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 raw_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 raw_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 raw_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 raw_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_var);
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 raw_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 char in_nohz_recently;
540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load;
542 unsigned long nr_load_updates;
548 #ifdef CONFIG_FAIR_GROUP_SCHED
549 /* list of leaf cfs_rq on this cpu: */
550 struct list_head leaf_cfs_rq_list;
552 #ifdef CONFIG_RT_GROUP_SCHED
553 struct list_head leaf_rt_rq_list;
557 * This is part of a global counter where only the total sum
558 * over all CPUs matters. A task can increase this counter on
559 * one CPU and if it got migrated afterwards it may decrease
560 * it on another CPU. Always updated under the runqueue lock:
562 unsigned long nr_uninterruptible;
564 struct task_struct *curr, *idle;
565 unsigned long next_balance;
566 struct mm_struct *prev_mm;
573 struct root_domain *rd;
574 struct sched_domain *sd;
576 unsigned char idle_at_tick;
577 /* For active balancing */
581 /* cpu of this runqueue: */
585 unsigned long avg_load_per_task;
587 struct task_struct *migration_thread;
588 struct list_head migration_queue;
596 /* calc_load related fields */
597 unsigned long calc_load_update;
598 long calc_load_active;
600 #ifdef CONFIG_SCHED_HRTICK
602 int hrtick_csd_pending;
603 struct call_single_data hrtick_csd;
605 struct hrtimer hrtick_timer;
608 #ifdef CONFIG_SCHEDSTATS
610 struct sched_info rq_sched_info;
611 unsigned long long rq_cpu_time;
612 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
614 /* sys_sched_yield() stats */
615 unsigned int yld_count;
617 /* schedule() stats */
618 unsigned int sched_switch;
619 unsigned int sched_count;
620 unsigned int sched_goidle;
622 /* try_to_wake_up() stats */
623 unsigned int ttwu_count;
624 unsigned int ttwu_local;
627 unsigned int bkl_count;
631 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
634 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
636 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
639 static inline int cpu_of(struct rq *rq)
648 #define rcu_dereference_check_sched_domain(p) \
649 rcu_dereference_check((p), \
650 rcu_read_lock_sched_held() || \
651 lockdep_is_held(&sched_domains_mutex))
654 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
655 * See detach_destroy_domains: synchronize_sched for details.
657 * The domain tree of any CPU may only be accessed from within
658 * preempt-disabled sections.
660 #define for_each_domain(cpu, __sd) \
661 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
663 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
664 #define this_rq() (&__get_cpu_var(runqueues))
665 #define task_rq(p) cpu_rq(task_cpu(p))
666 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
667 #define raw_rq() (&__raw_get_cpu_var(runqueues))
669 inline void update_rq_clock(struct rq *rq)
671 rq->clock = sched_clock_cpu(cpu_of(rq));
675 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
677 #ifdef CONFIG_SCHED_DEBUG
678 # define const_debug __read_mostly
680 # define const_debug static const
685 * @cpu: the processor in question.
687 * Returns true if the current cpu runqueue is locked.
688 * This interface allows printk to be called with the runqueue lock
689 * held and know whether or not it is OK to wake up the klogd.
691 int runqueue_is_locked(int cpu)
693 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
697 * Debugging: various feature bits
700 #define SCHED_FEAT(name, enabled) \
701 __SCHED_FEAT_##name ,
704 #include "sched_features.h"
709 #define SCHED_FEAT(name, enabled) \
710 (1UL << __SCHED_FEAT_##name) * enabled |
712 const_debug unsigned int sysctl_sched_features =
713 #include "sched_features.h"
718 #ifdef CONFIG_SCHED_DEBUG
719 #define SCHED_FEAT(name, enabled) \
722 static __read_mostly char *sched_feat_names[] = {
723 #include "sched_features.h"
729 static int sched_feat_show(struct seq_file *m, void *v)
733 for (i = 0; sched_feat_names[i]; i++) {
734 if (!(sysctl_sched_features & (1UL << i)))
736 seq_printf(m, "%s ", sched_feat_names[i]);
744 sched_feat_write(struct file *filp, const char __user *ubuf,
745 size_t cnt, loff_t *ppos)
755 if (copy_from_user(&buf, ubuf, cnt))
760 if (strncmp(buf, "NO_", 3) == 0) {
765 for (i = 0; sched_feat_names[i]; i++) {
766 int len = strlen(sched_feat_names[i]);
768 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
770 sysctl_sched_features &= ~(1UL << i);
772 sysctl_sched_features |= (1UL << i);
777 if (!sched_feat_names[i])
785 static int sched_feat_open(struct inode *inode, struct file *filp)
787 return single_open(filp, sched_feat_show, NULL);
790 static const struct file_operations sched_feat_fops = {
791 .open = sched_feat_open,
792 .write = sched_feat_write,
795 .release = single_release,
798 static __init int sched_init_debug(void)
800 debugfs_create_file("sched_features", 0644, NULL, NULL,
805 late_initcall(sched_init_debug);
809 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
812 * Number of tasks to iterate in a single balance run.
813 * Limited because this is done with IRQs disabled.
815 const_debug unsigned int sysctl_sched_nr_migrate = 32;
818 * ratelimit for updating the group shares.
821 unsigned int sysctl_sched_shares_ratelimit = 250000;
822 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
825 * Inject some fuzzyness into changing the per-cpu group shares
826 * this avoids remote rq-locks at the expense of fairness.
829 unsigned int sysctl_sched_shares_thresh = 4;
832 * period over which we average the RT time consumption, measured
837 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
840 * period over which we measure -rt task cpu usage in us.
843 unsigned int sysctl_sched_rt_period = 1000000;
845 static __read_mostly int scheduler_running;
848 * part of the period that we allow rt tasks to run in us.
851 int sysctl_sched_rt_runtime = 950000;
853 static inline u64 global_rt_period(void)
855 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
858 static inline u64 global_rt_runtime(void)
860 if (sysctl_sched_rt_runtime < 0)
863 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
866 #ifndef prepare_arch_switch
867 # define prepare_arch_switch(next) do { } while (0)
869 #ifndef finish_arch_switch
870 # define finish_arch_switch(prev) do { } while (0)
873 static inline int task_current(struct rq *rq, struct task_struct *p)
875 return rq->curr == p;
878 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
879 static inline int task_running(struct rq *rq, struct task_struct *p)
881 return task_current(rq, p);
884 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
888 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
890 #ifdef CONFIG_DEBUG_SPINLOCK
891 /* this is a valid case when another task releases the spinlock */
892 rq->lock.owner = current;
895 * If we are tracking spinlock dependencies then we have to
896 * fix up the runqueue lock - which gets 'carried over' from
899 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
901 raw_spin_unlock_irq(&rq->lock);
904 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
905 static inline int task_running(struct rq *rq, struct task_struct *p)
910 return task_current(rq, p);
914 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
918 * We can optimise this out completely for !SMP, because the
919 * SMP rebalancing from interrupt is the only thing that cares
924 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
925 raw_spin_unlock_irq(&rq->lock);
927 raw_spin_unlock(&rq->lock);
931 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
935 * After ->oncpu is cleared, the task can be moved to a different CPU.
936 * We must ensure this doesn't happen until the switch is completely
942 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
946 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
949 * Check whether the task is waking, we use this to synchronize against
950 * ttwu() so that task_cpu() reports a stable number.
952 * We need to make an exception for PF_STARTING tasks because the fork
953 * path might require task_rq_lock() to work, eg. it can call
954 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
956 static inline int task_is_waking(struct task_struct *p)
958 return unlikely((p->state == TASK_WAKING) && !(p->flags & PF_STARTING));
962 * __task_rq_lock - lock the runqueue a given task resides on.
963 * Must be called interrupts disabled.
965 static inline struct rq *__task_rq_lock(struct task_struct *p)
971 while (task_is_waking(p))
974 raw_spin_lock(&rq->lock);
975 if (likely(rq == task_rq(p) && !task_is_waking(p)))
977 raw_spin_unlock(&rq->lock);
982 * task_rq_lock - lock the runqueue a given task resides on and disable
983 * interrupts. Note the ordering: we can safely lookup the task_rq without
984 * explicitly disabling preemption.
986 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
992 while (task_is_waking(p))
994 local_irq_save(*flags);
996 raw_spin_lock(&rq->lock);
997 if (likely(rq == task_rq(p) && !task_is_waking(p)))
999 raw_spin_unlock_irqrestore(&rq->lock, *flags);
1003 void task_rq_unlock_wait(struct task_struct *p)
1005 struct rq *rq = task_rq(p);
1007 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1008 raw_spin_unlock_wait(&rq->lock);
1011 static void __task_rq_unlock(struct rq *rq)
1012 __releases(rq->lock)
1014 raw_spin_unlock(&rq->lock);
1017 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1018 __releases(rq->lock)
1020 raw_spin_unlock_irqrestore(&rq->lock, *flags);
1024 * this_rq_lock - lock this runqueue and disable interrupts.
1026 static struct rq *this_rq_lock(void)
1027 __acquires(rq->lock)
1031 local_irq_disable();
1033 raw_spin_lock(&rq->lock);
1038 #ifdef CONFIG_SCHED_HRTICK
1040 * Use HR-timers to deliver accurate preemption points.
1042 * Its all a bit involved since we cannot program an hrt while holding the
1043 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1046 * When we get rescheduled we reprogram the hrtick_timer outside of the
1052 * - enabled by features
1053 * - hrtimer is actually high res
1055 static inline int hrtick_enabled(struct rq *rq)
1057 if (!sched_feat(HRTICK))
1059 if (!cpu_active(cpu_of(rq)))
1061 return hrtimer_is_hres_active(&rq->hrtick_timer);
1064 static void hrtick_clear(struct rq *rq)
1066 if (hrtimer_active(&rq->hrtick_timer))
1067 hrtimer_cancel(&rq->hrtick_timer);
1071 * High-resolution timer tick.
1072 * Runs from hardirq context with interrupts disabled.
1074 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1076 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1078 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1080 raw_spin_lock(&rq->lock);
1081 update_rq_clock(rq);
1082 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1083 raw_spin_unlock(&rq->lock);
1085 return HRTIMER_NORESTART;
1090 * called from hardirq (IPI) context
1092 static void __hrtick_start(void *arg)
1094 struct rq *rq = arg;
1096 raw_spin_lock(&rq->lock);
1097 hrtimer_restart(&rq->hrtick_timer);
1098 rq->hrtick_csd_pending = 0;
1099 raw_spin_unlock(&rq->lock);
1103 * Called to set the hrtick timer state.
1105 * called with rq->lock held and irqs disabled
1107 static void hrtick_start(struct rq *rq, u64 delay)
1109 struct hrtimer *timer = &rq->hrtick_timer;
1110 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1112 hrtimer_set_expires(timer, time);
1114 if (rq == this_rq()) {
1115 hrtimer_restart(timer);
1116 } else if (!rq->hrtick_csd_pending) {
1117 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1118 rq->hrtick_csd_pending = 1;
1123 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1125 int cpu = (int)(long)hcpu;
1128 case CPU_UP_CANCELED:
1129 case CPU_UP_CANCELED_FROZEN:
1130 case CPU_DOWN_PREPARE:
1131 case CPU_DOWN_PREPARE_FROZEN:
1133 case CPU_DEAD_FROZEN:
1134 hrtick_clear(cpu_rq(cpu));
1141 static __init void init_hrtick(void)
1143 hotcpu_notifier(hotplug_hrtick, 0);
1147 * Called to set the hrtick timer state.
1149 * called with rq->lock held and irqs disabled
1151 static void hrtick_start(struct rq *rq, u64 delay)
1153 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1154 HRTIMER_MODE_REL_PINNED, 0);
1157 static inline void init_hrtick(void)
1160 #endif /* CONFIG_SMP */
1162 static void init_rq_hrtick(struct rq *rq)
1165 rq->hrtick_csd_pending = 0;
1167 rq->hrtick_csd.flags = 0;
1168 rq->hrtick_csd.func = __hrtick_start;
1169 rq->hrtick_csd.info = rq;
1172 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1173 rq->hrtick_timer.function = hrtick;
1175 #else /* CONFIG_SCHED_HRTICK */
1176 static inline void hrtick_clear(struct rq *rq)
1180 static inline void init_rq_hrtick(struct rq *rq)
1184 static inline void init_hrtick(void)
1187 #endif /* CONFIG_SCHED_HRTICK */
1190 * resched_task - mark a task 'to be rescheduled now'.
1192 * On UP this means the setting of the need_resched flag, on SMP it
1193 * might also involve a cross-CPU call to trigger the scheduler on
1198 #ifndef tsk_is_polling
1199 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1202 static void resched_task(struct task_struct *p)
1206 assert_raw_spin_locked(&task_rq(p)->lock);
1208 if (test_tsk_need_resched(p))
1211 set_tsk_need_resched(p);
1214 if (cpu == smp_processor_id())
1217 /* NEED_RESCHED must be visible before we test polling */
1219 if (!tsk_is_polling(p))
1220 smp_send_reschedule(cpu);
1223 static void resched_cpu(int cpu)
1225 struct rq *rq = cpu_rq(cpu);
1226 unsigned long flags;
1228 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1230 resched_task(cpu_curr(cpu));
1231 raw_spin_unlock_irqrestore(&rq->lock, flags);
1236 * When add_timer_on() enqueues a timer into the timer wheel of an
1237 * idle CPU then this timer might expire before the next timer event
1238 * which is scheduled to wake up that CPU. In case of a completely
1239 * idle system the next event might even be infinite time into the
1240 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1241 * leaves the inner idle loop so the newly added timer is taken into
1242 * account when the CPU goes back to idle and evaluates the timer
1243 * wheel for the next timer event.
1245 void wake_up_idle_cpu(int cpu)
1247 struct rq *rq = cpu_rq(cpu);
1249 if (cpu == smp_processor_id())
1253 * This is safe, as this function is called with the timer
1254 * wheel base lock of (cpu) held. When the CPU is on the way
1255 * to idle and has not yet set rq->curr to idle then it will
1256 * be serialized on the timer wheel base lock and take the new
1257 * timer into account automatically.
1259 if (rq->curr != rq->idle)
1263 * We can set TIF_RESCHED on the idle task of the other CPU
1264 * lockless. The worst case is that the other CPU runs the
1265 * idle task through an additional NOOP schedule()
1267 set_tsk_need_resched(rq->idle);
1269 /* NEED_RESCHED must be visible before we test polling */
1271 if (!tsk_is_polling(rq->idle))
1272 smp_send_reschedule(cpu);
1274 #endif /* CONFIG_NO_HZ */
1276 static u64 sched_avg_period(void)
1278 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1281 static void sched_avg_update(struct rq *rq)
1283 s64 period = sched_avg_period();
1285 while ((s64)(rq->clock - rq->age_stamp) > period) {
1286 rq->age_stamp += period;
1291 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1293 rq->rt_avg += rt_delta;
1294 sched_avg_update(rq);
1297 #else /* !CONFIG_SMP */
1298 static void resched_task(struct task_struct *p)
1300 assert_raw_spin_locked(&task_rq(p)->lock);
1301 set_tsk_need_resched(p);
1304 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1307 #endif /* CONFIG_SMP */
1309 #if BITS_PER_LONG == 32
1310 # define WMULT_CONST (~0UL)
1312 # define WMULT_CONST (1UL << 32)
1315 #define WMULT_SHIFT 32
1318 * Shift right and round:
1320 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1323 * delta *= weight / lw
1325 static unsigned long
1326 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1327 struct load_weight *lw)
1331 if (!lw->inv_weight) {
1332 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1335 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1339 tmp = (u64)delta_exec * weight;
1341 * Check whether we'd overflow the 64-bit multiplication:
1343 if (unlikely(tmp > WMULT_CONST))
1344 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1347 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1349 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1352 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1358 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1365 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1366 * of tasks with abnormal "nice" values across CPUs the contribution that
1367 * each task makes to its run queue's load is weighted according to its
1368 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1369 * scaled version of the new time slice allocation that they receive on time
1373 #define WEIGHT_IDLEPRIO 3
1374 #define WMULT_IDLEPRIO 1431655765
1377 * Nice levels are multiplicative, with a gentle 10% change for every
1378 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1379 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1380 * that remained on nice 0.
1382 * The "10% effect" is relative and cumulative: from _any_ nice level,
1383 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1384 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1385 * If a task goes up by ~10% and another task goes down by ~10% then
1386 * the relative distance between them is ~25%.)
1388 static const int prio_to_weight[40] = {
1389 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1390 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1391 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1392 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1393 /* 0 */ 1024, 820, 655, 526, 423,
1394 /* 5 */ 335, 272, 215, 172, 137,
1395 /* 10 */ 110, 87, 70, 56, 45,
1396 /* 15 */ 36, 29, 23, 18, 15,
1400 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1402 * In cases where the weight does not change often, we can use the
1403 * precalculated inverse to speed up arithmetics by turning divisions
1404 * into multiplications:
1406 static const u32 prio_to_wmult[40] = {
1407 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1408 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1409 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1410 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1411 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1412 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1413 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1414 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1417 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1420 * runqueue iterator, to support SMP load-balancing between different
1421 * scheduling classes, without having to expose their internal data
1422 * structures to the load-balancing proper:
1424 struct rq_iterator {
1426 struct task_struct *(*start)(void *);
1427 struct task_struct *(*next)(void *);
1431 static unsigned long
1432 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1433 unsigned long max_load_move, struct sched_domain *sd,
1434 enum cpu_idle_type idle, int *all_pinned,
1435 int *this_best_prio, struct rq_iterator *iterator);
1438 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1439 struct sched_domain *sd, enum cpu_idle_type idle,
1440 struct rq_iterator *iterator);
1443 /* Time spent by the tasks of the cpu accounting group executing in ... */
1444 enum cpuacct_stat_index {
1445 CPUACCT_STAT_USER, /* ... user mode */
1446 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1448 CPUACCT_STAT_NSTATS,
1451 #ifdef CONFIG_CGROUP_CPUACCT
1452 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1453 static void cpuacct_update_stats(struct task_struct *tsk,
1454 enum cpuacct_stat_index idx, cputime_t val);
1456 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1457 static inline void cpuacct_update_stats(struct task_struct *tsk,
1458 enum cpuacct_stat_index idx, cputime_t val) {}
1461 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1463 update_load_add(&rq->load, load);
1466 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1468 update_load_sub(&rq->load, load);
1471 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1472 typedef int (*tg_visitor)(struct task_group *, void *);
1475 * Iterate the full tree, calling @down when first entering a node and @up when
1476 * leaving it for the final time.
1478 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1480 struct task_group *parent, *child;
1484 parent = &root_task_group;
1486 ret = (*down)(parent, data);
1489 list_for_each_entry_rcu(child, &parent->children, siblings) {
1496 ret = (*up)(parent, data);
1501 parent = parent->parent;
1510 static int tg_nop(struct task_group *tg, void *data)
1517 /* Used instead of source_load when we know the type == 0 */
1518 static unsigned long weighted_cpuload(const int cpu)
1520 return cpu_rq(cpu)->load.weight;
1524 * Return a low guess at the load of a migration-source cpu weighted
1525 * according to the scheduling class and "nice" value.
1527 * We want to under-estimate the load of migration sources, to
1528 * balance conservatively.
1530 static unsigned long source_load(int cpu, int type)
1532 struct rq *rq = cpu_rq(cpu);
1533 unsigned long total = weighted_cpuload(cpu);
1535 if (type == 0 || !sched_feat(LB_BIAS))
1538 return min(rq->cpu_load[type-1], total);
1542 * Return a high guess at the load of a migration-target cpu weighted
1543 * according to the scheduling class and "nice" value.
1545 static unsigned long target_load(int cpu, int type)
1547 struct rq *rq = cpu_rq(cpu);
1548 unsigned long total = weighted_cpuload(cpu);
1550 if (type == 0 || !sched_feat(LB_BIAS))
1553 return max(rq->cpu_load[type-1], total);
1556 static struct sched_group *group_of(int cpu)
1558 struct sched_domain *sd = rcu_dereference_sched(cpu_rq(cpu)->sd);
1566 static unsigned long power_of(int cpu)
1568 struct sched_group *group = group_of(cpu);
1571 return SCHED_LOAD_SCALE;
1573 return group->cpu_power;
1576 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1578 static unsigned long cpu_avg_load_per_task(int cpu)
1580 struct rq *rq = cpu_rq(cpu);
1581 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1584 rq->avg_load_per_task = rq->load.weight / nr_running;
1586 rq->avg_load_per_task = 0;
1588 return rq->avg_load_per_task;
1591 #ifdef CONFIG_FAIR_GROUP_SCHED
1593 static __read_mostly unsigned long *update_shares_data;
1595 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1598 * Calculate and set the cpu's group shares.
1600 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1601 unsigned long sd_shares,
1602 unsigned long sd_rq_weight,
1603 unsigned long *usd_rq_weight)
1605 unsigned long shares, rq_weight;
1608 rq_weight = usd_rq_weight[cpu];
1611 rq_weight = NICE_0_LOAD;
1615 * \Sum_j shares_j * rq_weight_i
1616 * shares_i = -----------------------------
1617 * \Sum_j rq_weight_j
1619 shares = (sd_shares * rq_weight) / sd_rq_weight;
1620 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1622 if (abs(shares - tg->se[cpu]->load.weight) >
1623 sysctl_sched_shares_thresh) {
1624 struct rq *rq = cpu_rq(cpu);
1625 unsigned long flags;
1627 raw_spin_lock_irqsave(&rq->lock, flags);
1628 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1629 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1630 __set_se_shares(tg->se[cpu], shares);
1631 raw_spin_unlock_irqrestore(&rq->lock, flags);
1636 * Re-compute the task group their per cpu shares over the given domain.
1637 * This needs to be done in a bottom-up fashion because the rq weight of a
1638 * parent group depends on the shares of its child groups.
1640 static int tg_shares_up(struct task_group *tg, void *data)
1642 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1643 unsigned long *usd_rq_weight;
1644 struct sched_domain *sd = data;
1645 unsigned long flags;
1651 local_irq_save(flags);
1652 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1654 for_each_cpu(i, sched_domain_span(sd)) {
1655 weight = tg->cfs_rq[i]->load.weight;
1656 usd_rq_weight[i] = weight;
1658 rq_weight += weight;
1660 * If there are currently no tasks on the cpu pretend there
1661 * is one of average load so that when a new task gets to
1662 * run here it will not get delayed by group starvation.
1665 weight = NICE_0_LOAD;
1667 sum_weight += weight;
1668 shares += tg->cfs_rq[i]->shares;
1672 rq_weight = sum_weight;
1674 if ((!shares && rq_weight) || shares > tg->shares)
1675 shares = tg->shares;
1677 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1678 shares = tg->shares;
1680 for_each_cpu(i, sched_domain_span(sd))
1681 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1683 local_irq_restore(flags);
1689 * Compute the cpu's hierarchical load factor for each task group.
1690 * This needs to be done in a top-down fashion because the load of a child
1691 * group is a fraction of its parents load.
1693 static int tg_load_down(struct task_group *tg, void *data)
1696 long cpu = (long)data;
1699 load = cpu_rq(cpu)->load.weight;
1701 load = tg->parent->cfs_rq[cpu]->h_load;
1702 load *= tg->cfs_rq[cpu]->shares;
1703 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1706 tg->cfs_rq[cpu]->h_load = load;
1711 static void update_shares(struct sched_domain *sd)
1716 if (root_task_group_empty())
1719 now = cpu_clock(raw_smp_processor_id());
1720 elapsed = now - sd->last_update;
1722 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1723 sd->last_update = now;
1724 walk_tg_tree(tg_nop, tg_shares_up, sd);
1728 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1730 if (root_task_group_empty())
1733 raw_spin_unlock(&rq->lock);
1735 raw_spin_lock(&rq->lock);
1738 static void update_h_load(long cpu)
1740 if (root_task_group_empty())
1743 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1748 static inline void update_shares(struct sched_domain *sd)
1752 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1758 #ifdef CONFIG_PREEMPT
1760 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1763 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1764 * way at the expense of forcing extra atomic operations in all
1765 * invocations. This assures that the double_lock is acquired using the
1766 * same underlying policy as the spinlock_t on this architecture, which
1767 * reduces latency compared to the unfair variant below. However, it
1768 * also adds more overhead and therefore may reduce throughput.
1770 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1771 __releases(this_rq->lock)
1772 __acquires(busiest->lock)
1773 __acquires(this_rq->lock)
1775 raw_spin_unlock(&this_rq->lock);
1776 double_rq_lock(this_rq, busiest);
1783 * Unfair double_lock_balance: Optimizes throughput at the expense of
1784 * latency by eliminating extra atomic operations when the locks are
1785 * already in proper order on entry. This favors lower cpu-ids and will
1786 * grant the double lock to lower cpus over higher ids under contention,
1787 * regardless of entry order into the function.
1789 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1790 __releases(this_rq->lock)
1791 __acquires(busiest->lock)
1792 __acquires(this_rq->lock)
1796 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1797 if (busiest < this_rq) {
1798 raw_spin_unlock(&this_rq->lock);
1799 raw_spin_lock(&busiest->lock);
1800 raw_spin_lock_nested(&this_rq->lock,
1801 SINGLE_DEPTH_NESTING);
1804 raw_spin_lock_nested(&busiest->lock,
1805 SINGLE_DEPTH_NESTING);
1810 #endif /* CONFIG_PREEMPT */
1813 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1815 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1817 if (unlikely(!irqs_disabled())) {
1818 /* printk() doesn't work good under rq->lock */
1819 raw_spin_unlock(&this_rq->lock);
1823 return _double_lock_balance(this_rq, busiest);
1826 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1827 __releases(busiest->lock)
1829 raw_spin_unlock(&busiest->lock);
1830 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1834 #ifdef CONFIG_FAIR_GROUP_SCHED
1835 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1838 cfs_rq->shares = shares;
1843 static void calc_load_account_active(struct rq *this_rq);
1844 static void update_sysctl(void);
1845 static int get_update_sysctl_factor(void);
1847 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1849 set_task_rq(p, cpu);
1852 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1853 * successfuly executed on another CPU. We must ensure that updates of
1854 * per-task data have been completed by this moment.
1857 task_thread_info(p)->cpu = cpu;
1861 #include "sched_stats.h"
1862 #include "sched_idletask.c"
1863 #include "sched_fair.c"
1864 #include "sched_rt.c"
1865 #ifdef CONFIG_SCHED_DEBUG
1866 # include "sched_debug.c"
1869 #define sched_class_highest (&rt_sched_class)
1870 #define for_each_class(class) \
1871 for (class = sched_class_highest; class; class = class->next)
1873 static void inc_nr_running(struct rq *rq)
1878 static void dec_nr_running(struct rq *rq)
1883 static void set_load_weight(struct task_struct *p)
1885 if (task_has_rt_policy(p)) {
1886 p->se.load.weight = prio_to_weight[0] * 2;
1887 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1892 * SCHED_IDLE tasks get minimal weight:
1894 if (p->policy == SCHED_IDLE) {
1895 p->se.load.weight = WEIGHT_IDLEPRIO;
1896 p->se.load.inv_weight = WMULT_IDLEPRIO;
1900 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1901 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1904 static void update_avg(u64 *avg, u64 sample)
1906 s64 diff = sample - *avg;
1910 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1913 p->se.start_runtime = p->se.sum_exec_runtime;
1915 sched_info_queued(p);
1916 p->sched_class->enqueue_task(rq, p, wakeup);
1920 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1923 if (p->se.last_wakeup) {
1924 update_avg(&p->se.avg_overlap,
1925 p->se.sum_exec_runtime - p->se.last_wakeup);
1926 p->se.last_wakeup = 0;
1928 update_avg(&p->se.avg_wakeup,
1929 sysctl_sched_wakeup_granularity);
1933 sched_info_dequeued(p);
1934 p->sched_class->dequeue_task(rq, p, sleep);
1939 * __normal_prio - return the priority that is based on the static prio
1941 static inline int __normal_prio(struct task_struct *p)
1943 return p->static_prio;
1947 * Calculate the expected normal priority: i.e. priority
1948 * without taking RT-inheritance into account. Might be
1949 * boosted by interactivity modifiers. Changes upon fork,
1950 * setprio syscalls, and whenever the interactivity
1951 * estimator recalculates.
1953 static inline int normal_prio(struct task_struct *p)
1957 if (task_has_rt_policy(p))
1958 prio = MAX_RT_PRIO-1 - p->rt_priority;
1960 prio = __normal_prio(p);
1965 * Calculate the current priority, i.e. the priority
1966 * taken into account by the scheduler. This value might
1967 * be boosted by RT tasks, or might be boosted by
1968 * interactivity modifiers. Will be RT if the task got
1969 * RT-boosted. If not then it returns p->normal_prio.
1971 static int effective_prio(struct task_struct *p)
1973 p->normal_prio = normal_prio(p);
1975 * If we are RT tasks or we were boosted to RT priority,
1976 * keep the priority unchanged. Otherwise, update priority
1977 * to the normal priority:
1979 if (!rt_prio(p->prio))
1980 return p->normal_prio;
1985 * activate_task - move a task to the runqueue.
1987 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1989 if (task_contributes_to_load(p))
1990 rq->nr_uninterruptible--;
1992 enqueue_task(rq, p, wakeup);
1997 * deactivate_task - remove a task from the runqueue.
1999 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
2001 if (task_contributes_to_load(p))
2002 rq->nr_uninterruptible++;
2004 dequeue_task(rq, p, sleep);
2009 * task_curr - is this task currently executing on a CPU?
2010 * @p: the task in question.
2012 inline int task_curr(const struct task_struct *p)
2014 return cpu_curr(task_cpu(p)) == p;
2017 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2018 const struct sched_class *prev_class,
2019 int oldprio, int running)
2021 if (prev_class != p->sched_class) {
2022 if (prev_class->switched_from)
2023 prev_class->switched_from(rq, p, running);
2024 p->sched_class->switched_to(rq, p, running);
2026 p->sched_class->prio_changed(rq, p, oldprio, running);
2031 * Is this task likely cache-hot:
2034 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2038 if (p->sched_class != &fair_sched_class)
2042 * Buddy candidates are cache hot:
2044 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2045 (&p->se == cfs_rq_of(&p->se)->next ||
2046 &p->se == cfs_rq_of(&p->se)->last))
2049 if (sysctl_sched_migration_cost == -1)
2051 if (sysctl_sched_migration_cost == 0)
2054 delta = now - p->se.exec_start;
2056 return delta < (s64)sysctl_sched_migration_cost;
2059 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2061 #ifdef CONFIG_SCHED_DEBUG
2063 * We should never call set_task_cpu() on a blocked task,
2064 * ttwu() will sort out the placement.
2066 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2067 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2070 trace_sched_migrate_task(p, new_cpu);
2072 if (task_cpu(p) != new_cpu) {
2073 p->se.nr_migrations++;
2074 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2077 __set_task_cpu(p, new_cpu);
2080 struct migration_req {
2081 struct list_head list;
2083 struct task_struct *task;
2086 struct completion done;
2090 * The task's runqueue lock must be held.
2091 * Returns true if you have to wait for migration thread.
2094 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2096 struct rq *rq = task_rq(p);
2099 * If the task is not on a runqueue (and not running), then
2100 * the next wake-up will properly place the task.
2102 if (!p->se.on_rq && !task_running(rq, p))
2105 init_completion(&req->done);
2107 req->dest_cpu = dest_cpu;
2108 list_add(&req->list, &rq->migration_queue);
2114 * wait_task_context_switch - wait for a thread to complete at least one
2117 * @p must not be current.
2119 void wait_task_context_switch(struct task_struct *p)
2121 unsigned long nvcsw, nivcsw, flags;
2129 * The runqueue is assigned before the actual context
2130 * switch. We need to take the runqueue lock.
2132 * We could check initially without the lock but it is
2133 * very likely that we need to take the lock in every
2136 rq = task_rq_lock(p, &flags);
2137 running = task_running(rq, p);
2138 task_rq_unlock(rq, &flags);
2140 if (likely(!running))
2143 * The switch count is incremented before the actual
2144 * context switch. We thus wait for two switches to be
2145 * sure at least one completed.
2147 if ((p->nvcsw - nvcsw) > 1)
2149 if ((p->nivcsw - nivcsw) > 1)
2157 * wait_task_inactive - wait for a thread to unschedule.
2159 * If @match_state is nonzero, it's the @p->state value just checked and
2160 * not expected to change. If it changes, i.e. @p might have woken up,
2161 * then return zero. When we succeed in waiting for @p to be off its CPU,
2162 * we return a positive number (its total switch count). If a second call
2163 * a short while later returns the same number, the caller can be sure that
2164 * @p has remained unscheduled the whole time.
2166 * The caller must ensure that the task *will* unschedule sometime soon,
2167 * else this function might spin for a *long* time. This function can't
2168 * be called with interrupts off, or it may introduce deadlock with
2169 * smp_call_function() if an IPI is sent by the same process we are
2170 * waiting to become inactive.
2172 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2174 unsigned long flags;
2181 * We do the initial early heuristics without holding
2182 * any task-queue locks at all. We'll only try to get
2183 * the runqueue lock when things look like they will
2189 * If the task is actively running on another CPU
2190 * still, just relax and busy-wait without holding
2193 * NOTE! Since we don't hold any locks, it's not
2194 * even sure that "rq" stays as the right runqueue!
2195 * But we don't care, since "task_running()" will
2196 * return false if the runqueue has changed and p
2197 * is actually now running somewhere else!
2199 while (task_running(rq, p)) {
2200 if (match_state && unlikely(p->state != match_state))
2206 * Ok, time to look more closely! We need the rq
2207 * lock now, to be *sure*. If we're wrong, we'll
2208 * just go back and repeat.
2210 rq = task_rq_lock(p, &flags);
2211 trace_sched_wait_task(rq, p);
2212 running = task_running(rq, p);
2213 on_rq = p->se.on_rq;
2215 if (!match_state || p->state == match_state)
2216 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2217 task_rq_unlock(rq, &flags);
2220 * If it changed from the expected state, bail out now.
2222 if (unlikely(!ncsw))
2226 * Was it really running after all now that we
2227 * checked with the proper locks actually held?
2229 * Oops. Go back and try again..
2231 if (unlikely(running)) {
2237 * It's not enough that it's not actively running,
2238 * it must be off the runqueue _entirely_, and not
2241 * So if it was still runnable (but just not actively
2242 * running right now), it's preempted, and we should
2243 * yield - it could be a while.
2245 if (unlikely(on_rq)) {
2246 schedule_timeout_uninterruptible(1);
2251 * Ahh, all good. It wasn't running, and it wasn't
2252 * runnable, which means that it will never become
2253 * running in the future either. We're all done!
2262 * kick_process - kick a running thread to enter/exit the kernel
2263 * @p: the to-be-kicked thread
2265 * Cause a process which is running on another CPU to enter
2266 * kernel-mode, without any delay. (to get signals handled.)
2268 * NOTE: this function doesnt have to take the runqueue lock,
2269 * because all it wants to ensure is that the remote task enters
2270 * the kernel. If the IPI races and the task has been migrated
2271 * to another CPU then no harm is done and the purpose has been
2274 void kick_process(struct task_struct *p)
2280 if ((cpu != smp_processor_id()) && task_curr(p))
2281 smp_send_reschedule(cpu);
2284 EXPORT_SYMBOL_GPL(kick_process);
2285 #endif /* CONFIG_SMP */
2288 * task_oncpu_function_call - call a function on the cpu on which a task runs
2289 * @p: the task to evaluate
2290 * @func: the function to be called
2291 * @info: the function call argument
2293 * Calls the function @func when the task is currently running. This might
2294 * be on the current CPU, which just calls the function directly
2296 void task_oncpu_function_call(struct task_struct *p,
2297 void (*func) (void *info), void *info)
2304 smp_call_function_single(cpu, func, info, 1);
2309 static int select_fallback_rq(int cpu, struct task_struct *p)
2312 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2314 /* Look for allowed, online CPU in same node. */
2315 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2316 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2319 /* Any allowed, online CPU? */
2320 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2321 if (dest_cpu < nr_cpu_ids)
2324 /* No more Mr. Nice Guy. */
2325 if (dest_cpu >= nr_cpu_ids) {
2327 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2329 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2332 * Don't tell them about moving exiting tasks or
2333 * kernel threads (both mm NULL), since they never
2336 if (p->mm && printk_ratelimit()) {
2337 printk(KERN_INFO "process %d (%s) no "
2338 "longer affine to cpu%d\n",
2339 task_pid_nr(p), p->comm, cpu);
2347 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2348 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2351 * exec: is unstable, retry loop
2352 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2355 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2357 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2360 * In order not to call set_task_cpu() on a blocking task we need
2361 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2364 * Since this is common to all placement strategies, this lives here.
2366 * [ this allows ->select_task() to simply return task_cpu(p) and
2367 * not worry about this generic constraint ]
2369 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2371 cpu = select_fallback_rq(task_cpu(p), p);
2378 * try_to_wake_up - wake up a thread
2379 * @p: the to-be-woken-up thread
2380 * @state: the mask of task states that can be woken
2381 * @sync: do a synchronous wakeup?
2383 * Put it on the run-queue if it's not already there. The "current"
2384 * thread is always on the run-queue (except when the actual
2385 * re-schedule is in progress), and as such you're allowed to do
2386 * the simpler "current->state = TASK_RUNNING" to mark yourself
2387 * runnable without the overhead of this.
2389 * returns failure only if the task is already active.
2391 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2394 int cpu, orig_cpu, this_cpu, success = 0;
2395 unsigned long flags;
2396 struct rq *rq, *orig_rq;
2398 if (!sched_feat(SYNC_WAKEUPS))
2399 wake_flags &= ~WF_SYNC;
2401 this_cpu = get_cpu();
2404 rq = orig_rq = task_rq_lock(p, &flags);
2405 update_rq_clock(rq);
2406 if (!(p->state & state))
2416 if (unlikely(task_running(rq, p)))
2420 * In order to handle concurrent wakeups and release the rq->lock
2421 * we put the task in TASK_WAKING state.
2423 * First fix up the nr_uninterruptible count:
2425 if (task_contributes_to_load(p))
2426 rq->nr_uninterruptible--;
2427 p->state = TASK_WAKING;
2429 if (p->sched_class->task_waking)
2430 p->sched_class->task_waking(rq, p);
2432 __task_rq_unlock(rq);
2434 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2435 if (cpu != orig_cpu) {
2437 * Since we migrate the task without holding any rq->lock,
2438 * we need to be careful with task_rq_lock(), since that
2439 * might end up locking an invalid rq.
2441 set_task_cpu(p, cpu);
2445 raw_spin_lock(&rq->lock);
2446 update_rq_clock(rq);
2449 * We migrated the task without holding either rq->lock, however
2450 * since the task is not on the task list itself, nobody else
2451 * will try and migrate the task, hence the rq should match the
2452 * cpu we just moved it to.
2454 WARN_ON(task_cpu(p) != cpu);
2455 WARN_ON(p->state != TASK_WAKING);
2457 #ifdef CONFIG_SCHEDSTATS
2458 schedstat_inc(rq, ttwu_count);
2459 if (cpu == this_cpu)
2460 schedstat_inc(rq, ttwu_local);
2462 struct sched_domain *sd;
2463 for_each_domain(this_cpu, sd) {
2464 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2465 schedstat_inc(sd, ttwu_wake_remote);
2470 #endif /* CONFIG_SCHEDSTATS */
2473 #endif /* CONFIG_SMP */
2474 schedstat_inc(p, se.nr_wakeups);
2475 if (wake_flags & WF_SYNC)
2476 schedstat_inc(p, se.nr_wakeups_sync);
2477 if (orig_cpu != cpu)
2478 schedstat_inc(p, se.nr_wakeups_migrate);
2479 if (cpu == this_cpu)
2480 schedstat_inc(p, se.nr_wakeups_local);
2482 schedstat_inc(p, se.nr_wakeups_remote);
2483 activate_task(rq, p, 1);
2487 * Only attribute actual wakeups done by this task.
2489 if (!in_interrupt()) {
2490 struct sched_entity *se = ¤t->se;
2491 u64 sample = se->sum_exec_runtime;
2493 if (se->last_wakeup)
2494 sample -= se->last_wakeup;
2496 sample -= se->start_runtime;
2497 update_avg(&se->avg_wakeup, sample);
2499 se->last_wakeup = se->sum_exec_runtime;
2503 trace_sched_wakeup(rq, p, success);
2504 check_preempt_curr(rq, p, wake_flags);
2506 p->state = TASK_RUNNING;
2508 if (p->sched_class->task_woken)
2509 p->sched_class->task_woken(rq, p);
2511 if (unlikely(rq->idle_stamp)) {
2512 u64 delta = rq->clock - rq->idle_stamp;
2513 u64 max = 2*sysctl_sched_migration_cost;
2518 update_avg(&rq->avg_idle, delta);
2523 task_rq_unlock(rq, &flags);
2530 * wake_up_process - Wake up a specific process
2531 * @p: The process to be woken up.
2533 * Attempt to wake up the nominated process and move it to the set of runnable
2534 * processes. Returns 1 if the process was woken up, 0 if it was already
2537 * It may be assumed that this function implies a write memory barrier before
2538 * changing the task state if and only if any tasks are woken up.
2540 int wake_up_process(struct task_struct *p)
2542 return try_to_wake_up(p, TASK_ALL, 0);
2544 EXPORT_SYMBOL(wake_up_process);
2546 int wake_up_state(struct task_struct *p, unsigned int state)
2548 return try_to_wake_up(p, state, 0);
2552 * Perform scheduler related setup for a newly forked process p.
2553 * p is forked by current.
2555 * __sched_fork() is basic setup used by init_idle() too:
2557 static void __sched_fork(struct task_struct *p)
2559 p->se.exec_start = 0;
2560 p->se.sum_exec_runtime = 0;
2561 p->se.prev_sum_exec_runtime = 0;
2562 p->se.nr_migrations = 0;
2563 p->se.last_wakeup = 0;
2564 p->se.avg_overlap = 0;
2565 p->se.start_runtime = 0;
2566 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2568 #ifdef CONFIG_SCHEDSTATS
2569 p->se.wait_start = 0;
2571 p->se.wait_count = 0;
2574 p->se.sleep_start = 0;
2575 p->se.sleep_max = 0;
2576 p->se.sum_sleep_runtime = 0;
2578 p->se.block_start = 0;
2579 p->se.block_max = 0;
2581 p->se.slice_max = 0;
2583 p->se.nr_migrations_cold = 0;
2584 p->se.nr_failed_migrations_affine = 0;
2585 p->se.nr_failed_migrations_running = 0;
2586 p->se.nr_failed_migrations_hot = 0;
2587 p->se.nr_forced_migrations = 0;
2589 p->se.nr_wakeups = 0;
2590 p->se.nr_wakeups_sync = 0;
2591 p->se.nr_wakeups_migrate = 0;
2592 p->se.nr_wakeups_local = 0;
2593 p->se.nr_wakeups_remote = 0;
2594 p->se.nr_wakeups_affine = 0;
2595 p->se.nr_wakeups_affine_attempts = 0;
2596 p->se.nr_wakeups_passive = 0;
2597 p->se.nr_wakeups_idle = 0;
2601 INIT_LIST_HEAD(&p->rt.run_list);
2603 INIT_LIST_HEAD(&p->se.group_node);
2605 #ifdef CONFIG_PREEMPT_NOTIFIERS
2606 INIT_HLIST_HEAD(&p->preempt_notifiers);
2611 * fork()/clone()-time setup:
2613 void sched_fork(struct task_struct *p, int clone_flags)
2615 int cpu = get_cpu();
2619 * We mark the process as waking here. This guarantees that
2620 * nobody will actually run it, and a signal or other external
2621 * event cannot wake it up and insert it on the runqueue either.
2623 p->state = TASK_WAKING;
2626 * Revert to default priority/policy on fork if requested.
2628 if (unlikely(p->sched_reset_on_fork)) {
2629 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2630 p->policy = SCHED_NORMAL;
2631 p->normal_prio = p->static_prio;
2634 if (PRIO_TO_NICE(p->static_prio) < 0) {
2635 p->static_prio = NICE_TO_PRIO(0);
2636 p->normal_prio = p->static_prio;
2641 * We don't need the reset flag anymore after the fork. It has
2642 * fulfilled its duty:
2644 p->sched_reset_on_fork = 0;
2648 * Make sure we do not leak PI boosting priority to the child.
2650 p->prio = current->normal_prio;
2652 if (!rt_prio(p->prio))
2653 p->sched_class = &fair_sched_class;
2655 if (p->sched_class->task_fork)
2656 p->sched_class->task_fork(p);
2658 set_task_cpu(p, cpu);
2660 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2661 if (likely(sched_info_on()))
2662 memset(&p->sched_info, 0, sizeof(p->sched_info));
2664 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2667 #ifdef CONFIG_PREEMPT
2668 /* Want to start with kernel preemption disabled. */
2669 task_thread_info(p)->preempt_count = 1;
2671 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2677 * wake_up_new_task - wake up a newly created task for the first time.
2679 * This function will do some initial scheduler statistics housekeeping
2680 * that must be done for every newly created context, then puts the task
2681 * on the runqueue and wakes it.
2683 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2685 unsigned long flags;
2687 int cpu = get_cpu();
2691 * Fork balancing, do it here and not earlier because:
2692 * - cpus_allowed can change in the fork path
2693 * - any previously selected cpu might disappear through hotplug
2695 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2696 * ->cpus_allowed is stable, we have preemption disabled, meaning
2697 * cpu_online_mask is stable.
2699 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2700 set_task_cpu(p, cpu);
2704 * Since the task is not on the rq and we still have TASK_WAKING set
2705 * nobody else will migrate this task.
2708 raw_spin_lock_irqsave(&rq->lock, flags);
2710 BUG_ON(p->state != TASK_WAKING);
2711 p->state = TASK_RUNNING;
2712 update_rq_clock(rq);
2713 activate_task(rq, p, 0);
2714 trace_sched_wakeup_new(rq, p, 1);
2715 check_preempt_curr(rq, p, WF_FORK);
2717 if (p->sched_class->task_woken)
2718 p->sched_class->task_woken(rq, p);
2720 task_rq_unlock(rq, &flags);
2724 #ifdef CONFIG_PREEMPT_NOTIFIERS
2727 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2728 * @notifier: notifier struct to register
2730 void preempt_notifier_register(struct preempt_notifier *notifier)
2732 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2734 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2737 * preempt_notifier_unregister - no longer interested in preemption notifications
2738 * @notifier: notifier struct to unregister
2740 * This is safe to call from within a preemption notifier.
2742 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2744 hlist_del(¬ifier->link);
2746 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2748 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2750 struct preempt_notifier *notifier;
2751 struct hlist_node *node;
2753 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2754 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2758 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2759 struct task_struct *next)
2761 struct preempt_notifier *notifier;
2762 struct hlist_node *node;
2764 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2765 notifier->ops->sched_out(notifier, next);
2768 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2770 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2775 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2776 struct task_struct *next)
2780 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2783 * prepare_task_switch - prepare to switch tasks
2784 * @rq: the runqueue preparing to switch
2785 * @prev: the current task that is being switched out
2786 * @next: the task we are going to switch to.
2788 * This is called with the rq lock held and interrupts off. It must
2789 * be paired with a subsequent finish_task_switch after the context
2792 * prepare_task_switch sets up locking and calls architecture specific
2796 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2797 struct task_struct *next)
2799 fire_sched_out_preempt_notifiers(prev, next);
2800 prepare_lock_switch(rq, next);
2801 prepare_arch_switch(next);
2805 * finish_task_switch - clean up after a task-switch
2806 * @rq: runqueue associated with task-switch
2807 * @prev: the thread we just switched away from.
2809 * finish_task_switch must be called after the context switch, paired
2810 * with a prepare_task_switch call before the context switch.
2811 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2812 * and do any other architecture-specific cleanup actions.
2814 * Note that we may have delayed dropping an mm in context_switch(). If
2815 * so, we finish that here outside of the runqueue lock. (Doing it
2816 * with the lock held can cause deadlocks; see schedule() for
2819 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2820 __releases(rq->lock)
2822 struct mm_struct *mm = rq->prev_mm;
2828 * A task struct has one reference for the use as "current".
2829 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2830 * schedule one last time. The schedule call will never return, and
2831 * the scheduled task must drop that reference.
2832 * The test for TASK_DEAD must occur while the runqueue locks are
2833 * still held, otherwise prev could be scheduled on another cpu, die
2834 * there before we look at prev->state, and then the reference would
2836 * Manfred Spraul <manfred@colorfullife.com>
2838 prev_state = prev->state;
2839 finish_arch_switch(prev);
2840 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2841 local_irq_disable();
2842 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2843 perf_event_task_sched_in(current);
2844 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2846 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2847 finish_lock_switch(rq, prev);
2849 fire_sched_in_preempt_notifiers(current);
2852 if (unlikely(prev_state == TASK_DEAD)) {
2854 * Remove function-return probe instances associated with this
2855 * task and put them back on the free list.
2857 kprobe_flush_task(prev);
2858 put_task_struct(prev);
2864 /* assumes rq->lock is held */
2865 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2867 if (prev->sched_class->pre_schedule)
2868 prev->sched_class->pre_schedule(rq, prev);
2871 /* rq->lock is NOT held, but preemption is disabled */
2872 static inline void post_schedule(struct rq *rq)
2874 if (rq->post_schedule) {
2875 unsigned long flags;
2877 raw_spin_lock_irqsave(&rq->lock, flags);
2878 if (rq->curr->sched_class->post_schedule)
2879 rq->curr->sched_class->post_schedule(rq);
2880 raw_spin_unlock_irqrestore(&rq->lock, flags);
2882 rq->post_schedule = 0;
2888 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2892 static inline void post_schedule(struct rq *rq)
2899 * schedule_tail - first thing a freshly forked thread must call.
2900 * @prev: the thread we just switched away from.
2902 asmlinkage void schedule_tail(struct task_struct *prev)
2903 __releases(rq->lock)
2905 struct rq *rq = this_rq();
2907 finish_task_switch(rq, prev);
2910 * FIXME: do we need to worry about rq being invalidated by the
2915 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2916 /* In this case, finish_task_switch does not reenable preemption */
2919 if (current->set_child_tid)
2920 put_user(task_pid_vnr(current), current->set_child_tid);
2924 * context_switch - switch to the new MM and the new
2925 * thread's register state.
2928 context_switch(struct rq *rq, struct task_struct *prev,
2929 struct task_struct *next)
2931 struct mm_struct *mm, *oldmm;
2933 prepare_task_switch(rq, prev, next);
2934 trace_sched_switch(rq, prev, next);
2936 oldmm = prev->active_mm;
2938 * For paravirt, this is coupled with an exit in switch_to to
2939 * combine the page table reload and the switch backend into
2942 arch_start_context_switch(prev);
2945 next->active_mm = oldmm;
2946 atomic_inc(&oldmm->mm_count);
2947 enter_lazy_tlb(oldmm, next);
2949 switch_mm(oldmm, mm, next);
2951 if (likely(!prev->mm)) {
2952 prev->active_mm = NULL;
2953 rq->prev_mm = oldmm;
2956 * Since the runqueue lock will be released by the next
2957 * task (which is an invalid locking op but in the case
2958 * of the scheduler it's an obvious special-case), so we
2959 * do an early lockdep release here:
2961 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2962 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2965 /* Here we just switch the register state and the stack. */
2966 switch_to(prev, next, prev);
2970 * this_rq must be evaluated again because prev may have moved
2971 * CPUs since it called schedule(), thus the 'rq' on its stack
2972 * frame will be invalid.
2974 finish_task_switch(this_rq(), prev);
2978 * nr_running, nr_uninterruptible and nr_context_switches:
2980 * externally visible scheduler statistics: current number of runnable
2981 * threads, current number of uninterruptible-sleeping threads, total
2982 * number of context switches performed since bootup.
2984 unsigned long nr_running(void)
2986 unsigned long i, sum = 0;
2988 for_each_online_cpu(i)
2989 sum += cpu_rq(i)->nr_running;
2994 unsigned long nr_uninterruptible(void)
2996 unsigned long i, sum = 0;
2998 for_each_possible_cpu(i)
2999 sum += cpu_rq(i)->nr_uninterruptible;
3002 * Since we read the counters lockless, it might be slightly
3003 * inaccurate. Do not allow it to go below zero though:
3005 if (unlikely((long)sum < 0))
3011 unsigned long long nr_context_switches(void)
3014 unsigned long long sum = 0;
3016 for_each_possible_cpu(i)
3017 sum += cpu_rq(i)->nr_switches;
3022 unsigned long nr_iowait(void)
3024 unsigned long i, sum = 0;
3026 for_each_possible_cpu(i)
3027 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3032 unsigned long nr_iowait_cpu(void)
3034 struct rq *this = this_rq();
3035 return atomic_read(&this->nr_iowait);
3038 unsigned long this_cpu_load(void)
3040 struct rq *this = this_rq();
3041 return this->cpu_load[0];
3045 /* Variables and functions for calc_load */
3046 static atomic_long_t calc_load_tasks;
3047 static unsigned long calc_load_update;
3048 unsigned long avenrun[3];
3049 EXPORT_SYMBOL(avenrun);
3052 * get_avenrun - get the load average array
3053 * @loads: pointer to dest load array
3054 * @offset: offset to add
3055 * @shift: shift count to shift the result left
3057 * These values are estimates at best, so no need for locking.
3059 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3061 loads[0] = (avenrun[0] + offset) << shift;
3062 loads[1] = (avenrun[1] + offset) << shift;
3063 loads[2] = (avenrun[2] + offset) << shift;
3066 static unsigned long
3067 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3070 load += active * (FIXED_1 - exp);
3071 return load >> FSHIFT;
3075 * calc_load - update the avenrun load estimates 10 ticks after the
3076 * CPUs have updated calc_load_tasks.
3078 void calc_global_load(void)
3080 unsigned long upd = calc_load_update + 10;
3083 if (time_before(jiffies, upd))
3086 active = atomic_long_read(&calc_load_tasks);
3087 active = active > 0 ? active * FIXED_1 : 0;
3089 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3090 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3091 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3093 calc_load_update += LOAD_FREQ;
3097 * Either called from update_cpu_load() or from a cpu going idle
3099 static void calc_load_account_active(struct rq *this_rq)
3101 long nr_active, delta;
3103 nr_active = this_rq->nr_running;
3104 nr_active += (long) this_rq->nr_uninterruptible;
3106 if (nr_active != this_rq->calc_load_active) {
3107 delta = nr_active - this_rq->calc_load_active;
3108 this_rq->calc_load_active = nr_active;
3109 atomic_long_add(delta, &calc_load_tasks);
3114 * Update rq->cpu_load[] statistics. This function is usually called every
3115 * scheduler tick (TICK_NSEC).
3117 static void update_cpu_load(struct rq *this_rq)
3119 unsigned long this_load = this_rq->load.weight;
3122 this_rq->nr_load_updates++;
3124 /* Update our load: */
3125 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3126 unsigned long old_load, new_load;
3128 /* scale is effectively 1 << i now, and >> i divides by scale */
3130 old_load = this_rq->cpu_load[i];
3131 new_load = this_load;
3133 * Round up the averaging division if load is increasing. This
3134 * prevents us from getting stuck on 9 if the load is 10, for
3137 if (new_load > old_load)
3138 new_load += scale-1;
3139 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3142 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3143 this_rq->calc_load_update += LOAD_FREQ;
3144 calc_load_account_active(this_rq);
3151 * double_rq_lock - safely lock two runqueues
3153 * Note this does not disable interrupts like task_rq_lock,
3154 * you need to do so manually before calling.
3156 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3157 __acquires(rq1->lock)
3158 __acquires(rq2->lock)
3160 BUG_ON(!irqs_disabled());
3162 raw_spin_lock(&rq1->lock);
3163 __acquire(rq2->lock); /* Fake it out ;) */
3166 raw_spin_lock(&rq1->lock);
3167 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3169 raw_spin_lock(&rq2->lock);
3170 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3173 update_rq_clock(rq1);
3174 update_rq_clock(rq2);
3178 * double_rq_unlock - safely unlock two runqueues
3180 * Note this does not restore interrupts like task_rq_unlock,
3181 * you need to do so manually after calling.
3183 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3184 __releases(rq1->lock)
3185 __releases(rq2->lock)
3187 raw_spin_unlock(&rq1->lock);
3189 raw_spin_unlock(&rq2->lock);
3191 __release(rq2->lock);
3195 * sched_exec - execve() is a valuable balancing opportunity, because at
3196 * this point the task has the smallest effective memory and cache footprint.
3198 void sched_exec(void)
3200 struct task_struct *p = current;
3201 struct migration_req req;
3202 int dest_cpu, this_cpu;
3203 unsigned long flags;
3207 this_cpu = get_cpu();
3208 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3209 if (dest_cpu == this_cpu) {
3214 rq = task_rq_lock(p, &flags);
3218 * select_task_rq() can race against ->cpus_allowed
3220 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3221 || unlikely(!cpu_active(dest_cpu))) {
3222 task_rq_unlock(rq, &flags);
3226 /* force the process onto the specified CPU */
3227 if (migrate_task(p, dest_cpu, &req)) {
3228 /* Need to wait for migration thread (might exit: take ref). */
3229 struct task_struct *mt = rq->migration_thread;
3231 get_task_struct(mt);
3232 task_rq_unlock(rq, &flags);
3233 wake_up_process(mt);
3234 put_task_struct(mt);
3235 wait_for_completion(&req.done);
3239 task_rq_unlock(rq, &flags);
3243 * pull_task - move a task from a remote runqueue to the local runqueue.
3244 * Both runqueues must be locked.
3246 static void pull_task(struct rq *src_rq, struct task_struct *p,
3247 struct rq *this_rq, int this_cpu)
3249 deactivate_task(src_rq, p, 0);
3250 set_task_cpu(p, this_cpu);
3251 activate_task(this_rq, p, 0);
3252 check_preempt_curr(this_rq, p, 0);
3256 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3259 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3260 struct sched_domain *sd, enum cpu_idle_type idle,
3263 int tsk_cache_hot = 0;
3265 * We do not migrate tasks that are:
3266 * 1) running (obviously), or
3267 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3268 * 3) are cache-hot on their current CPU.
3270 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3271 schedstat_inc(p, se.nr_failed_migrations_affine);
3276 if (task_running(rq, p)) {
3277 schedstat_inc(p, se.nr_failed_migrations_running);
3282 * Aggressive migration if:
3283 * 1) task is cache cold, or
3284 * 2) too many balance attempts have failed.
3287 tsk_cache_hot = task_hot(p, rq->clock, sd);
3288 if (!tsk_cache_hot ||
3289 sd->nr_balance_failed > sd->cache_nice_tries) {
3290 #ifdef CONFIG_SCHEDSTATS
3291 if (tsk_cache_hot) {
3292 schedstat_inc(sd, lb_hot_gained[idle]);
3293 schedstat_inc(p, se.nr_forced_migrations);
3299 if (tsk_cache_hot) {
3300 schedstat_inc(p, se.nr_failed_migrations_hot);
3306 static unsigned long
3307 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3308 unsigned long max_load_move, struct sched_domain *sd,
3309 enum cpu_idle_type idle, int *all_pinned,
3310 int *this_best_prio, struct rq_iterator *iterator)
3312 int loops = 0, pulled = 0, pinned = 0;
3313 struct task_struct *p;
3314 long rem_load_move = max_load_move;
3316 if (max_load_move == 0)
3322 * Start the load-balancing iterator:
3324 p = iterator->start(iterator->arg);
3326 if (!p || loops++ > sysctl_sched_nr_migrate)
3329 if ((p->se.load.weight >> 1) > rem_load_move ||
3330 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3331 p = iterator->next(iterator->arg);
3335 pull_task(busiest, p, this_rq, this_cpu);
3337 rem_load_move -= p->se.load.weight;
3339 #ifdef CONFIG_PREEMPT
3341 * NEWIDLE balancing is a source of latency, so preemptible kernels
3342 * will stop after the first task is pulled to minimize the critical
3345 if (idle == CPU_NEWLY_IDLE)
3350 * We only want to steal up to the prescribed amount of weighted load.
3352 if (rem_load_move > 0) {
3353 if (p->prio < *this_best_prio)
3354 *this_best_prio = p->prio;
3355 p = iterator->next(iterator->arg);
3360 * Right now, this is one of only two places pull_task() is called,
3361 * so we can safely collect pull_task() stats here rather than
3362 * inside pull_task().
3364 schedstat_add(sd, lb_gained[idle], pulled);
3367 *all_pinned = pinned;
3369 return max_load_move - rem_load_move;
3373 * move_tasks tries to move up to max_load_move weighted load from busiest to
3374 * this_rq, as part of a balancing operation within domain "sd".
3375 * Returns 1 if successful and 0 otherwise.
3377 * Called with both runqueues locked.
3379 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3380 unsigned long max_load_move,
3381 struct sched_domain *sd, enum cpu_idle_type idle,
3384 const struct sched_class *class = sched_class_highest;
3385 unsigned long total_load_moved = 0;
3386 int this_best_prio = this_rq->curr->prio;
3390 class->load_balance(this_rq, this_cpu, busiest,
3391 max_load_move - total_load_moved,
3392 sd, idle, all_pinned, &this_best_prio);
3393 class = class->next;
3395 #ifdef CONFIG_PREEMPT
3397 * NEWIDLE balancing is a source of latency, so preemptible
3398 * kernels will stop after the first task is pulled to minimize
3399 * the critical section.
3401 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3404 } while (class && max_load_move > total_load_moved);
3406 return total_load_moved > 0;
3410 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3411 struct sched_domain *sd, enum cpu_idle_type idle,
3412 struct rq_iterator *iterator)
3414 struct task_struct *p = iterator->start(iterator->arg);
3418 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3419 pull_task(busiest, p, this_rq, this_cpu);
3421 * Right now, this is only the second place pull_task()
3422 * is called, so we can safely collect pull_task()
3423 * stats here rather than inside pull_task().
3425 schedstat_inc(sd, lb_gained[idle]);
3429 p = iterator->next(iterator->arg);
3436 * move_one_task tries to move exactly one task from busiest to this_rq, as
3437 * part of active balancing operations within "domain".
3438 * Returns 1 if successful and 0 otherwise.
3440 * Called with both runqueues locked.
3442 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3443 struct sched_domain *sd, enum cpu_idle_type idle)
3445 const struct sched_class *class;
3447 for_each_class(class) {
3448 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3454 /********** Helpers for find_busiest_group ************************/
3456 * sd_lb_stats - Structure to store the statistics of a sched_domain
3457 * during load balancing.
3459 struct sd_lb_stats {
3460 struct sched_group *busiest; /* Busiest group in this sd */
3461 struct sched_group *this; /* Local group in this sd */
3462 unsigned long total_load; /* Total load of all groups in sd */
3463 unsigned long total_pwr; /* Total power of all groups in sd */
3464 unsigned long avg_load; /* Average load across all groups in sd */
3466 /** Statistics of this group */
3467 unsigned long this_load;
3468 unsigned long this_load_per_task;
3469 unsigned long this_nr_running;
3471 /* Statistics of the busiest group */
3472 unsigned long max_load;
3473 unsigned long busiest_load_per_task;
3474 unsigned long busiest_nr_running;
3476 int group_imb; /* Is there imbalance in this sd */
3477 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3478 int power_savings_balance; /* Is powersave balance needed for this sd */
3479 struct sched_group *group_min; /* Least loaded group in sd */
3480 struct sched_group *group_leader; /* Group which relieves group_min */
3481 unsigned long min_load_per_task; /* load_per_task in group_min */
3482 unsigned long leader_nr_running; /* Nr running of group_leader */
3483 unsigned long min_nr_running; /* Nr running of group_min */
3488 * sg_lb_stats - stats of a sched_group required for load_balancing
3490 struct sg_lb_stats {
3491 unsigned long avg_load; /*Avg load across the CPUs of the group */
3492 unsigned long group_load; /* Total load over the CPUs of the group */
3493 unsigned long sum_nr_running; /* Nr tasks running in the group */
3494 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3495 unsigned long group_capacity;
3496 int group_imb; /* Is there an imbalance in the group ? */
3500 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3501 * @group: The group whose first cpu is to be returned.
3503 static inline unsigned int group_first_cpu(struct sched_group *group)
3505 return cpumask_first(sched_group_cpus(group));
3509 * get_sd_load_idx - Obtain the load index for a given sched domain.
3510 * @sd: The sched_domain whose load_idx is to be obtained.
3511 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3513 static inline int get_sd_load_idx(struct sched_domain *sd,
3514 enum cpu_idle_type idle)
3520 load_idx = sd->busy_idx;
3523 case CPU_NEWLY_IDLE:
3524 load_idx = sd->newidle_idx;
3527 load_idx = sd->idle_idx;
3535 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3537 * init_sd_power_savings_stats - Initialize power savings statistics for
3538 * the given sched_domain, during load balancing.
3540 * @sd: Sched domain whose power-savings statistics are to be initialized.
3541 * @sds: Variable containing the statistics for sd.
3542 * @idle: Idle status of the CPU at which we're performing load-balancing.
3544 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3545 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3548 * Busy processors will not participate in power savings
3551 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3552 sds->power_savings_balance = 0;
3554 sds->power_savings_balance = 1;
3555 sds->min_nr_running = ULONG_MAX;
3556 sds->leader_nr_running = 0;
3561 * update_sd_power_savings_stats - Update the power saving stats for a
3562 * sched_domain while performing load balancing.
3564 * @group: sched_group belonging to the sched_domain under consideration.
3565 * @sds: Variable containing the statistics of the sched_domain
3566 * @local_group: Does group contain the CPU for which we're performing
3568 * @sgs: Variable containing the statistics of the group.
3570 static inline void update_sd_power_savings_stats(struct sched_group *group,
3571 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3574 if (!sds->power_savings_balance)
3578 * If the local group is idle or completely loaded
3579 * no need to do power savings balance at this domain
3581 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3582 !sds->this_nr_running))
3583 sds->power_savings_balance = 0;
3586 * If a group is already running at full capacity or idle,
3587 * don't include that group in power savings calculations
3589 if (!sds->power_savings_balance ||
3590 sgs->sum_nr_running >= sgs->group_capacity ||
3591 !sgs->sum_nr_running)
3595 * Calculate the group which has the least non-idle load.
3596 * This is the group from where we need to pick up the load
3599 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3600 (sgs->sum_nr_running == sds->min_nr_running &&
3601 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3602 sds->group_min = group;
3603 sds->min_nr_running = sgs->sum_nr_running;
3604 sds->min_load_per_task = sgs->sum_weighted_load /
3605 sgs->sum_nr_running;
3609 * Calculate the group which is almost near its
3610 * capacity but still has some space to pick up some load
3611 * from other group and save more power
3613 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3616 if (sgs->sum_nr_running > sds->leader_nr_running ||
3617 (sgs->sum_nr_running == sds->leader_nr_running &&
3618 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3619 sds->group_leader = group;
3620 sds->leader_nr_running = sgs->sum_nr_running;
3625 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3626 * @sds: Variable containing the statistics of the sched_domain
3627 * under consideration.
3628 * @this_cpu: Cpu at which we're currently performing load-balancing.
3629 * @imbalance: Variable to store the imbalance.
3632 * Check if we have potential to perform some power-savings balance.
3633 * If yes, set the busiest group to be the least loaded group in the
3634 * sched_domain, so that it's CPUs can be put to idle.
3636 * Returns 1 if there is potential to perform power-savings balance.
3639 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3640 int this_cpu, unsigned long *imbalance)
3642 if (!sds->power_savings_balance)
3645 if (sds->this != sds->group_leader ||
3646 sds->group_leader == sds->group_min)
3649 *imbalance = sds->min_load_per_task;
3650 sds->busiest = sds->group_min;
3655 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3656 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3657 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3662 static inline void update_sd_power_savings_stats(struct sched_group *group,
3663 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3668 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3669 int this_cpu, unsigned long *imbalance)
3673 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3676 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3678 return SCHED_LOAD_SCALE;
3681 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3683 return default_scale_freq_power(sd, cpu);
3686 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3688 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3689 unsigned long smt_gain = sd->smt_gain;
3696 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3698 return default_scale_smt_power(sd, cpu);
3701 unsigned long scale_rt_power(int cpu)
3703 struct rq *rq = cpu_rq(cpu);
3704 u64 total, available;
3706 sched_avg_update(rq);
3708 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3709 available = total - rq->rt_avg;
3711 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3712 total = SCHED_LOAD_SCALE;
3714 total >>= SCHED_LOAD_SHIFT;
3716 return div_u64(available, total);
3719 static void update_cpu_power(struct sched_domain *sd, int cpu)
3721 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3722 unsigned long power = SCHED_LOAD_SCALE;
3723 struct sched_group *sdg = sd->groups;
3725 if (sched_feat(ARCH_POWER))
3726 power *= arch_scale_freq_power(sd, cpu);
3728 power *= default_scale_freq_power(sd, cpu);
3730 power >>= SCHED_LOAD_SHIFT;
3732 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3733 if (sched_feat(ARCH_POWER))
3734 power *= arch_scale_smt_power(sd, cpu);
3736 power *= default_scale_smt_power(sd, cpu);
3738 power >>= SCHED_LOAD_SHIFT;
3741 power *= scale_rt_power(cpu);
3742 power >>= SCHED_LOAD_SHIFT;
3747 sdg->cpu_power = power;
3750 static void update_group_power(struct sched_domain *sd, int cpu)
3752 struct sched_domain *child = sd->child;
3753 struct sched_group *group, *sdg = sd->groups;
3754 unsigned long power;
3757 update_cpu_power(sd, cpu);
3763 group = child->groups;
3765 power += group->cpu_power;
3766 group = group->next;
3767 } while (group != child->groups);
3769 sdg->cpu_power = power;
3773 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3774 * @sd: The sched_domain whose statistics are to be updated.
3775 * @group: sched_group whose statistics are to be updated.
3776 * @this_cpu: Cpu for which load balance is currently performed.
3777 * @idle: Idle status of this_cpu
3778 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3779 * @sd_idle: Idle status of the sched_domain containing group.
3780 * @local_group: Does group contain this_cpu.
3781 * @cpus: Set of cpus considered for load balancing.
3782 * @balance: Should we balance.
3783 * @sgs: variable to hold the statistics for this group.
3785 static inline void update_sg_lb_stats(struct sched_domain *sd,
3786 struct sched_group *group, int this_cpu,
3787 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3788 int local_group, const struct cpumask *cpus,
3789 int *balance, struct sg_lb_stats *sgs)
3791 unsigned long load, max_cpu_load, min_cpu_load;
3793 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3794 unsigned long sum_avg_load_per_task;
3795 unsigned long avg_load_per_task;
3798 balance_cpu = group_first_cpu(group);
3799 if (balance_cpu == this_cpu)
3800 update_group_power(sd, this_cpu);
3803 /* Tally up the load of all CPUs in the group */
3804 sum_avg_load_per_task = avg_load_per_task = 0;
3806 min_cpu_load = ~0UL;
3808 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3809 struct rq *rq = cpu_rq(i);
3811 if (*sd_idle && rq->nr_running)
3814 /* Bias balancing toward cpus of our domain */
3816 if (idle_cpu(i) && !first_idle_cpu) {
3821 load = target_load(i, load_idx);
3823 load = source_load(i, load_idx);
3824 if (load > max_cpu_load)
3825 max_cpu_load = load;
3826 if (min_cpu_load > load)
3827 min_cpu_load = load;
3830 sgs->group_load += load;
3831 sgs->sum_nr_running += rq->nr_running;
3832 sgs->sum_weighted_load += weighted_cpuload(i);
3834 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3838 * First idle cpu or the first cpu(busiest) in this sched group
3839 * is eligible for doing load balancing at this and above
3840 * domains. In the newly idle case, we will allow all the cpu's
3841 * to do the newly idle load balance.
3843 if (idle != CPU_NEWLY_IDLE && local_group &&
3844 balance_cpu != this_cpu && balance) {
3849 /* Adjust by relative CPU power of the group */
3850 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3854 * Consider the group unbalanced when the imbalance is larger
3855 * than the average weight of two tasks.
3857 * APZ: with cgroup the avg task weight can vary wildly and
3858 * might not be a suitable number - should we keep a
3859 * normalized nr_running number somewhere that negates
3862 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3865 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3868 sgs->group_capacity =
3869 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3873 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3874 * @sd: sched_domain whose statistics are to be updated.
3875 * @this_cpu: Cpu for which load balance is currently performed.
3876 * @idle: Idle status of this_cpu
3877 * @sd_idle: Idle status of the sched_domain containing group.
3878 * @cpus: Set of cpus considered for load balancing.
3879 * @balance: Should we balance.
3880 * @sds: variable to hold the statistics for this sched_domain.
3882 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3883 enum cpu_idle_type idle, int *sd_idle,
3884 const struct cpumask *cpus, int *balance,
3885 struct sd_lb_stats *sds)
3887 struct sched_domain *child = sd->child;
3888 struct sched_group *group = sd->groups;
3889 struct sg_lb_stats sgs;
3890 int load_idx, prefer_sibling = 0;
3892 if (child && child->flags & SD_PREFER_SIBLING)
3895 init_sd_power_savings_stats(sd, sds, idle);
3896 load_idx = get_sd_load_idx(sd, idle);
3901 local_group = cpumask_test_cpu(this_cpu,
3902 sched_group_cpus(group));
3903 memset(&sgs, 0, sizeof(sgs));
3904 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3905 local_group, cpus, balance, &sgs);
3907 if (local_group && balance && !(*balance))
3910 sds->total_load += sgs.group_load;
3911 sds->total_pwr += group->cpu_power;
3914 * In case the child domain prefers tasks go to siblings
3915 * first, lower the group capacity to one so that we'll try
3916 * and move all the excess tasks away.
3919 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3922 sds->this_load = sgs.avg_load;
3924 sds->this_nr_running = sgs.sum_nr_running;
3925 sds->this_load_per_task = sgs.sum_weighted_load;
3926 } else if (sgs.avg_load > sds->max_load &&
3927 (sgs.sum_nr_running > sgs.group_capacity ||
3929 sds->max_load = sgs.avg_load;
3930 sds->busiest = group;
3931 sds->busiest_nr_running = sgs.sum_nr_running;
3932 sds->busiest_load_per_task = sgs.sum_weighted_load;
3933 sds->group_imb = sgs.group_imb;
3936 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3937 group = group->next;
3938 } while (group != sd->groups);
3942 * fix_small_imbalance - Calculate the minor imbalance that exists
3943 * amongst the groups of a sched_domain, during
3945 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3946 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3947 * @imbalance: Variable to store the imbalance.
3949 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3950 int this_cpu, unsigned long *imbalance)
3952 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3953 unsigned int imbn = 2;
3955 if (sds->this_nr_running) {
3956 sds->this_load_per_task /= sds->this_nr_running;
3957 if (sds->busiest_load_per_task >
3958 sds->this_load_per_task)
3961 sds->this_load_per_task =
3962 cpu_avg_load_per_task(this_cpu);
3964 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3965 sds->busiest_load_per_task * imbn) {
3966 *imbalance = sds->busiest_load_per_task;
3971 * OK, we don't have enough imbalance to justify moving tasks,
3972 * however we may be able to increase total CPU power used by
3976 pwr_now += sds->busiest->cpu_power *
3977 min(sds->busiest_load_per_task, sds->max_load);
3978 pwr_now += sds->this->cpu_power *
3979 min(sds->this_load_per_task, sds->this_load);
3980 pwr_now /= SCHED_LOAD_SCALE;
3982 /* Amount of load we'd subtract */
3983 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3984 sds->busiest->cpu_power;
3985 if (sds->max_load > tmp)
3986 pwr_move += sds->busiest->cpu_power *
3987 min(sds->busiest_load_per_task, sds->max_load - tmp);
3989 /* Amount of load we'd add */
3990 if (sds->max_load * sds->busiest->cpu_power <
3991 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3992 tmp = (sds->max_load * sds->busiest->cpu_power) /
3993 sds->this->cpu_power;
3995 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3996 sds->this->cpu_power;
3997 pwr_move += sds->this->cpu_power *
3998 min(sds->this_load_per_task, sds->this_load + tmp);
3999 pwr_move /= SCHED_LOAD_SCALE;
4001 /* Move if we gain throughput */
4002 if (pwr_move > pwr_now)
4003 *imbalance = sds->busiest_load_per_task;
4007 * calculate_imbalance - Calculate the amount of imbalance present within the
4008 * groups of a given sched_domain during load balance.
4009 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4010 * @this_cpu: Cpu for which currently load balance is being performed.
4011 * @imbalance: The variable to store the imbalance.
4013 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
4014 unsigned long *imbalance)
4016 unsigned long max_pull;
4018 * In the presence of smp nice balancing, certain scenarios can have
4019 * max load less than avg load(as we skip the groups at or below
4020 * its cpu_power, while calculating max_load..)
4022 if (sds->max_load < sds->avg_load) {
4024 return fix_small_imbalance(sds, this_cpu, imbalance);
4027 /* Don't want to pull so many tasks that a group would go idle */
4028 max_pull = min(sds->max_load - sds->avg_load,
4029 sds->max_load - sds->busiest_load_per_task);
4031 /* How much load to actually move to equalise the imbalance */
4032 *imbalance = min(max_pull * sds->busiest->cpu_power,
4033 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
4037 * if *imbalance is less than the average load per runnable task
4038 * there is no gaurantee that any tasks will be moved so we'll have
4039 * a think about bumping its value to force at least one task to be
4042 if (*imbalance < sds->busiest_load_per_task)
4043 return fix_small_imbalance(sds, this_cpu, imbalance);
4046 /******* find_busiest_group() helpers end here *********************/
4049 * find_busiest_group - Returns the busiest group within the sched_domain
4050 * if there is an imbalance. If there isn't an imbalance, and
4051 * the user has opted for power-savings, it returns a group whose
4052 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4053 * such a group exists.
4055 * Also calculates the amount of weighted load which should be moved
4056 * to restore balance.
4058 * @sd: The sched_domain whose busiest group is to be returned.
4059 * @this_cpu: The cpu for which load balancing is currently being performed.
4060 * @imbalance: Variable which stores amount of weighted load which should
4061 * be moved to restore balance/put a group to idle.
4062 * @idle: The idle status of this_cpu.
4063 * @sd_idle: The idleness of sd
4064 * @cpus: The set of CPUs under consideration for load-balancing.
4065 * @balance: Pointer to a variable indicating if this_cpu
4066 * is the appropriate cpu to perform load balancing at this_level.
4068 * Returns: - the busiest group if imbalance exists.
4069 * - If no imbalance and user has opted for power-savings balance,
4070 * return the least loaded group whose CPUs can be
4071 * put to idle by rebalancing its tasks onto our group.
4073 static struct sched_group *
4074 find_busiest_group(struct sched_domain *sd, int this_cpu,
4075 unsigned long *imbalance, enum cpu_idle_type idle,
4076 int *sd_idle, const struct cpumask *cpus, int *balance)
4078 struct sd_lb_stats sds;
4080 memset(&sds, 0, sizeof(sds));
4083 * Compute the various statistics relavent for load balancing at
4086 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4089 /* Cases where imbalance does not exist from POV of this_cpu */
4090 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4092 * 2) There is no busy sibling group to pull from.
4093 * 3) This group is the busiest group.
4094 * 4) This group is more busy than the avg busieness at this
4096 * 5) The imbalance is within the specified limit.
4097 * 6) Any rebalance would lead to ping-pong
4099 if (balance && !(*balance))
4102 if (!sds.busiest || sds.busiest_nr_running == 0)
4105 if (sds.this_load >= sds.max_load)
4108 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4110 if (sds.this_load >= sds.avg_load)
4113 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4116 sds.busiest_load_per_task /= sds.busiest_nr_running;
4118 sds.busiest_load_per_task =
4119 min(sds.busiest_load_per_task, sds.avg_load);
4122 * We're trying to get all the cpus to the average_load, so we don't
4123 * want to push ourselves above the average load, nor do we wish to
4124 * reduce the max loaded cpu below the average load, as either of these
4125 * actions would just result in more rebalancing later, and ping-pong
4126 * tasks around. Thus we look for the minimum possible imbalance.
4127 * Negative imbalances (*we* are more loaded than anyone else) will
4128 * be counted as no imbalance for these purposes -- we can't fix that
4129 * by pulling tasks to us. Be careful of negative numbers as they'll
4130 * appear as very large values with unsigned longs.
4132 if (sds.max_load <= sds.busiest_load_per_task)
4135 /* Looks like there is an imbalance. Compute it */
4136 calculate_imbalance(&sds, this_cpu, imbalance);
4141 * There is no obvious imbalance. But check if we can do some balancing
4144 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4152 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4155 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4156 unsigned long imbalance, const struct cpumask *cpus)
4158 struct rq *busiest = NULL, *rq;
4159 unsigned long max_load = 0;
4162 for_each_cpu(i, sched_group_cpus(group)) {
4163 unsigned long power = power_of(i);
4164 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4167 if (!cpumask_test_cpu(i, cpus))
4171 wl = weighted_cpuload(i);
4174 * When comparing with imbalance, use weighted_cpuload()
4175 * which is not scaled with the cpu power.
4177 if (capacity && rq->nr_running == 1 && wl > imbalance)
4181 * For the load comparisons with the other cpu's, consider
4182 * the weighted_cpuload() scaled with the cpu power, so that
4183 * the load can be moved away from the cpu that is potentially
4184 * running at a lower capacity.
4186 wl = (wl * SCHED_LOAD_SCALE) / power;
4188 if (wl > max_load) {
4198 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4199 * so long as it is large enough.
4201 #define MAX_PINNED_INTERVAL 512
4203 /* Working cpumask for load_balance and load_balance_newidle. */
4204 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4207 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4208 * tasks if there is an imbalance.
4210 static int load_balance(int this_cpu, struct rq *this_rq,
4211 struct sched_domain *sd, enum cpu_idle_type idle,
4214 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4215 struct sched_group *group;
4216 unsigned long imbalance;
4218 unsigned long flags;
4219 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4221 cpumask_copy(cpus, cpu_active_mask);
4224 * When power savings policy is enabled for the parent domain, idle
4225 * sibling can pick up load irrespective of busy siblings. In this case,
4226 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4227 * portraying it as CPU_NOT_IDLE.
4229 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4230 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4233 schedstat_inc(sd, lb_count[idle]);
4237 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4244 schedstat_inc(sd, lb_nobusyg[idle]);
4248 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4250 schedstat_inc(sd, lb_nobusyq[idle]);
4254 BUG_ON(busiest == this_rq);
4256 schedstat_add(sd, lb_imbalance[idle], imbalance);
4259 if (busiest->nr_running > 1) {
4261 * Attempt to move tasks. If find_busiest_group has found
4262 * an imbalance but busiest->nr_running <= 1, the group is
4263 * still unbalanced. ld_moved simply stays zero, so it is
4264 * correctly treated as an imbalance.
4266 local_irq_save(flags);
4267 double_rq_lock(this_rq, busiest);
4268 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4269 imbalance, sd, idle, &all_pinned);
4270 double_rq_unlock(this_rq, busiest);
4271 local_irq_restore(flags);
4274 * some other cpu did the load balance for us.
4276 if (ld_moved && this_cpu != smp_processor_id())
4277 resched_cpu(this_cpu);
4279 /* All tasks on this runqueue were pinned by CPU affinity */
4280 if (unlikely(all_pinned)) {
4281 cpumask_clear_cpu(cpu_of(busiest), cpus);
4282 if (!cpumask_empty(cpus))
4289 schedstat_inc(sd, lb_failed[idle]);
4290 sd->nr_balance_failed++;
4292 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4294 raw_spin_lock_irqsave(&busiest->lock, flags);
4296 /* don't kick the migration_thread, if the curr
4297 * task on busiest cpu can't be moved to this_cpu
4299 if (!cpumask_test_cpu(this_cpu,
4300 &busiest->curr->cpus_allowed)) {
4301 raw_spin_unlock_irqrestore(&busiest->lock,
4304 goto out_one_pinned;
4307 if (!busiest->active_balance) {
4308 busiest->active_balance = 1;
4309 busiest->push_cpu = this_cpu;
4312 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4314 wake_up_process(busiest->migration_thread);
4317 * We've kicked active balancing, reset the failure
4320 sd->nr_balance_failed = sd->cache_nice_tries+1;
4323 sd->nr_balance_failed = 0;
4325 if (likely(!active_balance)) {
4326 /* We were unbalanced, so reset the balancing interval */
4327 sd->balance_interval = sd->min_interval;
4330 * If we've begun active balancing, start to back off. This
4331 * case may not be covered by the all_pinned logic if there
4332 * is only 1 task on the busy runqueue (because we don't call
4335 if (sd->balance_interval < sd->max_interval)
4336 sd->balance_interval *= 2;
4339 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4340 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4346 schedstat_inc(sd, lb_balanced[idle]);
4348 sd->nr_balance_failed = 0;
4351 /* tune up the balancing interval */
4352 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4353 (sd->balance_interval < sd->max_interval))
4354 sd->balance_interval *= 2;
4356 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4357 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4368 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4369 * tasks if there is an imbalance.
4371 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4372 * this_rq is locked.
4375 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4377 struct sched_group *group;
4378 struct rq *busiest = NULL;
4379 unsigned long imbalance;
4383 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4385 cpumask_copy(cpus, cpu_active_mask);
4388 * When power savings policy is enabled for the parent domain, idle
4389 * sibling can pick up load irrespective of busy siblings. In this case,
4390 * let the state of idle sibling percolate up as IDLE, instead of
4391 * portraying it as CPU_NOT_IDLE.
4393 if (sd->flags & SD_SHARE_CPUPOWER &&
4394 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4397 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4399 update_shares_locked(this_rq, sd);
4400 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4401 &sd_idle, cpus, NULL);
4403 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4407 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4409 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4413 BUG_ON(busiest == this_rq);
4415 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4418 if (busiest->nr_running > 1) {
4419 /* Attempt to move tasks */
4420 double_lock_balance(this_rq, busiest);
4421 /* this_rq->clock is already updated */
4422 update_rq_clock(busiest);
4423 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4424 imbalance, sd, CPU_NEWLY_IDLE,
4426 double_unlock_balance(this_rq, busiest);
4428 if (unlikely(all_pinned)) {
4429 cpumask_clear_cpu(cpu_of(busiest), cpus);
4430 if (!cpumask_empty(cpus))
4436 int active_balance = 0;
4438 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4439 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4440 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4443 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4446 if (sd->nr_balance_failed++ < 2)
4450 * The only task running in a non-idle cpu can be moved to this
4451 * cpu in an attempt to completely freeup the other CPU
4452 * package. The same method used to move task in load_balance()
4453 * have been extended for load_balance_newidle() to speedup
4454 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4456 * The package power saving logic comes from
4457 * find_busiest_group(). If there are no imbalance, then
4458 * f_b_g() will return NULL. However when sched_mc={1,2} then
4459 * f_b_g() will select a group from which a running task may be
4460 * pulled to this cpu in order to make the other package idle.
4461 * If there is no opportunity to make a package idle and if
4462 * there are no imbalance, then f_b_g() will return NULL and no
4463 * action will be taken in load_balance_newidle().
4465 * Under normal task pull operation due to imbalance, there
4466 * will be more than one task in the source run queue and
4467 * move_tasks() will succeed. ld_moved will be true and this
4468 * active balance code will not be triggered.
4471 /* Lock busiest in correct order while this_rq is held */
4472 double_lock_balance(this_rq, busiest);
4475 * don't kick the migration_thread, if the curr
4476 * task on busiest cpu can't be moved to this_cpu
4478 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4479 double_unlock_balance(this_rq, busiest);
4484 if (!busiest->active_balance) {
4485 busiest->active_balance = 1;
4486 busiest->push_cpu = this_cpu;
4490 double_unlock_balance(this_rq, busiest);
4492 * Should not call ttwu while holding a rq->lock
4494 raw_spin_unlock(&this_rq->lock);
4496 wake_up_process(busiest->migration_thread);
4497 raw_spin_lock(&this_rq->lock);
4500 sd->nr_balance_failed = 0;
4502 update_shares_locked(this_rq, sd);
4506 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4507 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4508 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4510 sd->nr_balance_failed = 0;
4516 * idle_balance is called by schedule() if this_cpu is about to become
4517 * idle. Attempts to pull tasks from other CPUs.
4519 static void idle_balance(int this_cpu, struct rq *this_rq)
4521 struct sched_domain *sd;
4522 int pulled_task = 0;
4523 unsigned long next_balance = jiffies + HZ;
4525 this_rq->idle_stamp = this_rq->clock;
4527 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4530 for_each_domain(this_cpu, sd) {
4531 unsigned long interval;
4533 if (!(sd->flags & SD_LOAD_BALANCE))
4536 if (sd->flags & SD_BALANCE_NEWIDLE)
4537 /* If we've pulled tasks over stop searching: */
4538 pulled_task = load_balance_newidle(this_cpu, this_rq,
4541 interval = msecs_to_jiffies(sd->balance_interval);
4542 if (time_after(next_balance, sd->last_balance + interval))
4543 next_balance = sd->last_balance + interval;
4545 this_rq->idle_stamp = 0;
4549 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4551 * We are going idle. next_balance may be set based on
4552 * a busy processor. So reset next_balance.
4554 this_rq->next_balance = next_balance;
4559 * active_load_balance is run by migration threads. It pushes running tasks
4560 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4561 * running on each physical CPU where possible, and avoids physical /
4562 * logical imbalances.
4564 * Called with busiest_rq locked.
4566 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4568 int target_cpu = busiest_rq->push_cpu;
4569 struct sched_domain *sd;
4570 struct rq *target_rq;
4572 /* Is there any task to move? */
4573 if (busiest_rq->nr_running <= 1)
4576 target_rq = cpu_rq(target_cpu);
4579 * This condition is "impossible", if it occurs
4580 * we need to fix it. Originally reported by
4581 * Bjorn Helgaas on a 128-cpu setup.
4583 BUG_ON(busiest_rq == target_rq);
4585 /* move a task from busiest_rq to target_rq */
4586 double_lock_balance(busiest_rq, target_rq);
4587 update_rq_clock(busiest_rq);
4588 update_rq_clock(target_rq);
4590 /* Search for an sd spanning us and the target CPU. */
4591 for_each_domain(target_cpu, sd) {
4592 if ((sd->flags & SD_LOAD_BALANCE) &&
4593 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4598 schedstat_inc(sd, alb_count);
4600 if (move_one_task(target_rq, target_cpu, busiest_rq,
4602 schedstat_inc(sd, alb_pushed);
4604 schedstat_inc(sd, alb_failed);
4606 double_unlock_balance(busiest_rq, target_rq);
4611 atomic_t load_balancer;
4612 cpumask_var_t cpu_mask;
4613 cpumask_var_t ilb_grp_nohz_mask;
4614 } nohz ____cacheline_aligned = {
4615 .load_balancer = ATOMIC_INIT(-1),
4618 int get_nohz_load_balancer(void)
4620 return atomic_read(&nohz.load_balancer);
4623 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4625 * lowest_flag_domain - Return lowest sched_domain containing flag.
4626 * @cpu: The cpu whose lowest level of sched domain is to
4628 * @flag: The flag to check for the lowest sched_domain
4629 * for the given cpu.
4631 * Returns the lowest sched_domain of a cpu which contains the given flag.
4633 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4635 struct sched_domain *sd;
4637 for_each_domain(cpu, sd)
4638 if (sd && (sd->flags & flag))
4645 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4646 * @cpu: The cpu whose domains we're iterating over.
4647 * @sd: variable holding the value of the power_savings_sd
4649 * @flag: The flag to filter the sched_domains to be iterated.
4651 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4652 * set, starting from the lowest sched_domain to the highest.
4654 #define for_each_flag_domain(cpu, sd, flag) \
4655 for (sd = lowest_flag_domain(cpu, flag); \
4656 (sd && (sd->flags & flag)); sd = sd->parent)
4659 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4660 * @ilb_group: group to be checked for semi-idleness
4662 * Returns: 1 if the group is semi-idle. 0 otherwise.
4664 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4665 * and atleast one non-idle CPU. This helper function checks if the given
4666 * sched_group is semi-idle or not.
4668 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4670 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4671 sched_group_cpus(ilb_group));
4674 * A sched_group is semi-idle when it has atleast one busy cpu
4675 * and atleast one idle cpu.
4677 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4680 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4686 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4687 * @cpu: The cpu which is nominating a new idle_load_balancer.
4689 * Returns: Returns the id of the idle load balancer if it exists,
4690 * Else, returns >= nr_cpu_ids.
4692 * This algorithm picks the idle load balancer such that it belongs to a
4693 * semi-idle powersavings sched_domain. The idea is to try and avoid
4694 * completely idle packages/cores just for the purpose of idle load balancing
4695 * when there are other idle cpu's which are better suited for that job.
4697 static int find_new_ilb(int cpu)
4699 struct sched_domain *sd;
4700 struct sched_group *ilb_group;
4703 * Have idle load balancer selection from semi-idle packages only
4704 * when power-aware load balancing is enabled
4706 if (!(sched_smt_power_savings || sched_mc_power_savings))
4710 * Optimize for the case when we have no idle CPUs or only one
4711 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4713 if (cpumask_weight(nohz.cpu_mask) < 2)
4716 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4717 ilb_group = sd->groups;
4720 if (is_semi_idle_group(ilb_group))
4721 return cpumask_first(nohz.ilb_grp_nohz_mask);
4723 ilb_group = ilb_group->next;
4725 } while (ilb_group != sd->groups);
4729 return cpumask_first(nohz.cpu_mask);
4731 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4732 static inline int find_new_ilb(int call_cpu)
4734 return cpumask_first(nohz.cpu_mask);
4739 * This routine will try to nominate the ilb (idle load balancing)
4740 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4741 * load balancing on behalf of all those cpus. If all the cpus in the system
4742 * go into this tickless mode, then there will be no ilb owner (as there is
4743 * no need for one) and all the cpus will sleep till the next wakeup event
4746 * For the ilb owner, tick is not stopped. And this tick will be used
4747 * for idle load balancing. ilb owner will still be part of
4750 * While stopping the tick, this cpu will become the ilb owner if there
4751 * is no other owner. And will be the owner till that cpu becomes busy
4752 * or if all cpus in the system stop their ticks at which point
4753 * there is no need for ilb owner.
4755 * When the ilb owner becomes busy, it nominates another owner, during the
4756 * next busy scheduler_tick()
4758 int select_nohz_load_balancer(int stop_tick)
4760 int cpu = smp_processor_id();
4763 cpu_rq(cpu)->in_nohz_recently = 1;
4765 if (!cpu_active(cpu)) {
4766 if (atomic_read(&nohz.load_balancer) != cpu)
4770 * If we are going offline and still the leader,
4773 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4779 cpumask_set_cpu(cpu, nohz.cpu_mask);
4781 /* time for ilb owner also to sleep */
4782 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4783 if (atomic_read(&nohz.load_balancer) == cpu)
4784 atomic_set(&nohz.load_balancer, -1);
4788 if (atomic_read(&nohz.load_balancer) == -1) {
4789 /* make me the ilb owner */
4790 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4792 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4795 if (!(sched_smt_power_savings ||
4796 sched_mc_power_savings))
4799 * Check to see if there is a more power-efficient
4802 new_ilb = find_new_ilb(cpu);
4803 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4804 atomic_set(&nohz.load_balancer, -1);
4805 resched_cpu(new_ilb);
4811 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4814 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4816 if (atomic_read(&nohz.load_balancer) == cpu)
4817 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4824 static DEFINE_SPINLOCK(balancing);
4827 * It checks each scheduling domain to see if it is due to be balanced,
4828 * and initiates a balancing operation if so.
4830 * Balancing parameters are set up in arch_init_sched_domains.
4832 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4835 struct rq *rq = cpu_rq(cpu);
4836 unsigned long interval;
4837 struct sched_domain *sd;
4838 /* Earliest time when we have to do rebalance again */
4839 unsigned long next_balance = jiffies + 60*HZ;
4840 int update_next_balance = 0;
4843 for_each_domain(cpu, sd) {
4844 if (!(sd->flags & SD_LOAD_BALANCE))
4847 interval = sd->balance_interval;
4848 if (idle != CPU_IDLE)
4849 interval *= sd->busy_factor;
4851 /* scale ms to jiffies */
4852 interval = msecs_to_jiffies(interval);
4853 if (unlikely(!interval))
4855 if (interval > HZ*NR_CPUS/10)
4856 interval = HZ*NR_CPUS/10;
4858 need_serialize = sd->flags & SD_SERIALIZE;
4860 if (need_serialize) {
4861 if (!spin_trylock(&balancing))
4865 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4866 if (load_balance(cpu, rq, sd, idle, &balance)) {
4868 * We've pulled tasks over so either we're no
4869 * longer idle, or one of our SMT siblings is
4872 idle = CPU_NOT_IDLE;
4874 sd->last_balance = jiffies;
4877 spin_unlock(&balancing);
4879 if (time_after(next_balance, sd->last_balance + interval)) {
4880 next_balance = sd->last_balance + interval;
4881 update_next_balance = 1;
4885 * Stop the load balance at this level. There is another
4886 * CPU in our sched group which is doing load balancing more
4894 * next_balance will be updated only when there is a need.
4895 * When the cpu is attached to null domain for ex, it will not be
4898 if (likely(update_next_balance))
4899 rq->next_balance = next_balance;
4903 * run_rebalance_domains is triggered when needed from the scheduler tick.
4904 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4905 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4907 static void run_rebalance_domains(struct softirq_action *h)
4909 int this_cpu = smp_processor_id();
4910 struct rq *this_rq = cpu_rq(this_cpu);
4911 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4912 CPU_IDLE : CPU_NOT_IDLE;
4914 rebalance_domains(this_cpu, idle);
4918 * If this cpu is the owner for idle load balancing, then do the
4919 * balancing on behalf of the other idle cpus whose ticks are
4922 if (this_rq->idle_at_tick &&
4923 atomic_read(&nohz.load_balancer) == this_cpu) {
4927 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4928 if (balance_cpu == this_cpu)
4932 * If this cpu gets work to do, stop the load balancing
4933 * work being done for other cpus. Next load
4934 * balancing owner will pick it up.
4939 rebalance_domains(balance_cpu, CPU_IDLE);
4941 rq = cpu_rq(balance_cpu);
4942 if (time_after(this_rq->next_balance, rq->next_balance))
4943 this_rq->next_balance = rq->next_balance;
4949 static inline int on_null_domain(int cpu)
4951 return !rcu_dereference_sched(cpu_rq(cpu)->sd);
4955 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4957 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4958 * idle load balancing owner or decide to stop the periodic load balancing,
4959 * if the whole system is idle.
4961 static inline void trigger_load_balance(struct rq *rq, int cpu)
4965 * If we were in the nohz mode recently and busy at the current
4966 * scheduler tick, then check if we need to nominate new idle
4969 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4970 rq->in_nohz_recently = 0;
4972 if (atomic_read(&nohz.load_balancer) == cpu) {
4973 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4974 atomic_set(&nohz.load_balancer, -1);
4977 if (atomic_read(&nohz.load_balancer) == -1) {
4978 int ilb = find_new_ilb(cpu);
4980 if (ilb < nr_cpu_ids)
4986 * If this cpu is idle and doing idle load balancing for all the
4987 * cpus with ticks stopped, is it time for that to stop?
4989 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4990 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4996 * If this cpu is idle and the idle load balancing is done by
4997 * someone else, then no need raise the SCHED_SOFTIRQ
4999 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
5000 cpumask_test_cpu(cpu, nohz.cpu_mask))
5003 /* Don't need to rebalance while attached to NULL domain */
5004 if (time_after_eq(jiffies, rq->next_balance) &&
5005 likely(!on_null_domain(cpu)))
5006 raise_softirq(SCHED_SOFTIRQ);
5009 #else /* CONFIG_SMP */
5012 * on UP we do not need to balance between CPUs:
5014 static inline void idle_balance(int cpu, struct rq *rq)
5020 DEFINE_PER_CPU(struct kernel_stat, kstat);
5022 EXPORT_PER_CPU_SYMBOL(kstat);
5025 * Return any ns on the sched_clock that have not yet been accounted in
5026 * @p in case that task is currently running.
5028 * Called with task_rq_lock() held on @rq.
5030 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
5034 if (task_current(rq, p)) {
5035 update_rq_clock(rq);
5036 ns = rq->clock - p->se.exec_start;
5044 unsigned long long task_delta_exec(struct task_struct *p)
5046 unsigned long flags;
5050 rq = task_rq_lock(p, &flags);
5051 ns = do_task_delta_exec(p, rq);
5052 task_rq_unlock(rq, &flags);
5058 * Return accounted runtime for the task.
5059 * In case the task is currently running, return the runtime plus current's
5060 * pending runtime that have not been accounted yet.
5062 unsigned long long task_sched_runtime(struct task_struct *p)
5064 unsigned long flags;
5068 rq = task_rq_lock(p, &flags);
5069 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5070 task_rq_unlock(rq, &flags);
5076 * Return sum_exec_runtime for the thread group.
5077 * In case the task is currently running, return the sum plus current's
5078 * pending runtime that have not been accounted yet.
5080 * Note that the thread group might have other running tasks as well,
5081 * so the return value not includes other pending runtime that other
5082 * running tasks might have.
5084 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5086 struct task_cputime totals;
5087 unsigned long flags;
5091 rq = task_rq_lock(p, &flags);
5092 thread_group_cputime(p, &totals);
5093 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5094 task_rq_unlock(rq, &flags);
5100 * Account user cpu time to a process.
5101 * @p: the process that the cpu time gets accounted to
5102 * @cputime: the cpu time spent in user space since the last update
5103 * @cputime_scaled: cputime scaled by cpu frequency
5105 void account_user_time(struct task_struct *p, cputime_t cputime,
5106 cputime_t cputime_scaled)
5108 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5111 /* Add user time to process. */
5112 p->utime = cputime_add(p->utime, cputime);
5113 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5114 account_group_user_time(p, cputime);
5116 /* Add user time to cpustat. */
5117 tmp = cputime_to_cputime64(cputime);
5118 if (TASK_NICE(p) > 0)
5119 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5121 cpustat->user = cputime64_add(cpustat->user, tmp);
5123 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5124 /* Account for user time used */
5125 acct_update_integrals(p);
5129 * Account guest cpu time to a process.
5130 * @p: the process that the cpu time gets accounted to
5131 * @cputime: the cpu time spent in virtual machine since the last update
5132 * @cputime_scaled: cputime scaled by cpu frequency
5134 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5135 cputime_t cputime_scaled)
5138 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5140 tmp = cputime_to_cputime64(cputime);
5142 /* Add guest 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);
5146 p->gtime = cputime_add(p->gtime, cputime);
5148 /* Add guest time to cpustat. */
5149 if (TASK_NICE(p) > 0) {
5150 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5151 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5153 cpustat->user = cputime64_add(cpustat->user, tmp);
5154 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5159 * Account system cpu time to a process.
5160 * @p: the process that the cpu time gets accounted to
5161 * @hardirq_offset: the offset to subtract from hardirq_count()
5162 * @cputime: the cpu time spent in kernel space since the last update
5163 * @cputime_scaled: cputime scaled by cpu frequency
5165 void account_system_time(struct task_struct *p, int hardirq_offset,
5166 cputime_t cputime, cputime_t cputime_scaled)
5168 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5171 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5172 account_guest_time(p, cputime, cputime_scaled);
5176 /* Add system time to process. */
5177 p->stime = cputime_add(p->stime, cputime);
5178 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5179 account_group_system_time(p, cputime);
5181 /* Add system time to cpustat. */
5182 tmp = cputime_to_cputime64(cputime);
5183 if (hardirq_count() - hardirq_offset)
5184 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5185 else if (softirq_count())
5186 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5188 cpustat->system = cputime64_add(cpustat->system, tmp);
5190 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5192 /* Account for system time used */
5193 acct_update_integrals(p);
5197 * Account for involuntary wait time.
5198 * @steal: the cpu time spent in involuntary wait
5200 void account_steal_time(cputime_t cputime)
5202 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5203 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5205 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5209 * Account for idle time.
5210 * @cputime: the cpu time spent in idle wait
5212 void account_idle_time(cputime_t cputime)
5214 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5215 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5216 struct rq *rq = this_rq();
5218 if (atomic_read(&rq->nr_iowait) > 0)
5219 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5221 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5224 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5227 * Account a single tick of cpu time.
5228 * @p: the process that the cpu time gets accounted to
5229 * @user_tick: indicates if the tick is a user or a system tick
5231 void account_process_tick(struct task_struct *p, int user_tick)
5233 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5234 struct rq *rq = this_rq();
5237 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5238 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5239 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5242 account_idle_time(cputime_one_jiffy);
5246 * Account multiple ticks of steal time.
5247 * @p: the process from which the cpu time has been stolen
5248 * @ticks: number of stolen ticks
5250 void account_steal_ticks(unsigned long ticks)
5252 account_steal_time(jiffies_to_cputime(ticks));
5256 * Account multiple ticks of idle time.
5257 * @ticks: number of stolen ticks
5259 void account_idle_ticks(unsigned long ticks)
5261 account_idle_time(jiffies_to_cputime(ticks));
5267 * Use precise platform statistics if available:
5269 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5270 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5276 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5278 struct task_cputime cputime;
5280 thread_group_cputime(p, &cputime);
5282 *ut = cputime.utime;
5283 *st = cputime.stime;
5287 #ifndef nsecs_to_cputime
5288 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5291 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5293 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5296 * Use CFS's precise accounting:
5298 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5303 temp = (u64)(rtime * utime);
5304 do_div(temp, total);
5305 utime = (cputime_t)temp;
5310 * Compare with previous values, to keep monotonicity:
5312 p->prev_utime = max(p->prev_utime, utime);
5313 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5315 *ut = p->prev_utime;
5316 *st = p->prev_stime;
5320 * Must be called with siglock held.
5322 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5324 struct signal_struct *sig = p->signal;
5325 struct task_cputime cputime;
5326 cputime_t rtime, utime, total;
5328 thread_group_cputime(p, &cputime);
5330 total = cputime_add(cputime.utime, cputime.stime);
5331 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5336 temp = (u64)(rtime * cputime.utime);
5337 do_div(temp, total);
5338 utime = (cputime_t)temp;
5342 sig->prev_utime = max(sig->prev_utime, utime);
5343 sig->prev_stime = max(sig->prev_stime,
5344 cputime_sub(rtime, sig->prev_utime));
5346 *ut = sig->prev_utime;
5347 *st = sig->prev_stime;
5352 * This function gets called by the timer code, with HZ frequency.
5353 * We call it with interrupts disabled.
5355 * It also gets called by the fork code, when changing the parent's
5358 void scheduler_tick(void)
5360 int cpu = smp_processor_id();
5361 struct rq *rq = cpu_rq(cpu);
5362 struct task_struct *curr = rq->curr;
5366 raw_spin_lock(&rq->lock);
5367 update_rq_clock(rq);
5368 update_cpu_load(rq);
5369 curr->sched_class->task_tick(rq, curr, 0);
5370 raw_spin_unlock(&rq->lock);
5372 perf_event_task_tick(curr);
5375 rq->idle_at_tick = idle_cpu(cpu);
5376 trigger_load_balance(rq, cpu);
5380 notrace unsigned long get_parent_ip(unsigned long addr)
5382 if (in_lock_functions(addr)) {
5383 addr = CALLER_ADDR2;
5384 if (in_lock_functions(addr))
5385 addr = CALLER_ADDR3;
5390 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5391 defined(CONFIG_PREEMPT_TRACER))
5393 void __kprobes add_preempt_count(int val)
5395 #ifdef CONFIG_DEBUG_PREEMPT
5399 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5402 preempt_count() += val;
5403 #ifdef CONFIG_DEBUG_PREEMPT
5405 * Spinlock count overflowing soon?
5407 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5410 if (preempt_count() == val)
5411 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5413 EXPORT_SYMBOL(add_preempt_count);
5415 void __kprobes sub_preempt_count(int val)
5417 #ifdef CONFIG_DEBUG_PREEMPT
5421 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5424 * Is the spinlock portion underflowing?
5426 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5427 !(preempt_count() & PREEMPT_MASK)))
5431 if (preempt_count() == val)
5432 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5433 preempt_count() -= val;
5435 EXPORT_SYMBOL(sub_preempt_count);
5440 * Print scheduling while atomic bug:
5442 static noinline void __schedule_bug(struct task_struct *prev)
5444 struct pt_regs *regs = get_irq_regs();
5446 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5447 prev->comm, prev->pid, preempt_count());
5449 debug_show_held_locks(prev);
5451 if (irqs_disabled())
5452 print_irqtrace_events(prev);
5461 * Various schedule()-time debugging checks and statistics:
5463 static inline void schedule_debug(struct task_struct *prev)
5466 * Test if we are atomic. Since do_exit() needs to call into
5467 * schedule() atomically, we ignore that path for now.
5468 * Otherwise, whine if we are scheduling when we should not be.
5470 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5471 __schedule_bug(prev);
5473 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5475 schedstat_inc(this_rq(), sched_count);
5476 #ifdef CONFIG_SCHEDSTATS
5477 if (unlikely(prev->lock_depth >= 0)) {
5478 schedstat_inc(this_rq(), bkl_count);
5479 schedstat_inc(prev, sched_info.bkl_count);
5484 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5486 if (prev->state == TASK_RUNNING) {
5487 u64 runtime = prev->se.sum_exec_runtime;
5489 runtime -= prev->se.prev_sum_exec_runtime;
5490 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5493 * In order to avoid avg_overlap growing stale when we are
5494 * indeed overlapping and hence not getting put to sleep, grow
5495 * the avg_overlap on preemption.
5497 * We use the average preemption runtime because that
5498 * correlates to the amount of cache footprint a task can
5501 update_avg(&prev->se.avg_overlap, runtime);
5503 prev->sched_class->put_prev_task(rq, prev);
5507 * Pick up the highest-prio task:
5509 static inline struct task_struct *
5510 pick_next_task(struct rq *rq)
5512 const struct sched_class *class;
5513 struct task_struct *p;
5516 * Optimization: we know that if all tasks are in
5517 * the fair class we can call that function directly:
5519 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5520 p = fair_sched_class.pick_next_task(rq);
5525 class = sched_class_highest;
5527 p = class->pick_next_task(rq);
5531 * Will never be NULL as the idle class always
5532 * returns a non-NULL p:
5534 class = class->next;
5539 * schedule() is the main scheduler function.
5541 asmlinkage void __sched schedule(void)
5543 struct task_struct *prev, *next;
5544 unsigned long *switch_count;
5550 cpu = smp_processor_id();
5554 switch_count = &prev->nivcsw;
5556 release_kernel_lock(prev);
5557 need_resched_nonpreemptible:
5559 schedule_debug(prev);
5561 if (sched_feat(HRTICK))
5564 raw_spin_lock_irq(&rq->lock);
5565 update_rq_clock(rq);
5566 clear_tsk_need_resched(prev);
5568 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5569 if (unlikely(signal_pending_state(prev->state, prev)))
5570 prev->state = TASK_RUNNING;
5572 deactivate_task(rq, prev, 1);
5573 switch_count = &prev->nvcsw;
5576 pre_schedule(rq, prev);
5578 if (unlikely(!rq->nr_running))
5579 idle_balance(cpu, rq);
5581 put_prev_task(rq, prev);
5582 next = pick_next_task(rq);
5584 if (likely(prev != next)) {
5585 sched_info_switch(prev, next);
5586 perf_event_task_sched_out(prev, next);
5592 context_switch(rq, prev, next); /* unlocks the rq */
5594 * the context switch might have flipped the stack from under
5595 * us, hence refresh the local variables.
5597 cpu = smp_processor_id();
5600 raw_spin_unlock_irq(&rq->lock);
5604 if (unlikely(reacquire_kernel_lock(current) < 0)) {
5606 switch_count = &prev->nivcsw;
5607 goto need_resched_nonpreemptible;
5610 preempt_enable_no_resched();
5614 EXPORT_SYMBOL(schedule);
5616 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5618 * Look out! "owner" is an entirely speculative pointer
5619 * access and not reliable.
5621 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5626 if (!sched_feat(OWNER_SPIN))
5629 #ifdef CONFIG_DEBUG_PAGEALLOC
5631 * Need to access the cpu field knowing that
5632 * DEBUG_PAGEALLOC could have unmapped it if
5633 * the mutex owner just released it and exited.
5635 if (probe_kernel_address(&owner->cpu, cpu))
5642 * Even if the access succeeded (likely case),
5643 * the cpu field may no longer be valid.
5645 if (cpu >= nr_cpumask_bits)
5649 * We need to validate that we can do a
5650 * get_cpu() and that we have the percpu area.
5652 if (!cpu_online(cpu))
5659 * Owner changed, break to re-assess state.
5661 if (lock->owner != owner)
5665 * Is that owner really running on that cpu?
5667 if (task_thread_info(rq->curr) != owner || need_resched())
5677 #ifdef CONFIG_PREEMPT
5679 * this is the entry point to schedule() from in-kernel preemption
5680 * off of preempt_enable. Kernel preemptions off return from interrupt
5681 * occur there and call schedule directly.
5683 asmlinkage void __sched preempt_schedule(void)
5685 struct thread_info *ti = current_thread_info();
5688 * If there is a non-zero preempt_count or interrupts are disabled,
5689 * we do not want to preempt the current task. Just return..
5691 if (likely(ti->preempt_count || irqs_disabled()))
5695 add_preempt_count(PREEMPT_ACTIVE);
5697 sub_preempt_count(PREEMPT_ACTIVE);
5700 * Check again in case we missed a preemption opportunity
5701 * between schedule and now.
5704 } while (need_resched());
5706 EXPORT_SYMBOL(preempt_schedule);
5709 * this is the entry point to schedule() from kernel preemption
5710 * off of irq context.
5711 * Note, that this is called and return with irqs disabled. This will
5712 * protect us against recursive calling from irq.
5714 asmlinkage void __sched preempt_schedule_irq(void)
5716 struct thread_info *ti = current_thread_info();
5718 /* Catch callers which need to be fixed */
5719 BUG_ON(ti->preempt_count || !irqs_disabled());
5722 add_preempt_count(PREEMPT_ACTIVE);
5725 local_irq_disable();
5726 sub_preempt_count(PREEMPT_ACTIVE);
5729 * Check again in case we missed a preemption opportunity
5730 * between schedule and now.
5733 } while (need_resched());
5736 #endif /* CONFIG_PREEMPT */
5738 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5741 return try_to_wake_up(curr->private, mode, wake_flags);
5743 EXPORT_SYMBOL(default_wake_function);
5746 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5747 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5748 * number) then we wake all the non-exclusive tasks and one exclusive task.
5750 * There are circumstances in which we can try to wake a task which has already
5751 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5752 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5754 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5755 int nr_exclusive, int wake_flags, void *key)
5757 wait_queue_t *curr, *next;
5759 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5760 unsigned flags = curr->flags;
5762 if (curr->func(curr, mode, wake_flags, key) &&
5763 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5769 * __wake_up - wake up threads blocked on a waitqueue.
5771 * @mode: which threads
5772 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5773 * @key: is directly passed to the wakeup function
5775 * It may be assumed that this function implies a write memory barrier before
5776 * changing the task state if and only if any tasks are woken up.
5778 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5779 int nr_exclusive, void *key)
5781 unsigned long flags;
5783 spin_lock_irqsave(&q->lock, flags);
5784 __wake_up_common(q, mode, nr_exclusive, 0, key);
5785 spin_unlock_irqrestore(&q->lock, flags);
5787 EXPORT_SYMBOL(__wake_up);
5790 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5792 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5794 __wake_up_common(q, mode, 1, 0, NULL);
5797 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5799 __wake_up_common(q, mode, 1, 0, key);
5803 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5805 * @mode: which threads
5806 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5807 * @key: opaque value to be passed to wakeup targets
5809 * The sync wakeup differs that the waker knows that it will schedule
5810 * away soon, so while the target thread will be woken up, it will not
5811 * be migrated to another CPU - ie. the two threads are 'synchronized'
5812 * with each other. This can prevent needless bouncing between CPUs.
5814 * On UP it can prevent extra preemption.
5816 * It may be assumed that this function implies a write memory barrier before
5817 * changing the task state if and only if any tasks are woken up.
5819 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5820 int nr_exclusive, void *key)
5822 unsigned long flags;
5823 int wake_flags = WF_SYNC;
5828 if (unlikely(!nr_exclusive))
5831 spin_lock_irqsave(&q->lock, flags);
5832 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5833 spin_unlock_irqrestore(&q->lock, flags);
5835 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5838 * __wake_up_sync - see __wake_up_sync_key()
5840 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5842 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5844 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5847 * complete: - signals a single thread waiting on this completion
5848 * @x: holds the state of this particular completion
5850 * This will wake up a single thread waiting on this completion. Threads will be
5851 * awakened in the same order in which they were queued.
5853 * See also complete_all(), wait_for_completion() and related routines.
5855 * It may be assumed that this function implies a write memory barrier before
5856 * changing the task state if and only if any tasks are woken up.
5858 void complete(struct completion *x)
5860 unsigned long flags;
5862 spin_lock_irqsave(&x->wait.lock, flags);
5864 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5865 spin_unlock_irqrestore(&x->wait.lock, flags);
5867 EXPORT_SYMBOL(complete);
5870 * complete_all: - signals all threads waiting on this completion
5871 * @x: holds the state of this particular completion
5873 * This will wake up all threads waiting on this particular completion event.
5875 * It may be assumed that this function implies a write memory barrier before
5876 * changing the task state if and only if any tasks are woken up.
5878 void complete_all(struct completion *x)
5880 unsigned long flags;
5882 spin_lock_irqsave(&x->wait.lock, flags);
5883 x->done += UINT_MAX/2;
5884 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5885 spin_unlock_irqrestore(&x->wait.lock, flags);
5887 EXPORT_SYMBOL(complete_all);
5889 static inline long __sched
5890 do_wait_for_common(struct completion *x, long timeout, int state)
5893 DECLARE_WAITQUEUE(wait, current);
5895 wait.flags |= WQ_FLAG_EXCLUSIVE;
5896 __add_wait_queue_tail(&x->wait, &wait);
5898 if (signal_pending_state(state, current)) {
5899 timeout = -ERESTARTSYS;
5902 __set_current_state(state);
5903 spin_unlock_irq(&x->wait.lock);
5904 timeout = schedule_timeout(timeout);
5905 spin_lock_irq(&x->wait.lock);
5906 } while (!x->done && timeout);
5907 __remove_wait_queue(&x->wait, &wait);
5912 return timeout ?: 1;
5916 wait_for_common(struct completion *x, long timeout, int state)
5920 spin_lock_irq(&x->wait.lock);
5921 timeout = do_wait_for_common(x, timeout, state);
5922 spin_unlock_irq(&x->wait.lock);
5927 * wait_for_completion: - waits for completion of a task
5928 * @x: holds the state of this particular completion
5930 * This waits to be signaled for completion of a specific task. It is NOT
5931 * interruptible and there is no timeout.
5933 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5934 * and interrupt capability. Also see complete().
5936 void __sched wait_for_completion(struct completion *x)
5938 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5940 EXPORT_SYMBOL(wait_for_completion);
5943 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5944 * @x: holds the state of this particular completion
5945 * @timeout: timeout value in jiffies
5947 * This waits for either a completion of a specific task to be signaled or for a
5948 * specified timeout to expire. The timeout is in jiffies. It is not
5951 unsigned long __sched
5952 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5954 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5956 EXPORT_SYMBOL(wait_for_completion_timeout);
5959 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5960 * @x: holds the state of this particular completion
5962 * This waits for completion of a specific task to be signaled. It is
5965 int __sched wait_for_completion_interruptible(struct completion *x)
5967 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5968 if (t == -ERESTARTSYS)
5972 EXPORT_SYMBOL(wait_for_completion_interruptible);
5975 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5976 * @x: holds the state of this particular completion
5977 * @timeout: timeout value in jiffies
5979 * This waits for either a completion of a specific task to be signaled or for a
5980 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5982 unsigned long __sched
5983 wait_for_completion_interruptible_timeout(struct completion *x,
5984 unsigned long timeout)
5986 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5988 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5991 * wait_for_completion_killable: - waits for completion of a task (killable)
5992 * @x: holds the state of this particular completion
5994 * This waits to be signaled for completion of a specific task. It can be
5995 * interrupted by a kill signal.
5997 int __sched wait_for_completion_killable(struct completion *x)
5999 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
6000 if (t == -ERESTARTSYS)
6004 EXPORT_SYMBOL(wait_for_completion_killable);
6007 * try_wait_for_completion - try to decrement a completion without blocking
6008 * @x: completion structure
6010 * Returns: 0 if a decrement cannot be done without blocking
6011 * 1 if a decrement succeeded.
6013 * If a completion is being used as a counting completion,
6014 * attempt to decrement the counter without blocking. This
6015 * enables us to avoid waiting if the resource the completion
6016 * is protecting is not available.
6018 bool try_wait_for_completion(struct completion *x)
6020 unsigned long flags;
6023 spin_lock_irqsave(&x->wait.lock, flags);
6028 spin_unlock_irqrestore(&x->wait.lock, flags);
6031 EXPORT_SYMBOL(try_wait_for_completion);
6034 * completion_done - Test to see if a completion has any waiters
6035 * @x: completion structure
6037 * Returns: 0 if there are waiters (wait_for_completion() in progress)
6038 * 1 if there are no waiters.
6041 bool completion_done(struct completion *x)
6043 unsigned long flags;
6046 spin_lock_irqsave(&x->wait.lock, flags);
6049 spin_unlock_irqrestore(&x->wait.lock, flags);
6052 EXPORT_SYMBOL(completion_done);
6055 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
6057 unsigned long flags;
6060 init_waitqueue_entry(&wait, current);
6062 __set_current_state(state);
6064 spin_lock_irqsave(&q->lock, flags);
6065 __add_wait_queue(q, &wait);
6066 spin_unlock(&q->lock);
6067 timeout = schedule_timeout(timeout);
6068 spin_lock_irq(&q->lock);
6069 __remove_wait_queue(q, &wait);
6070 spin_unlock_irqrestore(&q->lock, flags);
6075 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6077 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6079 EXPORT_SYMBOL(interruptible_sleep_on);
6082 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6084 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6086 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6088 void __sched sleep_on(wait_queue_head_t *q)
6090 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6092 EXPORT_SYMBOL(sleep_on);
6094 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6096 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6098 EXPORT_SYMBOL(sleep_on_timeout);
6100 #ifdef CONFIG_RT_MUTEXES
6103 * rt_mutex_setprio - set the current priority of a task
6105 * @prio: prio value (kernel-internal form)
6107 * This function changes the 'effective' priority of a task. It does
6108 * not touch ->normal_prio like __setscheduler().
6110 * Used by the rt_mutex code to implement priority inheritance logic.
6112 void rt_mutex_setprio(struct task_struct *p, int prio)
6114 unsigned long flags;
6115 int oldprio, on_rq, running;
6117 const struct sched_class *prev_class = p->sched_class;
6119 BUG_ON(prio < 0 || prio > MAX_PRIO);
6121 rq = task_rq_lock(p, &flags);
6122 update_rq_clock(rq);
6125 on_rq = p->se.on_rq;
6126 running = task_current(rq, p);
6128 dequeue_task(rq, p, 0);
6130 p->sched_class->put_prev_task(rq, p);
6133 p->sched_class = &rt_sched_class;
6135 p->sched_class = &fair_sched_class;
6140 p->sched_class->set_curr_task(rq);
6142 enqueue_task(rq, p, 0);
6144 check_class_changed(rq, p, prev_class, oldprio, running);
6146 task_rq_unlock(rq, &flags);
6151 void set_user_nice(struct task_struct *p, long nice)
6153 int old_prio, delta, on_rq;
6154 unsigned long flags;
6157 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6160 * We have to be careful, if called from sys_setpriority(),
6161 * the task might be in the middle of scheduling on another CPU.
6163 rq = task_rq_lock(p, &flags);
6164 update_rq_clock(rq);
6166 * The RT priorities are set via sched_setscheduler(), but we still
6167 * allow the 'normal' nice value to be set - but as expected
6168 * it wont have any effect on scheduling until the task is
6169 * SCHED_FIFO/SCHED_RR:
6171 if (task_has_rt_policy(p)) {
6172 p->static_prio = NICE_TO_PRIO(nice);
6175 on_rq = p->se.on_rq;
6177 dequeue_task(rq, p, 0);
6179 p->static_prio = NICE_TO_PRIO(nice);
6182 p->prio = effective_prio(p);
6183 delta = p->prio - old_prio;
6186 enqueue_task(rq, p, 0);
6188 * If the task increased its priority or is running and
6189 * lowered its priority, then reschedule its CPU:
6191 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6192 resched_task(rq->curr);
6195 task_rq_unlock(rq, &flags);
6197 EXPORT_SYMBOL(set_user_nice);
6200 * can_nice - check if a task can reduce its nice value
6204 int can_nice(const struct task_struct *p, const int nice)
6206 /* convert nice value [19,-20] to rlimit style value [1,40] */
6207 int nice_rlim = 20 - nice;
6209 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6210 capable(CAP_SYS_NICE));
6213 #ifdef __ARCH_WANT_SYS_NICE
6216 * sys_nice - change the priority of the current process.
6217 * @increment: priority increment
6219 * sys_setpriority is a more generic, but much slower function that
6220 * does similar things.
6222 SYSCALL_DEFINE1(nice, int, increment)
6227 * Setpriority might change our priority at the same moment.
6228 * We don't have to worry. Conceptually one call occurs first
6229 * and we have a single winner.
6231 if (increment < -40)
6236 nice = TASK_NICE(current) + increment;
6242 if (increment < 0 && !can_nice(current, nice))
6245 retval = security_task_setnice(current, nice);
6249 set_user_nice(current, nice);
6256 * task_prio - return the priority value of a given task.
6257 * @p: the task in question.
6259 * This is the priority value as seen by users in /proc.
6260 * RT tasks are offset by -200. Normal tasks are centered
6261 * around 0, value goes from -16 to +15.
6263 int task_prio(const struct task_struct *p)
6265 return p->prio - MAX_RT_PRIO;
6269 * task_nice - return the nice value of a given task.
6270 * @p: the task in question.
6272 int task_nice(const struct task_struct *p)
6274 return TASK_NICE(p);
6276 EXPORT_SYMBOL(task_nice);
6279 * idle_cpu - is a given cpu idle currently?
6280 * @cpu: the processor in question.
6282 int idle_cpu(int cpu)
6284 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6288 * idle_task - return the idle task for a given cpu.
6289 * @cpu: the processor in question.
6291 struct task_struct *idle_task(int cpu)
6293 return cpu_rq(cpu)->idle;
6297 * find_process_by_pid - find a process with a matching PID value.
6298 * @pid: the pid in question.
6300 static struct task_struct *find_process_by_pid(pid_t pid)
6302 return pid ? find_task_by_vpid(pid) : current;
6305 /* Actually do priority change: must hold rq lock. */
6307 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6309 BUG_ON(p->se.on_rq);
6312 p->rt_priority = prio;
6313 p->normal_prio = normal_prio(p);
6314 /* we are holding p->pi_lock already */
6315 p->prio = rt_mutex_getprio(p);
6316 if (rt_prio(p->prio))
6317 p->sched_class = &rt_sched_class;
6319 p->sched_class = &fair_sched_class;
6324 * check the target process has a UID that matches the current process's
6326 static bool check_same_owner(struct task_struct *p)
6328 const struct cred *cred = current_cred(), *pcred;
6332 pcred = __task_cred(p);
6333 match = (cred->euid == pcred->euid ||
6334 cred->euid == pcred->uid);
6339 static int __sched_setscheduler(struct task_struct *p, int policy,
6340 struct sched_param *param, bool user)
6342 int retval, oldprio, oldpolicy = -1, on_rq, running;
6343 unsigned long flags;
6344 const struct sched_class *prev_class = p->sched_class;
6348 /* may grab non-irq protected spin_locks */
6349 BUG_ON(in_interrupt());
6351 /* double check policy once rq lock held */
6353 reset_on_fork = p->sched_reset_on_fork;
6354 policy = oldpolicy = p->policy;
6356 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6357 policy &= ~SCHED_RESET_ON_FORK;
6359 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6360 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6361 policy != SCHED_IDLE)
6366 * Valid priorities for SCHED_FIFO and SCHED_RR are
6367 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6368 * SCHED_BATCH and SCHED_IDLE is 0.
6370 if (param->sched_priority < 0 ||
6371 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6372 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6374 if (rt_policy(policy) != (param->sched_priority != 0))
6378 * Allow unprivileged RT tasks to decrease priority:
6380 if (user && !capable(CAP_SYS_NICE)) {
6381 if (rt_policy(policy)) {
6382 unsigned long rlim_rtprio;
6384 if (!lock_task_sighand(p, &flags))
6386 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6387 unlock_task_sighand(p, &flags);
6389 /* can't set/change the rt policy */
6390 if (policy != p->policy && !rlim_rtprio)
6393 /* can't increase priority */
6394 if (param->sched_priority > p->rt_priority &&
6395 param->sched_priority > rlim_rtprio)
6399 * Like positive nice levels, dont allow tasks to
6400 * move out of SCHED_IDLE either:
6402 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6405 /* can't change other user's priorities */
6406 if (!check_same_owner(p))
6409 /* Normal users shall not reset the sched_reset_on_fork flag */
6410 if (p->sched_reset_on_fork && !reset_on_fork)
6415 #ifdef CONFIG_RT_GROUP_SCHED
6417 * Do not allow realtime tasks into groups that have no runtime
6420 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6421 task_group(p)->rt_bandwidth.rt_runtime == 0)
6425 retval = security_task_setscheduler(p, policy, param);
6431 * make sure no PI-waiters arrive (or leave) while we are
6432 * changing the priority of the task:
6434 raw_spin_lock_irqsave(&p->pi_lock, flags);
6436 * To be able to change p->policy safely, the apropriate
6437 * runqueue lock must be held.
6439 rq = __task_rq_lock(p);
6440 /* recheck policy now with rq lock held */
6441 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6442 policy = oldpolicy = -1;
6443 __task_rq_unlock(rq);
6444 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6447 update_rq_clock(rq);
6448 on_rq = p->se.on_rq;
6449 running = task_current(rq, p);
6451 deactivate_task(rq, p, 0);
6453 p->sched_class->put_prev_task(rq, p);
6455 p->sched_reset_on_fork = reset_on_fork;
6458 __setscheduler(rq, p, policy, param->sched_priority);
6461 p->sched_class->set_curr_task(rq);
6463 activate_task(rq, p, 0);
6465 check_class_changed(rq, p, prev_class, oldprio, running);
6467 __task_rq_unlock(rq);
6468 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6470 rt_mutex_adjust_pi(p);
6476 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6477 * @p: the task in question.
6478 * @policy: new policy.
6479 * @param: structure containing the new RT priority.
6481 * NOTE that the task may be already dead.
6483 int sched_setscheduler(struct task_struct *p, int policy,
6484 struct sched_param *param)
6486 return __sched_setscheduler(p, policy, param, true);
6488 EXPORT_SYMBOL_GPL(sched_setscheduler);
6491 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6492 * @p: the task in question.
6493 * @policy: new policy.
6494 * @param: structure containing the new RT priority.
6496 * Just like sched_setscheduler, only don't bother checking if the
6497 * current context has permission. For example, this is needed in
6498 * stop_machine(): we create temporary high priority worker threads,
6499 * but our caller might not have that capability.
6501 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6502 struct sched_param *param)
6504 return __sched_setscheduler(p, policy, param, false);
6508 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6510 struct sched_param lparam;
6511 struct task_struct *p;
6514 if (!param || pid < 0)
6516 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6521 p = find_process_by_pid(pid);
6523 retval = sched_setscheduler(p, policy, &lparam);
6530 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6531 * @pid: the pid in question.
6532 * @policy: new policy.
6533 * @param: structure containing the new RT priority.
6535 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6536 struct sched_param __user *, param)
6538 /* negative values for policy are not valid */
6542 return do_sched_setscheduler(pid, policy, param);
6546 * sys_sched_setparam - set/change the RT priority of a thread
6547 * @pid: the pid in question.
6548 * @param: structure containing the new RT priority.
6550 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6552 return do_sched_setscheduler(pid, -1, param);
6556 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6557 * @pid: the pid in question.
6559 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6561 struct task_struct *p;
6569 p = find_process_by_pid(pid);
6571 retval = security_task_getscheduler(p);
6574 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6581 * sys_sched_getparam - get the RT priority of a thread
6582 * @pid: the pid in question.
6583 * @param: structure containing the RT priority.
6585 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6587 struct sched_param lp;
6588 struct task_struct *p;
6591 if (!param || pid < 0)
6595 p = find_process_by_pid(pid);
6600 retval = security_task_getscheduler(p);
6604 lp.sched_priority = p->rt_priority;
6608 * This one might sleep, we cannot do it with a spinlock held ...
6610 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6619 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6621 cpumask_var_t cpus_allowed, new_mask;
6622 struct task_struct *p;
6628 p = find_process_by_pid(pid);
6635 /* Prevent p going away */
6639 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6643 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6645 goto out_free_cpus_allowed;
6648 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6651 retval = security_task_setscheduler(p, 0, NULL);
6655 cpuset_cpus_allowed(p, cpus_allowed);
6656 cpumask_and(new_mask, in_mask, cpus_allowed);
6658 retval = set_cpus_allowed_ptr(p, new_mask);
6661 cpuset_cpus_allowed(p, cpus_allowed);
6662 if (!cpumask_subset(new_mask, cpus_allowed)) {
6664 * We must have raced with a concurrent cpuset
6665 * update. Just reset the cpus_allowed to the
6666 * cpuset's cpus_allowed
6668 cpumask_copy(new_mask, cpus_allowed);
6673 free_cpumask_var(new_mask);
6674 out_free_cpus_allowed:
6675 free_cpumask_var(cpus_allowed);
6682 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6683 struct cpumask *new_mask)
6685 if (len < cpumask_size())
6686 cpumask_clear(new_mask);
6687 else if (len > cpumask_size())
6688 len = cpumask_size();
6690 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6694 * sys_sched_setaffinity - set the cpu affinity of a process
6695 * @pid: pid of the process
6696 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6697 * @user_mask_ptr: user-space pointer to the new cpu mask
6699 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6700 unsigned long __user *, user_mask_ptr)
6702 cpumask_var_t new_mask;
6705 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6708 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6710 retval = sched_setaffinity(pid, new_mask);
6711 free_cpumask_var(new_mask);
6715 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6717 struct task_struct *p;
6718 unsigned long flags;
6726 p = find_process_by_pid(pid);
6730 retval = security_task_getscheduler(p);
6734 rq = task_rq_lock(p, &flags);
6735 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6736 task_rq_unlock(rq, &flags);
6746 * sys_sched_getaffinity - get the cpu affinity of a process
6747 * @pid: pid of the process
6748 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6749 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6751 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6752 unsigned long __user *, user_mask_ptr)
6757 if (len < cpumask_size())
6760 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6763 ret = sched_getaffinity(pid, mask);
6765 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6768 ret = cpumask_size();
6770 free_cpumask_var(mask);
6776 * sys_sched_yield - yield the current processor to other threads.
6778 * This function yields the current CPU to other tasks. If there are no
6779 * other threads running on this CPU then this function will return.
6781 SYSCALL_DEFINE0(sched_yield)
6783 struct rq *rq = this_rq_lock();
6785 schedstat_inc(rq, yld_count);
6786 current->sched_class->yield_task(rq);
6789 * Since we are going to call schedule() anyway, there's
6790 * no need to preempt or enable interrupts:
6792 __release(rq->lock);
6793 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6794 do_raw_spin_unlock(&rq->lock);
6795 preempt_enable_no_resched();
6802 static inline int should_resched(void)
6804 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6807 static void __cond_resched(void)
6809 add_preempt_count(PREEMPT_ACTIVE);
6811 sub_preempt_count(PREEMPT_ACTIVE);
6814 int __sched _cond_resched(void)
6816 if (should_resched()) {
6822 EXPORT_SYMBOL(_cond_resched);
6825 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6826 * call schedule, and on return reacquire the lock.
6828 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6829 * operations here to prevent schedule() from being called twice (once via
6830 * spin_unlock(), once by hand).
6832 int __cond_resched_lock(spinlock_t *lock)
6834 int resched = should_resched();
6837 lockdep_assert_held(lock);
6839 if (spin_needbreak(lock) || resched) {
6850 EXPORT_SYMBOL(__cond_resched_lock);
6852 int __sched __cond_resched_softirq(void)
6854 BUG_ON(!in_softirq());
6856 if (should_resched()) {
6864 EXPORT_SYMBOL(__cond_resched_softirq);
6867 * yield - yield the current processor to other threads.
6869 * This is a shortcut for kernel-space yielding - it marks the
6870 * thread runnable and calls sys_sched_yield().
6872 void __sched yield(void)
6874 set_current_state(TASK_RUNNING);
6877 EXPORT_SYMBOL(yield);
6880 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6881 * that process accounting knows that this is a task in IO wait state.
6883 void __sched io_schedule(void)
6885 struct rq *rq = raw_rq();
6887 delayacct_blkio_start();
6888 atomic_inc(&rq->nr_iowait);
6889 current->in_iowait = 1;
6891 current->in_iowait = 0;
6892 atomic_dec(&rq->nr_iowait);
6893 delayacct_blkio_end();
6895 EXPORT_SYMBOL(io_schedule);
6897 long __sched io_schedule_timeout(long timeout)
6899 struct rq *rq = raw_rq();
6902 delayacct_blkio_start();
6903 atomic_inc(&rq->nr_iowait);
6904 current->in_iowait = 1;
6905 ret = schedule_timeout(timeout);
6906 current->in_iowait = 0;
6907 atomic_dec(&rq->nr_iowait);
6908 delayacct_blkio_end();
6913 * sys_sched_get_priority_max - return maximum RT priority.
6914 * @policy: scheduling class.
6916 * this syscall returns the maximum rt_priority that can be used
6917 * by a given scheduling class.
6919 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6926 ret = MAX_USER_RT_PRIO-1;
6938 * sys_sched_get_priority_min - return minimum RT priority.
6939 * @policy: scheduling class.
6941 * this syscall returns the minimum rt_priority that can be used
6942 * by a given scheduling class.
6944 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6962 * sys_sched_rr_get_interval - return the default timeslice of a process.
6963 * @pid: pid of the process.
6964 * @interval: userspace pointer to the timeslice value.
6966 * this syscall writes the default timeslice value of a given process
6967 * into the user-space timespec buffer. A value of '0' means infinity.
6969 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6970 struct timespec __user *, interval)
6972 struct task_struct *p;
6973 unsigned int time_slice;
6974 unsigned long flags;
6984 p = find_process_by_pid(pid);
6988 retval = security_task_getscheduler(p);
6992 rq = task_rq_lock(p, &flags);
6993 time_slice = p->sched_class->get_rr_interval(rq, p);
6994 task_rq_unlock(rq, &flags);
6997 jiffies_to_timespec(time_slice, &t);
6998 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
7006 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
7008 void sched_show_task(struct task_struct *p)
7010 unsigned long free = 0;
7013 state = p->state ? __ffs(p->state) + 1 : 0;
7014 printk(KERN_INFO "%-13.13s %c", p->comm,
7015 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
7016 #if BITS_PER_LONG == 32
7017 if (state == TASK_RUNNING)
7018 printk(KERN_CONT " running ");
7020 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
7022 if (state == TASK_RUNNING)
7023 printk(KERN_CONT " running task ");
7025 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
7027 #ifdef CONFIG_DEBUG_STACK_USAGE
7028 free = stack_not_used(p);
7030 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
7031 task_pid_nr(p), task_pid_nr(p->real_parent),
7032 (unsigned long)task_thread_info(p)->flags);
7034 show_stack(p, NULL);
7037 void show_state_filter(unsigned long state_filter)
7039 struct task_struct *g, *p;
7041 #if BITS_PER_LONG == 32
7043 " task PC stack pid father\n");
7046 " task PC stack pid father\n");
7048 read_lock(&tasklist_lock);
7049 do_each_thread(g, p) {
7051 * reset the NMI-timeout, listing all files on a slow
7052 * console might take alot of time:
7054 touch_nmi_watchdog();
7055 if (!state_filter || (p->state & state_filter))
7057 } while_each_thread(g, p);
7059 touch_all_softlockup_watchdogs();
7061 #ifdef CONFIG_SCHED_DEBUG
7062 sysrq_sched_debug_show();
7064 read_unlock(&tasklist_lock);
7066 * Only show locks if all tasks are dumped:
7069 debug_show_all_locks();
7072 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7074 idle->sched_class = &idle_sched_class;
7078 * init_idle - set up an idle thread for a given CPU
7079 * @idle: task in question
7080 * @cpu: cpu the idle task belongs to
7082 * NOTE: this function does not set the idle thread's NEED_RESCHED
7083 * flag, to make booting more robust.
7085 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7087 struct rq *rq = cpu_rq(cpu);
7088 unsigned long flags;
7090 raw_spin_lock_irqsave(&rq->lock, flags);
7093 idle->state = TASK_RUNNING;
7094 idle->se.exec_start = sched_clock();
7096 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7097 __set_task_cpu(idle, cpu);
7099 rq->curr = rq->idle = idle;
7100 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7103 raw_spin_unlock_irqrestore(&rq->lock, flags);
7105 /* Set the preempt count _outside_ the spinlocks! */
7106 #if defined(CONFIG_PREEMPT)
7107 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7109 task_thread_info(idle)->preempt_count = 0;
7112 * The idle tasks have their own, simple scheduling class:
7114 idle->sched_class = &idle_sched_class;
7115 ftrace_graph_init_task(idle);
7119 * In a system that switches off the HZ timer nohz_cpu_mask
7120 * indicates which cpus entered this state. This is used
7121 * in the rcu update to wait only for active cpus. For system
7122 * which do not switch off the HZ timer nohz_cpu_mask should
7123 * always be CPU_BITS_NONE.
7125 cpumask_var_t nohz_cpu_mask;
7128 * Increase the granularity value when there are more CPUs,
7129 * because with more CPUs the 'effective latency' as visible
7130 * to users decreases. But the relationship is not linear,
7131 * so pick a second-best guess by going with the log2 of the
7134 * This idea comes from the SD scheduler of Con Kolivas:
7136 static int get_update_sysctl_factor(void)
7138 unsigned int cpus = min_t(int, num_online_cpus(), 8);
7139 unsigned int factor;
7141 switch (sysctl_sched_tunable_scaling) {
7142 case SCHED_TUNABLESCALING_NONE:
7145 case SCHED_TUNABLESCALING_LINEAR:
7148 case SCHED_TUNABLESCALING_LOG:
7150 factor = 1 + ilog2(cpus);
7157 static void update_sysctl(void)
7159 unsigned int factor = get_update_sysctl_factor();
7161 #define SET_SYSCTL(name) \
7162 (sysctl_##name = (factor) * normalized_sysctl_##name)
7163 SET_SYSCTL(sched_min_granularity);
7164 SET_SYSCTL(sched_latency);
7165 SET_SYSCTL(sched_wakeup_granularity);
7166 SET_SYSCTL(sched_shares_ratelimit);
7170 static inline void sched_init_granularity(void)
7177 * This is how migration works:
7179 * 1) we queue a struct migration_req structure in the source CPU's
7180 * runqueue and wake up that CPU's migration thread.
7181 * 2) we down() the locked semaphore => thread blocks.
7182 * 3) migration thread wakes up (implicitly it forces the migrated
7183 * thread off the CPU)
7184 * 4) it gets the migration request and checks whether the migrated
7185 * task is still in the wrong runqueue.
7186 * 5) if it's in the wrong runqueue then the migration thread removes
7187 * it and puts it into the right queue.
7188 * 6) migration thread up()s the semaphore.
7189 * 7) we wake up and the migration is done.
7193 * Change a given task's CPU affinity. Migrate the thread to a
7194 * proper CPU and schedule it away if the CPU it's executing on
7195 * is removed from the allowed bitmask.
7197 * NOTE: the caller must have a valid reference to the task, the
7198 * task must not exit() & deallocate itself prematurely. The
7199 * call is not atomic; no spinlocks may be held.
7201 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7203 struct migration_req req;
7204 unsigned long flags;
7208 rq = task_rq_lock(p, &flags);
7210 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7215 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7216 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7221 if (p->sched_class->set_cpus_allowed)
7222 p->sched_class->set_cpus_allowed(p, new_mask);
7224 cpumask_copy(&p->cpus_allowed, new_mask);
7225 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7228 /* Can the task run on the task's current CPU? If so, we're done */
7229 if (cpumask_test_cpu(task_cpu(p), new_mask))
7232 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7233 /* Need help from migration thread: drop lock and wait. */
7234 struct task_struct *mt = rq->migration_thread;
7236 get_task_struct(mt);
7237 task_rq_unlock(rq, &flags);
7238 wake_up_process(rq->migration_thread);
7239 put_task_struct(mt);
7240 wait_for_completion(&req.done);
7241 tlb_migrate_finish(p->mm);
7245 task_rq_unlock(rq, &flags);
7249 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7252 * Move (not current) task off this cpu, onto dest cpu. We're doing
7253 * this because either it can't run here any more (set_cpus_allowed()
7254 * away from this CPU, or CPU going down), or because we're
7255 * attempting to rebalance this task on exec (sched_exec).
7257 * So we race with normal scheduler movements, but that's OK, as long
7258 * as the task is no longer on this CPU.
7260 * Returns non-zero if task was successfully migrated.
7262 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7264 struct rq *rq_dest, *rq_src;
7267 if (unlikely(!cpu_active(dest_cpu)))
7270 rq_src = cpu_rq(src_cpu);
7271 rq_dest = cpu_rq(dest_cpu);
7273 double_rq_lock(rq_src, rq_dest);
7274 /* Already moved. */
7275 if (task_cpu(p) != src_cpu)
7277 /* Affinity changed (again). */
7278 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7282 * If we're not on a rq, the next wake-up will ensure we're
7286 deactivate_task(rq_src, p, 0);
7287 set_task_cpu(p, dest_cpu);
7288 activate_task(rq_dest, p, 0);
7289 check_preempt_curr(rq_dest, p, 0);
7294 double_rq_unlock(rq_src, rq_dest);
7298 #define RCU_MIGRATION_IDLE 0
7299 #define RCU_MIGRATION_NEED_QS 1
7300 #define RCU_MIGRATION_GOT_QS 2
7301 #define RCU_MIGRATION_MUST_SYNC 3
7304 * migration_thread - this is a highprio system thread that performs
7305 * thread migration by bumping thread off CPU then 'pushing' onto
7308 static int migration_thread(void *data)
7311 int cpu = (long)data;
7315 BUG_ON(rq->migration_thread != current);
7317 set_current_state(TASK_INTERRUPTIBLE);
7318 while (!kthread_should_stop()) {
7319 struct migration_req *req;
7320 struct list_head *head;
7322 raw_spin_lock_irq(&rq->lock);
7324 if (cpu_is_offline(cpu)) {
7325 raw_spin_unlock_irq(&rq->lock);
7329 if (rq->active_balance) {
7330 active_load_balance(rq, cpu);
7331 rq->active_balance = 0;
7334 head = &rq->migration_queue;
7336 if (list_empty(head)) {
7337 raw_spin_unlock_irq(&rq->lock);
7339 set_current_state(TASK_INTERRUPTIBLE);
7342 req = list_entry(head->next, struct migration_req, list);
7343 list_del_init(head->next);
7345 if (req->task != NULL) {
7346 raw_spin_unlock(&rq->lock);
7347 __migrate_task(req->task, cpu, req->dest_cpu);
7348 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7349 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7350 raw_spin_unlock(&rq->lock);
7352 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7353 raw_spin_unlock(&rq->lock);
7354 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7358 complete(&req->done);
7360 __set_current_state(TASK_RUNNING);
7365 #ifdef CONFIG_HOTPLUG_CPU
7367 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7371 local_irq_disable();
7372 ret = __migrate_task(p, src_cpu, dest_cpu);
7378 * Figure out where task on dead CPU should go, use force if necessary.
7380 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7385 dest_cpu = select_fallback_rq(dead_cpu, p);
7387 /* It can have affinity changed while we were choosing. */
7388 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7393 * While a dead CPU has no uninterruptible tasks queued at this point,
7394 * it might still have a nonzero ->nr_uninterruptible counter, because
7395 * for performance reasons the counter is not stricly tracking tasks to
7396 * their home CPUs. So we just add the counter to another CPU's counter,
7397 * to keep the global sum constant after CPU-down:
7399 static void migrate_nr_uninterruptible(struct rq *rq_src)
7401 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7402 unsigned long flags;
7404 local_irq_save(flags);
7405 double_rq_lock(rq_src, rq_dest);
7406 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7407 rq_src->nr_uninterruptible = 0;
7408 double_rq_unlock(rq_src, rq_dest);
7409 local_irq_restore(flags);
7412 /* Run through task list and migrate tasks from the dead cpu. */
7413 static void migrate_live_tasks(int src_cpu)
7415 struct task_struct *p, *t;
7417 read_lock(&tasklist_lock);
7419 do_each_thread(t, p) {
7423 if (task_cpu(p) == src_cpu)
7424 move_task_off_dead_cpu(src_cpu, p);
7425 } while_each_thread(t, p);
7427 read_unlock(&tasklist_lock);
7431 * Schedules idle task to be the next runnable task on current CPU.
7432 * It does so by boosting its priority to highest possible.
7433 * Used by CPU offline code.
7435 void sched_idle_next(void)
7437 int this_cpu = smp_processor_id();
7438 struct rq *rq = cpu_rq(this_cpu);
7439 struct task_struct *p = rq->idle;
7440 unsigned long flags;
7442 /* cpu has to be offline */
7443 BUG_ON(cpu_online(this_cpu));
7446 * Strictly not necessary since rest of the CPUs are stopped by now
7447 * and interrupts disabled on the current cpu.
7449 raw_spin_lock_irqsave(&rq->lock, flags);
7451 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7453 update_rq_clock(rq);
7454 activate_task(rq, p, 0);
7456 raw_spin_unlock_irqrestore(&rq->lock, flags);
7460 * Ensures that the idle task is using init_mm right before its cpu goes
7463 void idle_task_exit(void)
7465 struct mm_struct *mm = current->active_mm;
7467 BUG_ON(cpu_online(smp_processor_id()));
7470 switch_mm(mm, &init_mm, current);
7474 /* called under rq->lock with disabled interrupts */
7475 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7477 struct rq *rq = cpu_rq(dead_cpu);
7479 /* Must be exiting, otherwise would be on tasklist. */
7480 BUG_ON(!p->exit_state);
7482 /* Cannot have done final schedule yet: would have vanished. */
7483 BUG_ON(p->state == TASK_DEAD);
7488 * Drop lock around migration; if someone else moves it,
7489 * that's OK. No task can be added to this CPU, so iteration is
7492 raw_spin_unlock_irq(&rq->lock);
7493 move_task_off_dead_cpu(dead_cpu, p);
7494 raw_spin_lock_irq(&rq->lock);
7499 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7500 static void migrate_dead_tasks(unsigned int dead_cpu)
7502 struct rq *rq = cpu_rq(dead_cpu);
7503 struct task_struct *next;
7506 if (!rq->nr_running)
7508 update_rq_clock(rq);
7509 next = pick_next_task(rq);
7512 next->sched_class->put_prev_task(rq, next);
7513 migrate_dead(dead_cpu, next);
7519 * remove the tasks which were accounted by rq from calc_load_tasks.
7521 static void calc_global_load_remove(struct rq *rq)
7523 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7524 rq->calc_load_active = 0;
7526 #endif /* CONFIG_HOTPLUG_CPU */
7528 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7530 static struct ctl_table sd_ctl_dir[] = {
7532 .procname = "sched_domain",
7538 static struct ctl_table sd_ctl_root[] = {
7540 .procname = "kernel",
7542 .child = sd_ctl_dir,
7547 static struct ctl_table *sd_alloc_ctl_entry(int n)
7549 struct ctl_table *entry =
7550 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7555 static void sd_free_ctl_entry(struct ctl_table **tablep)
7557 struct ctl_table *entry;
7560 * In the intermediate directories, both the child directory and
7561 * procname are dynamically allocated and could fail but the mode
7562 * will always be set. In the lowest directory the names are
7563 * static strings and all have proc handlers.
7565 for (entry = *tablep; entry->mode; entry++) {
7567 sd_free_ctl_entry(&entry->child);
7568 if (entry->proc_handler == NULL)
7569 kfree(entry->procname);
7577 set_table_entry(struct ctl_table *entry,
7578 const char *procname, void *data, int maxlen,
7579 mode_t mode, proc_handler *proc_handler)
7581 entry->procname = procname;
7583 entry->maxlen = maxlen;
7585 entry->proc_handler = proc_handler;
7588 static struct ctl_table *
7589 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7591 struct ctl_table *table = sd_alloc_ctl_entry(13);
7596 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7597 sizeof(long), 0644, proc_doulongvec_minmax);
7598 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7599 sizeof(long), 0644, proc_doulongvec_minmax);
7600 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7601 sizeof(int), 0644, proc_dointvec_minmax);
7602 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7603 sizeof(int), 0644, proc_dointvec_minmax);
7604 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7605 sizeof(int), 0644, proc_dointvec_minmax);
7606 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7607 sizeof(int), 0644, proc_dointvec_minmax);
7608 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7609 sizeof(int), 0644, proc_dointvec_minmax);
7610 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7611 sizeof(int), 0644, proc_dointvec_minmax);
7612 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7613 sizeof(int), 0644, proc_dointvec_minmax);
7614 set_table_entry(&table[9], "cache_nice_tries",
7615 &sd->cache_nice_tries,
7616 sizeof(int), 0644, proc_dointvec_minmax);
7617 set_table_entry(&table[10], "flags", &sd->flags,
7618 sizeof(int), 0644, proc_dointvec_minmax);
7619 set_table_entry(&table[11], "name", sd->name,
7620 CORENAME_MAX_SIZE, 0444, proc_dostring);
7621 /* &table[12] is terminator */
7626 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7628 struct ctl_table *entry, *table;
7629 struct sched_domain *sd;
7630 int domain_num = 0, i;
7633 for_each_domain(cpu, sd)
7635 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7640 for_each_domain(cpu, sd) {
7641 snprintf(buf, 32, "domain%d", i);
7642 entry->procname = kstrdup(buf, GFP_KERNEL);
7644 entry->child = sd_alloc_ctl_domain_table(sd);
7651 static struct ctl_table_header *sd_sysctl_header;
7652 static void register_sched_domain_sysctl(void)
7654 int i, cpu_num = num_possible_cpus();
7655 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7658 WARN_ON(sd_ctl_dir[0].child);
7659 sd_ctl_dir[0].child = entry;
7664 for_each_possible_cpu(i) {
7665 snprintf(buf, 32, "cpu%d", i);
7666 entry->procname = kstrdup(buf, GFP_KERNEL);
7668 entry->child = sd_alloc_ctl_cpu_table(i);
7672 WARN_ON(sd_sysctl_header);
7673 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7676 /* may be called multiple times per register */
7677 static void unregister_sched_domain_sysctl(void)
7679 if (sd_sysctl_header)
7680 unregister_sysctl_table(sd_sysctl_header);
7681 sd_sysctl_header = NULL;
7682 if (sd_ctl_dir[0].child)
7683 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7686 static void register_sched_domain_sysctl(void)
7689 static void unregister_sched_domain_sysctl(void)
7694 static void set_rq_online(struct rq *rq)
7697 const struct sched_class *class;
7699 cpumask_set_cpu(rq->cpu, rq->rd->online);
7702 for_each_class(class) {
7703 if (class->rq_online)
7704 class->rq_online(rq);
7709 static void set_rq_offline(struct rq *rq)
7712 const struct sched_class *class;
7714 for_each_class(class) {
7715 if (class->rq_offline)
7716 class->rq_offline(rq);
7719 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7725 * migration_call - callback that gets triggered when a CPU is added.
7726 * Here we can start up the necessary migration thread for the new CPU.
7728 static int __cpuinit
7729 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7731 struct task_struct *p;
7732 int cpu = (long)hcpu;
7733 unsigned long flags;
7738 case CPU_UP_PREPARE:
7739 case CPU_UP_PREPARE_FROZEN:
7740 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7743 kthread_bind(p, cpu);
7744 /* Must be high prio: stop_machine expects to yield to it. */
7745 rq = task_rq_lock(p, &flags);
7746 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7747 task_rq_unlock(rq, &flags);
7749 cpu_rq(cpu)->migration_thread = p;
7750 rq->calc_load_update = calc_load_update;
7754 case CPU_ONLINE_FROZEN:
7755 /* Strictly unnecessary, as first user will wake it. */
7756 wake_up_process(cpu_rq(cpu)->migration_thread);
7758 /* Update our root-domain */
7760 raw_spin_lock_irqsave(&rq->lock, flags);
7762 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7766 raw_spin_unlock_irqrestore(&rq->lock, flags);
7769 #ifdef CONFIG_HOTPLUG_CPU
7770 case CPU_UP_CANCELED:
7771 case CPU_UP_CANCELED_FROZEN:
7772 if (!cpu_rq(cpu)->migration_thread)
7774 /* Unbind it from offline cpu so it can run. Fall thru. */
7775 kthread_bind(cpu_rq(cpu)->migration_thread,
7776 cpumask_any(cpu_online_mask));
7777 kthread_stop(cpu_rq(cpu)->migration_thread);
7778 put_task_struct(cpu_rq(cpu)->migration_thread);
7779 cpu_rq(cpu)->migration_thread = NULL;
7783 case CPU_DEAD_FROZEN:
7784 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7785 migrate_live_tasks(cpu);
7787 kthread_stop(rq->migration_thread);
7788 put_task_struct(rq->migration_thread);
7789 rq->migration_thread = NULL;
7790 /* Idle task back to normal (off runqueue, low prio) */
7791 raw_spin_lock_irq(&rq->lock);
7792 update_rq_clock(rq);
7793 deactivate_task(rq, rq->idle, 0);
7794 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7795 rq->idle->sched_class = &idle_sched_class;
7796 migrate_dead_tasks(cpu);
7797 raw_spin_unlock_irq(&rq->lock);
7799 migrate_nr_uninterruptible(rq);
7800 BUG_ON(rq->nr_running != 0);
7801 calc_global_load_remove(rq);
7803 * No need to migrate the tasks: it was best-effort if
7804 * they didn't take sched_hotcpu_mutex. Just wake up
7807 raw_spin_lock_irq(&rq->lock);
7808 while (!list_empty(&rq->migration_queue)) {
7809 struct migration_req *req;
7811 req = list_entry(rq->migration_queue.next,
7812 struct migration_req, list);
7813 list_del_init(&req->list);
7814 raw_spin_unlock_irq(&rq->lock);
7815 complete(&req->done);
7816 raw_spin_lock_irq(&rq->lock);
7818 raw_spin_unlock_irq(&rq->lock);
7822 case CPU_DYING_FROZEN:
7823 /* Update our root-domain */
7825 raw_spin_lock_irqsave(&rq->lock, flags);
7827 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7830 raw_spin_unlock_irqrestore(&rq->lock, flags);
7838 * Register at high priority so that task migration (migrate_all_tasks)
7839 * happens before everything else. This has to be lower priority than
7840 * the notifier in the perf_event subsystem, though.
7842 static struct notifier_block __cpuinitdata migration_notifier = {
7843 .notifier_call = migration_call,
7847 static int __init migration_init(void)
7849 void *cpu = (void *)(long)smp_processor_id();
7852 /* Start one for the boot CPU: */
7853 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7854 BUG_ON(err == NOTIFY_BAD);
7855 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7856 register_cpu_notifier(&migration_notifier);
7860 early_initcall(migration_init);
7865 #ifdef CONFIG_SCHED_DEBUG
7867 static __read_mostly int sched_domain_debug_enabled;
7869 static int __init sched_domain_debug_setup(char *str)
7871 sched_domain_debug_enabled = 1;
7875 early_param("sched_debug", sched_domain_debug_setup);
7877 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7878 struct cpumask *groupmask)
7880 struct sched_group *group = sd->groups;
7883 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7884 cpumask_clear(groupmask);
7886 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7888 if (!(sd->flags & SD_LOAD_BALANCE)) {
7889 printk("does not load-balance\n");
7891 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7896 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7898 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7899 printk(KERN_ERR "ERROR: domain->span does not contain "
7902 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7903 printk(KERN_ERR "ERROR: domain->groups does not contain"
7907 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7911 printk(KERN_ERR "ERROR: group is NULL\n");
7915 if (!group->cpu_power) {
7916 printk(KERN_CONT "\n");
7917 printk(KERN_ERR "ERROR: domain->cpu_power not "
7922 if (!cpumask_weight(sched_group_cpus(group))) {
7923 printk(KERN_CONT "\n");
7924 printk(KERN_ERR "ERROR: empty group\n");
7928 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7929 printk(KERN_CONT "\n");
7930 printk(KERN_ERR "ERROR: repeated CPUs\n");
7934 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7936 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7938 printk(KERN_CONT " %s", str);
7939 if (group->cpu_power != SCHED_LOAD_SCALE) {
7940 printk(KERN_CONT " (cpu_power = %d)",
7944 group = group->next;
7945 } while (group != sd->groups);
7946 printk(KERN_CONT "\n");
7948 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7949 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7952 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7953 printk(KERN_ERR "ERROR: parent span is not a superset "
7954 "of domain->span\n");
7958 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7960 cpumask_var_t groupmask;
7963 if (!sched_domain_debug_enabled)
7967 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7971 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7973 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7974 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7979 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7986 free_cpumask_var(groupmask);
7988 #else /* !CONFIG_SCHED_DEBUG */
7989 # define sched_domain_debug(sd, cpu) do { } while (0)
7990 #endif /* CONFIG_SCHED_DEBUG */
7992 static int sd_degenerate(struct sched_domain *sd)
7994 if (cpumask_weight(sched_domain_span(sd)) == 1)
7997 /* Following flags need at least 2 groups */
7998 if (sd->flags & (SD_LOAD_BALANCE |
7999 SD_BALANCE_NEWIDLE |
8003 SD_SHARE_PKG_RESOURCES)) {
8004 if (sd->groups != sd->groups->next)
8008 /* Following flags don't use groups */
8009 if (sd->flags & (SD_WAKE_AFFINE))
8016 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
8018 unsigned long cflags = sd->flags, pflags = parent->flags;
8020 if (sd_degenerate(parent))
8023 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
8026 /* Flags needing groups don't count if only 1 group in parent */
8027 if (parent->groups == parent->groups->next) {
8028 pflags &= ~(SD_LOAD_BALANCE |
8029 SD_BALANCE_NEWIDLE |
8033 SD_SHARE_PKG_RESOURCES);
8034 if (nr_node_ids == 1)
8035 pflags &= ~SD_SERIALIZE;
8037 if (~cflags & pflags)
8043 static void free_rootdomain(struct root_domain *rd)
8045 synchronize_sched();
8047 cpupri_cleanup(&rd->cpupri);
8049 free_cpumask_var(rd->rto_mask);
8050 free_cpumask_var(rd->online);
8051 free_cpumask_var(rd->span);
8055 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8057 struct root_domain *old_rd = NULL;
8058 unsigned long flags;
8060 raw_spin_lock_irqsave(&rq->lock, flags);
8065 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8068 cpumask_clear_cpu(rq->cpu, old_rd->span);
8071 * If we dont want to free the old_rt yet then
8072 * set old_rd to NULL to skip the freeing later
8075 if (!atomic_dec_and_test(&old_rd->refcount))
8079 atomic_inc(&rd->refcount);
8082 cpumask_set_cpu(rq->cpu, rd->span);
8083 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8086 raw_spin_unlock_irqrestore(&rq->lock, flags);
8089 free_rootdomain(old_rd);
8092 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8094 gfp_t gfp = GFP_KERNEL;
8096 memset(rd, 0, sizeof(*rd));
8101 if (!alloc_cpumask_var(&rd->span, gfp))
8103 if (!alloc_cpumask_var(&rd->online, gfp))
8105 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8108 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8113 free_cpumask_var(rd->rto_mask);
8115 free_cpumask_var(rd->online);
8117 free_cpumask_var(rd->span);
8122 static void init_defrootdomain(void)
8124 init_rootdomain(&def_root_domain, true);
8126 atomic_set(&def_root_domain.refcount, 1);
8129 static struct root_domain *alloc_rootdomain(void)
8131 struct root_domain *rd;
8133 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8137 if (init_rootdomain(rd, false) != 0) {
8146 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8147 * hold the hotplug lock.
8150 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8152 struct rq *rq = cpu_rq(cpu);
8153 struct sched_domain *tmp;
8155 /* Remove the sched domains which do not contribute to scheduling. */
8156 for (tmp = sd; tmp; ) {
8157 struct sched_domain *parent = tmp->parent;
8161 if (sd_parent_degenerate(tmp, parent)) {
8162 tmp->parent = parent->parent;
8164 parent->parent->child = tmp;
8169 if (sd && sd_degenerate(sd)) {
8175 sched_domain_debug(sd, cpu);
8177 rq_attach_root(rq, rd);
8178 rcu_assign_pointer(rq->sd, sd);
8181 /* cpus with isolated domains */
8182 static cpumask_var_t cpu_isolated_map;
8184 /* Setup the mask of cpus configured for isolated domains */
8185 static int __init isolated_cpu_setup(char *str)
8187 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8188 cpulist_parse(str, cpu_isolated_map);
8192 __setup("isolcpus=", isolated_cpu_setup);
8195 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8196 * to a function which identifies what group(along with sched group) a CPU
8197 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8198 * (due to the fact that we keep track of groups covered with a struct cpumask).
8200 * init_sched_build_groups will build a circular linked list of the groups
8201 * covered by the given span, and will set each group's ->cpumask correctly,
8202 * and ->cpu_power to 0.
8205 init_sched_build_groups(const struct cpumask *span,
8206 const struct cpumask *cpu_map,
8207 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8208 struct sched_group **sg,
8209 struct cpumask *tmpmask),
8210 struct cpumask *covered, struct cpumask *tmpmask)
8212 struct sched_group *first = NULL, *last = NULL;
8215 cpumask_clear(covered);
8217 for_each_cpu(i, span) {
8218 struct sched_group *sg;
8219 int group = group_fn(i, cpu_map, &sg, tmpmask);
8222 if (cpumask_test_cpu(i, covered))
8225 cpumask_clear(sched_group_cpus(sg));
8228 for_each_cpu(j, span) {
8229 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8232 cpumask_set_cpu(j, covered);
8233 cpumask_set_cpu(j, sched_group_cpus(sg));
8244 #define SD_NODES_PER_DOMAIN 16
8249 * find_next_best_node - find the next node to include in a sched_domain
8250 * @node: node whose sched_domain we're building
8251 * @used_nodes: nodes already in the sched_domain
8253 * Find the next node to include in a given scheduling domain. Simply
8254 * finds the closest node not already in the @used_nodes map.
8256 * Should use nodemask_t.
8258 static int find_next_best_node(int node, nodemask_t *used_nodes)
8260 int i, n, val, min_val, best_node = 0;
8264 for (i = 0; i < nr_node_ids; i++) {
8265 /* Start at @node */
8266 n = (node + i) % nr_node_ids;
8268 if (!nr_cpus_node(n))
8271 /* Skip already used nodes */
8272 if (node_isset(n, *used_nodes))
8275 /* Simple min distance search */
8276 val = node_distance(node, n);
8278 if (val < min_val) {
8284 node_set(best_node, *used_nodes);
8289 * sched_domain_node_span - get a cpumask for a node's sched_domain
8290 * @node: node whose cpumask we're constructing
8291 * @span: resulting cpumask
8293 * Given a node, construct a good cpumask for its sched_domain to span. It
8294 * should be one that prevents unnecessary balancing, but also spreads tasks
8297 static void sched_domain_node_span(int node, struct cpumask *span)
8299 nodemask_t used_nodes;
8302 cpumask_clear(span);
8303 nodes_clear(used_nodes);
8305 cpumask_or(span, span, cpumask_of_node(node));
8306 node_set(node, used_nodes);
8308 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8309 int next_node = find_next_best_node(node, &used_nodes);
8311 cpumask_or(span, span, cpumask_of_node(next_node));
8314 #endif /* CONFIG_NUMA */
8316 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8319 * The cpus mask in sched_group and sched_domain hangs off the end.
8321 * ( See the the comments in include/linux/sched.h:struct sched_group
8322 * and struct sched_domain. )
8324 struct static_sched_group {
8325 struct sched_group sg;
8326 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8329 struct static_sched_domain {
8330 struct sched_domain sd;
8331 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8337 cpumask_var_t domainspan;
8338 cpumask_var_t covered;
8339 cpumask_var_t notcovered;
8341 cpumask_var_t nodemask;
8342 cpumask_var_t this_sibling_map;
8343 cpumask_var_t this_core_map;
8344 cpumask_var_t send_covered;
8345 cpumask_var_t tmpmask;
8346 struct sched_group **sched_group_nodes;
8347 struct root_domain *rd;
8351 sa_sched_groups = 0,
8356 sa_this_sibling_map,
8358 sa_sched_group_nodes,
8368 * SMT sched-domains:
8370 #ifdef CONFIG_SCHED_SMT
8371 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8372 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
8375 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8376 struct sched_group **sg, struct cpumask *unused)
8379 *sg = &per_cpu(sched_groups, cpu).sg;
8382 #endif /* CONFIG_SCHED_SMT */
8385 * multi-core sched-domains:
8387 #ifdef CONFIG_SCHED_MC
8388 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8389 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8390 #endif /* CONFIG_SCHED_MC */
8392 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8394 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8395 struct sched_group **sg, struct cpumask *mask)
8399 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8400 group = cpumask_first(mask);
8402 *sg = &per_cpu(sched_group_core, group).sg;
8405 #elif defined(CONFIG_SCHED_MC)
8407 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8408 struct sched_group **sg, struct cpumask *unused)
8411 *sg = &per_cpu(sched_group_core, cpu).sg;
8416 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8417 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8420 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8421 struct sched_group **sg, struct cpumask *mask)
8424 #ifdef CONFIG_SCHED_MC
8425 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8426 group = cpumask_first(mask);
8427 #elif defined(CONFIG_SCHED_SMT)
8428 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8429 group = cpumask_first(mask);
8434 *sg = &per_cpu(sched_group_phys, group).sg;
8440 * The init_sched_build_groups can't handle what we want to do with node
8441 * groups, so roll our own. Now each node has its own list of groups which
8442 * gets dynamically allocated.
8444 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8445 static struct sched_group ***sched_group_nodes_bycpu;
8447 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8448 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8450 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8451 struct sched_group **sg,
8452 struct cpumask *nodemask)
8456 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8457 group = cpumask_first(nodemask);
8460 *sg = &per_cpu(sched_group_allnodes, group).sg;
8464 static void init_numa_sched_groups_power(struct sched_group *group_head)
8466 struct sched_group *sg = group_head;
8472 for_each_cpu(j, sched_group_cpus(sg)) {
8473 struct sched_domain *sd;
8475 sd = &per_cpu(phys_domains, j).sd;
8476 if (j != group_first_cpu(sd->groups)) {
8478 * Only add "power" once for each
8484 sg->cpu_power += sd->groups->cpu_power;
8487 } while (sg != group_head);
8490 static int build_numa_sched_groups(struct s_data *d,
8491 const struct cpumask *cpu_map, int num)
8493 struct sched_domain *sd;
8494 struct sched_group *sg, *prev;
8497 cpumask_clear(d->covered);
8498 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8499 if (cpumask_empty(d->nodemask)) {
8500 d->sched_group_nodes[num] = NULL;
8504 sched_domain_node_span(num, d->domainspan);
8505 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8507 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8510 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8514 d->sched_group_nodes[num] = sg;
8516 for_each_cpu(j, d->nodemask) {
8517 sd = &per_cpu(node_domains, j).sd;
8522 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8524 cpumask_or(d->covered, d->covered, d->nodemask);
8527 for (j = 0; j < nr_node_ids; j++) {
8528 n = (num + j) % nr_node_ids;
8529 cpumask_complement(d->notcovered, d->covered);
8530 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8531 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8532 if (cpumask_empty(d->tmpmask))
8534 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8535 if (cpumask_empty(d->tmpmask))
8537 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8541 "Can not alloc domain group for node %d\n", j);
8545 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8546 sg->next = prev->next;
8547 cpumask_or(d->covered, d->covered, d->tmpmask);
8554 #endif /* CONFIG_NUMA */
8557 /* Free memory allocated for various sched_group structures */
8558 static void free_sched_groups(const struct cpumask *cpu_map,
8559 struct cpumask *nodemask)
8563 for_each_cpu(cpu, cpu_map) {
8564 struct sched_group **sched_group_nodes
8565 = sched_group_nodes_bycpu[cpu];
8567 if (!sched_group_nodes)
8570 for (i = 0; i < nr_node_ids; i++) {
8571 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8573 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8574 if (cpumask_empty(nodemask))
8584 if (oldsg != sched_group_nodes[i])
8587 kfree(sched_group_nodes);
8588 sched_group_nodes_bycpu[cpu] = NULL;
8591 #else /* !CONFIG_NUMA */
8592 static void free_sched_groups(const struct cpumask *cpu_map,
8593 struct cpumask *nodemask)
8596 #endif /* CONFIG_NUMA */
8599 * Initialize sched groups cpu_power.
8601 * cpu_power indicates the capacity of sched group, which is used while
8602 * distributing the load between different sched groups in a sched domain.
8603 * Typically cpu_power for all the groups in a sched domain will be same unless
8604 * there are asymmetries in the topology. If there are asymmetries, group
8605 * having more cpu_power will pickup more load compared to the group having
8608 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8610 struct sched_domain *child;
8611 struct sched_group *group;
8615 WARN_ON(!sd || !sd->groups);
8617 if (cpu != group_first_cpu(sd->groups))
8622 sd->groups->cpu_power = 0;
8625 power = SCHED_LOAD_SCALE;
8626 weight = cpumask_weight(sched_domain_span(sd));
8628 * SMT siblings share the power of a single core.
8629 * Usually multiple threads get a better yield out of
8630 * that one core than a single thread would have,
8631 * reflect that in sd->smt_gain.
8633 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8634 power *= sd->smt_gain;
8636 power >>= SCHED_LOAD_SHIFT;
8638 sd->groups->cpu_power += power;
8643 * Add cpu_power of each child group to this groups cpu_power.
8645 group = child->groups;
8647 sd->groups->cpu_power += group->cpu_power;
8648 group = group->next;
8649 } while (group != child->groups);
8653 * Initializers for schedule domains
8654 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8657 #ifdef CONFIG_SCHED_DEBUG
8658 # define SD_INIT_NAME(sd, type) sd->name = #type
8660 # define SD_INIT_NAME(sd, type) do { } while (0)
8663 #define SD_INIT(sd, type) sd_init_##type(sd)
8665 #define SD_INIT_FUNC(type) \
8666 static noinline void sd_init_##type(struct sched_domain *sd) \
8668 memset(sd, 0, sizeof(*sd)); \
8669 *sd = SD_##type##_INIT; \
8670 sd->level = SD_LV_##type; \
8671 SD_INIT_NAME(sd, type); \
8676 SD_INIT_FUNC(ALLNODES)
8679 #ifdef CONFIG_SCHED_SMT
8680 SD_INIT_FUNC(SIBLING)
8682 #ifdef CONFIG_SCHED_MC
8686 static int default_relax_domain_level = -1;
8688 static int __init setup_relax_domain_level(char *str)
8692 val = simple_strtoul(str, NULL, 0);
8693 if (val < SD_LV_MAX)
8694 default_relax_domain_level = val;
8698 __setup("relax_domain_level=", setup_relax_domain_level);
8700 static void set_domain_attribute(struct sched_domain *sd,
8701 struct sched_domain_attr *attr)
8705 if (!attr || attr->relax_domain_level < 0) {
8706 if (default_relax_domain_level < 0)
8709 request = default_relax_domain_level;
8711 request = attr->relax_domain_level;
8712 if (request < sd->level) {
8713 /* turn off idle balance on this domain */
8714 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8716 /* turn on idle balance on this domain */
8717 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8721 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8722 const struct cpumask *cpu_map)
8725 case sa_sched_groups:
8726 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8727 d->sched_group_nodes = NULL;
8729 free_rootdomain(d->rd); /* fall through */
8731 free_cpumask_var(d->tmpmask); /* fall through */
8732 case sa_send_covered:
8733 free_cpumask_var(d->send_covered); /* fall through */
8734 case sa_this_core_map:
8735 free_cpumask_var(d->this_core_map); /* fall through */
8736 case sa_this_sibling_map:
8737 free_cpumask_var(d->this_sibling_map); /* fall through */
8739 free_cpumask_var(d->nodemask); /* fall through */
8740 case sa_sched_group_nodes:
8742 kfree(d->sched_group_nodes); /* fall through */
8744 free_cpumask_var(d->notcovered); /* fall through */
8746 free_cpumask_var(d->covered); /* fall through */
8748 free_cpumask_var(d->domainspan); /* fall through */
8755 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8756 const struct cpumask *cpu_map)
8759 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8761 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8762 return sa_domainspan;
8763 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8765 /* Allocate the per-node list of sched groups */
8766 d->sched_group_nodes = kcalloc(nr_node_ids,
8767 sizeof(struct sched_group *), GFP_KERNEL);
8768 if (!d->sched_group_nodes) {
8769 printk(KERN_WARNING "Can not alloc sched group node list\n");
8770 return sa_notcovered;
8772 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8774 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8775 return sa_sched_group_nodes;
8776 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8778 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8779 return sa_this_sibling_map;
8780 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8781 return sa_this_core_map;
8782 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8783 return sa_send_covered;
8784 d->rd = alloc_rootdomain();
8786 printk(KERN_WARNING "Cannot alloc root domain\n");
8789 return sa_rootdomain;
8792 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8793 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8795 struct sched_domain *sd = NULL;
8797 struct sched_domain *parent;
8800 if (cpumask_weight(cpu_map) >
8801 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8802 sd = &per_cpu(allnodes_domains, i).sd;
8803 SD_INIT(sd, ALLNODES);
8804 set_domain_attribute(sd, attr);
8805 cpumask_copy(sched_domain_span(sd), cpu_map);
8806 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8811 sd = &per_cpu(node_domains, i).sd;
8813 set_domain_attribute(sd, attr);
8814 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8815 sd->parent = parent;
8818 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8823 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8824 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8825 struct sched_domain *parent, int i)
8827 struct sched_domain *sd;
8828 sd = &per_cpu(phys_domains, i).sd;
8830 set_domain_attribute(sd, attr);
8831 cpumask_copy(sched_domain_span(sd), d->nodemask);
8832 sd->parent = parent;
8835 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8839 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8840 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8841 struct sched_domain *parent, int i)
8843 struct sched_domain *sd = parent;
8844 #ifdef CONFIG_SCHED_MC
8845 sd = &per_cpu(core_domains, i).sd;
8847 set_domain_attribute(sd, attr);
8848 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8849 sd->parent = parent;
8851 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8856 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8857 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8858 struct sched_domain *parent, int i)
8860 struct sched_domain *sd = parent;
8861 #ifdef CONFIG_SCHED_SMT
8862 sd = &per_cpu(cpu_domains, i).sd;
8863 SD_INIT(sd, SIBLING);
8864 set_domain_attribute(sd, attr);
8865 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8866 sd->parent = parent;
8868 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8873 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8874 const struct cpumask *cpu_map, int cpu)
8877 #ifdef CONFIG_SCHED_SMT
8878 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8879 cpumask_and(d->this_sibling_map, cpu_map,
8880 topology_thread_cpumask(cpu));
8881 if (cpu == cpumask_first(d->this_sibling_map))
8882 init_sched_build_groups(d->this_sibling_map, cpu_map,
8884 d->send_covered, d->tmpmask);
8887 #ifdef CONFIG_SCHED_MC
8888 case SD_LV_MC: /* set up multi-core groups */
8889 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8890 if (cpu == cpumask_first(d->this_core_map))
8891 init_sched_build_groups(d->this_core_map, cpu_map,
8893 d->send_covered, d->tmpmask);
8896 case SD_LV_CPU: /* set up physical groups */
8897 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8898 if (!cpumask_empty(d->nodemask))
8899 init_sched_build_groups(d->nodemask, cpu_map,
8901 d->send_covered, d->tmpmask);
8904 case SD_LV_ALLNODES:
8905 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8906 d->send_covered, d->tmpmask);
8915 * Build sched domains for a given set of cpus and attach the sched domains
8916 * to the individual cpus
8918 static int __build_sched_domains(const struct cpumask *cpu_map,
8919 struct sched_domain_attr *attr)
8921 enum s_alloc alloc_state = sa_none;
8923 struct sched_domain *sd;
8929 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8930 if (alloc_state != sa_rootdomain)
8932 alloc_state = sa_sched_groups;
8935 * Set up domains for cpus specified by the cpu_map.
8937 for_each_cpu(i, cpu_map) {
8938 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8941 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8942 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8943 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8944 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8947 for_each_cpu(i, cpu_map) {
8948 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8949 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8952 /* Set up physical groups */
8953 for (i = 0; i < nr_node_ids; i++)
8954 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8957 /* Set up node groups */
8959 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8961 for (i = 0; i < nr_node_ids; i++)
8962 if (build_numa_sched_groups(&d, cpu_map, i))
8966 /* Calculate CPU power for physical packages and nodes */
8967 #ifdef CONFIG_SCHED_SMT
8968 for_each_cpu(i, cpu_map) {
8969 sd = &per_cpu(cpu_domains, i).sd;
8970 init_sched_groups_power(i, sd);
8973 #ifdef CONFIG_SCHED_MC
8974 for_each_cpu(i, cpu_map) {
8975 sd = &per_cpu(core_domains, i).sd;
8976 init_sched_groups_power(i, sd);
8980 for_each_cpu(i, cpu_map) {
8981 sd = &per_cpu(phys_domains, i).sd;
8982 init_sched_groups_power(i, sd);
8986 for (i = 0; i < nr_node_ids; i++)
8987 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8989 if (d.sd_allnodes) {
8990 struct sched_group *sg;
8992 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8994 init_numa_sched_groups_power(sg);
8998 /* Attach the domains */
8999 for_each_cpu(i, cpu_map) {
9000 #ifdef CONFIG_SCHED_SMT
9001 sd = &per_cpu(cpu_domains, i).sd;
9002 #elif defined(CONFIG_SCHED_MC)
9003 sd = &per_cpu(core_domains, i).sd;
9005 sd = &per_cpu(phys_domains, i).sd;
9007 cpu_attach_domain(sd, d.rd, i);
9010 d.sched_group_nodes = NULL; /* don't free this we still need it */
9011 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
9015 __free_domain_allocs(&d, alloc_state, cpu_map);
9019 static int build_sched_domains(const struct cpumask *cpu_map)
9021 return __build_sched_domains(cpu_map, NULL);
9024 static cpumask_var_t *doms_cur; /* current sched domains */
9025 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
9026 static struct sched_domain_attr *dattr_cur;
9027 /* attribues of custom domains in 'doms_cur' */
9030 * Special case: If a kmalloc of a doms_cur partition (array of
9031 * cpumask) fails, then fallback to a single sched domain,
9032 * as determined by the single cpumask fallback_doms.
9034 static cpumask_var_t fallback_doms;
9037 * arch_update_cpu_topology lets virtualized architectures update the
9038 * cpu core maps. It is supposed to return 1 if the topology changed
9039 * or 0 if it stayed the same.
9041 int __attribute__((weak)) arch_update_cpu_topology(void)
9046 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
9049 cpumask_var_t *doms;
9051 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
9054 for (i = 0; i < ndoms; i++) {
9055 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
9056 free_sched_domains(doms, i);
9063 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
9066 for (i = 0; i < ndoms; i++)
9067 free_cpumask_var(doms[i]);
9072 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9073 * For now this just excludes isolated cpus, but could be used to
9074 * exclude other special cases in the future.
9076 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9080 arch_update_cpu_topology();
9082 doms_cur = alloc_sched_domains(ndoms_cur);
9084 doms_cur = &fallback_doms;
9085 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
9087 err = build_sched_domains(doms_cur[0]);
9088 register_sched_domain_sysctl();
9093 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9094 struct cpumask *tmpmask)
9096 free_sched_groups(cpu_map, tmpmask);
9100 * Detach sched domains from a group of cpus specified in cpu_map
9101 * These cpus will now be attached to the NULL domain
9103 static void detach_destroy_domains(const struct cpumask *cpu_map)
9105 /* Save because hotplug lock held. */
9106 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9109 for_each_cpu(i, cpu_map)
9110 cpu_attach_domain(NULL, &def_root_domain, i);
9111 synchronize_sched();
9112 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9115 /* handle null as "default" */
9116 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9117 struct sched_domain_attr *new, int idx_new)
9119 struct sched_domain_attr tmp;
9126 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9127 new ? (new + idx_new) : &tmp,
9128 sizeof(struct sched_domain_attr));
9132 * Partition sched domains as specified by the 'ndoms_new'
9133 * cpumasks in the array doms_new[] of cpumasks. This compares
9134 * doms_new[] to the current sched domain partitioning, doms_cur[].
9135 * It destroys each deleted domain and builds each new domain.
9137 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9138 * The masks don't intersect (don't overlap.) We should setup one
9139 * sched domain for each mask. CPUs not in any of the cpumasks will
9140 * not be load balanced. If the same cpumask appears both in the
9141 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9144 * The passed in 'doms_new' should be allocated using
9145 * alloc_sched_domains. This routine takes ownership of it and will
9146 * free_sched_domains it when done with it. If the caller failed the
9147 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9148 * and partition_sched_domains() will fallback to the single partition
9149 * 'fallback_doms', it also forces the domains to be rebuilt.
9151 * If doms_new == NULL it will be replaced with cpu_online_mask.
9152 * ndoms_new == 0 is a special case for destroying existing domains,
9153 * and it will not create the default domain.
9155 * Call with hotplug lock held
9157 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9158 struct sched_domain_attr *dattr_new)
9163 mutex_lock(&sched_domains_mutex);
9165 /* always unregister in case we don't destroy any domains */
9166 unregister_sched_domain_sysctl();
9168 /* Let architecture update cpu core mappings. */
9169 new_topology = arch_update_cpu_topology();
9171 n = doms_new ? ndoms_new : 0;
9173 /* Destroy deleted domains */
9174 for (i = 0; i < ndoms_cur; i++) {
9175 for (j = 0; j < n && !new_topology; j++) {
9176 if (cpumask_equal(doms_cur[i], doms_new[j])
9177 && dattrs_equal(dattr_cur, i, dattr_new, j))
9180 /* no match - a current sched domain not in new doms_new[] */
9181 detach_destroy_domains(doms_cur[i]);
9186 if (doms_new == NULL) {
9188 doms_new = &fallback_doms;
9189 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9190 WARN_ON_ONCE(dattr_new);
9193 /* Build new domains */
9194 for (i = 0; i < ndoms_new; i++) {
9195 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9196 if (cpumask_equal(doms_new[i], doms_cur[j])
9197 && dattrs_equal(dattr_new, i, dattr_cur, j))
9200 /* no match - add a new doms_new */
9201 __build_sched_domains(doms_new[i],
9202 dattr_new ? dattr_new + i : NULL);
9207 /* Remember the new sched domains */
9208 if (doms_cur != &fallback_doms)
9209 free_sched_domains(doms_cur, ndoms_cur);
9210 kfree(dattr_cur); /* kfree(NULL) is safe */
9211 doms_cur = doms_new;
9212 dattr_cur = dattr_new;
9213 ndoms_cur = ndoms_new;
9215 register_sched_domain_sysctl();
9217 mutex_unlock(&sched_domains_mutex);
9220 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9221 static void arch_reinit_sched_domains(void)
9225 /* Destroy domains first to force the rebuild */
9226 partition_sched_domains(0, NULL, NULL);
9228 rebuild_sched_domains();
9232 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9234 unsigned int level = 0;
9236 if (sscanf(buf, "%u", &level) != 1)
9240 * level is always be positive so don't check for
9241 * level < POWERSAVINGS_BALANCE_NONE which is 0
9242 * What happens on 0 or 1 byte write,
9243 * need to check for count as well?
9246 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9250 sched_smt_power_savings = level;
9252 sched_mc_power_savings = level;
9254 arch_reinit_sched_domains();
9259 #ifdef CONFIG_SCHED_MC
9260 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9263 return sprintf(page, "%u\n", sched_mc_power_savings);
9265 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9266 const char *buf, size_t count)
9268 return sched_power_savings_store(buf, count, 0);
9270 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9271 sched_mc_power_savings_show,
9272 sched_mc_power_savings_store);
9275 #ifdef CONFIG_SCHED_SMT
9276 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9279 return sprintf(page, "%u\n", sched_smt_power_savings);
9281 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9282 const char *buf, size_t count)
9284 return sched_power_savings_store(buf, count, 1);
9286 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9287 sched_smt_power_savings_show,
9288 sched_smt_power_savings_store);
9291 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9295 #ifdef CONFIG_SCHED_SMT
9297 err = sysfs_create_file(&cls->kset.kobj,
9298 &attr_sched_smt_power_savings.attr);
9300 #ifdef CONFIG_SCHED_MC
9301 if (!err && mc_capable())
9302 err = sysfs_create_file(&cls->kset.kobj,
9303 &attr_sched_mc_power_savings.attr);
9307 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9309 #ifndef CONFIG_CPUSETS
9311 * Add online and remove offline CPUs from the scheduler domains.
9312 * When cpusets are enabled they take over this function.
9314 static int update_sched_domains(struct notifier_block *nfb,
9315 unsigned long action, void *hcpu)
9319 case CPU_ONLINE_FROZEN:
9320 case CPU_DOWN_PREPARE:
9321 case CPU_DOWN_PREPARE_FROZEN:
9322 case CPU_DOWN_FAILED:
9323 case CPU_DOWN_FAILED_FROZEN:
9324 partition_sched_domains(1, NULL, NULL);
9333 static int update_runtime(struct notifier_block *nfb,
9334 unsigned long action, void *hcpu)
9336 int cpu = (int)(long)hcpu;
9339 case CPU_DOWN_PREPARE:
9340 case CPU_DOWN_PREPARE_FROZEN:
9341 disable_runtime(cpu_rq(cpu));
9344 case CPU_DOWN_FAILED:
9345 case CPU_DOWN_FAILED_FROZEN:
9347 case CPU_ONLINE_FROZEN:
9348 enable_runtime(cpu_rq(cpu));
9356 void __init sched_init_smp(void)
9358 cpumask_var_t non_isolated_cpus;
9360 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9361 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9363 #if defined(CONFIG_NUMA)
9364 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9366 BUG_ON(sched_group_nodes_bycpu == NULL);
9369 mutex_lock(&sched_domains_mutex);
9370 arch_init_sched_domains(cpu_active_mask);
9371 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9372 if (cpumask_empty(non_isolated_cpus))
9373 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9374 mutex_unlock(&sched_domains_mutex);
9377 #ifndef CONFIG_CPUSETS
9378 /* XXX: Theoretical race here - CPU may be hotplugged now */
9379 hotcpu_notifier(update_sched_domains, 0);
9382 /* RT runtime code needs to handle some hotplug events */
9383 hotcpu_notifier(update_runtime, 0);
9387 /* Move init over to a non-isolated CPU */
9388 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9390 sched_init_granularity();
9391 free_cpumask_var(non_isolated_cpus);
9393 init_sched_rt_class();
9396 void __init sched_init_smp(void)
9398 sched_init_granularity();
9400 #endif /* CONFIG_SMP */
9402 const_debug unsigned int sysctl_timer_migration = 1;
9404 int in_sched_functions(unsigned long addr)
9406 return in_lock_functions(addr) ||
9407 (addr >= (unsigned long)__sched_text_start
9408 && addr < (unsigned long)__sched_text_end);
9411 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9413 cfs_rq->tasks_timeline = RB_ROOT;
9414 INIT_LIST_HEAD(&cfs_rq->tasks);
9415 #ifdef CONFIG_FAIR_GROUP_SCHED
9418 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9421 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9423 struct rt_prio_array *array;
9426 array = &rt_rq->active;
9427 for (i = 0; i < MAX_RT_PRIO; i++) {
9428 INIT_LIST_HEAD(array->queue + i);
9429 __clear_bit(i, array->bitmap);
9431 /* delimiter for bitsearch: */
9432 __set_bit(MAX_RT_PRIO, array->bitmap);
9434 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9435 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9437 rt_rq->highest_prio.next = MAX_RT_PRIO;
9441 rt_rq->rt_nr_migratory = 0;
9442 rt_rq->overloaded = 0;
9443 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
9447 rt_rq->rt_throttled = 0;
9448 rt_rq->rt_runtime = 0;
9449 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
9451 #ifdef CONFIG_RT_GROUP_SCHED
9452 rt_rq->rt_nr_boosted = 0;
9457 #ifdef CONFIG_FAIR_GROUP_SCHED
9458 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9459 struct sched_entity *se, int cpu, int add,
9460 struct sched_entity *parent)
9462 struct rq *rq = cpu_rq(cpu);
9463 tg->cfs_rq[cpu] = cfs_rq;
9464 init_cfs_rq(cfs_rq, rq);
9467 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9470 /* se could be NULL for init_task_group */
9475 se->cfs_rq = &rq->cfs;
9477 se->cfs_rq = parent->my_q;
9480 se->load.weight = tg->shares;
9481 se->load.inv_weight = 0;
9482 se->parent = parent;
9486 #ifdef CONFIG_RT_GROUP_SCHED
9487 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9488 struct sched_rt_entity *rt_se, int cpu, int add,
9489 struct sched_rt_entity *parent)
9491 struct rq *rq = cpu_rq(cpu);
9493 tg->rt_rq[cpu] = rt_rq;
9494 init_rt_rq(rt_rq, rq);
9496 rt_rq->rt_se = rt_se;
9497 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9499 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9501 tg->rt_se[cpu] = rt_se;
9506 rt_se->rt_rq = &rq->rt;
9508 rt_se->rt_rq = parent->my_q;
9510 rt_se->my_q = rt_rq;
9511 rt_se->parent = parent;
9512 INIT_LIST_HEAD(&rt_se->run_list);
9516 void __init sched_init(void)
9519 unsigned long alloc_size = 0, ptr;
9521 #ifdef CONFIG_FAIR_GROUP_SCHED
9522 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9524 #ifdef CONFIG_RT_GROUP_SCHED
9525 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9527 #ifdef CONFIG_USER_SCHED
9530 #ifdef CONFIG_CPUMASK_OFFSTACK
9531 alloc_size += num_possible_cpus() * cpumask_size();
9534 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9536 #ifdef CONFIG_FAIR_GROUP_SCHED
9537 init_task_group.se = (struct sched_entity **)ptr;
9538 ptr += nr_cpu_ids * sizeof(void **);
9540 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9541 ptr += nr_cpu_ids * sizeof(void **);
9543 #ifdef CONFIG_USER_SCHED
9544 root_task_group.se = (struct sched_entity **)ptr;
9545 ptr += nr_cpu_ids * sizeof(void **);
9547 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9548 ptr += nr_cpu_ids * sizeof(void **);
9549 #endif /* CONFIG_USER_SCHED */
9550 #endif /* CONFIG_FAIR_GROUP_SCHED */
9551 #ifdef CONFIG_RT_GROUP_SCHED
9552 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9553 ptr += nr_cpu_ids * sizeof(void **);
9555 init_task_group.rt_rq = (struct rt_rq **)ptr;
9556 ptr += nr_cpu_ids * sizeof(void **);
9558 #ifdef CONFIG_USER_SCHED
9559 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9560 ptr += nr_cpu_ids * sizeof(void **);
9562 root_task_group.rt_rq = (struct rt_rq **)ptr;
9563 ptr += nr_cpu_ids * sizeof(void **);
9564 #endif /* CONFIG_USER_SCHED */
9565 #endif /* CONFIG_RT_GROUP_SCHED */
9566 #ifdef CONFIG_CPUMASK_OFFSTACK
9567 for_each_possible_cpu(i) {
9568 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9569 ptr += cpumask_size();
9571 #endif /* CONFIG_CPUMASK_OFFSTACK */
9575 init_defrootdomain();
9578 init_rt_bandwidth(&def_rt_bandwidth,
9579 global_rt_period(), global_rt_runtime());
9581 #ifdef CONFIG_RT_GROUP_SCHED
9582 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9583 global_rt_period(), global_rt_runtime());
9584 #ifdef CONFIG_USER_SCHED
9585 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9586 global_rt_period(), RUNTIME_INF);
9587 #endif /* CONFIG_USER_SCHED */
9588 #endif /* CONFIG_RT_GROUP_SCHED */
9590 #ifdef CONFIG_GROUP_SCHED
9591 list_add(&init_task_group.list, &task_groups);
9592 INIT_LIST_HEAD(&init_task_group.children);
9594 #ifdef CONFIG_USER_SCHED
9595 INIT_LIST_HEAD(&root_task_group.children);
9596 init_task_group.parent = &root_task_group;
9597 list_add(&init_task_group.siblings, &root_task_group.children);
9598 #endif /* CONFIG_USER_SCHED */
9599 #endif /* CONFIG_GROUP_SCHED */
9601 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9602 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9603 __alignof__(unsigned long));
9605 for_each_possible_cpu(i) {
9609 raw_spin_lock_init(&rq->lock);
9611 rq->calc_load_active = 0;
9612 rq->calc_load_update = jiffies + LOAD_FREQ;
9613 init_cfs_rq(&rq->cfs, rq);
9614 init_rt_rq(&rq->rt, rq);
9615 #ifdef CONFIG_FAIR_GROUP_SCHED
9616 init_task_group.shares = init_task_group_load;
9617 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9618 #ifdef CONFIG_CGROUP_SCHED
9620 * How much cpu bandwidth does init_task_group get?
9622 * In case of task-groups formed thr' the cgroup filesystem, it
9623 * gets 100% of the cpu resources in the system. This overall
9624 * system cpu resource is divided among the tasks of
9625 * init_task_group and its child task-groups in a fair manner,
9626 * based on each entity's (task or task-group's) weight
9627 * (se->load.weight).
9629 * In other words, if init_task_group has 10 tasks of weight
9630 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9631 * then A0's share of the cpu resource is:
9633 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9635 * We achieve this by letting init_task_group's tasks sit
9636 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9638 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9639 #elif defined CONFIG_USER_SCHED
9640 root_task_group.shares = NICE_0_LOAD;
9641 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9643 * In case of task-groups formed thr' the user id of tasks,
9644 * init_task_group represents tasks belonging to root user.
9645 * Hence it forms a sibling of all subsequent groups formed.
9646 * In this case, init_task_group gets only a fraction of overall
9647 * system cpu resource, based on the weight assigned to root
9648 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9649 * by letting tasks of init_task_group sit in a separate cfs_rq
9650 * (init_tg_cfs_rq) and having one entity represent this group of
9651 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9653 init_tg_cfs_entry(&init_task_group,
9654 &per_cpu(init_tg_cfs_rq, i),
9655 &per_cpu(init_sched_entity, i), i, 1,
9656 root_task_group.se[i]);
9659 #endif /* CONFIG_FAIR_GROUP_SCHED */
9661 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9662 #ifdef CONFIG_RT_GROUP_SCHED
9663 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9664 #ifdef CONFIG_CGROUP_SCHED
9665 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9666 #elif defined CONFIG_USER_SCHED
9667 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9668 init_tg_rt_entry(&init_task_group,
9669 &per_cpu(init_rt_rq_var, i),
9670 &per_cpu(init_sched_rt_entity, i), i, 1,
9671 root_task_group.rt_se[i]);
9675 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9676 rq->cpu_load[j] = 0;
9680 rq->post_schedule = 0;
9681 rq->active_balance = 0;
9682 rq->next_balance = jiffies;
9686 rq->migration_thread = NULL;
9688 rq->avg_idle = 2*sysctl_sched_migration_cost;
9689 INIT_LIST_HEAD(&rq->migration_queue);
9690 rq_attach_root(rq, &def_root_domain);
9693 atomic_set(&rq->nr_iowait, 0);
9696 set_load_weight(&init_task);
9698 #ifdef CONFIG_PREEMPT_NOTIFIERS
9699 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9703 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9706 #ifdef CONFIG_RT_MUTEXES
9707 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
9711 * The boot idle thread does lazy MMU switching as well:
9713 atomic_inc(&init_mm.mm_count);
9714 enter_lazy_tlb(&init_mm, current);
9717 * Make us the idle thread. Technically, schedule() should not be
9718 * called from this thread, however somewhere below it might be,
9719 * but because we are the idle thread, we just pick up running again
9720 * when this runqueue becomes "idle".
9722 init_idle(current, smp_processor_id());
9724 calc_load_update = jiffies + LOAD_FREQ;
9727 * During early bootup we pretend to be a normal task:
9729 current->sched_class = &fair_sched_class;
9731 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9732 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9735 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9736 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9738 /* May be allocated at isolcpus cmdline parse time */
9739 if (cpu_isolated_map == NULL)
9740 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9745 scheduler_running = 1;
9748 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9749 static inline int preempt_count_equals(int preempt_offset)
9751 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
9753 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9756 void __might_sleep(char *file, int line, int preempt_offset)
9759 static unsigned long prev_jiffy; /* ratelimiting */
9761 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9762 system_state != SYSTEM_RUNNING || oops_in_progress)
9764 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9766 prev_jiffy = jiffies;
9769 "BUG: sleeping function called from invalid context at %s:%d\n",
9772 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9773 in_atomic(), irqs_disabled(),
9774 current->pid, current->comm);
9776 debug_show_held_locks(current);
9777 if (irqs_disabled())
9778 print_irqtrace_events(current);
9782 EXPORT_SYMBOL(__might_sleep);
9785 #ifdef CONFIG_MAGIC_SYSRQ
9786 static void normalize_task(struct rq *rq, struct task_struct *p)
9790 update_rq_clock(rq);
9791 on_rq = p->se.on_rq;
9793 deactivate_task(rq, p, 0);
9794 __setscheduler(rq, p, SCHED_NORMAL, 0);
9796 activate_task(rq, p, 0);
9797 resched_task(rq->curr);
9801 void normalize_rt_tasks(void)
9803 struct task_struct *g, *p;
9804 unsigned long flags;
9807 read_lock_irqsave(&tasklist_lock, flags);
9808 do_each_thread(g, p) {
9810 * Only normalize user tasks:
9815 p->se.exec_start = 0;
9816 #ifdef CONFIG_SCHEDSTATS
9817 p->se.wait_start = 0;
9818 p->se.sleep_start = 0;
9819 p->se.block_start = 0;
9824 * Renice negative nice level userspace
9827 if (TASK_NICE(p) < 0 && p->mm)
9828 set_user_nice(p, 0);
9832 raw_spin_lock(&p->pi_lock);
9833 rq = __task_rq_lock(p);
9835 normalize_task(rq, p);
9837 __task_rq_unlock(rq);
9838 raw_spin_unlock(&p->pi_lock);
9839 } while_each_thread(g, p);
9841 read_unlock_irqrestore(&tasklist_lock, flags);
9844 #endif /* CONFIG_MAGIC_SYSRQ */
9848 * These functions are only useful for the IA64 MCA handling.
9850 * They can only be called when the whole system has been
9851 * stopped - every CPU needs to be quiescent, and no scheduling
9852 * activity can take place. Using them for anything else would
9853 * be a serious bug, and as a result, they aren't even visible
9854 * under any other configuration.
9858 * curr_task - return the current task for a given cpu.
9859 * @cpu: the processor in question.
9861 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9863 struct task_struct *curr_task(int cpu)
9865 return cpu_curr(cpu);
9869 * set_curr_task - set the current task for a given cpu.
9870 * @cpu: the processor in question.
9871 * @p: the task pointer to set.
9873 * Description: This function must only be used when non-maskable interrupts
9874 * are serviced on a separate stack. It allows the architecture to switch the
9875 * notion of the current task on a cpu in a non-blocking manner. This function
9876 * must be called with all CPU's synchronized, and interrupts disabled, the
9877 * and caller must save the original value of the current task (see
9878 * curr_task() above) and restore that value before reenabling interrupts and
9879 * re-starting the system.
9881 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9883 void set_curr_task(int cpu, struct task_struct *p)
9890 #ifdef CONFIG_FAIR_GROUP_SCHED
9891 static void free_fair_sched_group(struct task_group *tg)
9895 for_each_possible_cpu(i) {
9897 kfree(tg->cfs_rq[i]);
9907 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9909 struct cfs_rq *cfs_rq;
9910 struct sched_entity *se;
9914 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9917 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9921 tg->shares = NICE_0_LOAD;
9923 for_each_possible_cpu(i) {
9926 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9927 GFP_KERNEL, cpu_to_node(i));
9931 se = kzalloc_node(sizeof(struct sched_entity),
9932 GFP_KERNEL, cpu_to_node(i));
9936 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9947 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9949 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9950 &cpu_rq(cpu)->leaf_cfs_rq_list);
9953 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9955 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9957 #else /* !CONFG_FAIR_GROUP_SCHED */
9958 static inline void free_fair_sched_group(struct task_group *tg)
9963 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9968 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9972 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9975 #endif /* CONFIG_FAIR_GROUP_SCHED */
9977 #ifdef CONFIG_RT_GROUP_SCHED
9978 static void free_rt_sched_group(struct task_group *tg)
9982 destroy_rt_bandwidth(&tg->rt_bandwidth);
9984 for_each_possible_cpu(i) {
9986 kfree(tg->rt_rq[i]);
9988 kfree(tg->rt_se[i]);
9996 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9998 struct rt_rq *rt_rq;
9999 struct sched_rt_entity *rt_se;
10003 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
10006 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
10010 init_rt_bandwidth(&tg->rt_bandwidth,
10011 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
10013 for_each_possible_cpu(i) {
10016 rt_rq = kzalloc_node(sizeof(struct rt_rq),
10017 GFP_KERNEL, cpu_to_node(i));
10021 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
10022 GFP_KERNEL, cpu_to_node(i));
10026 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
10037 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10039 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
10040 &cpu_rq(cpu)->leaf_rt_rq_list);
10043 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10045 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
10047 #else /* !CONFIG_RT_GROUP_SCHED */
10048 static inline void free_rt_sched_group(struct task_group *tg)
10053 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
10058 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10062 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10065 #endif /* CONFIG_RT_GROUP_SCHED */
10067 #ifdef CONFIG_GROUP_SCHED
10068 static void free_sched_group(struct task_group *tg)
10070 free_fair_sched_group(tg);
10071 free_rt_sched_group(tg);
10075 /* allocate runqueue etc for a new task group */
10076 struct task_group *sched_create_group(struct task_group *parent)
10078 struct task_group *tg;
10079 unsigned long flags;
10082 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10084 return ERR_PTR(-ENOMEM);
10086 if (!alloc_fair_sched_group(tg, parent))
10089 if (!alloc_rt_sched_group(tg, parent))
10092 spin_lock_irqsave(&task_group_lock, flags);
10093 for_each_possible_cpu(i) {
10094 register_fair_sched_group(tg, i);
10095 register_rt_sched_group(tg, i);
10097 list_add_rcu(&tg->list, &task_groups);
10099 WARN_ON(!parent); /* root should already exist */
10101 tg->parent = parent;
10102 INIT_LIST_HEAD(&tg->children);
10103 list_add_rcu(&tg->siblings, &parent->children);
10104 spin_unlock_irqrestore(&task_group_lock, flags);
10109 free_sched_group(tg);
10110 return ERR_PTR(-ENOMEM);
10113 /* rcu callback to free various structures associated with a task group */
10114 static void free_sched_group_rcu(struct rcu_head *rhp)
10116 /* now it should be safe to free those cfs_rqs */
10117 free_sched_group(container_of(rhp, struct task_group, rcu));
10120 /* Destroy runqueue etc associated with a task group */
10121 void sched_destroy_group(struct task_group *tg)
10123 unsigned long flags;
10126 spin_lock_irqsave(&task_group_lock, flags);
10127 for_each_possible_cpu(i) {
10128 unregister_fair_sched_group(tg, i);
10129 unregister_rt_sched_group(tg, i);
10131 list_del_rcu(&tg->list);
10132 list_del_rcu(&tg->siblings);
10133 spin_unlock_irqrestore(&task_group_lock, flags);
10135 /* wait for possible concurrent references to cfs_rqs complete */
10136 call_rcu(&tg->rcu, free_sched_group_rcu);
10139 /* change task's runqueue when it moves between groups.
10140 * The caller of this function should have put the task in its new group
10141 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10142 * reflect its new group.
10144 void sched_move_task(struct task_struct *tsk)
10146 int on_rq, running;
10147 unsigned long flags;
10150 rq = task_rq_lock(tsk, &flags);
10152 update_rq_clock(rq);
10154 running = task_current(rq, tsk);
10155 on_rq = tsk->se.on_rq;
10158 dequeue_task(rq, tsk, 0);
10159 if (unlikely(running))
10160 tsk->sched_class->put_prev_task(rq, tsk);
10162 set_task_rq(tsk, task_cpu(tsk));
10164 #ifdef CONFIG_FAIR_GROUP_SCHED
10165 if (tsk->sched_class->moved_group)
10166 tsk->sched_class->moved_group(tsk, on_rq);
10169 if (unlikely(running))
10170 tsk->sched_class->set_curr_task(rq);
10172 enqueue_task(rq, tsk, 0);
10174 task_rq_unlock(rq, &flags);
10176 #endif /* CONFIG_GROUP_SCHED */
10178 #ifdef CONFIG_FAIR_GROUP_SCHED
10179 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10181 struct cfs_rq *cfs_rq = se->cfs_rq;
10186 dequeue_entity(cfs_rq, se, 0);
10188 se->load.weight = shares;
10189 se->load.inv_weight = 0;
10192 enqueue_entity(cfs_rq, se, 0);
10195 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10197 struct cfs_rq *cfs_rq = se->cfs_rq;
10198 struct rq *rq = cfs_rq->rq;
10199 unsigned long flags;
10201 raw_spin_lock_irqsave(&rq->lock, flags);
10202 __set_se_shares(se, shares);
10203 raw_spin_unlock_irqrestore(&rq->lock, flags);
10206 static DEFINE_MUTEX(shares_mutex);
10208 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10211 unsigned long flags;
10214 * We can't change the weight of the root cgroup.
10219 if (shares < MIN_SHARES)
10220 shares = MIN_SHARES;
10221 else if (shares > MAX_SHARES)
10222 shares = MAX_SHARES;
10224 mutex_lock(&shares_mutex);
10225 if (tg->shares == shares)
10228 spin_lock_irqsave(&task_group_lock, flags);
10229 for_each_possible_cpu(i)
10230 unregister_fair_sched_group(tg, i);
10231 list_del_rcu(&tg->siblings);
10232 spin_unlock_irqrestore(&task_group_lock, flags);
10234 /* wait for any ongoing reference to this group to finish */
10235 synchronize_sched();
10238 * Now we are free to modify the group's share on each cpu
10239 * w/o tripping rebalance_share or load_balance_fair.
10241 tg->shares = shares;
10242 for_each_possible_cpu(i) {
10244 * force a rebalance
10246 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10247 set_se_shares(tg->se[i], shares);
10251 * Enable load balance activity on this group, by inserting it back on
10252 * each cpu's rq->leaf_cfs_rq_list.
10254 spin_lock_irqsave(&task_group_lock, flags);
10255 for_each_possible_cpu(i)
10256 register_fair_sched_group(tg, i);
10257 list_add_rcu(&tg->siblings, &tg->parent->children);
10258 spin_unlock_irqrestore(&task_group_lock, flags);
10260 mutex_unlock(&shares_mutex);
10264 unsigned long sched_group_shares(struct task_group *tg)
10270 #ifdef CONFIG_RT_GROUP_SCHED
10272 * Ensure that the real time constraints are schedulable.
10274 static DEFINE_MUTEX(rt_constraints_mutex);
10276 static unsigned long to_ratio(u64 period, u64 runtime)
10278 if (runtime == RUNTIME_INF)
10281 return div64_u64(runtime << 20, period);
10284 /* Must be called with tasklist_lock held */
10285 static inline int tg_has_rt_tasks(struct task_group *tg)
10287 struct task_struct *g, *p;
10289 do_each_thread(g, p) {
10290 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10292 } while_each_thread(g, p);
10297 struct rt_schedulable_data {
10298 struct task_group *tg;
10303 static int tg_schedulable(struct task_group *tg, void *data)
10305 struct rt_schedulable_data *d = data;
10306 struct task_group *child;
10307 unsigned long total, sum = 0;
10308 u64 period, runtime;
10310 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10311 runtime = tg->rt_bandwidth.rt_runtime;
10314 period = d->rt_period;
10315 runtime = d->rt_runtime;
10318 #ifdef CONFIG_USER_SCHED
10319 if (tg == &root_task_group) {
10320 period = global_rt_period();
10321 runtime = global_rt_runtime();
10326 * Cannot have more runtime than the period.
10328 if (runtime > period && runtime != RUNTIME_INF)
10332 * Ensure we don't starve existing RT tasks.
10334 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10337 total = to_ratio(period, runtime);
10340 * Nobody can have more than the global setting allows.
10342 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10346 * The sum of our children's runtime should not exceed our own.
10348 list_for_each_entry_rcu(child, &tg->children, siblings) {
10349 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10350 runtime = child->rt_bandwidth.rt_runtime;
10352 if (child == d->tg) {
10353 period = d->rt_period;
10354 runtime = d->rt_runtime;
10357 sum += to_ratio(period, runtime);
10366 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10368 struct rt_schedulable_data data = {
10370 .rt_period = period,
10371 .rt_runtime = runtime,
10374 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10377 static int tg_set_bandwidth(struct task_group *tg,
10378 u64 rt_period, u64 rt_runtime)
10382 mutex_lock(&rt_constraints_mutex);
10383 read_lock(&tasklist_lock);
10384 err = __rt_schedulable(tg, rt_period, rt_runtime);
10388 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10389 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10390 tg->rt_bandwidth.rt_runtime = rt_runtime;
10392 for_each_possible_cpu(i) {
10393 struct rt_rq *rt_rq = tg->rt_rq[i];
10395 raw_spin_lock(&rt_rq->rt_runtime_lock);
10396 rt_rq->rt_runtime = rt_runtime;
10397 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10399 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10401 read_unlock(&tasklist_lock);
10402 mutex_unlock(&rt_constraints_mutex);
10407 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10409 u64 rt_runtime, rt_period;
10411 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10412 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10413 if (rt_runtime_us < 0)
10414 rt_runtime = RUNTIME_INF;
10416 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10419 long sched_group_rt_runtime(struct task_group *tg)
10423 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10426 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10427 do_div(rt_runtime_us, NSEC_PER_USEC);
10428 return rt_runtime_us;
10431 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10433 u64 rt_runtime, rt_period;
10435 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10436 rt_runtime = tg->rt_bandwidth.rt_runtime;
10438 if (rt_period == 0)
10441 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10444 long sched_group_rt_period(struct task_group *tg)
10448 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10449 do_div(rt_period_us, NSEC_PER_USEC);
10450 return rt_period_us;
10453 static int sched_rt_global_constraints(void)
10455 u64 runtime, period;
10458 if (sysctl_sched_rt_period <= 0)
10461 runtime = global_rt_runtime();
10462 period = global_rt_period();
10465 * Sanity check on the sysctl variables.
10467 if (runtime > period && runtime != RUNTIME_INF)
10470 mutex_lock(&rt_constraints_mutex);
10471 read_lock(&tasklist_lock);
10472 ret = __rt_schedulable(NULL, 0, 0);
10473 read_unlock(&tasklist_lock);
10474 mutex_unlock(&rt_constraints_mutex);
10479 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10481 /* Don't accept realtime tasks when there is no way for them to run */
10482 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10488 #else /* !CONFIG_RT_GROUP_SCHED */
10489 static int sched_rt_global_constraints(void)
10491 unsigned long flags;
10494 if (sysctl_sched_rt_period <= 0)
10498 * There's always some RT tasks in the root group
10499 * -- migration, kstopmachine etc..
10501 if (sysctl_sched_rt_runtime == 0)
10504 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10505 for_each_possible_cpu(i) {
10506 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10508 raw_spin_lock(&rt_rq->rt_runtime_lock);
10509 rt_rq->rt_runtime = global_rt_runtime();
10510 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10512 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10516 #endif /* CONFIG_RT_GROUP_SCHED */
10518 int sched_rt_handler(struct ctl_table *table, int write,
10519 void __user *buffer, size_t *lenp,
10523 int old_period, old_runtime;
10524 static DEFINE_MUTEX(mutex);
10526 mutex_lock(&mutex);
10527 old_period = sysctl_sched_rt_period;
10528 old_runtime = sysctl_sched_rt_runtime;
10530 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10532 if (!ret && write) {
10533 ret = sched_rt_global_constraints();
10535 sysctl_sched_rt_period = old_period;
10536 sysctl_sched_rt_runtime = old_runtime;
10538 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10539 def_rt_bandwidth.rt_period =
10540 ns_to_ktime(global_rt_period());
10543 mutex_unlock(&mutex);
10548 #ifdef CONFIG_CGROUP_SCHED
10550 /* return corresponding task_group object of a cgroup */
10551 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10553 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10554 struct task_group, css);
10557 static struct cgroup_subsys_state *
10558 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10560 struct task_group *tg, *parent;
10562 if (!cgrp->parent) {
10563 /* This is early initialization for the top cgroup */
10564 return &init_task_group.css;
10567 parent = cgroup_tg(cgrp->parent);
10568 tg = sched_create_group(parent);
10570 return ERR_PTR(-ENOMEM);
10576 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10578 struct task_group *tg = cgroup_tg(cgrp);
10580 sched_destroy_group(tg);
10584 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10586 #ifdef CONFIG_RT_GROUP_SCHED
10587 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10590 /* We don't support RT-tasks being in separate groups */
10591 if (tsk->sched_class != &fair_sched_class)
10598 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10599 struct task_struct *tsk, bool threadgroup)
10601 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10605 struct task_struct *c;
10607 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10608 retval = cpu_cgroup_can_attach_task(cgrp, c);
10620 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10621 struct cgroup *old_cont, struct task_struct *tsk,
10624 sched_move_task(tsk);
10626 struct task_struct *c;
10628 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10629 sched_move_task(c);
10635 #ifdef CONFIG_FAIR_GROUP_SCHED
10636 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10639 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10642 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10644 struct task_group *tg = cgroup_tg(cgrp);
10646 return (u64) tg->shares;
10648 #endif /* CONFIG_FAIR_GROUP_SCHED */
10650 #ifdef CONFIG_RT_GROUP_SCHED
10651 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10654 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10657 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10659 return sched_group_rt_runtime(cgroup_tg(cgrp));
10662 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10665 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10668 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10670 return sched_group_rt_period(cgroup_tg(cgrp));
10672 #endif /* CONFIG_RT_GROUP_SCHED */
10674 static struct cftype cpu_files[] = {
10675 #ifdef CONFIG_FAIR_GROUP_SCHED
10678 .read_u64 = cpu_shares_read_u64,
10679 .write_u64 = cpu_shares_write_u64,
10682 #ifdef CONFIG_RT_GROUP_SCHED
10684 .name = "rt_runtime_us",
10685 .read_s64 = cpu_rt_runtime_read,
10686 .write_s64 = cpu_rt_runtime_write,
10689 .name = "rt_period_us",
10690 .read_u64 = cpu_rt_period_read_uint,
10691 .write_u64 = cpu_rt_period_write_uint,
10696 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10698 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10701 struct cgroup_subsys cpu_cgroup_subsys = {
10703 .create = cpu_cgroup_create,
10704 .destroy = cpu_cgroup_destroy,
10705 .can_attach = cpu_cgroup_can_attach,
10706 .attach = cpu_cgroup_attach,
10707 .populate = cpu_cgroup_populate,
10708 .subsys_id = cpu_cgroup_subsys_id,
10712 #endif /* CONFIG_CGROUP_SCHED */
10714 #ifdef CONFIG_CGROUP_CPUACCT
10717 * CPU accounting code for task groups.
10719 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10720 * (balbir@in.ibm.com).
10723 /* track cpu usage of a group of tasks and its child groups */
10725 struct cgroup_subsys_state css;
10726 /* cpuusage holds pointer to a u64-type object on every cpu */
10728 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10729 struct cpuacct *parent;
10732 struct cgroup_subsys cpuacct_subsys;
10734 /* return cpu accounting group corresponding to this container */
10735 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10737 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10738 struct cpuacct, css);
10741 /* return cpu accounting group to which this task belongs */
10742 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10744 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10745 struct cpuacct, css);
10748 /* create a new cpu accounting group */
10749 static struct cgroup_subsys_state *cpuacct_create(
10750 struct cgroup_subsys *ss, struct cgroup *cgrp)
10752 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10758 ca->cpuusage = alloc_percpu(u64);
10762 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10763 if (percpu_counter_init(&ca->cpustat[i], 0))
10764 goto out_free_counters;
10767 ca->parent = cgroup_ca(cgrp->parent);
10773 percpu_counter_destroy(&ca->cpustat[i]);
10774 free_percpu(ca->cpuusage);
10778 return ERR_PTR(-ENOMEM);
10781 /* destroy an existing cpu accounting group */
10783 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10785 struct cpuacct *ca = cgroup_ca(cgrp);
10788 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10789 percpu_counter_destroy(&ca->cpustat[i]);
10790 free_percpu(ca->cpuusage);
10794 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10796 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10799 #ifndef CONFIG_64BIT
10801 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10803 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10805 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10813 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10815 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10817 #ifndef CONFIG_64BIT
10819 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10821 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10823 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10829 /* return total cpu usage (in nanoseconds) of a group */
10830 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10832 struct cpuacct *ca = cgroup_ca(cgrp);
10833 u64 totalcpuusage = 0;
10836 for_each_present_cpu(i)
10837 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10839 return totalcpuusage;
10842 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10845 struct cpuacct *ca = cgroup_ca(cgrp);
10854 for_each_present_cpu(i)
10855 cpuacct_cpuusage_write(ca, i, 0);
10861 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10862 struct seq_file *m)
10864 struct cpuacct *ca = cgroup_ca(cgroup);
10868 for_each_present_cpu(i) {
10869 percpu = cpuacct_cpuusage_read(ca, i);
10870 seq_printf(m, "%llu ", (unsigned long long) percpu);
10872 seq_printf(m, "\n");
10876 static const char *cpuacct_stat_desc[] = {
10877 [CPUACCT_STAT_USER] = "user",
10878 [CPUACCT_STAT_SYSTEM] = "system",
10881 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10882 struct cgroup_map_cb *cb)
10884 struct cpuacct *ca = cgroup_ca(cgrp);
10887 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10888 s64 val = percpu_counter_read(&ca->cpustat[i]);
10889 val = cputime64_to_clock_t(val);
10890 cb->fill(cb, cpuacct_stat_desc[i], val);
10895 static struct cftype files[] = {
10898 .read_u64 = cpuusage_read,
10899 .write_u64 = cpuusage_write,
10902 .name = "usage_percpu",
10903 .read_seq_string = cpuacct_percpu_seq_read,
10907 .read_map = cpuacct_stats_show,
10911 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10913 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10917 * charge this task's execution time to its accounting group.
10919 * called with rq->lock held.
10921 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10923 struct cpuacct *ca;
10926 if (unlikely(!cpuacct_subsys.active))
10929 cpu = task_cpu(tsk);
10935 for (; ca; ca = ca->parent) {
10936 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10937 *cpuusage += cputime;
10944 * Charge the system/user time to the task's accounting group.
10946 static void cpuacct_update_stats(struct task_struct *tsk,
10947 enum cpuacct_stat_index idx, cputime_t val)
10949 struct cpuacct *ca;
10951 if (unlikely(!cpuacct_subsys.active))
10958 percpu_counter_add(&ca->cpustat[idx], val);
10964 struct cgroup_subsys cpuacct_subsys = {
10966 .create = cpuacct_create,
10967 .destroy = cpuacct_destroy,
10968 .populate = cpuacct_populate,
10969 .subsys_id = cpuacct_subsys_id,
10971 #endif /* CONFIG_CGROUP_CPUACCT */
10975 int rcu_expedited_torture_stats(char *page)
10979 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10981 void synchronize_sched_expedited(void)
10984 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10986 #else /* #ifndef CONFIG_SMP */
10988 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10989 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10991 #define RCU_EXPEDITED_STATE_POST -2
10992 #define RCU_EXPEDITED_STATE_IDLE -1
10994 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10996 int rcu_expedited_torture_stats(char *page)
11001 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
11002 for_each_online_cpu(cpu) {
11003 cnt += sprintf(&page[cnt], " %d:%d",
11004 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
11006 cnt += sprintf(&page[cnt], "\n");
11009 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
11011 static long synchronize_sched_expedited_count;
11014 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
11015 * approach to force grace period to end quickly. This consumes
11016 * significant time on all CPUs, and is thus not recommended for
11017 * any sort of common-case code.
11019 * Note that it is illegal to call this function while holding any
11020 * lock that is acquired by a CPU-hotplug notifier. Failing to
11021 * observe this restriction will result in deadlock.
11023 void synchronize_sched_expedited(void)
11026 unsigned long flags;
11027 bool need_full_sync = 0;
11029 struct migration_req *req;
11033 smp_mb(); /* ensure prior mod happens before capturing snap. */
11034 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
11036 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
11038 if (trycount++ < 10)
11039 udelay(trycount * num_online_cpus());
11041 synchronize_sched();
11044 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
11045 smp_mb(); /* ensure test happens before caller kfree */
11050 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
11051 for_each_online_cpu(cpu) {
11053 req = &per_cpu(rcu_migration_req, cpu);
11054 init_completion(&req->done);
11056 req->dest_cpu = RCU_MIGRATION_NEED_QS;
11057 raw_spin_lock_irqsave(&rq->lock, flags);
11058 list_add(&req->list, &rq->migration_queue);
11059 raw_spin_unlock_irqrestore(&rq->lock, flags);
11060 wake_up_process(rq->migration_thread);
11062 for_each_online_cpu(cpu) {
11063 rcu_expedited_state = cpu;
11064 req = &per_cpu(rcu_migration_req, cpu);
11066 wait_for_completion(&req->done);
11067 raw_spin_lock_irqsave(&rq->lock, flags);
11068 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
11069 need_full_sync = 1;
11070 req->dest_cpu = RCU_MIGRATION_IDLE;
11071 raw_spin_unlock_irqrestore(&rq->lock, flags);
11073 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11074 synchronize_sched_expedited_count++;
11075 mutex_unlock(&rcu_sched_expedited_mutex);
11077 if (need_full_sync)
11078 synchronize_sched();
11080 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11082 #endif /* #else #ifndef CONFIG_SMP */