4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct *user)
282 user->tg->uid = user->uid;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
301 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq);
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock);
312 #ifdef CONFIG_FAIR_GROUP_SCHED
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group.children);
321 #ifdef CONFIG_USER_SCHED
322 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
323 #else /* !CONFIG_USER_SCHED */
324 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 #endif /* CONFIG_USER_SCHED */
328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
336 #define MAX_SHARES (1UL << 18)
338 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
341 /* Default task group.
342 * Every task in system belong to this group at bootup.
344 struct task_group init_task_group;
346 /* return group to which a task belongs */
347 static inline struct task_group *task_group(struct task_struct *p)
349 struct task_group *tg;
351 #ifdef CONFIG_USER_SCHED
353 tg = __task_cred(p)->user->tg;
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
359 tg = &init_task_group;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
380 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
381 static inline struct task_group *task_group(struct task_struct *p)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load;
391 unsigned long nr_running;
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
399 struct list_head tasks;
400 struct list_head *balance_iterator;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity *curr, *next, *last;
408 unsigned int nr_spread_over;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load;
439 * this cpu's part of tg->shares
441 unsigned long shares;
444 * load.weight at the time we set shares
446 unsigned long rq_weight;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active;
454 unsigned long rt_nr_running;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
457 int curr; /* highest queued rt task prio */
459 int next; /* next highest */
464 unsigned long rt_nr_migratory;
465 unsigned long rt_nr_total;
467 struct plist_head pushable_tasks;
472 /* Nests inside the rq lock: */
473 spinlock_t rt_runtime_lock;
475 #ifdef CONFIG_RT_GROUP_SCHED
476 unsigned long rt_nr_boosted;
479 struct list_head leaf_rt_rq_list;
480 struct task_group *tg;
481 struct sched_rt_entity *rt_se;
488 * We add the notion of a root-domain which will be used to define per-domain
489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
491 * exclusive cpuset is created, we also create and attach a new root-domain
498 cpumask_var_t online;
501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
504 cpumask_var_t rto_mask;
507 struct cpupri cpupri;
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
515 static struct root_domain def_root_domain;
520 * This is the main, per-CPU runqueue data structure.
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
534 unsigned long nr_running;
535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
538 unsigned 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)
649 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
650 * See detach_destroy_domains: synchronize_sched for details.
652 * The domain tree of any CPU may only be accessed from within
653 * preempt-disabled sections.
655 #define for_each_domain(cpu, __sd) \
656 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
658 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
659 #define this_rq() (&__get_cpu_var(runqueues))
660 #define task_rq(p) cpu_rq(task_cpu(p))
661 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
662 #define raw_rq() (&__raw_get_cpu_var(runqueues))
664 inline void update_rq_clock(struct rq *rq)
666 rq->clock = sched_clock_cpu(cpu_of(rq));
670 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
672 #ifdef CONFIG_SCHED_DEBUG
673 # define const_debug __read_mostly
675 # define const_debug static const
680 * @cpu: the processor in question.
682 * Returns true if the current cpu runqueue is locked.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
686 int runqueue_is_locked(int cpu)
688 return spin_is_locked(&cpu_rq(cpu)->lock);
692 * Debugging: various feature bits
695 #define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
699 #include "sched_features.h"
704 #define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
707 const_debug unsigned int sysctl_sched_features =
708 #include "sched_features.h"
713 #ifdef CONFIG_SCHED_DEBUG
714 #define SCHED_FEAT(name, enabled) \
717 static __read_mostly char *sched_feat_names[] = {
718 #include "sched_features.h"
724 static int sched_feat_show(struct seq_file *m, void *v)
728 for (i = 0; sched_feat_names[i]; i++) {
729 if (!(sysctl_sched_features & (1UL << i)))
731 seq_printf(m, "%s ", sched_feat_names[i]);
739 sched_feat_write(struct file *filp, const char __user *ubuf,
740 size_t cnt, loff_t *ppos)
750 if (copy_from_user(&buf, ubuf, cnt))
755 if (strncmp(buf, "NO_", 3) == 0) {
760 for (i = 0; sched_feat_names[i]; i++) {
761 int len = strlen(sched_feat_names[i]);
763 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
765 sysctl_sched_features &= ~(1UL << i);
767 sysctl_sched_features |= (1UL << i);
772 if (!sched_feat_names[i])
780 static int sched_feat_open(struct inode *inode, struct file *filp)
782 return single_open(filp, sched_feat_show, NULL);
785 static const struct file_operations sched_feat_fops = {
786 .open = sched_feat_open,
787 .write = sched_feat_write,
790 .release = single_release,
793 static __init int sched_init_debug(void)
795 debugfs_create_file("sched_features", 0644, NULL, NULL,
800 late_initcall(sched_init_debug);
804 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
807 * Number of tasks to iterate in a single balance run.
808 * Limited because this is done with IRQs disabled.
810 const_debug unsigned int sysctl_sched_nr_migrate = 32;
813 * ratelimit for updating the group shares.
816 unsigned int sysctl_sched_shares_ratelimit = 250000;
817 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
820 * Inject some fuzzyness into changing the per-cpu group shares
821 * this avoids remote rq-locks at the expense of fairness.
824 unsigned int sysctl_sched_shares_thresh = 4;
827 * period over which we average the RT time consumption, measured
832 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
835 * period over which we measure -rt task cpu usage in us.
838 unsigned int sysctl_sched_rt_period = 1000000;
840 static __read_mostly int scheduler_running;
843 * part of the period that we allow rt tasks to run in us.
846 int sysctl_sched_rt_runtime = 950000;
848 static inline u64 global_rt_period(void)
850 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
853 static inline u64 global_rt_runtime(void)
855 if (sysctl_sched_rt_runtime < 0)
858 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
861 #ifndef prepare_arch_switch
862 # define prepare_arch_switch(next) do { } while (0)
864 #ifndef finish_arch_switch
865 # define finish_arch_switch(prev) do { } while (0)
868 static inline int task_current(struct rq *rq, struct task_struct *p)
870 return rq->curr == p;
873 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
874 static inline int task_running(struct rq *rq, struct task_struct *p)
876 return task_current(rq, p);
879 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
883 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
885 #ifdef CONFIG_DEBUG_SPINLOCK
886 /* this is a valid case when another task releases the spinlock */
887 rq->lock.owner = current;
890 * If we are tracking spinlock dependencies then we have to
891 * fix up the runqueue lock - which gets 'carried over' from
894 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
896 spin_unlock_irq(&rq->lock);
899 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
900 static inline int task_running(struct rq *rq, struct task_struct *p)
905 return task_current(rq, p);
909 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
913 * We can optimise this out completely for !SMP, because the
914 * SMP rebalancing from interrupt is the only thing that cares
919 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
920 spin_unlock_irq(&rq->lock);
922 spin_unlock(&rq->lock);
926 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
930 * After ->oncpu is cleared, the task can be moved to a different CPU.
931 * We must ensure this doesn't happen until the switch is completely
937 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
941 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
944 * __task_rq_lock - lock the runqueue a given task resides on.
945 * Must be called interrupts disabled.
947 static inline struct rq *__task_rq_lock(struct task_struct *p)
951 struct rq *rq = task_rq(p);
952 spin_lock(&rq->lock);
953 if (likely(rq == task_rq(p)))
955 spin_unlock(&rq->lock);
960 * task_rq_lock - lock the runqueue a given task resides on and disable
961 * interrupts. Note the ordering: we can safely lookup the task_rq without
962 * explicitly disabling preemption.
964 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
970 local_irq_save(*flags);
972 spin_lock(&rq->lock);
973 if (likely(rq == task_rq(p)))
975 spin_unlock_irqrestore(&rq->lock, *flags);
979 void task_rq_unlock_wait(struct task_struct *p)
981 struct rq *rq = task_rq(p);
983 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
984 spin_unlock_wait(&rq->lock);
987 static void __task_rq_unlock(struct rq *rq)
990 spin_unlock(&rq->lock);
993 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
996 spin_unlock_irqrestore(&rq->lock, *flags);
1000 * this_rq_lock - lock this runqueue and disable interrupts.
1002 static struct rq *this_rq_lock(void)
1003 __acquires(rq->lock)
1007 local_irq_disable();
1009 spin_lock(&rq->lock);
1014 #ifdef CONFIG_SCHED_HRTICK
1016 * Use HR-timers to deliver accurate preemption points.
1018 * Its all a bit involved since we cannot program an hrt while holding the
1019 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1022 * When we get rescheduled we reprogram the hrtick_timer outside of the
1028 * - enabled by features
1029 * - hrtimer is actually high res
1031 static inline int hrtick_enabled(struct rq *rq)
1033 if (!sched_feat(HRTICK))
1035 if (!cpu_active(cpu_of(rq)))
1037 return hrtimer_is_hres_active(&rq->hrtick_timer);
1040 static void hrtick_clear(struct rq *rq)
1042 if (hrtimer_active(&rq->hrtick_timer))
1043 hrtimer_cancel(&rq->hrtick_timer);
1047 * High-resolution timer tick.
1048 * Runs from hardirq context with interrupts disabled.
1050 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1052 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1054 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1056 spin_lock(&rq->lock);
1057 update_rq_clock(rq);
1058 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1059 spin_unlock(&rq->lock);
1061 return HRTIMER_NORESTART;
1066 * called from hardirq (IPI) context
1068 static void __hrtick_start(void *arg)
1070 struct rq *rq = arg;
1072 spin_lock(&rq->lock);
1073 hrtimer_restart(&rq->hrtick_timer);
1074 rq->hrtick_csd_pending = 0;
1075 spin_unlock(&rq->lock);
1079 * Called to set the hrtick timer state.
1081 * called with rq->lock held and irqs disabled
1083 static void hrtick_start(struct rq *rq, u64 delay)
1085 struct hrtimer *timer = &rq->hrtick_timer;
1086 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1088 hrtimer_set_expires(timer, time);
1090 if (rq == this_rq()) {
1091 hrtimer_restart(timer);
1092 } else if (!rq->hrtick_csd_pending) {
1093 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1094 rq->hrtick_csd_pending = 1;
1099 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1101 int cpu = (int)(long)hcpu;
1104 case CPU_UP_CANCELED:
1105 case CPU_UP_CANCELED_FROZEN:
1106 case CPU_DOWN_PREPARE:
1107 case CPU_DOWN_PREPARE_FROZEN:
1109 case CPU_DEAD_FROZEN:
1110 hrtick_clear(cpu_rq(cpu));
1117 static __init void init_hrtick(void)
1119 hotcpu_notifier(hotplug_hrtick, 0);
1123 * Called to set the hrtick timer state.
1125 * called with rq->lock held and irqs disabled
1127 static void hrtick_start(struct rq *rq, u64 delay)
1129 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1130 HRTIMER_MODE_REL_PINNED, 0);
1133 static inline void init_hrtick(void)
1136 #endif /* CONFIG_SMP */
1138 static void init_rq_hrtick(struct rq *rq)
1141 rq->hrtick_csd_pending = 0;
1143 rq->hrtick_csd.flags = 0;
1144 rq->hrtick_csd.func = __hrtick_start;
1145 rq->hrtick_csd.info = rq;
1148 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1149 rq->hrtick_timer.function = hrtick;
1151 #else /* CONFIG_SCHED_HRTICK */
1152 static inline void hrtick_clear(struct rq *rq)
1156 static inline void init_rq_hrtick(struct rq *rq)
1160 static inline void init_hrtick(void)
1163 #endif /* CONFIG_SCHED_HRTICK */
1166 * resched_task - mark a task 'to be rescheduled now'.
1168 * On UP this means the setting of the need_resched flag, on SMP it
1169 * might also involve a cross-CPU call to trigger the scheduler on
1174 #ifndef tsk_is_polling
1175 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1178 static void resched_task(struct task_struct *p)
1182 assert_spin_locked(&task_rq(p)->lock);
1184 if (test_tsk_need_resched(p))
1187 set_tsk_need_resched(p);
1190 if (cpu == smp_processor_id())
1193 /* NEED_RESCHED must be visible before we test polling */
1195 if (!tsk_is_polling(p))
1196 smp_send_reschedule(cpu);
1199 static void resched_cpu(int cpu)
1201 struct rq *rq = cpu_rq(cpu);
1202 unsigned long flags;
1204 if (!spin_trylock_irqsave(&rq->lock, flags))
1206 resched_task(cpu_curr(cpu));
1207 spin_unlock_irqrestore(&rq->lock, flags);
1212 * When add_timer_on() enqueues a timer into the timer wheel of an
1213 * idle CPU then this timer might expire before the next timer event
1214 * which is scheduled to wake up that CPU. In case of a completely
1215 * idle system the next event might even be infinite time into the
1216 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1217 * leaves the inner idle loop so the newly added timer is taken into
1218 * account when the CPU goes back to idle and evaluates the timer
1219 * wheel for the next timer event.
1221 void wake_up_idle_cpu(int cpu)
1223 struct rq *rq = cpu_rq(cpu);
1225 if (cpu == smp_processor_id())
1229 * This is safe, as this function is called with the timer
1230 * wheel base lock of (cpu) held. When the CPU is on the way
1231 * to idle and has not yet set rq->curr to idle then it will
1232 * be serialized on the timer wheel base lock and take the new
1233 * timer into account automatically.
1235 if (rq->curr != rq->idle)
1239 * We can set TIF_RESCHED on the idle task of the other CPU
1240 * lockless. The worst case is that the other CPU runs the
1241 * idle task through an additional NOOP schedule()
1243 set_tsk_need_resched(rq->idle);
1245 /* NEED_RESCHED must be visible before we test polling */
1247 if (!tsk_is_polling(rq->idle))
1248 smp_send_reschedule(cpu);
1250 #endif /* CONFIG_NO_HZ */
1252 static u64 sched_avg_period(void)
1254 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1257 static void sched_avg_update(struct rq *rq)
1259 s64 period = sched_avg_period();
1261 while ((s64)(rq->clock - rq->age_stamp) > period) {
1262 rq->age_stamp += period;
1267 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1269 rq->rt_avg += rt_delta;
1270 sched_avg_update(rq);
1273 #else /* !CONFIG_SMP */
1274 static void resched_task(struct task_struct *p)
1276 assert_spin_locked(&task_rq(p)->lock);
1277 set_tsk_need_resched(p);
1280 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1283 #endif /* CONFIG_SMP */
1285 #if BITS_PER_LONG == 32
1286 # define WMULT_CONST (~0UL)
1288 # define WMULT_CONST (1UL << 32)
1291 #define WMULT_SHIFT 32
1294 * Shift right and round:
1296 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1299 * delta *= weight / lw
1301 static unsigned long
1302 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1303 struct load_weight *lw)
1307 if (!lw->inv_weight) {
1308 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1311 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1315 tmp = (u64)delta_exec * weight;
1317 * Check whether we'd overflow the 64-bit multiplication:
1319 if (unlikely(tmp > WMULT_CONST))
1320 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1323 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1325 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1328 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1334 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1341 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1342 * of tasks with abnormal "nice" values across CPUs the contribution that
1343 * each task makes to its run queue's load is weighted according to its
1344 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1345 * scaled version of the new time slice allocation that they receive on time
1349 #define WEIGHT_IDLEPRIO 3
1350 #define WMULT_IDLEPRIO 1431655765
1353 * Nice levels are multiplicative, with a gentle 10% change for every
1354 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1355 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1356 * that remained on nice 0.
1358 * The "10% effect" is relative and cumulative: from _any_ nice level,
1359 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1360 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1361 * If a task goes up by ~10% and another task goes down by ~10% then
1362 * the relative distance between them is ~25%.)
1364 static const int prio_to_weight[40] = {
1365 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1366 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1367 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1368 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1369 /* 0 */ 1024, 820, 655, 526, 423,
1370 /* 5 */ 335, 272, 215, 172, 137,
1371 /* 10 */ 110, 87, 70, 56, 45,
1372 /* 15 */ 36, 29, 23, 18, 15,
1376 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1378 * In cases where the weight does not change often, we can use the
1379 * precalculated inverse to speed up arithmetics by turning divisions
1380 * into multiplications:
1382 static const u32 prio_to_wmult[40] = {
1383 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1384 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1385 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1386 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1387 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1388 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1389 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1390 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1393 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1396 * runqueue iterator, to support SMP load-balancing between different
1397 * scheduling classes, without having to expose their internal data
1398 * structures to the load-balancing proper:
1400 struct rq_iterator {
1402 struct task_struct *(*start)(void *);
1403 struct task_struct *(*next)(void *);
1407 static unsigned long
1408 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1409 unsigned long max_load_move, struct sched_domain *sd,
1410 enum cpu_idle_type idle, int *all_pinned,
1411 int *this_best_prio, struct rq_iterator *iterator);
1414 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1415 struct sched_domain *sd, enum cpu_idle_type idle,
1416 struct rq_iterator *iterator);
1419 /* Time spent by the tasks of the cpu accounting group executing in ... */
1420 enum cpuacct_stat_index {
1421 CPUACCT_STAT_USER, /* ... user mode */
1422 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1424 CPUACCT_STAT_NSTATS,
1427 #ifdef CONFIG_CGROUP_CPUACCT
1428 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1429 static void cpuacct_update_stats(struct task_struct *tsk,
1430 enum cpuacct_stat_index idx, cputime_t val);
1432 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1433 static inline void cpuacct_update_stats(struct task_struct *tsk,
1434 enum cpuacct_stat_index idx, cputime_t val) {}
1437 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1439 update_load_add(&rq->load, load);
1442 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1444 update_load_sub(&rq->load, load);
1447 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1448 typedef int (*tg_visitor)(struct task_group *, void *);
1451 * Iterate the full tree, calling @down when first entering a node and @up when
1452 * leaving it for the final time.
1454 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1456 struct task_group *parent, *child;
1460 parent = &root_task_group;
1462 ret = (*down)(parent, data);
1465 list_for_each_entry_rcu(child, &parent->children, siblings) {
1472 ret = (*up)(parent, data);
1477 parent = parent->parent;
1486 static int tg_nop(struct task_group *tg, void *data)
1493 /* Used instead of source_load when we know the type == 0 */
1494 static unsigned long weighted_cpuload(const int cpu)
1496 return cpu_rq(cpu)->load.weight;
1500 * Return a low guess at the load of a migration-source cpu weighted
1501 * according to the scheduling class and "nice" value.
1503 * We want to under-estimate the load of migration sources, to
1504 * balance conservatively.
1506 static unsigned long source_load(int cpu, int type)
1508 struct rq *rq = cpu_rq(cpu);
1509 unsigned long total = weighted_cpuload(cpu);
1511 if (type == 0 || !sched_feat(LB_BIAS))
1514 return min(rq->cpu_load[type-1], total);
1518 * Return a high guess at the load of a migration-target cpu weighted
1519 * according to the scheduling class and "nice" value.
1521 static unsigned long target_load(int cpu, int type)
1523 struct rq *rq = cpu_rq(cpu);
1524 unsigned long total = weighted_cpuload(cpu);
1526 if (type == 0 || !sched_feat(LB_BIAS))
1529 return max(rq->cpu_load[type-1], total);
1532 static struct sched_group *group_of(int cpu)
1534 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1542 static unsigned long power_of(int cpu)
1544 struct sched_group *group = group_of(cpu);
1547 return SCHED_LOAD_SCALE;
1549 return group->cpu_power;
1552 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1554 static unsigned long cpu_avg_load_per_task(int cpu)
1556 struct rq *rq = cpu_rq(cpu);
1557 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1560 rq->avg_load_per_task = rq->load.weight / nr_running;
1562 rq->avg_load_per_task = 0;
1564 return rq->avg_load_per_task;
1567 #ifdef CONFIG_FAIR_GROUP_SCHED
1569 static __read_mostly unsigned long *update_shares_data;
1571 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1574 * Calculate and set the cpu's group shares.
1576 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1577 unsigned long sd_shares,
1578 unsigned long sd_rq_weight,
1579 unsigned long *usd_rq_weight)
1581 unsigned long shares, rq_weight;
1584 rq_weight = usd_rq_weight[cpu];
1587 rq_weight = NICE_0_LOAD;
1591 * \Sum_j shares_j * rq_weight_i
1592 * shares_i = -----------------------------
1593 * \Sum_j rq_weight_j
1595 shares = (sd_shares * rq_weight) / sd_rq_weight;
1596 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1598 if (abs(shares - tg->se[cpu]->load.weight) >
1599 sysctl_sched_shares_thresh) {
1600 struct rq *rq = cpu_rq(cpu);
1601 unsigned long flags;
1603 spin_lock_irqsave(&rq->lock, flags);
1604 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1605 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1606 __set_se_shares(tg->se[cpu], shares);
1607 spin_unlock_irqrestore(&rq->lock, flags);
1612 * Re-compute the task group their per cpu shares over the given domain.
1613 * This needs to be done in a bottom-up fashion because the rq weight of a
1614 * parent group depends on the shares of its child groups.
1616 static int tg_shares_up(struct task_group *tg, void *data)
1618 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1619 unsigned long *usd_rq_weight;
1620 struct sched_domain *sd = data;
1621 unsigned long flags;
1627 local_irq_save(flags);
1628 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1630 for_each_cpu(i, sched_domain_span(sd)) {
1631 weight = tg->cfs_rq[i]->load.weight;
1632 usd_rq_weight[i] = weight;
1634 rq_weight += weight;
1636 * If there are currently no tasks on the cpu pretend there
1637 * is one of average load so that when a new task gets to
1638 * run here it will not get delayed by group starvation.
1641 weight = NICE_0_LOAD;
1643 sum_weight += weight;
1644 shares += tg->cfs_rq[i]->shares;
1648 rq_weight = sum_weight;
1650 if ((!shares && rq_weight) || shares > tg->shares)
1651 shares = tg->shares;
1653 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1654 shares = tg->shares;
1656 for_each_cpu(i, sched_domain_span(sd))
1657 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1659 local_irq_restore(flags);
1665 * Compute the cpu's hierarchical load factor for each task group.
1666 * This needs to be done in a top-down fashion because the load of a child
1667 * group is a fraction of its parents load.
1669 static int tg_load_down(struct task_group *tg, void *data)
1672 long cpu = (long)data;
1675 load = cpu_rq(cpu)->load.weight;
1677 load = tg->parent->cfs_rq[cpu]->h_load;
1678 load *= tg->cfs_rq[cpu]->shares;
1679 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1682 tg->cfs_rq[cpu]->h_load = load;
1687 static void update_shares(struct sched_domain *sd)
1692 if (root_task_group_empty())
1695 now = cpu_clock(raw_smp_processor_id());
1696 elapsed = now - sd->last_update;
1698 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1699 sd->last_update = now;
1700 walk_tg_tree(tg_nop, tg_shares_up, sd);
1704 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1706 if (root_task_group_empty())
1709 spin_unlock(&rq->lock);
1711 spin_lock(&rq->lock);
1714 static void update_h_load(long cpu)
1716 if (root_task_group_empty())
1719 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1724 static inline void update_shares(struct sched_domain *sd)
1728 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1734 #ifdef CONFIG_PREEMPT
1736 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1739 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1740 * way at the expense of forcing extra atomic operations in all
1741 * invocations. This assures that the double_lock is acquired using the
1742 * same underlying policy as the spinlock_t on this architecture, which
1743 * reduces latency compared to the unfair variant below. However, it
1744 * also adds more overhead and therefore may reduce throughput.
1746 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1747 __releases(this_rq->lock)
1748 __acquires(busiest->lock)
1749 __acquires(this_rq->lock)
1751 spin_unlock(&this_rq->lock);
1752 double_rq_lock(this_rq, busiest);
1759 * Unfair double_lock_balance: Optimizes throughput at the expense of
1760 * latency by eliminating extra atomic operations when the locks are
1761 * already in proper order on entry. This favors lower cpu-ids and will
1762 * grant the double lock to lower cpus over higher ids under contention,
1763 * regardless of entry order into the function.
1765 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1766 __releases(this_rq->lock)
1767 __acquires(busiest->lock)
1768 __acquires(this_rq->lock)
1772 if (unlikely(!spin_trylock(&busiest->lock))) {
1773 if (busiest < this_rq) {
1774 spin_unlock(&this_rq->lock);
1775 spin_lock(&busiest->lock);
1776 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1779 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1784 #endif /* CONFIG_PREEMPT */
1787 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1789 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1791 if (unlikely(!irqs_disabled())) {
1792 /* printk() doesn't work good under rq->lock */
1793 spin_unlock(&this_rq->lock);
1797 return _double_lock_balance(this_rq, busiest);
1800 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1801 __releases(busiest->lock)
1803 spin_unlock(&busiest->lock);
1804 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1808 #ifdef CONFIG_FAIR_GROUP_SCHED
1809 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1812 cfs_rq->shares = shares;
1817 static void calc_load_account_active(struct rq *this_rq);
1818 static void update_sysctl(void);
1820 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1822 set_task_rq(p, cpu);
1825 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1826 * successfuly executed on another CPU. We must ensure that updates of
1827 * per-task data have been completed by this moment.
1830 task_thread_info(p)->cpu = cpu;
1834 #include "sched_stats.h"
1835 #include "sched_idletask.c"
1836 #include "sched_fair.c"
1837 #include "sched_rt.c"
1838 #ifdef CONFIG_SCHED_DEBUG
1839 # include "sched_debug.c"
1842 #define sched_class_highest (&rt_sched_class)
1843 #define for_each_class(class) \
1844 for (class = sched_class_highest; class; class = class->next)
1846 static void inc_nr_running(struct rq *rq)
1851 static void dec_nr_running(struct rq *rq)
1856 static void set_load_weight(struct task_struct *p)
1858 if (task_has_rt_policy(p)) {
1859 p->se.load.weight = prio_to_weight[0] * 2;
1860 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1865 * SCHED_IDLE tasks get minimal weight:
1867 if (p->policy == SCHED_IDLE) {
1868 p->se.load.weight = WEIGHT_IDLEPRIO;
1869 p->se.load.inv_weight = WMULT_IDLEPRIO;
1873 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1874 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1877 static void update_avg(u64 *avg, u64 sample)
1879 s64 diff = sample - *avg;
1883 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1886 p->se.start_runtime = p->se.sum_exec_runtime;
1888 sched_info_queued(p);
1889 p->sched_class->enqueue_task(rq, p, wakeup);
1893 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1896 if (p->se.last_wakeup) {
1897 update_avg(&p->se.avg_overlap,
1898 p->se.sum_exec_runtime - p->se.last_wakeup);
1899 p->se.last_wakeup = 0;
1901 update_avg(&p->se.avg_wakeup,
1902 sysctl_sched_wakeup_granularity);
1906 sched_info_dequeued(p);
1907 p->sched_class->dequeue_task(rq, p, sleep);
1912 * __normal_prio - return the priority that is based on the static prio
1914 static inline int __normal_prio(struct task_struct *p)
1916 return p->static_prio;
1920 * Calculate the expected normal priority: i.e. priority
1921 * without taking RT-inheritance into account. Might be
1922 * boosted by interactivity modifiers. Changes upon fork,
1923 * setprio syscalls, and whenever the interactivity
1924 * estimator recalculates.
1926 static inline int normal_prio(struct task_struct *p)
1930 if (task_has_rt_policy(p))
1931 prio = MAX_RT_PRIO-1 - p->rt_priority;
1933 prio = __normal_prio(p);
1938 * Calculate the current priority, i.e. the priority
1939 * taken into account by the scheduler. This value might
1940 * be boosted by RT tasks, or might be boosted by
1941 * interactivity modifiers. Will be RT if the task got
1942 * RT-boosted. If not then it returns p->normal_prio.
1944 static int effective_prio(struct task_struct *p)
1946 p->normal_prio = normal_prio(p);
1948 * If we are RT tasks or we were boosted to RT priority,
1949 * keep the priority unchanged. Otherwise, update priority
1950 * to the normal priority:
1952 if (!rt_prio(p->prio))
1953 return p->normal_prio;
1958 * activate_task - move a task to the runqueue.
1960 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1962 if (task_contributes_to_load(p))
1963 rq->nr_uninterruptible--;
1965 enqueue_task(rq, p, wakeup);
1970 * deactivate_task - remove a task from the runqueue.
1972 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1974 if (task_contributes_to_load(p))
1975 rq->nr_uninterruptible++;
1977 dequeue_task(rq, p, sleep);
1982 * task_curr - is this task currently executing on a CPU?
1983 * @p: the task in question.
1985 inline int task_curr(const struct task_struct *p)
1987 return cpu_curr(task_cpu(p)) == p;
1990 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1991 const struct sched_class *prev_class,
1992 int oldprio, int running)
1994 if (prev_class != p->sched_class) {
1995 if (prev_class->switched_from)
1996 prev_class->switched_from(rq, p, running);
1997 p->sched_class->switched_to(rq, p, running);
1999 p->sched_class->prio_changed(rq, p, oldprio, running);
2003 * kthread_bind - bind a just-created kthread to a cpu.
2004 * @p: thread created by kthread_create().
2005 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2007 * Description: This function is equivalent to set_cpus_allowed(),
2008 * except that @cpu doesn't need to be online, and the thread must be
2009 * stopped (i.e., just returned from kthread_create()).
2011 * Function lives here instead of kthread.c because it messes with
2012 * scheduler internals which require locking.
2014 void kthread_bind(struct task_struct *p, unsigned int cpu)
2016 struct rq *rq = cpu_rq(cpu);
2017 unsigned long flags;
2019 /* Must have done schedule() in kthread() before we set_task_cpu */
2020 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2025 spin_lock_irqsave(&rq->lock, flags);
2026 update_rq_clock(rq);
2027 set_task_cpu(p, cpu);
2028 p->cpus_allowed = cpumask_of_cpu(cpu);
2029 p->rt.nr_cpus_allowed = 1;
2030 p->flags |= PF_THREAD_BOUND;
2031 spin_unlock_irqrestore(&rq->lock, flags);
2033 EXPORT_SYMBOL(kthread_bind);
2037 * Is this task likely cache-hot:
2040 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2045 * Buddy candidates are cache hot:
2047 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2048 (&p->se == cfs_rq_of(&p->se)->next ||
2049 &p->se == cfs_rq_of(&p->se)->last))
2052 if (p->sched_class != &fair_sched_class)
2055 if (sysctl_sched_migration_cost == -1)
2057 if (sysctl_sched_migration_cost == 0)
2060 delta = now - p->se.exec_start;
2062 return delta < (s64)sysctl_sched_migration_cost;
2066 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2068 int old_cpu = task_cpu(p);
2069 struct rq *old_rq = cpu_rq(old_cpu);
2070 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2071 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2073 trace_sched_migrate_task(p, new_cpu);
2075 if (old_cpu != new_cpu) {
2076 p->se.nr_migrations++;
2077 #ifdef CONFIG_SCHEDSTATS
2078 if (task_hot(p, old_rq->clock, NULL))
2079 schedstat_inc(p, se.nr_forced2_migrations);
2081 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2084 p->se.vruntime -= old_cfsrq->min_vruntime -
2085 new_cfsrq->min_vruntime;
2087 __set_task_cpu(p, new_cpu);
2090 struct migration_req {
2091 struct list_head list;
2093 struct task_struct *task;
2096 struct completion done;
2100 * The task's runqueue lock must be held.
2101 * Returns true if you have to wait for migration thread.
2104 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2106 struct rq *rq = task_rq(p);
2109 * If the task is not on a runqueue (and not running), then
2110 * it is sufficient to simply update the task's cpu field.
2112 if (!p->se.on_rq && !task_running(rq, p)) {
2113 update_rq_clock(rq);
2114 set_task_cpu(p, dest_cpu);
2118 init_completion(&req->done);
2120 req->dest_cpu = dest_cpu;
2121 list_add(&req->list, &rq->migration_queue);
2127 * wait_task_context_switch - wait for a thread to complete at least one
2130 * @p must not be current.
2132 void wait_task_context_switch(struct task_struct *p)
2134 unsigned long nvcsw, nivcsw, flags;
2142 * The runqueue is assigned before the actual context
2143 * switch. We need to take the runqueue lock.
2145 * We could check initially without the lock but it is
2146 * very likely that we need to take the lock in every
2149 rq = task_rq_lock(p, &flags);
2150 running = task_running(rq, p);
2151 task_rq_unlock(rq, &flags);
2153 if (likely(!running))
2156 * The switch count is incremented before the actual
2157 * context switch. We thus wait for two switches to be
2158 * sure at least one completed.
2160 if ((p->nvcsw - nvcsw) > 1)
2162 if ((p->nivcsw - nivcsw) > 1)
2170 * wait_task_inactive - wait for a thread to unschedule.
2172 * If @match_state is nonzero, it's the @p->state value just checked and
2173 * not expected to change. If it changes, i.e. @p might have woken up,
2174 * then return zero. When we succeed in waiting for @p to be off its CPU,
2175 * we return a positive number (its total switch count). If a second call
2176 * a short while later returns the same number, the caller can be sure that
2177 * @p has remained unscheduled the whole time.
2179 * The caller must ensure that the task *will* unschedule sometime soon,
2180 * else this function might spin for a *long* time. This function can't
2181 * be called with interrupts off, or it may introduce deadlock with
2182 * smp_call_function() if an IPI is sent by the same process we are
2183 * waiting to become inactive.
2185 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2187 unsigned long flags;
2194 * We do the initial early heuristics without holding
2195 * any task-queue locks at all. We'll only try to get
2196 * the runqueue lock when things look like they will
2202 * If the task is actively running on another CPU
2203 * still, just relax and busy-wait without holding
2206 * NOTE! Since we don't hold any locks, it's not
2207 * even sure that "rq" stays as the right runqueue!
2208 * But we don't care, since "task_running()" will
2209 * return false if the runqueue has changed and p
2210 * is actually now running somewhere else!
2212 while (task_running(rq, p)) {
2213 if (match_state && unlikely(p->state != match_state))
2219 * Ok, time to look more closely! We need the rq
2220 * lock now, to be *sure*. If we're wrong, we'll
2221 * just go back and repeat.
2223 rq = task_rq_lock(p, &flags);
2224 trace_sched_wait_task(rq, p);
2225 running = task_running(rq, p);
2226 on_rq = p->se.on_rq;
2228 if (!match_state || p->state == match_state)
2229 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2230 task_rq_unlock(rq, &flags);
2233 * If it changed from the expected state, bail out now.
2235 if (unlikely(!ncsw))
2239 * Was it really running after all now that we
2240 * checked with the proper locks actually held?
2242 * Oops. Go back and try again..
2244 if (unlikely(running)) {
2250 * It's not enough that it's not actively running,
2251 * it must be off the runqueue _entirely_, and not
2254 * So if it was still runnable (but just not actively
2255 * running right now), it's preempted, and we should
2256 * yield - it could be a while.
2258 if (unlikely(on_rq)) {
2259 schedule_timeout_uninterruptible(1);
2264 * Ahh, all good. It wasn't running, and it wasn't
2265 * runnable, which means that it will never become
2266 * running in the future either. We're all done!
2275 * kick_process - kick a running thread to enter/exit the kernel
2276 * @p: the to-be-kicked thread
2278 * Cause a process which is running on another CPU to enter
2279 * kernel-mode, without any delay. (to get signals handled.)
2281 * NOTE: this function doesnt have to take the runqueue lock,
2282 * because all it wants to ensure is that the remote task enters
2283 * the kernel. If the IPI races and the task has been migrated
2284 * to another CPU then no harm is done and the purpose has been
2287 void kick_process(struct task_struct *p)
2293 if ((cpu != smp_processor_id()) && task_curr(p))
2294 smp_send_reschedule(cpu);
2297 EXPORT_SYMBOL_GPL(kick_process);
2298 #endif /* CONFIG_SMP */
2301 * task_oncpu_function_call - call a function on the cpu on which a task runs
2302 * @p: the task to evaluate
2303 * @func: the function to be called
2304 * @info: the function call argument
2306 * Calls the function @func when the task is currently running. This might
2307 * be on the current CPU, which just calls the function directly
2309 void task_oncpu_function_call(struct task_struct *p,
2310 void (*func) (void *info), void *info)
2317 smp_call_function_single(cpu, func, info, 1);
2323 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2325 return p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2330 * try_to_wake_up - wake up a thread
2331 * @p: the to-be-woken-up thread
2332 * @state: the mask of task states that can be woken
2333 * @sync: do a synchronous wakeup?
2335 * Put it on the run-queue if it's not already there. The "current"
2336 * thread is always on the run-queue (except when the actual
2337 * re-schedule is in progress), and as such you're allowed to do
2338 * the simpler "current->state = TASK_RUNNING" to mark yourself
2339 * runnable without the overhead of this.
2341 * returns failure only if the task is already active.
2343 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2346 int cpu, orig_cpu, this_cpu, success = 0;
2347 unsigned long flags;
2348 struct rq *rq, *orig_rq;
2350 if (!sched_feat(SYNC_WAKEUPS))
2351 wake_flags &= ~WF_SYNC;
2353 this_cpu = get_cpu();
2356 rq = orig_rq = task_rq_lock(p, &flags);
2357 update_rq_clock(rq);
2358 if (!(p->state & state))
2368 if (unlikely(task_running(rq, p)))
2372 * In order to handle concurrent wakeups and release the rq->lock
2373 * we put the task in TASK_WAKING state.
2375 * First fix up the nr_uninterruptible count:
2377 if (task_contributes_to_load(p))
2378 rq->nr_uninterruptible--;
2379 p->state = TASK_WAKING;
2380 __task_rq_unlock(rq);
2382 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2383 if (cpu != orig_cpu)
2384 set_task_cpu(p, cpu);
2386 rq = __task_rq_lock(p);
2387 update_rq_clock(rq);
2389 WARN_ON(p->state != TASK_WAKING);
2392 #ifdef CONFIG_SCHEDSTATS
2393 schedstat_inc(rq, ttwu_count);
2394 if (cpu == this_cpu)
2395 schedstat_inc(rq, ttwu_local);
2397 struct sched_domain *sd;
2398 for_each_domain(this_cpu, sd) {
2399 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2400 schedstat_inc(sd, ttwu_wake_remote);
2405 #endif /* CONFIG_SCHEDSTATS */
2408 #endif /* CONFIG_SMP */
2409 schedstat_inc(p, se.nr_wakeups);
2410 if (wake_flags & WF_SYNC)
2411 schedstat_inc(p, se.nr_wakeups_sync);
2412 if (orig_cpu != cpu)
2413 schedstat_inc(p, se.nr_wakeups_migrate);
2414 if (cpu == this_cpu)
2415 schedstat_inc(p, se.nr_wakeups_local);
2417 schedstat_inc(p, se.nr_wakeups_remote);
2418 activate_task(rq, p, 1);
2422 * Only attribute actual wakeups done by this task.
2424 if (!in_interrupt()) {
2425 struct sched_entity *se = ¤t->se;
2426 u64 sample = se->sum_exec_runtime;
2428 if (se->last_wakeup)
2429 sample -= se->last_wakeup;
2431 sample -= se->start_runtime;
2432 update_avg(&se->avg_wakeup, sample);
2434 se->last_wakeup = se->sum_exec_runtime;
2438 trace_sched_wakeup(rq, p, success);
2439 check_preempt_curr(rq, p, wake_flags);
2441 p->state = TASK_RUNNING;
2443 if (p->sched_class->task_wake_up)
2444 p->sched_class->task_wake_up(rq, p);
2446 if (unlikely(rq->idle_stamp)) {
2447 u64 delta = rq->clock - rq->idle_stamp;
2448 u64 max = 2*sysctl_sched_migration_cost;
2453 update_avg(&rq->avg_idle, delta);
2458 task_rq_unlock(rq, &flags);
2465 * wake_up_process - Wake up a specific process
2466 * @p: The process to be woken up.
2468 * Attempt to wake up the nominated process and move it to the set of runnable
2469 * processes. Returns 1 if the process was woken up, 0 if it was already
2472 * It may be assumed that this function implies a write memory barrier before
2473 * changing the task state if and only if any tasks are woken up.
2475 int wake_up_process(struct task_struct *p)
2477 return try_to_wake_up(p, TASK_ALL, 0);
2479 EXPORT_SYMBOL(wake_up_process);
2481 int wake_up_state(struct task_struct *p, unsigned int state)
2483 return try_to_wake_up(p, state, 0);
2487 * Perform scheduler related setup for a newly forked process p.
2488 * p is forked by current.
2490 * __sched_fork() is basic setup used by init_idle() too:
2492 static void __sched_fork(struct task_struct *p)
2494 p->se.exec_start = 0;
2495 p->se.sum_exec_runtime = 0;
2496 p->se.prev_sum_exec_runtime = 0;
2497 p->se.nr_migrations = 0;
2498 p->se.last_wakeup = 0;
2499 p->se.avg_overlap = 0;
2500 p->se.start_runtime = 0;
2501 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2503 #ifdef CONFIG_SCHEDSTATS
2504 p->se.wait_start = 0;
2506 p->se.wait_count = 0;
2509 p->se.sleep_start = 0;
2510 p->se.sleep_max = 0;
2511 p->se.sum_sleep_runtime = 0;
2513 p->se.block_start = 0;
2514 p->se.block_max = 0;
2516 p->se.slice_max = 0;
2518 p->se.nr_migrations_cold = 0;
2519 p->se.nr_failed_migrations_affine = 0;
2520 p->se.nr_failed_migrations_running = 0;
2521 p->se.nr_failed_migrations_hot = 0;
2522 p->se.nr_forced_migrations = 0;
2523 p->se.nr_forced2_migrations = 0;
2525 p->se.nr_wakeups = 0;
2526 p->se.nr_wakeups_sync = 0;
2527 p->se.nr_wakeups_migrate = 0;
2528 p->se.nr_wakeups_local = 0;
2529 p->se.nr_wakeups_remote = 0;
2530 p->se.nr_wakeups_affine = 0;
2531 p->se.nr_wakeups_affine_attempts = 0;
2532 p->se.nr_wakeups_passive = 0;
2533 p->se.nr_wakeups_idle = 0;
2537 INIT_LIST_HEAD(&p->rt.run_list);
2539 INIT_LIST_HEAD(&p->se.group_node);
2541 #ifdef CONFIG_PREEMPT_NOTIFIERS
2542 INIT_HLIST_HEAD(&p->preempt_notifiers);
2546 * We mark the process as running here, but have not actually
2547 * inserted it onto the runqueue yet. This guarantees that
2548 * nobody will actually run it, and a signal or other external
2549 * event cannot wake it up and insert it on the runqueue either.
2551 p->state = TASK_RUNNING;
2555 * fork()/clone()-time setup:
2557 void sched_fork(struct task_struct *p, int clone_flags)
2559 int cpu = get_cpu();
2564 * Revert to default priority/policy on fork if requested.
2566 if (unlikely(p->sched_reset_on_fork)) {
2567 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2568 p->policy = SCHED_NORMAL;
2569 p->normal_prio = p->static_prio;
2572 if (PRIO_TO_NICE(p->static_prio) < 0) {
2573 p->static_prio = NICE_TO_PRIO(0);
2574 p->normal_prio = p->static_prio;
2579 * We don't need the reset flag anymore after the fork. It has
2580 * fulfilled its duty:
2582 p->sched_reset_on_fork = 0;
2586 * Make sure we do not leak PI boosting priority to the child.
2588 p->prio = current->normal_prio;
2590 if (!rt_prio(p->prio))
2591 p->sched_class = &fair_sched_class;
2593 if (p->sched_class->task_fork)
2594 p->sched_class->task_fork(p);
2597 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2599 set_task_cpu(p, cpu);
2601 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2602 if (likely(sched_info_on()))
2603 memset(&p->sched_info, 0, sizeof(p->sched_info));
2605 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2608 #ifdef CONFIG_PREEMPT
2609 /* Want to start with kernel preemption disabled. */
2610 task_thread_info(p)->preempt_count = 1;
2612 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2618 * wake_up_new_task - wake up a newly created task for the first time.
2620 * This function will do some initial scheduler statistics housekeeping
2621 * that must be done for every newly created context, then puts the task
2622 * on the runqueue and wakes it.
2624 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2626 unsigned long flags;
2629 rq = task_rq_lock(p, &flags);
2630 BUG_ON(p->state != TASK_RUNNING);
2631 update_rq_clock(rq);
2632 activate_task(rq, p, 0);
2633 trace_sched_wakeup_new(rq, p, 1);
2634 check_preempt_curr(rq, p, WF_FORK);
2636 if (p->sched_class->task_wake_up)
2637 p->sched_class->task_wake_up(rq, p);
2639 task_rq_unlock(rq, &flags);
2642 #ifdef CONFIG_PREEMPT_NOTIFIERS
2645 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2646 * @notifier: notifier struct to register
2648 void preempt_notifier_register(struct preempt_notifier *notifier)
2650 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2652 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2655 * preempt_notifier_unregister - no longer interested in preemption notifications
2656 * @notifier: notifier struct to unregister
2658 * This is safe to call from within a preemption notifier.
2660 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2662 hlist_del(¬ifier->link);
2664 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2666 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2668 struct preempt_notifier *notifier;
2669 struct hlist_node *node;
2671 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2672 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2676 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2677 struct task_struct *next)
2679 struct preempt_notifier *notifier;
2680 struct hlist_node *node;
2682 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2683 notifier->ops->sched_out(notifier, next);
2686 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2688 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2693 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2694 struct task_struct *next)
2698 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2701 * prepare_task_switch - prepare to switch tasks
2702 * @rq: the runqueue preparing to switch
2703 * @prev: the current task that is being switched out
2704 * @next: the task we are going to switch to.
2706 * This is called with the rq lock held and interrupts off. It must
2707 * be paired with a subsequent finish_task_switch after the context
2710 * prepare_task_switch sets up locking and calls architecture specific
2714 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2715 struct task_struct *next)
2717 fire_sched_out_preempt_notifiers(prev, next);
2718 prepare_lock_switch(rq, next);
2719 prepare_arch_switch(next);
2723 * finish_task_switch - clean up after a task-switch
2724 * @rq: runqueue associated with task-switch
2725 * @prev: the thread we just switched away from.
2727 * finish_task_switch must be called after the context switch, paired
2728 * with a prepare_task_switch call before the context switch.
2729 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2730 * and do any other architecture-specific cleanup actions.
2732 * Note that we may have delayed dropping an mm in context_switch(). If
2733 * so, we finish that here outside of the runqueue lock. (Doing it
2734 * with the lock held can cause deadlocks; see schedule() for
2737 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2738 __releases(rq->lock)
2740 struct mm_struct *mm = rq->prev_mm;
2746 * A task struct has one reference for the use as "current".
2747 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2748 * schedule one last time. The schedule call will never return, and
2749 * the scheduled task must drop that reference.
2750 * The test for TASK_DEAD must occur while the runqueue locks are
2751 * still held, otherwise prev could be scheduled on another cpu, die
2752 * there before we look at prev->state, and then the reference would
2754 * Manfred Spraul <manfred@colorfullife.com>
2756 prev_state = prev->state;
2757 finish_arch_switch(prev);
2758 perf_event_task_sched_in(current, cpu_of(rq));
2759 finish_lock_switch(rq, prev);
2761 fire_sched_in_preempt_notifiers(current);
2764 if (unlikely(prev_state == TASK_DEAD)) {
2766 * Remove function-return probe instances associated with this
2767 * task and put them back on the free list.
2769 kprobe_flush_task(prev);
2770 put_task_struct(prev);
2776 /* assumes rq->lock is held */
2777 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2779 if (prev->sched_class->pre_schedule)
2780 prev->sched_class->pre_schedule(rq, prev);
2783 /* rq->lock is NOT held, but preemption is disabled */
2784 static inline void post_schedule(struct rq *rq)
2786 if (rq->post_schedule) {
2787 unsigned long flags;
2789 spin_lock_irqsave(&rq->lock, flags);
2790 if (rq->curr->sched_class->post_schedule)
2791 rq->curr->sched_class->post_schedule(rq);
2792 spin_unlock_irqrestore(&rq->lock, flags);
2794 rq->post_schedule = 0;
2800 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2804 static inline void post_schedule(struct rq *rq)
2811 * schedule_tail - first thing a freshly forked thread must call.
2812 * @prev: the thread we just switched away from.
2814 asmlinkage void schedule_tail(struct task_struct *prev)
2815 __releases(rq->lock)
2817 struct rq *rq = this_rq();
2819 finish_task_switch(rq, prev);
2822 * FIXME: do we need to worry about rq being invalidated by the
2827 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2828 /* In this case, finish_task_switch does not reenable preemption */
2831 if (current->set_child_tid)
2832 put_user(task_pid_vnr(current), current->set_child_tid);
2836 * context_switch - switch to the new MM and the new
2837 * thread's register state.
2840 context_switch(struct rq *rq, struct task_struct *prev,
2841 struct task_struct *next)
2843 struct mm_struct *mm, *oldmm;
2845 prepare_task_switch(rq, prev, next);
2846 trace_sched_switch(rq, prev, next);
2848 oldmm = prev->active_mm;
2850 * For paravirt, this is coupled with an exit in switch_to to
2851 * combine the page table reload and the switch backend into
2854 arch_start_context_switch(prev);
2857 next->active_mm = oldmm;
2858 atomic_inc(&oldmm->mm_count);
2859 enter_lazy_tlb(oldmm, next);
2861 switch_mm(oldmm, mm, next);
2863 if (likely(!prev->mm)) {
2864 prev->active_mm = NULL;
2865 rq->prev_mm = oldmm;
2868 * Since the runqueue lock will be released by the next
2869 * task (which is an invalid locking op but in the case
2870 * of the scheduler it's an obvious special-case), so we
2871 * do an early lockdep release here:
2873 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2874 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2877 /* Here we just switch the register state and the stack. */
2878 switch_to(prev, next, prev);
2882 * this_rq must be evaluated again because prev may have moved
2883 * CPUs since it called schedule(), thus the 'rq' on its stack
2884 * frame will be invalid.
2886 finish_task_switch(this_rq(), prev);
2890 * nr_running, nr_uninterruptible and nr_context_switches:
2892 * externally visible scheduler statistics: current number of runnable
2893 * threads, current number of uninterruptible-sleeping threads, total
2894 * number of context switches performed since bootup.
2896 unsigned long nr_running(void)
2898 unsigned long i, sum = 0;
2900 for_each_online_cpu(i)
2901 sum += cpu_rq(i)->nr_running;
2906 unsigned long nr_uninterruptible(void)
2908 unsigned long i, sum = 0;
2910 for_each_possible_cpu(i)
2911 sum += cpu_rq(i)->nr_uninterruptible;
2914 * Since we read the counters lockless, it might be slightly
2915 * inaccurate. Do not allow it to go below zero though:
2917 if (unlikely((long)sum < 0))
2923 unsigned long long nr_context_switches(void)
2926 unsigned long long sum = 0;
2928 for_each_possible_cpu(i)
2929 sum += cpu_rq(i)->nr_switches;
2934 unsigned long nr_iowait(void)
2936 unsigned long i, sum = 0;
2938 for_each_possible_cpu(i)
2939 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2944 unsigned long nr_iowait_cpu(void)
2946 struct rq *this = this_rq();
2947 return atomic_read(&this->nr_iowait);
2950 unsigned long this_cpu_load(void)
2952 struct rq *this = this_rq();
2953 return this->cpu_load[0];
2957 /* Variables and functions for calc_load */
2958 static atomic_long_t calc_load_tasks;
2959 static unsigned long calc_load_update;
2960 unsigned long avenrun[3];
2961 EXPORT_SYMBOL(avenrun);
2964 * get_avenrun - get the load average array
2965 * @loads: pointer to dest load array
2966 * @offset: offset to add
2967 * @shift: shift count to shift the result left
2969 * These values are estimates at best, so no need for locking.
2971 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2973 loads[0] = (avenrun[0] + offset) << shift;
2974 loads[1] = (avenrun[1] + offset) << shift;
2975 loads[2] = (avenrun[2] + offset) << shift;
2978 static unsigned long
2979 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2982 load += active * (FIXED_1 - exp);
2983 return load >> FSHIFT;
2987 * calc_load - update the avenrun load estimates 10 ticks after the
2988 * CPUs have updated calc_load_tasks.
2990 void calc_global_load(void)
2992 unsigned long upd = calc_load_update + 10;
2995 if (time_before(jiffies, upd))
2998 active = atomic_long_read(&calc_load_tasks);
2999 active = active > 0 ? active * FIXED_1 : 0;
3001 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3002 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3003 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3005 calc_load_update += LOAD_FREQ;
3009 * Either called from update_cpu_load() or from a cpu going idle
3011 static void calc_load_account_active(struct rq *this_rq)
3013 long nr_active, delta;
3015 nr_active = this_rq->nr_running;
3016 nr_active += (long) this_rq->nr_uninterruptible;
3018 if (nr_active != this_rq->calc_load_active) {
3019 delta = nr_active - this_rq->calc_load_active;
3020 this_rq->calc_load_active = nr_active;
3021 atomic_long_add(delta, &calc_load_tasks);
3026 * Update rq->cpu_load[] statistics. This function is usually called every
3027 * scheduler tick (TICK_NSEC).
3029 static void update_cpu_load(struct rq *this_rq)
3031 unsigned long this_load = this_rq->load.weight;
3034 this_rq->nr_load_updates++;
3036 /* Update our load: */
3037 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3038 unsigned long old_load, new_load;
3040 /* scale is effectively 1 << i now, and >> i divides by scale */
3042 old_load = this_rq->cpu_load[i];
3043 new_load = this_load;
3045 * Round up the averaging division if load is increasing. This
3046 * prevents us from getting stuck on 9 if the load is 10, for
3049 if (new_load > old_load)
3050 new_load += scale-1;
3051 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3054 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3055 this_rq->calc_load_update += LOAD_FREQ;
3056 calc_load_account_active(this_rq);
3063 * double_rq_lock - safely lock two runqueues
3065 * Note this does not disable interrupts like task_rq_lock,
3066 * you need to do so manually before calling.
3068 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3069 __acquires(rq1->lock)
3070 __acquires(rq2->lock)
3072 BUG_ON(!irqs_disabled());
3074 spin_lock(&rq1->lock);
3075 __acquire(rq2->lock); /* Fake it out ;) */
3078 spin_lock(&rq1->lock);
3079 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3081 spin_lock(&rq2->lock);
3082 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3085 update_rq_clock(rq1);
3086 update_rq_clock(rq2);
3090 * double_rq_unlock - safely unlock two runqueues
3092 * Note this does not restore interrupts like task_rq_unlock,
3093 * you need to do so manually after calling.
3095 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3096 __releases(rq1->lock)
3097 __releases(rq2->lock)
3099 spin_unlock(&rq1->lock);
3101 spin_unlock(&rq2->lock);
3103 __release(rq2->lock);
3107 * If dest_cpu is allowed for this process, migrate the task to it.
3108 * This is accomplished by forcing the cpu_allowed mask to only
3109 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3110 * the cpu_allowed mask is restored.
3112 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3114 struct migration_req req;
3115 unsigned long flags;
3118 rq = task_rq_lock(p, &flags);
3119 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3120 || unlikely(!cpu_active(dest_cpu)))
3123 /* force the process onto the specified CPU */
3124 if (migrate_task(p, dest_cpu, &req)) {
3125 /* Need to wait for migration thread (might exit: take ref). */
3126 struct task_struct *mt = rq->migration_thread;
3128 get_task_struct(mt);
3129 task_rq_unlock(rq, &flags);
3130 wake_up_process(mt);
3131 put_task_struct(mt);
3132 wait_for_completion(&req.done);
3137 task_rq_unlock(rq, &flags);
3141 * sched_exec - execve() is a valuable balancing opportunity, because at
3142 * this point the task has the smallest effective memory and cache footprint.
3144 void sched_exec(void)
3146 int new_cpu, this_cpu = get_cpu();
3147 new_cpu = select_task_rq(current, SD_BALANCE_EXEC, 0);
3149 if (new_cpu != this_cpu)
3150 sched_migrate_task(current, new_cpu);
3154 * pull_task - move a task from a remote runqueue to the local runqueue.
3155 * Both runqueues must be locked.
3157 static void pull_task(struct rq *src_rq, struct task_struct *p,
3158 struct rq *this_rq, int this_cpu)
3160 deactivate_task(src_rq, p, 0);
3161 set_task_cpu(p, this_cpu);
3162 activate_task(this_rq, p, 0);
3163 check_preempt_curr(this_rq, p, 0);
3167 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3170 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3171 struct sched_domain *sd, enum cpu_idle_type idle,
3174 int tsk_cache_hot = 0;
3176 * We do not migrate tasks that are:
3177 * 1) running (obviously), or
3178 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3179 * 3) are cache-hot on their current CPU.
3181 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3182 schedstat_inc(p, se.nr_failed_migrations_affine);
3187 if (task_running(rq, p)) {
3188 schedstat_inc(p, se.nr_failed_migrations_running);
3193 * Aggressive migration if:
3194 * 1) task is cache cold, or
3195 * 2) too many balance attempts have failed.
3198 tsk_cache_hot = task_hot(p, rq->clock, sd);
3199 if (!tsk_cache_hot ||
3200 sd->nr_balance_failed > sd->cache_nice_tries) {
3201 #ifdef CONFIG_SCHEDSTATS
3202 if (tsk_cache_hot) {
3203 schedstat_inc(sd, lb_hot_gained[idle]);
3204 schedstat_inc(p, se.nr_forced_migrations);
3210 if (tsk_cache_hot) {
3211 schedstat_inc(p, se.nr_failed_migrations_hot);
3217 static unsigned long
3218 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3219 unsigned long max_load_move, struct sched_domain *sd,
3220 enum cpu_idle_type idle, int *all_pinned,
3221 int *this_best_prio, struct rq_iterator *iterator)
3223 int loops = 0, pulled = 0, pinned = 0;
3224 struct task_struct *p;
3225 long rem_load_move = max_load_move;
3227 if (max_load_move == 0)
3233 * Start the load-balancing iterator:
3235 p = iterator->start(iterator->arg);
3237 if (!p || loops++ > sysctl_sched_nr_migrate)
3240 if ((p->se.load.weight >> 1) > rem_load_move ||
3241 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3242 p = iterator->next(iterator->arg);
3246 pull_task(busiest, p, this_rq, this_cpu);
3248 rem_load_move -= p->se.load.weight;
3250 #ifdef CONFIG_PREEMPT
3252 * NEWIDLE balancing is a source of latency, so preemptible kernels
3253 * will stop after the first task is pulled to minimize the critical
3256 if (idle == CPU_NEWLY_IDLE)
3261 * We only want to steal up to the prescribed amount of weighted load.
3263 if (rem_load_move > 0) {
3264 if (p->prio < *this_best_prio)
3265 *this_best_prio = p->prio;
3266 p = iterator->next(iterator->arg);
3271 * Right now, this is one of only two places pull_task() is called,
3272 * so we can safely collect pull_task() stats here rather than
3273 * inside pull_task().
3275 schedstat_add(sd, lb_gained[idle], pulled);
3278 *all_pinned = pinned;
3280 return max_load_move - rem_load_move;
3284 * move_tasks tries to move up to max_load_move weighted load from busiest to
3285 * this_rq, as part of a balancing operation within domain "sd".
3286 * Returns 1 if successful and 0 otherwise.
3288 * Called with both runqueues locked.
3290 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3291 unsigned long max_load_move,
3292 struct sched_domain *sd, enum cpu_idle_type idle,
3295 const struct sched_class *class = sched_class_highest;
3296 unsigned long total_load_moved = 0;
3297 int this_best_prio = this_rq->curr->prio;
3301 class->load_balance(this_rq, this_cpu, busiest,
3302 max_load_move - total_load_moved,
3303 sd, idle, all_pinned, &this_best_prio);
3304 class = class->next;
3306 #ifdef CONFIG_PREEMPT
3308 * NEWIDLE balancing is a source of latency, so preemptible
3309 * kernels will stop after the first task is pulled to minimize
3310 * the critical section.
3312 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3315 } while (class && max_load_move > total_load_moved);
3317 return total_load_moved > 0;
3321 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3322 struct sched_domain *sd, enum cpu_idle_type idle,
3323 struct rq_iterator *iterator)
3325 struct task_struct *p = iterator->start(iterator->arg);
3329 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3330 pull_task(busiest, p, this_rq, this_cpu);
3332 * Right now, this is only the second place pull_task()
3333 * is called, so we can safely collect pull_task()
3334 * stats here rather than inside pull_task().
3336 schedstat_inc(sd, lb_gained[idle]);
3340 p = iterator->next(iterator->arg);
3347 * move_one_task tries to move exactly one task from busiest to this_rq, as
3348 * part of active balancing operations within "domain".
3349 * Returns 1 if successful and 0 otherwise.
3351 * Called with both runqueues locked.
3353 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3354 struct sched_domain *sd, enum cpu_idle_type idle)
3356 const struct sched_class *class;
3358 for_each_class(class) {
3359 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3365 /********** Helpers for find_busiest_group ************************/
3367 * sd_lb_stats - Structure to store the statistics of a sched_domain
3368 * during load balancing.
3370 struct sd_lb_stats {
3371 struct sched_group *busiest; /* Busiest group in this sd */
3372 struct sched_group *this; /* Local group in this sd */
3373 unsigned long total_load; /* Total load of all groups in sd */
3374 unsigned long total_pwr; /* Total power of all groups in sd */
3375 unsigned long avg_load; /* Average load across all groups in sd */
3377 /** Statistics of this group */
3378 unsigned long this_load;
3379 unsigned long this_load_per_task;
3380 unsigned long this_nr_running;
3382 /* Statistics of the busiest group */
3383 unsigned long max_load;
3384 unsigned long busiest_load_per_task;
3385 unsigned long busiest_nr_running;
3387 int group_imb; /* Is there imbalance in this sd */
3388 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3389 int power_savings_balance; /* Is powersave balance needed for this sd */
3390 struct sched_group *group_min; /* Least loaded group in sd */
3391 struct sched_group *group_leader; /* Group which relieves group_min */
3392 unsigned long min_load_per_task; /* load_per_task in group_min */
3393 unsigned long leader_nr_running; /* Nr running of group_leader */
3394 unsigned long min_nr_running; /* Nr running of group_min */
3399 * sg_lb_stats - stats of a sched_group required for load_balancing
3401 struct sg_lb_stats {
3402 unsigned long avg_load; /*Avg load across the CPUs of the group */
3403 unsigned long group_load; /* Total load over the CPUs of the group */
3404 unsigned long sum_nr_running; /* Nr tasks running in the group */
3405 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3406 unsigned long group_capacity;
3407 int group_imb; /* Is there an imbalance in the group ? */
3411 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3412 * @group: The group whose first cpu is to be returned.
3414 static inline unsigned int group_first_cpu(struct sched_group *group)
3416 return cpumask_first(sched_group_cpus(group));
3420 * get_sd_load_idx - Obtain the load index for a given sched domain.
3421 * @sd: The sched_domain whose load_idx is to be obtained.
3422 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3424 static inline int get_sd_load_idx(struct sched_domain *sd,
3425 enum cpu_idle_type idle)
3431 load_idx = sd->busy_idx;
3434 case CPU_NEWLY_IDLE:
3435 load_idx = sd->newidle_idx;
3438 load_idx = sd->idle_idx;
3446 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3448 * init_sd_power_savings_stats - Initialize power savings statistics for
3449 * the given sched_domain, during load balancing.
3451 * @sd: Sched domain whose power-savings statistics are to be initialized.
3452 * @sds: Variable containing the statistics for sd.
3453 * @idle: Idle status of the CPU at which we're performing load-balancing.
3455 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3456 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3459 * Busy processors will not participate in power savings
3462 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3463 sds->power_savings_balance = 0;
3465 sds->power_savings_balance = 1;
3466 sds->min_nr_running = ULONG_MAX;
3467 sds->leader_nr_running = 0;
3472 * update_sd_power_savings_stats - Update the power saving stats for a
3473 * sched_domain while performing load balancing.
3475 * @group: sched_group belonging to the sched_domain under consideration.
3476 * @sds: Variable containing the statistics of the sched_domain
3477 * @local_group: Does group contain the CPU for which we're performing
3479 * @sgs: Variable containing the statistics of the group.
3481 static inline void update_sd_power_savings_stats(struct sched_group *group,
3482 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3485 if (!sds->power_savings_balance)
3489 * If the local group is idle or completely loaded
3490 * no need to do power savings balance at this domain
3492 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3493 !sds->this_nr_running))
3494 sds->power_savings_balance = 0;
3497 * If a group is already running at full capacity or idle,
3498 * don't include that group in power savings calculations
3500 if (!sds->power_savings_balance ||
3501 sgs->sum_nr_running >= sgs->group_capacity ||
3502 !sgs->sum_nr_running)
3506 * Calculate the group which has the least non-idle load.
3507 * This is the group from where we need to pick up the load
3510 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3511 (sgs->sum_nr_running == sds->min_nr_running &&
3512 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3513 sds->group_min = group;
3514 sds->min_nr_running = sgs->sum_nr_running;
3515 sds->min_load_per_task = sgs->sum_weighted_load /
3516 sgs->sum_nr_running;
3520 * Calculate the group which is almost near its
3521 * capacity but still has some space to pick up some load
3522 * from other group and save more power
3524 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3527 if (sgs->sum_nr_running > sds->leader_nr_running ||
3528 (sgs->sum_nr_running == sds->leader_nr_running &&
3529 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3530 sds->group_leader = group;
3531 sds->leader_nr_running = sgs->sum_nr_running;
3536 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3537 * @sds: Variable containing the statistics of the sched_domain
3538 * under consideration.
3539 * @this_cpu: Cpu at which we're currently performing load-balancing.
3540 * @imbalance: Variable to store the imbalance.
3543 * Check if we have potential to perform some power-savings balance.
3544 * If yes, set the busiest group to be the least loaded group in the
3545 * sched_domain, so that it's CPUs can be put to idle.
3547 * Returns 1 if there is potential to perform power-savings balance.
3550 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3551 int this_cpu, unsigned long *imbalance)
3553 if (!sds->power_savings_balance)
3556 if (sds->this != sds->group_leader ||
3557 sds->group_leader == sds->group_min)
3560 *imbalance = sds->min_load_per_task;
3561 sds->busiest = sds->group_min;
3566 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3567 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3568 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3573 static inline void update_sd_power_savings_stats(struct sched_group *group,
3574 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3579 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3580 int this_cpu, unsigned long *imbalance)
3584 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3587 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3589 return SCHED_LOAD_SCALE;
3592 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3594 return default_scale_freq_power(sd, cpu);
3597 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3599 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3600 unsigned long smt_gain = sd->smt_gain;
3607 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3609 return default_scale_smt_power(sd, cpu);
3612 unsigned long scale_rt_power(int cpu)
3614 struct rq *rq = cpu_rq(cpu);
3615 u64 total, available;
3617 sched_avg_update(rq);
3619 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3620 available = total - rq->rt_avg;
3622 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3623 total = SCHED_LOAD_SCALE;
3625 total >>= SCHED_LOAD_SHIFT;
3627 return div_u64(available, total);
3630 static void update_cpu_power(struct sched_domain *sd, int cpu)
3632 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3633 unsigned long power = SCHED_LOAD_SCALE;
3634 struct sched_group *sdg = sd->groups;
3636 if (sched_feat(ARCH_POWER))
3637 power *= arch_scale_freq_power(sd, cpu);
3639 power *= default_scale_freq_power(sd, cpu);
3641 power >>= SCHED_LOAD_SHIFT;
3643 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3644 if (sched_feat(ARCH_POWER))
3645 power *= arch_scale_smt_power(sd, cpu);
3647 power *= default_scale_smt_power(sd, cpu);
3649 power >>= SCHED_LOAD_SHIFT;
3652 power *= scale_rt_power(cpu);
3653 power >>= SCHED_LOAD_SHIFT;
3658 sdg->cpu_power = power;
3661 static void update_group_power(struct sched_domain *sd, int cpu)
3663 struct sched_domain *child = sd->child;
3664 struct sched_group *group, *sdg = sd->groups;
3665 unsigned long power;
3668 update_cpu_power(sd, cpu);
3674 group = child->groups;
3676 power += group->cpu_power;
3677 group = group->next;
3678 } while (group != child->groups);
3680 sdg->cpu_power = power;
3684 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3685 * @sd: The sched_domain whose statistics are to be updated.
3686 * @group: sched_group whose statistics are to be updated.
3687 * @this_cpu: Cpu for which load balance is currently performed.
3688 * @idle: Idle status of this_cpu
3689 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3690 * @sd_idle: Idle status of the sched_domain containing group.
3691 * @local_group: Does group contain this_cpu.
3692 * @cpus: Set of cpus considered for load balancing.
3693 * @balance: Should we balance.
3694 * @sgs: variable to hold the statistics for this group.
3696 static inline void update_sg_lb_stats(struct sched_domain *sd,
3697 struct sched_group *group, int this_cpu,
3698 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3699 int local_group, const struct cpumask *cpus,
3700 int *balance, struct sg_lb_stats *sgs)
3702 unsigned long load, max_cpu_load, min_cpu_load;
3704 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3705 unsigned long sum_avg_load_per_task;
3706 unsigned long avg_load_per_task;
3709 balance_cpu = group_first_cpu(group);
3710 if (balance_cpu == this_cpu)
3711 update_group_power(sd, this_cpu);
3714 /* Tally up the load of all CPUs in the group */
3715 sum_avg_load_per_task = avg_load_per_task = 0;
3717 min_cpu_load = ~0UL;
3719 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3720 struct rq *rq = cpu_rq(i);
3722 if (*sd_idle && rq->nr_running)
3725 /* Bias balancing toward cpus of our domain */
3727 if (idle_cpu(i) && !first_idle_cpu) {
3732 load = target_load(i, load_idx);
3734 load = source_load(i, load_idx);
3735 if (load > max_cpu_load)
3736 max_cpu_load = load;
3737 if (min_cpu_load > load)
3738 min_cpu_load = load;
3741 sgs->group_load += load;
3742 sgs->sum_nr_running += rq->nr_running;
3743 sgs->sum_weighted_load += weighted_cpuload(i);
3745 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3749 * First idle cpu or the first cpu(busiest) in this sched group
3750 * is eligible for doing load balancing at this and above
3751 * domains. In the newly idle case, we will allow all the cpu's
3752 * to do the newly idle load balance.
3754 if (idle != CPU_NEWLY_IDLE && local_group &&
3755 balance_cpu != this_cpu && balance) {
3760 /* Adjust by relative CPU power of the group */
3761 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3765 * Consider the group unbalanced when the imbalance is larger
3766 * than the average weight of two tasks.
3768 * APZ: with cgroup the avg task weight can vary wildly and
3769 * might not be a suitable number - should we keep a
3770 * normalized nr_running number somewhere that negates
3773 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3776 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3779 sgs->group_capacity =
3780 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3784 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3785 * @sd: sched_domain whose statistics are to be updated.
3786 * @this_cpu: Cpu for which load balance is currently performed.
3787 * @idle: Idle status of this_cpu
3788 * @sd_idle: Idle status of the sched_domain containing group.
3789 * @cpus: Set of cpus considered for load balancing.
3790 * @balance: Should we balance.
3791 * @sds: variable to hold the statistics for this sched_domain.
3793 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3794 enum cpu_idle_type idle, int *sd_idle,
3795 const struct cpumask *cpus, int *balance,
3796 struct sd_lb_stats *sds)
3798 struct sched_domain *child = sd->child;
3799 struct sched_group *group = sd->groups;
3800 struct sg_lb_stats sgs;
3801 int load_idx, prefer_sibling = 0;
3803 if (child && child->flags & SD_PREFER_SIBLING)
3806 init_sd_power_savings_stats(sd, sds, idle);
3807 load_idx = get_sd_load_idx(sd, idle);
3812 local_group = cpumask_test_cpu(this_cpu,
3813 sched_group_cpus(group));
3814 memset(&sgs, 0, sizeof(sgs));
3815 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3816 local_group, cpus, balance, &sgs);
3818 if (local_group && balance && !(*balance))
3821 sds->total_load += sgs.group_load;
3822 sds->total_pwr += group->cpu_power;
3825 * In case the child domain prefers tasks go to siblings
3826 * first, lower the group capacity to one so that we'll try
3827 * and move all the excess tasks away.
3830 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3833 sds->this_load = sgs.avg_load;
3835 sds->this_nr_running = sgs.sum_nr_running;
3836 sds->this_load_per_task = sgs.sum_weighted_load;
3837 } else if (sgs.avg_load > sds->max_load &&
3838 (sgs.sum_nr_running > sgs.group_capacity ||
3840 sds->max_load = sgs.avg_load;
3841 sds->busiest = group;
3842 sds->busiest_nr_running = sgs.sum_nr_running;
3843 sds->busiest_load_per_task = sgs.sum_weighted_load;
3844 sds->group_imb = sgs.group_imb;
3847 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3848 group = group->next;
3849 } while (group != sd->groups);
3853 * fix_small_imbalance - Calculate the minor imbalance that exists
3854 * amongst the groups of a sched_domain, during
3856 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3857 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3858 * @imbalance: Variable to store the imbalance.
3860 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3861 int this_cpu, unsigned long *imbalance)
3863 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3864 unsigned int imbn = 2;
3866 if (sds->this_nr_running) {
3867 sds->this_load_per_task /= sds->this_nr_running;
3868 if (sds->busiest_load_per_task >
3869 sds->this_load_per_task)
3872 sds->this_load_per_task =
3873 cpu_avg_load_per_task(this_cpu);
3875 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3876 sds->busiest_load_per_task * imbn) {
3877 *imbalance = sds->busiest_load_per_task;
3882 * OK, we don't have enough imbalance to justify moving tasks,
3883 * however we may be able to increase total CPU power used by
3887 pwr_now += sds->busiest->cpu_power *
3888 min(sds->busiest_load_per_task, sds->max_load);
3889 pwr_now += sds->this->cpu_power *
3890 min(sds->this_load_per_task, sds->this_load);
3891 pwr_now /= SCHED_LOAD_SCALE;
3893 /* Amount of load we'd subtract */
3894 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3895 sds->busiest->cpu_power;
3896 if (sds->max_load > tmp)
3897 pwr_move += sds->busiest->cpu_power *
3898 min(sds->busiest_load_per_task, sds->max_load - tmp);
3900 /* Amount of load we'd add */
3901 if (sds->max_load * sds->busiest->cpu_power <
3902 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3903 tmp = (sds->max_load * sds->busiest->cpu_power) /
3904 sds->this->cpu_power;
3906 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3907 sds->this->cpu_power;
3908 pwr_move += sds->this->cpu_power *
3909 min(sds->this_load_per_task, sds->this_load + tmp);
3910 pwr_move /= SCHED_LOAD_SCALE;
3912 /* Move if we gain throughput */
3913 if (pwr_move > pwr_now)
3914 *imbalance = sds->busiest_load_per_task;
3918 * calculate_imbalance - Calculate the amount of imbalance present within the
3919 * groups of a given sched_domain during load balance.
3920 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3921 * @this_cpu: Cpu for which currently load balance is being performed.
3922 * @imbalance: The variable to store the imbalance.
3924 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3925 unsigned long *imbalance)
3927 unsigned long max_pull;
3929 * In the presence of smp nice balancing, certain scenarios can have
3930 * max load less than avg load(as we skip the groups at or below
3931 * its cpu_power, while calculating max_load..)
3933 if (sds->max_load < sds->avg_load) {
3935 return fix_small_imbalance(sds, this_cpu, imbalance);
3938 /* Don't want to pull so many tasks that a group would go idle */
3939 max_pull = min(sds->max_load - sds->avg_load,
3940 sds->max_load - sds->busiest_load_per_task);
3942 /* How much load to actually move to equalise the imbalance */
3943 *imbalance = min(max_pull * sds->busiest->cpu_power,
3944 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3948 * if *imbalance is less than the average load per runnable task
3949 * there is no gaurantee that any tasks will be moved so we'll have
3950 * a think about bumping its value to force at least one task to be
3953 if (*imbalance < sds->busiest_load_per_task)
3954 return fix_small_imbalance(sds, this_cpu, imbalance);
3957 /******* find_busiest_group() helpers end here *********************/
3960 * find_busiest_group - Returns the busiest group within the sched_domain
3961 * if there is an imbalance. If there isn't an imbalance, and
3962 * the user has opted for power-savings, it returns a group whose
3963 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3964 * such a group exists.
3966 * Also calculates the amount of weighted load which should be moved
3967 * to restore balance.
3969 * @sd: The sched_domain whose busiest group is to be returned.
3970 * @this_cpu: The cpu for which load balancing is currently being performed.
3971 * @imbalance: Variable which stores amount of weighted load which should
3972 * be moved to restore balance/put a group to idle.
3973 * @idle: The idle status of this_cpu.
3974 * @sd_idle: The idleness of sd
3975 * @cpus: The set of CPUs under consideration for load-balancing.
3976 * @balance: Pointer to a variable indicating if this_cpu
3977 * is the appropriate cpu to perform load balancing at this_level.
3979 * Returns: - the busiest group if imbalance exists.
3980 * - If no imbalance and user has opted for power-savings balance,
3981 * return the least loaded group whose CPUs can be
3982 * put to idle by rebalancing its tasks onto our group.
3984 static struct sched_group *
3985 find_busiest_group(struct sched_domain *sd, int this_cpu,
3986 unsigned long *imbalance, enum cpu_idle_type idle,
3987 int *sd_idle, const struct cpumask *cpus, int *balance)
3989 struct sd_lb_stats sds;
3991 memset(&sds, 0, sizeof(sds));
3994 * Compute the various statistics relavent for load balancing at
3997 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4000 /* Cases where imbalance does not exist from POV of this_cpu */
4001 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4003 * 2) There is no busy sibling group to pull from.
4004 * 3) This group is the busiest group.
4005 * 4) This group is more busy than the avg busieness at this
4007 * 5) The imbalance is within the specified limit.
4008 * 6) Any rebalance would lead to ping-pong
4010 if (balance && !(*balance))
4013 if (!sds.busiest || sds.busiest_nr_running == 0)
4016 if (sds.this_load >= sds.max_load)
4019 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4021 if (sds.this_load >= sds.avg_load)
4024 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4027 sds.busiest_load_per_task /= sds.busiest_nr_running;
4029 sds.busiest_load_per_task =
4030 min(sds.busiest_load_per_task, sds.avg_load);
4033 * We're trying to get all the cpus to the average_load, so we don't
4034 * want to push ourselves above the average load, nor do we wish to
4035 * reduce the max loaded cpu below the average load, as either of these
4036 * actions would just result in more rebalancing later, and ping-pong
4037 * tasks around. Thus we look for the minimum possible imbalance.
4038 * Negative imbalances (*we* are more loaded than anyone else) will
4039 * be counted as no imbalance for these purposes -- we can't fix that
4040 * by pulling tasks to us. Be careful of negative numbers as they'll
4041 * appear as very large values with unsigned longs.
4043 if (sds.max_load <= sds.busiest_load_per_task)
4046 /* Looks like there is an imbalance. Compute it */
4047 calculate_imbalance(&sds, this_cpu, imbalance);
4052 * There is no obvious imbalance. But check if we can do some balancing
4055 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4063 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4066 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4067 unsigned long imbalance, const struct cpumask *cpus)
4069 struct rq *busiest = NULL, *rq;
4070 unsigned long max_load = 0;
4073 for_each_cpu(i, sched_group_cpus(group)) {
4074 unsigned long power = power_of(i);
4075 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4078 if (!cpumask_test_cpu(i, cpus))
4082 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4085 if (capacity && rq->nr_running == 1 && wl > imbalance)
4088 if (wl > max_load) {
4098 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4099 * so long as it is large enough.
4101 #define MAX_PINNED_INTERVAL 512
4103 /* Working cpumask for load_balance and load_balance_newidle. */
4104 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4107 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4108 * tasks if there is an imbalance.
4110 static int load_balance(int this_cpu, struct rq *this_rq,
4111 struct sched_domain *sd, enum cpu_idle_type idle,
4114 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4115 struct sched_group *group;
4116 unsigned long imbalance;
4118 unsigned long flags;
4119 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4121 cpumask_copy(cpus, cpu_active_mask);
4124 * When power savings policy is enabled for the parent domain, idle
4125 * sibling can pick up load irrespective of busy siblings. In this case,
4126 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4127 * portraying it as CPU_NOT_IDLE.
4129 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4130 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4133 schedstat_inc(sd, lb_count[idle]);
4137 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4144 schedstat_inc(sd, lb_nobusyg[idle]);
4148 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4150 schedstat_inc(sd, lb_nobusyq[idle]);
4154 BUG_ON(busiest == this_rq);
4156 schedstat_add(sd, lb_imbalance[idle], imbalance);
4159 if (busiest->nr_running > 1) {
4161 * Attempt to move tasks. If find_busiest_group has found
4162 * an imbalance but busiest->nr_running <= 1, the group is
4163 * still unbalanced. ld_moved simply stays zero, so it is
4164 * correctly treated as an imbalance.
4166 local_irq_save(flags);
4167 double_rq_lock(this_rq, busiest);
4168 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4169 imbalance, sd, idle, &all_pinned);
4170 double_rq_unlock(this_rq, busiest);
4171 local_irq_restore(flags);
4174 * some other cpu did the load balance for us.
4176 if (ld_moved && this_cpu != smp_processor_id())
4177 resched_cpu(this_cpu);
4179 /* All tasks on this runqueue were pinned by CPU affinity */
4180 if (unlikely(all_pinned)) {
4181 cpumask_clear_cpu(cpu_of(busiest), cpus);
4182 if (!cpumask_empty(cpus))
4189 schedstat_inc(sd, lb_failed[idle]);
4190 sd->nr_balance_failed++;
4192 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4194 spin_lock_irqsave(&busiest->lock, flags);
4196 /* don't kick the migration_thread, if the curr
4197 * task on busiest cpu can't be moved to this_cpu
4199 if (!cpumask_test_cpu(this_cpu,
4200 &busiest->curr->cpus_allowed)) {
4201 spin_unlock_irqrestore(&busiest->lock, flags);
4203 goto out_one_pinned;
4206 if (!busiest->active_balance) {
4207 busiest->active_balance = 1;
4208 busiest->push_cpu = this_cpu;
4211 spin_unlock_irqrestore(&busiest->lock, flags);
4213 wake_up_process(busiest->migration_thread);
4216 * We've kicked active balancing, reset the failure
4219 sd->nr_balance_failed = sd->cache_nice_tries+1;
4222 sd->nr_balance_failed = 0;
4224 if (likely(!active_balance)) {
4225 /* We were unbalanced, so reset the balancing interval */
4226 sd->balance_interval = sd->min_interval;
4229 * If we've begun active balancing, start to back off. This
4230 * case may not be covered by the all_pinned logic if there
4231 * is only 1 task on the busy runqueue (because we don't call
4234 if (sd->balance_interval < sd->max_interval)
4235 sd->balance_interval *= 2;
4238 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4239 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4245 schedstat_inc(sd, lb_balanced[idle]);
4247 sd->nr_balance_failed = 0;
4250 /* tune up the balancing interval */
4251 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4252 (sd->balance_interval < sd->max_interval))
4253 sd->balance_interval *= 2;
4255 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4256 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4267 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4268 * tasks if there is an imbalance.
4270 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4271 * this_rq is locked.
4274 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4276 struct sched_group *group;
4277 struct rq *busiest = NULL;
4278 unsigned long imbalance;
4282 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4284 cpumask_copy(cpus, cpu_active_mask);
4287 * When power savings policy is enabled for the parent domain, idle
4288 * sibling can pick up load irrespective of busy siblings. In this case,
4289 * let the state of idle sibling percolate up as IDLE, instead of
4290 * portraying it as CPU_NOT_IDLE.
4292 if (sd->flags & SD_SHARE_CPUPOWER &&
4293 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4296 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4298 update_shares_locked(this_rq, sd);
4299 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4300 &sd_idle, cpus, NULL);
4302 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4306 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4308 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4312 BUG_ON(busiest == this_rq);
4314 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4317 if (busiest->nr_running > 1) {
4318 /* Attempt to move tasks */
4319 double_lock_balance(this_rq, busiest);
4320 /* this_rq->clock is already updated */
4321 update_rq_clock(busiest);
4322 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4323 imbalance, sd, CPU_NEWLY_IDLE,
4325 double_unlock_balance(this_rq, busiest);
4327 if (unlikely(all_pinned)) {
4328 cpumask_clear_cpu(cpu_of(busiest), cpus);
4329 if (!cpumask_empty(cpus))
4335 int active_balance = 0;
4337 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4338 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4339 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4342 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4345 if (sd->nr_balance_failed++ < 2)
4349 * The only task running in a non-idle cpu can be moved to this
4350 * cpu in an attempt to completely freeup the other CPU
4351 * package. The same method used to move task in load_balance()
4352 * have been extended for load_balance_newidle() to speedup
4353 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4355 * The package power saving logic comes from
4356 * find_busiest_group(). If there are no imbalance, then
4357 * f_b_g() will return NULL. However when sched_mc={1,2} then
4358 * f_b_g() will select a group from which a running task may be
4359 * pulled to this cpu in order to make the other package idle.
4360 * If there is no opportunity to make a package idle and if
4361 * there are no imbalance, then f_b_g() will return NULL and no
4362 * action will be taken in load_balance_newidle().
4364 * Under normal task pull operation due to imbalance, there
4365 * will be more than one task in the source run queue and
4366 * move_tasks() will succeed. ld_moved will be true and this
4367 * active balance code will not be triggered.
4370 /* Lock busiest in correct order while this_rq is held */
4371 double_lock_balance(this_rq, busiest);
4374 * don't kick the migration_thread, if the curr
4375 * task on busiest cpu can't be moved to this_cpu
4377 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4378 double_unlock_balance(this_rq, busiest);
4383 if (!busiest->active_balance) {
4384 busiest->active_balance = 1;
4385 busiest->push_cpu = this_cpu;
4389 double_unlock_balance(this_rq, busiest);
4391 * Should not call ttwu while holding a rq->lock
4393 spin_unlock(&this_rq->lock);
4395 wake_up_process(busiest->migration_thread);
4396 spin_lock(&this_rq->lock);
4399 sd->nr_balance_failed = 0;
4401 update_shares_locked(this_rq, sd);
4405 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4406 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4407 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4409 sd->nr_balance_failed = 0;
4415 * idle_balance is called by schedule() if this_cpu is about to become
4416 * idle. Attempts to pull tasks from other CPUs.
4418 static void idle_balance(int this_cpu, struct rq *this_rq)
4420 struct sched_domain *sd;
4421 int pulled_task = 0;
4422 unsigned long next_balance = jiffies + HZ;
4424 this_rq->idle_stamp = this_rq->clock;
4426 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4429 for_each_domain(this_cpu, sd) {
4430 unsigned long interval;
4432 if (!(sd->flags & SD_LOAD_BALANCE))
4435 if (sd->flags & SD_BALANCE_NEWIDLE)
4436 /* If we've pulled tasks over stop searching: */
4437 pulled_task = load_balance_newidle(this_cpu, this_rq,
4440 interval = msecs_to_jiffies(sd->balance_interval);
4441 if (time_after(next_balance, sd->last_balance + interval))
4442 next_balance = sd->last_balance + interval;
4444 this_rq->idle_stamp = 0;
4448 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4450 * We are going idle. next_balance may be set based on
4451 * a busy processor. So reset next_balance.
4453 this_rq->next_balance = next_balance;
4458 * active_load_balance is run by migration threads. It pushes running tasks
4459 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4460 * running on each physical CPU where possible, and avoids physical /
4461 * logical imbalances.
4463 * Called with busiest_rq locked.
4465 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4467 int target_cpu = busiest_rq->push_cpu;
4468 struct sched_domain *sd;
4469 struct rq *target_rq;
4471 /* Is there any task to move? */
4472 if (busiest_rq->nr_running <= 1)
4475 target_rq = cpu_rq(target_cpu);
4478 * This condition is "impossible", if it occurs
4479 * we need to fix it. Originally reported by
4480 * Bjorn Helgaas on a 128-cpu setup.
4482 BUG_ON(busiest_rq == target_rq);
4484 /* move a task from busiest_rq to target_rq */
4485 double_lock_balance(busiest_rq, target_rq);
4486 update_rq_clock(busiest_rq);
4487 update_rq_clock(target_rq);
4489 /* Search for an sd spanning us and the target CPU. */
4490 for_each_domain(target_cpu, sd) {
4491 if ((sd->flags & SD_LOAD_BALANCE) &&
4492 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4497 schedstat_inc(sd, alb_count);
4499 if (move_one_task(target_rq, target_cpu, busiest_rq,
4501 schedstat_inc(sd, alb_pushed);
4503 schedstat_inc(sd, alb_failed);
4505 double_unlock_balance(busiest_rq, target_rq);
4510 atomic_t load_balancer;
4511 cpumask_var_t cpu_mask;
4512 cpumask_var_t ilb_grp_nohz_mask;
4513 } nohz ____cacheline_aligned = {
4514 .load_balancer = ATOMIC_INIT(-1),
4517 int get_nohz_load_balancer(void)
4519 return atomic_read(&nohz.load_balancer);
4522 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4524 * lowest_flag_domain - Return lowest sched_domain containing flag.
4525 * @cpu: The cpu whose lowest level of sched domain is to
4527 * @flag: The flag to check for the lowest sched_domain
4528 * for the given cpu.
4530 * Returns the lowest sched_domain of a cpu which contains the given flag.
4532 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4534 struct sched_domain *sd;
4536 for_each_domain(cpu, sd)
4537 if (sd && (sd->flags & flag))
4544 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4545 * @cpu: The cpu whose domains we're iterating over.
4546 * @sd: variable holding the value of the power_savings_sd
4548 * @flag: The flag to filter the sched_domains to be iterated.
4550 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4551 * set, starting from the lowest sched_domain to the highest.
4553 #define for_each_flag_domain(cpu, sd, flag) \
4554 for (sd = lowest_flag_domain(cpu, flag); \
4555 (sd && (sd->flags & flag)); sd = sd->parent)
4558 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4559 * @ilb_group: group to be checked for semi-idleness
4561 * Returns: 1 if the group is semi-idle. 0 otherwise.
4563 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4564 * and atleast one non-idle CPU. This helper function checks if the given
4565 * sched_group is semi-idle or not.
4567 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4569 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4570 sched_group_cpus(ilb_group));
4573 * A sched_group is semi-idle when it has atleast one busy cpu
4574 * and atleast one idle cpu.
4576 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4579 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4585 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4586 * @cpu: The cpu which is nominating a new idle_load_balancer.
4588 * Returns: Returns the id of the idle load balancer if it exists,
4589 * Else, returns >= nr_cpu_ids.
4591 * This algorithm picks the idle load balancer such that it belongs to a
4592 * semi-idle powersavings sched_domain. The idea is to try and avoid
4593 * completely idle packages/cores just for the purpose of idle load balancing
4594 * when there are other idle cpu's which are better suited for that job.
4596 static int find_new_ilb(int cpu)
4598 struct sched_domain *sd;
4599 struct sched_group *ilb_group;
4602 * Have idle load balancer selection from semi-idle packages only
4603 * when power-aware load balancing is enabled
4605 if (!(sched_smt_power_savings || sched_mc_power_savings))
4609 * Optimize for the case when we have no idle CPUs or only one
4610 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4612 if (cpumask_weight(nohz.cpu_mask) < 2)
4615 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4616 ilb_group = sd->groups;
4619 if (is_semi_idle_group(ilb_group))
4620 return cpumask_first(nohz.ilb_grp_nohz_mask);
4622 ilb_group = ilb_group->next;
4624 } while (ilb_group != sd->groups);
4628 return cpumask_first(nohz.cpu_mask);
4630 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4631 static inline int find_new_ilb(int call_cpu)
4633 return cpumask_first(nohz.cpu_mask);
4638 * This routine will try to nominate the ilb (idle load balancing)
4639 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4640 * load balancing on behalf of all those cpus. If all the cpus in the system
4641 * go into this tickless mode, then there will be no ilb owner (as there is
4642 * no need for one) and all the cpus will sleep till the next wakeup event
4645 * For the ilb owner, tick is not stopped. And this tick will be used
4646 * for idle load balancing. ilb owner will still be part of
4649 * While stopping the tick, this cpu will become the ilb owner if there
4650 * is no other owner. And will be the owner till that cpu becomes busy
4651 * or if all cpus in the system stop their ticks at which point
4652 * there is no need for ilb owner.
4654 * When the ilb owner becomes busy, it nominates another owner, during the
4655 * next busy scheduler_tick()
4657 int select_nohz_load_balancer(int stop_tick)
4659 int cpu = smp_processor_id();
4662 cpu_rq(cpu)->in_nohz_recently = 1;
4664 if (!cpu_active(cpu)) {
4665 if (atomic_read(&nohz.load_balancer) != cpu)
4669 * If we are going offline and still the leader,
4672 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4678 cpumask_set_cpu(cpu, nohz.cpu_mask);
4680 /* time for ilb owner also to sleep */
4681 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4682 if (atomic_read(&nohz.load_balancer) == cpu)
4683 atomic_set(&nohz.load_balancer, -1);
4687 if (atomic_read(&nohz.load_balancer) == -1) {
4688 /* make me the ilb owner */
4689 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4691 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4694 if (!(sched_smt_power_savings ||
4695 sched_mc_power_savings))
4698 * Check to see if there is a more power-efficient
4701 new_ilb = find_new_ilb(cpu);
4702 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4703 atomic_set(&nohz.load_balancer, -1);
4704 resched_cpu(new_ilb);
4710 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4713 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4715 if (atomic_read(&nohz.load_balancer) == cpu)
4716 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4723 static DEFINE_SPINLOCK(balancing);
4726 * It checks each scheduling domain to see if it is due to be balanced,
4727 * and initiates a balancing operation if so.
4729 * Balancing parameters are set up in arch_init_sched_domains.
4731 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4734 struct rq *rq = cpu_rq(cpu);
4735 unsigned long interval;
4736 struct sched_domain *sd;
4737 /* Earliest time when we have to do rebalance again */
4738 unsigned long next_balance = jiffies + 60*HZ;
4739 int update_next_balance = 0;
4742 for_each_domain(cpu, sd) {
4743 if (!(sd->flags & SD_LOAD_BALANCE))
4746 interval = sd->balance_interval;
4747 if (idle != CPU_IDLE)
4748 interval *= sd->busy_factor;
4750 /* scale ms to jiffies */
4751 interval = msecs_to_jiffies(interval);
4752 if (unlikely(!interval))
4754 if (interval > HZ*NR_CPUS/10)
4755 interval = HZ*NR_CPUS/10;
4757 need_serialize = sd->flags & SD_SERIALIZE;
4759 if (need_serialize) {
4760 if (!spin_trylock(&balancing))
4764 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4765 if (load_balance(cpu, rq, sd, idle, &balance)) {
4767 * We've pulled tasks over so either we're no
4768 * longer idle, or one of our SMT siblings is
4771 idle = CPU_NOT_IDLE;
4773 sd->last_balance = jiffies;
4776 spin_unlock(&balancing);
4778 if (time_after(next_balance, sd->last_balance + interval)) {
4779 next_balance = sd->last_balance + interval;
4780 update_next_balance = 1;
4784 * Stop the load balance at this level. There is another
4785 * CPU in our sched group which is doing load balancing more
4793 * next_balance will be updated only when there is a need.
4794 * When the cpu is attached to null domain for ex, it will not be
4797 if (likely(update_next_balance))
4798 rq->next_balance = next_balance;
4802 * run_rebalance_domains is triggered when needed from the scheduler tick.
4803 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4804 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4806 static void run_rebalance_domains(struct softirq_action *h)
4808 int this_cpu = smp_processor_id();
4809 struct rq *this_rq = cpu_rq(this_cpu);
4810 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4811 CPU_IDLE : CPU_NOT_IDLE;
4813 rebalance_domains(this_cpu, idle);
4817 * If this cpu is the owner for idle load balancing, then do the
4818 * balancing on behalf of the other idle cpus whose ticks are
4821 if (this_rq->idle_at_tick &&
4822 atomic_read(&nohz.load_balancer) == this_cpu) {
4826 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4827 if (balance_cpu == this_cpu)
4831 * If this cpu gets work to do, stop the load balancing
4832 * work being done for other cpus. Next load
4833 * balancing owner will pick it up.
4838 rebalance_domains(balance_cpu, CPU_IDLE);
4840 rq = cpu_rq(balance_cpu);
4841 if (time_after(this_rq->next_balance, rq->next_balance))
4842 this_rq->next_balance = rq->next_balance;
4848 static inline int on_null_domain(int cpu)
4850 return !rcu_dereference(cpu_rq(cpu)->sd);
4854 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4856 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4857 * idle load balancing owner or decide to stop the periodic load balancing,
4858 * if the whole system is idle.
4860 static inline void trigger_load_balance(struct rq *rq, int cpu)
4864 * If we were in the nohz mode recently and busy at the current
4865 * scheduler tick, then check if we need to nominate new idle
4868 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4869 rq->in_nohz_recently = 0;
4871 if (atomic_read(&nohz.load_balancer) == cpu) {
4872 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4873 atomic_set(&nohz.load_balancer, -1);
4876 if (atomic_read(&nohz.load_balancer) == -1) {
4877 int ilb = find_new_ilb(cpu);
4879 if (ilb < nr_cpu_ids)
4885 * If this cpu is idle and doing idle load balancing for all the
4886 * cpus with ticks stopped, is it time for that to stop?
4888 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4889 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4895 * If this cpu is idle and the idle load balancing is done by
4896 * someone else, then no need raise the SCHED_SOFTIRQ
4898 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4899 cpumask_test_cpu(cpu, nohz.cpu_mask))
4902 /* Don't need to rebalance while attached to NULL domain */
4903 if (time_after_eq(jiffies, rq->next_balance) &&
4904 likely(!on_null_domain(cpu)))
4905 raise_softirq(SCHED_SOFTIRQ);
4908 #else /* CONFIG_SMP */
4911 * on UP we do not need to balance between CPUs:
4913 static inline void idle_balance(int cpu, struct rq *rq)
4919 DEFINE_PER_CPU(struct kernel_stat, kstat);
4921 EXPORT_PER_CPU_SYMBOL(kstat);
4924 * Return any ns on the sched_clock that have not yet been accounted in
4925 * @p in case that task is currently running.
4927 * Called with task_rq_lock() held on @rq.
4929 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4933 if (task_current(rq, p)) {
4934 update_rq_clock(rq);
4935 ns = rq->clock - p->se.exec_start;
4943 unsigned long long task_delta_exec(struct task_struct *p)
4945 unsigned long flags;
4949 rq = task_rq_lock(p, &flags);
4950 ns = do_task_delta_exec(p, rq);
4951 task_rq_unlock(rq, &flags);
4957 * Return accounted runtime for the task.
4958 * In case the task is currently running, return the runtime plus current's
4959 * pending runtime that have not been accounted yet.
4961 unsigned long long task_sched_runtime(struct task_struct *p)
4963 unsigned long flags;
4967 rq = task_rq_lock(p, &flags);
4968 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4969 task_rq_unlock(rq, &flags);
4975 * Return sum_exec_runtime for the thread group.
4976 * In case the task is currently running, return the sum plus current's
4977 * pending runtime that have not been accounted yet.
4979 * Note that the thread group might have other running tasks as well,
4980 * so the return value not includes other pending runtime that other
4981 * running tasks might have.
4983 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4985 struct task_cputime totals;
4986 unsigned long flags;
4990 rq = task_rq_lock(p, &flags);
4991 thread_group_cputime(p, &totals);
4992 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4993 task_rq_unlock(rq, &flags);
4999 * Account user cpu time to a process.
5000 * @p: the process that the cpu time gets accounted to
5001 * @cputime: the cpu time spent in user space since the last update
5002 * @cputime_scaled: cputime scaled by cpu frequency
5004 void account_user_time(struct task_struct *p, cputime_t cputime,
5005 cputime_t cputime_scaled)
5007 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5010 /* Add user time to process. */
5011 p->utime = cputime_add(p->utime, cputime);
5012 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5013 account_group_user_time(p, cputime);
5015 /* Add user time to cpustat. */
5016 tmp = cputime_to_cputime64(cputime);
5017 if (TASK_NICE(p) > 0)
5018 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5020 cpustat->user = cputime64_add(cpustat->user, tmp);
5022 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5023 /* Account for user time used */
5024 acct_update_integrals(p);
5028 * Account guest cpu time to a process.
5029 * @p: the process that the cpu time gets accounted to
5030 * @cputime: the cpu time spent in virtual machine since the last update
5031 * @cputime_scaled: cputime scaled by cpu frequency
5033 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5034 cputime_t cputime_scaled)
5037 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5039 tmp = cputime_to_cputime64(cputime);
5041 /* Add guest time to process. */
5042 p->utime = cputime_add(p->utime, cputime);
5043 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5044 account_group_user_time(p, cputime);
5045 p->gtime = cputime_add(p->gtime, cputime);
5047 /* Add guest time to cpustat. */
5048 if (TASK_NICE(p) > 0) {
5049 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5050 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5052 cpustat->user = cputime64_add(cpustat->user, tmp);
5053 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5058 * Account system cpu time to a process.
5059 * @p: the process that the cpu time gets accounted to
5060 * @hardirq_offset: the offset to subtract from hardirq_count()
5061 * @cputime: the cpu time spent in kernel space since the last update
5062 * @cputime_scaled: cputime scaled by cpu frequency
5064 void account_system_time(struct task_struct *p, int hardirq_offset,
5065 cputime_t cputime, cputime_t cputime_scaled)
5067 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5070 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5071 account_guest_time(p, cputime, cputime_scaled);
5075 /* Add system time to process. */
5076 p->stime = cputime_add(p->stime, cputime);
5077 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5078 account_group_system_time(p, cputime);
5080 /* Add system time to cpustat. */
5081 tmp = cputime_to_cputime64(cputime);
5082 if (hardirq_count() - hardirq_offset)
5083 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5084 else if (softirq_count())
5085 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5087 cpustat->system = cputime64_add(cpustat->system, tmp);
5089 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5091 /* Account for system time used */
5092 acct_update_integrals(p);
5096 * Account for involuntary wait time.
5097 * @steal: the cpu time spent in involuntary wait
5099 void account_steal_time(cputime_t cputime)
5101 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5102 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5104 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5108 * Account for idle time.
5109 * @cputime: the cpu time spent in idle wait
5111 void account_idle_time(cputime_t cputime)
5113 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5114 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5115 struct rq *rq = this_rq();
5117 if (atomic_read(&rq->nr_iowait) > 0)
5118 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5120 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5123 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5126 * Account a single tick of cpu time.
5127 * @p: the process that the cpu time gets accounted to
5128 * @user_tick: indicates if the tick is a user or a system tick
5130 void account_process_tick(struct task_struct *p, int user_tick)
5132 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5133 struct rq *rq = this_rq();
5136 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5137 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5138 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5141 account_idle_time(cputime_one_jiffy);
5145 * Account multiple ticks of steal time.
5146 * @p: the process from which the cpu time has been stolen
5147 * @ticks: number of stolen ticks
5149 void account_steal_ticks(unsigned long ticks)
5151 account_steal_time(jiffies_to_cputime(ticks));
5155 * Account multiple ticks of idle time.
5156 * @ticks: number of stolen ticks
5158 void account_idle_ticks(unsigned long ticks)
5160 account_idle_time(jiffies_to_cputime(ticks));
5166 * Use precise platform statistics if available:
5168 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5169 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5175 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5177 struct task_cputime cputime;
5179 thread_group_cputime(p, &cputime);
5181 *ut = cputime.utime;
5182 *st = cputime.stime;
5186 #ifndef nsecs_to_cputime
5187 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5190 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5192 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5195 * Use CFS's precise accounting:
5197 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5202 temp = (u64)(rtime * utime);
5203 do_div(temp, total);
5204 utime = (cputime_t)temp;
5209 * Compare with previous values, to keep monotonicity:
5211 p->prev_utime = max(p->prev_utime, utime);
5212 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5214 *ut = p->prev_utime;
5215 *st = p->prev_stime;
5219 * Must be called with siglock held.
5221 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5223 struct signal_struct *sig = p->signal;
5224 struct task_cputime cputime;
5225 cputime_t rtime, utime, total;
5227 thread_group_cputime(p, &cputime);
5229 total = cputime_add(cputime.utime, cputime.stime);
5230 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5235 temp = (u64)(rtime * cputime.utime);
5236 do_div(temp, total);
5237 utime = (cputime_t)temp;
5241 sig->prev_utime = max(sig->prev_utime, utime);
5242 sig->prev_stime = max(sig->prev_stime,
5243 cputime_sub(rtime, sig->prev_utime));
5245 *ut = sig->prev_utime;
5246 *st = sig->prev_stime;
5251 * This function gets called by the timer code, with HZ frequency.
5252 * We call it with interrupts disabled.
5254 * It also gets called by the fork code, when changing the parent's
5257 void scheduler_tick(void)
5259 int cpu = smp_processor_id();
5260 struct rq *rq = cpu_rq(cpu);
5261 struct task_struct *curr = rq->curr;
5265 spin_lock(&rq->lock);
5266 update_rq_clock(rq);
5267 update_cpu_load(rq);
5268 curr->sched_class->task_tick(rq, curr, 0);
5269 spin_unlock(&rq->lock);
5271 perf_event_task_tick(curr, cpu);
5274 rq->idle_at_tick = idle_cpu(cpu);
5275 trigger_load_balance(rq, cpu);
5279 notrace unsigned long get_parent_ip(unsigned long addr)
5281 if (in_lock_functions(addr)) {
5282 addr = CALLER_ADDR2;
5283 if (in_lock_functions(addr))
5284 addr = CALLER_ADDR3;
5289 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5290 defined(CONFIG_PREEMPT_TRACER))
5292 void __kprobes add_preempt_count(int val)
5294 #ifdef CONFIG_DEBUG_PREEMPT
5298 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5301 preempt_count() += val;
5302 #ifdef CONFIG_DEBUG_PREEMPT
5304 * Spinlock count overflowing soon?
5306 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5309 if (preempt_count() == val)
5310 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5312 EXPORT_SYMBOL(add_preempt_count);
5314 void __kprobes sub_preempt_count(int val)
5316 #ifdef CONFIG_DEBUG_PREEMPT
5320 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5323 * Is the spinlock portion underflowing?
5325 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5326 !(preempt_count() & PREEMPT_MASK)))
5330 if (preempt_count() == val)
5331 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5332 preempt_count() -= val;
5334 EXPORT_SYMBOL(sub_preempt_count);
5339 * Print scheduling while atomic bug:
5341 static noinline void __schedule_bug(struct task_struct *prev)
5343 struct pt_regs *regs = get_irq_regs();
5345 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5346 prev->comm, prev->pid, preempt_count());
5348 debug_show_held_locks(prev);
5350 if (irqs_disabled())
5351 print_irqtrace_events(prev);
5360 * Various schedule()-time debugging checks and statistics:
5362 static inline void schedule_debug(struct task_struct *prev)
5365 * Test if we are atomic. Since do_exit() needs to call into
5366 * schedule() atomically, we ignore that path for now.
5367 * Otherwise, whine if we are scheduling when we should not be.
5369 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5370 __schedule_bug(prev);
5372 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5374 schedstat_inc(this_rq(), sched_count);
5375 #ifdef CONFIG_SCHEDSTATS
5376 if (unlikely(prev->lock_depth >= 0)) {
5377 schedstat_inc(this_rq(), bkl_count);
5378 schedstat_inc(prev, sched_info.bkl_count);
5383 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5385 if (prev->state == TASK_RUNNING) {
5386 u64 runtime = prev->se.sum_exec_runtime;
5388 runtime -= prev->se.prev_sum_exec_runtime;
5389 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5392 * In order to avoid avg_overlap growing stale when we are
5393 * indeed overlapping and hence not getting put to sleep, grow
5394 * the avg_overlap on preemption.
5396 * We use the average preemption runtime because that
5397 * correlates to the amount of cache footprint a task can
5400 update_avg(&prev->se.avg_overlap, runtime);
5402 prev->sched_class->put_prev_task(rq, prev);
5406 * Pick up the highest-prio task:
5408 static inline struct task_struct *
5409 pick_next_task(struct rq *rq)
5411 const struct sched_class *class;
5412 struct task_struct *p;
5415 * Optimization: we know that if all tasks are in
5416 * the fair class we can call that function directly:
5418 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5419 p = fair_sched_class.pick_next_task(rq);
5424 class = sched_class_highest;
5426 p = class->pick_next_task(rq);
5430 * Will never be NULL as the idle class always
5431 * returns a non-NULL p:
5433 class = class->next;
5438 * schedule() is the main scheduler function.
5440 asmlinkage void __sched schedule(void)
5442 struct task_struct *prev, *next;
5443 unsigned long *switch_count;
5449 cpu = smp_processor_id();
5453 switch_count = &prev->nivcsw;
5455 release_kernel_lock(prev);
5456 need_resched_nonpreemptible:
5458 schedule_debug(prev);
5460 if (sched_feat(HRTICK))
5463 spin_lock_irq(&rq->lock);
5464 update_rq_clock(rq);
5465 clear_tsk_need_resched(prev);
5467 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5468 if (unlikely(signal_pending_state(prev->state, prev)))
5469 prev->state = TASK_RUNNING;
5471 deactivate_task(rq, prev, 1);
5472 switch_count = &prev->nvcsw;
5475 pre_schedule(rq, prev);
5477 if (unlikely(!rq->nr_running))
5478 idle_balance(cpu, rq);
5480 put_prev_task(rq, prev);
5481 next = pick_next_task(rq);
5483 if (likely(prev != next)) {
5484 sched_info_switch(prev, next);
5485 perf_event_task_sched_out(prev, next, cpu);
5491 context_switch(rq, prev, next); /* unlocks the rq */
5493 * the context switch might have flipped the stack from under
5494 * us, hence refresh the local variables.
5496 cpu = smp_processor_id();
5499 spin_unlock_irq(&rq->lock);
5503 if (unlikely(reacquire_kernel_lock(current) < 0))
5504 goto need_resched_nonpreemptible;
5506 preempt_enable_no_resched();
5510 EXPORT_SYMBOL(schedule);
5512 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5514 * Look out! "owner" is an entirely speculative pointer
5515 * access and not reliable.
5517 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5522 if (!sched_feat(OWNER_SPIN))
5525 #ifdef CONFIG_DEBUG_PAGEALLOC
5527 * Need to access the cpu field knowing that
5528 * DEBUG_PAGEALLOC could have unmapped it if
5529 * the mutex owner just released it and exited.
5531 if (probe_kernel_address(&owner->cpu, cpu))
5538 * Even if the access succeeded (likely case),
5539 * the cpu field may no longer be valid.
5541 if (cpu >= nr_cpumask_bits)
5545 * We need to validate that we can do a
5546 * get_cpu() and that we have the percpu area.
5548 if (!cpu_online(cpu))
5555 * Owner changed, break to re-assess state.
5557 if (lock->owner != owner)
5561 * Is that owner really running on that cpu?
5563 if (task_thread_info(rq->curr) != owner || need_resched())
5573 #ifdef CONFIG_PREEMPT
5575 * this is the entry point to schedule() from in-kernel preemption
5576 * off of preempt_enable. Kernel preemptions off return from interrupt
5577 * occur there and call schedule directly.
5579 asmlinkage void __sched preempt_schedule(void)
5581 struct thread_info *ti = current_thread_info();
5584 * If there is a non-zero preempt_count or interrupts are disabled,
5585 * we do not want to preempt the current task. Just return..
5587 if (likely(ti->preempt_count || irqs_disabled()))
5591 add_preempt_count(PREEMPT_ACTIVE);
5593 sub_preempt_count(PREEMPT_ACTIVE);
5596 * Check again in case we missed a preemption opportunity
5597 * between schedule and now.
5600 } while (need_resched());
5602 EXPORT_SYMBOL(preempt_schedule);
5605 * this is the entry point to schedule() from kernel preemption
5606 * off of irq context.
5607 * Note, that this is called and return with irqs disabled. This will
5608 * protect us against recursive calling from irq.
5610 asmlinkage void __sched preempt_schedule_irq(void)
5612 struct thread_info *ti = current_thread_info();
5614 /* Catch callers which need to be fixed */
5615 BUG_ON(ti->preempt_count || !irqs_disabled());
5618 add_preempt_count(PREEMPT_ACTIVE);
5621 local_irq_disable();
5622 sub_preempt_count(PREEMPT_ACTIVE);
5625 * Check again in case we missed a preemption opportunity
5626 * between schedule and now.
5629 } while (need_resched());
5632 #endif /* CONFIG_PREEMPT */
5634 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5637 return try_to_wake_up(curr->private, mode, wake_flags);
5639 EXPORT_SYMBOL(default_wake_function);
5642 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5643 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5644 * number) then we wake all the non-exclusive tasks and one exclusive task.
5646 * There are circumstances in which we can try to wake a task which has already
5647 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5648 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5650 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5651 int nr_exclusive, int wake_flags, void *key)
5653 wait_queue_t *curr, *next;
5655 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5656 unsigned flags = curr->flags;
5658 if (curr->func(curr, mode, wake_flags, key) &&
5659 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5665 * __wake_up - wake up threads blocked on a waitqueue.
5667 * @mode: which threads
5668 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5669 * @key: is directly passed to the wakeup function
5671 * It may be assumed that this function implies a write memory barrier before
5672 * changing the task state if and only if any tasks are woken up.
5674 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5675 int nr_exclusive, void *key)
5677 unsigned long flags;
5679 spin_lock_irqsave(&q->lock, flags);
5680 __wake_up_common(q, mode, nr_exclusive, 0, key);
5681 spin_unlock_irqrestore(&q->lock, flags);
5683 EXPORT_SYMBOL(__wake_up);
5686 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5688 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5690 __wake_up_common(q, mode, 1, 0, NULL);
5693 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5695 __wake_up_common(q, mode, 1, 0, key);
5699 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5701 * @mode: which threads
5702 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5703 * @key: opaque value to be passed to wakeup targets
5705 * The sync wakeup differs that the waker knows that it will schedule
5706 * away soon, so while the target thread will be woken up, it will not
5707 * be migrated to another CPU - ie. the two threads are 'synchronized'
5708 * with each other. This can prevent needless bouncing between CPUs.
5710 * On UP it can prevent extra preemption.
5712 * It may be assumed that this function implies a write memory barrier before
5713 * changing the task state if and only if any tasks are woken up.
5715 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5716 int nr_exclusive, void *key)
5718 unsigned long flags;
5719 int wake_flags = WF_SYNC;
5724 if (unlikely(!nr_exclusive))
5727 spin_lock_irqsave(&q->lock, flags);
5728 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5729 spin_unlock_irqrestore(&q->lock, flags);
5731 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5734 * __wake_up_sync - see __wake_up_sync_key()
5736 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5738 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5740 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5743 * complete: - signals a single thread waiting on this completion
5744 * @x: holds the state of this particular completion
5746 * This will wake up a single thread waiting on this completion. Threads will be
5747 * awakened in the same order in which they were queued.
5749 * See also complete_all(), wait_for_completion() and related routines.
5751 * It may be assumed that this function implies a write memory barrier before
5752 * changing the task state if and only if any tasks are woken up.
5754 void complete(struct completion *x)
5756 unsigned long flags;
5758 spin_lock_irqsave(&x->wait.lock, flags);
5760 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5761 spin_unlock_irqrestore(&x->wait.lock, flags);
5763 EXPORT_SYMBOL(complete);
5766 * complete_all: - signals all threads waiting on this completion
5767 * @x: holds the state of this particular completion
5769 * This will wake up all threads waiting on this particular completion event.
5771 * It may be assumed that this function implies a write memory barrier before
5772 * changing the task state if and only if any tasks are woken up.
5774 void complete_all(struct completion *x)
5776 unsigned long flags;
5778 spin_lock_irqsave(&x->wait.lock, flags);
5779 x->done += UINT_MAX/2;
5780 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5781 spin_unlock_irqrestore(&x->wait.lock, flags);
5783 EXPORT_SYMBOL(complete_all);
5785 static inline long __sched
5786 do_wait_for_common(struct completion *x, long timeout, int state)
5789 DECLARE_WAITQUEUE(wait, current);
5791 wait.flags |= WQ_FLAG_EXCLUSIVE;
5792 __add_wait_queue_tail(&x->wait, &wait);
5794 if (signal_pending_state(state, current)) {
5795 timeout = -ERESTARTSYS;
5798 __set_current_state(state);
5799 spin_unlock_irq(&x->wait.lock);
5800 timeout = schedule_timeout(timeout);
5801 spin_lock_irq(&x->wait.lock);
5802 } while (!x->done && timeout);
5803 __remove_wait_queue(&x->wait, &wait);
5808 return timeout ?: 1;
5812 wait_for_common(struct completion *x, long timeout, int state)
5816 spin_lock_irq(&x->wait.lock);
5817 timeout = do_wait_for_common(x, timeout, state);
5818 spin_unlock_irq(&x->wait.lock);
5823 * wait_for_completion: - waits for completion of a task
5824 * @x: holds the state of this particular completion
5826 * This waits to be signaled for completion of a specific task. It is NOT
5827 * interruptible and there is no timeout.
5829 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5830 * and interrupt capability. Also see complete().
5832 void __sched wait_for_completion(struct completion *x)
5834 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5836 EXPORT_SYMBOL(wait_for_completion);
5839 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5840 * @x: holds the state of this particular completion
5841 * @timeout: timeout value in jiffies
5843 * This waits for either a completion of a specific task to be signaled or for a
5844 * specified timeout to expire. The timeout is in jiffies. It is not
5847 unsigned long __sched
5848 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5850 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5852 EXPORT_SYMBOL(wait_for_completion_timeout);
5855 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5856 * @x: holds the state of this particular completion
5858 * This waits for completion of a specific task to be signaled. It is
5861 int __sched wait_for_completion_interruptible(struct completion *x)
5863 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5864 if (t == -ERESTARTSYS)
5868 EXPORT_SYMBOL(wait_for_completion_interruptible);
5871 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5872 * @x: holds the state of this particular completion
5873 * @timeout: timeout value in jiffies
5875 * This waits for either a completion of a specific task to be signaled or for a
5876 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5878 unsigned long __sched
5879 wait_for_completion_interruptible_timeout(struct completion *x,
5880 unsigned long timeout)
5882 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5884 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5887 * wait_for_completion_killable: - waits for completion of a task (killable)
5888 * @x: holds the state of this particular completion
5890 * This waits to be signaled for completion of a specific task. It can be
5891 * interrupted by a kill signal.
5893 int __sched wait_for_completion_killable(struct completion *x)
5895 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5896 if (t == -ERESTARTSYS)
5900 EXPORT_SYMBOL(wait_for_completion_killable);
5903 * try_wait_for_completion - try to decrement a completion without blocking
5904 * @x: completion structure
5906 * Returns: 0 if a decrement cannot be done without blocking
5907 * 1 if a decrement succeeded.
5909 * If a completion is being used as a counting completion,
5910 * attempt to decrement the counter without blocking. This
5911 * enables us to avoid waiting if the resource the completion
5912 * is protecting is not available.
5914 bool try_wait_for_completion(struct completion *x)
5918 spin_lock_irq(&x->wait.lock);
5923 spin_unlock_irq(&x->wait.lock);
5926 EXPORT_SYMBOL(try_wait_for_completion);
5929 * completion_done - Test to see if a completion has any waiters
5930 * @x: completion structure
5932 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5933 * 1 if there are no waiters.
5936 bool completion_done(struct completion *x)
5940 spin_lock_irq(&x->wait.lock);
5943 spin_unlock_irq(&x->wait.lock);
5946 EXPORT_SYMBOL(completion_done);
5949 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5951 unsigned long flags;
5954 init_waitqueue_entry(&wait, current);
5956 __set_current_state(state);
5958 spin_lock_irqsave(&q->lock, flags);
5959 __add_wait_queue(q, &wait);
5960 spin_unlock(&q->lock);
5961 timeout = schedule_timeout(timeout);
5962 spin_lock_irq(&q->lock);
5963 __remove_wait_queue(q, &wait);
5964 spin_unlock_irqrestore(&q->lock, flags);
5969 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5971 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5973 EXPORT_SYMBOL(interruptible_sleep_on);
5976 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5978 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5980 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5982 void __sched sleep_on(wait_queue_head_t *q)
5984 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5986 EXPORT_SYMBOL(sleep_on);
5988 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5990 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5992 EXPORT_SYMBOL(sleep_on_timeout);
5994 #ifdef CONFIG_RT_MUTEXES
5997 * rt_mutex_setprio - set the current priority of a task
5999 * @prio: prio value (kernel-internal form)
6001 * This function changes the 'effective' priority of a task. It does
6002 * not touch ->normal_prio like __setscheduler().
6004 * Used by the rt_mutex code to implement priority inheritance logic.
6006 void rt_mutex_setprio(struct task_struct *p, int prio)
6008 unsigned long flags;
6009 int oldprio, on_rq, running;
6011 const struct sched_class *prev_class = p->sched_class;
6013 BUG_ON(prio < 0 || prio > MAX_PRIO);
6015 rq = task_rq_lock(p, &flags);
6016 update_rq_clock(rq);
6019 on_rq = p->se.on_rq;
6020 running = task_current(rq, p);
6022 dequeue_task(rq, p, 0);
6024 p->sched_class->put_prev_task(rq, p);
6027 p->sched_class = &rt_sched_class;
6029 p->sched_class = &fair_sched_class;
6034 p->sched_class->set_curr_task(rq);
6036 enqueue_task(rq, p, 0);
6038 check_class_changed(rq, p, prev_class, oldprio, running);
6040 task_rq_unlock(rq, &flags);
6045 void set_user_nice(struct task_struct *p, long nice)
6047 int old_prio, delta, on_rq;
6048 unsigned long flags;
6051 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6054 * We have to be careful, if called from sys_setpriority(),
6055 * the task might be in the middle of scheduling on another CPU.
6057 rq = task_rq_lock(p, &flags);
6058 update_rq_clock(rq);
6060 * The RT priorities are set via sched_setscheduler(), but we still
6061 * allow the 'normal' nice value to be set - but as expected
6062 * it wont have any effect on scheduling until the task is
6063 * SCHED_FIFO/SCHED_RR:
6065 if (task_has_rt_policy(p)) {
6066 p->static_prio = NICE_TO_PRIO(nice);
6069 on_rq = p->se.on_rq;
6071 dequeue_task(rq, p, 0);
6073 p->static_prio = NICE_TO_PRIO(nice);
6076 p->prio = effective_prio(p);
6077 delta = p->prio - old_prio;
6080 enqueue_task(rq, p, 0);
6082 * If the task increased its priority or is running and
6083 * lowered its priority, then reschedule its CPU:
6085 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6086 resched_task(rq->curr);
6089 task_rq_unlock(rq, &flags);
6091 EXPORT_SYMBOL(set_user_nice);
6094 * can_nice - check if a task can reduce its nice value
6098 int can_nice(const struct task_struct *p, const int nice)
6100 /* convert nice value [19,-20] to rlimit style value [1,40] */
6101 int nice_rlim = 20 - nice;
6103 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6104 capable(CAP_SYS_NICE));
6107 #ifdef __ARCH_WANT_SYS_NICE
6110 * sys_nice - change the priority of the current process.
6111 * @increment: priority increment
6113 * sys_setpriority is a more generic, but much slower function that
6114 * does similar things.
6116 SYSCALL_DEFINE1(nice, int, increment)
6121 * Setpriority might change our priority at the same moment.
6122 * We don't have to worry. Conceptually one call occurs first
6123 * and we have a single winner.
6125 if (increment < -40)
6130 nice = TASK_NICE(current) + increment;
6136 if (increment < 0 && !can_nice(current, nice))
6139 retval = security_task_setnice(current, nice);
6143 set_user_nice(current, nice);
6150 * task_prio - return the priority value of a given task.
6151 * @p: the task in question.
6153 * This is the priority value as seen by users in /proc.
6154 * RT tasks are offset by -200. Normal tasks are centered
6155 * around 0, value goes from -16 to +15.
6157 int task_prio(const struct task_struct *p)
6159 return p->prio - MAX_RT_PRIO;
6163 * task_nice - return the nice value of a given task.
6164 * @p: the task in question.
6166 int task_nice(const struct task_struct *p)
6168 return TASK_NICE(p);
6170 EXPORT_SYMBOL(task_nice);
6173 * idle_cpu - is a given cpu idle currently?
6174 * @cpu: the processor in question.
6176 int idle_cpu(int cpu)
6178 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6182 * idle_task - return the idle task for a given cpu.
6183 * @cpu: the processor in question.
6185 struct task_struct *idle_task(int cpu)
6187 return cpu_rq(cpu)->idle;
6191 * find_process_by_pid - find a process with a matching PID value.
6192 * @pid: the pid in question.
6194 static struct task_struct *find_process_by_pid(pid_t pid)
6196 return pid ? find_task_by_vpid(pid) : current;
6199 /* Actually do priority change: must hold rq lock. */
6201 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6203 BUG_ON(p->se.on_rq);
6206 p->rt_priority = prio;
6207 p->normal_prio = normal_prio(p);
6208 /* we are holding p->pi_lock already */
6209 p->prio = rt_mutex_getprio(p);
6210 if (rt_prio(p->prio))
6211 p->sched_class = &rt_sched_class;
6213 p->sched_class = &fair_sched_class;
6218 * check the target process has a UID that matches the current process's
6220 static bool check_same_owner(struct task_struct *p)
6222 const struct cred *cred = current_cred(), *pcred;
6226 pcred = __task_cred(p);
6227 match = (cred->euid == pcred->euid ||
6228 cred->euid == pcred->uid);
6233 static int __sched_setscheduler(struct task_struct *p, int policy,
6234 struct sched_param *param, bool user)
6236 int retval, oldprio, oldpolicy = -1, on_rq, running;
6237 unsigned long flags;
6238 const struct sched_class *prev_class = p->sched_class;
6242 /* may grab non-irq protected spin_locks */
6243 BUG_ON(in_interrupt());
6245 /* double check policy once rq lock held */
6247 reset_on_fork = p->sched_reset_on_fork;
6248 policy = oldpolicy = p->policy;
6250 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6251 policy &= ~SCHED_RESET_ON_FORK;
6253 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6254 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6255 policy != SCHED_IDLE)
6260 * Valid priorities for SCHED_FIFO and SCHED_RR are
6261 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6262 * SCHED_BATCH and SCHED_IDLE is 0.
6264 if (param->sched_priority < 0 ||
6265 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6266 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6268 if (rt_policy(policy) != (param->sched_priority != 0))
6272 * Allow unprivileged RT tasks to decrease priority:
6274 if (user && !capable(CAP_SYS_NICE)) {
6275 if (rt_policy(policy)) {
6276 unsigned long rlim_rtprio;
6278 if (!lock_task_sighand(p, &flags))
6280 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6281 unlock_task_sighand(p, &flags);
6283 /* can't set/change the rt policy */
6284 if (policy != p->policy && !rlim_rtprio)
6287 /* can't increase priority */
6288 if (param->sched_priority > p->rt_priority &&
6289 param->sched_priority > rlim_rtprio)
6293 * Like positive nice levels, dont allow tasks to
6294 * move out of SCHED_IDLE either:
6296 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6299 /* can't change other user's priorities */
6300 if (!check_same_owner(p))
6303 /* Normal users shall not reset the sched_reset_on_fork flag */
6304 if (p->sched_reset_on_fork && !reset_on_fork)
6309 #ifdef CONFIG_RT_GROUP_SCHED
6311 * Do not allow realtime tasks into groups that have no runtime
6314 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6315 task_group(p)->rt_bandwidth.rt_runtime == 0)
6319 retval = security_task_setscheduler(p, policy, param);
6325 * make sure no PI-waiters arrive (or leave) while we are
6326 * changing the priority of the task:
6328 spin_lock_irqsave(&p->pi_lock, flags);
6330 * To be able to change p->policy safely, the apropriate
6331 * runqueue lock must be held.
6333 rq = __task_rq_lock(p);
6334 /* recheck policy now with rq lock held */
6335 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6336 policy = oldpolicy = -1;
6337 __task_rq_unlock(rq);
6338 spin_unlock_irqrestore(&p->pi_lock, flags);
6341 update_rq_clock(rq);
6342 on_rq = p->se.on_rq;
6343 running = task_current(rq, p);
6345 deactivate_task(rq, p, 0);
6347 p->sched_class->put_prev_task(rq, p);
6349 p->sched_reset_on_fork = reset_on_fork;
6352 __setscheduler(rq, p, policy, param->sched_priority);
6355 p->sched_class->set_curr_task(rq);
6357 activate_task(rq, p, 0);
6359 check_class_changed(rq, p, prev_class, oldprio, running);
6361 __task_rq_unlock(rq);
6362 spin_unlock_irqrestore(&p->pi_lock, flags);
6364 rt_mutex_adjust_pi(p);
6370 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6371 * @p: the task in question.
6372 * @policy: new policy.
6373 * @param: structure containing the new RT priority.
6375 * NOTE that the task may be already dead.
6377 int sched_setscheduler(struct task_struct *p, int policy,
6378 struct sched_param *param)
6380 return __sched_setscheduler(p, policy, param, true);
6382 EXPORT_SYMBOL_GPL(sched_setscheduler);
6385 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6386 * @p: the task in question.
6387 * @policy: new policy.
6388 * @param: structure containing the new RT priority.
6390 * Just like sched_setscheduler, only don't bother checking if the
6391 * current context has permission. For example, this is needed in
6392 * stop_machine(): we create temporary high priority worker threads,
6393 * but our caller might not have that capability.
6395 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6396 struct sched_param *param)
6398 return __sched_setscheduler(p, policy, param, false);
6402 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6404 struct sched_param lparam;
6405 struct task_struct *p;
6408 if (!param || pid < 0)
6410 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6415 p = find_process_by_pid(pid);
6417 retval = sched_setscheduler(p, policy, &lparam);
6424 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6425 * @pid: the pid in question.
6426 * @policy: new policy.
6427 * @param: structure containing the new RT priority.
6429 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6430 struct sched_param __user *, param)
6432 /* negative values for policy are not valid */
6436 return do_sched_setscheduler(pid, policy, param);
6440 * sys_sched_setparam - set/change the RT priority of a thread
6441 * @pid: the pid in question.
6442 * @param: structure containing the new RT priority.
6444 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6446 return do_sched_setscheduler(pid, -1, param);
6450 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6451 * @pid: the pid in question.
6453 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6455 struct task_struct *p;
6462 read_lock(&tasklist_lock);
6463 p = find_process_by_pid(pid);
6465 retval = security_task_getscheduler(p);
6468 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6470 read_unlock(&tasklist_lock);
6475 * sys_sched_getparam - get the RT priority of a thread
6476 * @pid: the pid in question.
6477 * @param: structure containing the RT priority.
6479 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6481 struct sched_param lp;
6482 struct task_struct *p;
6485 if (!param || pid < 0)
6488 read_lock(&tasklist_lock);
6489 p = find_process_by_pid(pid);
6494 retval = security_task_getscheduler(p);
6498 lp.sched_priority = p->rt_priority;
6499 read_unlock(&tasklist_lock);
6502 * This one might sleep, we cannot do it with a spinlock held ...
6504 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6509 read_unlock(&tasklist_lock);
6513 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6515 cpumask_var_t cpus_allowed, new_mask;
6516 struct task_struct *p;
6520 read_lock(&tasklist_lock);
6522 p = find_process_by_pid(pid);
6524 read_unlock(&tasklist_lock);
6530 * It is not safe to call set_cpus_allowed with the
6531 * tasklist_lock held. We will bump the task_struct's
6532 * usage count and then drop tasklist_lock.
6535 read_unlock(&tasklist_lock);
6537 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6541 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6543 goto out_free_cpus_allowed;
6546 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6549 retval = security_task_setscheduler(p, 0, NULL);
6553 cpuset_cpus_allowed(p, cpus_allowed);
6554 cpumask_and(new_mask, in_mask, cpus_allowed);
6556 retval = set_cpus_allowed_ptr(p, new_mask);
6559 cpuset_cpus_allowed(p, cpus_allowed);
6560 if (!cpumask_subset(new_mask, cpus_allowed)) {
6562 * We must have raced with a concurrent cpuset
6563 * update. Just reset the cpus_allowed to the
6564 * cpuset's cpus_allowed
6566 cpumask_copy(new_mask, cpus_allowed);
6571 free_cpumask_var(new_mask);
6572 out_free_cpus_allowed:
6573 free_cpumask_var(cpus_allowed);
6580 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6581 struct cpumask *new_mask)
6583 if (len < cpumask_size())
6584 cpumask_clear(new_mask);
6585 else if (len > cpumask_size())
6586 len = cpumask_size();
6588 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6592 * sys_sched_setaffinity - set the cpu affinity of a process
6593 * @pid: pid of the process
6594 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6595 * @user_mask_ptr: user-space pointer to the new cpu mask
6597 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6598 unsigned long __user *, user_mask_ptr)
6600 cpumask_var_t new_mask;
6603 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6606 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6608 retval = sched_setaffinity(pid, new_mask);
6609 free_cpumask_var(new_mask);
6613 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6615 struct task_struct *p;
6616 unsigned long flags;
6621 read_lock(&tasklist_lock);
6624 p = find_process_by_pid(pid);
6628 retval = security_task_getscheduler(p);
6632 rq = task_rq_lock(p, &flags);
6633 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6634 task_rq_unlock(rq, &flags);
6637 read_unlock(&tasklist_lock);
6644 * sys_sched_getaffinity - get the cpu affinity of a process
6645 * @pid: pid of the process
6646 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6647 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6649 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6650 unsigned long __user *, user_mask_ptr)
6655 if (len < cpumask_size())
6658 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6661 ret = sched_getaffinity(pid, mask);
6663 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6666 ret = cpumask_size();
6668 free_cpumask_var(mask);
6674 * sys_sched_yield - yield the current processor to other threads.
6676 * This function yields the current CPU to other tasks. If there are no
6677 * other threads running on this CPU then this function will return.
6679 SYSCALL_DEFINE0(sched_yield)
6681 struct rq *rq = this_rq_lock();
6683 schedstat_inc(rq, yld_count);
6684 current->sched_class->yield_task(rq);
6687 * Since we are going to call schedule() anyway, there's
6688 * no need to preempt or enable interrupts:
6690 __release(rq->lock);
6691 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6692 _raw_spin_unlock(&rq->lock);
6693 preempt_enable_no_resched();
6700 static inline int should_resched(void)
6702 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6705 static void __cond_resched(void)
6707 add_preempt_count(PREEMPT_ACTIVE);
6709 sub_preempt_count(PREEMPT_ACTIVE);
6712 int __sched _cond_resched(void)
6714 if (should_resched()) {
6720 EXPORT_SYMBOL(_cond_resched);
6723 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6724 * call schedule, and on return reacquire the lock.
6726 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6727 * operations here to prevent schedule() from being called twice (once via
6728 * spin_unlock(), once by hand).
6730 int __cond_resched_lock(spinlock_t *lock)
6732 int resched = should_resched();
6735 lockdep_assert_held(lock);
6737 if (spin_needbreak(lock) || resched) {
6748 EXPORT_SYMBOL(__cond_resched_lock);
6750 int __sched __cond_resched_softirq(void)
6752 BUG_ON(!in_softirq());
6754 if (should_resched()) {
6762 EXPORT_SYMBOL(__cond_resched_softirq);
6765 * yield - yield the current processor to other threads.
6767 * This is a shortcut for kernel-space yielding - it marks the
6768 * thread runnable and calls sys_sched_yield().
6770 void __sched yield(void)
6772 set_current_state(TASK_RUNNING);
6775 EXPORT_SYMBOL(yield);
6778 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6779 * that process accounting knows that this is a task in IO wait state.
6781 void __sched io_schedule(void)
6783 struct rq *rq = raw_rq();
6785 delayacct_blkio_start();
6786 atomic_inc(&rq->nr_iowait);
6787 current->in_iowait = 1;
6789 current->in_iowait = 0;
6790 atomic_dec(&rq->nr_iowait);
6791 delayacct_blkio_end();
6793 EXPORT_SYMBOL(io_schedule);
6795 long __sched io_schedule_timeout(long timeout)
6797 struct rq *rq = raw_rq();
6800 delayacct_blkio_start();
6801 atomic_inc(&rq->nr_iowait);
6802 current->in_iowait = 1;
6803 ret = schedule_timeout(timeout);
6804 current->in_iowait = 0;
6805 atomic_dec(&rq->nr_iowait);
6806 delayacct_blkio_end();
6811 * sys_sched_get_priority_max - return maximum RT priority.
6812 * @policy: scheduling class.
6814 * this syscall returns the maximum rt_priority that can be used
6815 * by a given scheduling class.
6817 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6824 ret = MAX_USER_RT_PRIO-1;
6836 * sys_sched_get_priority_min - return minimum RT priority.
6837 * @policy: scheduling class.
6839 * this syscall returns the minimum rt_priority that can be used
6840 * by a given scheduling class.
6842 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6860 * sys_sched_rr_get_interval - return the default timeslice of a process.
6861 * @pid: pid of the process.
6862 * @interval: userspace pointer to the timeslice value.
6864 * this syscall writes the default timeslice value of a given process
6865 * into the user-space timespec buffer. A value of '0' means infinity.
6867 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6868 struct timespec __user *, interval)
6870 struct task_struct *p;
6871 unsigned int time_slice;
6872 unsigned long flags;
6881 read_lock(&tasklist_lock);
6882 p = find_process_by_pid(pid);
6886 retval = security_task_getscheduler(p);
6890 rq = task_rq_lock(p, &flags);
6891 time_slice = p->sched_class->get_rr_interval(rq, p);
6892 task_rq_unlock(rq, &flags);
6894 read_unlock(&tasklist_lock);
6895 jiffies_to_timespec(time_slice, &t);
6896 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6900 read_unlock(&tasklist_lock);
6904 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6906 void sched_show_task(struct task_struct *p)
6908 unsigned long free = 0;
6911 state = p->state ? __ffs(p->state) + 1 : 0;
6912 printk(KERN_INFO "%-13.13s %c", p->comm,
6913 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6914 #if BITS_PER_LONG == 32
6915 if (state == TASK_RUNNING)
6916 printk(KERN_CONT " running ");
6918 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6920 if (state == TASK_RUNNING)
6921 printk(KERN_CONT " running task ");
6923 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6925 #ifdef CONFIG_DEBUG_STACK_USAGE
6926 free = stack_not_used(p);
6928 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6929 task_pid_nr(p), task_pid_nr(p->real_parent),
6930 (unsigned long)task_thread_info(p)->flags);
6932 show_stack(p, NULL);
6935 void show_state_filter(unsigned long state_filter)
6937 struct task_struct *g, *p;
6939 #if BITS_PER_LONG == 32
6941 " task PC stack pid father\n");
6944 " task PC stack pid father\n");
6946 read_lock(&tasklist_lock);
6947 do_each_thread(g, p) {
6949 * reset the NMI-timeout, listing all files on a slow
6950 * console might take alot of time:
6952 touch_nmi_watchdog();
6953 if (!state_filter || (p->state & state_filter))
6955 } while_each_thread(g, p);
6957 touch_all_softlockup_watchdogs();
6959 #ifdef CONFIG_SCHED_DEBUG
6960 sysrq_sched_debug_show();
6962 read_unlock(&tasklist_lock);
6964 * Only show locks if all tasks are dumped:
6967 debug_show_all_locks();
6970 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6972 idle->sched_class = &idle_sched_class;
6976 * init_idle - set up an idle thread for a given CPU
6977 * @idle: task in question
6978 * @cpu: cpu the idle task belongs to
6980 * NOTE: this function does not set the idle thread's NEED_RESCHED
6981 * flag, to make booting more robust.
6983 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6985 struct rq *rq = cpu_rq(cpu);
6986 unsigned long flags;
6988 spin_lock_irqsave(&rq->lock, flags);
6991 idle->se.exec_start = sched_clock();
6993 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6994 __set_task_cpu(idle, cpu);
6996 rq->curr = rq->idle = idle;
6997 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7000 spin_unlock_irqrestore(&rq->lock, flags);
7002 /* Set the preempt count _outside_ the spinlocks! */
7003 #if defined(CONFIG_PREEMPT)
7004 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7006 task_thread_info(idle)->preempt_count = 0;
7009 * The idle tasks have their own, simple scheduling class:
7011 idle->sched_class = &idle_sched_class;
7012 ftrace_graph_init_task(idle);
7016 * In a system that switches off the HZ timer nohz_cpu_mask
7017 * indicates which cpus entered this state. This is used
7018 * in the rcu update to wait only for active cpus. For system
7019 * which do not switch off the HZ timer nohz_cpu_mask should
7020 * always be CPU_BITS_NONE.
7022 cpumask_var_t nohz_cpu_mask;
7025 * Increase the granularity value when there are more CPUs,
7026 * because with more CPUs the 'effective latency' as visible
7027 * to users decreases. But the relationship is not linear,
7028 * so pick a second-best guess by going with the log2 of the
7031 * This idea comes from the SD scheduler of Con Kolivas:
7033 static void update_sysctl(void)
7035 unsigned int cpus = min(num_online_cpus(), 8U);
7036 unsigned int factor = 1 + ilog2(cpus);
7038 #define SET_SYSCTL(name) \
7039 (sysctl_##name = (factor) * normalized_sysctl_##name)
7040 SET_SYSCTL(sched_min_granularity);
7041 SET_SYSCTL(sched_latency);
7042 SET_SYSCTL(sched_wakeup_granularity);
7043 SET_SYSCTL(sched_shares_ratelimit);
7047 static inline void sched_init_granularity(void)
7054 * This is how migration works:
7056 * 1) we queue a struct migration_req structure in the source CPU's
7057 * runqueue and wake up that CPU's migration thread.
7058 * 2) we down() the locked semaphore => thread blocks.
7059 * 3) migration thread wakes up (implicitly it forces the migrated
7060 * thread off the CPU)
7061 * 4) it gets the migration request and checks whether the migrated
7062 * task is still in the wrong runqueue.
7063 * 5) if it's in the wrong runqueue then the migration thread removes
7064 * it and puts it into the right queue.
7065 * 6) migration thread up()s the semaphore.
7066 * 7) we wake up and the migration is done.
7070 * Change a given task's CPU affinity. Migrate the thread to a
7071 * proper CPU and schedule it away if the CPU it's executing on
7072 * is removed from the allowed bitmask.
7074 * NOTE: the caller must have a valid reference to the task, the
7075 * task must not exit() & deallocate itself prematurely. The
7076 * call is not atomic; no spinlocks may be held.
7078 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7080 struct migration_req req;
7081 unsigned long flags;
7085 rq = task_rq_lock(p, &flags);
7086 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7091 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7092 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7097 if (p->sched_class->set_cpus_allowed)
7098 p->sched_class->set_cpus_allowed(p, new_mask);
7100 cpumask_copy(&p->cpus_allowed, new_mask);
7101 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7104 /* Can the task run on the task's current CPU? If so, we're done */
7105 if (cpumask_test_cpu(task_cpu(p), new_mask))
7108 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7109 /* Need help from migration thread: drop lock and wait. */
7110 struct task_struct *mt = rq->migration_thread;
7112 get_task_struct(mt);
7113 task_rq_unlock(rq, &flags);
7114 wake_up_process(rq->migration_thread);
7115 put_task_struct(mt);
7116 wait_for_completion(&req.done);
7117 tlb_migrate_finish(p->mm);
7121 task_rq_unlock(rq, &flags);
7125 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7128 * Move (not current) task off this cpu, onto dest cpu. We're doing
7129 * this because either it can't run here any more (set_cpus_allowed()
7130 * away from this CPU, or CPU going down), or because we're
7131 * attempting to rebalance this task on exec (sched_exec).
7133 * So we race with normal scheduler movements, but that's OK, as long
7134 * as the task is no longer on this CPU.
7136 * Returns non-zero if task was successfully migrated.
7138 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7140 struct rq *rq_dest, *rq_src;
7143 if (unlikely(!cpu_active(dest_cpu)))
7146 rq_src = cpu_rq(src_cpu);
7147 rq_dest = cpu_rq(dest_cpu);
7149 double_rq_lock(rq_src, rq_dest);
7150 /* Already moved. */
7151 if (task_cpu(p) != src_cpu)
7153 /* Affinity changed (again). */
7154 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7157 on_rq = p->se.on_rq;
7159 deactivate_task(rq_src, p, 0);
7161 set_task_cpu(p, dest_cpu);
7163 activate_task(rq_dest, p, 0);
7164 check_preempt_curr(rq_dest, p, 0);
7169 double_rq_unlock(rq_src, rq_dest);
7173 #define RCU_MIGRATION_IDLE 0
7174 #define RCU_MIGRATION_NEED_QS 1
7175 #define RCU_MIGRATION_GOT_QS 2
7176 #define RCU_MIGRATION_MUST_SYNC 3
7179 * migration_thread - this is a highprio system thread that performs
7180 * thread migration by bumping thread off CPU then 'pushing' onto
7183 static int migration_thread(void *data)
7186 int cpu = (long)data;
7190 BUG_ON(rq->migration_thread != current);
7192 set_current_state(TASK_INTERRUPTIBLE);
7193 while (!kthread_should_stop()) {
7194 struct migration_req *req;
7195 struct list_head *head;
7197 spin_lock_irq(&rq->lock);
7199 if (cpu_is_offline(cpu)) {
7200 spin_unlock_irq(&rq->lock);
7204 if (rq->active_balance) {
7205 active_load_balance(rq, cpu);
7206 rq->active_balance = 0;
7209 head = &rq->migration_queue;
7211 if (list_empty(head)) {
7212 spin_unlock_irq(&rq->lock);
7214 set_current_state(TASK_INTERRUPTIBLE);
7217 req = list_entry(head->next, struct migration_req, list);
7218 list_del_init(head->next);
7220 if (req->task != NULL) {
7221 spin_unlock(&rq->lock);
7222 __migrate_task(req->task, cpu, req->dest_cpu);
7223 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7224 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7225 spin_unlock(&rq->lock);
7227 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7228 spin_unlock(&rq->lock);
7229 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7233 complete(&req->done);
7235 __set_current_state(TASK_RUNNING);
7240 #ifdef CONFIG_HOTPLUG_CPU
7242 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7246 local_irq_disable();
7247 ret = __migrate_task(p, src_cpu, dest_cpu);
7253 * Figure out where task on dead CPU should go, use force if necessary.
7255 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7258 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7261 /* Look for allowed, online CPU in same node. */
7262 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
7263 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7266 /* Any allowed, online CPU? */
7267 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
7268 if (dest_cpu < nr_cpu_ids)
7271 /* No more Mr. Nice Guy. */
7272 if (dest_cpu >= nr_cpu_ids) {
7273 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7274 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
7277 * Don't tell them about moving exiting tasks or
7278 * kernel threads (both mm NULL), since they never
7281 if (p->mm && printk_ratelimit()) {
7282 printk(KERN_INFO "process %d (%s) no "
7283 "longer affine to cpu%d\n",
7284 task_pid_nr(p), p->comm, dead_cpu);
7289 /* It can have affinity changed while we were choosing. */
7290 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7295 * While a dead CPU has no uninterruptible tasks queued at this point,
7296 * it might still have a nonzero ->nr_uninterruptible counter, because
7297 * for performance reasons the counter is not stricly tracking tasks to
7298 * their home CPUs. So we just add the counter to another CPU's counter,
7299 * to keep the global sum constant after CPU-down:
7301 static void migrate_nr_uninterruptible(struct rq *rq_src)
7303 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7304 unsigned long flags;
7306 local_irq_save(flags);
7307 double_rq_lock(rq_src, rq_dest);
7308 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7309 rq_src->nr_uninterruptible = 0;
7310 double_rq_unlock(rq_src, rq_dest);
7311 local_irq_restore(flags);
7314 /* Run through task list and migrate tasks from the dead cpu. */
7315 static void migrate_live_tasks(int src_cpu)
7317 struct task_struct *p, *t;
7319 read_lock(&tasklist_lock);
7321 do_each_thread(t, p) {
7325 if (task_cpu(p) == src_cpu)
7326 move_task_off_dead_cpu(src_cpu, p);
7327 } while_each_thread(t, p);
7329 read_unlock(&tasklist_lock);
7333 * Schedules idle task to be the next runnable task on current CPU.
7334 * It does so by boosting its priority to highest possible.
7335 * Used by CPU offline code.
7337 void sched_idle_next(void)
7339 int this_cpu = smp_processor_id();
7340 struct rq *rq = cpu_rq(this_cpu);
7341 struct task_struct *p = rq->idle;
7342 unsigned long flags;
7344 /* cpu has to be offline */
7345 BUG_ON(cpu_online(this_cpu));
7348 * Strictly not necessary since rest of the CPUs are stopped by now
7349 * and interrupts disabled on the current cpu.
7351 spin_lock_irqsave(&rq->lock, flags);
7353 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7355 update_rq_clock(rq);
7356 activate_task(rq, p, 0);
7358 spin_unlock_irqrestore(&rq->lock, flags);
7362 * Ensures that the idle task is using init_mm right before its cpu goes
7365 void idle_task_exit(void)
7367 struct mm_struct *mm = current->active_mm;
7369 BUG_ON(cpu_online(smp_processor_id()));
7372 switch_mm(mm, &init_mm, current);
7376 /* called under rq->lock with disabled interrupts */
7377 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7379 struct rq *rq = cpu_rq(dead_cpu);
7381 /* Must be exiting, otherwise would be on tasklist. */
7382 BUG_ON(!p->exit_state);
7384 /* Cannot have done final schedule yet: would have vanished. */
7385 BUG_ON(p->state == TASK_DEAD);
7390 * Drop lock around migration; if someone else moves it,
7391 * that's OK. No task can be added to this CPU, so iteration is
7394 spin_unlock_irq(&rq->lock);
7395 move_task_off_dead_cpu(dead_cpu, p);
7396 spin_lock_irq(&rq->lock);
7401 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7402 static void migrate_dead_tasks(unsigned int dead_cpu)
7404 struct rq *rq = cpu_rq(dead_cpu);
7405 struct task_struct *next;
7408 if (!rq->nr_running)
7410 update_rq_clock(rq);
7411 next = pick_next_task(rq);
7414 next->sched_class->put_prev_task(rq, next);
7415 migrate_dead(dead_cpu, next);
7421 * remove the tasks which were accounted by rq from calc_load_tasks.
7423 static void calc_global_load_remove(struct rq *rq)
7425 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7426 rq->calc_load_active = 0;
7428 #endif /* CONFIG_HOTPLUG_CPU */
7430 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7432 static struct ctl_table sd_ctl_dir[] = {
7434 .procname = "sched_domain",
7440 static struct ctl_table sd_ctl_root[] = {
7442 .ctl_name = CTL_KERN,
7443 .procname = "kernel",
7445 .child = sd_ctl_dir,
7450 static struct ctl_table *sd_alloc_ctl_entry(int n)
7452 struct ctl_table *entry =
7453 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7458 static void sd_free_ctl_entry(struct ctl_table **tablep)
7460 struct ctl_table *entry;
7463 * In the intermediate directories, both the child directory and
7464 * procname are dynamically allocated and could fail but the mode
7465 * will always be set. In the lowest directory the names are
7466 * static strings and all have proc handlers.
7468 for (entry = *tablep; entry->mode; entry++) {
7470 sd_free_ctl_entry(&entry->child);
7471 if (entry->proc_handler == NULL)
7472 kfree(entry->procname);
7480 set_table_entry(struct ctl_table *entry,
7481 const char *procname, void *data, int maxlen,
7482 mode_t mode, proc_handler *proc_handler)
7484 entry->procname = procname;
7486 entry->maxlen = maxlen;
7488 entry->proc_handler = proc_handler;
7491 static struct ctl_table *
7492 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7494 struct ctl_table *table = sd_alloc_ctl_entry(13);
7499 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7500 sizeof(long), 0644, proc_doulongvec_minmax);
7501 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7502 sizeof(long), 0644, proc_doulongvec_minmax);
7503 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7504 sizeof(int), 0644, proc_dointvec_minmax);
7505 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7506 sizeof(int), 0644, proc_dointvec_minmax);
7507 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7508 sizeof(int), 0644, proc_dointvec_minmax);
7509 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7510 sizeof(int), 0644, proc_dointvec_minmax);
7511 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7512 sizeof(int), 0644, proc_dointvec_minmax);
7513 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7514 sizeof(int), 0644, proc_dointvec_minmax);
7515 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7516 sizeof(int), 0644, proc_dointvec_minmax);
7517 set_table_entry(&table[9], "cache_nice_tries",
7518 &sd->cache_nice_tries,
7519 sizeof(int), 0644, proc_dointvec_minmax);
7520 set_table_entry(&table[10], "flags", &sd->flags,
7521 sizeof(int), 0644, proc_dointvec_minmax);
7522 set_table_entry(&table[11], "name", sd->name,
7523 CORENAME_MAX_SIZE, 0444, proc_dostring);
7524 /* &table[12] is terminator */
7529 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7531 struct ctl_table *entry, *table;
7532 struct sched_domain *sd;
7533 int domain_num = 0, i;
7536 for_each_domain(cpu, sd)
7538 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7543 for_each_domain(cpu, sd) {
7544 snprintf(buf, 32, "domain%d", i);
7545 entry->procname = kstrdup(buf, GFP_KERNEL);
7547 entry->child = sd_alloc_ctl_domain_table(sd);
7554 static struct ctl_table_header *sd_sysctl_header;
7555 static void register_sched_domain_sysctl(void)
7557 int i, cpu_num = num_possible_cpus();
7558 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7561 WARN_ON(sd_ctl_dir[0].child);
7562 sd_ctl_dir[0].child = entry;
7567 for_each_possible_cpu(i) {
7568 snprintf(buf, 32, "cpu%d", i);
7569 entry->procname = kstrdup(buf, GFP_KERNEL);
7571 entry->child = sd_alloc_ctl_cpu_table(i);
7575 WARN_ON(sd_sysctl_header);
7576 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7579 /* may be called multiple times per register */
7580 static void unregister_sched_domain_sysctl(void)
7582 if (sd_sysctl_header)
7583 unregister_sysctl_table(sd_sysctl_header);
7584 sd_sysctl_header = NULL;
7585 if (sd_ctl_dir[0].child)
7586 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7589 static void register_sched_domain_sysctl(void)
7592 static void unregister_sched_domain_sysctl(void)
7597 static void set_rq_online(struct rq *rq)
7600 const struct sched_class *class;
7602 cpumask_set_cpu(rq->cpu, rq->rd->online);
7605 for_each_class(class) {
7606 if (class->rq_online)
7607 class->rq_online(rq);
7612 static void set_rq_offline(struct rq *rq)
7615 const struct sched_class *class;
7617 for_each_class(class) {
7618 if (class->rq_offline)
7619 class->rq_offline(rq);
7622 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7628 * migration_call - callback that gets triggered when a CPU is added.
7629 * Here we can start up the necessary migration thread for the new CPU.
7631 static int __cpuinit
7632 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7634 struct task_struct *p;
7635 int cpu = (long)hcpu;
7636 unsigned long flags;
7641 case CPU_UP_PREPARE:
7642 case CPU_UP_PREPARE_FROZEN:
7643 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7646 kthread_bind(p, cpu);
7647 /* Must be high prio: stop_machine expects to yield to it. */
7648 rq = task_rq_lock(p, &flags);
7649 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7650 task_rq_unlock(rq, &flags);
7652 cpu_rq(cpu)->migration_thread = p;
7653 rq->calc_load_update = calc_load_update;
7657 case CPU_ONLINE_FROZEN:
7658 /* Strictly unnecessary, as first user will wake it. */
7659 wake_up_process(cpu_rq(cpu)->migration_thread);
7661 /* Update our root-domain */
7663 spin_lock_irqsave(&rq->lock, flags);
7665 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7669 spin_unlock_irqrestore(&rq->lock, flags);
7672 #ifdef CONFIG_HOTPLUG_CPU
7673 case CPU_UP_CANCELED:
7674 case CPU_UP_CANCELED_FROZEN:
7675 if (!cpu_rq(cpu)->migration_thread)
7677 /* Unbind it from offline cpu so it can run. Fall thru. */
7678 kthread_bind(cpu_rq(cpu)->migration_thread,
7679 cpumask_any(cpu_online_mask));
7680 kthread_stop(cpu_rq(cpu)->migration_thread);
7681 put_task_struct(cpu_rq(cpu)->migration_thread);
7682 cpu_rq(cpu)->migration_thread = NULL;
7686 case CPU_DEAD_FROZEN:
7687 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7688 migrate_live_tasks(cpu);
7690 kthread_stop(rq->migration_thread);
7691 put_task_struct(rq->migration_thread);
7692 rq->migration_thread = NULL;
7693 /* Idle task back to normal (off runqueue, low prio) */
7694 spin_lock_irq(&rq->lock);
7695 update_rq_clock(rq);
7696 deactivate_task(rq, rq->idle, 0);
7697 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7698 rq->idle->sched_class = &idle_sched_class;
7699 migrate_dead_tasks(cpu);
7700 spin_unlock_irq(&rq->lock);
7702 migrate_nr_uninterruptible(rq);
7703 BUG_ON(rq->nr_running != 0);
7704 calc_global_load_remove(rq);
7706 * No need to migrate the tasks: it was best-effort if
7707 * they didn't take sched_hotcpu_mutex. Just wake up
7710 spin_lock_irq(&rq->lock);
7711 while (!list_empty(&rq->migration_queue)) {
7712 struct migration_req *req;
7714 req = list_entry(rq->migration_queue.next,
7715 struct migration_req, list);
7716 list_del_init(&req->list);
7717 spin_unlock_irq(&rq->lock);
7718 complete(&req->done);
7719 spin_lock_irq(&rq->lock);
7721 spin_unlock_irq(&rq->lock);
7725 case CPU_DYING_FROZEN:
7726 /* Update our root-domain */
7728 spin_lock_irqsave(&rq->lock, flags);
7730 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7733 spin_unlock_irqrestore(&rq->lock, flags);
7741 * Register at high priority so that task migration (migrate_all_tasks)
7742 * happens before everything else. This has to be lower priority than
7743 * the notifier in the perf_event subsystem, though.
7745 static struct notifier_block __cpuinitdata migration_notifier = {
7746 .notifier_call = migration_call,
7750 static int __init migration_init(void)
7752 void *cpu = (void *)(long)smp_processor_id();
7755 /* Start one for the boot CPU: */
7756 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7757 BUG_ON(err == NOTIFY_BAD);
7758 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7759 register_cpu_notifier(&migration_notifier);
7763 early_initcall(migration_init);
7768 #ifdef CONFIG_SCHED_DEBUG
7770 static __read_mostly int sched_domain_debug_enabled;
7772 static int __init sched_domain_debug_setup(char *str)
7774 sched_domain_debug_enabled = 1;
7778 early_param("sched_debug", sched_domain_debug_setup);
7780 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7781 struct cpumask *groupmask)
7783 struct sched_group *group = sd->groups;
7786 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7787 cpumask_clear(groupmask);
7789 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7791 if (!(sd->flags & SD_LOAD_BALANCE)) {
7792 printk("does not load-balance\n");
7794 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7799 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7801 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7802 printk(KERN_ERR "ERROR: domain->span does not contain "
7805 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7806 printk(KERN_ERR "ERROR: domain->groups does not contain"
7810 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7814 printk(KERN_ERR "ERROR: group is NULL\n");
7818 if (!group->cpu_power) {
7819 printk(KERN_CONT "\n");
7820 printk(KERN_ERR "ERROR: domain->cpu_power not "
7825 if (!cpumask_weight(sched_group_cpus(group))) {
7826 printk(KERN_CONT "\n");
7827 printk(KERN_ERR "ERROR: empty group\n");
7831 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7832 printk(KERN_CONT "\n");
7833 printk(KERN_ERR "ERROR: repeated CPUs\n");
7837 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7839 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7841 printk(KERN_CONT " %s", str);
7842 if (group->cpu_power != SCHED_LOAD_SCALE) {
7843 printk(KERN_CONT " (cpu_power = %d)",
7847 group = group->next;
7848 } while (group != sd->groups);
7849 printk(KERN_CONT "\n");
7851 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7852 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7855 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7856 printk(KERN_ERR "ERROR: parent span is not a superset "
7857 "of domain->span\n");
7861 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7863 cpumask_var_t groupmask;
7866 if (!sched_domain_debug_enabled)
7870 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7874 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7876 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7877 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7882 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7889 free_cpumask_var(groupmask);
7891 #else /* !CONFIG_SCHED_DEBUG */
7892 # define sched_domain_debug(sd, cpu) do { } while (0)
7893 #endif /* CONFIG_SCHED_DEBUG */
7895 static int sd_degenerate(struct sched_domain *sd)
7897 if (cpumask_weight(sched_domain_span(sd)) == 1)
7900 /* Following flags need at least 2 groups */
7901 if (sd->flags & (SD_LOAD_BALANCE |
7902 SD_BALANCE_NEWIDLE |
7906 SD_SHARE_PKG_RESOURCES)) {
7907 if (sd->groups != sd->groups->next)
7911 /* Following flags don't use groups */
7912 if (sd->flags & (SD_WAKE_AFFINE))
7919 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7921 unsigned long cflags = sd->flags, pflags = parent->flags;
7923 if (sd_degenerate(parent))
7926 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7929 /* Flags needing groups don't count if only 1 group in parent */
7930 if (parent->groups == parent->groups->next) {
7931 pflags &= ~(SD_LOAD_BALANCE |
7932 SD_BALANCE_NEWIDLE |
7936 SD_SHARE_PKG_RESOURCES);
7937 if (nr_node_ids == 1)
7938 pflags &= ~SD_SERIALIZE;
7940 if (~cflags & pflags)
7946 static void free_rootdomain(struct root_domain *rd)
7948 synchronize_sched();
7950 cpupri_cleanup(&rd->cpupri);
7952 free_cpumask_var(rd->rto_mask);
7953 free_cpumask_var(rd->online);
7954 free_cpumask_var(rd->span);
7958 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7960 struct root_domain *old_rd = NULL;
7961 unsigned long flags;
7963 spin_lock_irqsave(&rq->lock, flags);
7968 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7971 cpumask_clear_cpu(rq->cpu, old_rd->span);
7974 * If we dont want to free the old_rt yet then
7975 * set old_rd to NULL to skip the freeing later
7978 if (!atomic_dec_and_test(&old_rd->refcount))
7982 atomic_inc(&rd->refcount);
7985 cpumask_set_cpu(rq->cpu, rd->span);
7986 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7989 spin_unlock_irqrestore(&rq->lock, flags);
7992 free_rootdomain(old_rd);
7995 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7997 gfp_t gfp = GFP_KERNEL;
7999 memset(rd, 0, sizeof(*rd));
8004 if (!alloc_cpumask_var(&rd->span, gfp))
8006 if (!alloc_cpumask_var(&rd->online, gfp))
8008 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8011 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8016 free_cpumask_var(rd->rto_mask);
8018 free_cpumask_var(rd->online);
8020 free_cpumask_var(rd->span);
8025 static void init_defrootdomain(void)
8027 init_rootdomain(&def_root_domain, true);
8029 atomic_set(&def_root_domain.refcount, 1);
8032 static struct root_domain *alloc_rootdomain(void)
8034 struct root_domain *rd;
8036 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8040 if (init_rootdomain(rd, false) != 0) {
8049 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8050 * hold the hotplug lock.
8053 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8055 struct rq *rq = cpu_rq(cpu);
8056 struct sched_domain *tmp;
8058 /* Remove the sched domains which do not contribute to scheduling. */
8059 for (tmp = sd; tmp; ) {
8060 struct sched_domain *parent = tmp->parent;
8064 if (sd_parent_degenerate(tmp, parent)) {
8065 tmp->parent = parent->parent;
8067 parent->parent->child = tmp;
8072 if (sd && sd_degenerate(sd)) {
8078 sched_domain_debug(sd, cpu);
8080 rq_attach_root(rq, rd);
8081 rcu_assign_pointer(rq->sd, sd);
8084 /* cpus with isolated domains */
8085 static cpumask_var_t cpu_isolated_map;
8087 /* Setup the mask of cpus configured for isolated domains */
8088 static int __init isolated_cpu_setup(char *str)
8090 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8091 cpulist_parse(str, cpu_isolated_map);
8095 __setup("isolcpus=", isolated_cpu_setup);
8098 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8099 * to a function which identifies what group(along with sched group) a CPU
8100 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8101 * (due to the fact that we keep track of groups covered with a struct cpumask).
8103 * init_sched_build_groups will build a circular linked list of the groups
8104 * covered by the given span, and will set each group's ->cpumask correctly,
8105 * and ->cpu_power to 0.
8108 init_sched_build_groups(const struct cpumask *span,
8109 const struct cpumask *cpu_map,
8110 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8111 struct sched_group **sg,
8112 struct cpumask *tmpmask),
8113 struct cpumask *covered, struct cpumask *tmpmask)
8115 struct sched_group *first = NULL, *last = NULL;
8118 cpumask_clear(covered);
8120 for_each_cpu(i, span) {
8121 struct sched_group *sg;
8122 int group = group_fn(i, cpu_map, &sg, tmpmask);
8125 if (cpumask_test_cpu(i, covered))
8128 cpumask_clear(sched_group_cpus(sg));
8131 for_each_cpu(j, span) {
8132 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8135 cpumask_set_cpu(j, covered);
8136 cpumask_set_cpu(j, sched_group_cpus(sg));
8147 #define SD_NODES_PER_DOMAIN 16
8152 * find_next_best_node - find the next node to include in a sched_domain
8153 * @node: node whose sched_domain we're building
8154 * @used_nodes: nodes already in the sched_domain
8156 * Find the next node to include in a given scheduling domain. Simply
8157 * finds the closest node not already in the @used_nodes map.
8159 * Should use nodemask_t.
8161 static int find_next_best_node(int node, nodemask_t *used_nodes)
8163 int i, n, val, min_val, best_node = 0;
8167 for (i = 0; i < nr_node_ids; i++) {
8168 /* Start at @node */
8169 n = (node + i) % nr_node_ids;
8171 if (!nr_cpus_node(n))
8174 /* Skip already used nodes */
8175 if (node_isset(n, *used_nodes))
8178 /* Simple min distance search */
8179 val = node_distance(node, n);
8181 if (val < min_val) {
8187 node_set(best_node, *used_nodes);
8192 * sched_domain_node_span - get a cpumask for a node's sched_domain
8193 * @node: node whose cpumask we're constructing
8194 * @span: resulting cpumask
8196 * Given a node, construct a good cpumask for its sched_domain to span. It
8197 * should be one that prevents unnecessary balancing, but also spreads tasks
8200 static void sched_domain_node_span(int node, struct cpumask *span)
8202 nodemask_t used_nodes;
8205 cpumask_clear(span);
8206 nodes_clear(used_nodes);
8208 cpumask_or(span, span, cpumask_of_node(node));
8209 node_set(node, used_nodes);
8211 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8212 int next_node = find_next_best_node(node, &used_nodes);
8214 cpumask_or(span, span, cpumask_of_node(next_node));
8217 #endif /* CONFIG_NUMA */
8219 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8222 * The cpus mask in sched_group and sched_domain hangs off the end.
8224 * ( See the the comments in include/linux/sched.h:struct sched_group
8225 * and struct sched_domain. )
8227 struct static_sched_group {
8228 struct sched_group sg;
8229 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8232 struct static_sched_domain {
8233 struct sched_domain sd;
8234 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8240 cpumask_var_t domainspan;
8241 cpumask_var_t covered;
8242 cpumask_var_t notcovered;
8244 cpumask_var_t nodemask;
8245 cpumask_var_t this_sibling_map;
8246 cpumask_var_t this_core_map;
8247 cpumask_var_t send_covered;
8248 cpumask_var_t tmpmask;
8249 struct sched_group **sched_group_nodes;
8250 struct root_domain *rd;
8254 sa_sched_groups = 0,
8259 sa_this_sibling_map,
8261 sa_sched_group_nodes,
8271 * SMT sched-domains:
8273 #ifdef CONFIG_SCHED_SMT
8274 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8275 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8278 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8279 struct sched_group **sg, struct cpumask *unused)
8282 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8285 #endif /* CONFIG_SCHED_SMT */
8288 * multi-core sched-domains:
8290 #ifdef CONFIG_SCHED_MC
8291 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8292 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8293 #endif /* CONFIG_SCHED_MC */
8295 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8297 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8298 struct sched_group **sg, struct cpumask *mask)
8302 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8303 group = cpumask_first(mask);
8305 *sg = &per_cpu(sched_group_core, group).sg;
8308 #elif defined(CONFIG_SCHED_MC)
8310 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8311 struct sched_group **sg, struct cpumask *unused)
8314 *sg = &per_cpu(sched_group_core, cpu).sg;
8319 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8320 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8323 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8324 struct sched_group **sg, struct cpumask *mask)
8327 #ifdef CONFIG_SCHED_MC
8328 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8329 group = cpumask_first(mask);
8330 #elif defined(CONFIG_SCHED_SMT)
8331 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8332 group = cpumask_first(mask);
8337 *sg = &per_cpu(sched_group_phys, group).sg;
8343 * The init_sched_build_groups can't handle what we want to do with node
8344 * groups, so roll our own. Now each node has its own list of groups which
8345 * gets dynamically allocated.
8347 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8348 static struct sched_group ***sched_group_nodes_bycpu;
8350 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8351 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8353 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8354 struct sched_group **sg,
8355 struct cpumask *nodemask)
8359 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8360 group = cpumask_first(nodemask);
8363 *sg = &per_cpu(sched_group_allnodes, group).sg;
8367 static void init_numa_sched_groups_power(struct sched_group *group_head)
8369 struct sched_group *sg = group_head;
8375 for_each_cpu(j, sched_group_cpus(sg)) {
8376 struct sched_domain *sd;
8378 sd = &per_cpu(phys_domains, j).sd;
8379 if (j != group_first_cpu(sd->groups)) {
8381 * Only add "power" once for each
8387 sg->cpu_power += sd->groups->cpu_power;
8390 } while (sg != group_head);
8393 static int build_numa_sched_groups(struct s_data *d,
8394 const struct cpumask *cpu_map, int num)
8396 struct sched_domain *sd;
8397 struct sched_group *sg, *prev;
8400 cpumask_clear(d->covered);
8401 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8402 if (cpumask_empty(d->nodemask)) {
8403 d->sched_group_nodes[num] = NULL;
8407 sched_domain_node_span(num, d->domainspan);
8408 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8410 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8413 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8417 d->sched_group_nodes[num] = sg;
8419 for_each_cpu(j, d->nodemask) {
8420 sd = &per_cpu(node_domains, j).sd;
8425 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8427 cpumask_or(d->covered, d->covered, d->nodemask);
8430 for (j = 0; j < nr_node_ids; j++) {
8431 n = (num + j) % nr_node_ids;
8432 cpumask_complement(d->notcovered, d->covered);
8433 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8434 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8435 if (cpumask_empty(d->tmpmask))
8437 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8438 if (cpumask_empty(d->tmpmask))
8440 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8444 "Can not alloc domain group for node %d\n", j);
8448 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8449 sg->next = prev->next;
8450 cpumask_or(d->covered, d->covered, d->tmpmask);
8457 #endif /* CONFIG_NUMA */
8460 /* Free memory allocated for various sched_group structures */
8461 static void free_sched_groups(const struct cpumask *cpu_map,
8462 struct cpumask *nodemask)
8466 for_each_cpu(cpu, cpu_map) {
8467 struct sched_group **sched_group_nodes
8468 = sched_group_nodes_bycpu[cpu];
8470 if (!sched_group_nodes)
8473 for (i = 0; i < nr_node_ids; i++) {
8474 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8476 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8477 if (cpumask_empty(nodemask))
8487 if (oldsg != sched_group_nodes[i])
8490 kfree(sched_group_nodes);
8491 sched_group_nodes_bycpu[cpu] = NULL;
8494 #else /* !CONFIG_NUMA */
8495 static void free_sched_groups(const struct cpumask *cpu_map,
8496 struct cpumask *nodemask)
8499 #endif /* CONFIG_NUMA */
8502 * Initialize sched groups cpu_power.
8504 * cpu_power indicates the capacity of sched group, which is used while
8505 * distributing the load between different sched groups in a sched domain.
8506 * Typically cpu_power for all the groups in a sched domain will be same unless
8507 * there are asymmetries in the topology. If there are asymmetries, group
8508 * having more cpu_power will pickup more load compared to the group having
8511 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8513 struct sched_domain *child;
8514 struct sched_group *group;
8518 WARN_ON(!sd || !sd->groups);
8520 if (cpu != group_first_cpu(sd->groups))
8525 sd->groups->cpu_power = 0;
8528 power = SCHED_LOAD_SCALE;
8529 weight = cpumask_weight(sched_domain_span(sd));
8531 * SMT siblings share the power of a single core.
8532 * Usually multiple threads get a better yield out of
8533 * that one core than a single thread would have,
8534 * reflect that in sd->smt_gain.
8536 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8537 power *= sd->smt_gain;
8539 power >>= SCHED_LOAD_SHIFT;
8541 sd->groups->cpu_power += power;
8546 * Add cpu_power of each child group to this groups cpu_power.
8548 group = child->groups;
8550 sd->groups->cpu_power += group->cpu_power;
8551 group = group->next;
8552 } while (group != child->groups);
8556 * Initializers for schedule domains
8557 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8560 #ifdef CONFIG_SCHED_DEBUG
8561 # define SD_INIT_NAME(sd, type) sd->name = #type
8563 # define SD_INIT_NAME(sd, type) do { } while (0)
8566 #define SD_INIT(sd, type) sd_init_##type(sd)
8568 #define SD_INIT_FUNC(type) \
8569 static noinline void sd_init_##type(struct sched_domain *sd) \
8571 memset(sd, 0, sizeof(*sd)); \
8572 *sd = SD_##type##_INIT; \
8573 sd->level = SD_LV_##type; \
8574 SD_INIT_NAME(sd, type); \
8579 SD_INIT_FUNC(ALLNODES)
8582 #ifdef CONFIG_SCHED_SMT
8583 SD_INIT_FUNC(SIBLING)
8585 #ifdef CONFIG_SCHED_MC
8589 static int default_relax_domain_level = -1;
8591 static int __init setup_relax_domain_level(char *str)
8595 val = simple_strtoul(str, NULL, 0);
8596 if (val < SD_LV_MAX)
8597 default_relax_domain_level = val;
8601 __setup("relax_domain_level=", setup_relax_domain_level);
8603 static void set_domain_attribute(struct sched_domain *sd,
8604 struct sched_domain_attr *attr)
8608 if (!attr || attr->relax_domain_level < 0) {
8609 if (default_relax_domain_level < 0)
8612 request = default_relax_domain_level;
8614 request = attr->relax_domain_level;
8615 if (request < sd->level) {
8616 /* turn off idle balance on this domain */
8617 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8619 /* turn on idle balance on this domain */
8620 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8624 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8625 const struct cpumask *cpu_map)
8628 case sa_sched_groups:
8629 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8630 d->sched_group_nodes = NULL;
8632 free_rootdomain(d->rd); /* fall through */
8634 free_cpumask_var(d->tmpmask); /* fall through */
8635 case sa_send_covered:
8636 free_cpumask_var(d->send_covered); /* fall through */
8637 case sa_this_core_map:
8638 free_cpumask_var(d->this_core_map); /* fall through */
8639 case sa_this_sibling_map:
8640 free_cpumask_var(d->this_sibling_map); /* fall through */
8642 free_cpumask_var(d->nodemask); /* fall through */
8643 case sa_sched_group_nodes:
8645 kfree(d->sched_group_nodes); /* fall through */
8647 free_cpumask_var(d->notcovered); /* fall through */
8649 free_cpumask_var(d->covered); /* fall through */
8651 free_cpumask_var(d->domainspan); /* fall through */
8658 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8659 const struct cpumask *cpu_map)
8662 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8664 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8665 return sa_domainspan;
8666 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8668 /* Allocate the per-node list of sched groups */
8669 d->sched_group_nodes = kcalloc(nr_node_ids,
8670 sizeof(struct sched_group *), GFP_KERNEL);
8671 if (!d->sched_group_nodes) {
8672 printk(KERN_WARNING "Can not alloc sched group node list\n");
8673 return sa_notcovered;
8675 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8677 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8678 return sa_sched_group_nodes;
8679 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8681 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8682 return sa_this_sibling_map;
8683 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8684 return sa_this_core_map;
8685 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8686 return sa_send_covered;
8687 d->rd = alloc_rootdomain();
8689 printk(KERN_WARNING "Cannot alloc root domain\n");
8692 return sa_rootdomain;
8695 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8696 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8698 struct sched_domain *sd = NULL;
8700 struct sched_domain *parent;
8703 if (cpumask_weight(cpu_map) >
8704 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8705 sd = &per_cpu(allnodes_domains, i).sd;
8706 SD_INIT(sd, ALLNODES);
8707 set_domain_attribute(sd, attr);
8708 cpumask_copy(sched_domain_span(sd), cpu_map);
8709 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8714 sd = &per_cpu(node_domains, i).sd;
8716 set_domain_attribute(sd, attr);
8717 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8718 sd->parent = parent;
8721 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8726 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8727 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8728 struct sched_domain *parent, int i)
8730 struct sched_domain *sd;
8731 sd = &per_cpu(phys_domains, i).sd;
8733 set_domain_attribute(sd, attr);
8734 cpumask_copy(sched_domain_span(sd), d->nodemask);
8735 sd->parent = parent;
8738 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8742 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8743 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8744 struct sched_domain *parent, int i)
8746 struct sched_domain *sd = parent;
8747 #ifdef CONFIG_SCHED_MC
8748 sd = &per_cpu(core_domains, i).sd;
8750 set_domain_attribute(sd, attr);
8751 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8752 sd->parent = parent;
8754 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8759 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8760 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8761 struct sched_domain *parent, int i)
8763 struct sched_domain *sd = parent;
8764 #ifdef CONFIG_SCHED_SMT
8765 sd = &per_cpu(cpu_domains, i).sd;
8766 SD_INIT(sd, SIBLING);
8767 set_domain_attribute(sd, attr);
8768 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8769 sd->parent = parent;
8771 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8776 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8777 const struct cpumask *cpu_map, int cpu)
8780 #ifdef CONFIG_SCHED_SMT
8781 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8782 cpumask_and(d->this_sibling_map, cpu_map,
8783 topology_thread_cpumask(cpu));
8784 if (cpu == cpumask_first(d->this_sibling_map))
8785 init_sched_build_groups(d->this_sibling_map, cpu_map,
8787 d->send_covered, d->tmpmask);
8790 #ifdef CONFIG_SCHED_MC
8791 case SD_LV_MC: /* set up multi-core groups */
8792 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8793 if (cpu == cpumask_first(d->this_core_map))
8794 init_sched_build_groups(d->this_core_map, cpu_map,
8796 d->send_covered, d->tmpmask);
8799 case SD_LV_CPU: /* set up physical groups */
8800 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8801 if (!cpumask_empty(d->nodemask))
8802 init_sched_build_groups(d->nodemask, cpu_map,
8804 d->send_covered, d->tmpmask);
8807 case SD_LV_ALLNODES:
8808 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8809 d->send_covered, d->tmpmask);
8818 * Build sched domains for a given set of cpus and attach the sched domains
8819 * to the individual cpus
8821 static int __build_sched_domains(const struct cpumask *cpu_map,
8822 struct sched_domain_attr *attr)
8824 enum s_alloc alloc_state = sa_none;
8826 struct sched_domain *sd;
8832 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8833 if (alloc_state != sa_rootdomain)
8835 alloc_state = sa_sched_groups;
8838 * Set up domains for cpus specified by the cpu_map.
8840 for_each_cpu(i, cpu_map) {
8841 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8844 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8845 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8846 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8847 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8850 for_each_cpu(i, cpu_map) {
8851 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8852 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8855 /* Set up physical groups */
8856 for (i = 0; i < nr_node_ids; i++)
8857 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8860 /* Set up node groups */
8862 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8864 for (i = 0; i < nr_node_ids; i++)
8865 if (build_numa_sched_groups(&d, cpu_map, i))
8869 /* Calculate CPU power for physical packages and nodes */
8870 #ifdef CONFIG_SCHED_SMT
8871 for_each_cpu(i, cpu_map) {
8872 sd = &per_cpu(cpu_domains, i).sd;
8873 init_sched_groups_power(i, sd);
8876 #ifdef CONFIG_SCHED_MC
8877 for_each_cpu(i, cpu_map) {
8878 sd = &per_cpu(core_domains, i).sd;
8879 init_sched_groups_power(i, sd);
8883 for_each_cpu(i, cpu_map) {
8884 sd = &per_cpu(phys_domains, i).sd;
8885 init_sched_groups_power(i, sd);
8889 for (i = 0; i < nr_node_ids; i++)
8890 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8892 if (d.sd_allnodes) {
8893 struct sched_group *sg;
8895 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8897 init_numa_sched_groups_power(sg);
8901 /* Attach the domains */
8902 for_each_cpu(i, cpu_map) {
8903 #ifdef CONFIG_SCHED_SMT
8904 sd = &per_cpu(cpu_domains, i).sd;
8905 #elif defined(CONFIG_SCHED_MC)
8906 sd = &per_cpu(core_domains, i).sd;
8908 sd = &per_cpu(phys_domains, i).sd;
8910 cpu_attach_domain(sd, d.rd, i);
8913 d.sched_group_nodes = NULL; /* don't free this we still need it */
8914 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8918 __free_domain_allocs(&d, alloc_state, cpu_map);
8922 static int build_sched_domains(const struct cpumask *cpu_map)
8924 return __build_sched_domains(cpu_map, NULL);
8927 static cpumask_var_t *doms_cur; /* current sched domains */
8928 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8929 static struct sched_domain_attr *dattr_cur;
8930 /* attribues of custom domains in 'doms_cur' */
8933 * Special case: If a kmalloc of a doms_cur partition (array of
8934 * cpumask) fails, then fallback to a single sched domain,
8935 * as determined by the single cpumask fallback_doms.
8937 static cpumask_var_t fallback_doms;
8940 * arch_update_cpu_topology lets virtualized architectures update the
8941 * cpu core maps. It is supposed to return 1 if the topology changed
8942 * or 0 if it stayed the same.
8944 int __attribute__((weak)) arch_update_cpu_topology(void)
8949 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8952 cpumask_var_t *doms;
8954 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8957 for (i = 0; i < ndoms; i++) {
8958 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
8959 free_sched_domains(doms, i);
8966 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
8969 for (i = 0; i < ndoms; i++)
8970 free_cpumask_var(doms[i]);
8975 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8976 * For now this just excludes isolated cpus, but could be used to
8977 * exclude other special cases in the future.
8979 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8983 arch_update_cpu_topology();
8985 doms_cur = alloc_sched_domains(ndoms_cur);
8987 doms_cur = &fallback_doms;
8988 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
8990 err = build_sched_domains(doms_cur[0]);
8991 register_sched_domain_sysctl();
8996 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8997 struct cpumask *tmpmask)
8999 free_sched_groups(cpu_map, tmpmask);
9003 * Detach sched domains from a group of cpus specified in cpu_map
9004 * These cpus will now be attached to the NULL domain
9006 static void detach_destroy_domains(const struct cpumask *cpu_map)
9008 /* Save because hotplug lock held. */
9009 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9012 for_each_cpu(i, cpu_map)
9013 cpu_attach_domain(NULL, &def_root_domain, i);
9014 synchronize_sched();
9015 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9018 /* handle null as "default" */
9019 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9020 struct sched_domain_attr *new, int idx_new)
9022 struct sched_domain_attr tmp;
9029 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9030 new ? (new + idx_new) : &tmp,
9031 sizeof(struct sched_domain_attr));
9035 * Partition sched domains as specified by the 'ndoms_new'
9036 * cpumasks in the array doms_new[] of cpumasks. This compares
9037 * doms_new[] to the current sched domain partitioning, doms_cur[].
9038 * It destroys each deleted domain and builds each new domain.
9040 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9041 * The masks don't intersect (don't overlap.) We should setup one
9042 * sched domain for each mask. CPUs not in any of the cpumasks will
9043 * not be load balanced. If the same cpumask appears both in the
9044 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9047 * The passed in 'doms_new' should be allocated using
9048 * alloc_sched_domains. This routine takes ownership of it and will
9049 * free_sched_domains it when done with it. If the caller failed the
9050 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9051 * and partition_sched_domains() will fallback to the single partition
9052 * 'fallback_doms', it also forces the domains to be rebuilt.
9054 * If doms_new == NULL it will be replaced with cpu_online_mask.
9055 * ndoms_new == 0 is a special case for destroying existing domains,
9056 * and it will not create the default domain.
9058 * Call with hotplug lock held
9060 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9061 struct sched_domain_attr *dattr_new)
9066 mutex_lock(&sched_domains_mutex);
9068 /* always unregister in case we don't destroy any domains */
9069 unregister_sched_domain_sysctl();
9071 /* Let architecture update cpu core mappings. */
9072 new_topology = arch_update_cpu_topology();
9074 n = doms_new ? ndoms_new : 0;
9076 /* Destroy deleted domains */
9077 for (i = 0; i < ndoms_cur; i++) {
9078 for (j = 0; j < n && !new_topology; j++) {
9079 if (cpumask_equal(doms_cur[i], doms_new[j])
9080 && dattrs_equal(dattr_cur, i, dattr_new, j))
9083 /* no match - a current sched domain not in new doms_new[] */
9084 detach_destroy_domains(doms_cur[i]);
9089 if (doms_new == NULL) {
9091 doms_new = &fallback_doms;
9092 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9093 WARN_ON_ONCE(dattr_new);
9096 /* Build new domains */
9097 for (i = 0; i < ndoms_new; i++) {
9098 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9099 if (cpumask_equal(doms_new[i], doms_cur[j])
9100 && dattrs_equal(dattr_new, i, dattr_cur, j))
9103 /* no match - add a new doms_new */
9104 __build_sched_domains(doms_new[i],
9105 dattr_new ? dattr_new + i : NULL);
9110 /* Remember the new sched domains */
9111 if (doms_cur != &fallback_doms)
9112 free_sched_domains(doms_cur, ndoms_cur);
9113 kfree(dattr_cur); /* kfree(NULL) is safe */
9114 doms_cur = doms_new;
9115 dattr_cur = dattr_new;
9116 ndoms_cur = ndoms_new;
9118 register_sched_domain_sysctl();
9120 mutex_unlock(&sched_domains_mutex);
9123 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9124 static void arch_reinit_sched_domains(void)
9128 /* Destroy domains first to force the rebuild */
9129 partition_sched_domains(0, NULL, NULL);
9131 rebuild_sched_domains();
9135 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9137 unsigned int level = 0;
9139 if (sscanf(buf, "%u", &level) != 1)
9143 * level is always be positive so don't check for
9144 * level < POWERSAVINGS_BALANCE_NONE which is 0
9145 * What happens on 0 or 1 byte write,
9146 * need to check for count as well?
9149 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9153 sched_smt_power_savings = level;
9155 sched_mc_power_savings = level;
9157 arch_reinit_sched_domains();
9162 #ifdef CONFIG_SCHED_MC
9163 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9166 return sprintf(page, "%u\n", sched_mc_power_savings);
9168 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9169 const char *buf, size_t count)
9171 return sched_power_savings_store(buf, count, 0);
9173 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9174 sched_mc_power_savings_show,
9175 sched_mc_power_savings_store);
9178 #ifdef CONFIG_SCHED_SMT
9179 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9182 return sprintf(page, "%u\n", sched_smt_power_savings);
9184 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9185 const char *buf, size_t count)
9187 return sched_power_savings_store(buf, count, 1);
9189 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9190 sched_smt_power_savings_show,
9191 sched_smt_power_savings_store);
9194 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9198 #ifdef CONFIG_SCHED_SMT
9200 err = sysfs_create_file(&cls->kset.kobj,
9201 &attr_sched_smt_power_savings.attr);
9203 #ifdef CONFIG_SCHED_MC
9204 if (!err && mc_capable())
9205 err = sysfs_create_file(&cls->kset.kobj,
9206 &attr_sched_mc_power_savings.attr);
9210 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9212 #ifndef CONFIG_CPUSETS
9214 * Add online and remove offline CPUs from the scheduler domains.
9215 * When cpusets are enabled they take over this function.
9217 static int update_sched_domains(struct notifier_block *nfb,
9218 unsigned long action, void *hcpu)
9222 case CPU_ONLINE_FROZEN:
9223 case CPU_DOWN_PREPARE:
9224 case CPU_DOWN_PREPARE_FROZEN:
9225 case CPU_DOWN_FAILED:
9226 case CPU_DOWN_FAILED_FROZEN:
9227 partition_sched_domains(1, NULL, NULL);
9236 static int update_runtime(struct notifier_block *nfb,
9237 unsigned long action, void *hcpu)
9239 int cpu = (int)(long)hcpu;
9242 case CPU_DOWN_PREPARE:
9243 case CPU_DOWN_PREPARE_FROZEN:
9244 disable_runtime(cpu_rq(cpu));
9247 case CPU_DOWN_FAILED:
9248 case CPU_DOWN_FAILED_FROZEN:
9250 case CPU_ONLINE_FROZEN:
9251 enable_runtime(cpu_rq(cpu));
9259 void __init sched_init_smp(void)
9261 cpumask_var_t non_isolated_cpus;
9263 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9264 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9266 #if defined(CONFIG_NUMA)
9267 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9269 BUG_ON(sched_group_nodes_bycpu == NULL);
9272 mutex_lock(&sched_domains_mutex);
9273 arch_init_sched_domains(cpu_active_mask);
9274 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9275 if (cpumask_empty(non_isolated_cpus))
9276 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9277 mutex_unlock(&sched_domains_mutex);
9280 #ifndef CONFIG_CPUSETS
9281 /* XXX: Theoretical race here - CPU may be hotplugged now */
9282 hotcpu_notifier(update_sched_domains, 0);
9285 /* RT runtime code needs to handle some hotplug events */
9286 hotcpu_notifier(update_runtime, 0);
9290 /* Move init over to a non-isolated CPU */
9291 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9293 sched_init_granularity();
9294 free_cpumask_var(non_isolated_cpus);
9296 init_sched_rt_class();
9299 void __init sched_init_smp(void)
9301 sched_init_granularity();
9303 #endif /* CONFIG_SMP */
9305 const_debug unsigned int sysctl_timer_migration = 1;
9307 int in_sched_functions(unsigned long addr)
9309 return in_lock_functions(addr) ||
9310 (addr >= (unsigned long)__sched_text_start
9311 && addr < (unsigned long)__sched_text_end);
9314 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9316 cfs_rq->tasks_timeline = RB_ROOT;
9317 INIT_LIST_HEAD(&cfs_rq->tasks);
9318 #ifdef CONFIG_FAIR_GROUP_SCHED
9321 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9324 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9326 struct rt_prio_array *array;
9329 array = &rt_rq->active;
9330 for (i = 0; i < MAX_RT_PRIO; i++) {
9331 INIT_LIST_HEAD(array->queue + i);
9332 __clear_bit(i, array->bitmap);
9334 /* delimiter for bitsearch: */
9335 __set_bit(MAX_RT_PRIO, array->bitmap);
9337 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9338 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9340 rt_rq->highest_prio.next = MAX_RT_PRIO;
9344 rt_rq->rt_nr_migratory = 0;
9345 rt_rq->overloaded = 0;
9346 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9350 rt_rq->rt_throttled = 0;
9351 rt_rq->rt_runtime = 0;
9352 spin_lock_init(&rt_rq->rt_runtime_lock);
9354 #ifdef CONFIG_RT_GROUP_SCHED
9355 rt_rq->rt_nr_boosted = 0;
9360 #ifdef CONFIG_FAIR_GROUP_SCHED
9361 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9362 struct sched_entity *se, int cpu, int add,
9363 struct sched_entity *parent)
9365 struct rq *rq = cpu_rq(cpu);
9366 tg->cfs_rq[cpu] = cfs_rq;
9367 init_cfs_rq(cfs_rq, rq);
9370 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9373 /* se could be NULL for init_task_group */
9378 se->cfs_rq = &rq->cfs;
9380 se->cfs_rq = parent->my_q;
9383 se->load.weight = tg->shares;
9384 se->load.inv_weight = 0;
9385 se->parent = parent;
9389 #ifdef CONFIG_RT_GROUP_SCHED
9390 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9391 struct sched_rt_entity *rt_se, int cpu, int add,
9392 struct sched_rt_entity *parent)
9394 struct rq *rq = cpu_rq(cpu);
9396 tg->rt_rq[cpu] = rt_rq;
9397 init_rt_rq(rt_rq, rq);
9399 rt_rq->rt_se = rt_se;
9400 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9402 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9404 tg->rt_se[cpu] = rt_se;
9409 rt_se->rt_rq = &rq->rt;
9411 rt_se->rt_rq = parent->my_q;
9413 rt_se->my_q = rt_rq;
9414 rt_se->parent = parent;
9415 INIT_LIST_HEAD(&rt_se->run_list);
9419 void __init sched_init(void)
9422 unsigned long alloc_size = 0, ptr;
9424 #ifdef CONFIG_FAIR_GROUP_SCHED
9425 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9427 #ifdef CONFIG_RT_GROUP_SCHED
9428 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9430 #ifdef CONFIG_USER_SCHED
9433 #ifdef CONFIG_CPUMASK_OFFSTACK
9434 alloc_size += num_possible_cpus() * cpumask_size();
9437 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9439 #ifdef CONFIG_FAIR_GROUP_SCHED
9440 init_task_group.se = (struct sched_entity **)ptr;
9441 ptr += nr_cpu_ids * sizeof(void **);
9443 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9444 ptr += nr_cpu_ids * sizeof(void **);
9446 #ifdef CONFIG_USER_SCHED
9447 root_task_group.se = (struct sched_entity **)ptr;
9448 ptr += nr_cpu_ids * sizeof(void **);
9450 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9451 ptr += nr_cpu_ids * sizeof(void **);
9452 #endif /* CONFIG_USER_SCHED */
9453 #endif /* CONFIG_FAIR_GROUP_SCHED */
9454 #ifdef CONFIG_RT_GROUP_SCHED
9455 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9456 ptr += nr_cpu_ids * sizeof(void **);
9458 init_task_group.rt_rq = (struct rt_rq **)ptr;
9459 ptr += nr_cpu_ids * sizeof(void **);
9461 #ifdef CONFIG_USER_SCHED
9462 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9463 ptr += nr_cpu_ids * sizeof(void **);
9465 root_task_group.rt_rq = (struct rt_rq **)ptr;
9466 ptr += nr_cpu_ids * sizeof(void **);
9467 #endif /* CONFIG_USER_SCHED */
9468 #endif /* CONFIG_RT_GROUP_SCHED */
9469 #ifdef CONFIG_CPUMASK_OFFSTACK
9470 for_each_possible_cpu(i) {
9471 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9472 ptr += cpumask_size();
9474 #endif /* CONFIG_CPUMASK_OFFSTACK */
9478 init_defrootdomain();
9481 init_rt_bandwidth(&def_rt_bandwidth,
9482 global_rt_period(), global_rt_runtime());
9484 #ifdef CONFIG_RT_GROUP_SCHED
9485 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9486 global_rt_period(), global_rt_runtime());
9487 #ifdef CONFIG_USER_SCHED
9488 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9489 global_rt_period(), RUNTIME_INF);
9490 #endif /* CONFIG_USER_SCHED */
9491 #endif /* CONFIG_RT_GROUP_SCHED */
9493 #ifdef CONFIG_GROUP_SCHED
9494 list_add(&init_task_group.list, &task_groups);
9495 INIT_LIST_HEAD(&init_task_group.children);
9497 #ifdef CONFIG_USER_SCHED
9498 INIT_LIST_HEAD(&root_task_group.children);
9499 init_task_group.parent = &root_task_group;
9500 list_add(&init_task_group.siblings, &root_task_group.children);
9501 #endif /* CONFIG_USER_SCHED */
9502 #endif /* CONFIG_GROUP_SCHED */
9504 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9505 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9506 __alignof__(unsigned long));
9508 for_each_possible_cpu(i) {
9512 spin_lock_init(&rq->lock);
9514 rq->calc_load_active = 0;
9515 rq->calc_load_update = jiffies + LOAD_FREQ;
9516 init_cfs_rq(&rq->cfs, rq);
9517 init_rt_rq(&rq->rt, rq);
9518 #ifdef CONFIG_FAIR_GROUP_SCHED
9519 init_task_group.shares = init_task_group_load;
9520 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9521 #ifdef CONFIG_CGROUP_SCHED
9523 * How much cpu bandwidth does init_task_group get?
9525 * In case of task-groups formed thr' the cgroup filesystem, it
9526 * gets 100% of the cpu resources in the system. This overall
9527 * system cpu resource is divided among the tasks of
9528 * init_task_group and its child task-groups in a fair manner,
9529 * based on each entity's (task or task-group's) weight
9530 * (se->load.weight).
9532 * In other words, if init_task_group has 10 tasks of weight
9533 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9534 * then A0's share of the cpu resource is:
9536 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9538 * We achieve this by letting init_task_group's tasks sit
9539 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9541 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9542 #elif defined CONFIG_USER_SCHED
9543 root_task_group.shares = NICE_0_LOAD;
9544 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9546 * In case of task-groups formed thr' the user id of tasks,
9547 * init_task_group represents tasks belonging to root user.
9548 * Hence it forms a sibling of all subsequent groups formed.
9549 * In this case, init_task_group gets only a fraction of overall
9550 * system cpu resource, based on the weight assigned to root
9551 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9552 * by letting tasks of init_task_group sit in a separate cfs_rq
9553 * (init_tg_cfs_rq) and having one entity represent this group of
9554 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9556 init_tg_cfs_entry(&init_task_group,
9557 &per_cpu(init_tg_cfs_rq, i),
9558 &per_cpu(init_sched_entity, i), i, 1,
9559 root_task_group.se[i]);
9562 #endif /* CONFIG_FAIR_GROUP_SCHED */
9564 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9565 #ifdef CONFIG_RT_GROUP_SCHED
9566 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9567 #ifdef CONFIG_CGROUP_SCHED
9568 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9569 #elif defined CONFIG_USER_SCHED
9570 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9571 init_tg_rt_entry(&init_task_group,
9572 &per_cpu(init_rt_rq, i),
9573 &per_cpu(init_sched_rt_entity, i), i, 1,
9574 root_task_group.rt_se[i]);
9578 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9579 rq->cpu_load[j] = 0;
9583 rq->post_schedule = 0;
9584 rq->active_balance = 0;
9585 rq->next_balance = jiffies;
9589 rq->migration_thread = NULL;
9591 rq->avg_idle = 2*sysctl_sched_migration_cost;
9592 INIT_LIST_HEAD(&rq->migration_queue);
9593 rq_attach_root(rq, &def_root_domain);
9596 atomic_set(&rq->nr_iowait, 0);
9599 set_load_weight(&init_task);
9601 #ifdef CONFIG_PREEMPT_NOTIFIERS
9602 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9606 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9609 #ifdef CONFIG_RT_MUTEXES
9610 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9614 * The boot idle thread does lazy MMU switching as well:
9616 atomic_inc(&init_mm.mm_count);
9617 enter_lazy_tlb(&init_mm, current);
9620 * Make us the idle thread. Technically, schedule() should not be
9621 * called from this thread, however somewhere below it might be,
9622 * but because we are the idle thread, we just pick up running again
9623 * when this runqueue becomes "idle".
9625 init_idle(current, smp_processor_id());
9627 calc_load_update = jiffies + LOAD_FREQ;
9630 * During early bootup we pretend to be a normal task:
9632 current->sched_class = &fair_sched_class;
9634 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9635 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9638 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9639 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9641 /* May be allocated at isolcpus cmdline parse time */
9642 if (cpu_isolated_map == NULL)
9643 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9648 scheduler_running = 1;
9651 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9652 static inline int preempt_count_equals(int preempt_offset)
9654 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9656 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9659 void __might_sleep(char *file, int line, int preempt_offset)
9662 static unsigned long prev_jiffy; /* ratelimiting */
9664 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9665 system_state != SYSTEM_RUNNING || oops_in_progress)
9667 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9669 prev_jiffy = jiffies;
9672 "BUG: sleeping function called from invalid context at %s:%d\n",
9675 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9676 in_atomic(), irqs_disabled(),
9677 current->pid, current->comm);
9679 debug_show_held_locks(current);
9680 if (irqs_disabled())
9681 print_irqtrace_events(current);
9685 EXPORT_SYMBOL(__might_sleep);
9688 #ifdef CONFIG_MAGIC_SYSRQ
9689 static void normalize_task(struct rq *rq, struct task_struct *p)
9693 update_rq_clock(rq);
9694 on_rq = p->se.on_rq;
9696 deactivate_task(rq, p, 0);
9697 __setscheduler(rq, p, SCHED_NORMAL, 0);
9699 activate_task(rq, p, 0);
9700 resched_task(rq->curr);
9704 void normalize_rt_tasks(void)
9706 struct task_struct *g, *p;
9707 unsigned long flags;
9710 read_lock_irqsave(&tasklist_lock, flags);
9711 do_each_thread(g, p) {
9713 * Only normalize user tasks:
9718 p->se.exec_start = 0;
9719 #ifdef CONFIG_SCHEDSTATS
9720 p->se.wait_start = 0;
9721 p->se.sleep_start = 0;
9722 p->se.block_start = 0;
9727 * Renice negative nice level userspace
9730 if (TASK_NICE(p) < 0 && p->mm)
9731 set_user_nice(p, 0);
9735 spin_lock(&p->pi_lock);
9736 rq = __task_rq_lock(p);
9738 normalize_task(rq, p);
9740 __task_rq_unlock(rq);
9741 spin_unlock(&p->pi_lock);
9742 } while_each_thread(g, p);
9744 read_unlock_irqrestore(&tasklist_lock, flags);
9747 #endif /* CONFIG_MAGIC_SYSRQ */
9751 * These functions are only useful for the IA64 MCA handling.
9753 * They can only be called when the whole system has been
9754 * stopped - every CPU needs to be quiescent, and no scheduling
9755 * activity can take place. Using them for anything else would
9756 * be a serious bug, and as a result, they aren't even visible
9757 * under any other configuration.
9761 * curr_task - return the current task for a given cpu.
9762 * @cpu: the processor in question.
9764 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9766 struct task_struct *curr_task(int cpu)
9768 return cpu_curr(cpu);
9772 * set_curr_task - set the current task for a given cpu.
9773 * @cpu: the processor in question.
9774 * @p: the task pointer to set.
9776 * Description: This function must only be used when non-maskable interrupts
9777 * are serviced on a separate stack. It allows the architecture to switch the
9778 * notion of the current task on a cpu in a non-blocking manner. This function
9779 * must be called with all CPU's synchronized, and interrupts disabled, the
9780 * and caller must save the original value of the current task (see
9781 * curr_task() above) and restore that value before reenabling interrupts and
9782 * re-starting the system.
9784 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9786 void set_curr_task(int cpu, struct task_struct *p)
9793 #ifdef CONFIG_FAIR_GROUP_SCHED
9794 static void free_fair_sched_group(struct task_group *tg)
9798 for_each_possible_cpu(i) {
9800 kfree(tg->cfs_rq[i]);
9810 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9812 struct cfs_rq *cfs_rq;
9813 struct sched_entity *se;
9817 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9820 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9824 tg->shares = NICE_0_LOAD;
9826 for_each_possible_cpu(i) {
9829 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9830 GFP_KERNEL, cpu_to_node(i));
9834 se = kzalloc_node(sizeof(struct sched_entity),
9835 GFP_KERNEL, cpu_to_node(i));
9839 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9848 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9850 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9851 &cpu_rq(cpu)->leaf_cfs_rq_list);
9854 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9856 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9858 #else /* !CONFG_FAIR_GROUP_SCHED */
9859 static inline void free_fair_sched_group(struct task_group *tg)
9864 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9869 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9873 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9876 #endif /* CONFIG_FAIR_GROUP_SCHED */
9878 #ifdef CONFIG_RT_GROUP_SCHED
9879 static void free_rt_sched_group(struct task_group *tg)
9883 destroy_rt_bandwidth(&tg->rt_bandwidth);
9885 for_each_possible_cpu(i) {
9887 kfree(tg->rt_rq[i]);
9889 kfree(tg->rt_se[i]);
9897 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9899 struct rt_rq *rt_rq;
9900 struct sched_rt_entity *rt_se;
9904 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9907 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9911 init_rt_bandwidth(&tg->rt_bandwidth,
9912 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9914 for_each_possible_cpu(i) {
9917 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9918 GFP_KERNEL, cpu_to_node(i));
9922 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9923 GFP_KERNEL, cpu_to_node(i));
9927 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9936 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9938 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9939 &cpu_rq(cpu)->leaf_rt_rq_list);
9942 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9944 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9946 #else /* !CONFIG_RT_GROUP_SCHED */
9947 static inline void free_rt_sched_group(struct task_group *tg)
9952 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9957 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9961 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9964 #endif /* CONFIG_RT_GROUP_SCHED */
9966 #ifdef CONFIG_GROUP_SCHED
9967 static void free_sched_group(struct task_group *tg)
9969 free_fair_sched_group(tg);
9970 free_rt_sched_group(tg);
9974 /* allocate runqueue etc for a new task group */
9975 struct task_group *sched_create_group(struct task_group *parent)
9977 struct task_group *tg;
9978 unsigned long flags;
9981 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9983 return ERR_PTR(-ENOMEM);
9985 if (!alloc_fair_sched_group(tg, parent))
9988 if (!alloc_rt_sched_group(tg, parent))
9991 spin_lock_irqsave(&task_group_lock, flags);
9992 for_each_possible_cpu(i) {
9993 register_fair_sched_group(tg, i);
9994 register_rt_sched_group(tg, i);
9996 list_add_rcu(&tg->list, &task_groups);
9998 WARN_ON(!parent); /* root should already exist */
10000 tg->parent = parent;
10001 INIT_LIST_HEAD(&tg->children);
10002 list_add_rcu(&tg->siblings, &parent->children);
10003 spin_unlock_irqrestore(&task_group_lock, flags);
10008 free_sched_group(tg);
10009 return ERR_PTR(-ENOMEM);
10012 /* rcu callback to free various structures associated with a task group */
10013 static void free_sched_group_rcu(struct rcu_head *rhp)
10015 /* now it should be safe to free those cfs_rqs */
10016 free_sched_group(container_of(rhp, struct task_group, rcu));
10019 /* Destroy runqueue etc associated with a task group */
10020 void sched_destroy_group(struct task_group *tg)
10022 unsigned long flags;
10025 spin_lock_irqsave(&task_group_lock, flags);
10026 for_each_possible_cpu(i) {
10027 unregister_fair_sched_group(tg, i);
10028 unregister_rt_sched_group(tg, i);
10030 list_del_rcu(&tg->list);
10031 list_del_rcu(&tg->siblings);
10032 spin_unlock_irqrestore(&task_group_lock, flags);
10034 /* wait for possible concurrent references to cfs_rqs complete */
10035 call_rcu(&tg->rcu, free_sched_group_rcu);
10038 /* change task's runqueue when it moves between groups.
10039 * The caller of this function should have put the task in its new group
10040 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10041 * reflect its new group.
10043 void sched_move_task(struct task_struct *tsk)
10045 int on_rq, running;
10046 unsigned long flags;
10049 rq = task_rq_lock(tsk, &flags);
10051 update_rq_clock(rq);
10053 running = task_current(rq, tsk);
10054 on_rq = tsk->se.on_rq;
10057 dequeue_task(rq, tsk, 0);
10058 if (unlikely(running))
10059 tsk->sched_class->put_prev_task(rq, tsk);
10061 set_task_rq(tsk, task_cpu(tsk));
10063 #ifdef CONFIG_FAIR_GROUP_SCHED
10064 if (tsk->sched_class->moved_group)
10065 tsk->sched_class->moved_group(tsk);
10068 if (unlikely(running))
10069 tsk->sched_class->set_curr_task(rq);
10071 enqueue_task(rq, tsk, 0);
10073 task_rq_unlock(rq, &flags);
10075 #endif /* CONFIG_GROUP_SCHED */
10077 #ifdef CONFIG_FAIR_GROUP_SCHED
10078 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10080 struct cfs_rq *cfs_rq = se->cfs_rq;
10085 dequeue_entity(cfs_rq, se, 0);
10087 se->load.weight = shares;
10088 se->load.inv_weight = 0;
10091 enqueue_entity(cfs_rq, se, 0);
10094 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10096 struct cfs_rq *cfs_rq = se->cfs_rq;
10097 struct rq *rq = cfs_rq->rq;
10098 unsigned long flags;
10100 spin_lock_irqsave(&rq->lock, flags);
10101 __set_se_shares(se, shares);
10102 spin_unlock_irqrestore(&rq->lock, flags);
10105 static DEFINE_MUTEX(shares_mutex);
10107 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10110 unsigned long flags;
10113 * We can't change the weight of the root cgroup.
10118 if (shares < MIN_SHARES)
10119 shares = MIN_SHARES;
10120 else if (shares > MAX_SHARES)
10121 shares = MAX_SHARES;
10123 mutex_lock(&shares_mutex);
10124 if (tg->shares == shares)
10127 spin_lock_irqsave(&task_group_lock, flags);
10128 for_each_possible_cpu(i)
10129 unregister_fair_sched_group(tg, i);
10130 list_del_rcu(&tg->siblings);
10131 spin_unlock_irqrestore(&task_group_lock, flags);
10133 /* wait for any ongoing reference to this group to finish */
10134 synchronize_sched();
10137 * Now we are free to modify the group's share on each cpu
10138 * w/o tripping rebalance_share or load_balance_fair.
10140 tg->shares = shares;
10141 for_each_possible_cpu(i) {
10143 * force a rebalance
10145 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10146 set_se_shares(tg->se[i], shares);
10150 * Enable load balance activity on this group, by inserting it back on
10151 * each cpu's rq->leaf_cfs_rq_list.
10153 spin_lock_irqsave(&task_group_lock, flags);
10154 for_each_possible_cpu(i)
10155 register_fair_sched_group(tg, i);
10156 list_add_rcu(&tg->siblings, &tg->parent->children);
10157 spin_unlock_irqrestore(&task_group_lock, flags);
10159 mutex_unlock(&shares_mutex);
10163 unsigned long sched_group_shares(struct task_group *tg)
10169 #ifdef CONFIG_RT_GROUP_SCHED
10171 * Ensure that the real time constraints are schedulable.
10173 static DEFINE_MUTEX(rt_constraints_mutex);
10175 static unsigned long to_ratio(u64 period, u64 runtime)
10177 if (runtime == RUNTIME_INF)
10180 return div64_u64(runtime << 20, period);
10183 /* Must be called with tasklist_lock held */
10184 static inline int tg_has_rt_tasks(struct task_group *tg)
10186 struct task_struct *g, *p;
10188 do_each_thread(g, p) {
10189 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10191 } while_each_thread(g, p);
10196 struct rt_schedulable_data {
10197 struct task_group *tg;
10202 static int tg_schedulable(struct task_group *tg, void *data)
10204 struct rt_schedulable_data *d = data;
10205 struct task_group *child;
10206 unsigned long total, sum = 0;
10207 u64 period, runtime;
10209 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10210 runtime = tg->rt_bandwidth.rt_runtime;
10213 period = d->rt_period;
10214 runtime = d->rt_runtime;
10217 #ifdef CONFIG_USER_SCHED
10218 if (tg == &root_task_group) {
10219 period = global_rt_period();
10220 runtime = global_rt_runtime();
10225 * Cannot have more runtime than the period.
10227 if (runtime > period && runtime != RUNTIME_INF)
10231 * Ensure we don't starve existing RT tasks.
10233 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10236 total = to_ratio(period, runtime);
10239 * Nobody can have more than the global setting allows.
10241 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10245 * The sum of our children's runtime should not exceed our own.
10247 list_for_each_entry_rcu(child, &tg->children, siblings) {
10248 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10249 runtime = child->rt_bandwidth.rt_runtime;
10251 if (child == d->tg) {
10252 period = d->rt_period;
10253 runtime = d->rt_runtime;
10256 sum += to_ratio(period, runtime);
10265 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10267 struct rt_schedulable_data data = {
10269 .rt_period = period,
10270 .rt_runtime = runtime,
10273 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10276 static int tg_set_bandwidth(struct task_group *tg,
10277 u64 rt_period, u64 rt_runtime)
10281 mutex_lock(&rt_constraints_mutex);
10282 read_lock(&tasklist_lock);
10283 err = __rt_schedulable(tg, rt_period, rt_runtime);
10287 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10288 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10289 tg->rt_bandwidth.rt_runtime = rt_runtime;
10291 for_each_possible_cpu(i) {
10292 struct rt_rq *rt_rq = tg->rt_rq[i];
10294 spin_lock(&rt_rq->rt_runtime_lock);
10295 rt_rq->rt_runtime = rt_runtime;
10296 spin_unlock(&rt_rq->rt_runtime_lock);
10298 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10300 read_unlock(&tasklist_lock);
10301 mutex_unlock(&rt_constraints_mutex);
10306 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10308 u64 rt_runtime, rt_period;
10310 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10311 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10312 if (rt_runtime_us < 0)
10313 rt_runtime = RUNTIME_INF;
10315 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10318 long sched_group_rt_runtime(struct task_group *tg)
10322 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10325 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10326 do_div(rt_runtime_us, NSEC_PER_USEC);
10327 return rt_runtime_us;
10330 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10332 u64 rt_runtime, rt_period;
10334 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10335 rt_runtime = tg->rt_bandwidth.rt_runtime;
10337 if (rt_period == 0)
10340 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10343 long sched_group_rt_period(struct task_group *tg)
10347 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10348 do_div(rt_period_us, NSEC_PER_USEC);
10349 return rt_period_us;
10352 static int sched_rt_global_constraints(void)
10354 u64 runtime, period;
10357 if (sysctl_sched_rt_period <= 0)
10360 runtime = global_rt_runtime();
10361 period = global_rt_period();
10364 * Sanity check on the sysctl variables.
10366 if (runtime > period && runtime != RUNTIME_INF)
10369 mutex_lock(&rt_constraints_mutex);
10370 read_lock(&tasklist_lock);
10371 ret = __rt_schedulable(NULL, 0, 0);
10372 read_unlock(&tasklist_lock);
10373 mutex_unlock(&rt_constraints_mutex);
10378 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10380 /* Don't accept realtime tasks when there is no way for them to run */
10381 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10387 #else /* !CONFIG_RT_GROUP_SCHED */
10388 static int sched_rt_global_constraints(void)
10390 unsigned long flags;
10393 if (sysctl_sched_rt_period <= 0)
10397 * There's always some RT tasks in the root group
10398 * -- migration, kstopmachine etc..
10400 if (sysctl_sched_rt_runtime == 0)
10403 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10404 for_each_possible_cpu(i) {
10405 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10407 spin_lock(&rt_rq->rt_runtime_lock);
10408 rt_rq->rt_runtime = global_rt_runtime();
10409 spin_unlock(&rt_rq->rt_runtime_lock);
10411 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10415 #endif /* CONFIG_RT_GROUP_SCHED */
10417 int sched_rt_handler(struct ctl_table *table, int write,
10418 void __user *buffer, size_t *lenp,
10422 int old_period, old_runtime;
10423 static DEFINE_MUTEX(mutex);
10425 mutex_lock(&mutex);
10426 old_period = sysctl_sched_rt_period;
10427 old_runtime = sysctl_sched_rt_runtime;
10429 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10431 if (!ret && write) {
10432 ret = sched_rt_global_constraints();
10434 sysctl_sched_rt_period = old_period;
10435 sysctl_sched_rt_runtime = old_runtime;
10437 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10438 def_rt_bandwidth.rt_period =
10439 ns_to_ktime(global_rt_period());
10442 mutex_unlock(&mutex);
10447 #ifdef CONFIG_CGROUP_SCHED
10449 /* return corresponding task_group object of a cgroup */
10450 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10452 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10453 struct task_group, css);
10456 static struct cgroup_subsys_state *
10457 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10459 struct task_group *tg, *parent;
10461 if (!cgrp->parent) {
10462 /* This is early initialization for the top cgroup */
10463 return &init_task_group.css;
10466 parent = cgroup_tg(cgrp->parent);
10467 tg = sched_create_group(parent);
10469 return ERR_PTR(-ENOMEM);
10475 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10477 struct task_group *tg = cgroup_tg(cgrp);
10479 sched_destroy_group(tg);
10483 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10485 #ifdef CONFIG_RT_GROUP_SCHED
10486 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10489 /* We don't support RT-tasks being in separate groups */
10490 if (tsk->sched_class != &fair_sched_class)
10497 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10498 struct task_struct *tsk, bool threadgroup)
10500 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10504 struct task_struct *c;
10506 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10507 retval = cpu_cgroup_can_attach_task(cgrp, c);
10519 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10520 struct cgroup *old_cont, struct task_struct *tsk,
10523 sched_move_task(tsk);
10525 struct task_struct *c;
10527 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10528 sched_move_task(c);
10534 #ifdef CONFIG_FAIR_GROUP_SCHED
10535 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10538 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10541 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10543 struct task_group *tg = cgroup_tg(cgrp);
10545 return (u64) tg->shares;
10547 #endif /* CONFIG_FAIR_GROUP_SCHED */
10549 #ifdef CONFIG_RT_GROUP_SCHED
10550 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10553 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10556 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10558 return sched_group_rt_runtime(cgroup_tg(cgrp));
10561 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10564 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10567 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10569 return sched_group_rt_period(cgroup_tg(cgrp));
10571 #endif /* CONFIG_RT_GROUP_SCHED */
10573 static struct cftype cpu_files[] = {
10574 #ifdef CONFIG_FAIR_GROUP_SCHED
10577 .read_u64 = cpu_shares_read_u64,
10578 .write_u64 = cpu_shares_write_u64,
10581 #ifdef CONFIG_RT_GROUP_SCHED
10583 .name = "rt_runtime_us",
10584 .read_s64 = cpu_rt_runtime_read,
10585 .write_s64 = cpu_rt_runtime_write,
10588 .name = "rt_period_us",
10589 .read_u64 = cpu_rt_period_read_uint,
10590 .write_u64 = cpu_rt_period_write_uint,
10595 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10597 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10600 struct cgroup_subsys cpu_cgroup_subsys = {
10602 .create = cpu_cgroup_create,
10603 .destroy = cpu_cgroup_destroy,
10604 .can_attach = cpu_cgroup_can_attach,
10605 .attach = cpu_cgroup_attach,
10606 .populate = cpu_cgroup_populate,
10607 .subsys_id = cpu_cgroup_subsys_id,
10611 #endif /* CONFIG_CGROUP_SCHED */
10613 #ifdef CONFIG_CGROUP_CPUACCT
10616 * CPU accounting code for task groups.
10618 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10619 * (balbir@in.ibm.com).
10622 /* track cpu usage of a group of tasks and its child groups */
10624 struct cgroup_subsys_state css;
10625 /* cpuusage holds pointer to a u64-type object on every cpu */
10627 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10628 struct cpuacct *parent;
10631 struct cgroup_subsys cpuacct_subsys;
10633 /* return cpu accounting group corresponding to this container */
10634 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10636 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10637 struct cpuacct, css);
10640 /* return cpu accounting group to which this task belongs */
10641 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10643 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10644 struct cpuacct, css);
10647 /* create a new cpu accounting group */
10648 static struct cgroup_subsys_state *cpuacct_create(
10649 struct cgroup_subsys *ss, struct cgroup *cgrp)
10651 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10657 ca->cpuusage = alloc_percpu(u64);
10661 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10662 if (percpu_counter_init(&ca->cpustat[i], 0))
10663 goto out_free_counters;
10666 ca->parent = cgroup_ca(cgrp->parent);
10672 percpu_counter_destroy(&ca->cpustat[i]);
10673 free_percpu(ca->cpuusage);
10677 return ERR_PTR(-ENOMEM);
10680 /* destroy an existing cpu accounting group */
10682 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10684 struct cpuacct *ca = cgroup_ca(cgrp);
10687 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10688 percpu_counter_destroy(&ca->cpustat[i]);
10689 free_percpu(ca->cpuusage);
10693 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10695 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10698 #ifndef CONFIG_64BIT
10700 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10702 spin_lock_irq(&cpu_rq(cpu)->lock);
10704 spin_unlock_irq(&cpu_rq(cpu)->lock);
10712 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10714 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10716 #ifndef CONFIG_64BIT
10718 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10720 spin_lock_irq(&cpu_rq(cpu)->lock);
10722 spin_unlock_irq(&cpu_rq(cpu)->lock);
10728 /* return total cpu usage (in nanoseconds) of a group */
10729 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10731 struct cpuacct *ca = cgroup_ca(cgrp);
10732 u64 totalcpuusage = 0;
10735 for_each_present_cpu(i)
10736 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10738 return totalcpuusage;
10741 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10744 struct cpuacct *ca = cgroup_ca(cgrp);
10753 for_each_present_cpu(i)
10754 cpuacct_cpuusage_write(ca, i, 0);
10760 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10761 struct seq_file *m)
10763 struct cpuacct *ca = cgroup_ca(cgroup);
10767 for_each_present_cpu(i) {
10768 percpu = cpuacct_cpuusage_read(ca, i);
10769 seq_printf(m, "%llu ", (unsigned long long) percpu);
10771 seq_printf(m, "\n");
10775 static const char *cpuacct_stat_desc[] = {
10776 [CPUACCT_STAT_USER] = "user",
10777 [CPUACCT_STAT_SYSTEM] = "system",
10780 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10781 struct cgroup_map_cb *cb)
10783 struct cpuacct *ca = cgroup_ca(cgrp);
10786 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10787 s64 val = percpu_counter_read(&ca->cpustat[i]);
10788 val = cputime64_to_clock_t(val);
10789 cb->fill(cb, cpuacct_stat_desc[i], val);
10794 static struct cftype files[] = {
10797 .read_u64 = cpuusage_read,
10798 .write_u64 = cpuusage_write,
10801 .name = "usage_percpu",
10802 .read_seq_string = cpuacct_percpu_seq_read,
10806 .read_map = cpuacct_stats_show,
10810 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10812 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10816 * charge this task's execution time to its accounting group.
10818 * called with rq->lock held.
10820 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10822 struct cpuacct *ca;
10825 if (unlikely(!cpuacct_subsys.active))
10828 cpu = task_cpu(tsk);
10834 for (; ca; ca = ca->parent) {
10835 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10836 *cpuusage += cputime;
10843 * Charge the system/user time to the task's accounting group.
10845 static void cpuacct_update_stats(struct task_struct *tsk,
10846 enum cpuacct_stat_index idx, cputime_t val)
10848 struct cpuacct *ca;
10850 if (unlikely(!cpuacct_subsys.active))
10857 percpu_counter_add(&ca->cpustat[idx], val);
10863 struct cgroup_subsys cpuacct_subsys = {
10865 .create = cpuacct_create,
10866 .destroy = cpuacct_destroy,
10867 .populate = cpuacct_populate,
10868 .subsys_id = cpuacct_subsys_id,
10870 #endif /* CONFIG_CGROUP_CPUACCT */
10874 int rcu_expedited_torture_stats(char *page)
10878 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10880 void synchronize_sched_expedited(void)
10883 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10885 #else /* #ifndef CONFIG_SMP */
10887 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10888 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10890 #define RCU_EXPEDITED_STATE_POST -2
10891 #define RCU_EXPEDITED_STATE_IDLE -1
10893 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10895 int rcu_expedited_torture_stats(char *page)
10900 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10901 for_each_online_cpu(cpu) {
10902 cnt += sprintf(&page[cnt], " %d:%d",
10903 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10905 cnt += sprintf(&page[cnt], "\n");
10908 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10910 static long synchronize_sched_expedited_count;
10913 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10914 * approach to force grace period to end quickly. This consumes
10915 * significant time on all CPUs, and is thus not recommended for
10916 * any sort of common-case code.
10918 * Note that it is illegal to call this function while holding any
10919 * lock that is acquired by a CPU-hotplug notifier. Failing to
10920 * observe this restriction will result in deadlock.
10922 void synchronize_sched_expedited(void)
10925 unsigned long flags;
10926 bool need_full_sync = 0;
10928 struct migration_req *req;
10932 smp_mb(); /* ensure prior mod happens before capturing snap. */
10933 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10935 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10937 if (trycount++ < 10)
10938 udelay(trycount * num_online_cpus());
10940 synchronize_sched();
10943 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10944 smp_mb(); /* ensure test happens before caller kfree */
10949 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10950 for_each_online_cpu(cpu) {
10952 req = &per_cpu(rcu_migration_req, cpu);
10953 init_completion(&req->done);
10955 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10956 spin_lock_irqsave(&rq->lock, flags);
10957 list_add(&req->list, &rq->migration_queue);
10958 spin_unlock_irqrestore(&rq->lock, flags);
10959 wake_up_process(rq->migration_thread);
10961 for_each_online_cpu(cpu) {
10962 rcu_expedited_state = cpu;
10963 req = &per_cpu(rcu_migration_req, cpu);
10965 wait_for_completion(&req->done);
10966 spin_lock_irqsave(&rq->lock, flags);
10967 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10968 need_full_sync = 1;
10969 req->dest_cpu = RCU_MIGRATION_IDLE;
10970 spin_unlock_irqrestore(&rq->lock, flags);
10972 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10973 synchronize_sched_expedited_count++;
10974 mutex_unlock(&rcu_sched_expedited_mutex);
10976 if (need_full_sync)
10977 synchronize_sched();
10979 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10981 #endif /* #else #ifndef CONFIG_SMP */