_struct *next)
2749 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2752 * prepare_task_switch - prepare to switch tasks
2753 * @rq: the runqueue preparing to switch
2754 * @prev: the current task that is being switched out
2755 * @next: the task we are going to switch to.
2757 * This is called with the rq lock held and interrupts off. It must
2758 * be paired with a subsequent finish_task_switch after the context
2761 * prepare_task_switch sets up locking and calls architecture specific
2765 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2766 struct task_struct *next)
2768 fire_sched_out_preempt_notifiers(prev, next);
2769 prepare_lock_switch(rq, next);
2770 prepare_arch_switch(next);
2774 * finish_task_switch - clean up after a task-switch
2775 * @rq: runqueue associated with task-switch
2776 * @prev: the thread we just switched away from.
2778 * finish_task_switch must be called after the context switch, paired
2779 * with a prepare_task_switch call before the context switch.
2780 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2781 * and do any other architecture-specific cleanup actions.
2783 * Note that we may have delayed dropping an mm in context_switch(). If
2784 * so, we finish that here outside of the runqueue lock. (Doing it
2785 * with the lock held can cause deadlocks; see schedule() for
2788 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2789 __releases(rq->lock)
2791 struct mm_struct *mm = rq->prev_mm;
2797 * A task struct has one reference for the use as "current".
2798 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2799 * schedule one last time. The schedule call will never return, and
2800 * the scheduled task must drop that reference.
2801 * The test for TASK_DEAD must occur while the runqueue locks are
2802 * still held, otherwise prev could be scheduled on another cpu, die
2803 * there before we look at prev->state, and then the reference would
2805 * Manfred Spraul <manfred@colorfullife.com>
2807 prev_state = prev->state;
2808 finish_arch_switch(prev);
2809 perf_event_task_sched_in(current, cpu_of(rq));
2810 finish_lock_switch(rq, prev);
2812 fire_sched_in_preempt_notifiers(current);
2815 if (unlikely(prev_state == TASK_DEAD)) {
2817 * Remove function-return probe instances associated with this
2818 * task and put them back on the free list.
2820 kprobe_flush_task(prev);
2821 put_task_struct(prev);
2827 /* assumes rq->lock is held */
2828 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2830 if (prev->sched_class->pre_schedule)
2831 prev->sched_class->pre_schedule(rq, prev);
2834 /* rq->lock is NOT held, but preemption is disabled */
2835 static inline void post_schedule(struct rq *rq)
2837 if (rq->post_schedule) {
2838 unsigned long flags;
2840 spin_lock_irqsave(&rq->lock, flags);
2841 if (rq->curr->sched_class->post_schedule)
2842 rq->curr->sched_class->post_schedule(rq);
2843 spin_unlock_irqrestore(&rq->lock, flags);
2845 rq->post_schedule = 0;
2851 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2855 static inline void post_schedule(struct rq *rq)
2862 * schedule_tail - first thing a freshly forked thread must call.
2863 * @prev: the thread we just switched away from.
2865 asmlinkage void schedule_tail(struct task_struct *prev)
2866 __releases(rq->lock)
2868 struct rq *rq = this_rq();
2870 finish_task_switch(rq, prev);
2873 * FIXME: do we need to worry about rq being invalidated by the
2878 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2879 /* In this case, finish_task_switch does not reenable preemption */
2882 if (current->set_child_tid)
2883 put_user(task_pid_vnr(current), current->set_child_tid);
2887 * context_switch - switch to the new MM and the new
2888 * thread's register state.
2891 context_switch(struct rq *rq, struct task_struct *prev,
2892 struct task_struct *next)
2894 struct mm_struct *mm, *oldmm;
2896 prepare_task_switch(rq, prev, next);
2897 trace_sched_switch(rq, prev, next);
2899 oldmm = prev->active_mm;
2901 * For paravirt, this is coupled with an exit in switch_to to
2902 * combine the page table reload and the switch backend into
2905 arch_start_context_switch(prev);
2907 if (unlikely(!mm)) {
2908 next->active_mm = oldmm;
2909 atomic_inc(&oldmm->mm_count);
2910 enter_lazy_tlb(oldmm, next);
2912 switch_mm(oldmm, mm, next);
2914 if (unlikely(!prev->mm)) {
2915 prev->active_mm = NULL;
2916 rq->prev_mm = oldmm;
2919 * Since the runqueue lock will be released by the next
2920 * task (which is an invalid locking op but in the case
2921 * of the scheduler it's an obvious special-case), so we
2922 * do an early lockdep release here:
2924 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2925 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2928 /* Here we just switch the register state and the stack. */
2929 switch_to(prev, next, prev);
2933 * this_rq must be evaluated again because prev may have moved
2934 * CPUs since it called schedule(), thus the 'rq' on its stack
2935 * frame will be invalid.
2937 finish_task_switch(this_rq(), prev);
2941 * nr_running, nr_uninterruptible and nr_context_switches:
2943 * externally visible scheduler statistics: current number of runnable
2944 * threads, current number of uninterruptible-sleeping threads, total
2945 * number of context switches performed since bootup.
2947 unsigned long nr_running(void)
2949 unsigned long i, sum = 0;
2951 for_each_online_cpu(i)
2952 sum += cpu_rq(i)->nr_running;
2957 unsigned long nr_uninterruptible(void)
2959 unsigned long i, sum = 0;
2961 for_each_possible_cpu(i)
2962 sum += cpu_rq(i)->nr_uninterruptible;
2965 * Since we read the counters lockless, it might be slightly
2966 * inaccurate. Do not allow it to go below zero though:
2968 if (unlikely((long)sum < 0))
2974 unsigned long long nr_context_switches(void)
2977 unsigned long long sum = 0;
2979 for_each_possible_cpu(i)
2980 sum += cpu_rq(i)->nr_switches;
2985 unsigned long nr_iowait(void)
2987 unsigned long i, sum = 0;
2989 for_each_possible_cpu(i)
2990 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2995 unsigned long nr_iowait_cpu(void)
2997 struct rq *this = this_rq();
2998 return atomic_read(&this->nr_iowait);
3001 unsigned long this_cpu_load(void)
3003 struct rq *this = this_rq();
3004 return this->cpu_load[0];
3008 /* Variables and functions for calc_load */
3009 static atomic_long_t calc_load_tasks;
3010 static unsigned long calc_load_update;
3011 unsigned long avenrun[3];
3012 EXPORT_SYMBOL(avenrun);
3015 * get_avenrun - get the load average array
3016 * @loads: pointer to dest load array
3017 * @offset: offset to add
3018 * @shift: shift count to shift the result left
3020 * These values are estimates at best, so no need for locking.
3022 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3024 loads[0] = (avenrun[0] + offset) << shift;
3025 loads[1] = (avenrun[1] + offset) << shift;
3026 loads[2] = (avenrun[2] + offset) << shift;
3029 static unsigned long
3030 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3033 load += active * (FIXED_1 - exp);
3034 return load >> FSHIFT;
3038 * calc_load - update the avenrun load estimates 10 ticks after the
3039 * CPUs have updated calc_load_tasks.
3041 void calc_global_load(void)
3043 unsigned long upd = calc_load_update + 10;
3046 if (time_before(jiffies, upd))
3049 active = atomic_long_read(&calc_load_tasks);
3050 active = active > 0 ? active * FIXED_1 : 0;
3052 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3053 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3054 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3056 calc_load_update += LOAD_FREQ;
3060 * Either called from update_cpu_load() or from a cpu going idle
3062 static void calc_load_account_active(struct rq *this_rq)
3064 long nr_active, delta;
3066 nr_active = this_rq->nr_running;
3067 nr_active += (long) this_rq->nr_uninterruptible;
3069 if (nr_active != this_rq->calc_load_active) {
3070 delta = nr_active - this_rq->calc_load_active;
3071 this_rq->calc_load_active = nr_active;
3072 atomic_long_add(delta, &calc_load_tasks);
3077 * Update rq->cpu_load[] statistics. This function is usually called every
3078 * scheduler tick (TICK_NSEC).
3080 static void update_cpu_load(struct rq *this_rq)
3082 unsigned long this_load = this_rq->load.weight;
3085 this_rq->nr_load_updates++;
3087 /* Update our load: */
3088 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3089 unsigned long old_load, new_load;
3091 /* scale is effectively 1 << i now, and >> i divides by scale */
3093 old_load = this_rq->cpu_load[i];
3094 new_load = this_load;
3096 * Round up the averaging division if load is increasing. This
3097 * prevents us from getting stuck on 9 if the load is 10, for
3100 if (new_load > old_load)
3101 new_load += scale-1;
3102 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3105 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3106 this_rq->calc_load_update += LOAD_FREQ;
3107 calc_load_account_active(this_rq);
3110 sched_avg_update(this_rq);
3116 * double_rq_lock - safely lock two runqueues
3118 * Note this does not disable interrupts like task_rq_lock,
3119 * you need to do so manually before calling.
3121 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3122 __acquires(rq1->lock)
3123 __acquires(rq2->lock)
3125 BUG_ON(!irqs_disabled());
3127 spin_lock(&rq1->lock);
3128 __acquire(rq2->lock); /* Fake it out ;) */
3131 spin_lock(&rq1->lock);
3132 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3134 spin_lock(&rq2->lock);
3135 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3138 update_rq_clock(rq1);
3139 update_rq_clock(rq2);
3143 * double_rq_unlock - safely unlock two runqueues
3145 * Note this does not restore interrupts like task_rq_unlock,
3146 * you need to do so manually after calling.
3148 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3149 __releases(rq1->lock)
3150 __releases(rq2->lock)
3152 spin_unlock(&rq1->lock);
3154 spin_unlock(&rq2->lock);
3156 __release(rq2->lock);
3160 * sched_exec - execve() is a valuable balancing opportunity, because at
3161 * this point the task has the smallest effective memory and cache footprint.
3163 void sched_exec(void)
3165 struct task_struct *p = current;
3166 struct migration_req req;
3167 unsigned long flags;
3171 rq = task_rq_lock(p, &flags);
3172 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3173 if (dest_cpu == smp_processor_id())
3177 * select_task_rq() can race against ->cpus_allowed
3179 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3180 likely(cpu_active(dest_cpu)) &&
3181 migrate_task(p, dest_cpu, &req)) {
3182 /* Need to wait for migration thread (might exit: take ref). */
3183 struct task_struct *mt = rq->migration_thread;
3185 get_task_struct(mt);
3186 task_rq_unlock(rq, &flags);
3187 wake_up_process(mt);
3188 put_task_struct(mt);
3189 wait_for_completion(&req.done);
3194 task_rq_unlock(rq, &flags);
3198 * pull_task - move a task from a remote runqueue to the local runqueue.
3199 * Both runqueues must be locked.
3201 static void pull_task(struct rq *src_rq, struct task_struct *p,
3202 struct rq *this_rq, int this_cpu)
3204 deactivate_task(src_rq, p, 0);
3205 set_task_cpu(p, this_cpu);
3206 activate_task(this_rq, p, 0);
3207 check_preempt_curr(this_rq, p, 0);
3211 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3214 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3215 struct sched_domain *sd, enum cpu_idle_type idle,
3218 int tsk_cache_hot = 0;
3220 * We do not migrate tasks that are:
3221 * 1) running (obviously), or
3222 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3223 * 3) are cache-hot on their current CPU.
3225 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3226 schedstat_inc(p, se.nr_failed_migrations_affine);
3231 if (task_running(rq, p)) {
3232 schedstat_inc(p, se.nr_failed_migrations_running);
3237 * Aggressive migration if:
3238 * 1) task is cache cold, or
3239 * 2) too many balance attempts have failed.
3242 tsk_cache_hot = task_hot(p, rq->clock, sd);
3243 if (!tsk_cache_hot ||
3244 sd->nr_balance_failed > sd->cache_nice_tries) {
3245 #ifdef CONFIG_SCHEDSTATS
3246 if (tsk_cache_hot) {
3247 schedstat_inc(sd, lb_hot_gained[idle]);
3248 schedstat_inc(p, se.nr_forced_migrations);
3254 if (tsk_cache_hot) {
3255 schedstat_inc(p, se.nr_failed_migrations_hot);
3261 static unsigned long
3262 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3263 unsigned long max_load_move, struct sched_domain *sd,
3264 enum cpu_idle_type idle, int *all_pinned,
3265 int *this_best_prio, struct rq_iterator *iterator)
3267 int loops = 0, pulled = 0, pinned = 0;
3268 struct task_struct *p;
3269 long rem_load_move = max_load_move;
3271 if (max_load_move == 0)
3277 * Start the load-balancing iterator:
3279 p = iterator->start(iterator->arg);
3281 if (!p || loops++ > sysctl_sched_nr_migrate)
3284 if ((p->se.load.weight >> 1) > rem_load_move ||
3285 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3286 p = iterator->next(iterator->arg);
3290 pull_task(busiest, p, this_rq, this_cpu);
3292 rem_load_move -= p->se.load.weight;
3294 #ifdef CONFIG_PREEMPT
3296 * NEWIDLE balancing is a source of latency, so preemptible kernels
3297 * will stop after the first task is pulled to minimize the critical
3300 if (idle == CPU_NEWLY_IDLE)
3305 * We only want to steal up to the prescribed amount of weighted load.
3307 if (rem_load_move > 0) {
3308 if (p->prio < *this_best_prio)
3309 *this_best_prio = p->prio;
3310 p = iterator->next(iterator->arg);
3315 * Right now, this is one of only two places pull_task() is called,
3316 * so we can safely collect pull_task() stats here rather than
3317 * inside pull_task().
3319 schedstat_add(sd, lb_gained[idle], pulled);
3322 *all_pinned = pinned;
3324 return max_load_move - rem_load_move;
3328 * move_tasks tries to move up to max_load_move weighted load from busiest to
3329 * this_rq, as part of a balancing operation within domain "sd".
3330 * Returns 1 if successful and 0 otherwise.
3332 * Called with both runqueues locked.
3334 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3335 unsigned long max_load_move,
3336 struct sched_domain *sd, enum cpu_idle_type idle,
3339 const struct sched_class *class = sched_class_highest;
3340 unsigned long total_load_moved = 0;
3341 int this_best_prio = this_rq->curr->prio;
3345 class->load_balance(this_rq, this_cpu, busiest,
3346 max_load_move - total_load_moved,
3347 sd, idle, all_pinned, &this_best_prio);
3348 class = class->next;
3350 #ifdef CONFIG_PREEMPT
3352 * NEWIDLE balancing is a source of latency, so preemptible
3353 * kernels will stop after the first task is pulled to minimize
3354 * the critical section.
3356 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3359 } while (class && max_load_move > total_load_moved);
3361 return total_load_moved > 0;
3365 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3366 struct sched_domain *sd, enum cpu_idle_type idle,
3367 struct rq_iterator *iterator)
3369 struct task_struct *p = iterator->start(iterator->arg);
3373 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3374 pull_task(busiest, p, this_rq, this_cpu);
3376 * Right now, this is only the second place pull_task()
3377 * is called, so we can safely collect pull_task()
3378 * stats here rather than inside pull_task().
3380 schedstat_inc(sd, lb_gained[idle]);
3384 p = iterator->next(iterator->arg);
3391 * move_one_task tries to move exactly one task from busiest to this_rq, as
3392 * part of active balancing operations within "domain".
3393 * Returns 1 if successful and 0 otherwise.
3395 * Called with both runqueues locked.
3397 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3398 struct sched_domain *sd, enum cpu_idle_type idle)
3400 const struct sched_class *class;
3402 for_each_class(class) {
3403 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3409 /********** Helpers for find_busiest_group ************************/
3411 * sd_lb_stats - Structure to store the statistics of a sched_domain
3412 * during load balancing.
3414 struct sd_lb_stats {
3415 struct sched_group *busiest; /* Busiest group in this sd */
3416 struct sched_group *this; /* Local group in this sd */
3417 unsigned long total_load; /* Total load of all groups in sd */
3418 unsigned long total_pwr; /* Total power of all groups in sd */
3419 unsigned long avg_load; /* Average load across all groups in sd */
3421 /** Statistics of this group */
3422 unsigned long this_load;
3423 unsigned long this_load_per_task;
3424 unsigned long this_nr_running;
3426 /* Statistics of the busiest group */
3427 unsigned long max_load;
3428 unsigned long busiest_load_per_task;
3429 unsigned long busiest_nr_running;
3430 unsigned long busiest_group_capacity;
3432 int group_imb; /* Is there imbalance in this sd */
3433 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3434 int power_savings_balance; /* Is powersave balance needed for this sd */
3435 struct sched_group *group_min; /* Least loaded group in sd */
3436 struct sched_group *group_leader; /* Group which relieves group_min */
3437 unsigned long min_load_per_task; /* load_per_task in group_min */
3438 unsigned long leader_nr_running; /* Nr running of group_leader */
3439 unsigned long min_nr_running; /* Nr running of group_min */
3444 * sg_lb_stats - stats of a sched_group required for load_balancing
3446 struct sg_lb_stats {
3447 unsigned long avg_load; /*Avg load across the CPUs of the group */
3448 unsigned long group_load; /* Total load over the CPUs of the group */
3449 unsigned long sum_nr_running; /* Nr tasks running in the group */
3450 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3451 unsigned long group_capacity;
3452 int group_imb; /* Is there an imbalance in the group ? */
3456 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3457 * @group: The group whose first cpu is to be returned.
3459 static inline unsigned int group_first_cpu(struct sched_group *group)
3461 return cpumask_first(sched_group_cpus(group));
3465 * get_sd_load_idx - Obtain the load index for a given sched domain.
3466 * @sd: The sched_domain whose load_idx is to be obtained.
3467 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3469 static inline int get_sd_load_idx(struct sched_domain *sd,
3470 enum cpu_idle_type idle)
3476 load_idx = sd->busy_idx;
3479 case CPU_NEWLY_IDLE:
3480 load_idx = sd->newidle_idx;
3483 load_idx = sd->idle_idx;
3491 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3493 * init_sd_power_savings_stats - Initialize power savings statistics for
3494 * the given sched_domain, during load balancing.
3496 * @sd: Sched domain whose power-savings statistics are to be initialized.
3497 * @sds: Variable containing the statistics for sd.
3498 * @idle: Idle status of the CPU at which we're performing load-balancing.
3500 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3501 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3504 * Busy processors will not participate in power savings
3507 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3508 sds->power_savings_balance = 0;
3510 sds->power_savings_balance = 1;
3511 sds->min_nr_running = ULONG_MAX;
3512 sds->leader_nr_running = 0;
3517 * update_sd_power_savings_stats - Update the power saving stats for a
3518 * sched_domain while performing load balancing.
3520 * @group: sched_group belonging to the sched_domain under consideration.
3521 * @sds: Variable containing the statistics of the sched_domain
3522 * @local_group: Does group contain the CPU for which we're performing
3524 * @sgs: Variable containing the statistics of the group.
3526 static inline void update_sd_power_savings_stats(struct sched_group *group,
3527 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3530 if (!sds->power_savings_balance)
3534 * If the local group is idle or completely loaded
3535 * no need to do power savings balance at this domain
3537 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3538 !sds->this_nr_running))
3539 sds->power_savings_balance = 0;
3542 * If a group is already running at full capacity or idle,
3543 * don't include that group in power savings calculations
3545 if (!sds->power_savings_balance ||
3546 sgs->sum_nr_running >= sgs->group_capacity ||
3547 !sgs->sum_nr_running)
3551 * Calculate the group which has the least non-idle load.
3552 * This is the group from where we need to pick up the load
3555 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3556 (sgs->sum_nr_running == sds->min_nr_running &&
3557 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3558 sds->group_min = group;
3559 sds->min_nr_running = sgs->sum_nr_running;
3560 sds->min_load_per_task = sgs->sum_weighted_load /
3561 sgs->sum_nr_running;
3565 * Calculate the group which is almost near its
3566 * capacity but still has some space to pick up some load
3567 * from other group and save more power
3569 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3572 if (sgs->sum_nr_running > sds->leader_nr_running ||
3573 (sgs->sum_nr_running == sds->leader_nr_running &&
3574 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3575 sds->group_leader = group;
3576 sds->leader_nr_running = sgs->sum_nr_running;
3581 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3582 * @sds: Variable containing the statistics of the sched_domain
3583 * under consideration.
3584 * @this_cpu: Cpu at which we're currently performing load-balancing.
3585 * @imbalance: Variable to store the imbalance.
3588 * Check if we have potential to perform some power-savings balance.
3589 * If yes, set the busiest group to be the least loaded group in the
3590 * sched_domain, so that it's CPUs can be put to idle.
3592 * Returns 1 if there is potential to perform power-savings balance.
3595 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3596 int this_cpu, unsigned long *imbalance)
3598 if (!sds->power_savings_balance)
3601 if (sds->this != sds->group_leader ||
3602 sds->group_leader == sds->group_min)
3605 *imbalance = sds->min_load_per_task;
3606 sds->busiest = sds->group_min;
3611 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3612 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3613 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3618 static inline void update_sd_power_savings_stats(struct sched_group *group,
3619 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3624 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3625 int this_cpu, unsigned long *imbalance)
3629 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3632 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3634 return SCHED_LOAD_SCALE;
3637 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3639 return default_scale_freq_power(sd, cpu);
3642 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3644 unsigned long weight = sd->span_weight;
3645 unsigned long smt_gain = sd->smt_gain;
3652 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3654 return default_scale_smt_power(sd, cpu);
3657 unsigned long scale_rt_power(int cpu)
3659 struct rq *rq = cpu_rq(cpu);
3660 u64 total, available;
3662 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3663 available = total - rq->rt_avg;
3665 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3666 total = SCHED_LOAD_SCALE;
3668 total >>= SCHED_LOAD_SHIFT;
3670 return div_u64(available, total);
3673 static void update_cpu_power(struct sched_domain *sd, int cpu)
3675 unsigned long weight = sd->span_weight;
3676 unsigned long power = SCHED_LOAD_SCALE;
3677 struct sched_group *sdg = sd->groups;
3679 if (sched_feat(ARCH_POWER))
3680 power *= arch_scale_freq_power(sd, cpu);
3682 power *= default_scale_freq_power(sd, cpu);
3684 power >>= SCHED_LOAD_SHIFT;
3686 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3687 if (sched_feat(ARCH_POWER))
3688 power *= arch_scale_smt_power(sd, cpu);
3690 power *= default_scale_smt_power(sd, cpu);
3692 power >>= SCHED_LOAD_SHIFT;
3695 power *= scale_rt_power(cpu);
3696 power >>= SCHED_LOAD_SHIFT;
3701 sdg->cpu_power = power;
3704 static void update_group_power(struct sched_domain *sd, int cpu)
3706 struct sched_domain *child = sd->child;
3707 struct sched_group *group, *sdg = sd->groups;
3708 unsigned long power;
3711 update_cpu_power(sd, cpu);
3717 group = child->groups;
3719 power += group->cpu_power;
3720 group = group->next;
3721 } while (group != child->groups);
3723 sdg->cpu_power = power;
3727 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3728 * @sd: The sched_domain whose statistics are to be updated.
3729 * @group: sched_group whose statistics are to be updated.
3730 * @this_cpu: Cpu for which load balance is currently performed.
3731 * @idle: Idle status of this_cpu
3732 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3733 * @sd_idle: Idle status of the sched_domain containing group.
3734 * @local_group: Does group contain this_cpu.
3735 * @cpus: Set of cpus considered for load balancing.
3736 * @balance: Should we balance.
3737 * @sgs: variable to hold the statistics for this group.
3739 static inline void update_sg_lb_stats(struct sched_domain *sd,
3740 struct sched_group *group, int this_cpu,
3741 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3742 int local_group, const struct cpumask *cpus,
3743 int *balance, struct sg_lb_stats *sgs)
3745 unsigned long load, max_cpu_load, min_cpu_load;
3747 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3748 unsigned long avg_load_per_task = 0;
3751 balance_cpu = group_first_cpu(group);
3752 if (balance_cpu == this_cpu)
3753 update_group_power(sd, this_cpu);
3756 /* Tally up the load of all CPUs in the group */
3758 min_cpu_load = ~0UL;
3760 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3761 struct rq *rq = cpu_rq(i);
3763 if (*sd_idle && rq->nr_running)
3766 /* Bias balancing toward cpus of our domain */
3768 if (idle_cpu(i) && !first_idle_cpu) {
3773 load = target_load(i, load_idx);
3775 load = source_load(i, load_idx);
3776 if (load > max_cpu_load)
3777 max_cpu_load = load;
3778 if (min_cpu_load > load)
3779 min_cpu_load = load;
3782 sgs->group_load += load;
3783 sgs->sum_nr_running += rq->nr_running;
3784 sgs->sum_weighted_load += weighted_cpuload(i);
3789 * First idle cpu or the first cpu(busiest) in this sched group
3790 * is eligible for doing load balancing at this and above
3791 * domains. In the newly idle case, we will allow all the cpu's
3792 * to do the newly idle load balance.
3794 if (idle != CPU_NEWLY_IDLE && local_group &&
3795 balance_cpu != this_cpu && balance) {
3800 /* Adjust by relative CPU power of the group */
3801 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3804 * Consider the group unbalanced when the imbalance is larger
3805 * than the average weight of two tasks.
3807 * APZ: with cgroup the avg task weight can vary wildly and
3808 * might not be a suitable number - should we keep a
3809 * normalized nr_running number somewhere that negates
3812 if (sgs->sum_nr_running)
3813 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3815 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3818 sgs->group_capacity =
3819 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3823 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3824 * @sd: sched_domain whose statistics are to be updated.
3825 * @this_cpu: Cpu for which load balance is currently performed.
3826 * @idle: Idle status of this_cpu
3827 * @sd_idle: Idle status of the sched_domain containing group.
3828 * @cpus: Set of cpus considered for load balancing.
3829 * @balance: Should we balance.
3830 * @sds: variable to hold the statistics for this sched_domain.
3832 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3833 enum cpu_idle_type idle, int *sd_idle,
3834 const struct cpumask *cpus, int *balance,
3835 struct sd_lb_stats *sds)
3837 struct sched_domain *child = sd->child;
3838 struct sched_group *group = sd->groups;
3839 struct sg_lb_stats sgs;
3840 int load_idx, prefer_sibling = 0;
3842 if (child && child->flags & SD_PREFER_SIBLING)
3845 init_sd_power_savings_stats(sd, sds, idle);
3846 load_idx = get_sd_load_idx(sd, idle);
3851 local_group = cpumask_test_cpu(this_cpu,
3852 sched_group_cpus(group));
3853 memset(&sgs, 0, sizeof(sgs));
3854 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3855 local_group, cpus, balance, &sgs);
3857 if (local_group && balance && !(*balance))
3860 sds->total_load += sgs.group_load;
3861 sds->total_pwr += group->cpu_power;
3864 * In case the child domain prefers tasks go to siblings
3865 * first, lower the group capacity to one so that we'll try
3866 * and move all the excess tasks away.
3869 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3872 sds->this_load = sgs.avg_load;
3874 sds->this_nr_running = sgs.sum_nr_running;
3875 sds->this_load_per_task = sgs.sum_weighted_load;
3876 } else if (sgs.avg_load > sds->max_load &&
3877 (sgs.sum_nr_running > sgs.group_capacity ||
3879 sds->max_load = sgs.avg_load;
3880 sds->busiest = group;
3881 sds->busiest_nr_running = sgs.sum_nr_running;
3882 sds->busiest_group_capacity = sgs.group_capacity;
3883 sds->busiest_load_per_task = sgs.sum_weighted_load;
3884 sds->group_imb = sgs.group_imb;
3887 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3888 group = group->next;
3889 } while (group != sd->groups);
3893 * fix_small_imbalance - Calculate the minor imbalance that exists
3894 * amongst the groups of a sched_domain, during
3896 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3897 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3898 * @imbalance: Variable to store the imbalance.
3900 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3901 int this_cpu, unsigned long *imbalance)
3903 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3904 unsigned int imbn = 2;
3905 unsigned long scaled_busy_load_per_task;
3907 if (sds->this_nr_running) {
3908 sds->this_load_per_task /= sds->this_nr_running;
3909 if (sds->busiest_load_per_task >
3910 sds->this_load_per_task)
3913 sds->this_load_per_task =
3914 cpu_avg_load_per_task(this_cpu);
3916 scaled_busy_load_per_task = sds->busiest_load_per_task
3918 scaled_busy_load_per_task /= sds->busiest->cpu_power;
3920 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3921 (scaled_busy_load_per_task * imbn)) {
3922 *imbalance = sds->busiest_load_per_task;
3927 * OK, we don't have enough imbalance to justify moving tasks,
3928 * however we may be able to increase total CPU power used by
3932 pwr_now += sds->busiest->cpu_power *
3933 min(sds->busiest_load_per_task, sds->max_load);
3934 pwr_now += sds->this->cpu_power *
3935 min(sds->this_load_per_task, sds->this_load);
3936 pwr_now /= SCHED_LOAD_SCALE;
3938 /* Amount of load we'd subtract */
3939 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3940 sds->busiest->cpu_power;
3941 if (sds->max_load > tmp)
3942 pwr_move += sds->busiest->cpu_power *
3943 min(sds->busiest_load_per_task, sds->max_load - tmp);
3945 /* Amount of load we'd add */
3946 if (sds->max_load * sds->busiest->cpu_power <
3947 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3948 tmp = (sds->max_load * sds->busiest->cpu_power) /
3949 sds->this->cpu_power;
3951 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3952 sds->this->cpu_power;
3953 pwr_move += sds->this->cpu_power *
3954 min(sds->this_load_per_task, sds->this_load + tmp);
3955 pwr_move /= SCHED_LOAD_SCALE;
3957 /* Move if we gain throughput */
3958 if (pwr_move > pwr_now)
3959 *imbalance = sds->busiest_load_per_task;
3963 * calculate_imbalance - Calculate the amount of imbalance present within the
3964 * groups of a given sched_domain during load balance.
3965 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3966 * @this_cpu: Cpu for which currently load balance is being performed.
3967 * @imbalance: The variable to store the imbalance.
3969 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3970 unsigned long *imbalance)
3972 unsigned long max_pull, load_above_capacity = ~0UL;
3974 sds->busiest_load_per_task /= sds->busiest_nr_running;
3975 if (sds->group_imb) {
3976 sds->busiest_load_per_task =
3977 min(sds->busiest_load_per_task, sds->avg_load);
3981 * In the presence of smp nice balancing, certain scenarios can have
3982 * max load less than avg load(as we skip the groups at or below
3983 * its cpu_power, while calculating max_load..)
3985 if (sds->max_load < sds->avg_load) {
3987 return fix_small_imbalance(sds, this_cpu, imbalance);
3990 if (!sds->group_imb) {
3992 * Don't want to pull so many tasks that a group would go idle.
3994 load_above_capacity = (sds->busiest_nr_running -
3995 sds->busiest_group_capacity);
3997 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
3999 load_above_capacity /= sds->busiest->cpu_power;
4003 * We're trying to get all the cpus to the average_load, so we don't
4004 * want to push ourselves above the average load, nor do we wish to
4005 * reduce the max loaded cpu below the average load. At the same time,
4006 * we also don't want to reduce the group load below the group capacity
4007 * (so that we can implement power-savings policies etc). Thus we look
4008 * for the minimum possible imbalance.
4009 * Be careful of negative numbers as they'll appear as very large values
4010 * with unsigned longs.
4012 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4014 /* How much load to actually move to equalise the imbalance */
4015 *imbalance = min(max_pull * sds->busiest->cpu_power,
4016 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
4020 * if *imbalance is less than the average load per runnable task
4021 * there is no gaurantee that any tasks will be moved so we'll have
4022 * a think about bumping its value to force at least one task to be
4025 if (*imbalance < sds->busiest_load_per_task)
4026 return fix_small_imbalance(sds, this_cpu, imbalance);
4029 /******* find_busiest_group() helpers end here *********************/
4032 * find_busiest_group - Returns the busiest group within the sched_domain
4033 * if there is an imbalance. If there isn't an imbalance, and
4034 * the user has opted for power-savings, it returns a group whose
4035 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4036 * such a group exists.
4038 * Also calculates the amount of weighted load which should be moved
4039 * to restore balance.
4041 * @sd: The sched_domain whose busiest group is to be returned.
4042 * @this_cpu: The cpu for which load balancing is currently being performed.
4043 * @imbalance: Variable which stores amount of weighted load which should
4044 * be moved to restore balance/put a group to idle.
4045 * @idle: The idle status of this_cpu.
4046 * @sd_idle: The idleness of sd
4047 * @cpus: The set of CPUs under consideration for load-balancing.
4048 * @balance: Pointer to a variable indicating if this_cpu
4049 * is the appropriate cpu to perform load balancing at this_level.
4051 * Returns: - the busiest group if imbalance exists.
4052 * - If no imbalance and user has opted for power-savings balance,
4053 * return the least loaded group whose CPUs can be
4054 * put to idle by rebalancing its tasks onto our group.
4056 static struct sched_group *
4057 find_busiest_group(struct sched_domain *sd, int this_cpu,
4058 unsigned long *imbalance, enum cpu_idle_type idle,
4059 int *sd_idle, const struct cpumask *cpus, int *balance)
4061 struct sd_lb_stats sds;
4063 memset(&sds, 0, sizeof(sds));
4066 * Compute the various statistics relavent for load balancing at
4069 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4072 /* Cases where imbalance does not exist from POV of this_cpu */
4073 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4075 * 2) There is no busy sibling group to pull from.
4076 * 3) This group is the busiest group.
4077 * 4) This group is more busy than the avg busieness at this
4079 * 5) The imbalance is within the specified limit.
4081 if (balance && !(*balance))
4084 if (!sds.busiest || sds.busiest_nr_running == 0)
4087 if (sds.this_load >= sds.max_load)
4090 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4092 if (sds.this_load >= sds.avg_load)
4095 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4098 /* Looks like there is an imbalance. Compute it */
4099 calculate_imbalance(&sds, this_cpu, imbalance);
4104 * There is no obvious imbalance. But check if we can do some balancing
4107 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4115 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4118 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4119 unsigned long imbalance, const struct cpumask *cpus)
4121 struct rq *busiest = NULL, *rq;
4122 unsigned long max_load = 0;
4125 for_each_cpu(i, sched_group_cpus(group)) {
4126 unsigned long power = power_of(i);
4127 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4130 if (!cpumask_test_cpu(i, cpus))
4134 wl = weighted_cpuload(i);
4137 * When comparing with imbalance, use weighted_cpuload()
4138 * which is not scaled with the cpu power.
4140 if (capacity && rq->nr_running == 1 && wl > imbalance)
4144 * For the load comparisons with the other cpu's, consider
4145 * the weighted_cpuload() scaled with the cpu power, so that
4146 * the load can be moved away from the cpu that is potentially
4147 * running at a lower capacity.
4149 wl = (wl * SCHED_LOAD_SCALE) / power;
4151 if (wl > max_load) {
4161 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4162 * so long as it is large enough.
4164 #define MAX_PINNED_INTERVAL 512
4166 /* Working cpumask for load_balance and load_balance_newidle. */
4167 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4170 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4171 * tasks if there is an imbalance.
4173 static int load_balance(int this_cpu, struct rq *this_rq,
4174 struct sched_domain *sd, enum cpu_idle_type idle,
4177 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4178 struct sched_group *group;
4179 unsigned long imbalance;
4181 unsigned long flags;
4182 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4184 cpumask_copy(cpus, cpu_active_mask);
4187 * When power savings policy is enabled for the parent domain, idle
4188 * sibling can pick up load irrespective of busy siblings. In this case,
4189 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4190 * portraying it as CPU_NOT_IDLE.
4192 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4193 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4196 schedstat_inc(sd, lb_count[idle]);
4200 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4207 schedstat_inc(sd, lb_nobusyg[idle]);
4211 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4213 schedstat_inc(sd, lb_nobusyq[idle]);
4217 BUG_ON(busiest == this_rq);
4219 schedstat_add(sd, lb_imbalance[idle], imbalance);
4222 if (busiest->nr_running > 1) {
4224 * Attempt to move tasks. If find_busiest_group has found
4225 * an imbalance but busiest->nr_running <= 1, the group is
4226 * still unbalanced. ld_moved simply stays zero, so it is
4227 * correctly treated as an imbalance.
4229 local_irq_save(flags);
4230 double_rq_lock(this_rq, busiest);
4231 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4232 imbalance, sd, idle, &all_pinned);
4233 double_rq_unlock(this_rq, busiest);
4234 local_irq_restore(flags);
4237 * some other cpu did the load balance for us.
4239 if (ld_moved && this_cpu != smp_processor_id())
4240 resched_cpu(this_cpu);
4242 /* All tasks on this runqueue were pinned by CPU affinity */
4243 if (unlikely(all_pinned)) {
4244 cpumask_clear_cpu(cpu_of(busiest), cpus);
4245 if (!cpumask_empty(cpus))
4252 schedstat_inc(sd, lb_failed[idle]);
4254 * Increment the failure counter only on periodic balance.
4255 * We do not want newidle balance, which can be very
4256 * frequent, pollute the failure counter causing
4257 * excessive cache_hot migrations and active balances.
4259 if (idle != CPU_NEWLY_IDLE)
4260 sd->nr_balance_failed++;
4262 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4264 spin_lock_irqsave(&busiest->lock, flags);
4266 /* don't kick the migration_thread, if the curr
4267 * task on busiest cpu can't be moved to this_cpu
4269 if (!cpumask_test_cpu(this_cpu,
4270 &busiest->curr->cpus_allowed)) {
4271 spin_unlock_irqrestore(&busiest->lock, flags);
4273 goto out_one_pinned;
4276 if (!busiest->active_balance) {
4277 busiest->active_balance = 1;
4278 busiest->push_cpu = this_cpu;
4281 spin_unlock_irqrestore(&busiest->lock, flags);
4283 wake_up_process(busiest->migration_thread);
4286 * We've kicked active balancing, reset the failure
4289 sd->nr_balance_failed = sd->cache_nice_tries+1;
4292 sd->nr_balance_failed = 0;
4294 if (likely(!active_balance)) {
4295 /* We were unbalanced, so reset the balancing interval */
4296 sd->balance_interval = sd->min_interval;
4299 * If we've begun active balancing, start to back off. This
4300 * case may not be covered by the all_pinned logic if there
4301 * is only 1 task on the busy runqueue (because we don't call
4304 if (sd->balance_interval < sd->max_interval)
4305 sd->balance_interval *= 2;
4308 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4309 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4315 schedstat_inc(sd, lb_balanced[idle]);
4317 sd->nr_balance_failed = 0;
4320 /* tune up the balancing interval */
4321 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4322 (sd->balance_interval < sd->max_interval))
4323 sd->balance_interval *= 2;
4325 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4326 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4337 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4338 * tasks if there is an imbalance.
4340 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4341 * this_rq is locked.
4344 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4346 struct sched_group *group;
4347 struct rq *busiest = NULL;
4348 unsigned long imbalance;
4352 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4354 cpumask_copy(cpus, cpu_active_mask);
4357 * When power savings policy is enabled for the parent domain, idle
4358 * sibling can pick up load irrespective of busy siblings. In this case,
4359 * let the state of idle sibling percolate up as IDLE, instead of
4360 * portraying it as CPU_NOT_IDLE.
4362 if (sd->flags & SD_SHARE_CPUPOWER &&
4363 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4366 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4368 update_shares_locked(this_rq, sd);
4369 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4370 &sd_idle, cpus, NULL);
4372 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4376 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4378 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4382 BUG_ON(busiest == this_rq);
4384 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4387 if (busiest->nr_running > 1) {
4388 /* Attempt to move tasks */
4389 double_lock_balance(this_rq, busiest);
4390 /* this_rq->clock is already updated */
4391 update_rq_clock(busiest);
4392 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4393 imbalance, sd, CPU_NEWLY_IDLE,
4395 double_unlock_balance(this_rq, busiest);
4397 if (unlikely(all_pinned)) {
4398 cpumask_clear_cpu(cpu_of(busiest), cpus);
4399 if (!cpumask_empty(cpus))
4405 int active_balance = 0;
4407 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4408 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4409 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4412 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4415 if (sd->nr_balance_failed++ < 2)
4419 * The only task running in a non-idle cpu can be moved to this
4420 * cpu in an attempt to completely freeup the other CPU
4421 * package. The same method used to move task in load_balance()
4422 * have been extended for load_balance_newidle() to speedup
4423 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4425 * The package power saving logic comes from
4426 * find_busiest_group(). If there are no imbalance, then
4427 * f_b_g() will return NULL. However when sched_mc={1,2} then
4428 * f_b_g() will select a group from which a running task may be
4429 * pulled to this cpu in order to make the other package idle.
4430 * If there is no opportunity to make a package idle and if
4431 * there are no imbalance, then f_b_g() will return NULL and no
4432 * action will be taken in load_balance_newidle().
4434 * Under normal task pull operation due to imbalance, there
4435 * will be more than one task in the source run queue and
4436 * move_tasks() will succeed. ld_moved will be true and this
4437 * active balance code will not be triggered.
4440 /* Lock busiest in correct order while this_rq is held */
4441 double_lock_balance(this_rq, busiest);
4444 * don't kick the migration_thread, if the curr
4445 * task on busiest cpu can't be moved to this_cpu
4447 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4448 double_unlock_balance(this_rq, busiest);
4453 if (!busiest->active_balance) {
4454 busiest->active_balance = 1;
4455 busiest->push_cpu = this_cpu;
4459 double_unlock_balance(this_rq, busiest);
4461 * Should not call ttwu while holding a rq->lock
4463 spin_unlock(&this_rq->lock);
4465 wake_up_process(busiest->migration_thread);
4466 spin_lock(&this_rq->lock);
4469 sd->nr_balance_failed = 0;
4471 update_shares_locked(this_rq, sd);
4475 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4476 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4477 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4479 sd->nr_balance_failed = 0;
4485 * idle_balance is called by schedule() if this_cpu is about to become
4486 * idle. Attempts to pull tasks from other CPUs.
4488 static void idle_balance(int this_cpu, struct rq *this_rq)
4490 struct sched_domain *sd;
4491 int pulled_task = 0;
4492 unsigned long next_balance = jiffies + HZ;
4494 this_rq->idle_stamp = this_rq->clock;
4496 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4499 for_each_domain(this_cpu, sd) {
4500 unsigned long interval;
4502 if (!(sd->flags & SD_LOAD_BALANCE))
4505 if (sd->flags & SD_BALANCE_NEWIDLE)
4506 /* If we've pulled tasks over stop searching: */
4507 pulled_task = load_balance_newidle(this_cpu, this_rq,
4510 interval = msecs_to_jiffies(sd->balance_interval);
4511 if (time_after(next_balance, sd->last_balance + interval))
4512 next_balance = sd->last_balance + interval;
4514 this_rq->idle_stamp = 0;
4518 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4520 * We are going idle. next_balance may be set based on
4521 * a busy processor. So reset next_balance.
4523 this_rq->next_balance = next_balance;
4528 * active_load_balance is run by migration threads. It pushes running tasks
4529 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4530 * running on each physical CPU where possible, and avoids physical /
4531 * logical imbalances.
4533 * Called with busiest_rq locked.
4535 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4537 int target_cpu = busiest_rq->push_cpu;
4538 struct sched_domain *sd;
4539 struct rq *target_rq;
4541 /* Is there any task to move? */
4542 if (busiest_rq->nr_running <= 1)
4545 target_rq = cpu_rq(target_cpu);
4548 * This condition is "impossible", if it occurs
4549 * we need to fix it. Originally reported by
4550 * Bjorn Helgaas on a 128-cpu setup.
4552 BUG_ON(busiest_rq == target_rq);
4554 /* move a task from busiest_rq to target_rq */
4555 double_lock_balance(busiest_rq, target_rq);
4556 update_rq_clock(busiest_rq);
4557 update_rq_clock(target_rq);
4559 /* Search for an sd spanning us and the target CPU. */
4560 for_each_domain(target_cpu, sd) {
4561 if ((sd->flags & SD_LOAD_BALANCE) &&
4562 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4567 schedstat_inc(sd, alb_count);
4569 if (move_one_task(target_rq, target_cpu, busiest_rq,
4571 schedstat_inc(sd, alb_pushed);
4573 schedstat_inc(sd, alb_failed);
4575 double_unlock_balance(busiest_rq, target_rq);
4580 atomic_t load_balancer;
4581 cpumask_var_t cpu_mask;
4582 cpumask_var_t ilb_grp_nohz_mask;
4583 } nohz ____cacheline_aligned = {
4584 .load_balancer = ATOMIC_INIT(-1),
4587 int get_nohz_load_balancer(void)
4589 return atomic_read(&nohz.load_balancer);
4592 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4594 * lowest_flag_domain - Return lowest sched_domain containing flag.
4595 * @cpu: The cpu whose lowest level of sched domain is to
4597 * @flag: The flag to check for the lowest sched_domain
4598 * for the given cpu.
4600 * Returns the lowest sched_domain of a cpu which contains the given flag.
4602 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4604 struct sched_domain *sd;
4606 for_each_domain(cpu, sd)
4607 if (sd && (sd->flags & flag))
4614 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4615 * @cpu: The cpu whose domains we're iterating over.
4616 * @sd: variable holding the value of the power_savings_sd
4618 * @flag: The flag to filter the sched_domains to be iterated.
4620 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4621 * set, starting from the lowest sched_domain to the highest.
4623 #define for_each_flag_domain(cpu, sd, flag) \
4624 for (sd = lowest_flag_domain(cpu, flag); \
4625 (sd && (sd->flags & flag)); sd = sd->parent)
4628 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4629 * @ilb_group: group to be checked for semi-idleness
4631 * Returns: 1 if the group is semi-idle. 0 otherwise.
4633 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4634 * and atleast one non-idle CPU. This helper function checks if the given
4635 * sched_group is semi-idle or not.
4637 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4639 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4640 sched_group_cpus(ilb_group));
4643 * A sched_group is semi-idle when it has atleast one busy cpu
4644 * and atleast one idle cpu.
4646 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4649 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4655 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4656 * @cpu: The cpu which is nominating a new idle_load_balancer.
4658 * Returns: Returns the id of the idle load balancer if it exists,
4659 * Else, returns >= nr_cpu_ids.
4661 * This algorithm picks the idle load balancer such that it belongs to a
4662 * semi-idle powersavings sched_domain. The idea is to try and avoid
4663 * completely idle packages/cores just for the purpose of idle load balancing
4664 * when there are other idle cpu's which are better suited for that job.
4666 static int find_new_ilb(int cpu)
4668 struct sched_domain *sd;
4669 struct sched_group *ilb_group;
4672 * Have idle load balancer selection from semi-idle packages only
4673 * when power-aware load balancing is enabled
4675 if (!(sched_smt_power_savings || sched_mc_power_savings))
4679 * Optimize for the case when we have no idle CPUs or only one
4680 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4682 if (cpumask_weight(nohz.cpu_mask) < 2)
4685 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4686 ilb_group = sd->groups;
4689 if (is_semi_idle_group(ilb_group))
4690 return cpumask_first(nohz.ilb_grp_nohz_mask);
4692 ilb_group = ilb_group->next;
4694 } while (ilb_group != sd->groups);
4698 return cpumask_first(nohz.cpu_mask);
4700 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4701 static inline int find_new_ilb(int call_cpu)
4703 return cpumask_first(nohz.cpu_mask);
4708 * This routine will try to nominate the ilb (idle load balancing)
4709 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4710 * load balancing on behalf of all those cpus. If all the cpus in the system
4711 * go into this tickless mode, then there will be no ilb owner (as there is
4712 * no need for one) and all the cpus will sleep till the next wakeup event
4715 * For the ilb owner, tick is not stopped. And this tick will be used
4716 * for idle load balancing. ilb owner will still be part of
4719 * While stopping the tick, this cpu will become the ilb owner if there
4720 * is no other owner. And will be the owner till that cpu becomes busy
4721 * or if all cpus in the system stop their ticks at which point
4722 * there is no need for ilb owner.
4724 * When the ilb owner becomes busy, it nominates another owner, during the
4725 * next busy scheduler_tick()
4727 int select_nohz_load_balancer(int stop_tick)
4729 int cpu = smp_processor_id();
4732 cpu_rq(cpu)->in_nohz_recently = 1;
4734 if (!cpu_active(cpu)) {
4735 if (atomic_read(&nohz.load_balancer) != cpu)
4739 * If we are going offline and still the leader,
4742 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4748 cpumask_set_cpu(cpu, nohz.cpu_mask);
4750 /* time for ilb owner also to sleep */
4751 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4752 if (atomic_read(&nohz.load_balancer) == cpu)
4753 atomic_set(&nohz.load_balancer, -1);
4757 if (atomic_read(&nohz.load_balancer) == -1) {
4758 /* make me the ilb owner */
4759 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4761 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4764 if (!(sched_smt_power_savings ||
4765 sched_mc_power_savings))
4768 * Check to see if there is a more power-efficient
4771 new_ilb = find_new_ilb(cpu);
4772 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4773 atomic_set(&nohz.load_balancer, -1);
4774 resched_cpu(new_ilb);
4780 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4783 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4785 if (atomic_read(&nohz.load_balancer) == cpu)
4786 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4793 static DEFINE_SPINLOCK(balancing);
4796 * It checks each scheduling domain to see if it is due to be balanced,
4797 * and initiates a balancing operation if so.
4799 * Balancing parameters are set up in arch_init_sched_domains.
4801 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4804 struct rq *rq = cpu_rq(cpu);
4805 unsigned long interval;
4806 struct sched_domain *sd;
4807 /* Earliest time when we have to do rebalance again */
4808 unsigned long next_balance = jiffies + 60*HZ;
4809 int update_next_balance = 0;
4812 for_each_domain(cpu, sd) {
4813 if (!(sd->flags & SD_LOAD_BALANCE))
4816 interval = sd->balance_interval;
4817 if (idle != CPU_IDLE)
4818 interval *= sd->busy_factor;
4820 /* scale ms to jiffies */
4821 interval = msecs_to_jiffies(interval);
4822 if (unlikely(!interval))
4824 if (interval > HZ*NR_CPUS/10)
4825 interval = HZ*NR_CPUS/10;
4827 need_serialize = sd->flags & SD_SERIALIZE;
4829 if (need_serialize) {
4830 if (!spin_trylock(&balancing))
4834 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4835 if (load_balance(cpu, rq, sd, idle, &balance)) {
4837 * We've pulled tasks over so either we're no
4838 * longer idle, or one of our SMT siblings is
4841 idle = CPU_NOT_IDLE;
4843 sd->last_balance = jiffies;
4846 spin_unlock(&balancing);
4848 if (time_after(next_balance, sd->last_balance + interval)) {
4849 next_balance = sd->last_balance + interval;
4850 update_next_balance = 1;
4854 * Stop the load balance at this level. There is another
4855 * CPU in our sched group which is doing load balancing more
4863 * next_balance will be updated only when there is a need.
4864 * When the cpu is attached to null domain for ex, it will not be
4867 if (likely(update_next_balance))
4868 rq->next_balance = next_balance;
4872 * run_rebalance_domains is triggered when needed from the scheduler tick.
4873 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4874 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4876 static void run_rebalance_domains(struct softirq_action *h)
4878 int this_cpu = smp_processor_id();
4879 struct rq *this_rq = cpu_rq(this_cpu);
4880 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4881 CPU_IDLE : CPU_NOT_IDLE;
4883 rebalance_domains(this_cpu, idle);
4887 * If this cpu is the owner for idle load balancing, then do the
4888 * balancing on behalf of the other idle cpus whose ticks are
4891 if (this_rq->idle_at_tick &&
4892 atomic_read(&nohz.load_balancer) == this_cpu) {
4896 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4897 if (balance_cpu == this_cpu)
4901 * If this cpu gets work to do, stop the load balancing
4902 * work being done for other cpus. Next load
4903 * balancing owner will pick it up.
4908 rebalance_domains(balance_cpu, CPU_IDLE);
4910 rq = cpu_rq(balance_cpu);
4911 if (time_after(this_rq->next_balance, rq->next_balance))
4912 this_rq->next_balance = rq->next_balance;
4918 static inline int on_null_domain(int cpu)
4920 return !rcu_dereference(cpu_rq(cpu)->sd);
4924 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4926 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4927 * idle load balancing owner or decide to stop the periodic load balancing,
4928 * if the whole system is idle.
4930 static inline void trigger_load_balance(struct rq *rq, int cpu)
4934 * If we were in the nohz mode recently and busy at the current
4935 * scheduler tick, then check if we need to nominate new idle
4938 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4939 rq->in_nohz_recently = 0;
4941 if (atomic_read(&nohz.load_balancer) == cpu) {
4942 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4943 atomic_set(&nohz.load_balancer, -1);
4946 if (atomic_read(&nohz.load_balancer) == -1) {
4947 int ilb = find_new_ilb(cpu);
4949 if (ilb < nr_cpu_ids)
4955 * If this cpu is idle and doing idle load balancing for all the
4956 * cpus with ticks stopped, is it time for that to stop?
4958 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4959 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4965 * If this cpu is idle and the idle load balancing is done by
4966 * someone else, then no need raise the SCHED_SOFTIRQ
4968 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4969 cpumask_test_cpu(cpu, nohz.cpu_mask))
4972 /* Don't need to rebalance while attached to NULL domain */
4973 if (time_after_eq(jiffies, rq->next_balance) &&
4974 likely(!on_null_domain(cpu)))
4975 raise_softirq(SCHED_SOFTIRQ);
4978 #else /* CONFIG_SMP */
4981 * on UP we do not need to balance between CPUs:
4983 static inline void idle_balance(int cpu, struct rq *rq)
4989 DEFINE_PER_CPU(struct kernel_stat, kstat);
4991 EXPORT_PER_CPU_SYMBOL(kstat);
4994 * Return any ns on the sched_clock that have not yet been accounted in
4995 * @p in case that task is currently running.
4997 * Called with task_rq_lock() held on @rq.
4999 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
5003 if (task_current(rq, p)) {
5004 update_rq_clock(rq);
5005 ns = rq->clock - p->se.exec_start;
5013 unsigned long long task_delta_exec(struct task_struct *p)
5015 unsigned long flags;
5019 rq = task_rq_lock(p, &flags);
5020 ns = do_task_delta_exec(p, rq);
5021 task_rq_unlock(rq, &flags);
5027 * Return accounted runtime for the task.
5028 * In case the task is currently running, return the runtime plus current's
5029 * pending runtime that have not been accounted yet.
5031 unsigned long long task_sched_runtime(struct task_struct *p)
5033 unsigned long flags;
5037 rq = task_rq_lock(p, &flags);
5038 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5039 task_rq_unlock(rq, &flags);
5045 * Return sum_exec_runtime for the thread group.
5046 * In case the task is currently running, return the sum plus current's
5047 * pending runtime that have not been accounted yet.
5049 * Note that the thread group might have other running tasks as well,
5050 * so the return value not includes other pending runtime that other
5051 * running tasks might have.
5053 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5055 struct task_cputime totals;
5056 unsigned long flags;
5060 rq = task_rq_lock(p, &flags);
5061 thread_group_cputime(p, &totals);
5062 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5063 task_rq_unlock(rq, &flags);
5069 * Account user cpu time to a process.
5070 * @p: the process that the cpu time gets accounted to
5071 * @cputime: the cpu time spent in user space since the last update
5072 * @cputime_scaled: cputime scaled by cpu frequency
5074 void account_user_time(struct task_struct *p, cputime_t cputime,
5075 cputime_t cputime_scaled)
5077 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5080 /* Add user time to process. */
5081 p->utime = cputime_add(p->utime, cputime);
5082 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5083 account_group_user_time(p, cputime);
5085 /* Add user time to cpustat. */
5086 tmp = cputime_to_cputime64(cputime);
5087 if (TASK_NICE(p) > 0)
5088 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5090 cpustat->user = cputime64_add(cpustat->user, tmp);
5092 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5093 /* Account for user time used */
5094 acct_update_integrals(p);
5098 * Account guest cpu time to a process.
5099 * @p: the process that the cpu time gets accounted to
5100 * @cputime: the cpu time spent in virtual machine since the last update
5101 * @cputime_scaled: cputime scaled by cpu frequency
5103 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5104 cputime_t cputime_scaled)
5107 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5109 tmp = cputime_to_cputime64(cputime);
5111 /* Add guest time to process. */
5112 p->utime = cputime_add(p->utime, cputime);
5113 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5114 account_group_user_time(p, cputime);
5115 p->gtime = cputime_add(p->gtime, cputime);
5117 /* Add guest time to cpustat. */
5118 cpustat->user = cputime64_add(cpustat->user, tmp);
5119 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5123 * Account system cpu time to a process.
5124 * @p: the process that the cpu time gets accounted to
5125 * @hardirq_offset: the offset to subtract from hardirq_count()
5126 * @cputime: the cpu time spent in kernel space since the last update
5127 * @cputime_scaled: cputime scaled by cpu frequency
5129 void account_system_time(struct task_struct *p, int hardirq_offset,
5130 cputime_t cputime, cputime_t cputime_scaled)
5132 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5135 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5136 account_guest_time(p, cputime, cputime_scaled);
5140 /* Add system time to process. */
5141 p->stime = cputime_add(p->stime, cputime);
5142 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5143 account_group_system_time(p, cputime);
5145 /* Add system time to cpustat. */
5146 tmp = cputime_to_cputime64(cputime);
5147 if (hardirq_count() - hardirq_offset)
5148 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5149 else if (softirq_count())
5150 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5152 cpustat->system = cputime64_add(cpustat->system, tmp);
5154 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5156 /* Account for system time used */
5157 acct_update_integrals(p);
5161 * Account for involuntary wait time.
5162 * @steal: the cpu time spent in involuntary wait
5164 void account_steal_time(cputime_t cputime)
5166 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5167 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5169 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5173 * Account for idle time.
5174 * @cputime: the cpu time spent in idle wait
5176 void account_idle_time(cputime_t cputime)
5178 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5179 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5180 struct rq *rq = this_rq();
5182 if (atomic_read(&rq->nr_iowait) > 0)
5183 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5185 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5188 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5191 * Account a single tick of cpu time.
5192 * @p: the process that the cpu time gets accounted to
5193 * @user_tick: indicates if the tick is a user or a system tick
5195 void account_process_tick(struct task_struct *p, int user_tick)
5197 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5198 struct rq *rq = this_rq();
5201 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5202 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5203 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5206 account_idle_time(cputime_one_jiffy);
5210 * Account multiple ticks of steal time.
5211 * @p: the process from which the cpu time has been stolen
5212 * @ticks: number of stolen ticks
5214 void account_steal_ticks(unsigned long ticks)
5216 account_steal_time(jiffies_to_cputime(ticks));
5220 * Account multiple ticks of idle time.
5221 * @ticks: number of stolen ticks
5223 void account_idle_ticks(unsigned long ticks)
5225 account_idle_time(jiffies_to_cputime(ticks));
5231 * Use precise platform statistics if available:
5233 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5234 cputime_t task_utime(struct task_struct *p)
5239 cputime_t task_stime(struct task_struct *p)
5244 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5246 struct task_cputime cputime;
5248 thread_group_cputime(p, &cputime);
5250 *ut = cputime.utime;
5251 *st = cputime.stime;
5255 #ifndef nsecs_to_cputime
5256 # define nsecs_to_cputime(__nsecs) \
5257 msecs_to_cputime(div_u64((__nsecs), NSEC_PER_MSEC))
5260 cputime_t task_utime(struct task_struct *p)
5262 cputime_t utime = p->utime, total = utime + p->stime;
5266 * Use CFS's precise accounting:
5268 temp = (u64)nsecs_to_cputime(p->se.sum_exec_runtime);
5272 do_div(temp, total);
5274 utime = (cputime_t)temp;
5276 p->prev_utime = max(p->prev_utime, utime);
5277 return p->prev_utime;
5280 cputime_t task_stime(struct task_struct *p)
5285 * Use CFS's precise accounting. (we subtract utime from
5286 * the total, to make sure the total observed by userspace
5287 * grows monotonically - apps rely on that):
5289 stime = nsecs_to_cputime(p->se.sum_exec_runtime) - task_utime(p);
5292 p->prev_stime = max(p->prev_stime, stime);
5294 return p->prev_stime;
5298 * Must be called with siglock held.
5300 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5302 struct signal_struct *sig = p->signal;
5303 struct task_cputime cputime;
5304 cputime_t rtime, utime, total;
5306 thread_group_cputime(p, &cputime);
5308 total = cputime_add(cputime.utime, cputime.stime);
5309 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5314 temp *= cputime.utime;
5315 do_div(temp, total);
5316 utime = (cputime_t)temp;
5320 sig->prev_utime = max(sig->prev_utime, utime);
5321 sig->prev_stime = max(sig->prev_stime,
5322 cputime_sub(rtime, sig->prev_utime));
5324 *ut = sig->prev_utime;
5325 *st = sig->prev_stime;
5329 inline cputime_t task_gtime(struct task_struct *p)
5335 * This function gets called by the timer code, with HZ frequency.
5336 * We call it with interrupts disabled.
5338 * It also gets called by the fork code, when changing the parent's
5341 void scheduler_tick(void)
5343 int cpu = smp_processor_id();
5344 struct rq *rq = cpu_rq(cpu);
5345 struct task_struct *curr = rq->curr;
5349 spin_lock(&rq->lock);
5350 update_rq_clock(rq);
5351 update_cpu_load(rq);
5352 curr->sched_class->task_tick(rq, curr, 0);
5353 spin_unlock(&rq->lock);
5355 perf_event_task_tick(curr, cpu);
5358 rq->idle_at_tick = idle_cpu(cpu);
5359 trigger_load_balance(rq, cpu);
5363 notrace unsigned long get_parent_ip(unsigned long addr)
5365 if (in_lock_functions(addr)) {
5366 addr = CALLER_ADDR2;
5367 if (in_lock_functions(addr))
5368 addr = CALLER_ADDR3;
5373 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5374 defined(CONFIG_PREEMPT_TRACER))
5376 void __kprobes add_preempt_count(int val)
5378 #ifdef CONFIG_DEBUG_PREEMPT
5382 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5385 preempt_count() += val;
5386 #ifdef CONFIG_DEBUG_PREEMPT
5388 * Spinlock count overflowing soon?
5390 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5393 if (preempt_count() == val)
5394 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5396 EXPORT_SYMBOL(add_preempt_count);
5398 void __kprobes sub_preempt_count(int val)
5400 #ifdef CONFIG_DEBUG_PREEMPT
5404 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5407 * Is the spinlock portion underflowing?
5409 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5410 !(preempt_count() & PREEMPT_MASK)))
5414 if (preempt_count() == val)
5415 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5416 preempt_count() -= val;
5418 EXPORT_SYMBOL(sub_preempt_count);
5423 * Print scheduling while atomic bug:
5425 static noinline void __schedule_bug(struct task_struct *prev)
5427 struct pt_regs *regs = get_irq_regs();
5429 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5430 prev->comm, prev->pid, preempt_count());
5432 debug_show_held_locks(prev);
5434 if (irqs_disabled())
5435 print_irqtrace_events(prev);
5444 * Various schedule()-time debugging checks and statistics:
5446 static inline void schedule_debug(struct task_struct *prev)
5449 * Test if we are atomic. Since do_exit() needs to call into
5450 * schedule() atomically, we ignore that path for now.
5451 * Otherwise, whine if we are scheduling when we should not be.
5453 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5454 __schedule_bug(prev);
5456 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5458 schedstat_inc(this_rq(), sched_count);
5459 #ifdef CONFIG_SCHEDSTATS
5460 if (unlikely(prev->lock_depth >= 0)) {
5461 schedstat_inc(this_rq(), bkl_count);
5462 schedstat_inc(prev, sched_info.bkl_count);
5467 static void put_prev_task(struct rq *rq, struct task_struct *p)
5469 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
5471 update_avg(&p->se.avg_running, runtime);
5473 if (p->state == TASK_RUNNING) {
5475 * In order to avoid avg_overlap growing stale when we are
5476 * indeed overlapping and hence not getting put to sleep, grow
5477 * the avg_overlap on preemption.
5479 * We use the average preemption runtime because that
5480 * correlates to the amount of cache footprint a task can
5483 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5484 update_avg(&p->se.avg_overlap, runtime);
5486 update_avg(&p->se.avg_running, 0);
5488 p->sched_class->put_prev_task(rq, p);
5492 * Pick up the highest-prio task:
5494 static inline struct task_struct *
5495 pick_next_task(struct rq *rq)
5497 const struct sched_class *class;
5498 struct task_struct *p;
5501 * Optimization: we know that if all tasks are in
5502 * the fair class we can call that function directly:
5504 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5505 p = fair_sched_class.pick_next_task(rq);
5510 class = sched_class_highest;
5512 p = class->pick_next_task(rq);
5516 * Will never be NULL as the idle class always
5517 * returns a non-NULL p:
5519 class = class->next;
5524 * schedule() is the main scheduler function.
5526 asmlinkage void __sched schedule(void)
5528 struct task_struct *prev, *next;
5529 unsigned long *switch_count;
5535 cpu = smp_processor_id();
5539 switch_count = &prev->nivcsw;
5541 release_kernel_lock(prev);
5542 need_resched_nonpreemptible:
5544 schedule_debug(prev);
5546 if (sched_feat(HRTICK))
5549 spin_lock_irq(&rq->lock);
5550 update_rq_clock(rq);
5551 clear_tsk_need_resched(prev);
5553 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5554 if (unlikely(signal_pending_state(prev->state, prev)))
5555 prev->state = TASK_RUNNING;
5557 deactivate_task(rq, prev, 1);
5558 switch_count = &prev->nvcsw;
5561 pre_schedule(rq, prev);
5563 if (unlikely(!rq->nr_running))
5564 idle_balance(cpu, rq);
5566 put_prev_task(rq, prev);
5567 next = pick_next_task(rq);
5569 if (likely(prev != next)) {
5570 sched_info_switch(prev, next);
5571 perf_event_task_sched_out(prev, next, cpu);
5577 context_switch(rq, prev, next); /* unlocks the rq */
5579 * the context switch might have flipped the stack from under
5580 * us, hence refresh the local variables.
5582 cpu = smp_processor_id();
5585 spin_unlock_irq(&rq->lock);
5589 if (unlikely(reacquire_kernel_lock(current) < 0))
5590 goto need_resched_nonpreemptible;
5592 preempt_enable_no_resched();
5596 EXPORT_SYMBOL(schedule);
5600 * Look out! "owner" is an entirely speculative pointer
5601 * access and not reliable.
5603 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5608 if (!sched_feat(OWNER_SPIN))
5611 #ifdef CONFIG_DEBUG_PAGEALLOC
5613 * Need to access the cpu field knowing that
5614 * DEBUG_PAGEALLOC could have unmapped it if
5615 * the mutex owner just released it and exited.
5617 if (probe_kernel_address(&owner->cpu, cpu))
5624 * Even if the access succeeded (likely case),
5625 * the cpu field may no longer be valid.
5627 if (cpu >= nr_cpumask_bits)
5631 * We need to validate that we can do a
5632 * get_cpu() and that we have the percpu area.
5634 if (!cpu_online(cpu))
5641 * Owner changed, break to re-assess state.
5643 if (lock->owner != owner)
5647 * Is that owner really running on that cpu?
5649 if (task_thread_info(rq->curr) != owner || need_resched())
5659 #ifdef CONFIG_PREEMPT
5661 * this is the entry point to schedule() from in-kernel preemption
5662 * off of preempt_enable. Kernel preemptions off return from interrupt
5663 * occur there and call schedule directly.
5665 asmlinkage void __sched preempt_schedule(void)
5667 struct thread_info *ti = current_thread_info();
5670 * If there is a non-zero preempt_count or interrupts are disabled,
5671 * we do not want to preempt the current task. Just return..
5673 if (likely(ti->preempt_count || irqs_disabled()))
5677 add_preempt_count(PREEMPT_ACTIVE);
5679 sub_preempt_count(PREEMPT_ACTIVE);
5682 * Check again in case we missed a preemption opportunity
5683 * between schedule and now.
5686 } while (need_resched());
5688 EXPORT_SYMBOL(preempt_schedule);
5691 * this is the entry point to schedule() from kernel preemption
5692 * off of irq context.
5693 * Note, that this is called and return with irqs disabled. This will
5694 * protect us against recursive calling from irq.
5696 asmlinkage void __sched preempt_schedule_irq(void)
5698 struct thread_info *ti = current_thread_info();
5700 /* Catch callers which need to be fixed */
5701 BUG_ON(ti->preempt_count || !irqs_disabled());
5704 add_preempt_count(PREEMPT_ACTIVE);
5707 local_irq_disable();
5708 sub_preempt_count(PREEMPT_ACTIVE);
5711 * Check again in case we missed a preemption opportunity
5712 * between schedule and now.
5715 } while (need_resched());
5718 #endif /* CONFIG_PREEMPT */
5720 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5723 return try_to_wake_up(curr->private, mode, wake_flags);
5725 EXPORT_SYMBOL(default_wake_function);
5728 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5729 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5730 * number) then we wake all the non-exclusive tasks and one exclusive task.
5732 * There are circumstances in which we can try to wake a task which has already
5733 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5734 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5736 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5737 int nr_exclusive, int wake_flags, void *key)
5739 wait_queue_t *curr, *next;
5741 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5742 unsigned flags = curr->flags;
5744 if (curr->func(curr, mode, wake_flags, key) &&
5745 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5751 * __wake_up - wake up threads blocked on a waitqueue.
5753 * @mode: which threads
5754 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5755 * @key: is directly passed to the wakeup function
5757 * It may be assumed that this function implies a write memory barrier before
5758 * changing the task state if and only if any tasks are woken up.
5760 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5761 int nr_exclusive, void *key)
5763 unsigned long flags;
5765 spin_lock_irqsave(&q->lock, flags);
5766 __wake_up_common(q, mode, nr_exclusive, 0, key);
5767 spin_unlock_irqrestore(&q->lock, flags);
5769 EXPORT_SYMBOL(__wake_up);
5772 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5774 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5776 __wake_up_common(q, mode, 1, 0, NULL);
5779 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5781 __wake_up_common(q, mode, 1, 0, key);
5785 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5787 * @mode: which threads
5788 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5789 * @key: opaque value to be passed to wakeup targets
5791 * The sync wakeup differs that the waker knows that it will schedule
5792 * away soon, so while the target thread will be woken up, it will not
5793 * be migrated to another CPU - ie. the two threads are 'synchronized'
5794 * with each other. This can prevent needless bouncing between CPUs.
5796 * On UP it can prevent extra preemption.
5798 * It may be assumed that this function implies a write memory barrier before
5799 * changing the task state if and only if any tasks are woken up.
5801 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5802 int nr_exclusive, void *key)
5804 unsigned long flags;
5805 int wake_flags = WF_SYNC;
5810 if (unlikely(!nr_exclusive))
5813 spin_lock_irqsave(&q->lock, flags);
5814 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5815 spin_unlock_irqrestore(&q->lock, flags);
5817 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5820 * __wake_up_sync - see __wake_up_sync_key()
5822 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5824 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5826 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5829 * complete: - signals a single thread waiting on this completion
5830 * @x: holds the state of this particular completion
5832 * This will wake up a single thread waiting on this completion. Threads will be
5833 * awakened in the same order in which they were queued.
5835 * See also complete_all(), wait_for_completion() and related routines.
5837 * It may be assumed that this function implies a write memory barrier before
5838 * changing the task state if and only if any tasks are woken up.
5840 void complete(struct completion *x)
5842 unsigned long flags;
5844 spin_lock_irqsave(&x->wait.lock, flags);
5846 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5847 spin_unlock_irqrestore(&x->wait.lock, flags);
5849 EXPORT_SYMBOL(complete);
5852 * complete_all: - signals all threads waiting on this completion
5853 * @x: holds the state of this particular completion
5855 * This will wake up all threads waiting on this particular completion event.
5857 * It may be assumed that this function implies a write memory barrier before
5858 * changing the task state if and only if any tasks are woken up.
5860 void complete_all(struct completion *x)
5862 unsigned long flags;
5864 spin_lock_irqsave(&x->wait.lock, flags);
5865 x->done += UINT_MAX/2;
5866 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5867 spin_unlock_irqrestore(&x->wait.lock, flags);
5869 EXPORT_SYMBOL(complete_all);
5871 static inline long __sched
5872 do_wait_for_common(struct completion *x, long timeout, int state)
5875 DECLARE_WAITQUEUE(wait, current);
5877 wait.flags |= WQ_FLAG_EXCLUSIVE;
5878 __add_wait_queue_tail(&x->wait, &wait);
5880 if (signal_pending_state(state, current)) {
5881 timeout = -ERESTARTSYS;
5884 __set_current_state(state);
5885 spin_unlock_irq(&x->wait.lock);
5886 timeout = schedule_timeout(timeout);
5887 spin_lock_irq(&x->wait.lock);
5888 } while (!x->done && timeout);
5889 __remove_wait_queue(&x->wait, &wait);
5894 return timeout ?: 1;
5898 wait_for_common(struct completion *x, long timeout, int state)
5902 spin_lock_irq(&x->wait.lock);
5903 timeout = do_wait_for_common(x, timeout, state);
5904 spin_unlock_irq(&x->wait.lock);
5909 * wait_for_completion: - waits for completion of a task
5910 * @x: holds the state of this particular completion
5912 * This waits to be signaled for completion of a specific task. It is NOT
5913 * interruptible and there is no timeout.
5915 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5916 * and interrupt capability. Also see complete().
5918 void __sched wait_for_completion(struct completion *x)
5920 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5922 EXPORT_SYMBOL(wait_for_completion);
5925 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5926 * @x: holds the state of this particular completion
5927 * @timeout: timeout value in jiffies
5929 * This waits for either a completion of a specific task to be signaled or for a
5930 * specified timeout to expire. The timeout is in jiffies. It is not
5933 unsigned long __sched
5934 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5936 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5938 EXPORT_SYMBOL(wait_for_completion_timeout);
5941 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5942 * @x: holds the state of this particular completion
5944 * This waits for completion of a specific task to be signaled. It is
5947 int __sched wait_for_completion_interruptible(struct completion *x)
5949 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5950 if (t == -ERESTARTSYS)
5954 EXPORT_SYMBOL(wait_for_completion_interruptible);
5957 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5958 * @x: holds the state of this particular completion
5959 * @timeout: timeout value in jiffies
5961 * This waits for either a completion of a specific task to be signaled or for a
5962 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5964 unsigned long __sched
5965 wait_for_completion_interruptible_timeout(struct completion *x,
5966 unsigned long timeout)
5968 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5970 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5973 * wait_for_completion_killable: - waits for completion of a task (killable)
5974 * @x: holds the state of this particular completion
5976 * This waits to be signaled for completion of a specific task. It can be
5977 * interrupted by a kill signal.
5979 int __sched wait_for_completion_killable(struct completion *x)
5981 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5982 if (t == -ERESTARTSYS)
5986 EXPORT_SYMBOL(wait_for_completion_killable);
5989 * try_wait_for_completion - try to decrement a completion without blocking
5990 * @x: completion structure
5992 * Returns: 0 if a decrement cannot be done without blocking
5993 * 1 if a decrement succeeded.
5995 * If a completion is being used as a counting completion,
5996 * attempt to decrement the counter without blocking. This
5997 * enables us to avoid waiting if the resource the completion
5998 * is protecting is not available.
6000 bool try_wait_for_completion(struct completion *x)
6002 unsigned long flags;
6005 spin_lock_irqsave(&x->wait.lock, flags);
6010 spin_unlock_irqrestore(&x->wait.lock, flags);
6013 EXPORT_SYMBOL(try_wait_for_completion);
6016 * completion_done - Test to see if a completion has any waiters
6017 * @x: completion structure
6019 * Returns: 0 if there are waiters (wait_for_completion() in progress)
6020 * 1 if there are no waiters.
6023 bool completion_done(struct completion *x)
6025 unsigned long flags;
6028 spin_lock_irqsave(&x->wait.lock, flags);
6031 spin_unlock_irqrestore(&x->wait.lock, flags);
6034 EXPORT_SYMBOL(completion_done);
6037 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
6039 unsigned long flags;
6042 init_waitqueue_entry(&wait, current);
6044 __set_current_state(state);
6046 spin_lock_irqsave(&q->lock, flags);
6047 __add_wait_queue(q, &wait);
6048 spin_unlock(&q->lock);
6049 timeout = schedule_timeout(timeout);
6050 spin_lock_irq(&q->lock);
6051 __remove_wait_queue(q, &wait);
6052 spin_unlock_irqrestore(&q->lock, flags);
6057 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6059 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6061 EXPORT_SYMBOL(interruptible_sleep_on);
6064 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6066 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6068 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6070 void __sched sleep_on(wait_queue_head_t *q)
6072 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6074 EXPORT_SYMBOL(sleep_on);
6076 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6078 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6080 EXPORT_SYMBOL(sleep_on_timeout);
6082 #ifdef CONFIG_RT_MUTEXES
6085 * rt_mutex_setprio - set the current priority of a task
6087 * @prio: prio value (kernel-internal form)
6089 * This function changes the 'effective' priority of a task. It does
6090 * not touch ->normal_prio like __setscheduler().
6092 * Used by the rt_mutex code to implement priority inheritance logic.
6094 void rt_mutex_setprio(struct task_struct *p, int prio)
6096 unsigned long flags;
6097 int oldprio, on_rq, running;
6099 const struct sched_class *prev_class;
6101 BUG_ON(prio < 0 || prio > MAX_PRIO);
6103 rq = task_rq_lock(p, &flags);
6104 update_rq_clock(rq);
6107 prev_class = p->sched_class;
6108 on_rq = p->se.on_rq;
6109 running = task_current(rq, p);
6111 dequeue_task(rq, p, 0);
6113 p->sched_class->put_prev_task(rq, p);
6116 p->sched_class = &rt_sched_class;
6118 p->sched_class = &fair_sched_class;
6123 p->sched_class->set_curr_task(rq);
6125 enqueue_task(rq, p, 0, oldprio < prio);
6127 check_class_changed(rq, p, prev_class, oldprio, running);
6129 task_rq_unlock(rq, &flags);
6134 void set_user_nice(struct task_struct *p, long nice)
6136 int old_prio, delta, on_rq;
6137 unsigned long flags;
6140 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6143 * We have to be careful, if called from sys_setpriority(),
6144 * the task might be in the middle of scheduling on another CPU.
6146 rq = task_rq_lock(p, &flags);
6147 update_rq_clock(rq);
6149 * The RT priorities are set via sched_setscheduler(), but we still
6150 * allow the 'normal' nice value to be set - but as expected
6151 * it wont have any effect on scheduling until the task is
6152 * SCHED_FIFO/SCHED_RR:
6154 if (task_has_rt_policy(p)) {
6155 p->static_prio = NICE_TO_PRIO(nice);
6158 on_rq = p->se.on_rq;
6160 dequeue_task(rq, p, 0);
6162 p->static_prio = NICE_TO_PRIO(nice);
6165 p->prio = effective_prio(p);
6166 delta = p->prio - old_prio;
6169 enqueue_task(rq, p, 0, false);
6171 * If the task increased its priority or is running and
6172 * lowered its priority, then reschedule its CPU:
6174 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6175 resched_task(rq->curr);
6178 task_rq_unlock(rq, &flags);
6180 EXPORT_SYMBOL(set_user_nice);
6183 * can_nice - check if a task can reduce its nice value
6187 int can_nice(const struct task_struct *p, const int nice)
6189 /* convert nice value [19,-20] to rlimit style value [1,40] */
6190 int nice_rlim = 20 - nice;
6192 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6193 capable(CAP_SYS_NICE));
6196 #ifdef __ARCH_WANT_SYS_NICE
6199 * sys_nice - change the priority of the current process.
6200 * @increment: priority increment
6202 * sys_setpriority is a more generic, but much slower function that
6203 * does similar things.
6205 SYSCALL_DEFINE1(nice, int, increment)
6210 * Setpriority might change our priority at the same moment.
6211 * We don't have to worry. Conceptually one call occurs first
6212 * and we have a single winner.
6214 if (increment < -40)
6219 nice = TASK_NICE(current) + increment;
6225 if (increment < 0 && !can_nice(current, nice))
6228 retval = security_task_setnice(current, nice);
6232 set_user_nice(current, nice);
6239 * task_prio - return the priority value of a given task.
6240 * @p: the task in question.
6242 * This is the priority value as seen by users in /proc.
6243 * RT tasks are offset by -200. Normal tasks are centered
6244 * around 0, value goes from -16 to +15.
6246 int task_prio(const struct task_struct *p)
6248 return p->prio - MAX_RT_PRIO;
6252 * task_nice - return the nice value of a given task.
6253 * @p: the task in question.
6255 int task_nice(const struct task_struct *p)
6257 return TASK_NICE(p);
6259 EXPORT_SYMBOL(task_nice);
6262 * idle_cpu - is a given cpu idle currently?
6263 * @cpu: the processor in question.
6265 int idle_cpu(int cpu)
6267 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6271 * idle_task - return the idle task for a given cpu.
6272 * @cpu: the processor in question.
6274 struct task_struct *idle_task(int cpu)
6276 return cpu_rq(cpu)->idle;
6280 * find_process_by_pid - find a process with a matching PID value.
6281 * @pid: the pid in question.
6283 static struct task_struct *find_process_by_pid(pid_t pid)
6285 return pid ? find_task_by_vpid(pid) : current;
6288 /* Actually do priority change: must hold rq lock. */
6290 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6292 BUG_ON(p->se.on_rq);
6295 switch (p->policy) {
6299 p->sched_class = &fair_sched_class;
6303 p->sched_class = &rt_sched_class;
6307 p->rt_priority = prio;
6308 p->normal_prio = normal_prio(p);
6309 /* we are holding p->pi_lock already */
6310 p->prio = rt_mutex_getprio(p);
6315 * check the target process has a UID that matches the current process's
6317 static bool check_same_owner(struct task_struct *p)
6319 const struct cred *cred = current_cred(), *pcred;
6323 pcred = __task_cred(p);
6324 match = (cred->euid == pcred->euid ||
6325 cred->euid == pcred->uid);
6330 static int __sched_setscheduler(struct task_struct *p, int policy,
6331 struct sched_param *param, bool user)
6333 int retval, oldprio, oldpolicy = -1, on_rq, running;
6334 unsigned long flags;
6335 const struct sched_class *prev_class;
6339 /* may grab non-irq protected spin_locks */
6340 BUG_ON(in_interrupt());
6342 /* double check policy once rq lock held */
6344 reset_on_fork = p->sched_reset_on_fork;
6345 policy = oldpolicy = p->policy;
6347 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6348 policy &= ~SCHED_RESET_ON_FORK;
6350 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6351 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6352 policy != SCHED_IDLE)
6357 * Valid priorities for SCHED_FIFO and SCHED_RR are
6358 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6359 * SCHED_BATCH and SCHED_IDLE is 0.
6361 if (param->sched_priority < 0 ||
6362 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6363 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6365 if (rt_policy(policy) != (param->sched_priority != 0))
6369 * Allow unprivileged RT tasks to decrease priority:
6371 if (user && !capable(CAP_SYS_NICE)) {
6372 if (rt_policy(policy)) {
6373 unsigned long rlim_rtprio;
6375 if (!lock_task_sighand(p, &flags))
6377 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6378 unlock_task_sighand(p, &flags);
6380 /* can't set/change the rt policy */
6381 if (policy != p->policy && !rlim_rtprio)
6384 /* can't increase priority */
6385 if (param->sched_priority > p->rt_priority &&
6386 param->sched_priority > rlim_rtprio)
6390 * Like positive nice levels, dont allow tasks to
6391 * move out of SCHED_IDLE either:
6393 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6396 /* can't change other user's priorities */
6397 if (!check_same_owner(p))
6400 /* Normal users shall not reset the sched_reset_on_fork flag */
6401 if (p->sched_reset_on_fork && !reset_on_fork)
6406 #ifdef CONFIG_RT_GROUP_SCHED
6408 * Do not allow realtime tasks into groups that have no runtime
6411 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6412 task_group(p)->rt_bandwidth.rt_runtime == 0)
6416 retval = security_task_setscheduler(p, policy, param);
6422 * make sure no PI-waiters arrive (or leave) while we are
6423 * changing the priority of the task:
6425 spin_lock_irqsave(&p->pi_lock, flags);
6427 * To be able to change p->policy safely, the apropriate
6428 * runqueue lock must be held.
6430 rq = __task_rq_lock(p);
6431 /* recheck policy now with rq lock held */
6432 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6433 policy = oldpolicy = -1;
6434 __task_rq_unlock(rq);
6435 spin_unlock_irqrestore(&p->pi_lock, flags);
6438 update_rq_clock(rq);
6439 on_rq = p->se.on_rq;
6440 running = task_current(rq, p);
6442 deactivate_task(rq, p, 0);
6444 p->sched_class->put_prev_task(rq, p);
6446 p->sched_reset_on_fork = reset_on_fork;
6449 prev_class = p->sched_class;
6450 __setscheduler(rq, p, policy, param->sched_priority);
6453 p->sched_class->set_curr_task(rq);
6455 activate_task(rq, p, 0);
6457 check_class_changed(rq, p, prev_class, oldprio, running);
6459 __task_rq_unlock(rq);
6460 spin_unlock_irqrestore(&p->pi_lock, flags);
6462 rt_mutex_adjust_pi(p);
6468 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6469 * @p: the task in question.
6470 * @policy: new policy.
6471 * @param: structure containing the new RT priority.
6473 * NOTE that the task may be already dead.
6475 int sched_setscheduler(struct task_struct *p, int policy,
6476 struct sched_param *param)
6478 return __sched_setscheduler(p, policy, param, true);
6480 EXPORT_SYMBOL_GPL(sched_setscheduler);
6483 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6484 * @p: the task in question.
6485 * @policy: new policy.
6486 * @param: structure containing the new RT priority.
6488 * Just like sched_setscheduler, only don't bother checking if the
6489 * current context has permission. For example, this is needed in
6490 * stop_machine(): we create temporary high priority worker threads,
6491 * but our caller might not have that capability.
6493 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6494 struct sched_param *param)
6496 return __sched_setscheduler(p, policy, param, false);
6500 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6502 struct sched_param lparam;
6503 struct task_struct *p;
6506 if (!param || pid < 0)
6508 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6513 p = find_process_by_pid(pid);
6515 retval = sched_setscheduler(p, policy, &lparam);
6522 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6523 * @pid: the pid in question.
6524 * @policy: new policy.
6525 * @param: structure containing the new RT priority.
6527 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6528 struct sched_param __user *, param)
6530 /* negative values for policy are not valid */
6534 return do_sched_setscheduler(pid, policy, param);
6538 * sys_sched_setparam - set/change the RT priority of a thread
6539 * @pid: the pid in question.
6540 * @param: structure containing the new RT priority.
6542 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6544 return do_sched_setscheduler(pid, -1, param);
6548 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6549 * @pid: the pid in question.
6551 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6553 struct task_struct *p;
6561 p = find_process_by_pid(pid);
6563 retval = security_task_getscheduler(p);
6566 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6573 * sys_sched_getparam - get the RT priority of a thread
6574 * @pid: the pid in question.
6575 * @param: structure containing the RT priority.
6577 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6579 struct sched_param lp;
6580 struct task_struct *p;
6583 if (!param || pid < 0)
6587 p = find_process_by_pid(pid);
6592 retval = security_task_getscheduler(p);
6596 lp.sched_priority = p->rt_priority;
6600 * This one might sleep, we cannot do it with a spinlock held ...
6602 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6611 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6613 cpumask_var_t cpus_allowed, new_mask;
6614 struct task_struct *p;
6620 p = find_process_by_pid(pid);
6627 /* Prevent p going away */
6631 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6635 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6637 goto out_free_cpus_allowed;
6640 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6643 retval = security_task_setscheduler(p, 0, NULL);
6647 cpuset_cpus_allowed(p, cpus_allowed);
6648 cpumask_and(new_mask, in_mask, cpus_allowed);
6650 retval = set_cpus_allowed_ptr(p, new_mask);
6653 cpuset_cpus_allowed(p, cpus_allowed);
6654 if (!cpumask_subset(new_mask, cpus_allowed)) {
6656 * We must have raced with a concurrent cpuset
6657 * update. Just reset the cpus_allowed to the
6658 * cpuset's cpus_allowed
6660 cpumask_copy(new_mask, cpus_allowed);
6665 free_cpumask_var(new_mask);
6666 out_free_cpus_allowed:
6667 free_cpumask_var(cpus_allowed);
6674 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6675 struct cpumask *new_mask)
6677 if (len < cpumask_size())
6678 cpumask_clear(new_mask);
6679 else if (len > cpumask_size())
6680 len = cpumask_size();
6682 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6686 * sys_sched_setaffinity - set the cpu affinity of a process
6687 * @pid: pid of the process
6688 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6689 * @user_mask_ptr: user-space pointer to the new cpu mask
6691 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6692 unsigned long __user *, user_mask_ptr)
6694 cpumask_var_t new_mask;
6697 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6700 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6702 retval = sched_setaffinity(pid, new_mask);
6703 free_cpumask_var(new_mask);
6707 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6709 struct task_struct *p;
6710 unsigned long flags;
6718 p = find_process_by_pid(pid);
6722 retval = security_task_getscheduler(p);
6726 rq = task_rq_lock(p, &flags);
6727 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6728 task_rq_unlock(rq, &flags);
6738 * sys_sched_getaffinity - get the cpu affinity of a process
6739 * @pid: pid of the process
6740 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6741 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6743 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6744 unsigned long __user *, user_mask_ptr)
6749 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
6751 if (len & (sizeof(unsigned long)-1))
6754 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6757 ret = sched_getaffinity(pid, mask);
6759 size_t retlen = min_t(size_t, len, cpumask_size());
6761 if (copy_to_user(user_mask_ptr, mask, retlen))
6766 free_cpumask_var(mask);
6772 * sys_sched_yield - yield the current processor to other threads.
6774 * This function yields the current CPU to other tasks. If there are no
6775 * other threads running on this CPU then this function will return.
6777 SYSCALL_DEFINE0(sched_yield)
6779 struct rq *rq = this_rq_lock();
6781 schedstat_inc(rq, yld_count);
6782 current->sched_class->yield_task(rq);
6785 * Since we are going to call schedule() anyway, there's
6786 * no need to preempt or enable interrupts:
6788 __release(rq->lock);
6789 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6790 _raw_spin_unlock(&rq->lock);
6791 preempt_enable_no_resched();
6798 static inline int should_resched(void)
6800 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6803 static void __cond_resched(void)
6805 add_preempt_count(PREEMPT_ACTIVE);
6807 sub_preempt_count(PREEMPT_ACTIVE);
6810 int __sched _cond_resched(void)
6812 if (should_resched()) {
6818 EXPORT_SYMBOL(_cond_resched);
6821 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6822 * call schedule, and on return reacquire the lock.
6824 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6825 * operations here to prevent schedule() from being called twice (once via
6826 * spin_unlock(), once by hand).
6828 int __cond_resched_lock(spinlock_t *lock)
6830 int resched = should_resched();
6833 lockdep_assert_held(lock);
6835 if (spin_needbreak(lock) || resched) {
6846 EXPORT_SYMBOL(__cond_resched_lock);
6848 int __sched __cond_resched_softirq(void)
6850 BUG_ON(!in_softirq());
6852 if (should_resched()) {
6860 EXPORT_SYMBOL(__cond_resched_softirq);
6863 * yield - yield the current processor to other threads.
6865 * This is a shortcut for kernel-space yielding - it marks the
6866 * thread runnable and calls sys_sched_yield().
6868 void __sched yield(void)
6870 set_current_state(TASK_RUNNING);
6873 EXPORT_SYMBOL(yield);
6876 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6877 * that process accounting knows that this is a task in IO wait state.
6879 void __sched io_schedule(void)
6881 struct rq *rq = raw_rq();
6883 delayacct_blkio_start();
6884 atomic_inc(&rq->nr_iowait);
6885 current->in_iowait = 1;
6887 current->in_iowait = 0;
6888 atomic_dec(&rq->nr_iowait);
6889 delayacct_blkio_end();
6891 EXPORT_SYMBOL(io_schedule);
6893 long __sched io_schedule_timeout(long timeout)
6895 struct rq *rq = raw_rq();
6898 delayacct_blkio_start();
6899 atomic_inc(&rq->nr_iowait);
6900 current->in_iowait = 1;
6901 ret = schedule_timeout(timeout);
6902 current->in_iowait = 0;
6903 atomic_dec(&rq->nr_iowait);
6904 delayacct_blkio_end();
6909 * sys_sched_get_priority_max - return maximum RT priority.
6910 * @policy: scheduling class.
6912 * this syscall returns the maximum rt_priority that can be used
6913 * by a given scheduling class.
6915 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6922 ret = MAX_USER_RT_PRIO-1;
6934 * sys_sched_get_priority_min - return minimum RT priority.
6935 * @policy: scheduling class.
6937 * this syscall returns the minimum rt_priority that can be used
6938 * by a given scheduling class.
6940 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6958 * sys_sched_rr_get_interval - return the default timeslice of a process.
6959 * @pid: pid of the process.
6960 * @interval: userspace pointer to the timeslice value.
6962 * this syscall writes the default timeslice value of a given process
6963 * into the user-space timespec buffer. A value of '0' means infinity.
6965 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6966 struct timespec __user *, interval)
6968 struct task_struct *p;
6969 unsigned int time_slice;
6970 unsigned long flags;
6980 p = find_process_by_pid(pid);
6984 retval = security_task_getscheduler(p);
6988 rq = task_rq_lock(p, &flags);
6989 time_slice = p->sched_class->get_rr_interval(rq, p);
6990 task_rq_unlock(rq, &flags);
6993 jiffies_to_timespec(time_slice, &t);
6994 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
7002 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
7004 void sched_show_task(struct task_struct *p)
7006 unsigned long free = 0;
7009 state = p->state ? __ffs(p->state) + 1 : 0;
7010 printk(KERN_INFO "%-13.13s %c", p->comm,
7011 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
7012 #if BITS_PER_LONG == 32
7013 if (state == TASK_RUNNING)
7014 printk(KERN_CONT " running ");
7016 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
7018 if (state == TASK_RUNNING)
7019 printk(KERN_CONT " running task ");
7021 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
7023 #ifdef CONFIG_DEBUG_STACK_USAGE
7024 free = stack_not_used(p);
7026 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
7027 task_pid_nr(p), task_pid_nr(p->real_parent),
7028 (unsigned long)task_thread_info(p)->flags);
7030 show_stack(p, NULL);
7033 void show_state_filter(unsigned long state_filter)
7035 struct task_struct *g, *p;
7037 #if BITS_PER_LONG == 32
7039 " task PC stack pid father\n");
7042 " task PC stack pid father\n");
7044 read_lock(&tasklist_lock);
7045 do_each_thread(g, p) {
7047 * reset the NMI-timeout, listing all files on a slow
7048 * console might take alot of time:
7050 touch_nmi_watchdog();
7051 if (!state_filter || (p->state & state_filter))
7053 } while_each_thread(g, p);
7055 touch_all_softlockup_watchdogs();
7057 #ifdef CONFIG_SCHED_DEBUG
7058 sysrq_sched_debug_show();
7060 read_unlock(&tasklist_lock);
7062 * Only show locks if all tasks are dumped:
7064 if (state_filter == -1)
7065 debug_show_all_locks();
7068 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7070 idle->sched_class = &idle_sched_class;
7074 * init_idle - set up an idle thread for a given CPU
7075 * @idle: task in question
7076 * @cpu: cpu the idle task belongs to
7078 * NOTE: this function does not set the idle thread's NEED_RESCHED
7079 * flag, to make booting more robust.
7081 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7083 struct rq *rq = cpu_rq(cpu);
7084 unsigned long flags;
7086 spin_lock_irqsave(&rq->lock, flags);
7089 idle->state = TASK_RUNNING;
7090 idle->se.exec_start = sched_clock();
7092 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7094 * We're having a chicken and egg problem, even though we are
7095 * holding rq->lock, the cpu isn't yet set to this cpu so the
7096 * lockdep check in task_group() will fail.
7098 * Similar case to sched_fork(). / Alternatively we could
7099 * use task_rq_lock() here and obtain the other rq->lock.
7104 __set_task_cpu(idle, cpu);
7107 rq->curr = rq->idle = idle;
7108 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7111 spin_unlock_irqrestore(&rq->lock, flags);
7113 /* Set the preempt count _outside_ the spinlocks! */
7114 #if defined(CONFIG_PREEMPT)
7115 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7117 task_thread_info(idle)->preempt_count = 0;
7120 * The idle tasks have their own, simple scheduling class:
7122 idle->sched_class = &idle_sched_class;
7123 ftrace_graph_init_task(idle);
7127 * In a system that switches off the HZ timer nohz_cpu_mask
7128 * indicates which cpus entered this state. This is used
7129 * in the rcu update to wait only for active cpus. For system
7130 * which do not switch off the HZ timer nohz_cpu_mask should
7131 * always be CPU_BITS_NONE.
7133 cpumask_var_t nohz_cpu_mask;
7136 * Increase the granularity value when there are more CPUs,
7137 * because with more CPUs the 'effective latency' as visible
7138 * to users decreases. But the relationship is not linear,
7139 * so pick a second-best guess by going with the log2 of the
7142 * This idea comes from the SD scheduler of Con Kolivas:
7144 static void update_sysctl(void)
7146 unsigned int cpus = min(num_online_cpus(), 8U);
7147 unsigned int factor = 1 + ilog2(cpus);
7149 #define SET_SYSCTL(name) \
7150 (sysctl_##name = (factor) * normalized_sysctl_##name)
7151 SET_SYSCTL(sched_min_granularity);
7152 SET_SYSCTL(sched_latency);
7153 SET_SYSCTL(sched_wakeup_granularity);
7154 SET_SYSCTL(sched_shares_ratelimit);
7158 static inline void sched_init_granularity(void)
7165 * This is how migration works:
7167 * 1) we queue a struct migration_req structure in the source CPU's
7168 * runqueue and wake up that CPU's migration thread.
7169 * 2) we down() the locked semaphore => thread blocks.
7170 * 3) migration thread wakes up (implicitly it forces the migrated
7171 * thread off the CPU)
7172 * 4) it gets the migration request and checks whether the migrated
7173 * task is still in the wrong runqueue.
7174 * 5) if it's in the wrong runqueue then the migration thread removes
7175 * it and puts it into the right queue.
7176 * 6) migration thread up()s the semaphore.
7177 * 7) we wake up and the migration is done.
7181 * Change a given task's CPU affinity. Migrate the thread to a
7182 * proper CPU and schedule it away if the CPU it's executing on
7183 * is removed from the allowed bitmask.
7185 * NOTE: the caller must have a valid reference to the task, the
7186 * task must not exit() & deallocate itself prematurely. The
7187 * call is not atomic; no spinlocks may be held.
7189 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7191 struct migration_req req;
7192 unsigned long flags;
7197 * Serialize against TASK_WAKING so that ttwu() and wunt() can
7198 * drop the rq->lock and still rely on ->cpus_allowed.
7201 while (task_is_waking(p))
7203 rq = task_rq_lock(p, &flags);
7204 if (task_is_waking(p)) {
7205 task_rq_unlock(rq, &flags);
7209 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7214 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7215 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7220 if (p->sched_class->set_cpus_allowed)
7221 p->sched_class->set_cpus_allowed(p, new_mask);
7223 cpumask_copy(&p->cpus_allowed, new_mask);
7224 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7227 /* Can the task run on the task's current CPU? If so, we're done */
7228 if (cpumask_test_cpu(task_cpu(p), new_mask))
7231 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7232 /* Need help from migration thread: drop lock and wait. */
7233 struct task_struct *mt = rq->migration_thread;
7235 get_task_struct(mt);
7236 task_rq_unlock(rq, &flags);
7237 wake_up_process(mt);
7238 put_task_struct(mt);
7239 wait_for_completion(&req.done);
7240 tlb_migrate_finish(p->mm);
7244 task_rq_unlock(rq, &flags);
7248 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7251 * Move (not current) task off this cpu, onto dest cpu. We're doing
7252 * this because either it can't run here any more (set_cpus_allowed()
7253 * away from this CPU, or CPU going down), or because we're
7254 * attempting to rebalance this task on exec (sched_exec).
7256 * So we race with normal scheduler movements, but that's OK, as long
7257 * as the task is no longer on this CPU.
7259 * Returns non-zero if task was successfully migrated.
7261 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7263 struct rq *rq_dest, *rq_src;
7266 if (unlikely(!cpu_active(dest_cpu)))
7269 rq_src = cpu_rq(src_cpu);
7270 rq_dest = cpu_rq(dest_cpu);
7272 double_rq_lock(rq_src, rq_dest);
7273 /* Already moved. */
7274 if (task_cpu(p) != src_cpu)
7276 /* Affinity changed (again). */
7277 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7281 * If we're not on a rq, the next wake-up will ensure we're
7285 deactivate_task(rq_src, p, 0);
7286 set_task_cpu(p, dest_cpu);
7287 activate_task(rq_dest, p, 0);
7288 check_preempt_curr(rq_dest, p, 0);
7293 double_rq_unlock(rq_src, rq_dest);
7297 #define RCU_MIGRATION_IDLE 0
7298 #define RCU_MIGRATION_NEED_QS 1
7299 #define RCU_MIGRATION_GOT_QS 2
7300 #define RCU_MIGRATION_MUST_SYNC 3
7303 * migration_thread - this is a highprio system thread that performs
7304 * thread migration by bumping thread off CPU then 'pushing' onto
7307 static int migration_thread(void *data)
7310 int cpu = (long)data;
7314 BUG_ON(rq->migration_thread != current);
7316 set_current_state(TASK_INTERRUPTIBLE);
7317 while (!kthread_should_stop()) {
7318 struct migration_req *req;
7319 struct list_head *head;
7321 spin_lock_irq(&rq->lock);
7323 if (cpu_is_offline(cpu)) {
7324 spin_unlock_irq(&rq->lock);
7328 if (rq->active_balance) {
7329 active_load_balance(rq, cpu);
7330 rq->active_balance = 0;
7333 head = &rq->migration_queue;
7335 if (list_empty(head)) {
7336 spin_unlock_irq(&rq->lock);
7338 set_current_state(TASK_INTERRUPTIBLE);
7341 req = list_entry(head->next, struct migration_req, list);
7342 list_del_init(head->next);
7344 if (req->task != NULL) {
7345 spin_unlock(&rq->lock);
7346 __migrate_task(req->task, cpu, req->dest_cpu);
7347 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7348 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7349 spin_unlock(&rq->lock);
7351 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7352 spin_unlock(&rq->lock);
7353 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7357 complete(&req->done);
7359 __set_current_state(TASK_RUNNING);
7364 #ifdef CONFIG_HOTPLUG_CPU
7366 * Figure out where task on dead CPU should go, use force if necessary.
7368 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7370 struct rq *rq = cpu_rq(dead_cpu);
7371 int needs_cpu, uninitialized_var(dest_cpu);
7372 unsigned long flags;
7374 local_irq_save(flags);
7376 spin_lock(&rq->lock);
7377 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
7379 dest_cpu = select_fallback_rq(dead_cpu, p);
7380 spin_unlock(&rq->lock);
7382 * It can only fail if we race with set_cpus_allowed(),
7383 * in the racer should migrate the task anyway.
7386 __migrate_task(p, dead_cpu, dest_cpu);
7387 local_irq_restore(flags);
7391 * While a dead CPU has no uninterruptible tasks queued at this point,
7392 * it might still have a nonzero ->nr_uninterruptible counter, because
7393 * for performance reasons the counter is not stricly tracking tasks to
7394 * their home CPUs. So we just add the counter to another CPU's counter,
7395 * to keep the global sum constant after CPU-down:
7397 static void migrate_nr_uninterruptible(struct rq *rq_src)
7399 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7400 unsigned long flags;
7402 local_irq_save(flags);
7403 double_rq_lock(rq_src, rq_dest);
7404 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7405 rq_src->nr_uninterruptible = 0;
7406 double_rq_unlock(rq_src, rq_dest);
7407 local_irq_restore(flags);
7410 /* Run through task list and migrate tasks from the dead cpu. */
7411 static void migrate_live_tasks(int src_cpu)
7413 struct task_struct *p, *t;
7415 read_lock(&tasklist_lock);
7417 do_each_thread(t, p) {
7421 if (task_cpu(p) == src_cpu)
7422 move_task_off_dead_cpu(src_cpu, p);
7423 } while_each_thread(t, p);
7425 read_unlock(&tasklist_lock);
7429 * Schedules idle task to be the next runnable task on current CPU.
7430 * It does so by boosting its priority to highest possible.
7431 * Used by CPU offline code.
7433 void sched_idle_next(void)
7435 int this_cpu = smp_processor_id();
7436 struct rq *rq = cpu_rq(this_cpu);
7437 struct task_struct *p = rq->idle;
7438 unsigned long flags;
7440 /* cpu has to be offline */
7441 BUG_ON(cpu_online(this_cpu));
7444 * Strictly not necessary since rest of the CPUs are stopped by now
7445 * and interrupts disabled on the current cpu.
7447 spin_lock_irqsave(&rq->lock, flags);
7449 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7451 update_rq_clock(rq);
7452 activate_task(rq, p, 0);
7454 spin_unlock_irqrestore(&rq->lock, flags);
7458 * Ensures that the idle task is using init_mm right before its cpu goes
7461 void idle_task_exit(void)
7463 struct mm_struct *mm = current->active_mm;
7465 BUG_ON(cpu_online(smp_processor_id()));
7468 switch_mm(mm, &init_mm, current);
7472 /* called under rq->lock with disabled interrupts */
7473 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7475 struct rq *rq = cpu_rq(dead_cpu);
7477 /* Must be exiting, otherwise would be on tasklist. */
7478 BUG_ON(!p->exit_state);
7480 /* Cannot have done final schedule yet: would have vanished. */
7481 BUG_ON(p->state == TASK_DEAD);
7486 * Drop lock around migration; if someone else moves it,
7487 * that's OK. No task can be added to this CPU, so iteration is
7490 spin_unlock_irq(&rq->lock);
7491 move_task_off_dead_cpu(dead_cpu, p);
7492 spin_lock_irq(&rq->lock);
7497 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7498 static void migrate_dead_tasks(unsigned int dead_cpu)
7500 struct rq *rq = cpu_rq(dead_cpu);
7501 struct task_struct *next;
7504 if (!rq->nr_running)
7506 update_rq_clock(rq);
7507 next = pick_next_task(rq);
7510 next->sched_class->put_prev_task(rq, next);
7511 migrate_dead(dead_cpu, next);
7517 * remove the tasks which were accounted by rq from calc_load_tasks.
7519 static void calc_global_load_remove(struct rq *rq)
7521 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7522 rq->calc_load_active = 0;
7524 #endif /* CONFIG_HOTPLUG_CPU */
7526 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7528 static struct ctl_table sd_ctl_dir[] = {
7530 .procname = "sched_domain",
7536 static struct ctl_table sd_ctl_root[] = {
7538 .ctl_name = CTL_KERN,
7539 .procname = "kernel",
7541 .child = sd_ctl_dir,
7546 static struct ctl_table *sd_alloc_ctl_entry(int n)
7548 struct ctl_table *entry =
7549 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7554 static void sd_free_ctl_entry(struct ctl_table **tablep)
7556 struct ctl_table *entry;
7559 * In the intermediate directories, both the child directory and
7560 * procname are dynamically allocated and could fail but the mode
7561 * will always be set. In the lowest directory the names are
7562 * static strings and all have proc handlers.
7564 for (entry = *tablep; entry->mode; entry++) {
7566 sd_free_ctl_entry(&entry->child);
7567 if (entry->proc_handler == NULL)
7568 kfree(entry->procname);
7576 set_table_entry(struct ctl_table *entry,
7577 const char *procname, void *data, int maxlen,
7578 mode_t mode, proc_handler *proc_handler)
7580 entry->procname = procname;
7582 entry->maxlen = maxlen;
7584 entry->proc_handler = proc_handler;
7587 static struct ctl_table *
7588 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7590 struct ctl_table *table = sd_alloc_ctl_entry(13);
7595 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7596 sizeof(long), 0644, proc_doulongvec_minmax);
7597 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7598 sizeof(long), 0644, proc_doulongvec_minmax);
7599 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7600 sizeof(int), 0644, proc_dointvec_minmax);
7601 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7602 sizeof(int), 0644, proc_dointvec_minmax);
7603 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7604 sizeof(int), 0644, proc_dointvec_minmax);
7605 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7606 sizeof(int), 0644, proc_dointvec_minmax);
7607 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7608 sizeof(int), 0644, proc_dointvec_minmax);
7609 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7610 sizeof(int), 0644, proc_dointvec_minmax);
7611 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7612 sizeof(int), 0644, proc_dointvec_minmax);
7613 set_table_entry(&table[9], "cache_nice_tries",
7614 &sd->cache_nice_tries,
7615 sizeof(int), 0644, proc_dointvec_minmax);
7616 set_table_entry(&table[10], "flags", &sd->flags,
7617 sizeof(int), 0644, proc_dointvec_minmax);
7618 set_table_entry(&table[11], "name", sd->name,
7619 CORENAME_MAX_SIZE, 0444, proc_dostring);
7620 /* &table[12] is terminator */
7625 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7627 struct ctl_table *entry, *table;
7628 struct sched_domain *sd;
7629 int domain_num = 0, i;
7632 for_each_domain(cpu, sd)
7634 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7639 for_each_domain(cpu, sd) {
7640 snprintf(buf, 32, "domain%d", i);
7641 entry->procname = kstrdup(buf, GFP_KERNEL);
7643 entry->child = sd_alloc_ctl_domain_table(sd);
7650 static struct ctl_table_header *sd_sysctl_header;
7651 static void register_sched_domain_sysctl(void)
7653 int i, cpu_num = num_possible_cpus();
7654 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7657 WARN_ON(sd_ctl_dir[0].child);
7658 sd_ctl_dir[0].child = entry;
7663 for_each_possible_cpu(i) {
7664 snprintf(buf, 32, "cpu%d", i);
7665 entry->procname = kstrdup(buf, GFP_KERNEL);
7667 entry->child = sd_alloc_ctl_cpu_table(i);
7671 WARN_ON(sd_sysctl_header);
7672 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7675 /* may be called multiple times per register */
7676 static void unregister_sched_domain_sysctl(void)
7678 if (sd_sysctl_header)
7679 unregister_sysctl_table(sd_sysctl_header);
7680 sd_sysctl_header = NULL;
7681 if (sd_ctl_dir[0].child)
7682 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7685 static void register_sched_domain_sysctl(void)
7688 static void unregister_sched_domain_sysctl(void)
7693 static void set_rq_online(struct rq *rq)
7696 const struct sched_class *class;
7698 cpumask_set_cpu(rq->cpu, rq->rd->online);
7701 for_each_class(class) {
7702 if (class->rq_online)
7703 class->rq_online(rq);
7708 static void set_rq_offline(struct rq *rq)
7711 const struct sched_class *class;
7713 for_each_class(class) {
7714 if (class->rq_offline)
7715 class->rq_offline(rq);
7718 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7724 * migration_call - callback that gets triggered when a CPU is added.
7725 * Here we can start up the necessary migration thread for the new CPU.
7727 static int __cpuinit
7728 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7730 struct task_struct *p;
7731 int cpu = (long)hcpu;
7732 unsigned long flags;
7735 switch (action & ~CPU_TASKS_FROZEN) {
7737 case CPU_UP_PREPARE:
7738 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7741 kthread_bind(p, cpu);
7742 /* Must be high prio: stop_machine expects to yield to it. */
7743 rq = task_rq_lock(p, &flags);
7744 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7745 task_rq_unlock(rq, &flags);
7747 cpu_rq(cpu)->migration_thread = p;
7748 rq->calc_load_update = calc_load_update;
7752 /* Strictly unnecessary, as first user will wake it. */
7753 wake_up_process(cpu_rq(cpu)->migration_thread);
7755 /* Update our root-domain */
7757 spin_lock_irqsave(&rq->lock, flags);
7759 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7763 spin_unlock_irqrestore(&rq->lock, flags);
7766 #ifdef CONFIG_HOTPLUG_CPU
7767 case CPU_UP_CANCELED:
7768 if (!cpu_rq(cpu)->migration_thread)
7770 /* Unbind it from offline cpu so it can run. Fall thru. */
7771 kthread_bind(cpu_rq(cpu)->migration_thread,
7772 cpumask_any(cpu_online_mask));
7773 kthread_stop(cpu_rq(cpu)->migration_thread);
7774 put_task_struct(cpu_rq(cpu)->migration_thread);
7775 cpu_rq(cpu)->migration_thread = NULL;
7780 * Bring the migration thread down in CPU_POST_DEAD event,
7781 * since the timers should have got migrated by now and thus
7782 * we should not see a deadlock between trying to kill the
7783 * migration thread and the sched_rt_period_timer.
7786 kthread_stop(rq->migration_thread);
7787 put_task_struct(rq->migration_thread);
7788 rq->migration_thread = NULL;
7792 migrate_live_tasks(cpu);
7794 /* Idle task back to normal (off runqueue, low prio) */
7795 spin_lock_irq(&rq->lock);
7796 update_rq_clock(rq);
7797 deactivate_task(rq, rq->idle, 0);
7798 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7799 rq->idle->sched_class = &idle_sched_class;
7800 migrate_dead_tasks(cpu);
7801 spin_unlock_irq(&rq->lock);
7802 migrate_nr_uninterruptible(rq);
7803 BUG_ON(rq->nr_running != 0);
7804 calc_global_load_remove(rq);
7806 * No need to migrate the tasks: it was best-effort if
7807 * they didn't take sched_hotcpu_mutex. Just wake up
7810 spin_lock_irq(&rq->lock);
7811 while (!list_empty(&rq->migration_queue)) {
7812 struct migration_req *req;
7814 req = list_entry(rq->migration_queue.next,
7815 struct migration_req, list);
7816 list_del_init(&req->list);
7817 spin_unlock_irq(&rq->lock);
7818 complete(&req->done);
7819 spin_lock_irq(&rq->lock);
7821 spin_unlock_irq(&rq->lock);
7825 /* Update our root-domain */
7827 spin_lock_irqsave(&rq->lock, flags);
7829 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7832 spin_unlock_irqrestore(&rq->lock, flags);
7840 * Register at high priority so that task migration (migrate_all_tasks)
7841 * happens before everything else. This has to be lower priority than
7842 * the notifier in the perf_event subsystem, though.
7844 static struct notifier_block __cpuinitdata migration_notifier = {
7845 .notifier_call = migration_call,
7849 static int __init migration_init(void)
7851 void *cpu = (void *)(long)smp_processor_id();
7854 /* Start one for the boot CPU: */
7855 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7856 BUG_ON(err == NOTIFY_BAD);
7857 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7858 register_cpu_notifier(&migration_notifier);
7862 early_initcall(migration_init);
7867 #ifdef CONFIG_SCHED_DEBUG
7869 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7870 struct cpumask *groupmask)
7872 struct sched_group *group = sd->groups;
7875 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7876 cpumask_clear(groupmask);
7878 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7880 if (!(sd->flags & SD_LOAD_BALANCE)) {
7881 printk("does not load-balance\n");
7883 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7888 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7890 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7891 printk(KERN_ERR "ERROR: domain->span does not contain "
7894 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7895 printk(KERN_ERR "ERROR: domain->groups does not contain"
7899 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7903 printk(KERN_ERR "ERROR: group is NULL\n");
7907 if (!group->cpu_power) {
7908 printk(KERN_CONT "\n");
7909 printk(KERN_ERR "ERROR: domain->cpu_power not "
7914 if (!cpumask_weight(sched_group_cpus(group))) {
7915 printk(KERN_CONT "\n");
7916 printk(KERN_ERR "ERROR: empty group\n");
7920 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7921 printk(KERN_CONT "\n");
7922 printk(KERN_ERR "ERROR: repeated CPUs\n");
7926 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7928 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7930 printk(KERN_CONT " %s", str);
7931 if (group->cpu_power != SCHED_LOAD_SCALE) {
7932 printk(KERN_CONT " (cpu_power = %d)",
7936 group = group->next;
7937 } while (group != sd->groups);
7938 printk(KERN_CONT "\n");
7940 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7941 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7944 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7945 printk(KERN_ERR "ERROR: parent span is not a superset "
7946 "of domain->span\n");
7950 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7952 cpumask_var_t groupmask;
7956 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7960 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7962 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7963 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7968 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7975 free_cpumask_var(groupmask);
7977 #else /* !CONFIG_SCHED_DEBUG */
7978 # define sched_domain_debug(sd, cpu) do { } while (0)
7979 #endif /* CONFIG_SCHED_DEBUG */
7981 static int sd_degenerate(struct sched_domain *sd)
7983 if (cpumask_weight(sched_domain_span(sd)) == 1)
7986 /* Following flags need at least 2 groups */
7987 if (sd->flags & (SD_LOAD_BALANCE |
7988 SD_BALANCE_NEWIDLE |
7992 SD_SHARE_PKG_RESOURCES)) {
7993 if (sd->groups != sd->groups->next)
7997 /* Following flags don't use groups */
7998 if (sd->flags & (SD_WAKE_AFFINE))
8005 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
8007 unsigned long cflags = sd->flags, pflags = parent->flags;
8009 if (sd_degenerate(parent))
8012 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
8015 /* Flags needing groups don't count if only 1 group in parent */
8016 if (parent->groups == parent->groups->next) {
8017 pflags &= ~(SD_LOAD_BALANCE |
8018 SD_BALANCE_NEWIDLE |
8022 SD_SHARE_PKG_RESOURCES);
8023 if (nr_node_ids == 1)
8024 pflags &= ~SD_SERIALIZE;
8026 if (~cflags & pflags)
8032 static void free_rootdomain(struct root_domain *rd)
8034 synchronize_sched();
8036 cpupri_cleanup(&rd->cpupri);
8038 free_cpumask_var(rd->rto_mask);
8039 free_cpumask_var(rd->online);
8040 free_cpumask_var(rd->span);
8044 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8046 struct root_domain *old_rd = NULL;
8047 unsigned long flags;
8049 spin_lock_irqsave(&rq->lock, flags);
8054 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8057 cpumask_clear_cpu(rq->cpu, old_rd->span);
8060 * If we dont want to free the old_rt yet then
8061 * set old_rd to NULL to skip the freeing later
8064 if (!atomic_dec_and_test(&old_rd->refcount))
8068 atomic_inc(&rd->refcount);
8071 cpumask_set_cpu(rq->cpu, rd->span);
8072 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8075 spin_unlock_irqrestore(&rq->lock, flags);
8078 free_rootdomain(old_rd);
8081 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8083 gfp_t gfp = GFP_KERNEL;
8085 memset(rd, 0, sizeof(*rd));
8090 if (!alloc_cpumask_var(&rd->span, gfp))
8092 if (!alloc_cpumask_var(&rd->online, gfp))
8094 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8097 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8102 free_cpumask_var(rd->rto_mask);
8104 free_cpumask_var(rd->online);
8106 free_cpumask_var(rd->span);
8111 static void init_defrootdomain(void)
8113 init_rootdomain(&def_root_domain, true);
8115 atomic_set(&def_root_domain.refcount, 1);
8118 static struct root_domain *alloc_rootdomain(void)
8120 struct root_domain *rd;
8122 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8126 if (init_rootdomain(rd, false) != 0) {
8135 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8136 * hold the hotplug lock.
8139 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8141 struct rq *rq = cpu_rq(cpu);
8142 struct sched_domain *tmp;
8144 for (tmp = sd; tmp; tmp = tmp->parent)
8145 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
8147 /* Remove the sched domains which do not contribute to scheduling. */
8148 for (tmp = sd; tmp; ) {
8149 struct sched_domain *parent = tmp->parent;
8153 if (sd_parent_degenerate(tmp, parent)) {
8154 tmp->parent = parent->parent;
8156 parent->parent->child = tmp;
8161 if (sd && sd_degenerate(sd)) {
8167 sched_domain_debug(sd, cpu);
8169 rq_attach_root(rq, rd);
8170 rcu_assign_pointer(rq->sd, sd);
8173 /* cpus with isolated domains */
8174 static cpumask_var_t cpu_isolated_map;
8176 /* Setup the mask of cpus configured for isolated domains */
8177 static int __init isolated_cpu_setup(char *str)
8179 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8180 cpulist_parse(str, cpu_isolated_map);
8184 __setup("isolcpus=", isolated_cpu_setup);
8187 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8188 * to a function which identifies what group(along with sched group) a CPU
8189 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8190 * (due to the fact that we keep track of groups covered with a struct cpumask).
8192 * init_sched_build_groups will build a circular linked list of the groups
8193 * covered by the given span, and will set each group's ->cpumask correctly,
8194 * and ->cpu_power to 0.
8197 init_sched_build_groups(const struct cpumask *span,
8198 const struct cpumask *cpu_map,
8199 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8200 struct sched_group **sg,
8201 struct cpumask *tmpmask),
8202 struct cpumask *covered, struct cpumask *tmpmask)
8204 struct sched_group *first = NULL, *last = NULL;
8207 cpumask_clear(covered);
8209 for_each_cpu(i, span) {
8210 struct sched_group *sg;
8211 int group = group_fn(i, cpu_map, &sg, tmpmask);
8214 if (cpumask_test_cpu(i, covered))
8217 cpumask_clear(sched_group_cpus(sg));
8220 for_each_cpu(j, span) {
8221 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8224 cpumask_set_cpu(j, covered);
8225 cpumask_set_cpu(j, sched_group_cpus(sg));
8236 #define SD_NODES_PER_DOMAIN 16
8241 * find_next_best_node - find the next node to include in a sched_domain
8242 * @node: node whose sched_domain we're building
8243 * @used_nodes: nodes already in the sched_domain
8245 * Find the next node to include in a given scheduling domain. Simply
8246 * finds the closest node not already in the @used_nodes map.
8248 * Should use nodemask_t.
8250 static int find_next_best_node(int node, nodemask_t *used_nodes)
8252 int i, n, val, min_val, best_node = 0;
8256 for (i = 0; i < nr_node_ids; i++) {
8257 /* Start at @node */
8258 n = (node + i) % nr_node_ids;
8260 if (!nr_cpus_node(n))
8263 /* Skip already used nodes */
8264 if (node_isset(n, *used_nodes))
8267 /* Simple min distance search */
8268 val = node_distance(node, n);
8270 if (val < min_val) {
8276 node_set(best_node, *used_nodes);
8281 * sched_domain_node_span - get a cpumask for a node's sched_domain
8282 * @node: node whose cpumask we're constructing
8283 * @span: resulting cpumask
8285 * Given a node, construct a good cpumask for its sched_domain to span. It
8286 * should be one that prevents unnecessary balancing, but also spreads tasks
8289 static void sched_domain_node_span(int node, struct cpumask *span)
8291 nodemask_t used_nodes;
8294 cpumask_clear(span);
8295 nodes_clear(used_nodes);
8297 cpumask_or(span, span, cpumask_of_node(node));
8298 node_set(node, used_nodes);
8300 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8301 int next_node = find_next_best_node(node, &used_nodes);
8303 cpumask_or(span, span, cpumask_of_node(next_node));
8306 #endif /* CONFIG_NUMA */
8308 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8311 * The cpus mask in sched_group and sched_domain hangs off the end.
8313 * ( See the the comments in include/linux/sched.h:struct sched_group
8314 * and struct sched_domain. )
8316 struct static_sched_group {
8317 struct sched_group sg;
8318 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8321 struct static_sched_domain {
8322 struct sched_domain sd;
8323 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8329 cpumask_var_t domainspan;
8330 cpumask_var_t covered;
8331 cpumask_var_t notcovered;
8333 cpumask_var_t nodemask;
8334 cpumask_var_t this_sibling_map;
8335 cpumask_var_t this_core_map;
8336 cpumask_var_t send_covered;
8337 cpumask_var_t tmpmask;
8338 struct sched_group **sched_group_nodes;
8339 struct root_domain *rd;
8343 sa_sched_groups = 0,
8348 sa_this_sibling_map,
8350 sa_sched_group_nodes,
8360 * SMT sched-domains:
8362 #ifdef CONFIG_SCHED_SMT
8363 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8364 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8367 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8368 struct sched_group **sg, struct cpumask *unused)
8371 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8374 #endif /* CONFIG_SCHED_SMT */
8377 * multi-core sched-domains:
8379 #ifdef CONFIG_SCHED_MC
8380 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8381 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8382 #endif /* CONFIG_SCHED_MC */
8384 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8386 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8387 struct sched_group **sg, struct cpumask *mask)
8391 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8392 group = cpumask_first(mask);
8394 *sg = &per_cpu(sched_group_core, group).sg;
8397 #elif defined(CONFIG_SCHED_MC)
8399 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8400 struct sched_group **sg, struct cpumask *unused)
8403 *sg = &per_cpu(sched_group_core, cpu).sg;
8408 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8409 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8412 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8413 struct sched_group **sg, struct cpumask *mask)
8416 #ifdef CONFIG_SCHED_MC
8417 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8418 group = cpumask_first(mask);
8419 #elif defined(CONFIG_SCHED_SMT)
8420 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8421 group = cpumask_first(mask);
8426 *sg = &per_cpu(sched_group_phys, group).sg;
8432 * The init_sched_build_groups can't handle what we want to do with node
8433 * groups, so roll our own. Now each node has its own list of groups which
8434 * gets dynamically allocated.
8436 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8437 static struct sched_group ***sched_group_nodes_bycpu;
8439 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8440 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8442 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8443 struct sched_group **sg,
8444 struct cpumask *nodemask)
8448 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8449 group = cpumask_first(nodemask);
8452 *sg = &per_cpu(sched_group_allnodes, group).sg;
8456 static void init_numa_sched_groups_power(struct sched_group *group_head)
8458 struct sched_group *sg = group_head;
8464 for_each_cpu(j, sched_group_cpus(sg)) {
8465 struct sched_domain *sd;
8467 sd = &per_cpu(phys_domains, j).sd;
8468 if (j != group_first_cpu(sd->groups)) {
8470 * Only add "power" once for each
8476 sg->cpu_power += sd->groups->cpu_power;
8479 } while (sg != group_head);
8482 static int build_numa_sched_groups(struct s_data *d,
8483 const struct cpumask *cpu_map, int num)
8485 struct sched_domain *sd;
8486 struct sched_group *sg, *prev;
8489 cpumask_clear(d->covered);
8490 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8491 if (cpumask_empty(d->nodemask)) {
8492 d->sched_group_nodes[num] = NULL;
8496 sched_domain_node_span(num, d->domainspan);
8497 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8499 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8502 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8506 d->sched_group_nodes[num] = sg;
8508 for_each_cpu(j, d->nodemask) {
8509 sd = &per_cpu(node_domains, j).sd;
8514 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8516 cpumask_or(d->covered, d->covered, d->nodemask);
8519 for (j = 0; j < nr_node_ids; j++) {
8520 n = (num + j) % nr_node_ids;
8521 cpumask_complement(d->notcovered, d->covered);
8522 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8523 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8524 if (cpumask_empty(d->tmpmask))
8526 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8527 if (cpumask_empty(d->tmpmask))
8529 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8533 "Can not alloc domain group for node %d\n", j);
8537 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8538 sg->next = prev->next;
8539 cpumask_or(d->covered, d->covered, d->tmpmask);
8546 #endif /* CONFIG_NUMA */
8549 /* Free memory allocated for various sched_group structures */
8550 static void free_sched_groups(const struct cpumask *cpu_map,
8551 struct cpumask *nodemask)
8555 for_each_cpu(cpu, cpu_map) {
8556 struct sched_group **sched_group_nodes
8557 = sched_group_nodes_bycpu[cpu];
8559 if (!sched_group_nodes)
8562 for (i = 0; i < nr_node_ids; i++) {
8563 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8565 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8566 if (cpumask_empty(nodemask))
8576 if (oldsg != sched_group_nodes[i])
8579 kfree(sched_group_nodes);
8580 sched_group_nodes_bycpu[cpu] = NULL;
8583 #else /* !CONFIG_NUMA */
8584 static void free_sched_groups(const struct cpumask *cpu_map,
8585 struct cpumask *nodemask)
8588 #endif /* CONFIG_NUMA */
8591 * Initialize sched groups cpu_power.
8593 * cpu_power indicates the capacity of sched group, which is used while
8594 * distributing the load between different sched groups in a sched domain.
8595 * Typically cpu_power for all the groups in a sched domain will be same unless
8596 * there are asymmetries in the topology. If there are asymmetries, group
8597 * having more cpu_power will pickup more load compared to the group having
8600 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8602 struct sched_domain *child;
8603 struct sched_group *group;
8607 WARN_ON(!sd || !sd->groups);
8609 if (cpu != group_first_cpu(sd->groups))
8614 sd->groups->cpu_power = 0;
8617 power = SCHED_LOAD_SCALE;
8618 weight = cpumask_weight(sched_domain_span(sd));
8620 * SMT siblings share the power of a single core.
8621 * Usually multiple threads get a better yield out of
8622 * that one core than a single thread would have,
8623 * reflect that in sd->smt_gain.
8625 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8626 power *= sd->smt_gain;
8628 power >>= SCHED_LOAD_SHIFT;
8630 sd->groups->cpu_power += power;
8635 * Add cpu_power of each child group to this groups cpu_power.
8637 group = child->groups;
8639 sd->groups->cpu_power += group->cpu_power;
8640 group = group->next;
8641 } while (group != child->groups);
8645 * Initializers for schedule domains
8646 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8649 #ifdef CONFIG_SCHED_DEBUG
8650 # define SD_INIT_NAME(sd, type) sd->name = #type
8652 # define SD_INIT_NAME(sd, type) do { } while (0)
8655 #define SD_INIT(sd, type) sd_init_##type(sd)
8657 #define SD_INIT_FUNC(type) \
8658 static noinline void sd_init_##type(struct sched_domain *sd) \
8660 memset(sd, 0, sizeof(*sd)); \
8661 *sd = SD_##type##_INIT; \
8662 sd->level = SD_LV_##type; \
8663 SD_INIT_NAME(sd, type); \
8668 SD_INIT_FUNC(ALLNODES)
8671 #ifdef CONFIG_SCHED_SMT
8672 SD_INIT_FUNC(SIBLING)
8674 #ifdef CONFIG_SCHED_MC
8678 static int default_relax_domain_level = -1;
8680 static int __init setup_relax_domain_level(char *str)
8684 val = simple_strtoul(str, NULL, 0);
8685 if (val < SD_LV_MAX)
8686 default_relax_domain_level = val;
8690 __setup("relax_domain_level=", setup_relax_domain_level);
8692 static void set_domain_attribute(struct sched_domain *sd,
8693 struct sched_domain_attr *attr)
8697 if (!attr || attr->relax_domain_level < 0) {
8698 if (default_relax_domain_level < 0)
8701 request = default_relax_domain_level;
8703 request = attr->relax_domain_level;
8704 if (request < sd->level) {
8705 /* turn off idle balance on this domain */
8706 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8708 /* turn on idle balance on this domain */
8709 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8713 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8714 const struct cpumask *cpu_map)
8717 case sa_sched_groups:
8718 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8719 d->sched_group_nodes = NULL;
8721 free_rootdomain(d->rd); /* fall through */
8723 free_cpumask_var(d->tmpmask); /* fall through */
8724 case sa_send_covered:
8725 free_cpumask_var(d->send_covered); /* fall through */
8726 case sa_this_core_map:
8727 free_cpumask_var(d->this_core_map); /* fall through */
8728 case sa_this_sibling_map:
8729 free_cpumask_var(d->this_sibling_map); /* fall through */
8731 free_cpumask_var(d->nodemask); /* fall through */
8732 case sa_sched_group_nodes:
8734 kfree(d->sched_group_nodes); /* fall through */
8736 free_cpumask_var(d->notcovered); /* fall through */
8738 free_cpumask_var(d->covered); /* fall through */
8740 free_cpumask_var(d->domainspan); /* fall through */
8747 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8748 const struct cpumask *cpu_map)
8751 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8753 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8754 return sa_domainspan;
8755 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8757 /* Allocate the per-node list of sched groups */
8758 d->sched_group_nodes = kcalloc(nr_node_ids,
8759 sizeof(struct sched_group *), GFP_KERNEL);
8760 if (!d->sched_group_nodes) {
8761 printk(KERN_WARNING "Can not alloc sched group node list\n");
8762 return sa_notcovered;
8764 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8766 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8767 return sa_sched_group_nodes;
8768 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8770 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8771 return sa_this_sibling_map;
8772 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8773 return sa_this_core_map;
8774 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8775 return sa_send_covered;
8776 d->rd = alloc_rootdomain();
8778 printk(KERN_WARNING "Cannot alloc root domain\n");
8781 return sa_rootdomain;
8784 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8785 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8787 struct sched_domain *sd = NULL;
8789 struct sched_domain *parent;
8792 if (cpumask_weight(cpu_map) >
8793 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8794 sd = &per_cpu(allnodes_domains, i).sd;
8795 SD_INIT(sd, ALLNODES);
8796 set_domain_attribute(sd, attr);
8797 cpumask_copy(sched_domain_span(sd), cpu_map);
8798 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8803 sd = &per_cpu(node_domains, i).sd;
8805 set_domain_attribute(sd, attr);
8806 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8807 sd->parent = parent;
8810 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8815 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8816 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8817 struct sched_domain *parent, int i)
8819 struct sched_domain *sd;
8820 sd = &per_cpu(phys_domains, i).sd;
8822 set_domain_attribute(sd, attr);
8823 cpumask_copy(sched_domain_span(sd), d->nodemask);
8824 sd->parent = parent;
8827 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8831 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8832 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8833 struct sched_domain *parent, int i)
8835 struct sched_domain *sd = parent;
8836 #ifdef CONFIG_SCHED_MC
8837 sd = &per_cpu(core_domains, i).sd;
8839 set_domain_attribute(sd, attr);
8840 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8841 sd->parent = parent;
8843 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8848 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8849 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8850 struct sched_domain *parent, int i)
8852 struct sched_domain *sd = parent;
8853 #ifdef CONFIG_SCHED_SMT
8854 sd = &per_cpu(cpu_domains, i).sd;
8855 SD_INIT(sd, SIBLING);
8856 set_domain_attribute(sd, attr);
8857 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8858 sd->parent = parent;
8860 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8865 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8866 const struct cpumask *cpu_map, int cpu)
8869 #ifdef CONFIG_SCHED_SMT
8870 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8871 cpumask_and(d->this_sibling_map, cpu_map,
8872 topology_thread_cpumask(cpu));
8873 if (cpu == cpumask_first(d->this_sibling_map))
8874 init_sched_build_groups(d->this_sibling_map, cpu_map,
8876 d->send_covered, d->tmpmask);
8879 #ifdef CONFIG_SCHED_MC
8880 case SD_LV_MC: /* set up multi-core groups */
8881 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8882 if (cpu == cpumask_first(d->this_core_map))
8883 init_sched_build_groups(d->this_core_map, cpu_map,
8885 d->send_covered, d->tmpmask);
8888 case SD_LV_CPU: /* set up physical groups */
8889 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8890 if (!cpumask_empty(d->nodemask))
8891 init_sched_build_groups(d->nodemask, cpu_map,
8893 d->send_covered, d->tmpmask);
8896 case SD_LV_ALLNODES:
8897 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8898 d->send_covered, d->tmpmask);
8907 * Build sched domains for a given set of cpus and attach the sched domains
8908 * to the individual cpus
8910 static int __build_sched_domains(const struct cpumask *cpu_map,
8911 struct sched_domain_attr *attr)
8913 enum s_alloc alloc_state = sa_none;
8915 struct sched_domain *sd;
8921 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8922 if (alloc_state != sa_rootdomain)
8924 alloc_state = sa_sched_groups;
8927 * Set up domains for cpus specified by the cpu_map.
8929 for_each_cpu(i, cpu_map) {
8930 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8933 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8934 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8935 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8936 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8939 for_each_cpu(i, cpu_map) {
8940 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8941 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8944 /* Set up physical groups */
8945 for (i = 0; i < nr_node_ids; i++)
8946 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8949 /* Set up node groups */
8951 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8953 for (i = 0; i < nr_node_ids; i++)
8954 if (build_numa_sched_groups(&d, cpu_map, i))
8958 /* Calculate CPU power for physical packages and nodes */
8959 #ifdef CONFIG_SCHED_SMT
8960 for_each_cpu(i, cpu_map) {
8961 sd = &per_cpu(cpu_domains, i).sd;
8962 init_sched_groups_power(i, sd);
8965 #ifdef CONFIG_SCHED_MC
8966 for_each_cpu(i, cpu_map) {
8967 sd = &per_cpu(core_domains, i).sd;
8968 init_sched_groups_power(i, sd);
8972 for_each_cpu(i, cpu_map) {
8973 sd = &per_cpu(phys_domains, i).sd;
8974 init_sched_groups_power(i, sd);
8978 for (i = 0; i < nr_node_ids; i++)
8979 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8981 if (d.sd_allnodes) {
8982 struct sched_group *sg;
8984 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8986 init_numa_sched_groups_power(sg);
8990 /* Attach the domains */
8991 for_each_cpu(i, cpu_map) {
8992 #ifdef CONFIG_SCHED_SMT
8993 sd = &per_cpu(cpu_domains, i).sd;
8994 #elif defined(CONFIG_SCHED_MC)
8995 sd = &per_cpu(core_domains, i).sd;
8997 sd = &per_cpu(phys_domains, i).sd;
8999 cpu_attach_domain(sd, d.rd, i);
9002 d.sched_group_nodes = NULL; /* don't free this we still need it */
9003 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
9007 __free_domain_allocs(&d, alloc_state, cpu_map);
9011 static int build_sched_domains(const struct cpumask *cpu_map)
9013 return __build_sched_domains(cpu_map, NULL);
9016 static struct cpumask *doms_cur; /* current sched domains */
9017 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
9018 static struct sched_domain_attr *dattr_cur;
9019 /* attribues of custom domains in 'doms_cur' */
9022 * Special case: If a kmalloc of a doms_cur partition (array of
9023 * cpumask) fails, then fallback to a single sched domain,
9024 * as determined by the single cpumask fallback_doms.
9026 static cpumask_var_t fallback_doms;
9029 * arch_update_cpu_topology lets virtualized architectures update the
9030 * cpu core maps. It is supposed to return 1 if the topology changed
9031 * or 0 if it stayed the same.
9033 int __attribute__((weak)) arch_update_cpu_topology(void)
9039 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9040 * For now this just excludes isolated cpus, but could be used to
9041 * exclude other special cases in the future.
9043 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9047 arch_update_cpu_topology();
9049 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
9051 doms_cur = fallback_doms;
9052 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
9054 err = build_sched_domains(doms_cur);
9055 register_sched_domain_sysctl();
9060 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9061 struct cpumask *tmpmask)
9063 free_sched_groups(cpu_map, tmpmask);
9067 * Detach sched domains from a group of cpus specified in cpu_map
9068 * These cpus will now be attached to the NULL domain
9070 static void detach_destroy_domains(const struct cpumask *cpu_map)
9072 /* Save because hotplug lock held. */
9073 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9076 for_each_cpu(i, cpu_map)
9077 cpu_attach_domain(NULL, &def_root_domain, i);
9078 synchronize_sched();
9079 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9082 /* handle null as "default" */
9083 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9084 struct sched_domain_attr *new, int idx_new)
9086 struct sched_domain_attr tmp;
9093 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9094 new ? (new + idx_new) : &tmp,
9095 sizeof(struct sched_domain_attr));
9099 * Partition sched domains as specified by the 'ndoms_new'
9100 * cpumasks in the array doms_new[] of cpumasks. This compares
9101 * doms_new[] to the current sched domain partitioning, doms_cur[].
9102 * It destroys each deleted domain and builds each new domain.
9104 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9105 * The masks don't intersect (don't overlap.) We should setup one
9106 * sched domain for each mask. CPUs not in any of the cpumasks will
9107 * not be load balanced. If the same cpumask appears both in the
9108 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9111 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9112 * ownership of it and will kfree it when done with it. If the caller
9113 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9114 * ndoms_new == 1, and partition_sched_domains() will fallback to
9115 * the single partition 'fallback_doms', it also forces the domains
9118 * If doms_new == NULL it will be replaced with cpu_online_mask.
9119 * ndoms_new == 0 is a special case for destroying existing domains,
9120 * and it will not create the default domain.
9122 * Call with hotplug lock held
9124 /* FIXME: Change to struct cpumask *doms_new[] */
9125 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
9126 struct sched_domain_attr *dattr_new)
9131 mutex_lock(&sched_domains_mutex);
9133 /* always unregister in case we don't destroy any domains */
9134 unregister_sched_domain_sysctl();
9136 /* Let architecture update cpu core mappings. */
9137 new_topology = arch_update_cpu_topology();
9139 n = doms_new ? ndoms_new : 0;
9141 /* Destroy deleted domains */
9142 for (i = 0; i < ndoms_cur; i++) {
9143 for (j = 0; j < n && !new_topology; j++) {
9144 if (cpumask_equal(&doms_cur[i], &doms_new[j])
9145 && dattrs_equal(dattr_cur, i, dattr_new, j))
9148 /* no match - a current sched domain not in new doms_new[] */
9149 detach_destroy_domains(doms_cur + i);
9154 if (doms_new == NULL) {
9156 doms_new = fallback_doms;
9157 cpumask_andnot(&doms_new[0], cpu_active_mask, cpu_isolated_map);
9158 WARN_ON_ONCE(dattr_new);
9161 /* Build new domains */
9162 for (i = 0; i < ndoms_new; i++) {
9163 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9164 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9165 && dattrs_equal(dattr_new, i, dattr_cur, j))
9168 /* no match - add a new doms_new */
9169 __build_sched_domains(doms_new + i,
9170 dattr_new ? dattr_new + i : NULL);
9175 /* Remember the new sched domains */
9176 if (doms_cur != fallback_doms)
9178 kfree(dattr_cur); /* kfree(NULL) is safe */
9179 doms_cur = doms_new;
9180 dattr_cur = dattr_new;
9181 ndoms_cur = ndoms_new;
9183 register_sched_domain_sysctl();
9185 mutex_unlock(&sched_domains_mutex);
9188 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9189 static void arch_reinit_sched_domains(void)
9193 /* Destroy domains first to force the rebuild */
9194 partition_sched_domains(0, NULL, NULL);
9196 rebuild_sched_domains();
9200 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9202 unsigned int level = 0;
9204 if (sscanf(buf, "%u", &level) != 1)
9208 * level is always be positive so don't check for
9209 * level < POWERSAVINGS_BALANCE_NONE which is 0
9210 * What happens on 0 or 1 byte write,
9211 * need to check for count as well?
9214 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9218 sched_smt_power_savings = level;
9220 sched_mc_power_savings = level;
9222 arch_reinit_sched_domains();
9227 #ifdef CONFIG_SCHED_MC
9228 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9231 return sprintf(page, "%u\n", sched_mc_power_savings);
9233 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9234 const char *buf, size_t count)
9236 return sched_power_savings_store(buf, count, 0);
9238 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9239 sched_mc_power_savings_show,
9240 sched_mc_power_savings_store);
9243 #ifdef CONFIG_SCHED_SMT
9244 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9247 return sprintf(page, "%u\n", sched_smt_power_savings);
9249 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9250 const char *buf, size_t count)
9252 return sched_power_savings_store(buf, count, 1);
9254 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9255 sched_smt_power_savings_show,
9256 sched_smt_power_savings_store);
9259 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9263 #ifdef CONFIG_SCHED_SMT
9265 err = sysfs_create_file(&cls->kset.kobj,
9266 &attr_sched_smt_power_savings.attr);
9268 #ifdef CONFIG_SCHED_MC
9269 if (!err && mc_capable())
9270 err = sysfs_create_file(&cls->kset.kobj,
9271 &attr_sched_mc_power_savings.attr);
9275 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9277 #ifndef CONFIG_CPUSETS
9279 * Add online and remove offline CPUs from the scheduler domains.
9280 * When cpusets are enabled they take over this function.
9282 static int update_sched_domains(struct notifier_block *nfb,
9283 unsigned long action, void *hcpu)
9287 case CPU_ONLINE_FROZEN:
9288 case CPU_DOWN_PREPARE:
9289 case CPU_DOWN_PREPARE_FROZEN:
9290 case CPU_DOWN_FAILED:
9291 case CPU_DOWN_FAILED_FROZEN:
9292 partition_sched_domains(1, NULL, NULL);
9301 static int update_runtime(struct notifier_block *nfb,
9302 unsigned long action, void *hcpu)
9304 int cpu = (int)(long)hcpu;
9307 case CPU_DOWN_PREPARE:
9308 case CPU_DOWN_PREPARE_FROZEN:
9309 disable_runtime(cpu_rq(cpu));
9312 case CPU_DOWN_FAILED:
9313 case CPU_DOWN_FAILED_FROZEN:
9315 case CPU_ONLINE_FROZEN:
9316 enable_runtime(cpu_rq(cpu));
9324 void __init sched_init_smp(void)
9326 cpumask_var_t non_isolated_cpus;
9328 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9329 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9331 #if defined(CONFIG_NUMA)
9332 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9334 BUG_ON(sched_group_nodes_bycpu == NULL);
9337 mutex_lock(&sched_domains_mutex);
9338 arch_init_sched_domains(cpu_active_mask);
9339 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9340 if (cpumask_empty(non_isolated_cpus))
9341 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9342 mutex_unlock(&sched_domains_mutex);
9345 #ifndef CONFIG_CPUSETS
9346 /* XXX: Theoretical race here - CPU may be hotplugged now */
9347 hotcpu_notifier(update_sched_domains, 0);
9350 /* RT runtime code needs to handle some hotplug events */
9351 hotcpu_notifier(update_runtime, 0);
9355 /* Move init over to a non-isolated CPU */
9356 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9358 sched_init_granularity();
9359 free_cpumask_var(non_isolated_cpus);
9361 init_sched_rt_class();
9364 void __init sched_init_smp(void)
9366 sched_init_granularity();
9368 #endif /* CONFIG_SMP */
9370 const_debug unsigned int sysctl_timer_migration = 1;
9372 int in_sched_functions(unsigned long addr)
9374 return in_lock_functions(addr) ||
9375 (addr >= (unsigned long)__sched_text_start
9376 && addr < (unsigned long)__sched_text_end);
9379 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9381 cfs_rq->tasks_timeline = RB_ROOT;
9382 INIT_LIST_HEAD(&cfs_rq->tasks);
9383 #ifdef CONFIG_FAIR_GROUP_SCHED
9386 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9389 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9391 struct rt_prio_array *array;
9394 array = &rt_rq->active;
9395 for (i = 0; i < MAX_RT_PRIO; i++) {
9396 INIT_LIST_HEAD(array->queue + i);
9397 __clear_bit(i, array->bitmap);
9399 /* delimiter for bitsearch: */
9400 __set_bit(MAX_RT_PRIO, array->bitmap);
9402 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9403 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9405 rt_rq->highest_prio.next = MAX_RT_PRIO;
9409 rt_rq->rt_nr_migratory = 0;
9410 rt_rq->overloaded = 0;
9411 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9415 rt_rq->rt_throttled = 0;
9416 rt_rq->rt_runtime = 0;
9417 spin_lock_init(&rt_rq->rt_runtime_lock);
9419 #ifdef CONFIG_RT_GROUP_SCHED
9420 rt_rq->rt_nr_boosted = 0;
9425 #ifdef CONFIG_FAIR_GROUP_SCHED
9426 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9427 struct sched_entity *se, int cpu, int add,
9428 struct sched_entity *parent)
9430 struct rq *rq = cpu_rq(cpu);
9431 tg->cfs_rq[cpu] = cfs_rq;
9432 init_cfs_rq(cfs_rq, rq);
9435 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9438 /* se could be NULL for init_task_group */
9443 se->cfs_rq = &rq->cfs;
9445 se->cfs_rq = parent->my_q;
9448 se->load.weight = tg->shares;
9449 se->load.inv_weight = 0;
9450 se->parent = parent;
9454 #ifdef CONFIG_RT_GROUP_SCHED
9455 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9456 struct sched_rt_entity *rt_se, int cpu, int add,
9457 struct sched_rt_entity *parent)
9459 struct rq *rq = cpu_rq(cpu);
9461 tg->rt_rq[cpu] = rt_rq;
9462 init_rt_rq(rt_rq, rq);
9464 rt_rq->rt_se = rt_se;
9465 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9467 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9469 tg->rt_se[cpu] = rt_se;
9474 rt_se->rt_rq = &rq->rt;
9476 rt_se->rt_rq = parent->my_q;
9478 rt_se->my_q = rt_rq;
9479 rt_se->parent = parent;
9480 INIT_LIST_HEAD(&rt_se->run_list);
9484 void __init sched_init(void)
9487 unsigned long alloc_size = 0, ptr;
9489 #ifdef CONFIG_FAIR_GROUP_SCHED
9490 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9492 #ifdef CONFIG_RT_GROUP_SCHED
9493 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9495 #ifdef CONFIG_CPUMASK_OFFSTACK
9496 alloc_size += num_possible_cpus() * cpumask_size();
9499 * As sched_init() is called before page_alloc is setup,
9500 * we use alloc_bootmem().
9503 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9505 #ifdef CONFIG_FAIR_GROUP_SCHED
9506 init_task_group.se = (struct sched_entity **)ptr;
9507 ptr += nr_cpu_ids * sizeof(void **);
9509 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9510 ptr += nr_cpu_ids * sizeof(void **);
9512 #endif /* CONFIG_FAIR_GROUP_SCHED */
9513 #ifdef CONFIG_RT_GROUP_SCHED
9514 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9515 ptr += nr_cpu_ids * sizeof(void **);
9517 init_task_group.rt_rq = (struct rt_rq **)ptr;
9518 ptr += nr_cpu_ids * sizeof(void **);
9520 #endif /* CONFIG_RT_GROUP_SCHED */
9521 #ifdef CONFIG_CPUMASK_OFFSTACK
9522 for_each_possible_cpu(i) {
9523 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9524 ptr += cpumask_size();
9526 #endif /* CONFIG_CPUMASK_OFFSTACK */
9530 init_defrootdomain();
9533 init_rt_bandwidth(&def_rt_bandwidth,
9534 global_rt_period(), global_rt_runtime());
9536 #ifdef CONFIG_RT_GROUP_SCHED
9537 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9538 global_rt_period(), global_rt_runtime());
9539 #endif /* CONFIG_RT_GROUP_SCHED */
9541 #ifdef CONFIG_CGROUP_SCHED
9542 list_add(&init_task_group.list, &task_groups);
9543 INIT_LIST_HEAD(&init_task_group.children);
9545 #endif /* CONFIG_CGROUP_SCHED */
9547 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9548 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9549 __alignof__(unsigned long));
9551 for_each_possible_cpu(i) {
9555 spin_lock_init(&rq->lock);
9557 rq->calc_load_active = 0;
9558 rq->calc_load_update = jiffies + LOAD_FREQ;
9559 init_cfs_rq(&rq->cfs, rq);
9560 init_rt_rq(&rq->rt, rq);
9561 #ifdef CONFIG_FAIR_GROUP_SCHED
9562 init_task_group.shares = init_task_group_load;
9563 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9564 #ifdef CONFIG_CGROUP_SCHED
9566 * How much cpu bandwidth does init_task_group get?
9568 * In case of task-groups formed thr' the cgroup filesystem, it
9569 * gets 100% of the cpu resources in the system. This overall
9570 * system cpu resource is divided among the tasks of
9571 * init_task_group and its child task-groups in a fair manner,
9572 * based on each entity's (task or task-group's) weight
9573 * (se->load.weight).
9575 * In other words, if init_task_group has 10 tasks of weight
9576 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9577 * then A0's share of the cpu resource is:
9579 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9581 * We achieve this by letting init_task_group's tasks sit
9582 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9584 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9586 #endif /* CONFIG_FAIR_GROUP_SCHED */
9588 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9589 #ifdef CONFIG_RT_GROUP_SCHED
9590 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9591 #ifdef CONFIG_CGROUP_SCHED
9592 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9593 #elif defined CONFIG_USER_SCHED
9594 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9595 init_tg_rt_entry(&init_task_group,
9596 &per_cpu(init_rt_rq, i),
9597 &per_cpu(init_sched_rt_entity, i), i, 1,
9598 root_task_group.rt_se[i]);
9602 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9603 rq->cpu_load[j] = 0;
9607 rq->post_schedule = 0;
9608 rq->active_balance = 0;
9609 rq->next_balance = jiffies;
9613 rq->migration_thread = NULL;
9615 rq->avg_idle = 2*sysctl_sched_migration_cost;
9616 INIT_LIST_HEAD(&rq->migration_queue);
9617 rq_attach_root(rq, &def_root_domain);
9620 atomic_set(&rq->nr_iowait, 0);
9623 set_load_weight(&init_task);
9625 #ifdef CONFIG_PREEMPT_NOTIFIERS
9626 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9630 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9633 #ifdef CONFIG_RT_MUTEXES
9634 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9638 * The boot idle thread does lazy MMU switching as well:
9640 atomic_inc(&init_mm.mm_count);
9641 enter_lazy_tlb(&init_mm, current);
9644 * Make us the idle thread. Technically, schedule() should not be
9645 * called from this thread, however somewhere below it might be,
9646 * but because we are the idle thread, we just pick up running again
9647 * when this runqueue becomes "idle".
9649 init_idle(current, smp_processor_id());
9651 calc_load_update = jiffies + LOAD_FREQ;
9654 * During early bootup we pretend to be a normal task:
9656 current->sched_class = &fair_sched_class;
9658 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9659 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9662 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9663 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9665 /* May be allocated at isolcpus cmdline parse time */
9666 if (cpu_isolated_map == NULL)
9667 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9672 scheduler_running = 1;
9675 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9676 static inline int preempt_count_equals(int preempt_offset)
9678 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9680 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9683 void __might_sleep(char *file, int line, int preempt_offset)
9686 static unsigned long prev_jiffy; /* ratelimiting */
9688 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9689 system_state != SYSTEM_RUNNING || oops_in_progress)
9691 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9693 prev_jiffy = jiffies;
9696 "BUG: sleeping function called from invalid context at %s:%d\n",
9699 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9700 in_atomic(), irqs_disabled(),
9701 current->pid, current->comm);
9703 debug_show_held_locks(current);
9704 if (irqs_disabled())
9705 print_irqtrace_events(current);
9709 EXPORT_SYMBOL(__might_sleep);
9712 #ifdef CONFIG_MAGIC_SYSRQ
9713 static void normalize_task(struct rq *rq, struct task_struct *p)
9717 update_rq_clock(rq);
9718 on_rq = p->se.on_rq;
9720 deactivate_task(rq, p, 0);
9721 __setscheduler(rq, p, SCHED_NORMAL, 0);
9723 activate_task(rq, p, 0);
9724 resched_task(rq->curr);
9728 void normalize_rt_tasks(void)
9730 struct task_struct *g, *p;
9731 unsigned long flags;
9734 read_lock_irqsave(&tasklist_lock, flags);
9735 do_each_thread(g, p) {
9737 * Only normalize user tasks:
9742 p->se.exec_start = 0;
9743 #ifdef CONFIG_SCHEDSTATS
9744 p->se.wait_start = 0;
9745 p->se.sleep_start = 0;
9746 p->se.block_start = 0;
9751 * Renice negative nice level userspace
9754 if (TASK_NICE(p) < 0 && p->mm)
9755 set_user_nice(p, 0);
9759 spin_lock(&p->pi_lock);
9760 rq = __task_rq_lock(p);
9762 normalize_task(rq, p);
9764 __task_rq_unlock(rq);
9765 spin_unlock(&p->pi_lock);
9766 } while_each_thread(g, p);
9768 read_unlock_irqrestore(&tasklist_lock, flags);
9771 #endif /* CONFIG_MAGIC_SYSRQ */
9775 * These functions are only useful for the IA64 MCA handling.
9777 * They can only be called when the whole system has been
9778 * stopped - every CPU needs to be quiescent, and no scheduling
9779 * activity can take place. Using them for anything else would
9780 * be a serious bug, and as a result, they aren't even visible
9781 * under any other configuration.
9785 * curr_task - return the current task for a given cpu.
9786 * @cpu: the processor in question.
9788 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9790 struct task_struct *curr_task(int cpu)
9792 return cpu_curr(cpu);
9796 * set_curr_task - set the current task for a given cpu.
9797 * @cpu: the processor in question.
9798 * @p: the task pointer to set.
9800 * Description: This function must only be used when non-maskable interrupts
9801 * are serviced on a separate stack. It allows the architecture to switch the
9802 * notion of the current task on a cpu in a non-blocking manner. This function
9803 * must be called with all CPU's synchronized, and interrupts disabled, the
9804 * and caller must save the original value of the current task (see
9805 * curr_task() above) and restore that value before reenabling interrupts and
9806 * re-starting the system.
9808 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9810 void set_curr_task(int cpu, struct task_struct *p)
9817 #ifdef CONFIG_FAIR_GROUP_SCHED
9818 static void free_fair_sched_group(struct task_group *tg)
9822 for_each_possible_cpu(i) {
9824 kfree(tg->cfs_rq[i]);
9834 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9836 struct cfs_rq *cfs_rq;
9837 struct sched_entity *se;
9841 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9844 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9848 tg->shares = NICE_0_LOAD;
9850 for_each_possible_cpu(i) {
9853 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9854 GFP_KERNEL, cpu_to_node(i));
9858 se = kzalloc_node(sizeof(struct sched_entity),
9859 GFP_KERNEL, cpu_to_node(i));
9863 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9872 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9874 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9875 &cpu_rq(cpu)->leaf_cfs_rq_list);
9878 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9880 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9882 #else /* !CONFG_FAIR_GROUP_SCHED */
9883 static inline void free_fair_sched_group(struct task_group *tg)
9888 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9893 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9897 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9900 #endif /* CONFIG_FAIR_GROUP_SCHED */
9902 #ifdef CONFIG_RT_GROUP_SCHED
9903 static void free_rt_sched_group(struct task_group *tg)
9907 destroy_rt_bandwidth(&tg->rt_bandwidth);
9909 for_each_possible_cpu(i) {
9911 kfree(tg->rt_rq[i]);
9913 kfree(tg->rt_se[i]);
9921 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9923 struct rt_rq *rt_rq;
9924 struct sched_rt_entity *rt_se;
9928 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9931 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9935 init_rt_bandwidth(&tg->rt_bandwidth,
9936 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9938 for_each_possible_cpu(i) {
9941 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9942 GFP_KERNEL, cpu_to_node(i));
9946 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9947 GFP_KERNEL, cpu_to_node(i));
9951 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9960 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9962 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9963 &cpu_rq(cpu)->leaf_rt_rq_list);
9966 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9968 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9970 #else /* !CONFIG_RT_GROUP_SCHED */
9971 static inline void free_rt_sched_group(struct task_group *tg)
9976 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9981 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9985 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9988 #endif /* CONFIG_RT_GROUP_SCHED */
9990 #ifdef CONFIG_CGROUP_SCHED
9991 static void free_sched_group(struct task_group *tg)
9993 free_fair_sched_group(tg);
9994 free_rt_sched_group(tg);
9998 /* allocate runqueue etc for a new task group */
9999 struct task_group *sched_create_group(struct task_group *parent)
10001 struct task_group *tg;
10002 unsigned long flags;
10005 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10007 return ERR_PTR(-ENOMEM);
10009 if (!alloc_fair_sched_group(tg, parent))
10012 if (!alloc_rt_sched_group(tg, parent))
10015 spin_lock_irqsave(&task_group_lock, flags);
10016 for_each_possible_cpu(i) {
10017 register_fair_sched_group(tg, i);
10018 register_rt_sched_group(tg, i);
10020 list_add_rcu(&tg->list, &task_groups);
10022 WARN_ON(!parent); /* root should already exist */
10024 tg->parent = parent;
10025 INIT_LIST_HEAD(&tg->children);
10026 list_add_rcu(&tg->siblings, &parent->children);
10027 spin_unlock_irqrestore(&task_group_lock, flags);
10032 free_sched_group(tg);
10033 return ERR_PTR(-ENOMEM);
10036 /* rcu callback to free various structures associated with a task group */
10037 static void free_sched_group_rcu(struct rcu_head *rhp)
10039 /* now it should be safe to free those cfs_rqs */
10040 free_sched_group(container_of(rhp, struct task_group, rcu));
10043 /* Destroy runqueue etc associated with a task group */
10044 void sched_destroy_group(struct task_group *tg)
10046 unsigned long flags;
10049 spin_lock_irqsave(&task_group_lock, flags);
10050 for_each_possible_cpu(i) {
10051 unregister_fair_sched_group(tg, i);
10052 unregister_rt_sched_group(tg, i);
10054 list_del_rcu(&tg->list);
10055 list_del_rcu(&tg->siblings);
10056 spin_unlock_irqrestore(&task_group_lock, flags);
10058 /* wait for possible concurrent references to cfs_rqs complete */
10059 call_rcu(&tg->rcu, free_sched_group_rcu);
10062 /* change task's runqueue when it moves between groups.
10063 * The caller of this function should have put the task in its new group
10064 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10065 * reflect its new group.
10067 void sched_move_task(struct task_struct *tsk)
10069 int on_rq, running;
10070 unsigned long flags;
10073 rq = task_rq_lock(tsk, &flags);
10075 update_rq_clock(rq);
10077 running = task_current(rq, tsk);
10078 on_rq = tsk->se.on_rq;
10081 dequeue_task(rq, tsk, 0);
10082 if (unlikely(running))
10083 tsk->sched_class->put_prev_task(rq, tsk);
10085 set_task_rq(tsk, task_cpu(tsk));
10087 #ifdef CONFIG_FAIR_GROUP_SCHED
10088 if (tsk->sched_class->moved_group)
10089 tsk->sched_class->moved_group(tsk, on_rq);
10092 if (unlikely(running))
10093 tsk->sched_class->set_curr_task(rq);
10095 enqueue_task(rq, tsk, 0, false);
10097 task_rq_unlock(rq, &flags);
10099 #endif /* CONFIG_CGROUP_SCHED */
10101 #ifdef CONFIG_FAIR_GROUP_SCHED
10102 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10104 struct cfs_rq *cfs_rq = se->cfs_rq;
10109 dequeue_entity(cfs_rq, se, 0);
10111 se->load.weight = shares;
10112 se->load.inv_weight = 0;
10115 enqueue_entity(cfs_rq, se, 0);
10118 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10120 struct cfs_rq *cfs_rq = se->cfs_rq;
10121 struct rq *rq = cfs_rq->rq;
10122 unsigned long flags;
10124 spin_lock_irqsave(&rq->lock, flags);
10125 __set_se_shares(se, shares);
10126 spin_unlock_irqrestore(&rq->lock, flags);
10129 static DEFINE_MUTEX(shares_mutex);
10131 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10134 unsigned long flags;
10137 * We can't change the weight of the root cgroup.
10142 if (shares < MIN_SHARES)
10143 shares = MIN_SHARES;
10144 else if (shares > MAX_SHARES)
10145 shares = MAX_SHARES;
10147 mutex_lock(&shares_mutex);
10148 if (tg->shares == shares)
10151 spin_lock_irqsave(&task_group_lock, flags);
10152 for_each_possible_cpu(i)
10153 unregister_fair_sched_group(tg, i);
10154 list_del_rcu(&tg->siblings);
10155 spin_unlock_irqrestore(&task_group_lock, flags);
10157 /* wait for any ongoing reference to this group to finish */
10158 synchronize_sched();
10161 * Now we are free to modify the group's share on each cpu
10162 * w/o tripping rebalance_share or load_balance_fair.
10164 tg->shares = shares;
10165 for_each_possible_cpu(i) {
10167 * force a rebalance
10169 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10170 set_se_shares(tg->se[i], shares);
10174 * Enable load balance activity on this group, by inserting it back on
10175 * each cpu's rq->leaf_cfs_rq_list.
10177 spin_lock_irqsave(&task_group_lock, flags);
10178 for_each_possible_cpu(i)
10179 register_fair_sched_group(tg, i);
10180 list_add_rcu(&tg->siblings, &tg->parent->children);
10181 spin_unlock_irqrestore(&task_group_lock, flags);
10183 mutex_unlock(&shares_mutex);
10187 unsigned long sched_group_shares(struct task_group *tg)
10193 #ifdef CONFIG_RT_GROUP_SCHED
10195 * Ensure that the real time constraints are schedulable.
10197 static DEFINE_MUTEX(rt_constraints_mutex);
10199 static unsigned long to_ratio(u64 period, u64 runtime)
10201 if (runtime == RUNTIME_INF)
10204 return div64_u64(runtime << 20, period);
10207 /* Must be called with tasklist_lock held */
10208 static inline int tg_has_rt_tasks(struct task_group *tg)
10210 struct task_struct *g, *p;
10212 do_each_thread(g, p) {
10213 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10215 } while_each_thread(g, p);
10220 struct rt_schedulable_data {
10221 struct task_group *tg;
10226 static int tg_schedulable(struct task_group *tg, void *data)
10228 struct rt_schedulable_data *d = data;
10229 struct task_group *child;
10230 unsigned long total, sum = 0;
10231 u64 period, runtime;
10233 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10234 runtime = tg->rt_bandwidth.rt_runtime;
10237 period = d->rt_period;
10238 runtime = d->rt_runtime;
10242 * Cannot have more runtime than the period.
10244 if (runtime > period && runtime != RUNTIME_INF)
10248 * Ensure we don't starve existing RT tasks.
10250 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10253 total = to_ratio(period, runtime);
10256 * Nobody can have more than the global setting allows.
10258 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10262 * The sum of our children's runtime should not exceed our own.
10264 list_for_each_entry_rcu(child, &tg->children, siblings) {
10265 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10266 runtime = child->rt_bandwidth.rt_runtime;
10268 if (child == d->tg) {
10269 period = d->rt_period;
10270 runtime = d->rt_runtime;
10273 sum += to_ratio(period, runtime);
10282 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10284 struct rt_schedulable_data data = {
10286 .rt_period = period,
10287 .rt_runtime = runtime,
10290 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10293 static int tg_set_bandwidth(struct task_group *tg,
10294 u64 rt_period, u64 rt_runtime)
10298 mutex_lock(&rt_constraints_mutex);
10299 read_lock(&tasklist_lock);
10300 err = __rt_schedulable(tg, rt_period, rt_runtime);
10304 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10305 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10306 tg->rt_bandwidth.rt_runtime = rt_runtime;
10308 for_each_possible_cpu(i) {
10309 struct rt_rq *rt_rq = tg->rt_rq[i];
10311 spin_lock(&rt_rq->rt_runtime_lock);
10312 rt_rq->rt_runtime = rt_runtime;
10313 spin_unlock(&rt_rq->rt_runtime_lock);
10315 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10317 read_unlock(&tasklist_lock);
10318 mutex_unlock(&rt_constraints_mutex);
10323 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10325 u64 rt_runtime, rt_period;
10327 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10328 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10329 if (rt_runtime_us < 0)
10330 rt_runtime = RUNTIME_INF;
10332 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10335 long sched_group_rt_runtime(struct task_group *tg)
10339 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10342 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10343 do_div(rt_runtime_us, NSEC_PER_USEC);
10344 return rt_runtime_us;
10347 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10349 u64 rt_runtime, rt_period;
10351 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10352 rt_runtime = tg->rt_bandwidth.rt_runtime;
10354 if (rt_period == 0)
10357 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10360 long sched_group_rt_period(struct task_group *tg)
10364 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10365 do_div(rt_period_us, NSEC_PER_USEC);
10366 return rt_period_us;
10369 static int sched_rt_global_constraints(void)
10371 u64 runtime, period;
10374 if (sysctl_sched_rt_period <= 0)
10377 runtime = global_rt_runtime();
10378 period = global_rt_period();
10381 * Sanity check on the sysctl variables.
10383 if (runtime > period && runtime != RUNTIME_INF)
10386 mutex_lock(&rt_constraints_mutex);
10387 read_lock(&tasklist_lock);
10388 ret = __rt_schedulable(NULL, 0, 0);
10389 read_unlock(&tasklist_lock);
10390 mutex_unlock(&rt_constraints_mutex);
10395 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10397 /* Don't accept realtime tasks when there is no way for them to run */
10398 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10404 #else /* !CONFIG_RT_GROUP_SCHED */
10405 static int sched_rt_global_constraints(void)
10407 unsigned long flags;
10410 if (sysctl_sched_rt_period <= 0)
10414 * There's always some RT tasks in the root group
10415 * -- migration, kstopmachine etc..
10417 if (sysctl_sched_rt_runtime == 0)
10420 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10421 for_each_possible_cpu(i) {
10422 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10424 spin_lock(&rt_rq->rt_runtime_lock);
10425 rt_rq->rt_runtime = global_rt_runtime();
10426 spin_unlock(&rt_rq->rt_runtime_lock);
10428 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10432 #endif /* CONFIG_RT_GROUP_SCHED */
10434 int sched_rt_handler(struct ctl_table *table, int write,
10435 void __user *buffer, size_t *lenp,
10439 int old_period, old_runtime;
10440 static DEFINE_MUTEX(mutex);
10442 mutex_lock(&mutex);
10443 old_period = sysctl_sched_rt_period;
10444 old_runtime = sysctl_sched_rt_runtime;
10446 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10448 if (!ret && write) {
10449 ret = sched_rt_global_constraints();
10451 sysctl_sched_rt_period = old_period;
10452 sysctl_sched_rt_runtime = old_runtime;
10454 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10455 def_rt_bandwidth.rt_period =
10456 ns_to_ktime(global_rt_period());
10459 mutex_unlock(&mutex);
10464 #ifdef CONFIG_CGROUP_SCHED
10466 /* return corresponding task_group object of a cgroup */
10467 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10469 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10470 struct task_group, css);
10473 static struct cgroup_subsys_state *
10474 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10476 struct task_group *tg, *parent;
10478 if (!cgrp->parent) {
10479 /* This is early initialization for the top cgroup */
10480 return &init_task_group.css;
10483 parent = cgroup_tg(cgrp->parent);
10484 tg = sched_create_group(parent);
10486 return ERR_PTR(-ENOMEM);
10492 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10494 struct task_group *tg = cgroup_tg(cgrp);
10496 sched_destroy_group(tg);
10500 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10502 #ifdef CONFIG_RT_GROUP_SCHED
10503 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10506 /* We don't support RT-tasks being in separate groups */
10507 if (tsk->sched_class != &fair_sched_class)
10514 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10515 struct task_struct *tsk, bool threadgroup)
10517 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10521 struct task_struct *c;
10523 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10524 retval = cpu_cgroup_can_attach_task(cgrp, c);
10536 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10537 struct cgroup *old_cont, struct task_struct *tsk,
10540 sched_move_task(tsk);
10542 struct task_struct *c;
10544 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10545 sched_move_task(c);
10551 #ifdef CONFIG_FAIR_GROUP_SCHED
10552 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10555 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10558 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10560 struct task_group *tg = cgroup_tg(cgrp);
10562 return (u64) tg->shares;
10564 #endif /* CONFIG_FAIR_GROUP_SCHED */
10566 #ifdef CONFIG_RT_GROUP_SCHED
10567 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10570 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10573 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10575 return sched_group_rt_runtime(cgroup_tg(cgrp));
10578 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10581 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10584 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10586 return sched_group_rt_period(cgroup_tg(cgrp));
10588 #endif /* CONFIG_RT_GROUP_SCHED */
10590 static struct cftype cpu_files[] = {
10591 #ifdef CONFIG_FAIR_GROUP_SCHED
10594 .read_u64 = cpu_shares_read_u64,
10595 .write_u64 = cpu_shares_write_u64,
10598 #ifdef CONFIG_RT_GROUP_SCHED
10600 .name = "rt_runtime_us",
10601 .read_s64 = cpu_rt_runtime_read,
10602 .write_s64 = cpu_rt_runtime_write,
10605 .name = "rt_period_us",
10606 .read_u64 = cpu_rt_period_read_uint,
10607 .write_u64 = cpu_rt_period_write_uint,
10612 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10614 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10617 struct cgroup_subsys cpu_cgroup_subsys = {
10619 .create = cpu_cgroup_create,
10620 .destroy = cpu_cgroup_destroy,
10621 .can_attach = cpu_cgroup_can_attach,
10622 .attach = cpu_cgroup_attach,
10623 .populate = cpu_cgroup_populate,
10624 .subsys_id = cpu_cgroup_subsys_id,
10628 #endif /* CONFIG_CGROUP_SCHED */
10630 #ifdef CONFIG_CGROUP_CPUACCT
10633 * CPU accounting code for task groups.
10635 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10636 * (balbir@in.ibm.com).
10639 /* track cpu usage of a group of tasks and its child groups */
10641 struct cgroup_subsys_state css;
10642 /* cpuusage holds pointer to a u64-type object on every cpu */
10644 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10645 struct cpuacct *parent;
10648 struct cgroup_subsys cpuacct_subsys;
10650 /* return cpu accounting group corresponding to this container */
10651 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10653 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10654 struct cpuacct, css);
10657 /* return cpu accounting group to which this task belongs */
10658 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10660 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10661 struct cpuacct, css);
10664 /* create a new cpu accounting group */
10665 static struct cgroup_subsys_state *cpuacct_create(
10666 struct cgroup_subsys *ss, struct cgroup *cgrp)
10668 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10674 ca->cpuusage = alloc_percpu(u64);
10678 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10679 if (percpu_counter_init(&ca->cpustat[i], 0))
10680 goto out_free_counters;
10683 ca->parent = cgroup_ca(cgrp->parent);
10689 percpu_counter_destroy(&ca->cpustat[i]);
10690 free_percpu(ca->cpuusage);
10694 return ERR_PTR(-ENOMEM);
10697 /* destroy an existing cpu accounting group */
10699 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10701 struct cpuacct *ca = cgroup_ca(cgrp);
10704 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10705 percpu_counter_destroy(&ca->cpustat[i]);
10706 free_percpu(ca->cpuusage);
10710 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10712 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10715 #ifndef CONFIG_64BIT
10717 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10719 spin_lock_irq(&cpu_rq(cpu)->lock);
10721 spin_unlock_irq(&cpu_rq(cpu)->lock);
10729 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10731 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10733 #ifndef CONFIG_64BIT
10735 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10737 spin_lock_irq(&cpu_rq(cpu)->lock);
10739 spin_unlock_irq(&cpu_rq(cpu)->lock);
10745 /* return total cpu usage (in nanoseconds) of a group */
10746 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10748 struct cpuacct *ca = cgroup_ca(cgrp);
10749 u64 totalcpuusage = 0;
10752 for_each_present_cpu(i)
10753 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10755 return totalcpuusage;
10758 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10761 struct cpuacct *ca = cgroup_ca(cgrp);
10770 for_each_present_cpu(i)
10771 cpuacct_cpuusage_write(ca, i, 0);
10777 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10778 struct seq_file *m)
10780 struct cpuacct *ca = cgroup_ca(cgroup);
10784 for_each_present_cpu(i) {
10785 percpu = cpuacct_cpuusage_read(ca, i);
10786 seq_printf(m, "%llu ", (unsigned long long) percpu);
10788 seq_printf(m, "\n");
10792 static const char *cpuacct_stat_desc[] = {
10793 [CPUACCT_STAT_USER] = "user",
10794 [CPUACCT_STAT_SYSTEM] = "system",
10797 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10798 struct cgroup_map_cb *cb)
10800 struct cpuacct *ca = cgroup_ca(cgrp);
10803 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10804 s64 val = percpu_counter_read(&ca->cpustat[i]);
10805 val = cputime64_to_clock_t(val);
10806 cb->fill(cb, cpuacct_stat_desc[i], val);
10811 static struct cftype files[] = {
10814 .read_u64 = cpuusage_read,
10815 .write_u64 = cpuusage_write,
10818 .name = "usage_percpu",
10819 .read_seq_string = cpuacct_percpu_seq_read,
10823 .read_map = cpuacct_stats_show,
10827 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10829 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10833 * charge this task's execution time to its accounting group.
10835 * called with rq->lock held.
10837 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10839 struct cpuacct *ca;
10842 if (unlikely(!cpuacct_subsys.active))
10845 cpu = task_cpu(tsk);
10851 for (; ca; ca = ca->parent) {
10852 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10853 *cpuusage += cputime;
10860 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
10861 * in cputime_t units. As a result, cpuacct_update_stats calls
10862 * percpu_counter_add with values large enough to always overflow the
10863 * per cpu batch limit causing bad SMP scalability.
10865 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
10866 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
10867 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
10870 #define CPUACCT_BATCH \
10871 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
10873 #define CPUACCT_BATCH 0
10877 * Charge the system/user time to the task's accounting group.
10879 static void cpuacct_update_stats(struct task_struct *tsk,
10880 enum cpuacct_stat_index idx, cputime_t val)
10882 struct cpuacct *ca;
10883 int batch = CPUACCT_BATCH;
10885 if (unlikely(!cpuacct_subsys.active))
10892 __percpu_counter_add(&ca->cpustat[idx], val, batch);
10898 struct cgroup_subsys cpuacct_subsys = {
10900 .create = cpuacct_create,
10901 .destroy = cpuacct_destroy,
10902 .populate = cpuacct_populate,
10903 .subsys_id = cpuacct_subsys_id,
10905 #endif /* CONFIG_CGROUP_CPUACCT */
10909 int rcu_expedited_torture_stats(char *page)
10913 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10915 void synchronize_sched_expedited(void)
10918 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10920 #else /* #ifndef CONFIG_SMP */
10922 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10923 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10925 #define RCU_EXPEDITED_STATE_POST -2
10926 #define RCU_EXPEDITED_STATE_IDLE -1
10928 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10930 int rcu_expedited_torture_stats(char *page)
10935 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10936 for_each_online_cpu(cpu) {
10937 cnt += sprintf(&page[cnt], " %d:%d",
10938 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10940 cnt += sprintf(&page[cnt], "\n");
10943 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10945 static long synchronize_sched_expedited_count;
10948 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10949 * approach to force grace period to end quickly. This consumes
10950 * significant time on all CPUs, and is thus not recommended for
10951 * any sort of common-case code.
10953 * Note that it is illegal to call this function while holding any
10954 * lock that is acquired by a CPU-hotplug notifier. Failing to
10955 * observe this restriction will result in deadlock.
10957 void synchronize_sched_expedited(void)
10960 unsigned long flags;
10961 bool need_full_sync = 0;
10963 struct migration_req *req;
10967 smp_mb(); /* ensure prior mod happens before capturing snap. */
10968 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10970 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10972 if (trycount++ < 10)
10973 udelay(trycount * num_online_cpus());
10975 synchronize_sched();
10978 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10979 smp_mb(); /* ensure test happens before caller kfree */
10984 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10985 for_each_online_cpu(cpu) {
10987 req = &per_cpu(rcu_migration_req, cpu);
10988 init_completion(&req->done);
10990 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10991 spin_lock_irqsave(&rq->lock, flags);
10992 list_add(&req->list, &rq->migration_queue);
10993 spin_unlock_irqrestore(&rq->lock, flags);
10994 wake_up_process(rq->migration_thread);
10996 for_each_online_cpu(cpu) {
10997 rcu_expedited_state = cpu;
10998 req = &per_cpu(rcu_migration_req, cpu);
11000 wait_for_completion(&req->done);
11001 spin_lock_irqsave(&rq->lock, flags);
11002 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
11003 need_full_sync = 1;
11004 req->dest_cpu = RCU_MIGRATION_IDLE;
11005 spin_unlock_irqrestore(&rq->lock, flags);
11007 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11008 mutex_unlock(&rcu_sched_expedited_mutex);
11010 if (need_full_sync)
11011 synchronize_sched();
11013 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11015 #endif /* #else #ifndef CONFIG_SMP */