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_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct *user)
282 user->tg->uid = user->uid;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group.children);
321 #ifdef CONFIG_USER_SCHED
322 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
323 #else /* !CONFIG_USER_SCHED */
324 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 #endif /* CONFIG_USER_SCHED */
328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
336 #define MAX_SHARES (1UL << 18)
338 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
341 /* Default task group.
342 * Every task in system belong to this group at bootup.
344 struct task_group init_task_group;
346 /* return group to which a task belongs */
347 static inline struct task_group *task_group(struct task_struct *p)
349 struct task_group *tg;
351 #ifdef CONFIG_USER_SCHED
353 tg = __task_cred(p)->user->tg;
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
359 tg = &init_task_group;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
380 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
381 static inline struct task_group *task_group(struct task_struct *p)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load;
391 unsigned long nr_running;
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
399 struct list_head tasks;
400 struct list_head *balance_iterator;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity *curr, *next, *last;
408 unsigned int nr_spread_over;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load;
439 * this cpu's part of tg->shares
441 unsigned long shares;
444 * load.weight at the time we set shares
446 unsigned long rq_weight;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active;
454 unsigned long rt_nr_running;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
457 int curr; /* highest queued rt task prio */
459 int next; /* next highest */
464 unsigned long rt_nr_migratory;
465 unsigned long rt_nr_total;
467 struct plist_head pushable_tasks;
472 /* Nests inside the rq lock: */
473 spinlock_t rt_runtime_lock;
475 #ifdef CONFIG_RT_GROUP_SCHED
476 unsigned long rt_nr_boosted;
479 struct list_head leaf_rt_rq_list;
480 struct task_group *tg;
481 struct sched_rt_entity *rt_se;
488 * We add the notion of a root-domain which will be used to define per-domain
489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
491 * exclusive cpuset is created, we also create and attach a new root-domain
498 cpumask_var_t online;
501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
504 cpumask_var_t rto_mask;
507 struct cpupri cpupri;
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
515 static struct root_domain def_root_domain;
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned long nr_running;
535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
538 unsigned long last_tick_seen;
539 unsigned char in_nohz_recently;
541 /* capture load from *all* tasks on this cpu: */
542 struct load_weight load;
543 unsigned long nr_load_updates;
549 #ifdef CONFIG_FAIR_GROUP_SCHED
550 /* list of leaf cfs_rq on this cpu: */
551 struct list_head leaf_cfs_rq_list;
553 #ifdef CONFIG_RT_GROUP_SCHED
554 struct list_head leaf_rt_rq_list;
558 * This is part of a global counter where only the total sum
559 * over all CPUs matters. A task can increase this counter on
560 * one CPU and if it got migrated afterwards it may decrease
561 * it on another CPU. Always updated under the runqueue lock:
563 unsigned long nr_uninterruptible;
565 struct task_struct *curr, *idle;
566 unsigned long next_balance;
567 struct mm_struct *prev_mm;
574 struct root_domain *rd;
575 struct sched_domain *sd;
577 unsigned char idle_at_tick;
578 /* For active balancing */
582 /* cpu of this runqueue: */
586 unsigned long avg_load_per_task;
588 struct task_struct *migration_thread;
589 struct list_head migration_queue;
597 /* calc_load related fields */
598 unsigned long calc_load_update;
599 long calc_load_active;
601 #ifdef CONFIG_SCHED_HRTICK
603 int hrtick_csd_pending;
604 struct call_single_data hrtick_csd;
606 struct hrtimer hrtick_timer;
609 #ifdef CONFIG_SCHEDSTATS
611 struct sched_info rq_sched_info;
612 unsigned long long rq_cpu_time;
613 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
615 /* sys_sched_yield() stats */
616 unsigned int yld_count;
618 /* schedule() stats */
619 unsigned int sched_switch;
620 unsigned int sched_count;
621 unsigned int sched_goidle;
623 /* try_to_wake_up() stats */
624 unsigned int ttwu_count;
625 unsigned int ttwu_local;
628 unsigned int bkl_count;
632 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
635 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
637 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
640 static inline int cpu_of(struct rq *rq)
650 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
651 * See detach_destroy_domains: synchronize_sched for details.
653 * The domain tree of any CPU may only be accessed from within
654 * preempt-disabled sections.
656 #define for_each_domain(cpu, __sd) \
657 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
659 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
660 #define this_rq() (&__get_cpu_var(runqueues))
661 #define task_rq(p) cpu_rq(task_cpu(p))
662 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
663 #define raw_rq() (&__raw_get_cpu_var(runqueues))
665 inline void update_rq_clock(struct rq *rq)
667 rq->clock = sched_clock_cpu(cpu_of(rq));
671 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
673 #ifdef CONFIG_SCHED_DEBUG
674 # define const_debug __read_mostly
676 # define const_debug static const
681 * @cpu: the processor in question.
683 * Returns true if the current cpu runqueue is locked.
684 * This interface allows printk to be called with the runqueue lock
685 * held and know whether or not it is OK to wake up the klogd.
687 int runqueue_is_locked(int cpu)
689 return spin_is_locked(&cpu_rq(cpu)->lock);
693 * Debugging: various feature bits
696 #define SCHED_FEAT(name, enabled) \
697 __SCHED_FEAT_##name ,
700 #include "sched_features.h"
705 #define SCHED_FEAT(name, enabled) \
706 (1UL << __SCHED_FEAT_##name) * enabled |
708 const_debug unsigned int sysctl_sched_features =
709 #include "sched_features.h"
714 #ifdef CONFIG_SCHED_DEBUG
715 #define SCHED_FEAT(name, enabled) \
718 static __read_mostly char *sched_feat_names[] = {
719 #include "sched_features.h"
725 static int sched_feat_show(struct seq_file *m, void *v)
729 for (i = 0; sched_feat_names[i]; i++) {
730 if (!(sysctl_sched_features & (1UL << i)))
732 seq_printf(m, "%s ", sched_feat_names[i]);
740 sched_feat_write(struct file *filp, const char __user *ubuf,
741 size_t cnt, loff_t *ppos)
751 if (copy_from_user(&buf, ubuf, cnt))
756 if (strncmp(buf, "NO_", 3) == 0) {
761 for (i = 0; sched_feat_names[i]; i++) {
762 int len = strlen(sched_feat_names[i]);
764 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
766 sysctl_sched_features &= ~(1UL << i);
768 sysctl_sched_features |= (1UL << i);
773 if (!sched_feat_names[i])
781 static int sched_feat_open(struct inode *inode, struct file *filp)
783 return single_open(filp, sched_feat_show, NULL);
786 static const struct file_operations sched_feat_fops = {
787 .open = sched_feat_open,
788 .write = sched_feat_write,
791 .release = single_release,
794 static __init int sched_init_debug(void)
796 debugfs_create_file("sched_features", 0644, NULL, NULL,
801 late_initcall(sched_init_debug);
805 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
808 * Number of tasks to iterate in a single balance run.
809 * Limited because this is done with IRQs disabled.
811 const_debug unsigned int sysctl_sched_nr_migrate = 32;
814 * ratelimit for updating the group shares.
817 unsigned int sysctl_sched_shares_ratelimit = 250000;
818 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
821 * Inject some fuzzyness into changing the per-cpu group shares
822 * this avoids remote rq-locks at the expense of fairness.
825 unsigned int sysctl_sched_shares_thresh = 4;
828 * period over which we average the RT time consumption, measured
833 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
836 * period over which we measure -rt task cpu usage in us.
839 unsigned int sysctl_sched_rt_period = 1000000;
841 static __read_mostly int scheduler_running;
844 * part of the period that we allow rt tasks to run in us.
847 int sysctl_sched_rt_runtime = 950000;
849 static inline u64 global_rt_period(void)
851 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
854 static inline u64 global_rt_runtime(void)
856 if (sysctl_sched_rt_runtime < 0)
859 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
862 #ifndef prepare_arch_switch
863 # define prepare_arch_switch(next) do { } while (0)
865 #ifndef finish_arch_switch
866 # define finish_arch_switch(prev) do { } while (0)
869 static inline int task_current(struct rq *rq, struct task_struct *p)
871 return rq->curr == p;
874 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
875 static inline int task_running(struct rq *rq, struct task_struct *p)
877 return task_current(rq, p);
880 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
884 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
886 #ifdef CONFIG_DEBUG_SPINLOCK
887 /* this is a valid case when another task releases the spinlock */
888 rq->lock.owner = current;
891 * If we are tracking spinlock dependencies then we have to
892 * fix up the runqueue lock - which gets 'carried over' from
895 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
897 spin_unlock_irq(&rq->lock);
900 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
901 static inline int task_running(struct rq *rq, struct task_struct *p)
906 return task_current(rq, p);
910 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
914 * We can optimise this out completely for !SMP, because the
915 * SMP rebalancing from interrupt is the only thing that cares
920 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 spin_unlock_irq(&rq->lock);
923 spin_unlock(&rq->lock);
927 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
931 * After ->oncpu is cleared, the task can be moved to a different CPU.
932 * We must ensure this doesn't happen until the switch is completely
938 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
942 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
945 * Check whether the task is waking, we use this to synchronize against
946 * ttwu() so that task_cpu() reports a stable number.
948 * We need to make an exception for PF_STARTING tasks because the fork
949 * path might require task_rq_lock() to work, eg. it can call
950 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
952 static inline int task_is_waking(struct task_struct *p)
954 return unlikely((p->state == TASK_WAKING) && !(p->flags & PF_STARTING));
958 * __task_rq_lock - lock the runqueue a given task resides on.
959 * Must be called interrupts disabled.
961 static inline struct rq *__task_rq_lock(struct task_struct *p)
967 while (task_is_waking(p))
970 spin_lock(&rq->lock);
971 if (likely(rq == task_rq(p) && !task_is_waking(p)))
973 spin_unlock(&rq->lock);
978 * task_rq_lock - lock the runqueue a given task resides on and disable
979 * interrupts. Note the ordering: we can safely lookup the task_rq without
980 * explicitly disabling preemption.
982 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
988 while (task_is_waking(p))
990 local_irq_save(*flags);
992 spin_lock(&rq->lock);
993 if (likely(rq == task_rq(p) && !task_is_waking(p)))
995 spin_unlock_irqrestore(&rq->lock, *flags);
999 void task_rq_unlock_wait(struct task_struct *p)
1001 struct rq *rq = task_rq(p);
1003 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1004 spin_unlock_wait(&rq->lock);
1007 static void __task_rq_unlock(struct rq *rq)
1008 __releases(rq->lock)
1010 spin_unlock(&rq->lock);
1013 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1014 __releases(rq->lock)
1016 spin_unlock_irqrestore(&rq->lock, *flags);
1020 * this_rq_lock - lock this runqueue and disable interrupts.
1022 static struct rq *this_rq_lock(void)
1023 __acquires(rq->lock)
1027 local_irq_disable();
1029 spin_lock(&rq->lock);
1034 #ifdef CONFIG_SCHED_HRTICK
1036 * Use HR-timers to deliver accurate preemption points.
1038 * Its all a bit involved since we cannot program an hrt while holding the
1039 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1042 * When we get rescheduled we reprogram the hrtick_timer outside of the
1048 * - enabled by features
1049 * - hrtimer is actually high res
1051 static inline int hrtick_enabled(struct rq *rq)
1053 if (!sched_feat(HRTICK))
1055 if (!cpu_active(cpu_of(rq)))
1057 return hrtimer_is_hres_active(&rq->hrtick_timer);
1060 static void hrtick_clear(struct rq *rq)
1062 if (hrtimer_active(&rq->hrtick_timer))
1063 hrtimer_cancel(&rq->hrtick_timer);
1067 * High-resolution timer tick.
1068 * Runs from hardirq context with interrupts disabled.
1070 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1072 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1074 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1076 spin_lock(&rq->lock);
1077 update_rq_clock(rq);
1078 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1079 spin_unlock(&rq->lock);
1081 return HRTIMER_NORESTART;
1086 * called from hardirq (IPI) context
1088 static void __hrtick_start(void *arg)
1090 struct rq *rq = arg;
1092 spin_lock(&rq->lock);
1093 hrtimer_restart(&rq->hrtick_timer);
1094 rq->hrtick_csd_pending = 0;
1095 spin_unlock(&rq->lock);
1099 * Called to set the hrtick timer state.
1101 * called with rq->lock held and irqs disabled
1103 static void hrtick_start(struct rq *rq, u64 delay)
1105 struct hrtimer *timer = &rq->hrtick_timer;
1106 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1108 hrtimer_set_expires(timer, time);
1110 if (rq == this_rq()) {
1111 hrtimer_restart(timer);
1112 } else if (!rq->hrtick_csd_pending) {
1113 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1114 rq->hrtick_csd_pending = 1;
1119 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1121 int cpu = (int)(long)hcpu;
1124 case CPU_UP_CANCELED:
1125 case CPU_UP_CANCELED_FROZEN:
1126 case CPU_DOWN_PREPARE:
1127 case CPU_DOWN_PREPARE_FROZEN:
1129 case CPU_DEAD_FROZEN:
1130 hrtick_clear(cpu_rq(cpu));
1137 static __init void init_hrtick(void)
1139 hotcpu_notifier(hotplug_hrtick, 0);
1143 * Called to set the hrtick timer state.
1145 * called with rq->lock held and irqs disabled
1147 static void hrtick_start(struct rq *rq, u64 delay)
1149 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1150 HRTIMER_MODE_REL_PINNED, 0);
1153 static inline void init_hrtick(void)
1156 #endif /* CONFIG_SMP */
1158 static void init_rq_hrtick(struct rq *rq)
1161 rq->hrtick_csd_pending = 0;
1163 rq->hrtick_csd.flags = 0;
1164 rq->hrtick_csd.func = __hrtick_start;
1165 rq->hrtick_csd.info = rq;
1168 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1169 rq->hrtick_timer.function = hrtick;
1171 #else /* CONFIG_SCHED_HRTICK */
1172 static inline void hrtick_clear(struct rq *rq)
1176 static inline void init_rq_hrtick(struct rq *rq)
1180 static inline void init_hrtick(void)
1183 #endif /* CONFIG_SCHED_HRTICK */
1186 * resched_task - mark a task 'to be rescheduled now'.
1188 * On UP this means the setting of the need_resched flag, on SMP it
1189 * might also involve a cross-CPU call to trigger the scheduler on
1194 #ifndef tsk_is_polling
1195 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1198 static void resched_task(struct task_struct *p)
1202 assert_spin_locked(&task_rq(p)->lock);
1204 if (test_tsk_need_resched(p))
1207 set_tsk_need_resched(p);
1210 if (cpu == smp_processor_id())
1213 /* NEED_RESCHED must be visible before we test polling */
1215 if (!tsk_is_polling(p))
1216 smp_send_reschedule(cpu);
1219 static void resched_cpu(int cpu)
1221 struct rq *rq = cpu_rq(cpu);
1222 unsigned long flags;
1224 if (!spin_trylock_irqsave(&rq->lock, flags))
1226 resched_task(cpu_curr(cpu));
1227 spin_unlock_irqrestore(&rq->lock, flags);
1232 * When add_timer_on() enqueues a timer into the timer wheel of an
1233 * idle CPU then this timer might expire before the next timer event
1234 * which is scheduled to wake up that CPU. In case of a completely
1235 * idle system the next event might even be infinite time into the
1236 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1237 * leaves the inner idle loop so the newly added timer is taken into
1238 * account when the CPU goes back to idle and evaluates the timer
1239 * wheel for the next timer event.
1241 void wake_up_idle_cpu(int cpu)
1243 struct rq *rq = cpu_rq(cpu);
1245 if (cpu == smp_processor_id())
1249 * This is safe, as this function is called with the timer
1250 * wheel base lock of (cpu) held. When the CPU is on the way
1251 * to idle and has not yet set rq->curr to idle then it will
1252 * be serialized on the timer wheel base lock and take the new
1253 * timer into account automatically.
1255 if (rq->curr != rq->idle)
1259 * We can set TIF_RESCHED on the idle task of the other CPU
1260 * lockless. The worst case is that the other CPU runs the
1261 * idle task through an additional NOOP schedule()
1263 set_tsk_need_resched(rq->idle);
1265 /* NEED_RESCHED must be visible before we test polling */
1267 if (!tsk_is_polling(rq->idle))
1268 smp_send_reschedule(cpu);
1270 #endif /* CONFIG_NO_HZ */
1272 static u64 sched_avg_period(void)
1274 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1277 static void sched_avg_update(struct rq *rq)
1279 s64 period = sched_avg_period();
1281 while ((s64)(rq->clock - rq->age_stamp) > period) {
1283 * Inline assembly required to prevent the compiler
1284 * optimising this loop into a divmod call.
1285 * See __iter_div_u64_rem() for another example of this.
1287 asm("" : "+rm" (rq->age_stamp));
1288 rq->age_stamp += period;
1293 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1295 rq->rt_avg += rt_delta;
1296 sched_avg_update(rq);
1299 #else /* !CONFIG_SMP */
1300 static void resched_task(struct task_struct *p)
1302 assert_spin_locked(&task_rq(p)->lock);
1303 set_tsk_need_resched(p);
1306 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1309 #endif /* CONFIG_SMP */
1311 #if BITS_PER_LONG == 32
1312 # define WMULT_CONST (~0UL)
1314 # define WMULT_CONST (1UL << 32)
1317 #define WMULT_SHIFT 32
1320 * Shift right and round:
1322 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1325 * delta *= weight / lw
1327 static unsigned long
1328 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1329 struct load_weight *lw)
1333 if (!lw->inv_weight) {
1334 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1337 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1341 tmp = (u64)delta_exec * weight;
1343 * Check whether we'd overflow the 64-bit multiplication:
1345 if (unlikely(tmp > WMULT_CONST))
1346 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1349 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1351 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1354 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1360 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1367 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1368 * of tasks with abnormal "nice" values across CPUs the contribution that
1369 * each task makes to its run queue's load is weighted according to its
1370 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1371 * scaled version of the new time slice allocation that they receive on time
1375 #define WEIGHT_IDLEPRIO 3
1376 #define WMULT_IDLEPRIO 1431655765
1379 * Nice levels are multiplicative, with a gentle 10% change for every
1380 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1381 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1382 * that remained on nice 0.
1384 * The "10% effect" is relative and cumulative: from _any_ nice level,
1385 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1386 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1387 * If a task goes up by ~10% and another task goes down by ~10% then
1388 * the relative distance between them is ~25%.)
1390 static const int prio_to_weight[40] = {
1391 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1392 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1393 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1394 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1395 /* 0 */ 1024, 820, 655, 526, 423,
1396 /* 5 */ 335, 272, 215, 172, 137,
1397 /* 10 */ 110, 87, 70, 56, 45,
1398 /* 15 */ 36, 29, 23, 18, 15,
1402 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1404 * In cases where the weight does not change often, we can use the
1405 * precalculated inverse to speed up arithmetics by turning divisions
1406 * into multiplications:
1408 static const u32 prio_to_wmult[40] = {
1409 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1410 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1411 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1412 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1413 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1414 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1415 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1416 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1419 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1422 * runqueue iterator, to support SMP load-balancing between different
1423 * scheduling classes, without having to expose their internal data
1424 * structures to the load-balancing proper:
1426 struct rq_iterator {
1428 struct task_struct *(*start)(void *);
1429 struct task_struct *(*next)(void *);
1433 static unsigned long
1434 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1435 unsigned long max_load_move, struct sched_domain *sd,
1436 enum cpu_idle_type idle, int *all_pinned,
1437 int *this_best_prio, struct rq_iterator *iterator);
1440 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1441 struct sched_domain *sd, enum cpu_idle_type idle,
1442 struct rq_iterator *iterator);
1445 /* Time spent by the tasks of the cpu accounting group executing in ... */
1446 enum cpuacct_stat_index {
1447 CPUACCT_STAT_USER, /* ... user mode */
1448 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1450 CPUACCT_STAT_NSTATS,
1453 #ifdef CONFIG_CGROUP_CPUACCT
1454 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1455 static void cpuacct_update_stats(struct task_struct *tsk,
1456 enum cpuacct_stat_index idx, cputime_t val);
1458 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1459 static inline void cpuacct_update_stats(struct task_struct *tsk,
1460 enum cpuacct_stat_index idx, cputime_t val) {}
1463 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1465 update_load_add(&rq->load, load);
1468 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1470 update_load_sub(&rq->load, load);
1473 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1474 typedef int (*tg_visitor)(struct task_group *, void *);
1477 * Iterate the full tree, calling @down when first entering a node and @up when
1478 * leaving it for the final time.
1480 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1482 struct task_group *parent, *child;
1486 parent = &root_task_group;
1488 ret = (*down)(parent, data);
1491 list_for_each_entry_rcu(child, &parent->children, siblings) {
1498 ret = (*up)(parent, data);
1503 parent = parent->parent;
1512 static int tg_nop(struct task_group *tg, void *data)
1519 /* Used instead of source_load when we know the type == 0 */
1520 static unsigned long weighted_cpuload(const int cpu)
1522 return cpu_rq(cpu)->load.weight;
1526 * Return a low guess at the load of a migration-source cpu weighted
1527 * according to the scheduling class and "nice" value.
1529 * We want to under-estimate the load of migration sources, to
1530 * balance conservatively.
1532 static unsigned long source_load(int cpu, int type)
1534 struct rq *rq = cpu_rq(cpu);
1535 unsigned long total = weighted_cpuload(cpu);
1537 if (type == 0 || !sched_feat(LB_BIAS))
1540 return min(rq->cpu_load[type-1], total);
1544 * Return a high guess at the load of a migration-target cpu weighted
1545 * according to the scheduling class and "nice" value.
1547 static unsigned long target_load(int cpu, int type)
1549 struct rq *rq = cpu_rq(cpu);
1550 unsigned long total = weighted_cpuload(cpu);
1552 if (type == 0 || !sched_feat(LB_BIAS))
1555 return max(rq->cpu_load[type-1], total);
1558 static struct sched_group *group_of(int cpu)
1560 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1568 static unsigned long power_of(int cpu)
1570 struct sched_group *group = group_of(cpu);
1573 return SCHED_LOAD_SCALE;
1575 return group->cpu_power;
1578 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1580 static unsigned long cpu_avg_load_per_task(int cpu)
1582 struct rq *rq = cpu_rq(cpu);
1583 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1586 rq->avg_load_per_task = rq->load.weight / nr_running;
1588 rq->avg_load_per_task = 0;
1590 return rq->avg_load_per_task;
1593 #ifdef CONFIG_FAIR_GROUP_SCHED
1595 static __read_mostly unsigned long *update_shares_data;
1597 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1600 * Calculate and set the cpu's group shares.
1602 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1603 unsigned long sd_shares,
1604 unsigned long sd_rq_weight,
1605 unsigned long *usd_rq_weight)
1607 unsigned long shares, rq_weight;
1610 rq_weight = usd_rq_weight[cpu];
1613 rq_weight = NICE_0_LOAD;
1617 * \Sum_j shares_j * rq_weight_i
1618 * shares_i = -----------------------------
1619 * \Sum_j rq_weight_j
1621 shares = (sd_shares * rq_weight) / sd_rq_weight;
1622 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1624 if (abs(shares - tg->se[cpu]->load.weight) >
1625 sysctl_sched_shares_thresh) {
1626 struct rq *rq = cpu_rq(cpu);
1627 unsigned long flags;
1629 spin_lock_irqsave(&rq->lock, flags);
1630 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1631 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1632 __set_se_shares(tg->se[cpu], shares);
1633 spin_unlock_irqrestore(&rq->lock, flags);
1638 * Re-compute the task group their per cpu shares over the given domain.
1639 * This needs to be done in a bottom-up fashion because the rq weight of a
1640 * parent group depends on the shares of its child groups.
1642 static int tg_shares_up(struct task_group *tg, void *data)
1644 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1645 unsigned long *usd_rq_weight;
1646 struct sched_domain *sd = data;
1647 unsigned long flags;
1653 local_irq_save(flags);
1654 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1656 for_each_cpu(i, sched_domain_span(sd)) {
1657 weight = tg->cfs_rq[i]->load.weight;
1658 usd_rq_weight[i] = weight;
1660 rq_weight += weight;
1662 * If there are currently no tasks on the cpu pretend there
1663 * is one of average load so that when a new task gets to
1664 * run here it will not get delayed by group starvation.
1667 weight = NICE_0_LOAD;
1669 sum_weight += weight;
1670 shares += tg->cfs_rq[i]->shares;
1674 rq_weight = sum_weight;
1676 if ((!shares && rq_weight) || shares > tg->shares)
1677 shares = tg->shares;
1679 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1680 shares = tg->shares;
1682 for_each_cpu(i, sched_domain_span(sd))
1683 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1685 local_irq_restore(flags);
1691 * Compute the cpu's hierarchical load factor for each task group.
1692 * This needs to be done in a top-down fashion because the load of a child
1693 * group is a fraction of its parents load.
1695 static int tg_load_down(struct task_group *tg, void *data)
1698 long cpu = (long)data;
1701 load = cpu_rq(cpu)->load.weight;
1703 load = tg->parent->cfs_rq[cpu]->h_load;
1704 load *= tg->cfs_rq[cpu]->shares;
1705 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1708 tg->cfs_rq[cpu]->h_load = load;
1713 static void update_shares(struct sched_domain *sd)
1718 if (root_task_group_empty())
1721 now = cpu_clock(raw_smp_processor_id());
1722 elapsed = now - sd->last_update;
1724 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1725 sd->last_update = now;
1726 walk_tg_tree(tg_nop, tg_shares_up, sd);
1730 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1732 if (root_task_group_empty())
1735 spin_unlock(&rq->lock);
1737 spin_lock(&rq->lock);
1740 static void update_h_load(long cpu)
1742 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1747 static inline void update_shares(struct sched_domain *sd)
1751 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1757 #ifdef CONFIG_PREEMPT
1759 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1762 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1763 * way at the expense of forcing extra atomic operations in all
1764 * invocations. This assures that the double_lock is acquired using the
1765 * same underlying policy as the spinlock_t on this architecture, which
1766 * reduces latency compared to the unfair variant below. However, it
1767 * also adds more overhead and therefore may reduce throughput.
1769 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1770 __releases(this_rq->lock)
1771 __acquires(busiest->lock)
1772 __acquires(this_rq->lock)
1774 spin_unlock(&this_rq->lock);
1775 double_rq_lock(this_rq, busiest);
1782 * Unfair double_lock_balance: Optimizes throughput at the expense of
1783 * latency by eliminating extra atomic operations when the locks are
1784 * already in proper order on entry. This favors lower cpu-ids and will
1785 * grant the double lock to lower cpus over higher ids under contention,
1786 * regardless of entry order into the function.
1788 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1789 __releases(this_rq->lock)
1790 __acquires(busiest->lock)
1791 __acquires(this_rq->lock)
1795 if (unlikely(!spin_trylock(&busiest->lock))) {
1796 if (busiest < this_rq) {
1797 spin_unlock(&this_rq->lock);
1798 spin_lock(&busiest->lock);
1799 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1802 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1807 #endif /* CONFIG_PREEMPT */
1810 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1812 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1814 if (unlikely(!irqs_disabled())) {
1815 /* printk() doesn't work good under rq->lock */
1816 spin_unlock(&this_rq->lock);
1820 return _double_lock_balance(this_rq, busiest);
1823 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1824 __releases(busiest->lock)
1826 spin_unlock(&busiest->lock);
1827 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1831 #ifdef CONFIG_FAIR_GROUP_SCHED
1832 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1835 cfs_rq->shares = shares;
1840 static void calc_load_account_active(struct rq *this_rq);
1841 static void update_sysctl(void);
1843 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1845 set_task_rq(p, cpu);
1848 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1849 * successfuly executed on another CPU. We must ensure that updates of
1850 * per-task data have been completed by this moment.
1853 task_thread_info(p)->cpu = cpu;
1857 #include "sched_stats.h"
1858 #include "sched_idletask.c"
1859 #include "sched_fair.c"
1860 #include "sched_rt.c"
1861 #ifdef CONFIG_SCHED_DEBUG
1862 # include "sched_debug.c"
1865 #define sched_class_highest (&rt_sched_class)
1866 #define for_each_class(class) \
1867 for (class = sched_class_highest; class; class = class->next)
1869 static void inc_nr_running(struct rq *rq)
1874 static void dec_nr_running(struct rq *rq)
1879 static void set_load_weight(struct task_struct *p)
1881 if (task_has_rt_policy(p)) {
1882 p->se.load.weight = prio_to_weight[0] * 2;
1883 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1888 * SCHED_IDLE tasks get minimal weight:
1890 if (p->policy == SCHED_IDLE) {
1891 p->se.load.weight = WEIGHT_IDLEPRIO;
1892 p->se.load.inv_weight = WMULT_IDLEPRIO;
1896 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1897 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1900 static void update_avg(u64 *avg, u64 sample)
1902 s64 diff = sample - *avg;
1907 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1910 p->se.start_runtime = p->se.sum_exec_runtime;
1912 sched_info_queued(p);
1913 p->sched_class->enqueue_task(rq, p, wakeup, head);
1917 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1920 if (p->se.last_wakeup) {
1921 update_avg(&p->se.avg_overlap,
1922 p->se.sum_exec_runtime - p->se.last_wakeup);
1923 p->se.last_wakeup = 0;
1925 update_avg(&p->se.avg_wakeup,
1926 sysctl_sched_wakeup_granularity);
1930 sched_info_dequeued(p);
1931 p->sched_class->dequeue_task(rq, p, sleep);
1936 * __normal_prio - return the priority that is based on the static prio
1938 static inline int __normal_prio(struct task_struct *p)
1940 return p->static_prio;
1944 * Calculate the expected normal priority: i.e. priority
1945 * without taking RT-inheritance into account. Might be
1946 * boosted by interactivity modifiers. Changes upon fork,
1947 * setprio syscalls, and whenever the interactivity
1948 * estimator recalculates.
1950 static inline int normal_prio(struct task_struct *p)
1954 if (task_has_rt_policy(p))
1955 prio = MAX_RT_PRIO-1 - p->rt_priority;
1957 prio = __normal_prio(p);
1962 * Calculate the current priority, i.e. the priority
1963 * taken into account by the scheduler. This value might
1964 * be boosted by RT tasks, or might be boosted by
1965 * interactivity modifiers. Will be RT if the task got
1966 * RT-boosted. If not then it returns p->normal_prio.
1968 static int effective_prio(struct task_struct *p)
1970 p->normal_prio = normal_prio(p);
1972 * If we are RT tasks or we were boosted to RT priority,
1973 * keep the priority unchanged. Otherwise, update priority
1974 * to the normal priority:
1976 if (!rt_prio(p->prio))
1977 return p->normal_prio;
1982 * activate_task - move a task to the runqueue.
1984 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1986 if (task_contributes_to_load(p))
1987 rq->nr_uninterruptible--;
1989 enqueue_task(rq, p, wakeup, false);
1994 * deactivate_task - remove a task from the runqueue.
1996 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1998 if (task_contributes_to_load(p))
1999 rq->nr_uninterruptible++;
2001 dequeue_task(rq, p, sleep);
2006 * task_curr - is this task currently executing on a CPU?
2007 * @p: the task in question.
2009 inline int task_curr(const struct task_struct *p)
2011 return cpu_curr(task_cpu(p)) == p;
2014 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2015 const struct sched_class *prev_class,
2016 int oldprio, int running)
2018 if (prev_class != p->sched_class) {
2019 if (prev_class->switched_from)
2020 prev_class->switched_from(rq, p, running);
2021 p->sched_class->switched_to(rq, p, running);
2023 p->sched_class->prio_changed(rq, p, oldprio, running);
2027 * kthread_bind - bind a just-created kthread to a cpu.
2028 * @p: thread created by kthread_create().
2029 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2031 * Description: This function is equivalent to set_cpus_allowed(),
2032 * except that @cpu doesn't need to be online, and the thread must be
2033 * stopped (i.e., just returned from kthread_create()).
2035 * Function lives here instead of kthread.c because it messes with
2036 * scheduler internals which require locking.
2038 void kthread_bind(struct task_struct *p, unsigned int cpu)
2040 /* Must have done schedule() in kthread() before we set_task_cpu */
2041 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2046 p->cpus_allowed = cpumask_of_cpu(cpu);
2047 p->rt.nr_cpus_allowed = 1;
2048 p->flags |= PF_THREAD_BOUND;
2050 EXPORT_SYMBOL(kthread_bind);
2054 * Is this task likely cache-hot:
2057 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2061 if (p->sched_class != &fair_sched_class)
2065 * Buddy candidates are cache hot:
2067 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2068 (&p->se == cfs_rq_of(&p->se)->next ||
2069 &p->se == cfs_rq_of(&p->se)->last))
2072 if (sysctl_sched_migration_cost == -1)
2074 if (sysctl_sched_migration_cost == 0)
2077 delta = now - p->se.exec_start;
2079 return delta < (s64)sysctl_sched_migration_cost;
2083 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2085 int old_cpu = task_cpu(p);
2087 #ifdef CONFIG_SCHED_DEBUG
2089 * We should never call set_task_cpu() on a blocked task,
2090 * ttwu() will sort out the placement.
2092 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2093 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2096 trace_sched_migrate_task(p, new_cpu);
2098 if (old_cpu != new_cpu) {
2099 p->se.nr_migrations++;
2100 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2104 __set_task_cpu(p, new_cpu);
2107 struct migration_req {
2108 struct list_head list;
2110 struct task_struct *task;
2113 struct completion done;
2117 * The task's runqueue lock must be held.
2118 * Returns true if you have to wait for migration thread.
2121 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2123 struct rq *rq = task_rq(p);
2126 * If the task is not on a runqueue (and not running), then
2127 * the next wake-up will properly place the task.
2129 if (!p->se.on_rq && !task_running(rq, p))
2132 init_completion(&req->done);
2134 req->dest_cpu = dest_cpu;
2135 list_add(&req->list, &rq->migration_queue);
2141 * wait_task_context_switch - wait for a thread to complete at least one
2144 * @p must not be current.
2146 void wait_task_context_switch(struct task_struct *p)
2148 unsigned long nvcsw, nivcsw, flags;
2156 * The runqueue is assigned before the actual context
2157 * switch. We need to take the runqueue lock.
2159 * We could check initially without the lock but it is
2160 * very likely that we need to take the lock in every
2163 rq = task_rq_lock(p, &flags);
2164 running = task_running(rq, p);
2165 task_rq_unlock(rq, &flags);
2167 if (likely(!running))
2170 * The switch count is incremented before the actual
2171 * context switch. We thus wait for two switches to be
2172 * sure at least one completed.
2174 if ((p->nvcsw - nvcsw) > 1)
2176 if ((p->nivcsw - nivcsw) > 1)
2184 * wait_task_inactive - wait for a thread to unschedule.
2186 * If @match_state is nonzero, it's the @p->state value just checked and
2187 * not expected to change. If it changes, i.e. @p might have woken up,
2188 * then return zero. When we succeed in waiting for @p to be off its CPU,
2189 * we return a positive number (its total switch count). If a second call
2190 * a short while later returns the same number, the caller can be sure that
2191 * @p has remained unscheduled the whole time.
2193 * The caller must ensure that the task *will* unschedule sometime soon,
2194 * else this function might spin for a *long* time. This function can't
2195 * be called with interrupts off, or it may introduce deadlock with
2196 * smp_call_function() if an IPI is sent by the same process we are
2197 * waiting to become inactive.
2199 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2201 unsigned long flags;
2208 * We do the initial early heuristics without holding
2209 * any task-queue locks at all. We'll only try to get
2210 * the runqueue lock when things look like they will
2216 * If the task is actively running on another CPU
2217 * still, just relax and busy-wait without holding
2220 * NOTE! Since we don't hold any locks, it's not
2221 * even sure that "rq" stays as the right runqueue!
2222 * But we don't care, since "task_running()" will
2223 * return false if the runqueue has changed and p
2224 * is actually now running somewhere else!
2226 while (task_running(rq, p)) {
2227 if (match_state && unlikely(p->state != match_state))
2233 * Ok, time to look more closely! We need the rq
2234 * lock now, to be *sure*. If we're wrong, we'll
2235 * just go back and repeat.
2237 rq = task_rq_lock(p, &flags);
2238 trace_sched_wait_task(rq, p);
2239 running = task_running(rq, p);
2240 on_rq = p->se.on_rq;
2242 if (!match_state || p->state == match_state)
2243 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2244 task_rq_unlock(rq, &flags);
2247 * If it changed from the expected state, bail out now.
2249 if (unlikely(!ncsw))
2253 * Was it really running after all now that we
2254 * checked with the proper locks actually held?
2256 * Oops. Go back and try again..
2258 if (unlikely(running)) {
2264 * It's not enough that it's not actively running,
2265 * it must be off the runqueue _entirely_, and not
2268 * So if it was still runnable (but just not actively
2269 * running right now), it's preempted, and we should
2270 * yield - it could be a while.
2272 if (unlikely(on_rq)) {
2273 schedule_timeout_uninterruptible(1);
2278 * Ahh, all good. It wasn't running, and it wasn't
2279 * runnable, which means that it will never become
2280 * running in the future either. We're all done!
2289 * kick_process - kick a running thread to enter/exit the kernel
2290 * @p: the to-be-kicked thread
2292 * Cause a process which is running on another CPU to enter
2293 * kernel-mode, without any delay. (to get signals handled.)
2295 * NOTE: this function doesnt have to take the runqueue lock,
2296 * because all it wants to ensure is that the remote task enters
2297 * the kernel. If the IPI races and the task has been migrated
2298 * to another CPU then no harm is done and the purpose has been
2301 void kick_process(struct task_struct *p)
2307 if ((cpu != smp_processor_id()) && task_curr(p))
2308 smp_send_reschedule(cpu);
2311 EXPORT_SYMBOL_GPL(kick_process);
2312 #endif /* CONFIG_SMP */
2315 * task_oncpu_function_call - call a function on the cpu on which a task runs
2316 * @p: the task to evaluate
2317 * @func: the function to be called
2318 * @info: the function call argument
2320 * Calls the function @func when the task is currently running. This might
2321 * be on the current CPU, which just calls the function directly
2323 void task_oncpu_function_call(struct task_struct *p,
2324 void (*func) (void *info), void *info)
2331 smp_call_function_single(cpu, func, info, 1);
2336 static int select_fallback_rq(int cpu, struct task_struct *p)
2339 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2341 /* Look for allowed, online CPU in same node. */
2342 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2343 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2346 /* Any allowed, online CPU? */
2347 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2348 if (dest_cpu < nr_cpu_ids)
2351 /* No more Mr. Nice Guy. */
2352 if (dest_cpu >= nr_cpu_ids) {
2354 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2356 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2359 * Don't tell them about moving exiting tasks or
2360 * kernel threads (both mm NULL), since they never
2363 if (p->mm && printk_ratelimit()) {
2364 printk(KERN_INFO "process %d (%s) no "
2365 "longer affine to cpu%d\n",
2366 task_pid_nr(p), p->comm, cpu);
2374 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2375 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2378 * exec: is unstable, retry loop
2379 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2382 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2384 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2387 * In order not to call set_task_cpu() on a blocking task we need
2388 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2391 * Since this is common to all placement strategies, this lives here.
2393 * [ this allows ->select_task() to simply return task_cpu(p) and
2394 * not worry about this generic constraint ]
2396 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2398 cpu = select_fallback_rq(task_cpu(p), p);
2405 * try_to_wake_up - wake up a thread
2406 * @p: the to-be-woken-up thread
2407 * @state: the mask of task states that can be woken
2408 * @sync: do a synchronous wakeup?
2410 * Put it on the run-queue if it's not already there. The "current"
2411 * thread is always on the run-queue (except when the actual
2412 * re-schedule is in progress), and as such you're allowed to do
2413 * the simpler "current->state = TASK_RUNNING" to mark yourself
2414 * runnable without the overhead of this.
2416 * returns failure only if the task is already active.
2418 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2421 int cpu, orig_cpu, this_cpu, success = 0;
2422 unsigned long flags;
2423 struct rq *rq, *orig_rq;
2425 if (!sched_feat(SYNC_WAKEUPS))
2426 wake_flags &= ~WF_SYNC;
2428 this_cpu = get_cpu();
2431 rq = orig_rq = task_rq_lock(p, &flags);
2432 update_rq_clock(rq);
2433 if (!(p->state & state))
2443 if (unlikely(task_running(rq, p)))
2447 * In order to handle concurrent wakeups and release the rq->lock
2448 * we put the task in TASK_WAKING state.
2450 * First fix up the nr_uninterruptible count:
2452 if (task_contributes_to_load(p))
2453 rq->nr_uninterruptible--;
2454 p->state = TASK_WAKING;
2456 if (p->sched_class->task_waking)
2457 p->sched_class->task_waking(rq, p);
2459 __task_rq_unlock(rq);
2461 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2462 if (cpu != orig_cpu) {
2464 * Since we migrate the task without holding any rq->lock,
2465 * we need to be careful with task_rq_lock(), since that
2466 * might end up locking an invalid rq.
2468 set_task_cpu(p, cpu);
2472 spin_lock(&rq->lock);
2473 update_rq_clock(rq);
2476 * We migrated the task without holding either rq->lock, however
2477 * since the task is not on the task list itself, nobody else
2478 * will try and migrate the task, hence the rq should match the
2479 * cpu we just moved it to.
2481 WARN_ON(task_cpu(p) != cpu);
2482 WARN_ON(p->state != TASK_WAKING);
2484 #ifdef CONFIG_SCHEDSTATS
2485 schedstat_inc(rq, ttwu_count);
2486 if (cpu == this_cpu)
2487 schedstat_inc(rq, ttwu_local);
2489 struct sched_domain *sd;
2490 for_each_domain(this_cpu, sd) {
2491 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2492 schedstat_inc(sd, ttwu_wake_remote);
2497 #endif /* CONFIG_SCHEDSTATS */
2500 #endif /* CONFIG_SMP */
2501 schedstat_inc(p, se.nr_wakeups);
2502 if (wake_flags & WF_SYNC)
2503 schedstat_inc(p, se.nr_wakeups_sync);
2504 if (orig_cpu != cpu)
2505 schedstat_inc(p, se.nr_wakeups_migrate);
2506 if (cpu == this_cpu)
2507 schedstat_inc(p, se.nr_wakeups_local);
2509 schedstat_inc(p, se.nr_wakeups_remote);
2510 activate_task(rq, p, 1);
2514 * Only attribute actual wakeups done by this task.
2516 if (!in_interrupt()) {
2517 struct sched_entity *se = ¤t->se;
2518 u64 sample = se->sum_exec_runtime;
2520 if (se->last_wakeup)
2521 sample -= se->last_wakeup;
2523 sample -= se->start_runtime;
2524 update_avg(&se->avg_wakeup, sample);
2526 se->last_wakeup = se->sum_exec_runtime;
2530 trace_sched_wakeup(rq, p, success);
2531 check_preempt_curr(rq, p, wake_flags);
2533 p->state = TASK_RUNNING;
2535 if (p->sched_class->task_woken)
2536 p->sched_class->task_woken(rq, p);
2538 if (unlikely(rq->idle_stamp)) {
2539 u64 delta = rq->clock - rq->idle_stamp;
2540 u64 max = 2*sysctl_sched_migration_cost;
2545 update_avg(&rq->avg_idle, delta);
2550 task_rq_unlock(rq, &flags);
2557 * wake_up_process - Wake up a specific process
2558 * @p: The process to be woken up.
2560 * Attempt to wake up the nominated process and move it to the set of runnable
2561 * processes. Returns 1 if the process was woken up, 0 if it was already
2564 * It may be assumed that this function implies a write memory barrier before
2565 * changing the task state if and only if any tasks are woken up.
2567 int wake_up_process(struct task_struct *p)
2569 return try_to_wake_up(p, TASK_ALL, 0);
2571 EXPORT_SYMBOL(wake_up_process);
2573 int wake_up_state(struct task_struct *p, unsigned int state)
2575 return try_to_wake_up(p, state, 0);
2579 * Perform scheduler related setup for a newly forked process p.
2580 * p is forked by current.
2582 * __sched_fork() is basic setup used by init_idle() too:
2584 static void __sched_fork(struct task_struct *p)
2586 p->se.exec_start = 0;
2587 p->se.sum_exec_runtime = 0;
2588 p->se.prev_sum_exec_runtime = 0;
2589 p->se.nr_migrations = 0;
2590 p->se.last_wakeup = 0;
2591 p->se.avg_overlap = 0;
2592 p->se.start_runtime = 0;
2593 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2594 p->se.avg_running = 0;
2596 #ifdef CONFIG_SCHEDSTATS
2597 p->se.wait_start = 0;
2599 p->se.wait_count = 0;
2602 p->se.sleep_start = 0;
2603 p->se.sleep_max = 0;
2604 p->se.sum_sleep_runtime = 0;
2606 p->se.block_start = 0;
2607 p->se.block_max = 0;
2609 p->se.slice_max = 0;
2611 p->se.nr_migrations_cold = 0;
2612 p->se.nr_failed_migrations_affine = 0;
2613 p->se.nr_failed_migrations_running = 0;
2614 p->se.nr_failed_migrations_hot = 0;
2615 p->se.nr_forced_migrations = 0;
2617 p->se.nr_wakeups = 0;
2618 p->se.nr_wakeups_sync = 0;
2619 p->se.nr_wakeups_migrate = 0;
2620 p->se.nr_wakeups_local = 0;
2621 p->se.nr_wakeups_remote = 0;
2622 p->se.nr_wakeups_affine = 0;
2623 p->se.nr_wakeups_affine_attempts = 0;
2624 p->se.nr_wakeups_passive = 0;
2625 p->se.nr_wakeups_idle = 0;
2629 INIT_LIST_HEAD(&p->rt.run_list);
2631 INIT_LIST_HEAD(&p->se.group_node);
2633 #ifdef CONFIG_PREEMPT_NOTIFIERS
2634 INIT_HLIST_HEAD(&p->preempt_notifiers);
2639 * fork()/clone()-time setup:
2641 void sched_fork(struct task_struct *p, int clone_flags)
2643 int cpu = get_cpu();
2647 * We mark the process as waking here. This guarantees that
2648 * nobody will actually run it, and a signal or other external
2649 * event cannot wake it up and insert it on the runqueue either.
2651 p->state = TASK_WAKING;
2654 * Revert to default priority/policy on fork if requested.
2656 if (unlikely(p->sched_reset_on_fork)) {
2657 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2658 p->policy = SCHED_NORMAL;
2659 p->normal_prio = p->static_prio;
2662 if (PRIO_TO_NICE(p->static_prio) < 0) {
2663 p->static_prio = NICE_TO_PRIO(0);
2664 p->normal_prio = p->static_prio;
2669 * We don't need the reset flag anymore after the fork. It has
2670 * fulfilled its duty:
2672 p->sched_reset_on_fork = 0;
2676 * Make sure we do not leak PI boosting priority to the child.
2678 p->prio = current->normal_prio;
2680 if (!rt_prio(p->prio))
2681 p->sched_class = &fair_sched_class;
2683 if (p->sched_class->task_fork)
2684 p->sched_class->task_fork(p);
2686 set_task_cpu(p, cpu);
2688 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2689 if (likely(sched_info_on()))
2690 memset(&p->sched_info, 0, sizeof(p->sched_info));
2692 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2695 #ifdef CONFIG_PREEMPT
2696 /* Want to start with kernel preemption disabled. */
2697 task_thread_info(p)->preempt_count = 1;
2699 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2705 * wake_up_new_task - wake up a newly created task for the first time.
2707 * This function will do some initial scheduler statistics housekeeping
2708 * that must be done for every newly created context, then puts the task
2709 * on the runqueue and wakes it.
2711 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2713 unsigned long flags;
2715 int cpu = get_cpu();
2719 * Fork balancing, do it here and not earlier because:
2720 * - cpus_allowed can change in the fork path
2721 * - any previously selected cpu might disappear through hotplug
2723 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2724 * ->cpus_allowed is stable, we have preemption disabled, meaning
2725 * cpu_online_mask is stable.
2727 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2728 set_task_cpu(p, cpu);
2732 * Since the task is not on the rq and we still have TASK_WAKING set
2733 * nobody else will migrate this task.
2736 spin_lock_irqsave(&rq->lock, flags);
2738 BUG_ON(p->state != TASK_WAKING);
2739 p->state = TASK_RUNNING;
2740 update_rq_clock(rq);
2741 activate_task(rq, p, 0);
2742 trace_sched_wakeup_new(rq, p, 1);
2743 check_preempt_curr(rq, p, WF_FORK);
2745 if (p->sched_class->task_woken)
2746 p->sched_class->task_woken(rq, p);
2748 task_rq_unlock(rq, &flags);
2752 #ifdef CONFIG_PREEMPT_NOTIFIERS
2755 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2756 * @notifier: notifier struct to register
2758 void preempt_notifier_register(struct preempt_notifier *notifier)
2760 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2762 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2765 * preempt_notifier_unregister - no longer interested in preemption notifications
2766 * @notifier: notifier struct to unregister
2768 * This is safe to call from within a preemption notifier.
2770 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2772 hlist_del(¬ifier->link);
2774 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2776 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2778 struct preempt_notifier *notifier;
2779 struct hlist_node *node;
2781 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2782 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2786 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2787 struct task_struct *next)
2789 struct preempt_notifier *notifier;
2790 struct hlist_node *node;
2792 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2793 notifier->ops->sched_out(notifier, next);
2796 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2798 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2803 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2804 struct task_struct *next)
2808 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2811 * prepare_task_switch - prepare to switch tasks
2812 * @rq: the runqueue preparing to switch
2813 * @prev: the current task that is being switched out
2814 * @next: the task we are going to switch to.
2816 * This is called with the rq lock held and interrupts off. It must
2817 * be paired with a subsequent finish_task_switch after the context
2820 * prepare_task_switch sets up locking and calls architecture specific
2824 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2825 struct task_struct *next)
2827 fire_sched_out_preempt_notifiers(prev, next);
2828 prepare_lock_switch(rq, next);
2829 prepare_arch_switch(next);
2833 * finish_task_switch - clean up after a task-switch
2834 * @rq: runqueue associated with task-switch
2835 * @prev: the thread we just switched away from.
2837 * finish_task_switch must be called after the context switch, paired
2838 * with a prepare_task_switch call before the context switch.
2839 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2840 * and do any other architecture-specific cleanup actions.
2842 * Note that we may have delayed dropping an mm in context_switch(). If
2843 * so, we finish that here outside of the runqueue lock. (Doing it
2844 * with the lock held can cause deadlocks; see schedule() for
2847 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2848 __releases(rq->lock)
2850 struct mm_struct *mm = rq->prev_mm;
2856 * A task struct has one reference for the use as "current".
2857 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2858 * schedule one last time. The schedule call will never return, and
2859 * the scheduled task must drop that reference.
2860 * The test for TASK_DEAD must occur while the runqueue locks are
2861 * still held, otherwise prev could be scheduled on another cpu, die
2862 * there before we look at prev->state, and then the reference would
2864 * Manfred Spraul <manfred@colorfullife.com>
2866 prev_state = prev->state;
2867 finish_arch_switch(prev);
2868 perf_event_task_sched_in(current, cpu_of(rq));
2869 finish_lock_switch(rq, prev);
2871 fire_sched_in_preempt_notifiers(current);
2874 if (unlikely(prev_state == TASK_DEAD)) {
2876 * Remove function-return probe instances associated with this
2877 * task and put them back on the free list.
2879 kprobe_flush_task(prev);
2880 put_task_struct(prev);
2886 /* assumes rq->lock is held */
2887 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2889 if (prev->sched_class->pre_schedule)
2890 prev->sched_class->pre_schedule(rq, prev);
2893 /* rq->lock is NOT held, but preemption is disabled */
2894 static inline void post_schedule(struct rq *rq)
2896 if (rq->post_schedule) {
2897 unsigned long flags;
2899 spin_lock_irqsave(&rq->lock, flags);
2900 if (rq->curr->sched_class->post_schedule)
2901 rq->curr->sched_class->post_schedule(rq);
2902 spin_unlock_irqrestore(&rq->lock, flags);
2904 rq->post_schedule = 0;
2910 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2914 static inline void post_schedule(struct rq *rq)
2921 * schedule_tail - first thing a freshly forked thread must call.
2922 * @prev: the thread we just switched away from.
2924 asmlinkage void schedule_tail(struct task_struct *prev)
2925 __releases(rq->lock)
2927 struct rq *rq = this_rq();
2929 finish_task_switch(rq, prev);
2932 * FIXME: do we need to worry about rq being invalidated by the
2937 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2938 /* In this case, finish_task_switch does not reenable preemption */
2941 if (current->set_child_tid)
2942 put_user(task_pid_vnr(current), current->set_child_tid);
2946 * context_switch - switch to the new MM and the new
2947 * thread's register state.
2950 context_switch(struct rq *rq, struct task_struct *prev,
2951 struct task_struct *next)
2953 struct mm_struct *mm, *oldmm;
2955 prepare_task_switch(rq, prev, next);
2956 trace_sched_switch(rq, prev, next);
2958 oldmm = prev->active_mm;
2960 * For paravirt, this is coupled with an exit in switch_to to
2961 * combine the page table reload and the switch backend into
2964 arch_start_context_switch(prev);
2966 if (unlikely(!mm)) {
2967 next->active_mm = oldmm;
2968 atomic_inc(&oldmm->mm_count);
2969 enter_lazy_tlb(oldmm, next);
2971 switch_mm(oldmm, mm, next);
2973 if (unlikely(!prev->mm)) {
2974 prev->active_mm = NULL;
2975 rq->prev_mm = oldmm;
2978 * Since the runqueue lock will be released by the next
2979 * task (which is an invalid locking op but in the case
2980 * of the scheduler it's an obvious special-case), so we
2981 * do an early lockdep release here:
2983 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2984 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2987 /* Here we just switch the register state and the stack. */
2988 switch_to(prev, next, prev);
2992 * this_rq must be evaluated again because prev may have moved
2993 * CPUs since it called schedule(), thus the 'rq' on its stack
2994 * frame will be invalid.
2996 finish_task_switch(this_rq(), prev);
3000 * nr_running, nr_uninterruptible and nr_context_switches:
3002 * externally visible scheduler statistics: current number of runnable
3003 * threads, current number of uninterruptible-sleeping threads, total
3004 * number of context switches performed since bootup.
3006 unsigned long nr_running(void)
3008 unsigned long i, sum = 0;
3010 for_each_online_cpu(i)
3011 sum += cpu_rq(i)->nr_running;
3016 unsigned long nr_uninterruptible(void)
3018 unsigned long i, sum = 0;
3020 for_each_possible_cpu(i)
3021 sum += cpu_rq(i)->nr_uninterruptible;
3024 * Since we read the counters lockless, it might be slightly
3025 * inaccurate. Do not allow it to go below zero though:
3027 if (unlikely((long)sum < 0))
3033 unsigned long long nr_context_switches(void)
3036 unsigned long long sum = 0;
3038 for_each_possible_cpu(i)
3039 sum += cpu_rq(i)->nr_switches;
3044 unsigned long nr_iowait(void)
3046 unsigned long i, sum = 0;
3048 for_each_possible_cpu(i)
3049 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3054 unsigned long nr_iowait_cpu(void)
3056 struct rq *this = this_rq();
3057 return atomic_read(&this->nr_iowait);
3060 unsigned long this_cpu_load(void)
3062 struct rq *this = this_rq();
3063 return this->cpu_load[0];
3067 /* Variables and functions for calc_load */
3068 static atomic_long_t calc_load_tasks;
3069 static unsigned long calc_load_update;
3070 unsigned long avenrun[3];
3071 EXPORT_SYMBOL(avenrun);
3074 * get_avenrun - get the load average array
3075 * @loads: pointer to dest load array
3076 * @offset: offset to add
3077 * @shift: shift count to shift the result left
3079 * These values are estimates at best, so no need for locking.
3081 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3083 loads[0] = (avenrun[0] + offset) << shift;
3084 loads[1] = (avenrun[1] + offset) << shift;
3085 loads[2] = (avenrun[2] + offset) << shift;
3088 static unsigned long
3089 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3092 load += active * (FIXED_1 - exp);
3093 return load >> FSHIFT;
3097 * calc_load - update the avenrun load estimates 10 ticks after the
3098 * CPUs have updated calc_load_tasks.
3100 void calc_global_load(void)
3102 unsigned long upd = calc_load_update + 10;
3105 if (time_before(jiffies, upd))
3108 active = atomic_long_read(&calc_load_tasks);
3109 active = active > 0 ? active * FIXED_1 : 0;
3111 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3112 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3113 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3115 calc_load_update += LOAD_FREQ;
3119 * Either called from update_cpu_load() or from a cpu going idle
3121 static void calc_load_account_active(struct rq *this_rq)
3123 long nr_active, delta;
3125 nr_active = this_rq->nr_running;
3126 nr_active += (long) this_rq->nr_uninterruptible;
3128 if (nr_active != this_rq->calc_load_active) {
3129 delta = nr_active - this_rq->calc_load_active;
3130 this_rq->calc_load_active = nr_active;
3131 atomic_long_add(delta, &calc_load_tasks);
3136 * Update rq->cpu_load[] statistics. This function is usually called every
3137 * scheduler tick (TICK_NSEC).
3139 static void update_cpu_load(struct rq *this_rq)
3141 unsigned long this_load = this_rq->load.weight;
3144 this_rq->nr_load_updates++;
3146 /* Update our load: */
3147 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3148 unsigned long old_load, new_load;
3150 /* scale is effectively 1 << i now, and >> i divides by scale */
3152 old_load = this_rq->cpu_load[i];
3153 new_load = this_load;
3155 * Round up the averaging division if load is increasing. This
3156 * prevents us from getting stuck on 9 if the load is 10, for
3159 if (new_load > old_load)
3160 new_load += scale-1;
3161 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3164 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3165 this_rq->calc_load_update += LOAD_FREQ;
3166 calc_load_account_active(this_rq);
3173 * double_rq_lock - safely lock two runqueues
3175 * Note this does not disable interrupts like task_rq_lock,
3176 * you need to do so manually before calling.
3178 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3179 __acquires(rq1->lock)
3180 __acquires(rq2->lock)
3182 BUG_ON(!irqs_disabled());
3184 spin_lock(&rq1->lock);
3185 __acquire(rq2->lock); /* Fake it out ;) */
3188 spin_lock(&rq1->lock);
3189 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3191 spin_lock(&rq2->lock);
3192 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3195 update_rq_clock(rq1);
3196 update_rq_clock(rq2);
3200 * double_rq_unlock - safely unlock two runqueues
3202 * Note this does not restore interrupts like task_rq_unlock,
3203 * you need to do so manually after calling.
3205 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3206 __releases(rq1->lock)
3207 __releases(rq2->lock)
3209 spin_unlock(&rq1->lock);
3211 spin_unlock(&rq2->lock);
3213 __release(rq2->lock);
3217 * sched_exec - execve() is a valuable balancing opportunity, because at
3218 * this point the task has the smallest effective memory and cache footprint.
3220 void sched_exec(void)
3222 struct task_struct *p = current;
3223 struct migration_req req;
3224 int dest_cpu, this_cpu;
3225 unsigned long flags;
3229 this_cpu = get_cpu();
3230 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3231 if (dest_cpu == this_cpu) {
3236 rq = task_rq_lock(p, &flags);
3240 * select_task_rq() can race against ->cpus_allowed
3242 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3243 || unlikely(!cpu_active(dest_cpu))) {
3244 task_rq_unlock(rq, &flags);
3248 /* force the process onto the specified CPU */
3249 if (migrate_task(p, dest_cpu, &req)) {
3250 /* Need to wait for migration thread (might exit: take ref). */
3251 struct task_struct *mt = rq->migration_thread;
3253 get_task_struct(mt);
3254 task_rq_unlock(rq, &flags);
3255 wake_up_process(mt);
3256 put_task_struct(mt);
3257 wait_for_completion(&req.done);
3261 task_rq_unlock(rq, &flags);
3265 * pull_task - move a task from a remote runqueue to the local runqueue.
3266 * Both runqueues must be locked.
3268 static void pull_task(struct rq *src_rq, struct task_struct *p,
3269 struct rq *this_rq, int this_cpu)
3271 deactivate_task(src_rq, p, 0);
3272 set_task_cpu(p, this_cpu);
3273 activate_task(this_rq, p, 0);
3274 check_preempt_curr(this_rq, p, 0);
3278 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3281 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3282 struct sched_domain *sd, enum cpu_idle_type idle,
3285 int tsk_cache_hot = 0;
3287 * We do not migrate tasks that are:
3288 * 1) running (obviously), or
3289 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3290 * 3) are cache-hot on their current CPU.
3292 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3293 schedstat_inc(p, se.nr_failed_migrations_affine);
3298 if (task_running(rq, p)) {
3299 schedstat_inc(p, se.nr_failed_migrations_running);
3304 * Aggressive migration if:
3305 * 1) task is cache cold, or
3306 * 2) too many balance attempts have failed.
3309 tsk_cache_hot = task_hot(p, rq->clock, sd);
3310 if (!tsk_cache_hot ||
3311 sd->nr_balance_failed > sd->cache_nice_tries) {
3312 #ifdef CONFIG_SCHEDSTATS
3313 if (tsk_cache_hot) {
3314 schedstat_inc(sd, lb_hot_gained[idle]);
3315 schedstat_inc(p, se.nr_forced_migrations);
3321 if (tsk_cache_hot) {
3322 schedstat_inc(p, se.nr_failed_migrations_hot);
3328 static unsigned long
3329 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3330 unsigned long max_load_move, struct sched_domain *sd,
3331 enum cpu_idle_type idle, int *all_pinned,
3332 int *this_best_prio, struct rq_iterator *iterator)
3334 int loops = 0, pulled = 0, pinned = 0;
3335 struct task_struct *p;
3336 long rem_load_move = max_load_move;
3338 if (max_load_move == 0)
3344 * Start the load-balancing iterator:
3346 p = iterator->start(iterator->arg);
3348 if (!p || loops++ > sysctl_sched_nr_migrate)
3351 if ((p->se.load.weight >> 1) > rem_load_move ||
3352 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3353 p = iterator->next(iterator->arg);
3357 pull_task(busiest, p, this_rq, this_cpu);
3359 rem_load_move -= p->se.load.weight;
3361 #ifdef CONFIG_PREEMPT
3363 * NEWIDLE balancing is a source of latency, so preemptible kernels
3364 * will stop after the first task is pulled to minimize the critical
3367 if (idle == CPU_NEWLY_IDLE)
3372 * We only want to steal up to the prescribed amount of weighted load.
3374 if (rem_load_move > 0) {
3375 if (p->prio < *this_best_prio)
3376 *this_best_prio = p->prio;
3377 p = iterator->next(iterator->arg);
3382 * Right now, this is one of only two places pull_task() is called,
3383 * so we can safely collect pull_task() stats here rather than
3384 * inside pull_task().
3386 schedstat_add(sd, lb_gained[idle], pulled);
3389 *all_pinned = pinned;
3391 return max_load_move - rem_load_move;
3395 * move_tasks tries to move up to max_load_move weighted load from busiest to
3396 * this_rq, as part of a balancing operation within domain "sd".
3397 * Returns 1 if successful and 0 otherwise.
3399 * Called with both runqueues locked.
3401 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3402 unsigned long max_load_move,
3403 struct sched_domain *sd, enum cpu_idle_type idle,
3406 const struct sched_class *class = sched_class_highest;
3407 unsigned long total_load_moved = 0;
3408 int this_best_prio = this_rq->curr->prio;
3412 class->load_balance(this_rq, this_cpu, busiest,
3413 max_load_move - total_load_moved,
3414 sd, idle, all_pinned, &this_best_prio);
3415 class = class->next;
3417 #ifdef CONFIG_PREEMPT
3419 * NEWIDLE balancing is a source of latency, so preemptible
3420 * kernels will stop after the first task is pulled to minimize
3421 * the critical section.
3423 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3426 } while (class && max_load_move > total_load_moved);
3428 return total_load_moved > 0;
3432 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3433 struct sched_domain *sd, enum cpu_idle_type idle,
3434 struct rq_iterator *iterator)
3436 struct task_struct *p = iterator->start(iterator->arg);
3440 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3441 pull_task(busiest, p, this_rq, this_cpu);
3443 * Right now, this is only the second place pull_task()
3444 * is called, so we can safely collect pull_task()
3445 * stats here rather than inside pull_task().
3447 schedstat_inc(sd, lb_gained[idle]);
3451 p = iterator->next(iterator->arg);
3458 * move_one_task tries to move exactly one task from busiest to this_rq, as
3459 * part of active balancing operations within "domain".
3460 * Returns 1 if successful and 0 otherwise.
3462 * Called with both runqueues locked.
3464 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3465 struct sched_domain *sd, enum cpu_idle_type idle)
3467 const struct sched_class *class;
3469 for_each_class(class) {
3470 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3476 /********** Helpers for find_busiest_group ************************/
3478 * sd_lb_stats - Structure to store the statistics of a sched_domain
3479 * during load balancing.
3481 struct sd_lb_stats {
3482 struct sched_group *busiest; /* Busiest group in this sd */
3483 struct sched_group *this; /* Local group in this sd */
3484 unsigned long total_load; /* Total load of all groups in sd */
3485 unsigned long total_pwr; /* Total power of all groups in sd */
3486 unsigned long avg_load; /* Average load across all groups in sd */
3488 /** Statistics of this group */
3489 unsigned long this_load;
3490 unsigned long this_load_per_task;
3491 unsigned long this_nr_running;
3493 /* Statistics of the busiest group */
3494 unsigned long max_load;
3495 unsigned long busiest_load_per_task;
3496 unsigned long busiest_nr_running;
3497 unsigned long busiest_group_capacity;
3499 int group_imb; /* Is there imbalance in this sd */
3500 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3501 int power_savings_balance; /* Is powersave balance needed for this sd */
3502 struct sched_group *group_min; /* Least loaded group in sd */
3503 struct sched_group *group_leader; /* Group which relieves group_min */
3504 unsigned long min_load_per_task; /* load_per_task in group_min */
3505 unsigned long leader_nr_running; /* Nr running of group_leader */
3506 unsigned long min_nr_running; /* Nr running of group_min */
3511 * sg_lb_stats - stats of a sched_group required for load_balancing
3513 struct sg_lb_stats {
3514 unsigned long avg_load; /*Avg load across the CPUs of the group */
3515 unsigned long group_load; /* Total load over the CPUs of the group */
3516 unsigned long sum_nr_running; /* Nr tasks running in the group */
3517 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3518 unsigned long group_capacity;
3519 int group_imb; /* Is there an imbalance in the group ? */
3523 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3524 * @group: The group whose first cpu is to be returned.
3526 static inline unsigned int group_first_cpu(struct sched_group *group)
3528 return cpumask_first(sched_group_cpus(group));
3532 * get_sd_load_idx - Obtain the load index for a given sched domain.
3533 * @sd: The sched_domain whose load_idx is to be obtained.
3534 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3536 static inline int get_sd_load_idx(struct sched_domain *sd,
3537 enum cpu_idle_type idle)
3543 load_idx = sd->busy_idx;
3546 case CPU_NEWLY_IDLE:
3547 load_idx = sd->newidle_idx;
3550 load_idx = sd->idle_idx;
3558 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3560 * init_sd_power_savings_stats - Initialize power savings statistics for
3561 * the given sched_domain, during load balancing.
3563 * @sd: Sched domain whose power-savings statistics are to be initialized.
3564 * @sds: Variable containing the statistics for sd.
3565 * @idle: Idle status of the CPU at which we're performing load-balancing.
3567 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3568 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3571 * Busy processors will not participate in power savings
3574 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3575 sds->power_savings_balance = 0;
3577 sds->power_savings_balance = 1;
3578 sds->min_nr_running = ULONG_MAX;
3579 sds->leader_nr_running = 0;
3584 * update_sd_power_savings_stats - Update the power saving stats for a
3585 * sched_domain while performing load balancing.
3587 * @group: sched_group belonging to the sched_domain under consideration.
3588 * @sds: Variable containing the statistics of the sched_domain
3589 * @local_group: Does group contain the CPU for which we're performing
3591 * @sgs: Variable containing the statistics of the group.
3593 static inline void update_sd_power_savings_stats(struct sched_group *group,
3594 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3597 if (!sds->power_savings_balance)
3601 * If the local group is idle or completely loaded
3602 * no need to do power savings balance at this domain
3604 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3605 !sds->this_nr_running))
3606 sds->power_savings_balance = 0;
3609 * If a group is already running at full capacity or idle,
3610 * don't include that group in power savings calculations
3612 if (!sds->power_savings_balance ||
3613 sgs->sum_nr_running >= sgs->group_capacity ||
3614 !sgs->sum_nr_running)
3618 * Calculate the group which has the least non-idle load.
3619 * This is the group from where we need to pick up the load
3622 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3623 (sgs->sum_nr_running == sds->min_nr_running &&
3624 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3625 sds->group_min = group;
3626 sds->min_nr_running = sgs->sum_nr_running;
3627 sds->min_load_per_task = sgs->sum_weighted_load /
3628 sgs->sum_nr_running;
3632 * Calculate the group which is almost near its
3633 * capacity but still has some space to pick up some load
3634 * from other group and save more power
3636 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3639 if (sgs->sum_nr_running > sds->leader_nr_running ||
3640 (sgs->sum_nr_running == sds->leader_nr_running &&
3641 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3642 sds->group_leader = group;
3643 sds->leader_nr_running = sgs->sum_nr_running;
3648 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3649 * @sds: Variable containing the statistics of the sched_domain
3650 * under consideration.
3651 * @this_cpu: Cpu at which we're currently performing load-balancing.
3652 * @imbalance: Variable to store the imbalance.
3655 * Check if we have potential to perform some power-savings balance.
3656 * If yes, set the busiest group to be the least loaded group in the
3657 * sched_domain, so that it's CPUs can be put to idle.
3659 * Returns 1 if there is potential to perform power-savings balance.
3662 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3663 int this_cpu, unsigned long *imbalance)
3665 if (!sds->power_savings_balance)
3668 if (sds->this != sds->group_leader ||
3669 sds->group_leader == sds->group_min)
3672 *imbalance = sds->min_load_per_task;
3673 sds->busiest = sds->group_min;
3678 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3679 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3680 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3685 static inline void update_sd_power_savings_stats(struct sched_group *group,
3686 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3691 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3692 int this_cpu, unsigned long *imbalance)
3696 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3699 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3701 return SCHED_LOAD_SCALE;
3704 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3706 return default_scale_freq_power(sd, cpu);
3709 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3711 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3712 unsigned long smt_gain = sd->smt_gain;
3719 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3721 return default_scale_smt_power(sd, cpu);
3724 unsigned long scale_rt_power(int cpu)
3726 struct rq *rq = cpu_rq(cpu);
3727 u64 total, available;
3729 sched_avg_update(rq);
3731 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3732 available = total - rq->rt_avg;
3734 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3735 total = SCHED_LOAD_SCALE;
3737 total >>= SCHED_LOAD_SHIFT;
3739 return div_u64(available, total);
3742 static void update_cpu_power(struct sched_domain *sd, int cpu)
3744 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3745 unsigned long power = SCHED_LOAD_SCALE;
3746 struct sched_group *sdg = sd->groups;
3748 if (sched_feat(ARCH_POWER))
3749 power *= arch_scale_freq_power(sd, cpu);
3751 power *= default_scale_freq_power(sd, cpu);
3753 power >>= SCHED_LOAD_SHIFT;
3755 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3756 if (sched_feat(ARCH_POWER))
3757 power *= arch_scale_smt_power(sd, cpu);
3759 power *= default_scale_smt_power(sd, cpu);
3761 power >>= SCHED_LOAD_SHIFT;
3764 power *= scale_rt_power(cpu);
3765 power >>= SCHED_LOAD_SHIFT;
3770 sdg->cpu_power = power;
3773 static void update_group_power(struct sched_domain *sd, int cpu)
3775 struct sched_domain *child = sd->child;
3776 struct sched_group *group, *sdg = sd->groups;
3777 unsigned long power;
3780 update_cpu_power(sd, cpu);
3786 group = child->groups;
3788 power += group->cpu_power;
3789 group = group->next;
3790 } while (group != child->groups);
3792 sdg->cpu_power = power;
3796 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3797 * @sd: The sched_domain whose statistics are to be updated.
3798 * @group: sched_group whose statistics are to be updated.
3799 * @this_cpu: Cpu for which load balance is currently performed.
3800 * @idle: Idle status of this_cpu
3801 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3802 * @sd_idle: Idle status of the sched_domain containing group.
3803 * @local_group: Does group contain this_cpu.
3804 * @cpus: Set of cpus considered for load balancing.
3805 * @balance: Should we balance.
3806 * @sgs: variable to hold the statistics for this group.
3808 static inline void update_sg_lb_stats(struct sched_domain *sd,
3809 struct sched_group *group, int this_cpu,
3810 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3811 int local_group, const struct cpumask *cpus,
3812 int *balance, struct sg_lb_stats *sgs)
3814 unsigned long load, max_cpu_load, min_cpu_load;
3816 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3817 unsigned long avg_load_per_task = 0;
3820 balance_cpu = group_first_cpu(group);
3821 if (balance_cpu == this_cpu)
3822 update_group_power(sd, this_cpu);
3825 /* Tally up the load of all CPUs in the group */
3827 min_cpu_load = ~0UL;
3829 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3830 struct rq *rq = cpu_rq(i);
3832 if (*sd_idle && rq->nr_running)
3835 /* Bias balancing toward cpus of our domain */
3837 if (idle_cpu(i) && !first_idle_cpu) {
3842 load = target_load(i, load_idx);
3844 load = source_load(i, load_idx);
3845 if (load > max_cpu_load)
3846 max_cpu_load = load;
3847 if (min_cpu_load > load)
3848 min_cpu_load = load;
3851 sgs->group_load += load;
3852 sgs->sum_nr_running += rq->nr_running;
3853 sgs->sum_weighted_load += weighted_cpuload(i);
3858 * First idle cpu or the first cpu(busiest) in this sched group
3859 * is eligible for doing load balancing at this and above
3860 * domains. In the newly idle case, we will allow all the cpu's
3861 * to do the newly idle load balance.
3863 if (idle != CPU_NEWLY_IDLE && local_group &&
3864 balance_cpu != this_cpu && balance) {
3869 /* Adjust by relative CPU power of the group */
3870 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3873 * Consider the group unbalanced when the imbalance is larger
3874 * than the average weight of two tasks.
3876 * APZ: with cgroup the avg task weight can vary wildly and
3877 * might not be a suitable number - should we keep a
3878 * normalized nr_running number somewhere that negates
3881 if (sgs->sum_nr_running)
3882 avg_load_per_task = sgs->sum_weighted_load / sgs->sum_nr_running;
3884 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3887 sgs->group_capacity =
3888 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3892 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3893 * @sd: sched_domain whose statistics are to be updated.
3894 * @this_cpu: Cpu for which load balance is currently performed.
3895 * @idle: Idle status of this_cpu
3896 * @sd_idle: Idle status of the sched_domain containing group.
3897 * @cpus: Set of cpus considered for load balancing.
3898 * @balance: Should we balance.
3899 * @sds: variable to hold the statistics for this sched_domain.
3901 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3902 enum cpu_idle_type idle, int *sd_idle,
3903 const struct cpumask *cpus, int *balance,
3904 struct sd_lb_stats *sds)
3906 struct sched_domain *child = sd->child;
3907 struct sched_group *group = sd->groups;
3908 struct sg_lb_stats sgs;
3909 int load_idx, prefer_sibling = 0;
3911 if (child && child->flags & SD_PREFER_SIBLING)
3914 init_sd_power_savings_stats(sd, sds, idle);
3915 load_idx = get_sd_load_idx(sd, idle);
3920 local_group = cpumask_test_cpu(this_cpu,
3921 sched_group_cpus(group));
3922 memset(&sgs, 0, sizeof(sgs));
3923 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3924 local_group, cpus, balance, &sgs);
3926 if (local_group && balance && !(*balance))
3929 sds->total_load += sgs.group_load;
3930 sds->total_pwr += group->cpu_power;
3933 * In case the child domain prefers tasks go to siblings
3934 * first, lower the group capacity to one so that we'll try
3935 * and move all the excess tasks away.
3938 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3941 sds->this_load = sgs.avg_load;
3943 sds->this_nr_running = sgs.sum_nr_running;
3944 sds->this_load_per_task = sgs.sum_weighted_load;
3945 } else if (sgs.avg_load > sds->max_load &&
3946 (sgs.sum_nr_running > sgs.group_capacity ||
3948 sds->max_load = sgs.avg_load;
3949 sds->busiest = group;
3950 sds->busiest_nr_running = sgs.sum_nr_running;
3951 sds->busiest_group_capacity = sgs.group_capacity;
3952 sds->busiest_load_per_task = sgs.sum_weighted_load;
3953 sds->group_imb = sgs.group_imb;
3956 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3957 group = group->next;
3958 } while (group != sd->groups);
3962 * fix_small_imbalance - Calculate the minor imbalance that exists
3963 * amongst the groups of a sched_domain, during
3965 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3966 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3967 * @imbalance: Variable to store the imbalance.
3969 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3970 int this_cpu, unsigned long *imbalance)
3972 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3973 unsigned int imbn = 2;
3974 unsigned long scaled_busy_load_per_task;
3976 if (sds->this_nr_running) {
3977 sds->this_load_per_task /= sds->this_nr_running;
3978 if (sds->busiest_load_per_task >
3979 sds->this_load_per_task)
3982 sds->this_load_per_task =
3983 cpu_avg_load_per_task(this_cpu);
3985 scaled_busy_load_per_task = sds->busiest_load_per_task
3987 scaled_busy_load_per_task /= sds->busiest->cpu_power;
3989 if (sds->max_load - sds->this_load + scaled_busy_load_per_task >=
3990 (scaled_busy_load_per_task * imbn)) {
3991 *imbalance = sds->busiest_load_per_task;
3996 * OK, we don't have enough imbalance to justify moving tasks,
3997 * however we may be able to increase total CPU power used by
4001 pwr_now += sds->busiest->cpu_power *
4002 min(sds->busiest_load_per_task, sds->max_load);
4003 pwr_now += sds->this->cpu_power *
4004 min(sds->this_load_per_task, sds->this_load);
4005 pwr_now /= SCHED_LOAD_SCALE;
4007 /* Amount of load we'd subtract */
4008 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
4009 sds->busiest->cpu_power;
4010 if (sds->max_load > tmp)
4011 pwr_move += sds->busiest->cpu_power *
4012 min(sds->busiest_load_per_task, sds->max_load - tmp);
4014 /* Amount of load we'd add */
4015 if (sds->max_load * sds->busiest->cpu_power <
4016 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
4017 tmp = (sds->max_load * sds->busiest->cpu_power) /
4018 sds->this->cpu_power;
4020 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
4021 sds->this->cpu_power;
4022 pwr_move += sds->this->cpu_power *
4023 min(sds->this_load_per_task, sds->this_load + tmp);
4024 pwr_move /= SCHED_LOAD_SCALE;
4026 /* Move if we gain throughput */
4027 if (pwr_move > pwr_now)
4028 *imbalance = sds->busiest_load_per_task;
4032 * calculate_imbalance - Calculate the amount of imbalance present within the
4033 * groups of a given sched_domain during load balance.
4034 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4035 * @this_cpu: Cpu for which currently load balance is being performed.
4036 * @imbalance: The variable to store the imbalance.
4038 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
4039 unsigned long *imbalance)
4041 unsigned long max_pull, load_above_capacity = ~0UL;
4043 sds->busiest_load_per_task /= sds->busiest_nr_running;
4044 if (sds->group_imb) {
4045 sds->busiest_load_per_task =
4046 min(sds->busiest_load_per_task, sds->avg_load);
4050 * In the presence of smp nice balancing, certain scenarios can have
4051 * max load less than avg load(as we skip the groups at or below
4052 * its cpu_power, while calculating max_load..)
4054 if (sds->max_load < sds->avg_load) {
4056 return fix_small_imbalance(sds, this_cpu, imbalance);
4059 if (!sds->group_imb) {
4061 * Don't want to pull so many tasks that a group would go idle.
4063 load_above_capacity = (sds->busiest_nr_running -
4064 sds->busiest_group_capacity);
4066 load_above_capacity *= (SCHED_LOAD_SCALE * SCHED_LOAD_SCALE);
4068 load_above_capacity /= sds->busiest->cpu_power;
4072 * We're trying to get all the cpus to the average_load, so we don't
4073 * want to push ourselves above the average load, nor do we wish to
4074 * reduce the max loaded cpu below the average load. At the same time,
4075 * we also don't want to reduce the group load below the group capacity
4076 * (so that we can implement power-savings policies etc). Thus we look
4077 * for the minimum possible imbalance.
4078 * Be careful of negative numbers as they'll appear as very large values
4079 * with unsigned longs.
4081 max_pull = min(sds->max_load - sds->avg_load, load_above_capacity);
4083 /* How much load to actually move to equalise the imbalance */
4084 *imbalance = min(max_pull * sds->busiest->cpu_power,
4085 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
4089 * if *imbalance is less than the average load per runnable task
4090 * there is no gaurantee that any tasks will be moved so we'll have
4091 * a think about bumping its value to force at least one task to be
4094 if (*imbalance < sds->busiest_load_per_task)
4095 return fix_small_imbalance(sds, this_cpu, imbalance);
4098 /******* find_busiest_group() helpers end here *********************/
4101 * find_busiest_group - Returns the busiest group within the sched_domain
4102 * if there is an imbalance. If there isn't an imbalance, and
4103 * the user has opted for power-savings, it returns a group whose
4104 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4105 * such a group exists.
4107 * Also calculates the amount of weighted load which should be moved
4108 * to restore balance.
4110 * @sd: The sched_domain whose busiest group is to be returned.
4111 * @this_cpu: The cpu for which load balancing is currently being performed.
4112 * @imbalance: Variable which stores amount of weighted load which should
4113 * be moved to restore balance/put a group to idle.
4114 * @idle: The idle status of this_cpu.
4115 * @sd_idle: The idleness of sd
4116 * @cpus: The set of CPUs under consideration for load-balancing.
4117 * @balance: Pointer to a variable indicating if this_cpu
4118 * is the appropriate cpu to perform load balancing at this_level.
4120 * Returns: - the busiest group if imbalance exists.
4121 * - If no imbalance and user has opted for power-savings balance,
4122 * return the least loaded group whose CPUs can be
4123 * put to idle by rebalancing its tasks onto our group.
4125 static struct sched_group *
4126 find_busiest_group(struct sched_domain *sd, int this_cpu,
4127 unsigned long *imbalance, enum cpu_idle_type idle,
4128 int *sd_idle, const struct cpumask *cpus, int *balance)
4130 struct sd_lb_stats sds;
4132 memset(&sds, 0, sizeof(sds));
4135 * Compute the various statistics relavent for load balancing at
4138 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4141 /* Cases where imbalance does not exist from POV of this_cpu */
4142 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4144 * 2) There is no busy sibling group to pull from.
4145 * 3) This group is the busiest group.
4146 * 4) This group is more busy than the avg busieness at this
4148 * 5) The imbalance is within the specified limit.
4150 if (balance && !(*balance))
4153 if (!sds.busiest || sds.busiest_nr_running == 0)
4156 if (sds.this_load >= sds.max_load)
4159 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4161 if (sds.this_load >= sds.avg_load)
4164 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4167 /* Looks like there is an imbalance. Compute it */
4168 calculate_imbalance(&sds, this_cpu, imbalance);
4173 * There is no obvious imbalance. But check if we can do some balancing
4176 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4184 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4187 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4188 unsigned long imbalance, const struct cpumask *cpus)
4190 struct rq *busiest = NULL, *rq;
4191 unsigned long max_load = 0;
4194 for_each_cpu(i, sched_group_cpus(group)) {
4195 unsigned long power = power_of(i);
4196 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4199 if (!cpumask_test_cpu(i, cpus))
4203 wl = weighted_cpuload(i);
4206 * When comparing with imbalance, use weighted_cpuload()
4207 * which is not scaled with the cpu power.
4209 if (capacity && rq->nr_running == 1 && wl > imbalance)
4213 * For the load comparisons with the other cpu's, consider
4214 * the weighted_cpuload() scaled with the cpu power, so that
4215 * the load can be moved away from the cpu that is potentially
4216 * running at a lower capacity.
4218 wl = (wl * SCHED_LOAD_SCALE) / power;
4220 if (wl > max_load) {
4230 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4231 * so long as it is large enough.
4233 #define MAX_PINNED_INTERVAL 512
4235 /* Working cpumask for load_balance and load_balance_newidle. */
4236 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4239 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4240 * tasks if there is an imbalance.
4242 static int load_balance(int this_cpu, struct rq *this_rq,
4243 struct sched_domain *sd, enum cpu_idle_type idle,
4246 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4247 struct sched_group *group;
4248 unsigned long imbalance;
4250 unsigned long flags;
4251 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4253 cpumask_copy(cpus, cpu_active_mask);
4256 * When power savings policy is enabled for the parent domain, idle
4257 * sibling can pick up load irrespective of busy siblings. In this case,
4258 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4259 * portraying it as CPU_NOT_IDLE.
4261 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4262 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4265 schedstat_inc(sd, lb_count[idle]);
4269 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4276 schedstat_inc(sd, lb_nobusyg[idle]);
4280 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4282 schedstat_inc(sd, lb_nobusyq[idle]);
4286 BUG_ON(busiest == this_rq);
4288 schedstat_add(sd, lb_imbalance[idle], imbalance);
4291 if (busiest->nr_running > 1) {
4293 * Attempt to move tasks. If find_busiest_group has found
4294 * an imbalance but busiest->nr_running <= 1, the group is
4295 * still unbalanced. ld_moved simply stays zero, so it is
4296 * correctly treated as an imbalance.
4298 local_irq_save(flags);
4299 double_rq_lock(this_rq, busiest);
4300 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4301 imbalance, sd, idle, &all_pinned);
4302 double_rq_unlock(this_rq, busiest);
4303 local_irq_restore(flags);
4306 * some other cpu did the load balance for us.
4308 if (ld_moved && this_cpu != smp_processor_id())
4309 resched_cpu(this_cpu);
4311 /* All tasks on this runqueue were pinned by CPU affinity */
4312 if (unlikely(all_pinned)) {
4313 cpumask_clear_cpu(cpu_of(busiest), cpus);
4314 if (!cpumask_empty(cpus))
4321 schedstat_inc(sd, lb_failed[idle]);
4322 sd->nr_balance_failed++;
4324 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4326 spin_lock_irqsave(&busiest->lock, flags);
4328 /* don't kick the migration_thread, if the curr
4329 * task on busiest cpu can't be moved to this_cpu
4331 if (!cpumask_test_cpu(this_cpu,
4332 &busiest->curr->cpus_allowed)) {
4333 spin_unlock_irqrestore(&busiest->lock, flags);
4335 goto out_one_pinned;
4338 if (!busiest->active_balance) {
4339 busiest->active_balance = 1;
4340 busiest->push_cpu = this_cpu;
4343 spin_unlock_irqrestore(&busiest->lock, flags);
4345 wake_up_process(busiest->migration_thread);
4348 * We've kicked active balancing, reset the failure
4351 sd->nr_balance_failed = sd->cache_nice_tries+1;
4354 sd->nr_balance_failed = 0;
4356 if (likely(!active_balance)) {
4357 /* We were unbalanced, so reset the balancing interval */
4358 sd->balance_interval = sd->min_interval;
4361 * If we've begun active balancing, start to back off. This
4362 * case may not be covered by the all_pinned logic if there
4363 * is only 1 task on the busy runqueue (because we don't call
4366 if (sd->balance_interval < sd->max_interval)
4367 sd->balance_interval *= 2;
4370 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4371 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4377 schedstat_inc(sd, lb_balanced[idle]);
4379 sd->nr_balance_failed = 0;
4382 /* tune up the balancing interval */
4383 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4384 (sd->balance_interval < sd->max_interval))
4385 sd->balance_interval *= 2;
4387 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4388 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4399 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4400 * tasks if there is an imbalance.
4402 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4403 * this_rq is locked.
4406 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4408 struct sched_group *group;
4409 struct rq *busiest = NULL;
4410 unsigned long imbalance;
4414 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4416 cpumask_copy(cpus, cpu_active_mask);
4419 * When power savings policy is enabled for the parent domain, idle
4420 * sibling can pick up load irrespective of busy siblings. In this case,
4421 * let the state of idle sibling percolate up as IDLE, instead of
4422 * portraying it as CPU_NOT_IDLE.
4424 if (sd->flags & SD_SHARE_CPUPOWER &&
4425 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4428 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4430 update_shares_locked(this_rq, sd);
4431 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4432 &sd_idle, cpus, NULL);
4434 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4438 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4440 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4444 BUG_ON(busiest == this_rq);
4446 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4449 if (busiest->nr_running > 1) {
4450 /* Attempt to move tasks */
4451 double_lock_balance(this_rq, busiest);
4452 /* this_rq->clock is already updated */
4453 update_rq_clock(busiest);
4454 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4455 imbalance, sd, CPU_NEWLY_IDLE,
4457 double_unlock_balance(this_rq, busiest);
4459 if (unlikely(all_pinned)) {
4460 cpumask_clear_cpu(cpu_of(busiest), cpus);
4461 if (!cpumask_empty(cpus))
4467 int active_balance = 0;
4469 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4470 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4471 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4474 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4477 if (sd->nr_balance_failed++ < 2)
4481 * The only task running in a non-idle cpu can be moved to this
4482 * cpu in an attempt to completely freeup the other CPU
4483 * package. The same method used to move task in load_balance()
4484 * have been extended for load_balance_newidle() to speedup
4485 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4487 * The package power saving logic comes from
4488 * find_busiest_group(). If there are no imbalance, then
4489 * f_b_g() will return NULL. However when sched_mc={1,2} then
4490 * f_b_g() will select a group from which a running task may be
4491 * pulled to this cpu in order to make the other package idle.
4492 * If there is no opportunity to make a package idle and if
4493 * there are no imbalance, then f_b_g() will return NULL and no
4494 * action will be taken in load_balance_newidle().
4496 * Under normal task pull operation due to imbalance, there
4497 * will be more than one task in the source run queue and
4498 * move_tasks() will succeed. ld_moved will be true and this
4499 * active balance code will not be triggered.
4502 /* Lock busiest in correct order while this_rq is held */
4503 double_lock_balance(this_rq, busiest);
4506 * don't kick the migration_thread, if the curr
4507 * task on busiest cpu can't be moved to this_cpu
4509 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4510 double_unlock_balance(this_rq, busiest);
4515 if (!busiest->active_balance) {
4516 busiest->active_balance = 1;
4517 busiest->push_cpu = this_cpu;
4521 double_unlock_balance(this_rq, busiest);
4523 * Should not call ttwu while holding a rq->lock
4525 spin_unlock(&this_rq->lock);
4527 wake_up_process(busiest->migration_thread);
4528 spin_lock(&this_rq->lock);
4531 sd->nr_balance_failed = 0;
4533 update_shares_locked(this_rq, sd);
4537 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4538 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4539 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4541 sd->nr_balance_failed = 0;
4547 * idle_balance is called by schedule() if this_cpu is about to become
4548 * idle. Attempts to pull tasks from other CPUs.
4550 static void idle_balance(int this_cpu, struct rq *this_rq)
4552 struct sched_domain *sd;
4553 int pulled_task = 0;
4554 unsigned long next_balance = jiffies + HZ;
4556 this_rq->idle_stamp = this_rq->clock;
4558 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4561 for_each_domain(this_cpu, sd) {
4562 unsigned long interval;
4564 if (!(sd->flags & SD_LOAD_BALANCE))
4567 if (sd->flags & SD_BALANCE_NEWIDLE)
4568 /* If we've pulled tasks over stop searching: */
4569 pulled_task = load_balance_newidle(this_cpu, this_rq,
4572 interval = msecs_to_jiffies(sd->balance_interval);
4573 if (time_after(next_balance, sd->last_balance + interval))
4574 next_balance = sd->last_balance + interval;
4576 this_rq->idle_stamp = 0;
4580 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4582 * We are going idle. next_balance may be set based on
4583 * a busy processor. So reset next_balance.
4585 this_rq->next_balance = next_balance;
4590 * active_load_balance is run by migration threads. It pushes running tasks
4591 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4592 * running on each physical CPU where possible, and avoids physical /
4593 * logical imbalances.
4595 * Called with busiest_rq locked.
4597 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4599 int target_cpu = busiest_rq->push_cpu;
4600 struct sched_domain *sd;
4601 struct rq *target_rq;
4603 /* Is there any task to move? */
4604 if (busiest_rq->nr_running <= 1)
4607 target_rq = cpu_rq(target_cpu);
4610 * This condition is "impossible", if it occurs
4611 * we need to fix it. Originally reported by
4612 * Bjorn Helgaas on a 128-cpu setup.
4614 BUG_ON(busiest_rq == target_rq);
4616 /* move a task from busiest_rq to target_rq */
4617 double_lock_balance(busiest_rq, target_rq);
4618 update_rq_clock(busiest_rq);
4619 update_rq_clock(target_rq);
4621 /* Search for an sd spanning us and the target CPU. */
4622 for_each_domain(target_cpu, sd) {
4623 if ((sd->flags & SD_LOAD_BALANCE) &&
4624 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4629 schedstat_inc(sd, alb_count);
4631 if (move_one_task(target_rq, target_cpu, busiest_rq,
4633 schedstat_inc(sd, alb_pushed);
4635 schedstat_inc(sd, alb_failed);
4637 double_unlock_balance(busiest_rq, target_rq);
4642 atomic_t load_balancer;
4643 cpumask_var_t cpu_mask;
4644 cpumask_var_t ilb_grp_nohz_mask;
4645 } nohz ____cacheline_aligned = {
4646 .load_balancer = ATOMIC_INIT(-1),
4649 int get_nohz_load_balancer(void)
4651 return atomic_read(&nohz.load_balancer);
4654 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4656 * lowest_flag_domain - Return lowest sched_domain containing flag.
4657 * @cpu: The cpu whose lowest level of sched domain is to
4659 * @flag: The flag to check for the lowest sched_domain
4660 * for the given cpu.
4662 * Returns the lowest sched_domain of a cpu which contains the given flag.
4664 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4666 struct sched_domain *sd;
4668 for_each_domain(cpu, sd)
4669 if (sd && (sd->flags & flag))
4676 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4677 * @cpu: The cpu whose domains we're iterating over.
4678 * @sd: variable holding the value of the power_savings_sd
4680 * @flag: The flag to filter the sched_domains to be iterated.
4682 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4683 * set, starting from the lowest sched_domain to the highest.
4685 #define for_each_flag_domain(cpu, sd, flag) \
4686 for (sd = lowest_flag_domain(cpu, flag); \
4687 (sd && (sd->flags & flag)); sd = sd->parent)
4690 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4691 * @ilb_group: group to be checked for semi-idleness
4693 * Returns: 1 if the group is semi-idle. 0 otherwise.
4695 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4696 * and atleast one non-idle CPU. This helper function checks if the given
4697 * sched_group is semi-idle or not.
4699 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4701 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4702 sched_group_cpus(ilb_group));
4705 * A sched_group is semi-idle when it has atleast one busy cpu
4706 * and atleast one idle cpu.
4708 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4711 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4717 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4718 * @cpu: The cpu which is nominating a new idle_load_balancer.
4720 * Returns: Returns the id of the idle load balancer if it exists,
4721 * Else, returns >= nr_cpu_ids.
4723 * This algorithm picks the idle load balancer such that it belongs to a
4724 * semi-idle powersavings sched_domain. The idea is to try and avoid
4725 * completely idle packages/cores just for the purpose of idle load balancing
4726 * when there are other idle cpu's which are better suited for that job.
4728 static int find_new_ilb(int cpu)
4730 struct sched_domain *sd;
4731 struct sched_group *ilb_group;
4734 * Have idle load balancer selection from semi-idle packages only
4735 * when power-aware load balancing is enabled
4737 if (!(sched_smt_power_savings || sched_mc_power_savings))
4741 * Optimize for the case when we have no idle CPUs or only one
4742 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4744 if (cpumask_weight(nohz.cpu_mask) < 2)
4747 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4748 ilb_group = sd->groups;
4751 if (is_semi_idle_group(ilb_group))
4752 return cpumask_first(nohz.ilb_grp_nohz_mask);
4754 ilb_group = ilb_group->next;
4756 } while (ilb_group != sd->groups);
4760 return cpumask_first(nohz.cpu_mask);
4762 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4763 static inline int find_new_ilb(int call_cpu)
4765 return cpumask_first(nohz.cpu_mask);
4770 * This routine will try to nominate the ilb (idle load balancing)
4771 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4772 * load balancing on behalf of all those cpus. If all the cpus in the system
4773 * go into this tickless mode, then there will be no ilb owner (as there is
4774 * no need for one) and all the cpus will sleep till the next wakeup event
4777 * For the ilb owner, tick is not stopped. And this tick will be used
4778 * for idle load balancing. ilb owner will still be part of
4781 * While stopping the tick, this cpu will become the ilb owner if there
4782 * is no other owner. And will be the owner till that cpu becomes busy
4783 * or if all cpus in the system stop their ticks at which point
4784 * there is no need for ilb owner.
4786 * When the ilb owner becomes busy, it nominates another owner, during the
4787 * next busy scheduler_tick()
4789 int select_nohz_load_balancer(int stop_tick)
4791 int cpu = smp_processor_id();
4794 cpu_rq(cpu)->in_nohz_recently = 1;
4796 if (!cpu_active(cpu)) {
4797 if (atomic_read(&nohz.load_balancer) != cpu)
4801 * If we are going offline and still the leader,
4804 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4810 cpumask_set_cpu(cpu, nohz.cpu_mask);
4812 /* time for ilb owner also to sleep */
4813 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4814 if (atomic_read(&nohz.load_balancer) == cpu)
4815 atomic_set(&nohz.load_balancer, -1);
4819 if (atomic_read(&nohz.load_balancer) == -1) {
4820 /* make me the ilb owner */
4821 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4823 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4826 if (!(sched_smt_power_savings ||
4827 sched_mc_power_savings))
4830 * Check to see if there is a more power-efficient
4833 new_ilb = find_new_ilb(cpu);
4834 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4835 atomic_set(&nohz.load_balancer, -1);
4836 resched_cpu(new_ilb);
4842 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4845 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4847 if (atomic_read(&nohz.load_balancer) == cpu)
4848 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4855 static DEFINE_SPINLOCK(balancing);
4858 * It checks each scheduling domain to see if it is due to be balanced,
4859 * and initiates a balancing operation if so.
4861 * Balancing parameters are set up in arch_init_sched_domains.
4863 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4866 struct rq *rq = cpu_rq(cpu);
4867 unsigned long interval;
4868 struct sched_domain *sd;
4869 /* Earliest time when we have to do rebalance again */
4870 unsigned long next_balance = jiffies + 60*HZ;
4871 int update_next_balance = 0;
4874 for_each_domain(cpu, sd) {
4875 if (!(sd->flags & SD_LOAD_BALANCE))
4878 interval = sd->balance_interval;
4879 if (idle != CPU_IDLE)
4880 interval *= sd->busy_factor;
4882 /* scale ms to jiffies */
4883 interval = msecs_to_jiffies(interval);
4884 if (unlikely(!interval))
4886 if (interval > HZ*NR_CPUS/10)
4887 interval = HZ*NR_CPUS/10;
4889 need_serialize = sd->flags & SD_SERIALIZE;
4891 if (need_serialize) {
4892 if (!spin_trylock(&balancing))
4896 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4897 if (load_balance(cpu, rq, sd, idle, &balance)) {
4899 * We've pulled tasks over so either we're no
4900 * longer idle, or one of our SMT siblings is
4903 idle = CPU_NOT_IDLE;
4905 sd->last_balance = jiffies;
4908 spin_unlock(&balancing);
4910 if (time_after(next_balance, sd->last_balance + interval)) {
4911 next_balance = sd->last_balance + interval;
4912 update_next_balance = 1;
4916 * Stop the load balance at this level. There is another
4917 * CPU in our sched group which is doing load balancing more
4925 * next_balance will be updated only when there is a need.
4926 * When the cpu is attached to null domain for ex, it will not be
4929 if (likely(update_next_balance))
4930 rq->next_balance = next_balance;
4934 * run_rebalance_domains is triggered when needed from the scheduler tick.
4935 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4936 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4938 static void run_rebalance_domains(struct softirq_action *h)
4940 int this_cpu = smp_processor_id();
4941 struct rq *this_rq = cpu_rq(this_cpu);
4942 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4943 CPU_IDLE : CPU_NOT_IDLE;
4945 rebalance_domains(this_cpu, idle);
4949 * If this cpu is the owner for idle load balancing, then do the
4950 * balancing on behalf of the other idle cpus whose ticks are
4953 if (this_rq->idle_at_tick &&
4954 atomic_read(&nohz.load_balancer) == this_cpu) {
4958 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4959 if (balance_cpu == this_cpu)
4963 * If this cpu gets work to do, stop the load balancing
4964 * work being done for other cpus. Next load
4965 * balancing owner will pick it up.
4970 rebalance_domains(balance_cpu, CPU_IDLE);
4972 rq = cpu_rq(balance_cpu);
4973 if (time_after(this_rq->next_balance, rq->next_balance))
4974 this_rq->next_balance = rq->next_balance;
4980 static inline int on_null_domain(int cpu)
4982 return !rcu_dereference(cpu_rq(cpu)->sd);
4986 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4988 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4989 * idle load balancing owner or decide to stop the periodic load balancing,
4990 * if the whole system is idle.
4992 static inline void trigger_load_balance(struct rq *rq, int cpu)
4996 * If we were in the nohz mode recently and busy at the current
4997 * scheduler tick, then check if we need to nominate new idle
5000 if (rq->in_nohz_recently && !rq->idle_at_tick) {
5001 rq->in_nohz_recently = 0;
5003 if (atomic_read(&nohz.load_balancer) == cpu) {
5004 cpumask_clear_cpu(cpu, nohz.cpu_mask);
5005 atomic_set(&nohz.load_balancer, -1);
5008 if (atomic_read(&nohz.load_balancer) == -1) {
5009 int ilb = find_new_ilb(cpu);
5011 if (ilb < nr_cpu_ids)
5017 * If this cpu is idle and doing idle load balancing for all the
5018 * cpus with ticks stopped, is it time for that to stop?
5020 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
5021 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
5027 * If this cpu is idle and the idle load balancing is done by
5028 * someone else, then no need raise the SCHED_SOFTIRQ
5030 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
5031 cpumask_test_cpu(cpu, nohz.cpu_mask))
5034 /* Don't need to rebalance while attached to NULL domain */
5035 if (time_after_eq(jiffies, rq->next_balance) &&
5036 likely(!on_null_domain(cpu)))
5037 raise_softirq(SCHED_SOFTIRQ);
5040 #else /* CONFIG_SMP */
5043 * on UP we do not need to balance between CPUs:
5045 static inline void idle_balance(int cpu, struct rq *rq)
5051 DEFINE_PER_CPU(struct kernel_stat, kstat);
5053 EXPORT_PER_CPU_SYMBOL(kstat);
5056 * Return any ns on the sched_clock that have not yet been accounted in
5057 * @p in case that task is currently running.
5059 * Called with task_rq_lock() held on @rq.
5061 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
5065 if (task_current(rq, p)) {
5066 update_rq_clock(rq);
5067 ns = rq->clock - p->se.exec_start;
5075 unsigned long long task_delta_exec(struct task_struct *p)
5077 unsigned long flags;
5081 rq = task_rq_lock(p, &flags);
5082 ns = do_task_delta_exec(p, rq);
5083 task_rq_unlock(rq, &flags);
5089 * Return accounted runtime for the task.
5090 * In case the task is currently running, return the runtime plus current's
5091 * pending runtime that have not been accounted yet.
5093 unsigned long long task_sched_runtime(struct task_struct *p)
5095 unsigned long flags;
5099 rq = task_rq_lock(p, &flags);
5100 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5101 task_rq_unlock(rq, &flags);
5107 * Return sum_exec_runtime for the thread group.
5108 * In case the task is currently running, return the sum plus current's
5109 * pending runtime that have not been accounted yet.
5111 * Note that the thread group might have other running tasks as well,
5112 * so the return value not includes other pending runtime that other
5113 * running tasks might have.
5115 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5117 struct task_cputime totals;
5118 unsigned long flags;
5122 rq = task_rq_lock(p, &flags);
5123 thread_group_cputime(p, &totals);
5124 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5125 task_rq_unlock(rq, &flags);
5131 * Account user cpu time to a process.
5132 * @p: the process that the cpu time gets accounted to
5133 * @cputime: the cpu time spent in user space since the last update
5134 * @cputime_scaled: cputime scaled by cpu frequency
5136 void account_user_time(struct task_struct *p, cputime_t cputime,
5137 cputime_t cputime_scaled)
5139 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5142 /* Add user time to process. */
5143 p->utime = cputime_add(p->utime, cputime);
5144 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5145 account_group_user_time(p, cputime);
5147 /* Add user time to cpustat. */
5148 tmp = cputime_to_cputime64(cputime);
5149 if (TASK_NICE(p) > 0)
5150 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5152 cpustat->user = cputime64_add(cpustat->user, tmp);
5154 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5155 /* Account for user time used */
5156 acct_update_integrals(p);
5160 * Account guest cpu time to a process.
5161 * @p: the process that the cpu time gets accounted to
5162 * @cputime: the cpu time spent in virtual machine since the last update
5163 * @cputime_scaled: cputime scaled by cpu frequency
5165 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5166 cputime_t cputime_scaled)
5169 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5171 tmp = cputime_to_cputime64(cputime);
5173 /* Add guest time to process. */
5174 p->utime = cputime_add(p->utime, cputime);
5175 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5176 account_group_user_time(p, cputime);
5177 p->gtime = cputime_add(p->gtime, cputime);
5179 /* Add guest time to cpustat. */
5180 cpustat->user = cputime64_add(cpustat->user, tmp);
5181 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5185 * Account system cpu time to a process.
5186 * @p: the process that the cpu time gets accounted to
5187 * @hardirq_offset: the offset to subtract from hardirq_count()
5188 * @cputime: the cpu time spent in kernel space since the last update
5189 * @cputime_scaled: cputime scaled by cpu frequency
5191 void account_system_time(struct task_struct *p, int hardirq_offset,
5192 cputime_t cputime, cputime_t cputime_scaled)
5194 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5197 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5198 account_guest_time(p, cputime, cputime_scaled);
5202 /* Add system time to process. */
5203 p->stime = cputime_add(p->stime, cputime);
5204 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5205 account_group_system_time(p, cputime);
5207 /* Add system time to cpustat. */
5208 tmp = cputime_to_cputime64(cputime);
5209 if (hardirq_count() - hardirq_offset)
5210 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5211 else if (softirq_count())
5212 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5214 cpustat->system = cputime64_add(cpustat->system, tmp);
5216 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5218 /* Account for system time used */
5219 acct_update_integrals(p);
5223 * Account for involuntary wait time.
5224 * @steal: the cpu time spent in involuntary wait
5226 void account_steal_time(cputime_t cputime)
5228 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5229 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5231 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5235 * Account for idle time.
5236 * @cputime: the cpu time spent in idle wait
5238 void account_idle_time(cputime_t cputime)
5240 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5241 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5242 struct rq *rq = this_rq();
5244 if (atomic_read(&rq->nr_iowait) > 0)
5245 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5247 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5250 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5253 * Account a single tick of cpu time.
5254 * @p: the process that the cpu time gets accounted to
5255 * @user_tick: indicates if the tick is a user or a system tick
5257 void account_process_tick(struct task_struct *p, int user_tick)
5259 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5260 struct rq *rq = this_rq();
5263 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5264 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5265 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5268 account_idle_time(cputime_one_jiffy);
5272 * Account multiple ticks of steal time.
5273 * @p: the process from which the cpu time has been stolen
5274 * @ticks: number of stolen ticks
5276 void account_steal_ticks(unsigned long ticks)
5278 account_steal_time(jiffies_to_cputime(ticks));
5282 * Account multiple ticks of idle time.
5283 * @ticks: number of stolen ticks
5285 void account_idle_ticks(unsigned long ticks)
5287 account_idle_time(jiffies_to_cputime(ticks));
5293 * Use precise platform statistics if available:
5295 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5296 cputime_t task_utime(struct task_struct *p)
5301 cputime_t task_stime(struct task_struct *p)
5306 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5308 struct task_cputime cputime;
5310 thread_group_cputime(p, &cputime);
5312 *ut = cputime.utime;
5313 *st = cputime.stime;
5317 #ifndef nsecs_to_cputime
5318 # define nsecs_to_cputime(__nsecs) \
5319 msecs_to_cputime(div_u64((__nsecs), NSEC_PER_MSEC))
5322 cputime_t task_utime(struct task_struct *p)
5324 cputime_t utime = p->utime, total = utime + p->stime;
5328 * Use CFS's precise accounting:
5330 temp = (u64)nsecs_to_cputime(p->se.sum_exec_runtime);
5334 do_div(temp, total);
5336 utime = (cputime_t)temp;
5338 p->prev_utime = max(p->prev_utime, utime);
5339 return p->prev_utime;
5342 cputime_t task_stime(struct task_struct *p)
5347 * Use CFS's precise accounting. (we subtract utime from
5348 * the total, to make sure the total observed by userspace
5349 * grows monotonically - apps rely on that):
5351 stime = nsecs_to_cputime(p->se.sum_exec_runtime) - task_utime(p);
5354 p->prev_stime = max(p->prev_stime, stime);
5356 return p->prev_stime;
5360 * Must be called with siglock held.
5362 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5364 struct signal_struct *sig = p->signal;
5365 struct task_cputime cputime;
5366 cputime_t rtime, utime, total;
5368 thread_group_cputime(p, &cputime);
5370 total = cputime_add(cputime.utime, cputime.stime);
5371 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5376 temp = (u64)(rtime * cputime.utime);
5377 do_div(temp, total);
5378 utime = (cputime_t)temp;
5382 sig->prev_utime = max(sig->prev_utime, utime);
5383 sig->prev_stime = max(sig->prev_stime,
5384 cputime_sub(rtime, sig->prev_utime));
5386 *ut = sig->prev_utime;
5387 *st = sig->prev_stime;
5391 inline cputime_t task_gtime(struct task_struct *p)
5397 * This function gets called by the timer code, with HZ frequency.
5398 * We call it with interrupts disabled.
5400 * It also gets called by the fork code, when changing the parent's
5403 void scheduler_tick(void)
5405 int cpu = smp_processor_id();
5406 struct rq *rq = cpu_rq(cpu);
5407 struct task_struct *curr = rq->curr;
5411 spin_lock(&rq->lock);
5412 update_rq_clock(rq);
5413 update_cpu_load(rq);
5414 curr->sched_class->task_tick(rq, curr, 0);
5415 spin_unlock(&rq->lock);
5417 perf_event_task_tick(curr, cpu);
5420 rq->idle_at_tick = idle_cpu(cpu);
5421 trigger_load_balance(rq, cpu);
5425 notrace unsigned long get_parent_ip(unsigned long addr)
5427 if (in_lock_functions(addr)) {
5428 addr = CALLER_ADDR2;
5429 if (in_lock_functions(addr))
5430 addr = CALLER_ADDR3;
5435 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5436 defined(CONFIG_PREEMPT_TRACER))
5438 void __kprobes add_preempt_count(int val)
5440 #ifdef CONFIG_DEBUG_PREEMPT
5444 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5447 preempt_count() += val;
5448 #ifdef CONFIG_DEBUG_PREEMPT
5450 * Spinlock count overflowing soon?
5452 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5455 if (preempt_count() == val)
5456 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5458 EXPORT_SYMBOL(add_preempt_count);
5460 void __kprobes sub_preempt_count(int val)
5462 #ifdef CONFIG_DEBUG_PREEMPT
5466 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5469 * Is the spinlock portion underflowing?
5471 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5472 !(preempt_count() & PREEMPT_MASK)))
5476 if (preempt_count() == val)
5477 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5478 preempt_count() -= val;
5480 EXPORT_SYMBOL(sub_preempt_count);
5485 * Print scheduling while atomic bug:
5487 static noinline void __schedule_bug(struct task_struct *prev)
5489 struct pt_regs *regs = get_irq_regs();
5491 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5492 prev->comm, prev->pid, preempt_count());
5494 debug_show_held_locks(prev);
5496 if (irqs_disabled())
5497 print_irqtrace_events(prev);
5506 * Various schedule()-time debugging checks and statistics:
5508 static inline void schedule_debug(struct task_struct *prev)
5511 * Test if we are atomic. Since do_exit() needs to call into
5512 * schedule() atomically, we ignore that path for now.
5513 * Otherwise, whine if we are scheduling when we should not be.
5515 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5516 __schedule_bug(prev);
5518 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5520 schedstat_inc(this_rq(), sched_count);
5521 #ifdef CONFIG_SCHEDSTATS
5522 if (unlikely(prev->lock_depth >= 0)) {
5523 schedstat_inc(this_rq(), bkl_count);
5524 schedstat_inc(prev, sched_info.bkl_count);
5529 static void put_prev_task(struct rq *rq, struct task_struct *p)
5531 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
5533 update_avg(&p->se.avg_running, runtime);
5535 if (p->state == TASK_RUNNING) {
5537 * In order to avoid avg_overlap growing stale when we are
5538 * indeed overlapping and hence not getting put to sleep, grow
5539 * the avg_overlap on preemption.
5541 * We use the average preemption runtime because that
5542 * correlates to the amount of cache footprint a task can
5545 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5546 update_avg(&p->se.avg_overlap, runtime);
5548 update_avg(&p->se.avg_running, 0);
5550 p->sched_class->put_prev_task(rq, p);
5554 * Pick up the highest-prio task:
5556 static inline struct task_struct *
5557 pick_next_task(struct rq *rq)
5559 const struct sched_class *class;
5560 struct task_struct *p;
5563 * Optimization: we know that if all tasks are in
5564 * the fair class we can call that function directly:
5566 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5567 p = fair_sched_class.pick_next_task(rq);
5572 class = sched_class_highest;
5574 p = class->pick_next_task(rq);
5578 * Will never be NULL as the idle class always
5579 * returns a non-NULL p:
5581 class = class->next;
5586 * schedule() is the main scheduler function.
5588 asmlinkage void __sched schedule(void)
5590 struct task_struct *prev, *next;
5591 unsigned long *switch_count;
5597 cpu = smp_processor_id();
5601 switch_count = &prev->nivcsw;
5603 release_kernel_lock(prev);
5604 need_resched_nonpreemptible:
5606 schedule_debug(prev);
5608 if (sched_feat(HRTICK))
5611 spin_lock_irq(&rq->lock);
5612 update_rq_clock(rq);
5613 clear_tsk_need_resched(prev);
5615 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5616 if (unlikely(signal_pending_state(prev->state, prev)))
5617 prev->state = TASK_RUNNING;
5619 deactivate_task(rq, prev, 1);
5620 switch_count = &prev->nvcsw;
5623 pre_schedule(rq, prev);
5625 if (unlikely(!rq->nr_running))
5626 idle_balance(cpu, rq);
5628 put_prev_task(rq, prev);
5629 next = pick_next_task(rq);
5631 if (likely(prev != next)) {
5632 sched_info_switch(prev, next);
5633 perf_event_task_sched_out(prev, next, cpu);
5639 context_switch(rq, prev, next); /* unlocks the rq */
5641 * the context switch might have flipped the stack from under
5642 * us, hence refresh the local variables.
5644 cpu = smp_processor_id();
5647 spin_unlock_irq(&rq->lock);
5651 if (unlikely(reacquire_kernel_lock(current) < 0))
5652 goto need_resched_nonpreemptible;
5654 preempt_enable_no_resched();
5658 EXPORT_SYMBOL(schedule);
5662 * Look out! "owner" is an entirely speculative pointer
5663 * access and not reliable.
5665 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5670 if (!sched_feat(OWNER_SPIN))
5673 #ifdef CONFIG_DEBUG_PAGEALLOC
5675 * Need to access the cpu field knowing that
5676 * DEBUG_PAGEALLOC could have unmapped it if
5677 * the mutex owner just released it and exited.
5679 if (probe_kernel_address(&owner->cpu, cpu))
5686 * Even if the access succeeded (likely case),
5687 * the cpu field may no longer be valid.
5689 if (cpu >= nr_cpumask_bits)
5693 * We need to validate that we can do a
5694 * get_cpu() and that we have the percpu area.
5696 if (!cpu_online(cpu))
5703 * Owner changed, break to re-assess state.
5705 if (lock->owner != owner)
5709 * Is that owner really running on that cpu?
5711 if (task_thread_info(rq->curr) != owner || need_resched())
5721 #ifdef CONFIG_PREEMPT
5723 * this is the entry point to schedule() from in-kernel preemption
5724 * off of preempt_enable. Kernel preemptions off return from interrupt
5725 * occur there and call schedule directly.
5727 asmlinkage void __sched preempt_schedule(void)
5729 struct thread_info *ti = current_thread_info();
5732 * If there is a non-zero preempt_count or interrupts are disabled,
5733 * we do not want to preempt the current task. Just return..
5735 if (likely(ti->preempt_count || irqs_disabled()))
5739 add_preempt_count(PREEMPT_ACTIVE);
5741 sub_preempt_count(PREEMPT_ACTIVE);
5744 * Check again in case we missed a preemption opportunity
5745 * between schedule and now.
5748 } while (need_resched());
5750 EXPORT_SYMBOL(preempt_schedule);
5753 * this is the entry point to schedule() from kernel preemption
5754 * off of irq context.
5755 * Note, that this is called and return with irqs disabled. This will
5756 * protect us against recursive calling from irq.
5758 asmlinkage void __sched preempt_schedule_irq(void)
5760 struct thread_info *ti = current_thread_info();
5762 /* Catch callers which need to be fixed */
5763 BUG_ON(ti->preempt_count || !irqs_disabled());
5766 add_preempt_count(PREEMPT_ACTIVE);
5769 local_irq_disable();
5770 sub_preempt_count(PREEMPT_ACTIVE);
5773 * Check again in case we missed a preemption opportunity
5774 * between schedule and now.
5777 } while (need_resched());
5780 #endif /* CONFIG_PREEMPT */
5782 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5785 return try_to_wake_up(curr->private, mode, wake_flags);
5787 EXPORT_SYMBOL(default_wake_function);
5790 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5791 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5792 * number) then we wake all the non-exclusive tasks and one exclusive task.
5794 * There are circumstances in which we can try to wake a task which has already
5795 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5796 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5798 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5799 int nr_exclusive, int wake_flags, void *key)
5801 wait_queue_t *curr, *next;
5803 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5804 unsigned flags = curr->flags;
5806 if (curr->func(curr, mode, wake_flags, key) &&
5807 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5813 * __wake_up - wake up threads blocked on a waitqueue.
5815 * @mode: which threads
5816 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5817 * @key: is directly passed to the wakeup function
5819 * It may be assumed that this function implies a write memory barrier before
5820 * changing the task state if and only if any tasks are woken up.
5822 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5823 int nr_exclusive, void *key)
5825 unsigned long flags;
5827 spin_lock_irqsave(&q->lock, flags);
5828 __wake_up_common(q, mode, nr_exclusive, 0, key);
5829 spin_unlock_irqrestore(&q->lock, flags);
5831 EXPORT_SYMBOL(__wake_up);
5834 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5836 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5838 __wake_up_common(q, mode, 1, 0, NULL);
5841 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5843 __wake_up_common(q, mode, 1, 0, key);
5847 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5849 * @mode: which threads
5850 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5851 * @key: opaque value to be passed to wakeup targets
5853 * The sync wakeup differs that the waker knows that it will schedule
5854 * away soon, so while the target thread will be woken up, it will not
5855 * be migrated to another CPU - ie. the two threads are 'synchronized'
5856 * with each other. This can prevent needless bouncing between CPUs.
5858 * On UP it can prevent extra preemption.
5860 * It may be assumed that this function implies a write memory barrier before
5861 * changing the task state if and only if any tasks are woken up.
5863 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5864 int nr_exclusive, void *key)
5866 unsigned long flags;
5867 int wake_flags = WF_SYNC;
5872 if (unlikely(!nr_exclusive))
5875 spin_lock_irqsave(&q->lock, flags);
5876 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5877 spin_unlock_irqrestore(&q->lock, flags);
5879 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5882 * __wake_up_sync - see __wake_up_sync_key()
5884 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5886 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5888 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5891 * complete: - signals a single thread waiting on this completion
5892 * @x: holds the state of this particular completion
5894 * This will wake up a single thread waiting on this completion. Threads will be
5895 * awakened in the same order in which they were queued.
5897 * See also complete_all(), wait_for_completion() and related routines.
5899 * It may be assumed that this function implies a write memory barrier before
5900 * changing the task state if and only if any tasks are woken up.
5902 void complete(struct completion *x)
5904 unsigned long flags;
5906 spin_lock_irqsave(&x->wait.lock, flags);
5908 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5909 spin_unlock_irqrestore(&x->wait.lock, flags);
5911 EXPORT_SYMBOL(complete);
5914 * complete_all: - signals all threads waiting on this completion
5915 * @x: holds the state of this particular completion
5917 * This will wake up all threads waiting on this particular completion event.
5919 * It may be assumed that this function implies a write memory barrier before
5920 * changing the task state if and only if any tasks are woken up.
5922 void complete_all(struct completion *x)
5924 unsigned long flags;
5926 spin_lock_irqsave(&x->wait.lock, flags);
5927 x->done += UINT_MAX/2;
5928 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5929 spin_unlock_irqrestore(&x->wait.lock, flags);
5931 EXPORT_SYMBOL(complete_all);
5933 static inline long __sched
5934 do_wait_for_common(struct completion *x, long timeout, int state)
5937 DECLARE_WAITQUEUE(wait, current);
5939 wait.flags |= WQ_FLAG_EXCLUSIVE;
5940 __add_wait_queue_tail(&x->wait, &wait);
5942 if (signal_pending_state(state, current)) {
5943 timeout = -ERESTARTSYS;
5946 __set_current_state(state);
5947 spin_unlock_irq(&x->wait.lock);
5948 timeout = schedule_timeout(timeout);
5949 spin_lock_irq(&x->wait.lock);
5950 } while (!x->done && timeout);
5951 __remove_wait_queue(&x->wait, &wait);
5956 return timeout ?: 1;
5960 wait_for_common(struct completion *x, long timeout, int state)
5964 spin_lock_irq(&x->wait.lock);
5965 timeout = do_wait_for_common(x, timeout, state);
5966 spin_unlock_irq(&x->wait.lock);
5971 * wait_for_completion: - waits for completion of a task
5972 * @x: holds the state of this particular completion
5974 * This waits to be signaled for completion of a specific task. It is NOT
5975 * interruptible and there is no timeout.
5977 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5978 * and interrupt capability. Also see complete().
5980 void __sched wait_for_completion(struct completion *x)
5982 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5984 EXPORT_SYMBOL(wait_for_completion);
5987 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5988 * @x: holds the state of this particular completion
5989 * @timeout: timeout value in jiffies
5991 * This waits for either a completion of a specific task to be signaled or for a
5992 * specified timeout to expire. The timeout is in jiffies. It is not
5995 unsigned long __sched
5996 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5998 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
6000 EXPORT_SYMBOL(wait_for_completion_timeout);
6003 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
6004 * @x: holds the state of this particular completion
6006 * This waits for completion of a specific task to be signaled. It is
6009 int __sched wait_for_completion_interruptible(struct completion *x)
6011 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
6012 if (t == -ERESTARTSYS)
6016 EXPORT_SYMBOL(wait_for_completion_interruptible);
6019 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
6020 * @x: holds the state of this particular completion
6021 * @timeout: timeout value in jiffies
6023 * This waits for either a completion of a specific task to be signaled or for a
6024 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
6026 unsigned long __sched
6027 wait_for_completion_interruptible_timeout(struct completion *x,
6028 unsigned long timeout)
6030 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
6032 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
6035 * wait_for_completion_killable: - waits for completion of a task (killable)
6036 * @x: holds the state of this particular completion
6038 * This waits to be signaled for completion of a specific task. It can be
6039 * interrupted by a kill signal.
6041 int __sched wait_for_completion_killable(struct completion *x)
6043 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
6044 if (t == -ERESTARTSYS)
6048 EXPORT_SYMBOL(wait_for_completion_killable);
6051 * try_wait_for_completion - try to decrement a completion without blocking
6052 * @x: completion structure
6054 * Returns: 0 if a decrement cannot be done without blocking
6055 * 1 if a decrement succeeded.
6057 * If a completion is being used as a counting completion,
6058 * attempt to decrement the counter without blocking. This
6059 * enables us to avoid waiting if the resource the completion
6060 * is protecting is not available.
6062 bool try_wait_for_completion(struct completion *x)
6064 unsigned long flags;
6067 spin_lock_irqsave(&x->wait.lock, flags);
6072 spin_unlock_irqrestore(&x->wait.lock, flags);
6075 EXPORT_SYMBOL(try_wait_for_completion);
6078 * completion_done - Test to see if a completion has any waiters
6079 * @x: completion structure
6081 * Returns: 0 if there are waiters (wait_for_completion() in progress)
6082 * 1 if there are no waiters.
6085 bool completion_done(struct completion *x)
6087 unsigned long flags;
6090 spin_lock_irqsave(&x->wait.lock, flags);
6093 spin_unlock_irqrestore(&x->wait.lock, flags);
6096 EXPORT_SYMBOL(completion_done);
6099 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
6101 unsigned long flags;
6104 init_waitqueue_entry(&wait, current);
6106 __set_current_state(state);
6108 spin_lock_irqsave(&q->lock, flags);
6109 __add_wait_queue(q, &wait);
6110 spin_unlock(&q->lock);
6111 timeout = schedule_timeout(timeout);
6112 spin_lock_irq(&q->lock);
6113 __remove_wait_queue(q, &wait);
6114 spin_unlock_irqrestore(&q->lock, flags);
6119 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6121 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6123 EXPORT_SYMBOL(interruptible_sleep_on);
6126 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6128 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6130 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6132 void __sched sleep_on(wait_queue_head_t *q)
6134 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6136 EXPORT_SYMBOL(sleep_on);
6138 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6140 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6142 EXPORT_SYMBOL(sleep_on_timeout);
6144 #ifdef CONFIG_RT_MUTEXES
6147 * rt_mutex_setprio - set the current priority of a task
6149 * @prio: prio value (kernel-internal form)
6151 * This function changes the 'effective' priority of a task. It does
6152 * not touch ->normal_prio like __setscheduler().
6154 * Used by the rt_mutex code to implement priority inheritance logic.
6156 void rt_mutex_setprio(struct task_struct *p, int prio)
6158 unsigned long flags;
6159 int oldprio, on_rq, running;
6161 const struct sched_class *prev_class;
6163 BUG_ON(prio < 0 || prio > MAX_PRIO);
6165 rq = task_rq_lock(p, &flags);
6166 update_rq_clock(rq);
6169 prev_class = p->sched_class;
6170 on_rq = p->se.on_rq;
6171 running = task_current(rq, p);
6173 dequeue_task(rq, p, 0);
6175 p->sched_class->put_prev_task(rq, p);
6178 p->sched_class = &rt_sched_class;
6180 p->sched_class = &fair_sched_class;
6185 p->sched_class->set_curr_task(rq);
6187 enqueue_task(rq, p, 0, oldprio < prio);
6189 check_class_changed(rq, p, prev_class, oldprio, running);
6191 task_rq_unlock(rq, &flags);
6196 void set_user_nice(struct task_struct *p, long nice)
6198 int old_prio, delta, on_rq;
6199 unsigned long flags;
6202 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6205 * We have to be careful, if called from sys_setpriority(),
6206 * the task might be in the middle of scheduling on another CPU.
6208 rq = task_rq_lock(p, &flags);
6209 update_rq_clock(rq);
6211 * The RT priorities are set via sched_setscheduler(), but we still
6212 * allow the 'normal' nice value to be set - but as expected
6213 * it wont have any effect on scheduling until the task is
6214 * SCHED_FIFO/SCHED_RR:
6216 if (task_has_rt_policy(p)) {
6217 p->static_prio = NICE_TO_PRIO(nice);
6220 on_rq = p->se.on_rq;
6222 dequeue_task(rq, p, 0);
6224 p->static_prio = NICE_TO_PRIO(nice);
6227 p->prio = effective_prio(p);
6228 delta = p->prio - old_prio;
6231 enqueue_task(rq, p, 0, false);
6233 * If the task increased its priority or is running and
6234 * lowered its priority, then reschedule its CPU:
6236 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6237 resched_task(rq->curr);
6240 task_rq_unlock(rq, &flags);
6242 EXPORT_SYMBOL(set_user_nice);
6245 * can_nice - check if a task can reduce its nice value
6249 int can_nice(const struct task_struct *p, const int nice)
6251 /* convert nice value [19,-20] to rlimit style value [1,40] */
6252 int nice_rlim = 20 - nice;
6254 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6255 capable(CAP_SYS_NICE));
6258 #ifdef __ARCH_WANT_SYS_NICE
6261 * sys_nice - change the priority of the current process.
6262 * @increment: priority increment
6264 * sys_setpriority is a more generic, but much slower function that
6265 * does similar things.
6267 SYSCALL_DEFINE1(nice, int, increment)
6272 * Setpriority might change our priority at the same moment.
6273 * We don't have to worry. Conceptually one call occurs first
6274 * and we have a single winner.
6276 if (increment < -40)
6281 nice = TASK_NICE(current) + increment;
6287 if (increment < 0 && !can_nice(current, nice))
6290 retval = security_task_setnice(current, nice);
6294 set_user_nice(current, nice);
6301 * task_prio - return the priority value of a given task.
6302 * @p: the task in question.
6304 * This is the priority value as seen by users in /proc.
6305 * RT tasks are offset by -200. Normal tasks are centered
6306 * around 0, value goes from -16 to +15.
6308 int task_prio(const struct task_struct *p)
6310 return p->prio - MAX_RT_PRIO;
6314 * task_nice - return the nice value of a given task.
6315 * @p: the task in question.
6317 int task_nice(const struct task_struct *p)
6319 return TASK_NICE(p);
6321 EXPORT_SYMBOL(task_nice);
6324 * idle_cpu - is a given cpu idle currently?
6325 * @cpu: the processor in question.
6327 int idle_cpu(int cpu)
6329 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6333 * idle_task - return the idle task for a given cpu.
6334 * @cpu: the processor in question.
6336 struct task_struct *idle_task(int cpu)
6338 return cpu_rq(cpu)->idle;
6342 * find_process_by_pid - find a process with a matching PID value.
6343 * @pid: the pid in question.
6345 static struct task_struct *find_process_by_pid(pid_t pid)
6347 return pid ? find_task_by_vpid(pid) : current;
6350 /* Actually do priority change: must hold rq lock. */
6352 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6354 BUG_ON(p->se.on_rq);
6357 switch (p->policy) {
6361 p->sched_class = &fair_sched_class;
6365 p->sched_class = &rt_sched_class;
6369 p->rt_priority = prio;
6370 p->normal_prio = normal_prio(p);
6371 /* we are holding p->pi_lock already */
6372 p->prio = rt_mutex_getprio(p);
6377 * check the target process has a UID that matches the current process's
6379 static bool check_same_owner(struct task_struct *p)
6381 const struct cred *cred = current_cred(), *pcred;
6385 pcred = __task_cred(p);
6386 match = (cred->euid == pcred->euid ||
6387 cred->euid == pcred->uid);
6392 static int __sched_setscheduler(struct task_struct *p, int policy,
6393 struct sched_param *param, bool user)
6395 int retval, oldprio, oldpolicy = -1, on_rq, running;
6396 unsigned long flags;
6397 const struct sched_class *prev_class;
6401 /* may grab non-irq protected spin_locks */
6402 BUG_ON(in_interrupt());
6404 /* double check policy once rq lock held */
6406 reset_on_fork = p->sched_reset_on_fork;
6407 policy = oldpolicy = p->policy;
6409 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6410 policy &= ~SCHED_RESET_ON_FORK;
6412 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6413 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6414 policy != SCHED_IDLE)
6419 * Valid priorities for SCHED_FIFO and SCHED_RR are
6420 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6421 * SCHED_BATCH and SCHED_IDLE is 0.
6423 if (param->sched_priority < 0 ||
6424 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6425 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6427 if (rt_policy(policy) != (param->sched_priority != 0))
6431 * Allow unprivileged RT tasks to decrease priority:
6433 if (user && !capable(CAP_SYS_NICE)) {
6434 if (rt_policy(policy)) {
6435 unsigned long rlim_rtprio;
6437 if (!lock_task_sighand(p, &flags))
6439 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6440 unlock_task_sighand(p, &flags);
6442 /* can't set/change the rt policy */
6443 if (policy != p->policy && !rlim_rtprio)
6446 /* can't increase priority */
6447 if (param->sched_priority > p->rt_priority &&
6448 param->sched_priority > rlim_rtprio)
6452 * Like positive nice levels, dont allow tasks to
6453 * move out of SCHED_IDLE either:
6455 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6458 /* can't change other user's priorities */
6459 if (!check_same_owner(p))
6462 /* Normal users shall not reset the sched_reset_on_fork flag */
6463 if (p->sched_reset_on_fork && !reset_on_fork)
6468 #ifdef CONFIG_RT_GROUP_SCHED
6470 * Do not allow realtime tasks into groups that have no runtime
6473 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6474 task_group(p)->rt_bandwidth.rt_runtime == 0)
6478 retval = security_task_setscheduler(p, policy, param);
6484 * make sure no PI-waiters arrive (or leave) while we are
6485 * changing the priority of the task:
6487 spin_lock_irqsave(&p->pi_lock, flags);
6489 * To be able to change p->policy safely, the apropriate
6490 * runqueue lock must be held.
6492 rq = __task_rq_lock(p);
6493 /* recheck policy now with rq lock held */
6494 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6495 policy = oldpolicy = -1;
6496 __task_rq_unlock(rq);
6497 spin_unlock_irqrestore(&p->pi_lock, flags);
6500 update_rq_clock(rq);
6501 on_rq = p->se.on_rq;
6502 running = task_current(rq, p);
6504 deactivate_task(rq, p, 0);
6506 p->sched_class->put_prev_task(rq, p);
6508 p->sched_reset_on_fork = reset_on_fork;
6511 prev_class = p->sched_class;
6512 __setscheduler(rq, p, policy, param->sched_priority);
6515 p->sched_class->set_curr_task(rq);
6517 activate_task(rq, p, 0);
6519 check_class_changed(rq, p, prev_class, oldprio, running);
6521 __task_rq_unlock(rq);
6522 spin_unlock_irqrestore(&p->pi_lock, flags);
6524 rt_mutex_adjust_pi(p);
6530 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6531 * @p: the task in question.
6532 * @policy: new policy.
6533 * @param: structure containing the new RT priority.
6535 * NOTE that the task may be already dead.
6537 int sched_setscheduler(struct task_struct *p, int policy,
6538 struct sched_param *param)
6540 return __sched_setscheduler(p, policy, param, true);
6542 EXPORT_SYMBOL_GPL(sched_setscheduler);
6545 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6546 * @p: the task in question.
6547 * @policy: new policy.
6548 * @param: structure containing the new RT priority.
6550 * Just like sched_setscheduler, only don't bother checking if the
6551 * current context has permission. For example, this is needed in
6552 * stop_machine(): we create temporary high priority worker threads,
6553 * but our caller might not have that capability.
6555 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6556 struct sched_param *param)
6558 return __sched_setscheduler(p, policy, param, false);
6562 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6564 struct sched_param lparam;
6565 struct task_struct *p;
6568 if (!param || pid < 0)
6570 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6575 p = find_process_by_pid(pid);
6577 retval = sched_setscheduler(p, policy, &lparam);
6584 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6585 * @pid: the pid in question.
6586 * @policy: new policy.
6587 * @param: structure containing the new RT priority.
6589 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6590 struct sched_param __user *, param)
6592 /* negative values for policy are not valid */
6596 return do_sched_setscheduler(pid, policy, param);
6600 * sys_sched_setparam - set/change the RT priority of a thread
6601 * @pid: the pid in question.
6602 * @param: structure containing the new RT priority.
6604 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6606 return do_sched_setscheduler(pid, -1, param);
6610 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6611 * @pid: the pid in question.
6613 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6615 struct task_struct *p;
6623 p = find_process_by_pid(pid);
6625 retval = security_task_getscheduler(p);
6628 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6635 * sys_sched_getparam - get the RT priority of a thread
6636 * @pid: the pid in question.
6637 * @param: structure containing the RT priority.
6639 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6641 struct sched_param lp;
6642 struct task_struct *p;
6645 if (!param || pid < 0)
6649 p = find_process_by_pid(pid);
6654 retval = security_task_getscheduler(p);
6658 lp.sched_priority = p->rt_priority;
6662 * This one might sleep, we cannot do it with a spinlock held ...
6664 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6673 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6675 cpumask_var_t cpus_allowed, new_mask;
6676 struct task_struct *p;
6682 p = find_process_by_pid(pid);
6689 /* Prevent p going away */
6693 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6697 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6699 goto out_free_cpus_allowed;
6702 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6705 retval = security_task_setscheduler(p, 0, NULL);
6709 cpuset_cpus_allowed(p, cpus_allowed);
6710 cpumask_and(new_mask, in_mask, cpus_allowed);
6712 retval = set_cpus_allowed_ptr(p, new_mask);
6715 cpuset_cpus_allowed(p, cpus_allowed);
6716 if (!cpumask_subset(new_mask, cpus_allowed)) {
6718 * We must have raced with a concurrent cpuset
6719 * update. Just reset the cpus_allowed to the
6720 * cpuset's cpus_allowed
6722 cpumask_copy(new_mask, cpus_allowed);
6727 free_cpumask_var(new_mask);
6728 out_free_cpus_allowed:
6729 free_cpumask_var(cpus_allowed);
6736 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6737 struct cpumask *new_mask)
6739 if (len < cpumask_size())
6740 cpumask_clear(new_mask);
6741 else if (len > cpumask_size())
6742 len = cpumask_size();
6744 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6748 * sys_sched_setaffinity - set the cpu affinity of a process
6749 * @pid: pid of the process
6750 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6751 * @user_mask_ptr: user-space pointer to the new cpu mask
6753 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6754 unsigned long __user *, user_mask_ptr)
6756 cpumask_var_t new_mask;
6759 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6762 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6764 retval = sched_setaffinity(pid, new_mask);
6765 free_cpumask_var(new_mask);
6769 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6771 struct task_struct *p;
6772 unsigned long flags;
6780 p = find_process_by_pid(pid);
6784 retval = security_task_getscheduler(p);
6788 rq = task_rq_lock(p, &flags);
6789 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6790 task_rq_unlock(rq, &flags);
6800 * sys_sched_getaffinity - get the cpu affinity of a process
6801 * @pid: pid of the process
6802 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6803 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6805 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6806 unsigned long __user *, user_mask_ptr)
6811 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
6813 if (len & (sizeof(unsigned long)-1))
6816 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6819 ret = sched_getaffinity(pid, mask);
6821 size_t retlen = min_t(size_t, len, cpumask_size());
6823 if (copy_to_user(user_mask_ptr, mask, retlen))
6828 free_cpumask_var(mask);
6834 * sys_sched_yield - yield the current processor to other threads.
6836 * This function yields the current CPU to other tasks. If there are no
6837 * other threads running on this CPU then this function will return.
6839 SYSCALL_DEFINE0(sched_yield)
6841 struct rq *rq = this_rq_lock();
6843 schedstat_inc(rq, yld_count);
6844 current->sched_class->yield_task(rq);
6847 * Since we are going to call schedule() anyway, there's
6848 * no need to preempt or enable interrupts:
6850 __release(rq->lock);
6851 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6852 _raw_spin_unlock(&rq->lock);
6853 preempt_enable_no_resched();
6860 static inline int should_resched(void)
6862 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6865 static void __cond_resched(void)
6867 add_preempt_count(PREEMPT_ACTIVE);
6869 sub_preempt_count(PREEMPT_ACTIVE);
6872 int __sched _cond_resched(void)
6874 if (should_resched()) {
6880 EXPORT_SYMBOL(_cond_resched);
6883 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6884 * call schedule, and on return reacquire the lock.
6886 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6887 * operations here to prevent schedule() from being called twice (once via
6888 * spin_unlock(), once by hand).
6890 int __cond_resched_lock(spinlock_t *lock)
6892 int resched = should_resched();
6895 lockdep_assert_held(lock);
6897 if (spin_needbreak(lock) || resched) {
6908 EXPORT_SYMBOL(__cond_resched_lock);
6910 int __sched __cond_resched_softirq(void)
6912 BUG_ON(!in_softirq());
6914 if (should_resched()) {
6922 EXPORT_SYMBOL(__cond_resched_softirq);
6925 * yield - yield the current processor to other threads.
6927 * This is a shortcut for kernel-space yielding - it marks the
6928 * thread runnable and calls sys_sched_yield().
6930 void __sched yield(void)
6932 set_current_state(TASK_RUNNING);
6935 EXPORT_SYMBOL(yield);
6938 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6939 * that process accounting knows that this is a task in IO wait state.
6941 void __sched io_schedule(void)
6943 struct rq *rq = raw_rq();
6945 delayacct_blkio_start();
6946 atomic_inc(&rq->nr_iowait);
6947 current->in_iowait = 1;
6949 current->in_iowait = 0;
6950 atomic_dec(&rq->nr_iowait);
6951 delayacct_blkio_end();
6953 EXPORT_SYMBOL(io_schedule);
6955 long __sched io_schedule_timeout(long timeout)
6957 struct rq *rq = raw_rq();
6960 delayacct_blkio_start();
6961 atomic_inc(&rq->nr_iowait);
6962 current->in_iowait = 1;
6963 ret = schedule_timeout(timeout);
6964 current->in_iowait = 0;
6965 atomic_dec(&rq->nr_iowait);
6966 delayacct_blkio_end();
6971 * sys_sched_get_priority_max - return maximum RT priority.
6972 * @policy: scheduling class.
6974 * this syscall returns the maximum rt_priority that can be used
6975 * by a given scheduling class.
6977 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6984 ret = MAX_USER_RT_PRIO-1;
6996 * sys_sched_get_priority_min - return minimum RT priority.
6997 * @policy: scheduling class.
6999 * this syscall returns the minimum rt_priority that can be used
7000 * by a given scheduling class.
7002 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
7020 * sys_sched_rr_get_interval - return the default timeslice of a process.
7021 * @pid: pid of the process.
7022 * @interval: userspace pointer to the timeslice value.
7024 * this syscall writes the default timeslice value of a given process
7025 * into the user-space timespec buffer. A value of '0' means infinity.
7027 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
7028 struct timespec __user *, interval)
7030 struct task_struct *p;
7031 unsigned int time_slice;
7032 unsigned long flags;
7042 p = find_process_by_pid(pid);
7046 retval = security_task_getscheduler(p);
7050 rq = task_rq_lock(p, &flags);
7051 time_slice = p->sched_class->get_rr_interval(rq, p);
7052 task_rq_unlock(rq, &flags);
7055 jiffies_to_timespec(time_slice, &t);
7056 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
7064 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
7066 void sched_show_task(struct task_struct *p)
7068 unsigned long free = 0;
7071 state = p->state ? __ffs(p->state) + 1 : 0;
7072 printk(KERN_INFO "%-13.13s %c", p->comm,
7073 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
7074 #if BITS_PER_LONG == 32
7075 if (state == TASK_RUNNING)
7076 printk(KERN_CONT " running ");
7078 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
7080 if (state == TASK_RUNNING)
7081 printk(KERN_CONT " running task ");
7083 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
7085 #ifdef CONFIG_DEBUG_STACK_USAGE
7086 free = stack_not_used(p);
7088 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
7089 task_pid_nr(p), task_pid_nr(p->real_parent),
7090 (unsigned long)task_thread_info(p)->flags);
7092 show_stack(p, NULL);
7095 void show_state_filter(unsigned long state_filter)
7097 struct task_struct *g, *p;
7099 #if BITS_PER_LONG == 32
7101 " task PC stack pid father\n");
7104 " task PC stack pid father\n");
7106 read_lock(&tasklist_lock);
7107 do_each_thread(g, p) {
7109 * reset the NMI-timeout, listing all files on a slow
7110 * console might take alot of time:
7112 touch_nmi_watchdog();
7113 if (!state_filter || (p->state & state_filter))
7115 } while_each_thread(g, p);
7117 touch_all_softlockup_watchdogs();
7119 #ifdef CONFIG_SCHED_DEBUG
7120 sysrq_sched_debug_show();
7122 read_unlock(&tasklist_lock);
7124 * Only show locks if all tasks are dumped:
7126 if (state_filter == -1)
7127 debug_show_all_locks();
7130 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7132 idle->sched_class = &idle_sched_class;
7136 * init_idle - set up an idle thread for a given CPU
7137 * @idle: task in question
7138 * @cpu: cpu the idle task belongs to
7140 * NOTE: this function does not set the idle thread's NEED_RESCHED
7141 * flag, to make booting more robust.
7143 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7145 struct rq *rq = cpu_rq(cpu);
7146 unsigned long flags;
7148 spin_lock_irqsave(&rq->lock, flags);
7151 idle->state = TASK_RUNNING;
7152 idle->se.exec_start = sched_clock();
7154 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7155 __set_task_cpu(idle, cpu);
7157 rq->curr = rq->idle = idle;
7158 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7161 spin_unlock_irqrestore(&rq->lock, flags);
7163 /* Set the preempt count _outside_ the spinlocks! */
7164 #if defined(CONFIG_PREEMPT)
7165 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7167 task_thread_info(idle)->preempt_count = 0;
7170 * The idle tasks have their own, simple scheduling class:
7172 idle->sched_class = &idle_sched_class;
7173 ftrace_graph_init_task(idle);
7177 * In a system that switches off the HZ timer nohz_cpu_mask
7178 * indicates which cpus entered this state. This is used
7179 * in the rcu update to wait only for active cpus. For system
7180 * which do not switch off the HZ timer nohz_cpu_mask should
7181 * always be CPU_BITS_NONE.
7183 cpumask_var_t nohz_cpu_mask;
7186 * Increase the granularity value when there are more CPUs,
7187 * because with more CPUs the 'effective latency' as visible
7188 * to users decreases. But the relationship is not linear,
7189 * so pick a second-best guess by going with the log2 of the
7192 * This idea comes from the SD scheduler of Con Kolivas:
7194 static void update_sysctl(void)
7196 unsigned int cpus = min(num_online_cpus(), 8U);
7197 unsigned int factor = 1 + ilog2(cpus);
7199 #define SET_SYSCTL(name) \
7200 (sysctl_##name = (factor) * normalized_sysctl_##name)
7201 SET_SYSCTL(sched_min_granularity);
7202 SET_SYSCTL(sched_latency);
7203 SET_SYSCTL(sched_wakeup_granularity);
7204 SET_SYSCTL(sched_shares_ratelimit);
7208 static inline void sched_init_granularity(void)
7215 * This is how migration works:
7217 * 1) we queue a struct migration_req structure in the source CPU's
7218 * runqueue and wake up that CPU's migration thread.
7219 * 2) we down() the locked semaphore => thread blocks.
7220 * 3) migration thread wakes up (implicitly it forces the migrated
7221 * thread off the CPU)
7222 * 4) it gets the migration request and checks whether the migrated
7223 * task is still in the wrong runqueue.
7224 * 5) if it's in the wrong runqueue then the migration thread removes
7225 * it and puts it into the right queue.
7226 * 6) migration thread up()s the semaphore.
7227 * 7) we wake up and the migration is done.
7231 * Change a given task's CPU affinity. Migrate the thread to a
7232 * proper CPU and schedule it away if the CPU it's executing on
7233 * is removed from the allowed bitmask.
7235 * NOTE: the caller must have a valid reference to the task, the
7236 * task must not exit() & deallocate itself prematurely. The
7237 * call is not atomic; no spinlocks may be held.
7239 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7241 struct migration_req req;
7242 unsigned long flags;
7246 rq = task_rq_lock(p, &flags);
7248 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7253 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7254 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7259 if (p->sched_class->set_cpus_allowed)
7260 p->sched_class->set_cpus_allowed(p, new_mask);
7262 cpumask_copy(&p->cpus_allowed, new_mask);
7263 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7266 /* Can the task run on the task's current CPU? If so, we're done */
7267 if (cpumask_test_cpu(task_cpu(p), new_mask))
7270 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7271 /* Need help from migration thread: drop lock and wait. */
7272 struct task_struct *mt = rq->migration_thread;
7274 get_task_struct(mt);
7275 task_rq_unlock(rq, &flags);
7276 wake_up_process(mt);
7277 put_task_struct(mt);
7278 wait_for_completion(&req.done);
7279 tlb_migrate_finish(p->mm);
7283 task_rq_unlock(rq, &flags);
7287 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7290 * Move (not current) task off this cpu, onto dest cpu. We're doing
7291 * this because either it can't run here any more (set_cpus_allowed()
7292 * away from this CPU, or CPU going down), or because we're
7293 * attempting to rebalance this task on exec (sched_exec).
7295 * So we race with normal scheduler movements, but that's OK, as long
7296 * as the task is no longer on this CPU.
7298 * Returns non-zero if task was successfully migrated.
7300 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7302 struct rq *rq_dest, *rq_src;
7305 if (unlikely(!cpu_active(dest_cpu)))
7308 rq_src = cpu_rq(src_cpu);
7309 rq_dest = cpu_rq(dest_cpu);
7311 double_rq_lock(rq_src, rq_dest);
7312 /* Already moved. */
7313 if (task_cpu(p) != src_cpu)
7315 /* Affinity changed (again). */
7316 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7320 * If we're not on a rq, the next wake-up will ensure we're
7324 deactivate_task(rq_src, p, 0);
7325 set_task_cpu(p, dest_cpu);
7326 activate_task(rq_dest, p, 0);
7327 check_preempt_curr(rq_dest, p, 0);
7332 double_rq_unlock(rq_src, rq_dest);
7336 #define RCU_MIGRATION_IDLE 0
7337 #define RCU_MIGRATION_NEED_QS 1
7338 #define RCU_MIGRATION_GOT_QS 2
7339 #define RCU_MIGRATION_MUST_SYNC 3
7342 * migration_thread - this is a highprio system thread that performs
7343 * thread migration by bumping thread off CPU then 'pushing' onto
7346 static int migration_thread(void *data)
7349 int cpu = (long)data;
7353 BUG_ON(rq->migration_thread != current);
7355 set_current_state(TASK_INTERRUPTIBLE);
7356 while (!kthread_should_stop()) {
7357 struct migration_req *req;
7358 struct list_head *head;
7360 spin_lock_irq(&rq->lock);
7362 if (cpu_is_offline(cpu)) {
7363 spin_unlock_irq(&rq->lock);
7367 if (rq->active_balance) {
7368 active_load_balance(rq, cpu);
7369 rq->active_balance = 0;
7372 head = &rq->migration_queue;
7374 if (list_empty(head)) {
7375 spin_unlock_irq(&rq->lock);
7377 set_current_state(TASK_INTERRUPTIBLE);
7380 req = list_entry(head->next, struct migration_req, list);
7381 list_del_init(head->next);
7383 if (req->task != NULL) {
7384 spin_unlock(&rq->lock);
7385 __migrate_task(req->task, cpu, req->dest_cpu);
7386 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7387 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7388 spin_unlock(&rq->lock);
7390 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7391 spin_unlock(&rq->lock);
7392 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7396 complete(&req->done);
7398 __set_current_state(TASK_RUNNING);
7403 #ifdef CONFIG_HOTPLUG_CPU
7405 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7409 local_irq_disable();
7410 ret = __migrate_task(p, src_cpu, dest_cpu);
7416 * Figure out where task on dead CPU should go, use force if necessary.
7418 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7423 dest_cpu = select_fallback_rq(dead_cpu, p);
7425 /* It can have affinity changed while we were choosing. */
7426 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7431 * While a dead CPU has no uninterruptible tasks queued at this point,
7432 * it might still have a nonzero ->nr_uninterruptible counter, because
7433 * for performance reasons the counter is not stricly tracking tasks to
7434 * their home CPUs. So we just add the counter to another CPU's counter,
7435 * to keep the global sum constant after CPU-down:
7437 static void migrate_nr_uninterruptible(struct rq *rq_src)
7439 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7440 unsigned long flags;
7442 local_irq_save(flags);
7443 double_rq_lock(rq_src, rq_dest);
7444 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7445 rq_src->nr_uninterruptible = 0;
7446 double_rq_unlock(rq_src, rq_dest);
7447 local_irq_restore(flags);
7450 /* Run through task list and migrate tasks from the dead cpu. */
7451 static void migrate_live_tasks(int src_cpu)
7453 struct task_struct *p, *t;
7455 read_lock(&tasklist_lock);
7457 do_each_thread(t, p) {
7461 if (task_cpu(p) == src_cpu)
7462 move_task_off_dead_cpu(src_cpu, p);
7463 } while_each_thread(t, p);
7465 read_unlock(&tasklist_lock);
7469 * Schedules idle task to be the next runnable task on current CPU.
7470 * It does so by boosting its priority to highest possible.
7471 * Used by CPU offline code.
7473 void sched_idle_next(void)
7475 int this_cpu = smp_processor_id();
7476 struct rq *rq = cpu_rq(this_cpu);
7477 struct task_struct *p = rq->idle;
7478 unsigned long flags;
7480 /* cpu has to be offline */
7481 BUG_ON(cpu_online(this_cpu));
7484 * Strictly not necessary since rest of the CPUs are stopped by now
7485 * and interrupts disabled on the current cpu.
7487 spin_lock_irqsave(&rq->lock, flags);
7489 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7491 update_rq_clock(rq);
7492 activate_task(rq, p, 0);
7494 spin_unlock_irqrestore(&rq->lock, flags);
7498 * Ensures that the idle task is using init_mm right before its cpu goes
7501 void idle_task_exit(void)
7503 struct mm_struct *mm = current->active_mm;
7505 BUG_ON(cpu_online(smp_processor_id()));
7508 switch_mm(mm, &init_mm, current);
7512 /* called under rq->lock with disabled interrupts */
7513 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7515 struct rq *rq = cpu_rq(dead_cpu);
7517 /* Must be exiting, otherwise would be on tasklist. */
7518 BUG_ON(!p->exit_state);
7520 /* Cannot have done final schedule yet: would have vanished. */
7521 BUG_ON(p->state == TASK_DEAD);
7526 * Drop lock around migration; if someone else moves it,
7527 * that's OK. No task can be added to this CPU, so iteration is
7530 spin_unlock_irq(&rq->lock);
7531 move_task_off_dead_cpu(dead_cpu, p);
7532 spin_lock_irq(&rq->lock);
7537 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7538 static void migrate_dead_tasks(unsigned int dead_cpu)
7540 struct rq *rq = cpu_rq(dead_cpu);
7541 struct task_struct *next;
7544 if (!rq->nr_running)
7546 update_rq_clock(rq);
7547 next = pick_next_task(rq);
7550 next->sched_class->put_prev_task(rq, next);
7551 migrate_dead(dead_cpu, next);
7557 * remove the tasks which were accounted by rq from calc_load_tasks.
7559 static void calc_global_load_remove(struct rq *rq)
7561 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7562 rq->calc_load_active = 0;
7564 #endif /* CONFIG_HOTPLUG_CPU */
7566 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7568 static struct ctl_table sd_ctl_dir[] = {
7570 .procname = "sched_domain",
7576 static struct ctl_table sd_ctl_root[] = {
7578 .ctl_name = CTL_KERN,
7579 .procname = "kernel",
7581 .child = sd_ctl_dir,
7586 static struct ctl_table *sd_alloc_ctl_entry(int n)
7588 struct ctl_table *entry =
7589 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7594 static void sd_free_ctl_entry(struct ctl_table **tablep)
7596 struct ctl_table *entry;
7599 * In the intermediate directories, both the child directory and
7600 * procname are dynamically allocated and could fail but the mode
7601 * will always be set. In the lowest directory the names are
7602 * static strings and all have proc handlers.
7604 for (entry = *tablep; entry->mode; entry++) {
7606 sd_free_ctl_entry(&entry->child);
7607 if (entry->proc_handler == NULL)
7608 kfree(entry->procname);
7616 set_table_entry(struct ctl_table *entry,
7617 const char *procname, void *data, int maxlen,
7618 mode_t mode, proc_handler *proc_handler)
7620 entry->procname = procname;
7622 entry->maxlen = maxlen;
7624 entry->proc_handler = proc_handler;
7627 static struct ctl_table *
7628 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7630 struct ctl_table *table = sd_alloc_ctl_entry(13);
7635 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7636 sizeof(long), 0644, proc_doulongvec_minmax);
7637 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7638 sizeof(long), 0644, proc_doulongvec_minmax);
7639 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7640 sizeof(int), 0644, proc_dointvec_minmax);
7641 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7642 sizeof(int), 0644, proc_dointvec_minmax);
7643 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7644 sizeof(int), 0644, proc_dointvec_minmax);
7645 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7646 sizeof(int), 0644, proc_dointvec_minmax);
7647 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7648 sizeof(int), 0644, proc_dointvec_minmax);
7649 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7650 sizeof(int), 0644, proc_dointvec_minmax);
7651 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7652 sizeof(int), 0644, proc_dointvec_minmax);
7653 set_table_entry(&table[9], "cache_nice_tries",
7654 &sd->cache_nice_tries,
7655 sizeof(int), 0644, proc_dointvec_minmax);
7656 set_table_entry(&table[10], "flags", &sd->flags,
7657 sizeof(int), 0644, proc_dointvec_minmax);
7658 set_table_entry(&table[11], "name", sd->name,
7659 CORENAME_MAX_SIZE, 0444, proc_dostring);
7660 /* &table[12] is terminator */
7665 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7667 struct ctl_table *entry, *table;
7668 struct sched_domain *sd;
7669 int domain_num = 0, i;
7672 for_each_domain(cpu, sd)
7674 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7679 for_each_domain(cpu, sd) {
7680 snprintf(buf, 32, "domain%d", i);
7681 entry->procname = kstrdup(buf, GFP_KERNEL);
7683 entry->child = sd_alloc_ctl_domain_table(sd);
7690 static struct ctl_table_header *sd_sysctl_header;
7691 static void register_sched_domain_sysctl(void)
7693 int i, cpu_num = num_possible_cpus();
7694 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7697 WARN_ON(sd_ctl_dir[0].child);
7698 sd_ctl_dir[0].child = entry;
7703 for_each_possible_cpu(i) {
7704 snprintf(buf, 32, "cpu%d", i);
7705 entry->procname = kstrdup(buf, GFP_KERNEL);
7707 entry->child = sd_alloc_ctl_cpu_table(i);
7711 WARN_ON(sd_sysctl_header);
7712 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7715 /* may be called multiple times per register */
7716 static void unregister_sched_domain_sysctl(void)
7718 if (sd_sysctl_header)
7719 unregister_sysctl_table(sd_sysctl_header);
7720 sd_sysctl_header = NULL;
7721 if (sd_ctl_dir[0].child)
7722 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7725 static void register_sched_domain_sysctl(void)
7728 static void unregister_sched_domain_sysctl(void)
7733 static void set_rq_online(struct rq *rq)
7736 const struct sched_class *class;
7738 cpumask_set_cpu(rq->cpu, rq->rd->online);
7741 for_each_class(class) {
7742 if (class->rq_online)
7743 class->rq_online(rq);
7748 static void set_rq_offline(struct rq *rq)
7751 const struct sched_class *class;
7753 for_each_class(class) {
7754 if (class->rq_offline)
7755 class->rq_offline(rq);
7758 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7764 * migration_call - callback that gets triggered when a CPU is added.
7765 * Here we can start up the necessary migration thread for the new CPU.
7767 static int __cpuinit
7768 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7770 struct task_struct *p;
7771 int cpu = (long)hcpu;
7772 unsigned long flags;
7777 case CPU_UP_PREPARE:
7778 case CPU_UP_PREPARE_FROZEN:
7779 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7782 kthread_bind(p, cpu);
7783 /* Must be high prio: stop_machine expects to yield to it. */
7784 rq = task_rq_lock(p, &flags);
7785 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7786 task_rq_unlock(rq, &flags);
7788 cpu_rq(cpu)->migration_thread = p;
7789 rq->calc_load_update = calc_load_update;
7793 case CPU_ONLINE_FROZEN:
7794 /* Strictly unnecessary, as first user will wake it. */
7795 wake_up_process(cpu_rq(cpu)->migration_thread);
7797 /* Update our root-domain */
7799 spin_lock_irqsave(&rq->lock, flags);
7801 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7805 spin_unlock_irqrestore(&rq->lock, flags);
7808 #ifdef CONFIG_HOTPLUG_CPU
7809 case CPU_UP_CANCELED:
7810 case CPU_UP_CANCELED_FROZEN:
7811 if (!cpu_rq(cpu)->migration_thread)
7813 /* Unbind it from offline cpu so it can run. Fall thru. */
7814 kthread_bind(cpu_rq(cpu)->migration_thread,
7815 cpumask_any(cpu_online_mask));
7816 kthread_stop(cpu_rq(cpu)->migration_thread);
7817 put_task_struct(cpu_rq(cpu)->migration_thread);
7818 cpu_rq(cpu)->migration_thread = NULL;
7823 * Bring the migration thread down in CPU_POST_DEAD event,
7824 * since the timers should have got migrated by now and thus
7825 * we should not see a deadlock between trying to kill the
7826 * migration thread and the sched_rt_period_timer.
7829 kthread_stop(rq->migration_thread);
7830 put_task_struct(rq->migration_thread);
7831 rq->migration_thread = NULL;
7835 case CPU_DEAD_FROZEN:
7836 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7837 migrate_live_tasks(cpu);
7839 /* Idle task back to normal (off runqueue, low prio) */
7840 spin_lock_irq(&rq->lock);
7841 update_rq_clock(rq);
7842 deactivate_task(rq, rq->idle, 0);
7843 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7844 rq->idle->sched_class = &idle_sched_class;
7845 migrate_dead_tasks(cpu);
7846 spin_unlock_irq(&rq->lock);
7848 migrate_nr_uninterruptible(rq);
7849 BUG_ON(rq->nr_running != 0);
7850 calc_global_load_remove(rq);
7852 * No need to migrate the tasks: it was best-effort if
7853 * they didn't take sched_hotcpu_mutex. Just wake up
7856 spin_lock_irq(&rq->lock);
7857 while (!list_empty(&rq->migration_queue)) {
7858 struct migration_req *req;
7860 req = list_entry(rq->migration_queue.next,
7861 struct migration_req, list);
7862 list_del_init(&req->list);
7863 spin_unlock_irq(&rq->lock);
7864 complete(&req->done);
7865 spin_lock_irq(&rq->lock);
7867 spin_unlock_irq(&rq->lock);
7871 case CPU_DYING_FROZEN:
7872 /* Update our root-domain */
7874 spin_lock_irqsave(&rq->lock, flags);
7876 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7879 spin_unlock_irqrestore(&rq->lock, flags);
7887 * Register at high priority so that task migration (migrate_all_tasks)
7888 * happens before everything else. This has to be lower priority than
7889 * the notifier in the perf_event subsystem, though.
7891 static struct notifier_block __cpuinitdata migration_notifier = {
7892 .notifier_call = migration_call,
7896 static int __init migration_init(void)
7898 void *cpu = (void *)(long)smp_processor_id();
7901 /* Start one for the boot CPU: */
7902 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7903 BUG_ON(err == NOTIFY_BAD);
7904 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7905 register_cpu_notifier(&migration_notifier);
7909 early_initcall(migration_init);
7914 #ifdef CONFIG_SCHED_DEBUG
7916 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7917 struct cpumask *groupmask)
7919 struct sched_group *group = sd->groups;
7922 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7923 cpumask_clear(groupmask);
7925 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7927 if (!(sd->flags & SD_LOAD_BALANCE)) {
7928 printk("does not load-balance\n");
7930 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7935 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7937 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7938 printk(KERN_ERR "ERROR: domain->span does not contain "
7941 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7942 printk(KERN_ERR "ERROR: domain->groups does not contain"
7946 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7950 printk(KERN_ERR "ERROR: group is NULL\n");
7954 if (!group->cpu_power) {
7955 printk(KERN_CONT "\n");
7956 printk(KERN_ERR "ERROR: domain->cpu_power not "
7961 if (!cpumask_weight(sched_group_cpus(group))) {
7962 printk(KERN_CONT "\n");
7963 printk(KERN_ERR "ERROR: empty group\n");
7967 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7968 printk(KERN_CONT "\n");
7969 printk(KERN_ERR "ERROR: repeated CPUs\n");
7973 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7975 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7977 printk(KERN_CONT " %s", str);
7978 if (group->cpu_power != SCHED_LOAD_SCALE) {
7979 printk(KERN_CONT " (cpu_power = %d)",
7983 group = group->next;
7984 } while (group != sd->groups);
7985 printk(KERN_CONT "\n");
7987 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7988 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7991 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7992 printk(KERN_ERR "ERROR: parent span is not a superset "
7993 "of domain->span\n");
7997 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7999 cpumask_var_t groupmask;
8003 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
8007 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
8009 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
8010 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
8015 if (sched_domain_debug_one(sd, cpu, level, groupmask))
8022 free_cpumask_var(groupmask);
8024 #else /* !CONFIG_SCHED_DEBUG */
8025 # define sched_domain_debug(sd, cpu) do { } while (0)
8026 #endif /* CONFIG_SCHED_DEBUG */
8028 static int sd_degenerate(struct sched_domain *sd)
8030 if (cpumask_weight(sched_domain_span(sd)) == 1)
8033 /* Following flags need at least 2 groups */
8034 if (sd->flags & (SD_LOAD_BALANCE |
8035 SD_BALANCE_NEWIDLE |
8039 SD_SHARE_PKG_RESOURCES)) {
8040 if (sd->groups != sd->groups->next)
8044 /* Following flags don't use groups */
8045 if (sd->flags & (SD_WAKE_AFFINE))
8052 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
8054 unsigned long cflags = sd->flags, pflags = parent->flags;
8056 if (sd_degenerate(parent))
8059 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
8062 /* Flags needing groups don't count if only 1 group in parent */
8063 if (parent->groups == parent->groups->next) {
8064 pflags &= ~(SD_LOAD_BALANCE |
8065 SD_BALANCE_NEWIDLE |
8069 SD_SHARE_PKG_RESOURCES);
8070 if (nr_node_ids == 1)
8071 pflags &= ~SD_SERIALIZE;
8073 if (~cflags & pflags)
8079 static void free_rootdomain(struct root_domain *rd)
8081 synchronize_sched();
8083 cpupri_cleanup(&rd->cpupri);
8085 free_cpumask_var(rd->rto_mask);
8086 free_cpumask_var(rd->online);
8087 free_cpumask_var(rd->span);
8091 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8093 struct root_domain *old_rd = NULL;
8094 unsigned long flags;
8096 spin_lock_irqsave(&rq->lock, flags);
8101 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8104 cpumask_clear_cpu(rq->cpu, old_rd->span);
8107 * If we dont want to free the old_rt yet then
8108 * set old_rd to NULL to skip the freeing later
8111 if (!atomic_dec_and_test(&old_rd->refcount))
8115 atomic_inc(&rd->refcount);
8118 cpumask_set_cpu(rq->cpu, rd->span);
8119 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8122 spin_unlock_irqrestore(&rq->lock, flags);
8125 free_rootdomain(old_rd);
8128 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8130 gfp_t gfp = GFP_KERNEL;
8132 memset(rd, 0, sizeof(*rd));
8137 if (!alloc_cpumask_var(&rd->span, gfp))
8139 if (!alloc_cpumask_var(&rd->online, gfp))
8141 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8144 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8149 free_cpumask_var(rd->rto_mask);
8151 free_cpumask_var(rd->online);
8153 free_cpumask_var(rd->span);
8158 static void init_defrootdomain(void)
8160 init_rootdomain(&def_root_domain, true);
8162 atomic_set(&def_root_domain.refcount, 1);
8165 static struct root_domain *alloc_rootdomain(void)
8167 struct root_domain *rd;
8169 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8173 if (init_rootdomain(rd, false) != 0) {
8182 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8183 * hold the hotplug lock.
8186 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8188 struct rq *rq = cpu_rq(cpu);
8189 struct sched_domain *tmp;
8191 /* Remove the sched domains which do not contribute to scheduling. */
8192 for (tmp = sd; tmp; ) {
8193 struct sched_domain *parent = tmp->parent;
8197 if (sd_parent_degenerate(tmp, parent)) {
8198 tmp->parent = parent->parent;
8200 parent->parent->child = tmp;
8205 if (sd && sd_degenerate(sd)) {
8211 sched_domain_debug(sd, cpu);
8213 rq_attach_root(rq, rd);
8214 rcu_assign_pointer(rq->sd, sd);
8217 /* cpus with isolated domains */
8218 static cpumask_var_t cpu_isolated_map;
8220 /* Setup the mask of cpus configured for isolated domains */
8221 static int __init isolated_cpu_setup(char *str)
8223 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8224 cpulist_parse(str, cpu_isolated_map);
8228 __setup("isolcpus=", isolated_cpu_setup);
8231 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8232 * to a function which identifies what group(along with sched group) a CPU
8233 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8234 * (due to the fact that we keep track of groups covered with a struct cpumask).
8236 * init_sched_build_groups will build a circular linked list of the groups
8237 * covered by the given span, and will set each group's ->cpumask correctly,
8238 * and ->cpu_power to 0.
8241 init_sched_build_groups(const struct cpumask *span,
8242 const struct cpumask *cpu_map,
8243 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8244 struct sched_group **sg,
8245 struct cpumask *tmpmask),
8246 struct cpumask *covered, struct cpumask *tmpmask)
8248 struct sched_group *first = NULL, *last = NULL;
8251 cpumask_clear(covered);
8253 for_each_cpu(i, span) {
8254 struct sched_group *sg;
8255 int group = group_fn(i, cpu_map, &sg, tmpmask);
8258 if (cpumask_test_cpu(i, covered))
8261 cpumask_clear(sched_group_cpus(sg));
8264 for_each_cpu(j, span) {
8265 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8268 cpumask_set_cpu(j, covered);
8269 cpumask_set_cpu(j, sched_group_cpus(sg));
8280 #define SD_NODES_PER_DOMAIN 16
8285 * find_next_best_node - find the next node to include in a sched_domain
8286 * @node: node whose sched_domain we're building
8287 * @used_nodes: nodes already in the sched_domain
8289 * Find the next node to include in a given scheduling domain. Simply
8290 * finds the closest node not already in the @used_nodes map.
8292 * Should use nodemask_t.
8294 static int find_next_best_node(int node, nodemask_t *used_nodes)
8296 int i, n, val, min_val, best_node = 0;
8300 for (i = 0; i < nr_node_ids; i++) {
8301 /* Start at @node */
8302 n = (node + i) % nr_node_ids;
8304 if (!nr_cpus_node(n))
8307 /* Skip already used nodes */
8308 if (node_isset(n, *used_nodes))
8311 /* Simple min distance search */
8312 val = node_distance(node, n);
8314 if (val < min_val) {
8320 node_set(best_node, *used_nodes);
8325 * sched_domain_node_span - get a cpumask for a node's sched_domain
8326 * @node: node whose cpumask we're constructing
8327 * @span: resulting cpumask
8329 * Given a node, construct a good cpumask for its sched_domain to span. It
8330 * should be one that prevents unnecessary balancing, but also spreads tasks
8333 static void sched_domain_node_span(int node, struct cpumask *span)
8335 nodemask_t used_nodes;
8338 cpumask_clear(span);
8339 nodes_clear(used_nodes);
8341 cpumask_or(span, span, cpumask_of_node(node));
8342 node_set(node, used_nodes);
8344 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8345 int next_node = find_next_best_node(node, &used_nodes);
8347 cpumask_or(span, span, cpumask_of_node(next_node));
8350 #endif /* CONFIG_NUMA */
8352 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8355 * The cpus mask in sched_group and sched_domain hangs off the end.
8357 * ( See the the comments in include/linux/sched.h:struct sched_group
8358 * and struct sched_domain. )
8360 struct static_sched_group {
8361 struct sched_group sg;
8362 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8365 struct static_sched_domain {
8366 struct sched_domain sd;
8367 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8373 cpumask_var_t domainspan;
8374 cpumask_var_t covered;
8375 cpumask_var_t notcovered;
8377 cpumask_var_t nodemask;
8378 cpumask_var_t this_sibling_map;
8379 cpumask_var_t this_core_map;
8380 cpumask_var_t send_covered;
8381 cpumask_var_t tmpmask;
8382 struct sched_group **sched_group_nodes;
8383 struct root_domain *rd;
8387 sa_sched_groups = 0,
8392 sa_this_sibling_map,
8394 sa_sched_group_nodes,
8404 * SMT sched-domains:
8406 #ifdef CONFIG_SCHED_SMT
8407 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8408 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8411 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8412 struct sched_group **sg, struct cpumask *unused)
8415 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8418 #endif /* CONFIG_SCHED_SMT */
8421 * multi-core sched-domains:
8423 #ifdef CONFIG_SCHED_MC
8424 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8425 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8426 #endif /* CONFIG_SCHED_MC */
8428 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8430 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8431 struct sched_group **sg, struct cpumask *mask)
8435 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8436 group = cpumask_first(mask);
8438 *sg = &per_cpu(sched_group_core, group).sg;
8441 #elif defined(CONFIG_SCHED_MC)
8443 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8444 struct sched_group **sg, struct cpumask *unused)
8447 *sg = &per_cpu(sched_group_core, cpu).sg;
8452 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8453 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8456 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8457 struct sched_group **sg, struct cpumask *mask)
8460 #ifdef CONFIG_SCHED_MC
8461 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8462 group = cpumask_first(mask);
8463 #elif defined(CONFIG_SCHED_SMT)
8464 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8465 group = cpumask_first(mask);
8470 *sg = &per_cpu(sched_group_phys, group).sg;
8476 * The init_sched_build_groups can't handle what we want to do with node
8477 * groups, so roll our own. Now each node has its own list of groups which
8478 * gets dynamically allocated.
8480 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8481 static struct sched_group ***sched_group_nodes_bycpu;
8483 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8484 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8486 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8487 struct sched_group **sg,
8488 struct cpumask *nodemask)
8492 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8493 group = cpumask_first(nodemask);
8496 *sg = &per_cpu(sched_group_allnodes, group).sg;
8500 static void init_numa_sched_groups_power(struct sched_group *group_head)
8502 struct sched_group *sg = group_head;
8508 for_each_cpu(j, sched_group_cpus(sg)) {
8509 struct sched_domain *sd;
8511 sd = &per_cpu(phys_domains, j).sd;
8512 if (j != group_first_cpu(sd->groups)) {
8514 * Only add "power" once for each
8520 sg->cpu_power += sd->groups->cpu_power;
8523 } while (sg != group_head);
8526 static int build_numa_sched_groups(struct s_data *d,
8527 const struct cpumask *cpu_map, int num)
8529 struct sched_domain *sd;
8530 struct sched_group *sg, *prev;
8533 cpumask_clear(d->covered);
8534 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8535 if (cpumask_empty(d->nodemask)) {
8536 d->sched_group_nodes[num] = NULL;
8540 sched_domain_node_span(num, d->domainspan);
8541 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8543 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8546 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8550 d->sched_group_nodes[num] = sg;
8552 for_each_cpu(j, d->nodemask) {
8553 sd = &per_cpu(node_domains, j).sd;
8558 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8560 cpumask_or(d->covered, d->covered, d->nodemask);
8563 for (j = 0; j < nr_node_ids; j++) {
8564 n = (num + j) % nr_node_ids;
8565 cpumask_complement(d->notcovered, d->covered);
8566 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8567 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8568 if (cpumask_empty(d->tmpmask))
8570 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8571 if (cpumask_empty(d->tmpmask))
8573 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8577 "Can not alloc domain group for node %d\n", j);
8581 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8582 sg->next = prev->next;
8583 cpumask_or(d->covered, d->covered, d->tmpmask);
8590 #endif /* CONFIG_NUMA */
8593 /* Free memory allocated for various sched_group structures */
8594 static void free_sched_groups(const struct cpumask *cpu_map,
8595 struct cpumask *nodemask)
8599 for_each_cpu(cpu, cpu_map) {
8600 struct sched_group **sched_group_nodes
8601 = sched_group_nodes_bycpu[cpu];
8603 if (!sched_group_nodes)
8606 for (i = 0; i < nr_node_ids; i++) {
8607 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8609 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8610 if (cpumask_empty(nodemask))
8620 if (oldsg != sched_group_nodes[i])
8623 kfree(sched_group_nodes);
8624 sched_group_nodes_bycpu[cpu] = NULL;
8627 #else /* !CONFIG_NUMA */
8628 static void free_sched_groups(const struct cpumask *cpu_map,
8629 struct cpumask *nodemask)
8632 #endif /* CONFIG_NUMA */
8635 * Initialize sched groups cpu_power.
8637 * cpu_power indicates the capacity of sched group, which is used while
8638 * distributing the load between different sched groups in a sched domain.
8639 * Typically cpu_power for all the groups in a sched domain will be same unless
8640 * there are asymmetries in the topology. If there are asymmetries, group
8641 * having more cpu_power will pickup more load compared to the group having
8644 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8646 struct sched_domain *child;
8647 struct sched_group *group;
8651 WARN_ON(!sd || !sd->groups);
8653 if (cpu != group_first_cpu(sd->groups))
8658 sd->groups->cpu_power = 0;
8661 power = SCHED_LOAD_SCALE;
8662 weight = cpumask_weight(sched_domain_span(sd));
8664 * SMT siblings share the power of a single core.
8665 * Usually multiple threads get a better yield out of
8666 * that one core than a single thread would have,
8667 * reflect that in sd->smt_gain.
8669 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8670 power *= sd->smt_gain;
8672 power >>= SCHED_LOAD_SHIFT;
8674 sd->groups->cpu_power += power;
8679 * Add cpu_power of each child group to this groups cpu_power.
8681 group = child->groups;
8683 sd->groups->cpu_power += group->cpu_power;
8684 group = group->next;
8685 } while (group != child->groups);
8689 * Initializers for schedule domains
8690 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8693 #ifdef CONFIG_SCHED_DEBUG
8694 # define SD_INIT_NAME(sd, type) sd->name = #type
8696 # define SD_INIT_NAME(sd, type) do { } while (0)
8699 #define SD_INIT(sd, type) sd_init_##type(sd)
8701 #define SD_INIT_FUNC(type) \
8702 static noinline void sd_init_##type(struct sched_domain *sd) \
8704 memset(sd, 0, sizeof(*sd)); \
8705 *sd = SD_##type##_INIT; \
8706 sd->level = SD_LV_##type; \
8707 SD_INIT_NAME(sd, type); \
8712 SD_INIT_FUNC(ALLNODES)
8715 #ifdef CONFIG_SCHED_SMT
8716 SD_INIT_FUNC(SIBLING)
8718 #ifdef CONFIG_SCHED_MC
8722 static int default_relax_domain_level = -1;
8724 static int __init setup_relax_domain_level(char *str)
8728 val = simple_strtoul(str, NULL, 0);
8729 if (val < SD_LV_MAX)
8730 default_relax_domain_level = val;
8734 __setup("relax_domain_level=", setup_relax_domain_level);
8736 static void set_domain_attribute(struct sched_domain *sd,
8737 struct sched_domain_attr *attr)
8741 if (!attr || attr->relax_domain_level < 0) {
8742 if (default_relax_domain_level < 0)
8745 request = default_relax_domain_level;
8747 request = attr->relax_domain_level;
8748 if (request < sd->level) {
8749 /* turn off idle balance on this domain */
8750 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8752 /* turn on idle balance on this domain */
8753 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8757 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8758 const struct cpumask *cpu_map)
8761 case sa_sched_groups:
8762 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8763 d->sched_group_nodes = NULL;
8765 free_rootdomain(d->rd); /* fall through */
8767 free_cpumask_var(d->tmpmask); /* fall through */
8768 case sa_send_covered:
8769 free_cpumask_var(d->send_covered); /* fall through */
8770 case sa_this_core_map:
8771 free_cpumask_var(d->this_core_map); /* fall through */
8772 case sa_this_sibling_map:
8773 free_cpumask_var(d->this_sibling_map); /* fall through */
8775 free_cpumask_var(d->nodemask); /* fall through */
8776 case sa_sched_group_nodes:
8778 kfree(d->sched_group_nodes); /* fall through */
8780 free_cpumask_var(d->notcovered); /* fall through */
8782 free_cpumask_var(d->covered); /* fall through */
8784 free_cpumask_var(d->domainspan); /* fall through */
8791 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8792 const struct cpumask *cpu_map)
8795 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8797 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8798 return sa_domainspan;
8799 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8801 /* Allocate the per-node list of sched groups */
8802 d->sched_group_nodes = kcalloc(nr_node_ids,
8803 sizeof(struct sched_group *), GFP_KERNEL);
8804 if (!d->sched_group_nodes) {
8805 printk(KERN_WARNING "Can not alloc sched group node list\n");
8806 return sa_notcovered;
8808 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8810 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8811 return sa_sched_group_nodes;
8812 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8814 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8815 return sa_this_sibling_map;
8816 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8817 return sa_this_core_map;
8818 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8819 return sa_send_covered;
8820 d->rd = alloc_rootdomain();
8822 printk(KERN_WARNING "Cannot alloc root domain\n");
8825 return sa_rootdomain;
8828 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8829 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8831 struct sched_domain *sd = NULL;
8833 struct sched_domain *parent;
8836 if (cpumask_weight(cpu_map) >
8837 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8838 sd = &per_cpu(allnodes_domains, i).sd;
8839 SD_INIT(sd, ALLNODES);
8840 set_domain_attribute(sd, attr);
8841 cpumask_copy(sched_domain_span(sd), cpu_map);
8842 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8847 sd = &per_cpu(node_domains, i).sd;
8849 set_domain_attribute(sd, attr);
8850 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8851 sd->parent = parent;
8854 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8859 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8860 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8861 struct sched_domain *parent, int i)
8863 struct sched_domain *sd;
8864 sd = &per_cpu(phys_domains, i).sd;
8866 set_domain_attribute(sd, attr);
8867 cpumask_copy(sched_domain_span(sd), d->nodemask);
8868 sd->parent = parent;
8871 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8875 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8876 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8877 struct sched_domain *parent, int i)
8879 struct sched_domain *sd = parent;
8880 #ifdef CONFIG_SCHED_MC
8881 sd = &per_cpu(core_domains, i).sd;
8883 set_domain_attribute(sd, attr);
8884 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8885 sd->parent = parent;
8887 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8892 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8893 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8894 struct sched_domain *parent, int i)
8896 struct sched_domain *sd = parent;
8897 #ifdef CONFIG_SCHED_SMT
8898 sd = &per_cpu(cpu_domains, i).sd;
8899 SD_INIT(sd, SIBLING);
8900 set_domain_attribute(sd, attr);
8901 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8902 sd->parent = parent;
8904 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8909 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8910 const struct cpumask *cpu_map, int cpu)
8913 #ifdef CONFIG_SCHED_SMT
8914 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8915 cpumask_and(d->this_sibling_map, cpu_map,
8916 topology_thread_cpumask(cpu));
8917 if (cpu == cpumask_first(d->this_sibling_map))
8918 init_sched_build_groups(d->this_sibling_map, cpu_map,
8920 d->send_covered, d->tmpmask);
8923 #ifdef CONFIG_SCHED_MC
8924 case SD_LV_MC: /* set up multi-core groups */
8925 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8926 if (cpu == cpumask_first(d->this_core_map))
8927 init_sched_build_groups(d->this_core_map, cpu_map,
8929 d->send_covered, d->tmpmask);
8932 case SD_LV_CPU: /* set up physical groups */
8933 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8934 if (!cpumask_empty(d->nodemask))
8935 init_sched_build_groups(d->nodemask, cpu_map,
8937 d->send_covered, d->tmpmask);
8940 case SD_LV_ALLNODES:
8941 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8942 d->send_covered, d->tmpmask);
8951 * Build sched domains for a given set of cpus and attach the sched domains
8952 * to the individual cpus
8954 static int __build_sched_domains(const struct cpumask *cpu_map,
8955 struct sched_domain_attr *attr)
8957 enum s_alloc alloc_state = sa_none;
8959 struct sched_domain *sd;
8965 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8966 if (alloc_state != sa_rootdomain)
8968 alloc_state = sa_sched_groups;
8971 * Set up domains for cpus specified by the cpu_map.
8973 for_each_cpu(i, cpu_map) {
8974 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8977 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8978 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8979 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8980 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8983 for_each_cpu(i, cpu_map) {
8984 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8985 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8988 /* Set up physical groups */
8989 for (i = 0; i < nr_node_ids; i++)
8990 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8993 /* Set up node groups */
8995 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8997 for (i = 0; i < nr_node_ids; i++)
8998 if (build_numa_sched_groups(&d, cpu_map, i))
9002 /* Calculate CPU power for physical packages and nodes */
9003 #ifdef CONFIG_SCHED_SMT
9004 for_each_cpu(i, cpu_map) {
9005 sd = &per_cpu(cpu_domains, i).sd;
9006 init_sched_groups_power(i, sd);
9009 #ifdef CONFIG_SCHED_MC
9010 for_each_cpu(i, cpu_map) {
9011 sd = &per_cpu(core_domains, i).sd;
9012 init_sched_groups_power(i, sd);
9016 for_each_cpu(i, cpu_map) {
9017 sd = &per_cpu(phys_domains, i).sd;
9018 init_sched_groups_power(i, sd);
9022 for (i = 0; i < nr_node_ids; i++)
9023 init_numa_sched_groups_power(d.sched_group_nodes[i]);
9025 if (d.sd_allnodes) {
9026 struct sched_group *sg;
9028 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
9030 init_numa_sched_groups_power(sg);
9034 /* Attach the domains */
9035 for_each_cpu(i, cpu_map) {
9036 #ifdef CONFIG_SCHED_SMT
9037 sd = &per_cpu(cpu_domains, i).sd;
9038 #elif defined(CONFIG_SCHED_MC)
9039 sd = &per_cpu(core_domains, i).sd;
9041 sd = &per_cpu(phys_domains, i).sd;
9043 cpu_attach_domain(sd, d.rd, i);
9046 d.sched_group_nodes = NULL; /* don't free this we still need it */
9047 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
9051 __free_domain_allocs(&d, alloc_state, cpu_map);
9055 static int build_sched_domains(const struct cpumask *cpu_map)
9057 return __build_sched_domains(cpu_map, NULL);
9060 static struct cpumask *doms_cur; /* current sched domains */
9061 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
9062 static struct sched_domain_attr *dattr_cur;
9063 /* attribues of custom domains in 'doms_cur' */
9066 * Special case: If a kmalloc of a doms_cur partition (array of
9067 * cpumask) fails, then fallback to a single sched domain,
9068 * as determined by the single cpumask fallback_doms.
9070 static cpumask_var_t fallback_doms;
9073 * arch_update_cpu_topology lets virtualized architectures update the
9074 * cpu core maps. It is supposed to return 1 if the topology changed
9075 * or 0 if it stayed the same.
9077 int __attribute__((weak)) arch_update_cpu_topology(void)
9083 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9084 * For now this just excludes isolated cpus, but could be used to
9085 * exclude other special cases in the future.
9087 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9091 arch_update_cpu_topology();
9093 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
9095 doms_cur = fallback_doms;
9096 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
9098 err = build_sched_domains(doms_cur);
9099 register_sched_domain_sysctl();
9104 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9105 struct cpumask *tmpmask)
9107 free_sched_groups(cpu_map, tmpmask);
9111 * Detach sched domains from a group of cpus specified in cpu_map
9112 * These cpus will now be attached to the NULL domain
9114 static void detach_destroy_domains(const struct cpumask *cpu_map)
9116 /* Save because hotplug lock held. */
9117 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9120 for_each_cpu(i, cpu_map)
9121 cpu_attach_domain(NULL, &def_root_domain, i);
9122 synchronize_sched();
9123 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9126 /* handle null as "default" */
9127 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9128 struct sched_domain_attr *new, int idx_new)
9130 struct sched_domain_attr tmp;
9137 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9138 new ? (new + idx_new) : &tmp,
9139 sizeof(struct sched_domain_attr));
9143 * Partition sched domains as specified by the 'ndoms_new'
9144 * cpumasks in the array doms_new[] of cpumasks. This compares
9145 * doms_new[] to the current sched domain partitioning, doms_cur[].
9146 * It destroys each deleted domain and builds each new domain.
9148 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9149 * The masks don't intersect (don't overlap.) We should setup one
9150 * sched domain for each mask. CPUs not in any of the cpumasks will
9151 * not be load balanced. If the same cpumask appears both in the
9152 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9155 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9156 * ownership of it and will kfree it when done with it. If the caller
9157 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9158 * ndoms_new == 1, and partition_sched_domains() will fallback to
9159 * the single partition 'fallback_doms', it also forces the domains
9162 * If doms_new == NULL it will be replaced with cpu_online_mask.
9163 * ndoms_new == 0 is a special case for destroying existing domains,
9164 * and it will not create the default domain.
9166 * Call with hotplug lock held
9168 /* FIXME: Change to struct cpumask *doms_new[] */
9169 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
9170 struct sched_domain_attr *dattr_new)
9175 mutex_lock(&sched_domains_mutex);
9177 /* always unregister in case we don't destroy any domains */
9178 unregister_sched_domain_sysctl();
9180 /* Let architecture update cpu core mappings. */
9181 new_topology = arch_update_cpu_topology();
9183 n = doms_new ? ndoms_new : 0;
9185 /* Destroy deleted domains */
9186 for (i = 0; i < ndoms_cur; i++) {
9187 for (j = 0; j < n && !new_topology; j++) {
9188 if (cpumask_equal(&doms_cur[i], &doms_new[j])
9189 && dattrs_equal(dattr_cur, i, dattr_new, j))
9192 /* no match - a current sched domain not in new doms_new[] */
9193 detach_destroy_domains(doms_cur + i);
9198 if (doms_new == NULL) {
9200 doms_new = fallback_doms;
9201 cpumask_andnot(&doms_new[0], cpu_active_mask, cpu_isolated_map);
9202 WARN_ON_ONCE(dattr_new);
9205 /* Build new domains */
9206 for (i = 0; i < ndoms_new; i++) {
9207 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9208 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9209 && dattrs_equal(dattr_new, i, dattr_cur, j))
9212 /* no match - add a new doms_new */
9213 __build_sched_domains(doms_new + i,
9214 dattr_new ? dattr_new + i : NULL);
9219 /* Remember the new sched domains */
9220 if (doms_cur != fallback_doms)
9222 kfree(dattr_cur); /* kfree(NULL) is safe */
9223 doms_cur = doms_new;
9224 dattr_cur = dattr_new;
9225 ndoms_cur = ndoms_new;
9227 register_sched_domain_sysctl();
9229 mutex_unlock(&sched_domains_mutex);
9232 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9233 static void arch_reinit_sched_domains(void)
9237 /* Destroy domains first to force the rebuild */
9238 partition_sched_domains(0, NULL, NULL);
9240 rebuild_sched_domains();
9244 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9246 unsigned int level = 0;
9248 if (sscanf(buf, "%u", &level) != 1)
9252 * level is always be positive so don't check for
9253 * level < POWERSAVINGS_BALANCE_NONE which is 0
9254 * What happens on 0 or 1 byte write,
9255 * need to check for count as well?
9258 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9262 sched_smt_power_savings = level;
9264 sched_mc_power_savings = level;
9266 arch_reinit_sched_domains();
9271 #ifdef CONFIG_SCHED_MC
9272 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9275 return sprintf(page, "%u\n", sched_mc_power_savings);
9277 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9278 const char *buf, size_t count)
9280 return sched_power_savings_store(buf, count, 0);
9282 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9283 sched_mc_power_savings_show,
9284 sched_mc_power_savings_store);
9287 #ifdef CONFIG_SCHED_SMT
9288 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9291 return sprintf(page, "%u\n", sched_smt_power_savings);
9293 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9294 const char *buf, size_t count)
9296 return sched_power_savings_store(buf, count, 1);
9298 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9299 sched_smt_power_savings_show,
9300 sched_smt_power_savings_store);
9303 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9307 #ifdef CONFIG_SCHED_SMT
9309 err = sysfs_create_file(&cls->kset.kobj,
9310 &attr_sched_smt_power_savings.attr);
9312 #ifdef CONFIG_SCHED_MC
9313 if (!err && mc_capable())
9314 err = sysfs_create_file(&cls->kset.kobj,
9315 &attr_sched_mc_power_savings.attr);
9319 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9321 #ifndef CONFIG_CPUSETS
9323 * Add online and remove offline CPUs from the scheduler domains.
9324 * When cpusets are enabled they take over this function.
9326 static int update_sched_domains(struct notifier_block *nfb,
9327 unsigned long action, void *hcpu)
9331 case CPU_ONLINE_FROZEN:
9332 case CPU_DOWN_PREPARE:
9333 case CPU_DOWN_PREPARE_FROZEN:
9334 case CPU_DOWN_FAILED:
9335 case CPU_DOWN_FAILED_FROZEN:
9336 partition_sched_domains(1, NULL, NULL);
9345 static int update_runtime(struct notifier_block *nfb,
9346 unsigned long action, void *hcpu)
9348 int cpu = (int)(long)hcpu;
9351 case CPU_DOWN_PREPARE:
9352 case CPU_DOWN_PREPARE_FROZEN:
9353 disable_runtime(cpu_rq(cpu));
9356 case CPU_DOWN_FAILED:
9357 case CPU_DOWN_FAILED_FROZEN:
9359 case CPU_ONLINE_FROZEN:
9360 enable_runtime(cpu_rq(cpu));
9368 void __init sched_init_smp(void)
9370 cpumask_var_t non_isolated_cpus;
9372 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9373 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9375 #if defined(CONFIG_NUMA)
9376 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9378 BUG_ON(sched_group_nodes_bycpu == NULL);
9381 mutex_lock(&sched_domains_mutex);
9382 arch_init_sched_domains(cpu_active_mask);
9383 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9384 if (cpumask_empty(non_isolated_cpus))
9385 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9386 mutex_unlock(&sched_domains_mutex);
9389 #ifndef CONFIG_CPUSETS
9390 /* XXX: Theoretical race here - CPU may be hotplugged now */
9391 hotcpu_notifier(update_sched_domains, 0);
9394 /* RT runtime code needs to handle some hotplug events */
9395 hotcpu_notifier(update_runtime, 0);
9399 /* Move init over to a non-isolated CPU */
9400 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9402 sched_init_granularity();
9403 free_cpumask_var(non_isolated_cpus);
9405 init_sched_rt_class();
9408 void __init sched_init_smp(void)
9410 sched_init_granularity();
9412 #endif /* CONFIG_SMP */
9414 const_debug unsigned int sysctl_timer_migration = 1;
9416 int in_sched_functions(unsigned long addr)
9418 return in_lock_functions(addr) ||
9419 (addr >= (unsigned long)__sched_text_start
9420 && addr < (unsigned long)__sched_text_end);
9423 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9425 cfs_rq->tasks_timeline = RB_ROOT;
9426 INIT_LIST_HEAD(&cfs_rq->tasks);
9427 #ifdef CONFIG_FAIR_GROUP_SCHED
9430 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9433 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9435 struct rt_prio_array *array;
9438 array = &rt_rq->active;
9439 for (i = 0; i < MAX_RT_PRIO; i++) {
9440 INIT_LIST_HEAD(array->queue + i);
9441 __clear_bit(i, array->bitmap);
9443 /* delimiter for bitsearch: */
9444 __set_bit(MAX_RT_PRIO, array->bitmap);
9446 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9447 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9449 rt_rq->highest_prio.next = MAX_RT_PRIO;
9453 rt_rq->rt_nr_migratory = 0;
9454 rt_rq->overloaded = 0;
9455 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9459 rt_rq->rt_throttled = 0;
9460 rt_rq->rt_runtime = 0;
9461 spin_lock_init(&rt_rq->rt_runtime_lock);
9463 #ifdef CONFIG_RT_GROUP_SCHED
9464 rt_rq->rt_nr_boosted = 0;
9469 #ifdef CONFIG_FAIR_GROUP_SCHED
9470 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9471 struct sched_entity *se, int cpu, int add,
9472 struct sched_entity *parent)
9474 struct rq *rq = cpu_rq(cpu);
9475 tg->cfs_rq[cpu] = cfs_rq;
9476 init_cfs_rq(cfs_rq, rq);
9479 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9482 /* se could be NULL for init_task_group */
9487 se->cfs_rq = &rq->cfs;
9489 se->cfs_rq = parent->my_q;
9492 se->load.weight = tg->shares;
9493 se->load.inv_weight = 0;
9494 se->parent = parent;
9498 #ifdef CONFIG_RT_GROUP_SCHED
9499 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9500 struct sched_rt_entity *rt_se, int cpu, int add,
9501 struct sched_rt_entity *parent)
9503 struct rq *rq = cpu_rq(cpu);
9505 tg->rt_rq[cpu] = rt_rq;
9506 init_rt_rq(rt_rq, rq);
9508 rt_rq->rt_se = rt_se;
9509 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9511 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9513 tg->rt_se[cpu] = rt_se;
9518 rt_se->rt_rq = &rq->rt;
9520 rt_se->rt_rq = parent->my_q;
9522 rt_se->my_q = rt_rq;
9523 rt_se->parent = parent;
9524 INIT_LIST_HEAD(&rt_se->run_list);
9528 void __init sched_init(void)
9531 unsigned long alloc_size = 0, ptr;
9533 #ifdef CONFIG_FAIR_GROUP_SCHED
9534 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9536 #ifdef CONFIG_RT_GROUP_SCHED
9537 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9539 #ifdef CONFIG_USER_SCHED
9542 #ifdef CONFIG_CPUMASK_OFFSTACK
9543 alloc_size += num_possible_cpus() * cpumask_size();
9546 * As sched_init() is called before page_alloc is setup,
9547 * we use alloc_bootmem().
9550 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9552 #ifdef CONFIG_FAIR_GROUP_SCHED
9553 init_task_group.se = (struct sched_entity **)ptr;
9554 ptr += nr_cpu_ids * sizeof(void **);
9556 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9557 ptr += nr_cpu_ids * sizeof(void **);
9559 #ifdef CONFIG_USER_SCHED
9560 root_task_group.se = (struct sched_entity **)ptr;
9561 ptr += nr_cpu_ids * sizeof(void **);
9563 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9564 ptr += nr_cpu_ids * sizeof(void **);
9565 #endif /* CONFIG_USER_SCHED */
9566 #endif /* CONFIG_FAIR_GROUP_SCHED */
9567 #ifdef CONFIG_RT_GROUP_SCHED
9568 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9569 ptr += nr_cpu_ids * sizeof(void **);
9571 init_task_group.rt_rq = (struct rt_rq **)ptr;
9572 ptr += nr_cpu_ids * sizeof(void **);
9574 #ifdef CONFIG_USER_SCHED
9575 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9576 ptr += nr_cpu_ids * sizeof(void **);
9578 root_task_group.rt_rq = (struct rt_rq **)ptr;
9579 ptr += nr_cpu_ids * sizeof(void **);
9580 #endif /* CONFIG_USER_SCHED */
9581 #endif /* CONFIG_RT_GROUP_SCHED */
9582 #ifdef CONFIG_CPUMASK_OFFSTACK
9583 for_each_possible_cpu(i) {
9584 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9585 ptr += cpumask_size();
9587 #endif /* CONFIG_CPUMASK_OFFSTACK */
9591 init_defrootdomain();
9594 init_rt_bandwidth(&def_rt_bandwidth,
9595 global_rt_period(), global_rt_runtime());
9597 #ifdef CONFIG_RT_GROUP_SCHED
9598 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9599 global_rt_period(), global_rt_runtime());
9600 #ifdef CONFIG_USER_SCHED
9601 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9602 global_rt_period(), RUNTIME_INF);
9603 #endif /* CONFIG_USER_SCHED */
9604 #endif /* CONFIG_RT_GROUP_SCHED */
9606 #ifdef CONFIG_GROUP_SCHED
9607 list_add(&init_task_group.list, &task_groups);
9608 INIT_LIST_HEAD(&init_task_group.children);
9610 #ifdef CONFIG_USER_SCHED
9611 INIT_LIST_HEAD(&root_task_group.children);
9612 init_task_group.parent = &root_task_group;
9613 list_add(&init_task_group.siblings, &root_task_group.children);
9614 #endif /* CONFIG_USER_SCHED */
9615 #endif /* CONFIG_GROUP_SCHED */
9617 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9618 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9619 __alignof__(unsigned long));
9621 for_each_possible_cpu(i) {
9625 spin_lock_init(&rq->lock);
9627 rq->calc_load_active = 0;
9628 rq->calc_load_update = jiffies + LOAD_FREQ;
9629 init_cfs_rq(&rq->cfs, rq);
9630 init_rt_rq(&rq->rt, rq);
9631 #ifdef CONFIG_FAIR_GROUP_SCHED
9632 init_task_group.shares = init_task_group_load;
9633 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9634 #ifdef CONFIG_CGROUP_SCHED
9636 * How much cpu bandwidth does init_task_group get?
9638 * In case of task-groups formed thr' the cgroup filesystem, it
9639 * gets 100% of the cpu resources in the system. This overall
9640 * system cpu resource is divided among the tasks of
9641 * init_task_group and its child task-groups in a fair manner,
9642 * based on each entity's (task or task-group's) weight
9643 * (se->load.weight).
9645 * In other words, if init_task_group has 10 tasks of weight
9646 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9647 * then A0's share of the cpu resource is:
9649 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9651 * We achieve this by letting init_task_group's tasks sit
9652 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9654 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9655 #elif defined CONFIG_USER_SCHED
9656 root_task_group.shares = NICE_0_LOAD;
9657 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9659 * In case of task-groups formed thr' the user id of tasks,
9660 * init_task_group represents tasks belonging to root user.
9661 * Hence it forms a sibling of all subsequent groups formed.
9662 * In this case, init_task_group gets only a fraction of overall
9663 * system cpu resource, based on the weight assigned to root
9664 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9665 * by letting tasks of init_task_group sit in a separate cfs_rq
9666 * (init_tg_cfs_rq) and having one entity represent this group of
9667 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9669 init_tg_cfs_entry(&init_task_group,
9670 &per_cpu(init_tg_cfs_rq, i),
9671 &per_cpu(init_sched_entity, i), i, 1,
9672 root_task_group.se[i]);
9675 #endif /* CONFIG_FAIR_GROUP_SCHED */
9677 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9678 #ifdef CONFIG_RT_GROUP_SCHED
9679 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9680 #ifdef CONFIG_CGROUP_SCHED
9681 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9682 #elif defined CONFIG_USER_SCHED
9683 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9684 init_tg_rt_entry(&init_task_group,
9685 &per_cpu(init_rt_rq, i),
9686 &per_cpu(init_sched_rt_entity, i), i, 1,
9687 root_task_group.rt_se[i]);
9691 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9692 rq->cpu_load[j] = 0;
9696 rq->post_schedule = 0;
9697 rq->active_balance = 0;
9698 rq->next_balance = jiffies;
9702 rq->migration_thread = NULL;
9704 rq->avg_idle = 2*sysctl_sched_migration_cost;
9705 INIT_LIST_HEAD(&rq->migration_queue);
9706 rq_attach_root(rq, &def_root_domain);
9709 atomic_set(&rq->nr_iowait, 0);
9712 set_load_weight(&init_task);
9714 #ifdef CONFIG_PREEMPT_NOTIFIERS
9715 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9719 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9722 #ifdef CONFIG_RT_MUTEXES
9723 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9727 * The boot idle thread does lazy MMU switching as well:
9729 atomic_inc(&init_mm.mm_count);
9730 enter_lazy_tlb(&init_mm, current);
9733 * Make us the idle thread. Technically, schedule() should not be
9734 * called from this thread, however somewhere below it might be,
9735 * but because we are the idle thread, we just pick up running again
9736 * when this runqueue becomes "idle".
9738 init_idle(current, smp_processor_id());
9740 calc_load_update = jiffies + LOAD_FREQ;
9743 * During early bootup we pretend to be a normal task:
9745 current->sched_class = &fair_sched_class;
9747 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9748 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9751 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9752 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9754 /* May be allocated at isolcpus cmdline parse time */
9755 if (cpu_isolated_map == NULL)
9756 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9761 scheduler_running = 1;
9764 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9765 static inline int preempt_count_equals(int preempt_offset)
9767 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9769 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9772 void __might_sleep(char *file, int line, int preempt_offset)
9775 static unsigned long prev_jiffy; /* ratelimiting */
9777 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9778 system_state != SYSTEM_RUNNING || oops_in_progress)
9780 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9782 prev_jiffy = jiffies;
9785 "BUG: sleeping function called from invalid context at %s:%d\n",
9788 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9789 in_atomic(), irqs_disabled(),
9790 current->pid, current->comm);
9792 debug_show_held_locks(current);
9793 if (irqs_disabled())
9794 print_irqtrace_events(current);
9798 EXPORT_SYMBOL(__might_sleep);
9801 #ifdef CONFIG_MAGIC_SYSRQ
9802 static void normalize_task(struct rq *rq, struct task_struct *p)
9806 update_rq_clock(rq);
9807 on_rq = p->se.on_rq;
9809 deactivate_task(rq, p, 0);
9810 __setscheduler(rq, p, SCHED_NORMAL, 0);
9812 activate_task(rq, p, 0);
9813 resched_task(rq->curr);
9817 void normalize_rt_tasks(void)
9819 struct task_struct *g, *p;
9820 unsigned long flags;
9823 read_lock_irqsave(&tasklist_lock, flags);
9824 do_each_thread(g, p) {
9826 * Only normalize user tasks:
9831 p->se.exec_start = 0;
9832 #ifdef CONFIG_SCHEDSTATS
9833 p->se.wait_start = 0;
9834 p->se.sleep_start = 0;
9835 p->se.block_start = 0;
9840 * Renice negative nice level userspace
9843 if (TASK_NICE(p) < 0 && p->mm)
9844 set_user_nice(p, 0);
9848 spin_lock(&p->pi_lock);
9849 rq = __task_rq_lock(p);
9851 normalize_task(rq, p);
9853 __task_rq_unlock(rq);
9854 spin_unlock(&p->pi_lock);
9855 } while_each_thread(g, p);
9857 read_unlock_irqrestore(&tasklist_lock, flags);
9860 #endif /* CONFIG_MAGIC_SYSRQ */
9864 * These functions are only useful for the IA64 MCA handling.
9866 * They can only be called when the whole system has been
9867 * stopped - every CPU needs to be quiescent, and no scheduling
9868 * activity can take place. Using them for anything else would
9869 * be a serious bug, and as a result, they aren't even visible
9870 * under any other configuration.
9874 * curr_task - return the current task for a given cpu.
9875 * @cpu: the processor in question.
9877 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9879 struct task_struct *curr_task(int cpu)
9881 return cpu_curr(cpu);
9885 * set_curr_task - set the current task for a given cpu.
9886 * @cpu: the processor in question.
9887 * @p: the task pointer to set.
9889 * Description: This function must only be used when non-maskable interrupts
9890 * are serviced on a separate stack. It allows the architecture to switch the
9891 * notion of the current task on a cpu in a non-blocking manner. This function
9892 * must be called with all CPU's synchronized, and interrupts disabled, the
9893 * and caller must save the original value of the current task (see
9894 * curr_task() above) and restore that value before reenabling interrupts and
9895 * re-starting the system.
9897 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9899 void set_curr_task(int cpu, struct task_struct *p)
9906 #ifdef CONFIG_FAIR_GROUP_SCHED
9907 static void free_fair_sched_group(struct task_group *tg)
9911 for_each_possible_cpu(i) {
9913 kfree(tg->cfs_rq[i]);
9923 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9925 struct cfs_rq *cfs_rq;
9926 struct sched_entity *se;
9930 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9933 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9937 tg->shares = NICE_0_LOAD;
9939 for_each_possible_cpu(i) {
9942 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9943 GFP_KERNEL, cpu_to_node(i));
9947 se = kzalloc_node(sizeof(struct sched_entity),
9948 GFP_KERNEL, cpu_to_node(i));
9952 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9961 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9963 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9964 &cpu_rq(cpu)->leaf_cfs_rq_list);
9967 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9969 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9971 #else /* !CONFG_FAIR_GROUP_SCHED */
9972 static inline void free_fair_sched_group(struct task_group *tg)
9977 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9982 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9986 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9989 #endif /* CONFIG_FAIR_GROUP_SCHED */
9991 #ifdef CONFIG_RT_GROUP_SCHED
9992 static void free_rt_sched_group(struct task_group *tg)
9996 destroy_rt_bandwidth(&tg->rt_bandwidth);
9998 for_each_possible_cpu(i) {
10000 kfree(tg->rt_rq[i]);
10002 kfree(tg->rt_se[i]);
10010 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
10012 struct rt_rq *rt_rq;
10013 struct sched_rt_entity *rt_se;
10017 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
10020 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
10024 init_rt_bandwidth(&tg->rt_bandwidth,
10025 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
10027 for_each_possible_cpu(i) {
10030 rt_rq = kzalloc_node(sizeof(struct rt_rq),
10031 GFP_KERNEL, cpu_to_node(i));
10035 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
10036 GFP_KERNEL, cpu_to_node(i));
10040 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
10049 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10051 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
10052 &cpu_rq(cpu)->leaf_rt_rq_list);
10055 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10057 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
10059 #else /* !CONFIG_RT_GROUP_SCHED */
10060 static inline void free_rt_sched_group(struct task_group *tg)
10065 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
10070 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10074 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10077 #endif /* CONFIG_RT_GROUP_SCHED */
10079 #ifdef CONFIG_GROUP_SCHED
10080 static void free_sched_group(struct task_group *tg)
10082 free_fair_sched_group(tg);
10083 free_rt_sched_group(tg);
10087 /* allocate runqueue etc for a new task group */
10088 struct task_group *sched_create_group(struct task_group *parent)
10090 struct task_group *tg;
10091 unsigned long flags;
10094 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10096 return ERR_PTR(-ENOMEM);
10098 if (!alloc_fair_sched_group(tg, parent))
10101 if (!alloc_rt_sched_group(tg, parent))
10104 spin_lock_irqsave(&task_group_lock, flags);
10105 for_each_possible_cpu(i) {
10106 register_fair_sched_group(tg, i);
10107 register_rt_sched_group(tg, i);
10109 list_add_rcu(&tg->list, &task_groups);
10111 WARN_ON(!parent); /* root should already exist */
10113 tg->parent = parent;
10114 INIT_LIST_HEAD(&tg->children);
10115 list_add_rcu(&tg->siblings, &parent->children);
10116 spin_unlock_irqrestore(&task_group_lock, flags);
10121 free_sched_group(tg);
10122 return ERR_PTR(-ENOMEM);
10125 /* rcu callback to free various structures associated with a task group */
10126 static void free_sched_group_rcu(struct rcu_head *rhp)
10128 /* now it should be safe to free those cfs_rqs */
10129 free_sched_group(container_of(rhp, struct task_group, rcu));
10132 /* Destroy runqueue etc associated with a task group */
10133 void sched_destroy_group(struct task_group *tg)
10135 unsigned long flags;
10138 spin_lock_irqsave(&task_group_lock, flags);
10139 for_each_possible_cpu(i) {
10140 unregister_fair_sched_group(tg, i);
10141 unregister_rt_sched_group(tg, i);
10143 list_del_rcu(&tg->list);
10144 list_del_rcu(&tg->siblings);
10145 spin_unlock_irqrestore(&task_group_lock, flags);
10147 /* wait for possible concurrent references to cfs_rqs complete */
10148 call_rcu(&tg->rcu, free_sched_group_rcu);
10151 /* change task's runqueue when it moves between groups.
10152 * The caller of this function should have put the task in its new group
10153 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10154 * reflect its new group.
10156 void sched_move_task(struct task_struct *tsk)
10158 int on_rq, running;
10159 unsigned long flags;
10162 rq = task_rq_lock(tsk, &flags);
10164 update_rq_clock(rq);
10166 running = task_current(rq, tsk);
10167 on_rq = tsk->se.on_rq;
10170 dequeue_task(rq, tsk, 0);
10171 if (unlikely(running))
10172 tsk->sched_class->put_prev_task(rq, tsk);
10174 set_task_rq(tsk, task_cpu(tsk));
10176 #ifdef CONFIG_FAIR_GROUP_SCHED
10177 if (tsk->sched_class->moved_group)
10178 tsk->sched_class->moved_group(tsk, on_rq);
10181 if (unlikely(running))
10182 tsk->sched_class->set_curr_task(rq);
10184 enqueue_task(rq, tsk, 0, false);
10186 task_rq_unlock(rq, &flags);
10188 #endif /* CONFIG_GROUP_SCHED */
10190 #ifdef CONFIG_FAIR_GROUP_SCHED
10191 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10193 struct cfs_rq *cfs_rq = se->cfs_rq;
10198 dequeue_entity(cfs_rq, se, 0);
10200 se->load.weight = shares;
10201 se->load.inv_weight = 0;
10204 enqueue_entity(cfs_rq, se, 0);
10207 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10209 struct cfs_rq *cfs_rq = se->cfs_rq;
10210 struct rq *rq = cfs_rq->rq;
10211 unsigned long flags;
10213 spin_lock_irqsave(&rq->lock, flags);
10214 __set_se_shares(se, shares);
10215 spin_unlock_irqrestore(&rq->lock, flags);
10218 static DEFINE_MUTEX(shares_mutex);
10220 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10223 unsigned long flags;
10226 * We can't change the weight of the root cgroup.
10231 if (shares < MIN_SHARES)
10232 shares = MIN_SHARES;
10233 else if (shares > MAX_SHARES)
10234 shares = MAX_SHARES;
10236 mutex_lock(&shares_mutex);
10237 if (tg->shares == shares)
10240 spin_lock_irqsave(&task_group_lock, flags);
10241 for_each_possible_cpu(i)
10242 unregister_fair_sched_group(tg, i);
10243 list_del_rcu(&tg->siblings);
10244 spin_unlock_irqrestore(&task_group_lock, flags);
10246 /* wait for any ongoing reference to this group to finish */
10247 synchronize_sched();
10250 * Now we are free to modify the group's share on each cpu
10251 * w/o tripping rebalance_share or load_balance_fair.
10253 tg->shares = shares;
10254 for_each_possible_cpu(i) {
10256 * force a rebalance
10258 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10259 set_se_shares(tg->se[i], shares);
10263 * Enable load balance activity on this group, by inserting it back on
10264 * each cpu's rq->leaf_cfs_rq_list.
10266 spin_lock_irqsave(&task_group_lock, flags);
10267 for_each_possible_cpu(i)
10268 register_fair_sched_group(tg, i);
10269 list_add_rcu(&tg->siblings, &tg->parent->children);
10270 spin_unlock_irqrestore(&task_group_lock, flags);
10272 mutex_unlock(&shares_mutex);
10276 unsigned long sched_group_shares(struct task_group *tg)
10282 #ifdef CONFIG_RT_GROUP_SCHED
10284 * Ensure that the real time constraints are schedulable.
10286 static DEFINE_MUTEX(rt_constraints_mutex);
10288 static unsigned long to_ratio(u64 period, u64 runtime)
10290 if (runtime == RUNTIME_INF)
10293 return div64_u64(runtime << 20, period);
10296 /* Must be called with tasklist_lock held */
10297 static inline int tg_has_rt_tasks(struct task_group *tg)
10299 struct task_struct *g, *p;
10301 do_each_thread(g, p) {
10302 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10304 } while_each_thread(g, p);
10309 struct rt_schedulable_data {
10310 struct task_group *tg;
10315 static int tg_schedulable(struct task_group *tg, void *data)
10317 struct rt_schedulable_data *d = data;
10318 struct task_group *child;
10319 unsigned long total, sum = 0;
10320 u64 period, runtime;
10322 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10323 runtime = tg->rt_bandwidth.rt_runtime;
10326 period = d->rt_period;
10327 runtime = d->rt_runtime;
10330 #ifdef CONFIG_USER_SCHED
10331 if (tg == &root_task_group) {
10332 period = global_rt_period();
10333 runtime = global_rt_runtime();
10338 * Cannot have more runtime than the period.
10340 if (runtime > period && runtime != RUNTIME_INF)
10344 * Ensure we don't starve existing RT tasks.
10346 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10349 total = to_ratio(period, runtime);
10352 * Nobody can have more than the global setting allows.
10354 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10358 * The sum of our children's runtime should not exceed our own.
10360 list_for_each_entry_rcu(child, &tg->children, siblings) {
10361 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10362 runtime = child->rt_bandwidth.rt_runtime;
10364 if (child == d->tg) {
10365 period = d->rt_period;
10366 runtime = d->rt_runtime;
10369 sum += to_ratio(period, runtime);
10378 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10380 struct rt_schedulable_data data = {
10382 .rt_period = period,
10383 .rt_runtime = runtime,
10386 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10389 static int tg_set_bandwidth(struct task_group *tg,
10390 u64 rt_period, u64 rt_runtime)
10394 mutex_lock(&rt_constraints_mutex);
10395 read_lock(&tasklist_lock);
10396 err = __rt_schedulable(tg, rt_period, rt_runtime);
10400 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10401 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10402 tg->rt_bandwidth.rt_runtime = rt_runtime;
10404 for_each_possible_cpu(i) {
10405 struct rt_rq *rt_rq = tg->rt_rq[i];
10407 spin_lock(&rt_rq->rt_runtime_lock);
10408 rt_rq->rt_runtime = rt_runtime;
10409 spin_unlock(&rt_rq->rt_runtime_lock);
10411 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10413 read_unlock(&tasklist_lock);
10414 mutex_unlock(&rt_constraints_mutex);
10419 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10421 u64 rt_runtime, rt_period;
10423 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10424 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10425 if (rt_runtime_us < 0)
10426 rt_runtime = RUNTIME_INF;
10428 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10431 long sched_group_rt_runtime(struct task_group *tg)
10435 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10438 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10439 do_div(rt_runtime_us, NSEC_PER_USEC);
10440 return rt_runtime_us;
10443 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10445 u64 rt_runtime, rt_period;
10447 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10448 rt_runtime = tg->rt_bandwidth.rt_runtime;
10450 if (rt_period == 0)
10453 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10456 long sched_group_rt_period(struct task_group *tg)
10460 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10461 do_div(rt_period_us, NSEC_PER_USEC);
10462 return rt_period_us;
10465 static int sched_rt_global_constraints(void)
10467 u64 runtime, period;
10470 if (sysctl_sched_rt_period <= 0)
10473 runtime = global_rt_runtime();
10474 period = global_rt_period();
10477 * Sanity check on the sysctl variables.
10479 if (runtime > period && runtime != RUNTIME_INF)
10482 mutex_lock(&rt_constraints_mutex);
10483 read_lock(&tasklist_lock);
10484 ret = __rt_schedulable(NULL, 0, 0);
10485 read_unlock(&tasklist_lock);
10486 mutex_unlock(&rt_constraints_mutex);
10491 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10493 /* Don't accept realtime tasks when there is no way for them to run */
10494 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10500 #else /* !CONFIG_RT_GROUP_SCHED */
10501 static int sched_rt_global_constraints(void)
10503 unsigned long flags;
10506 if (sysctl_sched_rt_period <= 0)
10510 * There's always some RT tasks in the root group
10511 * -- migration, kstopmachine etc..
10513 if (sysctl_sched_rt_runtime == 0)
10516 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10517 for_each_possible_cpu(i) {
10518 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10520 spin_lock(&rt_rq->rt_runtime_lock);
10521 rt_rq->rt_runtime = global_rt_runtime();
10522 spin_unlock(&rt_rq->rt_runtime_lock);
10524 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10528 #endif /* CONFIG_RT_GROUP_SCHED */
10530 int sched_rt_handler(struct ctl_table *table, int write,
10531 void __user *buffer, size_t *lenp,
10535 int old_period, old_runtime;
10536 static DEFINE_MUTEX(mutex);
10538 mutex_lock(&mutex);
10539 old_period = sysctl_sched_rt_period;
10540 old_runtime = sysctl_sched_rt_runtime;
10542 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10544 if (!ret && write) {
10545 ret = sched_rt_global_constraints();
10547 sysctl_sched_rt_period = old_period;
10548 sysctl_sched_rt_runtime = old_runtime;
10550 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10551 def_rt_bandwidth.rt_period =
10552 ns_to_ktime(global_rt_period());
10555 mutex_unlock(&mutex);
10560 #ifdef CONFIG_CGROUP_SCHED
10562 /* return corresponding task_group object of a cgroup */
10563 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10565 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10566 struct task_group, css);
10569 static struct cgroup_subsys_state *
10570 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10572 struct task_group *tg, *parent;
10574 if (!cgrp->parent) {
10575 /* This is early initialization for the top cgroup */
10576 return &init_task_group.css;
10579 parent = cgroup_tg(cgrp->parent);
10580 tg = sched_create_group(parent);
10582 return ERR_PTR(-ENOMEM);
10588 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10590 struct task_group *tg = cgroup_tg(cgrp);
10592 sched_destroy_group(tg);
10596 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10598 #ifdef CONFIG_RT_GROUP_SCHED
10599 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10602 /* We don't support RT-tasks being in separate groups */
10603 if (tsk->sched_class != &fair_sched_class)
10610 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10611 struct task_struct *tsk, bool threadgroup)
10613 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10617 struct task_struct *c;
10619 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10620 retval = cpu_cgroup_can_attach_task(cgrp, c);
10632 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10633 struct cgroup *old_cont, struct task_struct *tsk,
10636 sched_move_task(tsk);
10638 struct task_struct *c;
10640 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10641 sched_move_task(c);
10647 #ifdef CONFIG_FAIR_GROUP_SCHED
10648 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10651 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10654 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10656 struct task_group *tg = cgroup_tg(cgrp);
10658 return (u64) tg->shares;
10660 #endif /* CONFIG_FAIR_GROUP_SCHED */
10662 #ifdef CONFIG_RT_GROUP_SCHED
10663 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10666 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10669 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10671 return sched_group_rt_runtime(cgroup_tg(cgrp));
10674 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10677 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10680 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10682 return sched_group_rt_period(cgroup_tg(cgrp));
10684 #endif /* CONFIG_RT_GROUP_SCHED */
10686 static struct cftype cpu_files[] = {
10687 #ifdef CONFIG_FAIR_GROUP_SCHED
10690 .read_u64 = cpu_shares_read_u64,
10691 .write_u64 = cpu_shares_write_u64,
10694 #ifdef CONFIG_RT_GROUP_SCHED
10696 .name = "rt_runtime_us",
10697 .read_s64 = cpu_rt_runtime_read,
10698 .write_s64 = cpu_rt_runtime_write,
10701 .name = "rt_period_us",
10702 .read_u64 = cpu_rt_period_read_uint,
10703 .write_u64 = cpu_rt_period_write_uint,
10708 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10710 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10713 struct cgroup_subsys cpu_cgroup_subsys = {
10715 .create = cpu_cgroup_create,
10716 .destroy = cpu_cgroup_destroy,
10717 .can_attach = cpu_cgroup_can_attach,
10718 .attach = cpu_cgroup_attach,
10719 .populate = cpu_cgroup_populate,
10720 .subsys_id = cpu_cgroup_subsys_id,
10724 #endif /* CONFIG_CGROUP_SCHED */
10726 #ifdef CONFIG_CGROUP_CPUACCT
10729 * CPU accounting code for task groups.
10731 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10732 * (balbir@in.ibm.com).
10735 /* track cpu usage of a group of tasks and its child groups */
10737 struct cgroup_subsys_state css;
10738 /* cpuusage holds pointer to a u64-type object on every cpu */
10740 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10741 struct cpuacct *parent;
10744 struct cgroup_subsys cpuacct_subsys;
10746 /* return cpu accounting group corresponding to this container */
10747 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10749 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10750 struct cpuacct, css);
10753 /* return cpu accounting group to which this task belongs */
10754 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10756 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10757 struct cpuacct, css);
10760 /* create a new cpu accounting group */
10761 static struct cgroup_subsys_state *cpuacct_create(
10762 struct cgroup_subsys *ss, struct cgroup *cgrp)
10764 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10770 ca->cpuusage = alloc_percpu(u64);
10774 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10775 if (percpu_counter_init(&ca->cpustat[i], 0))
10776 goto out_free_counters;
10779 ca->parent = cgroup_ca(cgrp->parent);
10785 percpu_counter_destroy(&ca->cpustat[i]);
10786 free_percpu(ca->cpuusage);
10790 return ERR_PTR(-ENOMEM);
10793 /* destroy an existing cpu accounting group */
10795 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10797 struct cpuacct *ca = cgroup_ca(cgrp);
10800 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10801 percpu_counter_destroy(&ca->cpustat[i]);
10802 free_percpu(ca->cpuusage);
10806 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10808 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10811 #ifndef CONFIG_64BIT
10813 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10815 spin_lock_irq(&cpu_rq(cpu)->lock);
10817 spin_unlock_irq(&cpu_rq(cpu)->lock);
10825 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10827 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10829 #ifndef CONFIG_64BIT
10831 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10833 spin_lock_irq(&cpu_rq(cpu)->lock);
10835 spin_unlock_irq(&cpu_rq(cpu)->lock);
10841 /* return total cpu usage (in nanoseconds) of a group */
10842 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10844 struct cpuacct *ca = cgroup_ca(cgrp);
10845 u64 totalcpuusage = 0;
10848 for_each_present_cpu(i)
10849 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10851 return totalcpuusage;
10854 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10857 struct cpuacct *ca = cgroup_ca(cgrp);
10866 for_each_present_cpu(i)
10867 cpuacct_cpuusage_write(ca, i, 0);
10873 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10874 struct seq_file *m)
10876 struct cpuacct *ca = cgroup_ca(cgroup);
10880 for_each_present_cpu(i) {
10881 percpu = cpuacct_cpuusage_read(ca, i);
10882 seq_printf(m, "%llu ", (unsigned long long) percpu);
10884 seq_printf(m, "\n");
10888 static const char *cpuacct_stat_desc[] = {
10889 [CPUACCT_STAT_USER] = "user",
10890 [CPUACCT_STAT_SYSTEM] = "system",
10893 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10894 struct cgroup_map_cb *cb)
10896 struct cpuacct *ca = cgroup_ca(cgrp);
10899 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10900 s64 val = percpu_counter_read(&ca->cpustat[i]);
10901 val = cputime64_to_clock_t(val);
10902 cb->fill(cb, cpuacct_stat_desc[i], val);
10907 static struct cftype files[] = {
10910 .read_u64 = cpuusage_read,
10911 .write_u64 = cpuusage_write,
10914 .name = "usage_percpu",
10915 .read_seq_string = cpuacct_percpu_seq_read,
10919 .read_map = cpuacct_stats_show,
10923 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10925 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10929 * charge this task's execution time to its accounting group.
10931 * called with rq->lock held.
10933 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10935 struct cpuacct *ca;
10938 if (unlikely(!cpuacct_subsys.active))
10941 cpu = task_cpu(tsk);
10947 for (; ca; ca = ca->parent) {
10948 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10949 *cpuusage += cputime;
10956 * Charge the system/user time to the task's accounting group.
10958 static void cpuacct_update_stats(struct task_struct *tsk,
10959 enum cpuacct_stat_index idx, cputime_t val)
10961 struct cpuacct *ca;
10963 if (unlikely(!cpuacct_subsys.active))
10970 percpu_counter_add(&ca->cpustat[idx], val);
10976 struct cgroup_subsys cpuacct_subsys = {
10978 .create = cpuacct_create,
10979 .destroy = cpuacct_destroy,
10980 .populate = cpuacct_populate,
10981 .subsys_id = cpuacct_subsys_id,
10983 #endif /* CONFIG_CGROUP_CPUACCT */
10987 int rcu_expedited_torture_stats(char *page)
10991 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10993 void synchronize_sched_expedited(void)
10996 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10998 #else /* #ifndef CONFIG_SMP */
11000 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
11001 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
11003 #define RCU_EXPEDITED_STATE_POST -2
11004 #define RCU_EXPEDITED_STATE_IDLE -1
11006 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11008 int rcu_expedited_torture_stats(char *page)
11013 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
11014 for_each_online_cpu(cpu) {
11015 cnt += sprintf(&page[cnt], " %d:%d",
11016 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
11018 cnt += sprintf(&page[cnt], "\n");
11021 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
11023 static long synchronize_sched_expedited_count;
11026 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
11027 * approach to force grace period to end quickly. This consumes
11028 * significant time on all CPUs, and is thus not recommended for
11029 * any sort of common-case code.
11031 * Note that it is illegal to call this function while holding any
11032 * lock that is acquired by a CPU-hotplug notifier. Failing to
11033 * observe this restriction will result in deadlock.
11035 void synchronize_sched_expedited(void)
11038 unsigned long flags;
11039 bool need_full_sync = 0;
11041 struct migration_req *req;
11045 smp_mb(); /* ensure prior mod happens before capturing snap. */
11046 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
11048 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
11050 if (trycount++ < 10)
11051 udelay(trycount * num_online_cpus());
11053 synchronize_sched();
11056 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
11057 smp_mb(); /* ensure test happens before caller kfree */
11062 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
11063 for_each_online_cpu(cpu) {
11065 req = &per_cpu(rcu_migration_req, cpu);
11066 init_completion(&req->done);
11068 req->dest_cpu = RCU_MIGRATION_NEED_QS;
11069 spin_lock_irqsave(&rq->lock, flags);
11070 list_add(&req->list, &rq->migration_queue);
11071 spin_unlock_irqrestore(&rq->lock, flags);
11072 wake_up_process(rq->migration_thread);
11074 for_each_online_cpu(cpu) {
11075 rcu_expedited_state = cpu;
11076 req = &per_cpu(rcu_migration_req, cpu);
11078 wait_for_completion(&req->done);
11079 spin_lock_irqsave(&rq->lock, flags);
11080 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
11081 need_full_sync = 1;
11082 req->dest_cpu = RCU_MIGRATION_IDLE;
11083 spin_unlock_irqrestore(&rq->lock, flags);
11085 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11086 mutex_unlock(&rcu_sched_expedited_mutex);
11088 if (need_full_sync)
11089 synchronize_sched();
11091 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11093 #endif /* #else #ifndef CONFIG_SMP */