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);
313 static int root_task_group_empty(void)
315 return list_empty(&root_task_group.children);
319 #ifdef CONFIG_FAIR_GROUP_SCHED
320 #ifdef CONFIG_USER_SCHED
321 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
322 #else /* !CONFIG_USER_SCHED */
323 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
324 #endif /* CONFIG_USER_SCHED */
327 * A weight of 0 or 1 can cause arithmetics problems.
328 * A weight of a cfs_rq is the sum of weights of which entities
329 * are queued on this cfs_rq, so a weight of a entity should not be
330 * too large, so as the shares value of a task group.
331 * (The default weight is 1024 - so there's no practical
332 * limitation from this.)
335 #define MAX_SHARES (1UL << 18)
337 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
340 /* Default task group.
341 * Every task in system belong to this group at bootup.
343 struct task_group init_task_group;
345 /* return group to which a task belongs */
346 static inline struct task_group *task_group(struct task_struct *p)
348 struct task_group *tg;
350 #ifdef CONFIG_USER_SCHED
352 tg = __task_cred(p)->user->tg;
354 #elif defined(CONFIG_CGROUP_SCHED)
355 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
356 struct task_group, css);
358 tg = &init_task_group;
363 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
364 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
366 #ifdef CONFIG_FAIR_GROUP_SCHED
367 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
368 p->se.parent = task_group(p)->se[cpu];
371 #ifdef CONFIG_RT_GROUP_SCHED
372 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
373 p->rt.parent = task_group(p)->rt_se[cpu];
379 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
380 static inline struct task_group *task_group(struct task_struct *p)
385 #endif /* CONFIG_GROUP_SCHED */
387 /* CFS-related fields in a runqueue */
389 struct load_weight load;
390 unsigned long nr_running;
395 struct rb_root tasks_timeline;
396 struct rb_node *rb_leftmost;
398 struct list_head tasks;
399 struct list_head *balance_iterator;
402 * 'curr' points to currently running entity on this cfs_rq.
403 * It is set to NULL otherwise (i.e when none are currently running).
405 struct sched_entity *curr, *next, *last;
407 unsigned int nr_spread_over;
409 #ifdef CONFIG_FAIR_GROUP_SCHED
410 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
413 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
414 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
415 * (like users, containers etc.)
417 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
418 * list is used during load balance.
420 struct list_head leaf_cfs_rq_list;
421 struct task_group *tg; /* group that "owns" this runqueue */
425 * the part of load.weight contributed by tasks
427 unsigned long task_weight;
430 * h_load = weight * f(tg)
432 * Where f(tg) is the recursive weight fraction assigned to
435 unsigned long h_load;
438 * this cpu's part of tg->shares
440 unsigned long shares;
443 * load.weight at the time we set shares
445 unsigned long rq_weight;
450 /* Real-Time classes' related field in a runqueue: */
452 struct rt_prio_array active;
453 unsigned long rt_nr_running;
454 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
456 int curr; /* highest queued rt task prio */
458 int next; /* next highest */
463 unsigned long rt_nr_migratory;
464 unsigned long rt_nr_total;
466 struct plist_head pushable_tasks;
471 /* Nests inside the rq lock: */
472 spinlock_t rt_runtime_lock;
474 #ifdef CONFIG_RT_GROUP_SCHED
475 unsigned long rt_nr_boosted;
478 struct list_head leaf_rt_rq_list;
479 struct task_group *tg;
480 struct sched_rt_entity *rt_se;
487 * We add the notion of a root-domain which will be used to define per-domain
488 * variables. Each exclusive cpuset essentially defines an island domain by
489 * fully partitioning the member cpus from any other cpuset. Whenever a new
490 * exclusive cpuset is created, we also create and attach a new root-domain
497 cpumask_var_t online;
500 * The "RT overload" flag: it gets set if a CPU has more than
501 * one runnable RT task.
503 cpumask_var_t rto_mask;
506 struct cpupri cpupri;
511 * By default the system creates a single root-domain with all cpus as
512 * members (mimicking the global state we have today).
514 static struct root_domain def_root_domain;
519 * This is the main, per-CPU runqueue data structure.
521 * Locking rule: those places that want to lock multiple runqueues
522 * (such as the load balancing or the thread migration code), lock
523 * acquire operations must be ordered by ascending &runqueue.
530 * nr_running and cpu_load should be in the same cacheline because
531 * remote CPUs use both these fields when doing load calculation.
533 unsigned long nr_running;
534 #define CPU_LOAD_IDX_MAX 5
535 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
537 unsigned long last_tick_seen;
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;
544 u64 nr_migrations_in;
549 #ifdef CONFIG_FAIR_GROUP_SCHED
550 /* list of leaf cfs_rq on this cpu: */
551 struct list_head leaf_cfs_rq_list;
553 #ifdef CONFIG_RT_GROUP_SCHED
554 struct list_head leaf_rt_rq_list;
558 * This is part of a global counter where only the total sum
559 * over all CPUs matters. A task can increase this counter on
560 * one CPU and if it got migrated afterwards it may decrease
561 * it on another CPU. Always updated under the runqueue lock:
563 unsigned long nr_uninterruptible;
565 struct task_struct *curr, *idle;
566 unsigned long next_balance;
567 struct mm_struct *prev_mm;
574 struct root_domain *rd;
575 struct sched_domain *sd;
577 unsigned char idle_at_tick;
578 /* For active balancing */
582 /* cpu of this runqueue: */
586 unsigned long avg_load_per_task;
588 struct task_struct *migration_thread;
589 struct list_head migration_queue;
595 /* calc_load related fields */
596 unsigned long calc_load_update;
597 long calc_load_active;
599 #ifdef CONFIG_SCHED_HRTICK
601 int hrtick_csd_pending;
602 struct call_single_data hrtick_csd;
604 struct hrtimer hrtick_timer;
607 #ifdef CONFIG_SCHEDSTATS
609 struct sched_info rq_sched_info;
610 unsigned long long rq_cpu_time;
611 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
613 /* sys_sched_yield() stats */
614 unsigned int yld_count;
616 /* schedule() stats */
617 unsigned int sched_switch;
618 unsigned int sched_count;
619 unsigned int sched_goidle;
621 /* try_to_wake_up() stats */
622 unsigned int ttwu_count;
623 unsigned int ttwu_local;
626 unsigned int bkl_count;
630 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
633 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
635 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
638 static inline int cpu_of(struct rq *rq)
648 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
649 * See detach_destroy_domains: synchronize_sched for details.
651 * The domain tree of any CPU may only be accessed from within
652 * preempt-disabled sections.
654 #define for_each_domain(cpu, __sd) \
655 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
657 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
658 #define this_rq() (&__get_cpu_var(runqueues))
659 #define task_rq(p) cpu_rq(task_cpu(p))
660 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
661 #define raw_rq() (&__raw_get_cpu_var(runqueues))
663 inline void update_rq_clock(struct rq *rq)
665 rq->clock = sched_clock_cpu(cpu_of(rq));
669 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
671 #ifdef CONFIG_SCHED_DEBUG
672 # define const_debug __read_mostly
674 # define const_debug static const
680 * Returns true if the current cpu runqueue is locked.
681 * This interface allows printk to be called with the runqueue lock
682 * held and know whether or not it is OK to wake up the klogd.
684 int runqueue_is_locked(int cpu)
686 return spin_is_locked(&cpu_rq(cpu)->lock);
690 * Debugging: various feature bits
693 #define SCHED_FEAT(name, enabled) \
694 __SCHED_FEAT_##name ,
697 #include "sched_features.h"
702 #define SCHED_FEAT(name, enabled) \
703 (1UL << __SCHED_FEAT_##name) * enabled |
705 const_debug unsigned int sysctl_sched_features =
706 #include "sched_features.h"
711 #ifdef CONFIG_SCHED_DEBUG
712 #define SCHED_FEAT(name, enabled) \
715 static __read_mostly char *sched_feat_names[] = {
716 #include "sched_features.h"
722 static int sched_feat_show(struct seq_file *m, void *v)
726 for (i = 0; sched_feat_names[i]; i++) {
727 if (!(sysctl_sched_features & (1UL << i)))
729 seq_printf(m, "%s ", sched_feat_names[i]);
737 sched_feat_write(struct file *filp, const char __user *ubuf,
738 size_t cnt, loff_t *ppos)
748 if (copy_from_user(&buf, ubuf, cnt))
753 if (strncmp(buf, "NO_", 3) == 0) {
758 for (i = 0; sched_feat_names[i]; i++) {
759 int len = strlen(sched_feat_names[i]);
761 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
763 sysctl_sched_features &= ~(1UL << i);
765 sysctl_sched_features |= (1UL << i);
770 if (!sched_feat_names[i])
778 static int sched_feat_open(struct inode *inode, struct file *filp)
780 return single_open(filp, sched_feat_show, NULL);
783 static const struct file_operations sched_feat_fops = {
784 .open = sched_feat_open,
785 .write = sched_feat_write,
788 .release = single_release,
791 static __init int sched_init_debug(void)
793 debugfs_create_file("sched_features", 0644, NULL, NULL,
798 late_initcall(sched_init_debug);
802 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
805 * Number of tasks to iterate in a single balance run.
806 * Limited because this is done with IRQs disabled.
808 const_debug unsigned int sysctl_sched_nr_migrate = 32;
811 * ratelimit for updating the group shares.
814 unsigned int sysctl_sched_shares_ratelimit = 250000;
817 * Inject some fuzzyness into changing the per-cpu group shares
818 * this avoids remote rq-locks at the expense of fairness.
821 unsigned int sysctl_sched_shares_thresh = 4;
824 * period over which we average the RT time consumption, measured
829 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
832 * period over which we measure -rt task cpu usage in us.
835 unsigned int sysctl_sched_rt_period = 1000000;
837 static __read_mostly int scheduler_running;
840 * part of the period that we allow rt tasks to run in us.
843 int sysctl_sched_rt_runtime = 950000;
845 static inline u64 global_rt_period(void)
847 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
850 static inline u64 global_rt_runtime(void)
852 if (sysctl_sched_rt_runtime < 0)
855 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
858 #ifndef prepare_arch_switch
859 # define prepare_arch_switch(next) do { } while (0)
861 #ifndef finish_arch_switch
862 # define finish_arch_switch(prev) do { } while (0)
865 static inline int task_current(struct rq *rq, struct task_struct *p)
867 return rq->curr == p;
870 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
871 static inline int task_running(struct rq *rq, struct task_struct *p)
873 return task_current(rq, p);
876 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
880 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
882 #ifdef CONFIG_DEBUG_SPINLOCK
883 /* this is a valid case when another task releases the spinlock */
884 rq->lock.owner = current;
887 * If we are tracking spinlock dependencies then we have to
888 * fix up the runqueue lock - which gets 'carried over' from
891 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
893 spin_unlock_irq(&rq->lock);
896 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
897 static inline int task_running(struct rq *rq, struct task_struct *p)
902 return task_current(rq, p);
906 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
910 * We can optimise this out completely for !SMP, because the
911 * SMP rebalancing from interrupt is the only thing that cares
916 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
917 spin_unlock_irq(&rq->lock);
919 spin_unlock(&rq->lock);
923 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
927 * After ->oncpu is cleared, the task can be moved to a different CPU.
928 * We must ensure this doesn't happen until the switch is completely
934 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
938 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
941 * __task_rq_lock - lock the runqueue a given task resides on.
942 * Must be called interrupts disabled.
944 static inline struct rq *__task_rq_lock(struct task_struct *p)
948 struct rq *rq = task_rq(p);
949 spin_lock(&rq->lock);
950 if (likely(rq == task_rq(p)))
952 spin_unlock(&rq->lock);
957 * task_rq_lock - lock the runqueue a given task resides on and disable
958 * interrupts. Note the ordering: we can safely lookup the task_rq without
959 * explicitly disabling preemption.
961 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
967 local_irq_save(*flags);
969 spin_lock(&rq->lock);
970 if (likely(rq == task_rq(p)))
972 spin_unlock_irqrestore(&rq->lock, *flags);
976 void task_rq_unlock_wait(struct task_struct *p)
978 struct rq *rq = task_rq(p);
980 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
981 spin_unlock_wait(&rq->lock);
984 static void __task_rq_unlock(struct rq *rq)
987 spin_unlock(&rq->lock);
990 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
993 spin_unlock_irqrestore(&rq->lock, *flags);
997 * this_rq_lock - lock this runqueue and disable interrupts.
999 static struct rq *this_rq_lock(void)
1000 __acquires(rq->lock)
1004 local_irq_disable();
1006 spin_lock(&rq->lock);
1011 #ifdef CONFIG_SCHED_HRTICK
1013 * Use HR-timers to deliver accurate preemption points.
1015 * Its all a bit involved since we cannot program an hrt while holding the
1016 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1019 * When we get rescheduled we reprogram the hrtick_timer outside of the
1025 * - enabled by features
1026 * - hrtimer is actually high res
1028 static inline int hrtick_enabled(struct rq *rq)
1030 if (!sched_feat(HRTICK))
1032 if (!cpu_active(cpu_of(rq)))
1034 return hrtimer_is_hres_active(&rq->hrtick_timer);
1037 static void hrtick_clear(struct rq *rq)
1039 if (hrtimer_active(&rq->hrtick_timer))
1040 hrtimer_cancel(&rq->hrtick_timer);
1044 * High-resolution timer tick.
1045 * Runs from hardirq context with interrupts disabled.
1047 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1049 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1051 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1053 spin_lock(&rq->lock);
1054 update_rq_clock(rq);
1055 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1056 spin_unlock(&rq->lock);
1058 return HRTIMER_NORESTART;
1063 * called from hardirq (IPI) context
1065 static void __hrtick_start(void *arg)
1067 struct rq *rq = arg;
1069 spin_lock(&rq->lock);
1070 hrtimer_restart(&rq->hrtick_timer);
1071 rq->hrtick_csd_pending = 0;
1072 spin_unlock(&rq->lock);
1076 * Called to set the hrtick timer state.
1078 * called with rq->lock held and irqs disabled
1080 static void hrtick_start(struct rq *rq, u64 delay)
1082 struct hrtimer *timer = &rq->hrtick_timer;
1083 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1085 hrtimer_set_expires(timer, time);
1087 if (rq == this_rq()) {
1088 hrtimer_restart(timer);
1089 } else if (!rq->hrtick_csd_pending) {
1090 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1091 rq->hrtick_csd_pending = 1;
1096 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1098 int cpu = (int)(long)hcpu;
1101 case CPU_UP_CANCELED:
1102 case CPU_UP_CANCELED_FROZEN:
1103 case CPU_DOWN_PREPARE:
1104 case CPU_DOWN_PREPARE_FROZEN:
1106 case CPU_DEAD_FROZEN:
1107 hrtick_clear(cpu_rq(cpu));
1114 static __init void init_hrtick(void)
1116 hotcpu_notifier(hotplug_hrtick, 0);
1120 * Called to set the hrtick timer state.
1122 * called with rq->lock held and irqs disabled
1124 static void hrtick_start(struct rq *rq, u64 delay)
1126 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1127 HRTIMER_MODE_REL_PINNED, 0);
1130 static inline void init_hrtick(void)
1133 #endif /* CONFIG_SMP */
1135 static void init_rq_hrtick(struct rq *rq)
1138 rq->hrtick_csd_pending = 0;
1140 rq->hrtick_csd.flags = 0;
1141 rq->hrtick_csd.func = __hrtick_start;
1142 rq->hrtick_csd.info = rq;
1145 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1146 rq->hrtick_timer.function = hrtick;
1148 #else /* CONFIG_SCHED_HRTICK */
1149 static inline void hrtick_clear(struct rq *rq)
1153 static inline void init_rq_hrtick(struct rq *rq)
1157 static inline void init_hrtick(void)
1160 #endif /* CONFIG_SCHED_HRTICK */
1163 * resched_task - mark a task 'to be rescheduled now'.
1165 * On UP this means the setting of the need_resched flag, on SMP it
1166 * might also involve a cross-CPU call to trigger the scheduler on
1171 #ifndef tsk_is_polling
1172 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1175 static void resched_task(struct task_struct *p)
1179 assert_spin_locked(&task_rq(p)->lock);
1181 if (test_tsk_need_resched(p))
1184 set_tsk_need_resched(p);
1187 if (cpu == smp_processor_id())
1190 /* NEED_RESCHED must be visible before we test polling */
1192 if (!tsk_is_polling(p))
1193 smp_send_reschedule(cpu);
1196 static void resched_cpu(int cpu)
1198 struct rq *rq = cpu_rq(cpu);
1199 unsigned long flags;
1201 if (!spin_trylock_irqsave(&rq->lock, flags))
1203 resched_task(cpu_curr(cpu));
1204 spin_unlock_irqrestore(&rq->lock, flags);
1209 * When add_timer_on() enqueues a timer into the timer wheel of an
1210 * idle CPU then this timer might expire before the next timer event
1211 * which is scheduled to wake up that CPU. In case of a completely
1212 * idle system the next event might even be infinite time into the
1213 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1214 * leaves the inner idle loop so the newly added timer is taken into
1215 * account when the CPU goes back to idle and evaluates the timer
1216 * wheel for the next timer event.
1218 void wake_up_idle_cpu(int cpu)
1220 struct rq *rq = cpu_rq(cpu);
1222 if (cpu == smp_processor_id())
1226 * This is safe, as this function is called with the timer
1227 * wheel base lock of (cpu) held. When the CPU is on the way
1228 * to idle and has not yet set rq->curr to idle then it will
1229 * be serialized on the timer wheel base lock and take the new
1230 * timer into account automatically.
1232 if (rq->curr != rq->idle)
1236 * We can set TIF_RESCHED on the idle task of the other CPU
1237 * lockless. The worst case is that the other CPU runs the
1238 * idle task through an additional NOOP schedule()
1240 set_tsk_need_resched(rq->idle);
1242 /* NEED_RESCHED must be visible before we test polling */
1244 if (!tsk_is_polling(rq->idle))
1245 smp_send_reschedule(cpu);
1247 #endif /* CONFIG_NO_HZ */
1249 static u64 sched_avg_period(void)
1251 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1254 static void sched_avg_update(struct rq *rq)
1256 s64 period = sched_avg_period();
1258 while ((s64)(rq->clock - rq->age_stamp) > period) {
1259 rq->age_stamp += period;
1264 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1266 rq->rt_avg += rt_delta;
1267 sched_avg_update(rq);
1270 #else /* !CONFIG_SMP */
1271 static void resched_task(struct task_struct *p)
1273 assert_spin_locked(&task_rq(p)->lock);
1274 set_tsk_need_resched(p);
1277 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1280 #endif /* CONFIG_SMP */
1282 #if BITS_PER_LONG == 32
1283 # define WMULT_CONST (~0UL)
1285 # define WMULT_CONST (1UL << 32)
1288 #define WMULT_SHIFT 32
1291 * Shift right and round:
1293 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1296 * delta *= weight / lw
1298 static unsigned long
1299 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1300 struct load_weight *lw)
1304 if (!lw->inv_weight) {
1305 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1308 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1312 tmp = (u64)delta_exec * weight;
1314 * Check whether we'd overflow the 64-bit multiplication:
1316 if (unlikely(tmp > WMULT_CONST))
1317 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1320 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1322 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1325 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1331 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1338 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1339 * of tasks with abnormal "nice" values across CPUs the contribution that
1340 * each task makes to its run queue's load is weighted according to its
1341 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1342 * scaled version of the new time slice allocation that they receive on time
1346 #define WEIGHT_IDLEPRIO 3
1347 #define WMULT_IDLEPRIO 1431655765
1350 * Nice levels are multiplicative, with a gentle 10% change for every
1351 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1352 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1353 * that remained on nice 0.
1355 * The "10% effect" is relative and cumulative: from _any_ nice level,
1356 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1357 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1358 * If a task goes up by ~10% and another task goes down by ~10% then
1359 * the relative distance between them is ~25%.)
1361 static const int prio_to_weight[40] = {
1362 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1363 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1364 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1365 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1366 /* 0 */ 1024, 820, 655, 526, 423,
1367 /* 5 */ 335, 272, 215, 172, 137,
1368 /* 10 */ 110, 87, 70, 56, 45,
1369 /* 15 */ 36, 29, 23, 18, 15,
1373 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1375 * In cases where the weight does not change often, we can use the
1376 * precalculated inverse to speed up arithmetics by turning divisions
1377 * into multiplications:
1379 static const u32 prio_to_wmult[40] = {
1380 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1381 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1382 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1383 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1384 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1385 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1386 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1387 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1390 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1393 * runqueue iterator, to support SMP load-balancing between different
1394 * scheduling classes, without having to expose their internal data
1395 * structures to the load-balancing proper:
1397 struct rq_iterator {
1399 struct task_struct *(*start)(void *);
1400 struct task_struct *(*next)(void *);
1404 static unsigned long
1405 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1406 unsigned long max_load_move, struct sched_domain *sd,
1407 enum cpu_idle_type idle, int *all_pinned,
1408 int *this_best_prio, struct rq_iterator *iterator);
1411 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1412 struct sched_domain *sd, enum cpu_idle_type idle,
1413 struct rq_iterator *iterator);
1416 /* Time spent by the tasks of the cpu accounting group executing in ... */
1417 enum cpuacct_stat_index {
1418 CPUACCT_STAT_USER, /* ... user mode */
1419 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1421 CPUACCT_STAT_NSTATS,
1424 #ifdef CONFIG_CGROUP_CPUACCT
1425 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1426 static void cpuacct_update_stats(struct task_struct *tsk,
1427 enum cpuacct_stat_index idx, cputime_t val);
1429 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1430 static inline void cpuacct_update_stats(struct task_struct *tsk,
1431 enum cpuacct_stat_index idx, cputime_t val) {}
1434 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1436 update_load_add(&rq->load, load);
1439 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1441 update_load_sub(&rq->load, load);
1444 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1445 typedef int (*tg_visitor)(struct task_group *, void *);
1448 * Iterate the full tree, calling @down when first entering a node and @up when
1449 * leaving it for the final time.
1451 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1453 struct task_group *parent, *child;
1457 parent = &root_task_group;
1459 ret = (*down)(parent, data);
1462 list_for_each_entry_rcu(child, &parent->children, siblings) {
1469 ret = (*up)(parent, data);
1474 parent = parent->parent;
1483 static int tg_nop(struct task_group *tg, void *data)
1490 /* Used instead of source_load when we know the type == 0 */
1491 static unsigned long weighted_cpuload(const int cpu)
1493 return cpu_rq(cpu)->load.weight;
1497 * Return a low guess at the load of a migration-source cpu weighted
1498 * according to the scheduling class and "nice" value.
1500 * We want to under-estimate the load of migration sources, to
1501 * balance conservatively.
1503 static unsigned long source_load(int cpu, int type)
1505 struct rq *rq = cpu_rq(cpu);
1506 unsigned long total = weighted_cpuload(cpu);
1508 if (type == 0 || !sched_feat(LB_BIAS))
1511 return min(rq->cpu_load[type-1], total);
1515 * Return a high guess at the load of a migration-target cpu weighted
1516 * according to the scheduling class and "nice" value.
1518 static unsigned long target_load(int cpu, int type)
1520 struct rq *rq = cpu_rq(cpu);
1521 unsigned long total = weighted_cpuload(cpu);
1523 if (type == 0 || !sched_feat(LB_BIAS))
1526 return max(rq->cpu_load[type-1], total);
1529 static struct sched_group *group_of(int cpu)
1531 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1539 static unsigned long power_of(int cpu)
1541 struct sched_group *group = group_of(cpu);
1544 return SCHED_LOAD_SCALE;
1546 return group->cpu_power;
1549 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1551 static unsigned long cpu_avg_load_per_task(int cpu)
1553 struct rq *rq = cpu_rq(cpu);
1554 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1557 rq->avg_load_per_task = rq->load.weight / nr_running;
1559 rq->avg_load_per_task = 0;
1561 return rq->avg_load_per_task;
1564 #ifdef CONFIG_FAIR_GROUP_SCHED
1566 struct update_shares_data {
1567 unsigned long rq_weight[NR_CPUS];
1570 static DEFINE_PER_CPU(struct update_shares_data, update_shares_data);
1572 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1575 * Calculate and set the cpu's group shares.
1577 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1578 unsigned long sd_shares,
1579 unsigned long sd_rq_weight,
1580 struct update_shares_data *usd)
1582 unsigned long shares, rq_weight;
1585 rq_weight = usd->rq_weight[cpu];
1588 rq_weight = NICE_0_LOAD;
1592 * \Sum_j shares_j * rq_weight_i
1593 * shares_i = -----------------------------
1594 * \Sum_j rq_weight_j
1596 shares = (sd_shares * rq_weight) / sd_rq_weight;
1597 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1599 if (abs(shares - tg->se[cpu]->load.weight) >
1600 sysctl_sched_shares_thresh) {
1601 struct rq *rq = cpu_rq(cpu);
1602 unsigned long flags;
1604 spin_lock_irqsave(&rq->lock, flags);
1605 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1606 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1607 __set_se_shares(tg->se[cpu], shares);
1608 spin_unlock_irqrestore(&rq->lock, flags);
1613 * Re-compute the task group their per cpu shares over the given domain.
1614 * This needs to be done in a bottom-up fashion because the rq weight of a
1615 * parent group depends on the shares of its child groups.
1617 static int tg_shares_up(struct task_group *tg, void *data)
1619 unsigned long weight, rq_weight = 0, shares = 0;
1620 struct update_shares_data *usd;
1621 struct sched_domain *sd = data;
1622 unsigned long flags;
1628 local_irq_save(flags);
1629 usd = &__get_cpu_var(update_shares_data);
1631 for_each_cpu(i, sched_domain_span(sd)) {
1632 weight = tg->cfs_rq[i]->load.weight;
1633 usd->rq_weight[i] = 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 rq_weight += weight;
1644 shares += tg->cfs_rq[i]->shares;
1647 if ((!shares && rq_weight) || shares > tg->shares)
1648 shares = tg->shares;
1650 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1651 shares = tg->shares;
1653 for_each_cpu(i, sched_domain_span(sd))
1654 update_group_shares_cpu(tg, i, shares, rq_weight, usd);
1656 local_irq_restore(flags);
1662 * Compute the cpu's hierarchical load factor for each task group.
1663 * This needs to be done in a top-down fashion because the load of a child
1664 * group is a fraction of its parents load.
1666 static int tg_load_down(struct task_group *tg, void *data)
1669 long cpu = (long)data;
1672 load = cpu_rq(cpu)->load.weight;
1674 load = tg->parent->cfs_rq[cpu]->h_load;
1675 load *= tg->cfs_rq[cpu]->shares;
1676 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1679 tg->cfs_rq[cpu]->h_load = load;
1684 static void update_shares(struct sched_domain *sd)
1689 if (root_task_group_empty())
1692 now = cpu_clock(raw_smp_processor_id());
1693 elapsed = now - sd->last_update;
1695 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1696 sd->last_update = now;
1697 walk_tg_tree(tg_nop, tg_shares_up, sd);
1701 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1703 if (root_task_group_empty())
1706 spin_unlock(&rq->lock);
1708 spin_lock(&rq->lock);
1711 static void update_h_load(long cpu)
1713 if (root_task_group_empty())
1716 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1721 static inline void update_shares(struct sched_domain *sd)
1725 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1731 #ifdef CONFIG_PREEMPT
1733 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1736 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1737 * way at the expense of forcing extra atomic operations in all
1738 * invocations. This assures that the double_lock is acquired using the
1739 * same underlying policy as the spinlock_t on this architecture, which
1740 * reduces latency compared to the unfair variant below. However, it
1741 * also adds more overhead and therefore may reduce throughput.
1743 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1744 __releases(this_rq->lock)
1745 __acquires(busiest->lock)
1746 __acquires(this_rq->lock)
1748 spin_unlock(&this_rq->lock);
1749 double_rq_lock(this_rq, busiest);
1756 * Unfair double_lock_balance: Optimizes throughput at the expense of
1757 * latency by eliminating extra atomic operations when the locks are
1758 * already in proper order on entry. This favors lower cpu-ids and will
1759 * grant the double lock to lower cpus over higher ids under contention,
1760 * regardless of entry order into the function.
1762 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1763 __releases(this_rq->lock)
1764 __acquires(busiest->lock)
1765 __acquires(this_rq->lock)
1769 if (unlikely(!spin_trylock(&busiest->lock))) {
1770 if (busiest < this_rq) {
1771 spin_unlock(&this_rq->lock);
1772 spin_lock(&busiest->lock);
1773 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1776 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1781 #endif /* CONFIG_PREEMPT */
1784 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1786 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1788 if (unlikely(!irqs_disabled())) {
1789 /* printk() doesn't work good under rq->lock */
1790 spin_unlock(&this_rq->lock);
1794 return _double_lock_balance(this_rq, busiest);
1797 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1798 __releases(busiest->lock)
1800 spin_unlock(&busiest->lock);
1801 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1805 #ifdef CONFIG_FAIR_GROUP_SCHED
1806 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1809 cfs_rq->shares = shares;
1814 static void calc_load_account_active(struct rq *this_rq);
1816 #include "sched_stats.h"
1817 #include "sched_idletask.c"
1818 #include "sched_fair.c"
1819 #include "sched_rt.c"
1820 #ifdef CONFIG_SCHED_DEBUG
1821 # include "sched_debug.c"
1824 #define sched_class_highest (&rt_sched_class)
1825 #define for_each_class(class) \
1826 for (class = sched_class_highest; class; class = class->next)
1828 static void inc_nr_running(struct rq *rq)
1833 static void dec_nr_running(struct rq *rq)
1838 static void set_load_weight(struct task_struct *p)
1840 if (task_has_rt_policy(p)) {
1841 p->se.load.weight = prio_to_weight[0] * 2;
1842 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1847 * SCHED_IDLE tasks get minimal weight:
1849 if (p->policy == SCHED_IDLE) {
1850 p->se.load.weight = WEIGHT_IDLEPRIO;
1851 p->se.load.inv_weight = WMULT_IDLEPRIO;
1855 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1856 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1859 static void update_avg(u64 *avg, u64 sample)
1861 s64 diff = sample - *avg;
1865 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1868 p->se.start_runtime = p->se.sum_exec_runtime;
1870 sched_info_queued(p);
1871 p->sched_class->enqueue_task(rq, p, wakeup);
1875 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1878 if (p->se.last_wakeup) {
1879 update_avg(&p->se.avg_overlap,
1880 p->se.sum_exec_runtime - p->se.last_wakeup);
1881 p->se.last_wakeup = 0;
1883 update_avg(&p->se.avg_wakeup,
1884 sysctl_sched_wakeup_granularity);
1888 sched_info_dequeued(p);
1889 p->sched_class->dequeue_task(rq, p, sleep);
1894 * __normal_prio - return the priority that is based on the static prio
1896 static inline int __normal_prio(struct task_struct *p)
1898 return p->static_prio;
1902 * Calculate the expected normal priority: i.e. priority
1903 * without taking RT-inheritance into account. Might be
1904 * boosted by interactivity modifiers. Changes upon fork,
1905 * setprio syscalls, and whenever the interactivity
1906 * estimator recalculates.
1908 static inline int normal_prio(struct task_struct *p)
1912 if (task_has_rt_policy(p))
1913 prio = MAX_RT_PRIO-1 - p->rt_priority;
1915 prio = __normal_prio(p);
1920 * Calculate the current priority, i.e. the priority
1921 * taken into account by the scheduler. This value might
1922 * be boosted by RT tasks, or might be boosted by
1923 * interactivity modifiers. Will be RT if the task got
1924 * RT-boosted. If not then it returns p->normal_prio.
1926 static int effective_prio(struct task_struct *p)
1928 p->normal_prio = normal_prio(p);
1930 * If we are RT tasks or we were boosted to RT priority,
1931 * keep the priority unchanged. Otherwise, update priority
1932 * to the normal priority:
1934 if (!rt_prio(p->prio))
1935 return p->normal_prio;
1940 * activate_task - move a task to the runqueue.
1942 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1944 if (task_contributes_to_load(p))
1945 rq->nr_uninterruptible--;
1947 enqueue_task(rq, p, wakeup);
1952 * deactivate_task - remove a task from the runqueue.
1954 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1956 if (task_contributes_to_load(p))
1957 rq->nr_uninterruptible++;
1959 dequeue_task(rq, p, sleep);
1964 * task_curr - is this task currently executing on a CPU?
1965 * @p: the task in question.
1967 inline int task_curr(const struct task_struct *p)
1969 return cpu_curr(task_cpu(p)) == p;
1972 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1974 set_task_rq(p, cpu);
1977 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1978 * successfuly executed on another CPU. We must ensure that updates of
1979 * per-task data have been completed by this moment.
1982 task_thread_info(p)->cpu = cpu;
1986 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1987 const struct sched_class *prev_class,
1988 int oldprio, int running)
1990 if (prev_class != p->sched_class) {
1991 if (prev_class->switched_from)
1992 prev_class->switched_from(rq, p, running);
1993 p->sched_class->switched_to(rq, p, running);
1995 p->sched_class->prio_changed(rq, p, oldprio, running);
2000 * Is this task likely cache-hot:
2003 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2008 * Buddy candidates are cache hot:
2010 if (sched_feat(CACHE_HOT_BUDDY) &&
2011 (&p->se == cfs_rq_of(&p->se)->next ||
2012 &p->se == cfs_rq_of(&p->se)->last))
2015 if (p->sched_class != &fair_sched_class)
2018 if (sysctl_sched_migration_cost == -1)
2020 if (sysctl_sched_migration_cost == 0)
2023 delta = now - p->se.exec_start;
2025 return delta < (s64)sysctl_sched_migration_cost;
2029 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2031 int old_cpu = task_cpu(p);
2032 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2033 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2034 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2037 clock_offset = old_rq->clock - new_rq->clock;
2039 trace_sched_migrate_task(p, new_cpu);
2041 #ifdef CONFIG_SCHEDSTATS
2042 if (p->se.wait_start)
2043 p->se.wait_start -= clock_offset;
2044 if (p->se.sleep_start)
2045 p->se.sleep_start -= clock_offset;
2046 if (p->se.block_start)
2047 p->se.block_start -= clock_offset;
2049 if (old_cpu != new_cpu) {
2050 p->se.nr_migrations++;
2051 new_rq->nr_migrations_in++;
2052 #ifdef CONFIG_SCHEDSTATS
2053 if (task_hot(p, old_rq->clock, NULL))
2054 schedstat_inc(p, se.nr_forced2_migrations);
2056 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2059 p->se.vruntime -= old_cfsrq->min_vruntime -
2060 new_cfsrq->min_vruntime;
2062 __set_task_cpu(p, new_cpu);
2065 struct migration_req {
2066 struct list_head list;
2068 struct task_struct *task;
2071 struct completion done;
2075 * The task's runqueue lock must be held.
2076 * Returns true if you have to wait for migration thread.
2079 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2081 struct rq *rq = task_rq(p);
2084 * If the task is not on a runqueue (and not running), then
2085 * it is sufficient to simply update the task's cpu field.
2087 if (!p->se.on_rq && !task_running(rq, p)) {
2088 set_task_cpu(p, dest_cpu);
2092 init_completion(&req->done);
2094 req->dest_cpu = dest_cpu;
2095 list_add(&req->list, &rq->migration_queue);
2101 * wait_task_context_switch - wait for a thread to complete at least one
2104 * @p must not be current.
2106 void wait_task_context_switch(struct task_struct *p)
2108 unsigned long nvcsw, nivcsw, flags;
2116 * The runqueue is assigned before the actual context
2117 * switch. We need to take the runqueue lock.
2119 * We could check initially without the lock but it is
2120 * very likely that we need to take the lock in every
2123 rq = task_rq_lock(p, &flags);
2124 running = task_running(rq, p);
2125 task_rq_unlock(rq, &flags);
2127 if (likely(!running))
2130 * The switch count is incremented before the actual
2131 * context switch. We thus wait for two switches to be
2132 * sure at least one completed.
2134 if ((p->nvcsw - nvcsw) > 1)
2136 if ((p->nivcsw - nivcsw) > 1)
2144 * wait_task_inactive - wait for a thread to unschedule.
2146 * If @match_state is nonzero, it's the @p->state value just checked and
2147 * not expected to change. If it changes, i.e. @p might have woken up,
2148 * then return zero. When we succeed in waiting for @p to be off its CPU,
2149 * we return a positive number (its total switch count). If a second call
2150 * a short while later returns the same number, the caller can be sure that
2151 * @p has remained unscheduled the whole time.
2153 * The caller must ensure that the task *will* unschedule sometime soon,
2154 * else this function might spin for a *long* time. This function can't
2155 * be called with interrupts off, or it may introduce deadlock with
2156 * smp_call_function() if an IPI is sent by the same process we are
2157 * waiting to become inactive.
2159 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2161 unsigned long flags;
2168 * We do the initial early heuristics without holding
2169 * any task-queue locks at all. We'll only try to get
2170 * the runqueue lock when things look like they will
2176 * If the task is actively running on another CPU
2177 * still, just relax and busy-wait without holding
2180 * NOTE! Since we don't hold any locks, it's not
2181 * even sure that "rq" stays as the right runqueue!
2182 * But we don't care, since "task_running()" will
2183 * return false if the runqueue has changed and p
2184 * is actually now running somewhere else!
2186 while (task_running(rq, p)) {
2187 if (match_state && unlikely(p->state != match_state))
2193 * Ok, time to look more closely! We need the rq
2194 * lock now, to be *sure*. If we're wrong, we'll
2195 * just go back and repeat.
2197 rq = task_rq_lock(p, &flags);
2198 trace_sched_wait_task(rq, p);
2199 running = task_running(rq, p);
2200 on_rq = p->se.on_rq;
2202 if (!match_state || p->state == match_state)
2203 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2204 task_rq_unlock(rq, &flags);
2207 * If it changed from the expected state, bail out now.
2209 if (unlikely(!ncsw))
2213 * Was it really running after all now that we
2214 * checked with the proper locks actually held?
2216 * Oops. Go back and try again..
2218 if (unlikely(running)) {
2224 * It's not enough that it's not actively running,
2225 * it must be off the runqueue _entirely_, and not
2228 * So if it was still runnable (but just not actively
2229 * running right now), it's preempted, and we should
2230 * yield - it could be a while.
2232 if (unlikely(on_rq)) {
2233 schedule_timeout_uninterruptible(1);
2238 * Ahh, all good. It wasn't running, and it wasn't
2239 * runnable, which means that it will never become
2240 * running in the future either. We're all done!
2249 * kick_process - kick a running thread to enter/exit the kernel
2250 * @p: the to-be-kicked thread
2252 * Cause a process which is running on another CPU to enter
2253 * kernel-mode, without any delay. (to get signals handled.)
2255 * NOTE: this function doesnt have to take the runqueue lock,
2256 * because all it wants to ensure is that the remote task enters
2257 * the kernel. If the IPI races and the task has been migrated
2258 * to another CPU then no harm is done and the purpose has been
2261 void kick_process(struct task_struct *p)
2267 if ((cpu != smp_processor_id()) && task_curr(p))
2268 smp_send_reschedule(cpu);
2271 EXPORT_SYMBOL_GPL(kick_process);
2272 #endif /* CONFIG_SMP */
2275 * task_oncpu_function_call - call a function on the cpu on which a task runs
2276 * @p: the task to evaluate
2277 * @func: the function to be called
2278 * @info: the function call argument
2280 * Calls the function @func when the task is currently running. This might
2281 * be on the current CPU, which just calls the function directly
2283 void task_oncpu_function_call(struct task_struct *p,
2284 void (*func) (void *info), void *info)
2291 smp_call_function_single(cpu, func, info, 1);
2296 * try_to_wake_up - wake up a thread
2297 * @p: the to-be-woken-up thread
2298 * @state: the mask of task states that can be woken
2299 * @sync: do a synchronous wakeup?
2301 * Put it on the run-queue if it's not already there. The "current"
2302 * thread is always on the run-queue (except when the actual
2303 * re-schedule is in progress), and as such you're allowed to do
2304 * the simpler "current->state = TASK_RUNNING" to mark yourself
2305 * runnable without the overhead of this.
2307 * returns failure only if the task is already active.
2309 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2312 int cpu, orig_cpu, this_cpu, success = 0;
2313 unsigned long flags;
2314 struct rq *rq, *orig_rq;
2316 if (!sched_feat(SYNC_WAKEUPS))
2317 wake_flags &= ~WF_SYNC;
2319 this_cpu = get_cpu();
2322 rq = orig_rq = task_rq_lock(p, &flags);
2323 update_rq_clock(rq);
2324 if (!(p->state & state))
2334 if (unlikely(task_running(rq, p)))
2338 * In order to handle concurrent wakeups and release the rq->lock
2339 * we put the task in TASK_WAKING state.
2341 * First fix up the nr_uninterruptible count:
2343 if (task_contributes_to_load(p))
2344 rq->nr_uninterruptible--;
2345 p->state = TASK_WAKING;
2346 task_rq_unlock(rq, &flags);
2348 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2349 if (cpu != orig_cpu)
2350 set_task_cpu(p, cpu);
2352 rq = task_rq_lock(p, &flags);
2355 update_rq_clock(rq);
2357 WARN_ON(p->state != TASK_WAKING);
2360 #ifdef CONFIG_SCHEDSTATS
2361 schedstat_inc(rq, ttwu_count);
2362 if (cpu == this_cpu)
2363 schedstat_inc(rq, ttwu_local);
2365 struct sched_domain *sd;
2366 for_each_domain(this_cpu, sd) {
2367 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2368 schedstat_inc(sd, ttwu_wake_remote);
2373 #endif /* CONFIG_SCHEDSTATS */
2376 #endif /* CONFIG_SMP */
2377 schedstat_inc(p, se.nr_wakeups);
2378 if (wake_flags & WF_SYNC)
2379 schedstat_inc(p, se.nr_wakeups_sync);
2380 if (orig_cpu != cpu)
2381 schedstat_inc(p, se.nr_wakeups_migrate);
2382 if (cpu == this_cpu)
2383 schedstat_inc(p, se.nr_wakeups_local);
2385 schedstat_inc(p, se.nr_wakeups_remote);
2386 activate_task(rq, p, 1);
2390 * Only attribute actual wakeups done by this task.
2392 if (!in_interrupt()) {
2393 struct sched_entity *se = ¤t->se;
2394 u64 sample = se->sum_exec_runtime;
2396 if (se->last_wakeup)
2397 sample -= se->last_wakeup;
2399 sample -= se->start_runtime;
2400 update_avg(&se->avg_wakeup, sample);
2402 se->last_wakeup = se->sum_exec_runtime;
2406 trace_sched_wakeup(rq, p, success);
2407 check_preempt_curr(rq, p, wake_flags);
2409 p->state = TASK_RUNNING;
2411 if (p->sched_class->task_wake_up)
2412 p->sched_class->task_wake_up(rq, p);
2415 task_rq_unlock(rq, &flags);
2422 * wake_up_process - Wake up a specific process
2423 * @p: The process to be woken up.
2425 * Attempt to wake up the nominated process and move it to the set of runnable
2426 * processes. Returns 1 if the process was woken up, 0 if it was already
2429 * It may be assumed that this function implies a write memory barrier before
2430 * changing the task state if and only if any tasks are woken up.
2432 int wake_up_process(struct task_struct *p)
2434 return try_to_wake_up(p, TASK_ALL, 0);
2436 EXPORT_SYMBOL(wake_up_process);
2438 int wake_up_state(struct task_struct *p, unsigned int state)
2440 return try_to_wake_up(p, state, 0);
2444 * Perform scheduler related setup for a newly forked process p.
2445 * p is forked by current.
2447 * __sched_fork() is basic setup used by init_idle() too:
2449 static void __sched_fork(struct task_struct *p)
2451 p->se.exec_start = 0;
2452 p->se.sum_exec_runtime = 0;
2453 p->se.prev_sum_exec_runtime = 0;
2454 p->se.nr_migrations = 0;
2455 p->se.last_wakeup = 0;
2456 p->se.avg_overlap = 0;
2457 p->se.start_runtime = 0;
2458 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2459 p->se.avg_running = 0;
2461 #ifdef CONFIG_SCHEDSTATS
2462 p->se.wait_start = 0;
2464 p->se.wait_count = 0;
2467 p->se.sleep_start = 0;
2468 p->se.sleep_max = 0;
2469 p->se.sum_sleep_runtime = 0;
2471 p->se.block_start = 0;
2472 p->se.block_max = 0;
2474 p->se.slice_max = 0;
2476 p->se.nr_migrations_cold = 0;
2477 p->se.nr_failed_migrations_affine = 0;
2478 p->se.nr_failed_migrations_running = 0;
2479 p->se.nr_failed_migrations_hot = 0;
2480 p->se.nr_forced_migrations = 0;
2481 p->se.nr_forced2_migrations = 0;
2483 p->se.nr_wakeups = 0;
2484 p->se.nr_wakeups_sync = 0;
2485 p->se.nr_wakeups_migrate = 0;
2486 p->se.nr_wakeups_local = 0;
2487 p->se.nr_wakeups_remote = 0;
2488 p->se.nr_wakeups_affine = 0;
2489 p->se.nr_wakeups_affine_attempts = 0;
2490 p->se.nr_wakeups_passive = 0;
2491 p->se.nr_wakeups_idle = 0;
2495 INIT_LIST_HEAD(&p->rt.run_list);
2497 INIT_LIST_HEAD(&p->se.group_node);
2499 #ifdef CONFIG_PREEMPT_NOTIFIERS
2500 INIT_HLIST_HEAD(&p->preempt_notifiers);
2504 * We mark the process as running here, but have not actually
2505 * inserted it onto the runqueue yet. This guarantees that
2506 * nobody will actually run it, and a signal or other external
2507 * event cannot wake it up and insert it on the runqueue either.
2509 p->state = TASK_RUNNING;
2513 * fork()/clone()-time setup:
2515 void sched_fork(struct task_struct *p, int clone_flags)
2517 int cpu = get_cpu();
2522 * Revert to default priority/policy on fork if requested.
2524 if (unlikely(p->sched_reset_on_fork)) {
2525 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2526 p->policy = SCHED_NORMAL;
2527 p->normal_prio = p->static_prio;
2530 if (PRIO_TO_NICE(p->static_prio) < 0) {
2531 p->static_prio = NICE_TO_PRIO(0);
2532 p->normal_prio = p->static_prio;
2537 * We don't need the reset flag anymore after the fork. It has
2538 * fulfilled its duty:
2540 p->sched_reset_on_fork = 0;
2544 * Make sure we do not leak PI boosting priority to the child.
2546 p->prio = current->normal_prio;
2548 if (!rt_prio(p->prio))
2549 p->sched_class = &fair_sched_class;
2552 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_FORK, 0);
2554 set_task_cpu(p, cpu);
2556 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2557 if (likely(sched_info_on()))
2558 memset(&p->sched_info, 0, sizeof(p->sched_info));
2560 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2563 #ifdef CONFIG_PREEMPT
2564 /* Want to start with kernel preemption disabled. */
2565 task_thread_info(p)->preempt_count = 1;
2567 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2573 * wake_up_new_task - wake up a newly created task for the first time.
2575 * This function will do some initial scheduler statistics housekeeping
2576 * that must be done for every newly created context, then puts the task
2577 * on the runqueue and wakes it.
2579 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2581 unsigned long flags;
2584 rq = task_rq_lock(p, &flags);
2585 BUG_ON(p->state != TASK_RUNNING);
2586 update_rq_clock(rq);
2588 if (!p->sched_class->task_new || !current->se.on_rq) {
2589 activate_task(rq, p, 0);
2592 * Let the scheduling class do new task startup
2593 * management (if any):
2595 p->sched_class->task_new(rq, p);
2598 trace_sched_wakeup_new(rq, p, 1);
2599 check_preempt_curr(rq, p, WF_FORK);
2601 if (p->sched_class->task_wake_up)
2602 p->sched_class->task_wake_up(rq, p);
2604 task_rq_unlock(rq, &flags);
2607 #ifdef CONFIG_PREEMPT_NOTIFIERS
2610 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2611 * @notifier: notifier struct to register
2613 void preempt_notifier_register(struct preempt_notifier *notifier)
2615 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2617 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2620 * preempt_notifier_unregister - no longer interested in preemption notifications
2621 * @notifier: notifier struct to unregister
2623 * This is safe to call from within a preemption notifier.
2625 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2627 hlist_del(¬ifier->link);
2629 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2631 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2633 struct preempt_notifier *notifier;
2634 struct hlist_node *node;
2636 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2637 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2641 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2642 struct task_struct *next)
2644 struct preempt_notifier *notifier;
2645 struct hlist_node *node;
2647 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2648 notifier->ops->sched_out(notifier, next);
2651 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2653 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2658 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2659 struct task_struct *next)
2663 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2666 * prepare_task_switch - prepare to switch tasks
2667 * @rq: the runqueue preparing to switch
2668 * @prev: the current task that is being switched out
2669 * @next: the task we are going to switch to.
2671 * This is called with the rq lock held and interrupts off. It must
2672 * be paired with a subsequent finish_task_switch after the context
2675 * prepare_task_switch sets up locking and calls architecture specific
2679 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2680 struct task_struct *next)
2682 fire_sched_out_preempt_notifiers(prev, next);
2683 prepare_lock_switch(rq, next);
2684 prepare_arch_switch(next);
2688 * finish_task_switch - clean up after a task-switch
2689 * @rq: runqueue associated with task-switch
2690 * @prev: the thread we just switched away from.
2692 * finish_task_switch must be called after the context switch, paired
2693 * with a prepare_task_switch call before the context switch.
2694 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2695 * and do any other architecture-specific cleanup actions.
2697 * Note that we may have delayed dropping an mm in context_switch(). If
2698 * so, we finish that here outside of the runqueue lock. (Doing it
2699 * with the lock held can cause deadlocks; see schedule() for
2702 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2703 __releases(rq->lock)
2705 struct mm_struct *mm = rq->prev_mm;
2711 * A task struct has one reference for the use as "current".
2712 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2713 * schedule one last time. The schedule call will never return, and
2714 * the scheduled task must drop that reference.
2715 * The test for TASK_DEAD must occur while the runqueue locks are
2716 * still held, otherwise prev could be scheduled on another cpu, die
2717 * there before we look at prev->state, and then the reference would
2719 * Manfred Spraul <manfred@colorfullife.com>
2721 prev_state = prev->state;
2722 finish_arch_switch(prev);
2723 perf_event_task_sched_in(current, cpu_of(rq));
2724 finish_lock_switch(rq, prev);
2726 fire_sched_in_preempt_notifiers(current);
2729 if (unlikely(prev_state == TASK_DEAD)) {
2731 * Remove function-return probe instances associated with this
2732 * task and put them back on the free list.
2734 kprobe_flush_task(prev);
2735 put_task_struct(prev);
2741 /* assumes rq->lock is held */
2742 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2744 if (prev->sched_class->pre_schedule)
2745 prev->sched_class->pre_schedule(rq, prev);
2748 /* rq->lock is NOT held, but preemption is disabled */
2749 static inline void post_schedule(struct rq *rq)
2751 if (rq->post_schedule) {
2752 unsigned long flags;
2754 spin_lock_irqsave(&rq->lock, flags);
2755 if (rq->curr->sched_class->post_schedule)
2756 rq->curr->sched_class->post_schedule(rq);
2757 spin_unlock_irqrestore(&rq->lock, flags);
2759 rq->post_schedule = 0;
2765 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2769 static inline void post_schedule(struct rq *rq)
2776 * schedule_tail - first thing a freshly forked thread must call.
2777 * @prev: the thread we just switched away from.
2779 asmlinkage void schedule_tail(struct task_struct *prev)
2780 __releases(rq->lock)
2782 struct rq *rq = this_rq();
2784 finish_task_switch(rq, prev);
2787 * FIXME: do we need to worry about rq being invalidated by the
2792 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2793 /* In this case, finish_task_switch does not reenable preemption */
2796 if (current->set_child_tid)
2797 put_user(task_pid_vnr(current), current->set_child_tid);
2801 * context_switch - switch to the new MM and the new
2802 * thread's register state.
2805 context_switch(struct rq *rq, struct task_struct *prev,
2806 struct task_struct *next)
2808 struct mm_struct *mm, *oldmm;
2810 prepare_task_switch(rq, prev, next);
2811 trace_sched_switch(rq, prev, next);
2813 oldmm = prev->active_mm;
2815 * For paravirt, this is coupled with an exit in switch_to to
2816 * combine the page table reload and the switch backend into
2819 arch_start_context_switch(prev);
2821 if (unlikely(!mm)) {
2822 next->active_mm = oldmm;
2823 atomic_inc(&oldmm->mm_count);
2824 enter_lazy_tlb(oldmm, next);
2826 switch_mm(oldmm, mm, next);
2828 if (unlikely(!prev->mm)) {
2829 prev->active_mm = NULL;
2830 rq->prev_mm = oldmm;
2833 * Since the runqueue lock will be released by the next
2834 * task (which is an invalid locking op but in the case
2835 * of the scheduler it's an obvious special-case), so we
2836 * do an early lockdep release here:
2838 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2839 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2842 /* Here we just switch the register state and the stack. */
2843 switch_to(prev, next, prev);
2847 * this_rq must be evaluated again because prev may have moved
2848 * CPUs since it called schedule(), thus the 'rq' on its stack
2849 * frame will be invalid.
2851 finish_task_switch(this_rq(), prev);
2855 * nr_running, nr_uninterruptible and nr_context_switches:
2857 * externally visible scheduler statistics: current number of runnable
2858 * threads, current number of uninterruptible-sleeping threads, total
2859 * number of context switches performed since bootup.
2861 unsigned long nr_running(void)
2863 unsigned long i, sum = 0;
2865 for_each_online_cpu(i)
2866 sum += cpu_rq(i)->nr_running;
2871 unsigned long nr_uninterruptible(void)
2873 unsigned long i, sum = 0;
2875 for_each_possible_cpu(i)
2876 sum += cpu_rq(i)->nr_uninterruptible;
2879 * Since we read the counters lockless, it might be slightly
2880 * inaccurate. Do not allow it to go below zero though:
2882 if (unlikely((long)sum < 0))
2888 unsigned long long nr_context_switches(void)
2891 unsigned long long sum = 0;
2893 for_each_possible_cpu(i)
2894 sum += cpu_rq(i)->nr_switches;
2899 unsigned long nr_iowait(void)
2901 unsigned long i, sum = 0;
2903 for_each_possible_cpu(i)
2904 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2909 unsigned long nr_iowait_cpu(void)
2911 struct rq *this = this_rq();
2912 return atomic_read(&this->nr_iowait);
2915 unsigned long this_cpu_load(void)
2917 struct rq *this = this_rq();
2918 return this->cpu_load[0];
2922 /* Variables and functions for calc_load */
2923 static atomic_long_t calc_load_tasks;
2924 static unsigned long calc_load_update;
2925 unsigned long avenrun[3];
2926 EXPORT_SYMBOL(avenrun);
2929 * get_avenrun - get the load average array
2930 * @loads: pointer to dest load array
2931 * @offset: offset to add
2932 * @shift: shift count to shift the result left
2934 * These values are estimates at best, so no need for locking.
2936 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2938 loads[0] = (avenrun[0] + offset) << shift;
2939 loads[1] = (avenrun[1] + offset) << shift;
2940 loads[2] = (avenrun[2] + offset) << shift;
2943 static unsigned long
2944 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2947 load += active * (FIXED_1 - exp);
2948 return load >> FSHIFT;
2952 * calc_load - update the avenrun load estimates 10 ticks after the
2953 * CPUs have updated calc_load_tasks.
2955 void calc_global_load(void)
2957 unsigned long upd = calc_load_update + 10;
2960 if (time_before(jiffies, upd))
2963 active = atomic_long_read(&calc_load_tasks);
2964 active = active > 0 ? active * FIXED_1 : 0;
2966 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2967 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2968 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2970 calc_load_update += LOAD_FREQ;
2974 * Either called from update_cpu_load() or from a cpu going idle
2976 static void calc_load_account_active(struct rq *this_rq)
2978 long nr_active, delta;
2980 nr_active = this_rq->nr_running;
2981 nr_active += (long) this_rq->nr_uninterruptible;
2983 if (nr_active != this_rq->calc_load_active) {
2984 delta = nr_active - this_rq->calc_load_active;
2985 this_rq->calc_load_active = nr_active;
2986 atomic_long_add(delta, &calc_load_tasks);
2991 * Externally visible per-cpu scheduler statistics:
2992 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2994 u64 cpu_nr_migrations(int cpu)
2996 return cpu_rq(cpu)->nr_migrations_in;
3000 * Update rq->cpu_load[] statistics. This function is usually called every
3001 * scheduler tick (TICK_NSEC).
3003 static void update_cpu_load(struct rq *this_rq)
3005 unsigned long this_load = this_rq->load.weight;
3008 this_rq->nr_load_updates++;
3010 /* Update our load: */
3011 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3012 unsigned long old_load, new_load;
3014 /* scale is effectively 1 << i now, and >> i divides by scale */
3016 old_load = this_rq->cpu_load[i];
3017 new_load = this_load;
3019 * Round up the averaging division if load is increasing. This
3020 * prevents us from getting stuck on 9 if the load is 10, for
3023 if (new_load > old_load)
3024 new_load += scale-1;
3025 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3028 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3029 this_rq->calc_load_update += LOAD_FREQ;
3030 calc_load_account_active(this_rq);
3037 * double_rq_lock - safely lock two runqueues
3039 * Note this does not disable interrupts like task_rq_lock,
3040 * you need to do so manually before calling.
3042 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3043 __acquires(rq1->lock)
3044 __acquires(rq2->lock)
3046 BUG_ON(!irqs_disabled());
3048 spin_lock(&rq1->lock);
3049 __acquire(rq2->lock); /* Fake it out ;) */
3052 spin_lock(&rq1->lock);
3053 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3055 spin_lock(&rq2->lock);
3056 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3059 update_rq_clock(rq1);
3060 update_rq_clock(rq2);
3064 * double_rq_unlock - safely unlock two runqueues
3066 * Note this does not restore interrupts like task_rq_unlock,
3067 * you need to do so manually after calling.
3069 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3070 __releases(rq1->lock)
3071 __releases(rq2->lock)
3073 spin_unlock(&rq1->lock);
3075 spin_unlock(&rq2->lock);
3077 __release(rq2->lock);
3081 * If dest_cpu is allowed for this process, migrate the task to it.
3082 * This is accomplished by forcing the cpu_allowed mask to only
3083 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3084 * the cpu_allowed mask is restored.
3086 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3088 struct migration_req req;
3089 unsigned long flags;
3092 rq = task_rq_lock(p, &flags);
3093 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3094 || unlikely(!cpu_active(dest_cpu)))
3097 /* force the process onto the specified CPU */
3098 if (migrate_task(p, dest_cpu, &req)) {
3099 /* Need to wait for migration thread (might exit: take ref). */
3100 struct task_struct *mt = rq->migration_thread;
3102 get_task_struct(mt);
3103 task_rq_unlock(rq, &flags);
3104 wake_up_process(mt);
3105 put_task_struct(mt);
3106 wait_for_completion(&req.done);
3111 task_rq_unlock(rq, &flags);
3115 * sched_exec - execve() is a valuable balancing opportunity, because at
3116 * this point the task has the smallest effective memory and cache footprint.
3118 void sched_exec(void)
3120 int new_cpu, this_cpu = get_cpu();
3121 new_cpu = current->sched_class->select_task_rq(current, SD_BALANCE_EXEC, 0);
3123 if (new_cpu != this_cpu)
3124 sched_migrate_task(current, new_cpu);
3128 * pull_task - move a task from a remote runqueue to the local runqueue.
3129 * Both runqueues must be locked.
3131 static void pull_task(struct rq *src_rq, struct task_struct *p,
3132 struct rq *this_rq, int this_cpu)
3134 deactivate_task(src_rq, p, 0);
3135 set_task_cpu(p, this_cpu);
3136 activate_task(this_rq, p, 0);
3138 * Note that idle threads have a prio of MAX_PRIO, for this test
3139 * to be always true for them.
3141 check_preempt_curr(this_rq, p, 0);
3145 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3148 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3149 struct sched_domain *sd, enum cpu_idle_type idle,
3152 int tsk_cache_hot = 0;
3154 * We do not migrate tasks that are:
3155 * 1) running (obviously), or
3156 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3157 * 3) are cache-hot on their current CPU.
3159 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3160 schedstat_inc(p, se.nr_failed_migrations_affine);
3165 if (task_running(rq, p)) {
3166 schedstat_inc(p, se.nr_failed_migrations_running);
3171 * Aggressive migration if:
3172 * 1) task is cache cold, or
3173 * 2) too many balance attempts have failed.
3176 tsk_cache_hot = task_hot(p, rq->clock, sd);
3177 if (!tsk_cache_hot ||
3178 sd->nr_balance_failed > sd->cache_nice_tries) {
3179 #ifdef CONFIG_SCHEDSTATS
3180 if (tsk_cache_hot) {
3181 schedstat_inc(sd, lb_hot_gained[idle]);
3182 schedstat_inc(p, se.nr_forced_migrations);
3188 if (tsk_cache_hot) {
3189 schedstat_inc(p, se.nr_failed_migrations_hot);
3195 static unsigned long
3196 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3197 unsigned long max_load_move, struct sched_domain *sd,
3198 enum cpu_idle_type idle, int *all_pinned,
3199 int *this_best_prio, struct rq_iterator *iterator)
3201 int loops = 0, pulled = 0, pinned = 0;
3202 struct task_struct *p;
3203 long rem_load_move = max_load_move;
3205 if (max_load_move == 0)
3211 * Start the load-balancing iterator:
3213 p = iterator->start(iterator->arg);
3215 if (!p || loops++ > sysctl_sched_nr_migrate)
3218 if ((p->se.load.weight >> 1) > rem_load_move ||
3219 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3220 p = iterator->next(iterator->arg);
3224 pull_task(busiest, p, this_rq, this_cpu);
3226 rem_load_move -= p->se.load.weight;
3228 #ifdef CONFIG_PREEMPT
3230 * NEWIDLE balancing is a source of latency, so preemptible kernels
3231 * will stop after the first task is pulled to minimize the critical
3234 if (idle == CPU_NEWLY_IDLE)
3239 * We only want to steal up to the prescribed amount of weighted load.
3241 if (rem_load_move > 0) {
3242 if (p->prio < *this_best_prio)
3243 *this_best_prio = p->prio;
3244 p = iterator->next(iterator->arg);
3249 * Right now, this is one of only two places pull_task() is called,
3250 * so we can safely collect pull_task() stats here rather than
3251 * inside pull_task().
3253 schedstat_add(sd, lb_gained[idle], pulled);
3256 *all_pinned = pinned;
3258 return max_load_move - rem_load_move;
3262 * move_tasks tries to move up to max_load_move weighted load from busiest to
3263 * this_rq, as part of a balancing operation within domain "sd".
3264 * Returns 1 if successful and 0 otherwise.
3266 * Called with both runqueues locked.
3268 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3269 unsigned long max_load_move,
3270 struct sched_domain *sd, enum cpu_idle_type idle,
3273 const struct sched_class *class = sched_class_highest;
3274 unsigned long total_load_moved = 0;
3275 int this_best_prio = this_rq->curr->prio;
3279 class->load_balance(this_rq, this_cpu, busiest,
3280 max_load_move - total_load_moved,
3281 sd, idle, all_pinned, &this_best_prio);
3282 class = class->next;
3284 #ifdef CONFIG_PREEMPT
3286 * NEWIDLE balancing is a source of latency, so preemptible
3287 * kernels will stop after the first task is pulled to minimize
3288 * the critical section.
3290 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3293 } while (class && max_load_move > total_load_moved);
3295 return total_load_moved > 0;
3299 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3300 struct sched_domain *sd, enum cpu_idle_type idle,
3301 struct rq_iterator *iterator)
3303 struct task_struct *p = iterator->start(iterator->arg);
3307 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3308 pull_task(busiest, p, this_rq, this_cpu);
3310 * Right now, this is only the second place pull_task()
3311 * is called, so we can safely collect pull_task()
3312 * stats here rather than inside pull_task().
3314 schedstat_inc(sd, lb_gained[idle]);
3318 p = iterator->next(iterator->arg);
3325 * move_one_task tries to move exactly one task from busiest to this_rq, as
3326 * part of active balancing operations within "domain".
3327 * Returns 1 if successful and 0 otherwise.
3329 * Called with both runqueues locked.
3331 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3332 struct sched_domain *sd, enum cpu_idle_type idle)
3334 const struct sched_class *class;
3336 for_each_class(class) {
3337 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3343 /********** Helpers for find_busiest_group ************************/
3345 * sd_lb_stats - Structure to store the statistics of a sched_domain
3346 * during load balancing.
3348 struct sd_lb_stats {
3349 struct sched_group *busiest; /* Busiest group in this sd */
3350 struct sched_group *this; /* Local group in this sd */
3351 unsigned long total_load; /* Total load of all groups in sd */
3352 unsigned long total_pwr; /* Total power of all groups in sd */
3353 unsigned long avg_load; /* Average load across all groups in sd */
3355 /** Statistics of this group */
3356 unsigned long this_load;
3357 unsigned long this_load_per_task;
3358 unsigned long this_nr_running;
3360 /* Statistics of the busiest group */
3361 unsigned long max_load;
3362 unsigned long busiest_load_per_task;
3363 unsigned long busiest_nr_running;
3365 int group_imb; /* Is there imbalance in this sd */
3366 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3367 int power_savings_balance; /* Is powersave balance needed for this sd */
3368 struct sched_group *group_min; /* Least loaded group in sd */
3369 struct sched_group *group_leader; /* Group which relieves group_min */
3370 unsigned long min_load_per_task; /* load_per_task in group_min */
3371 unsigned long leader_nr_running; /* Nr running of group_leader */
3372 unsigned long min_nr_running; /* Nr running of group_min */
3377 * sg_lb_stats - stats of a sched_group required for load_balancing
3379 struct sg_lb_stats {
3380 unsigned long avg_load; /*Avg load across the CPUs of the group */
3381 unsigned long group_load; /* Total load over the CPUs of the group */
3382 unsigned long sum_nr_running; /* Nr tasks running in the group */
3383 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3384 unsigned long group_capacity;
3385 int group_imb; /* Is there an imbalance in the group ? */
3389 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3390 * @group: The group whose first cpu is to be returned.
3392 static inline unsigned int group_first_cpu(struct sched_group *group)
3394 return cpumask_first(sched_group_cpus(group));
3398 * get_sd_load_idx - Obtain the load index for a given sched domain.
3399 * @sd: The sched_domain whose load_idx is to be obtained.
3400 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3402 static inline int get_sd_load_idx(struct sched_domain *sd,
3403 enum cpu_idle_type idle)
3409 load_idx = sd->busy_idx;
3412 case CPU_NEWLY_IDLE:
3413 load_idx = sd->newidle_idx;
3416 load_idx = sd->idle_idx;
3424 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3426 * init_sd_power_savings_stats - Initialize power savings statistics for
3427 * the given sched_domain, during load balancing.
3429 * @sd: Sched domain whose power-savings statistics are to be initialized.
3430 * @sds: Variable containing the statistics for sd.
3431 * @idle: Idle status of the CPU at which we're performing load-balancing.
3433 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3434 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3437 * Busy processors will not participate in power savings
3440 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3441 sds->power_savings_balance = 0;
3443 sds->power_savings_balance = 1;
3444 sds->min_nr_running = ULONG_MAX;
3445 sds->leader_nr_running = 0;
3450 * update_sd_power_savings_stats - Update the power saving stats for a
3451 * sched_domain while performing load balancing.
3453 * @group: sched_group belonging to the sched_domain under consideration.
3454 * @sds: Variable containing the statistics of the sched_domain
3455 * @local_group: Does group contain the CPU for which we're performing
3457 * @sgs: Variable containing the statistics of the group.
3459 static inline void update_sd_power_savings_stats(struct sched_group *group,
3460 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3463 if (!sds->power_savings_balance)
3467 * If the local group is idle or completely loaded
3468 * no need to do power savings balance at this domain
3470 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3471 !sds->this_nr_running))
3472 sds->power_savings_balance = 0;
3475 * If a group is already running at full capacity or idle,
3476 * don't include that group in power savings calculations
3478 if (!sds->power_savings_balance ||
3479 sgs->sum_nr_running >= sgs->group_capacity ||
3480 !sgs->sum_nr_running)
3484 * Calculate the group which has the least non-idle load.
3485 * This is the group from where we need to pick up the load
3488 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3489 (sgs->sum_nr_running == sds->min_nr_running &&
3490 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3491 sds->group_min = group;
3492 sds->min_nr_running = sgs->sum_nr_running;
3493 sds->min_load_per_task = sgs->sum_weighted_load /
3494 sgs->sum_nr_running;
3498 * Calculate the group which is almost near its
3499 * capacity but still has some space to pick up some load
3500 * from other group and save more power
3502 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3505 if (sgs->sum_nr_running > sds->leader_nr_running ||
3506 (sgs->sum_nr_running == sds->leader_nr_running &&
3507 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3508 sds->group_leader = group;
3509 sds->leader_nr_running = sgs->sum_nr_running;
3514 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3515 * @sds: Variable containing the statistics of the sched_domain
3516 * under consideration.
3517 * @this_cpu: Cpu at which we're currently performing load-balancing.
3518 * @imbalance: Variable to store the imbalance.
3521 * Check if we have potential to perform some power-savings balance.
3522 * If yes, set the busiest group to be the least loaded group in the
3523 * sched_domain, so that it's CPUs can be put to idle.
3525 * Returns 1 if there is potential to perform power-savings balance.
3528 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3529 int this_cpu, unsigned long *imbalance)
3531 if (!sds->power_savings_balance)
3534 if (sds->this != sds->group_leader ||
3535 sds->group_leader == sds->group_min)
3538 *imbalance = sds->min_load_per_task;
3539 sds->busiest = sds->group_min;
3544 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3545 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3546 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3551 static inline void update_sd_power_savings_stats(struct sched_group *group,
3552 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3557 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3558 int this_cpu, unsigned long *imbalance)
3562 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3565 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3567 return SCHED_LOAD_SCALE;
3570 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3572 return default_scale_freq_power(sd, cpu);
3575 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3577 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3578 unsigned long smt_gain = sd->smt_gain;
3585 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3587 return default_scale_smt_power(sd, cpu);
3590 unsigned long scale_rt_power(int cpu)
3592 struct rq *rq = cpu_rq(cpu);
3593 u64 total, available;
3595 sched_avg_update(rq);
3597 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3598 available = total - rq->rt_avg;
3600 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3601 total = SCHED_LOAD_SCALE;
3603 total >>= SCHED_LOAD_SHIFT;
3605 return div_u64(available, total);
3608 static void update_cpu_power(struct sched_domain *sd, int cpu)
3610 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3611 unsigned long power = SCHED_LOAD_SCALE;
3612 struct sched_group *sdg = sd->groups;
3614 if (sched_feat(ARCH_POWER))
3615 power *= arch_scale_freq_power(sd, cpu);
3617 power *= default_scale_freq_power(sd, cpu);
3619 power >>= SCHED_LOAD_SHIFT;
3621 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3622 if (sched_feat(ARCH_POWER))
3623 power *= arch_scale_smt_power(sd, cpu);
3625 power *= default_scale_smt_power(sd, cpu);
3627 power >>= SCHED_LOAD_SHIFT;
3630 power *= scale_rt_power(cpu);
3631 power >>= SCHED_LOAD_SHIFT;
3636 sdg->cpu_power = power;
3639 static void update_group_power(struct sched_domain *sd, int cpu)
3641 struct sched_domain *child = sd->child;
3642 struct sched_group *group, *sdg = sd->groups;
3643 unsigned long power;
3646 update_cpu_power(sd, cpu);
3652 group = child->groups;
3654 power += group->cpu_power;
3655 group = group->next;
3656 } while (group != child->groups);
3658 sdg->cpu_power = power;
3662 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3663 * @group: sched_group whose statistics are to be updated.
3664 * @this_cpu: Cpu for which load balance is currently performed.
3665 * @idle: Idle status of this_cpu
3666 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3667 * @sd_idle: Idle status of the sched_domain containing group.
3668 * @local_group: Does group contain this_cpu.
3669 * @cpus: Set of cpus considered for load balancing.
3670 * @balance: Should we balance.
3671 * @sgs: variable to hold the statistics for this group.
3673 static inline void update_sg_lb_stats(struct sched_domain *sd,
3674 struct sched_group *group, int this_cpu,
3675 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3676 int local_group, const struct cpumask *cpus,
3677 int *balance, struct sg_lb_stats *sgs)
3679 unsigned long load, max_cpu_load, min_cpu_load;
3681 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3682 unsigned long sum_avg_load_per_task;
3683 unsigned long avg_load_per_task;
3686 balance_cpu = group_first_cpu(group);
3687 if (balance_cpu == this_cpu)
3688 update_group_power(sd, this_cpu);
3691 /* Tally up the load of all CPUs in the group */
3692 sum_avg_load_per_task = avg_load_per_task = 0;
3694 min_cpu_load = ~0UL;
3696 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3697 struct rq *rq = cpu_rq(i);
3699 if (*sd_idle && rq->nr_running)
3702 /* Bias balancing toward cpus of our domain */
3704 if (idle_cpu(i) && !first_idle_cpu) {
3709 load = target_load(i, load_idx);
3711 load = source_load(i, load_idx);
3712 if (load > max_cpu_load)
3713 max_cpu_load = load;
3714 if (min_cpu_load > load)
3715 min_cpu_load = load;
3718 sgs->group_load += load;
3719 sgs->sum_nr_running += rq->nr_running;
3720 sgs->sum_weighted_load += weighted_cpuload(i);
3722 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3726 * First idle cpu or the first cpu(busiest) in this sched group
3727 * is eligible for doing load balancing at this and above
3728 * domains. In the newly idle case, we will allow all the cpu's
3729 * to do the newly idle load balance.
3731 if (idle != CPU_NEWLY_IDLE && local_group &&
3732 balance_cpu != this_cpu && balance) {
3737 /* Adjust by relative CPU power of the group */
3738 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3742 * Consider the group unbalanced when the imbalance is larger
3743 * than the average weight of two tasks.
3745 * APZ: with cgroup the avg task weight can vary wildly and
3746 * might not be a suitable number - should we keep a
3747 * normalized nr_running number somewhere that negates
3750 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3753 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3756 sgs->group_capacity =
3757 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3761 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3762 * @sd: sched_domain whose statistics are to be updated.
3763 * @this_cpu: Cpu for which load balance is currently performed.
3764 * @idle: Idle status of this_cpu
3765 * @sd_idle: Idle status of the sched_domain containing group.
3766 * @cpus: Set of cpus considered for load balancing.
3767 * @balance: Should we balance.
3768 * @sds: variable to hold the statistics for this sched_domain.
3770 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3771 enum cpu_idle_type idle, int *sd_idle,
3772 const struct cpumask *cpus, int *balance,
3773 struct sd_lb_stats *sds)
3775 struct sched_domain *child = sd->child;
3776 struct sched_group *group = sd->groups;
3777 struct sg_lb_stats sgs;
3778 int load_idx, prefer_sibling = 0;
3780 if (child && child->flags & SD_PREFER_SIBLING)
3783 init_sd_power_savings_stats(sd, sds, idle);
3784 load_idx = get_sd_load_idx(sd, idle);
3789 local_group = cpumask_test_cpu(this_cpu,
3790 sched_group_cpus(group));
3791 memset(&sgs, 0, sizeof(sgs));
3792 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3793 local_group, cpus, balance, &sgs);
3795 if (local_group && balance && !(*balance))
3798 sds->total_load += sgs.group_load;
3799 sds->total_pwr += group->cpu_power;
3802 * In case the child domain prefers tasks go to siblings
3803 * first, lower the group capacity to one so that we'll try
3804 * and move all the excess tasks away.
3807 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3810 sds->this_load = sgs.avg_load;
3812 sds->this_nr_running = sgs.sum_nr_running;
3813 sds->this_load_per_task = sgs.sum_weighted_load;
3814 } else if (sgs.avg_load > sds->max_load &&
3815 (sgs.sum_nr_running > sgs.group_capacity ||
3817 sds->max_load = sgs.avg_load;
3818 sds->busiest = group;
3819 sds->busiest_nr_running = sgs.sum_nr_running;
3820 sds->busiest_load_per_task = sgs.sum_weighted_load;
3821 sds->group_imb = sgs.group_imb;
3824 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3825 group = group->next;
3826 } while (group != sd->groups);
3830 * fix_small_imbalance - Calculate the minor imbalance that exists
3831 * amongst the groups of a sched_domain, during
3833 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3834 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3835 * @imbalance: Variable to store the imbalance.
3837 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3838 int this_cpu, unsigned long *imbalance)
3840 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3841 unsigned int imbn = 2;
3843 if (sds->this_nr_running) {
3844 sds->this_load_per_task /= sds->this_nr_running;
3845 if (sds->busiest_load_per_task >
3846 sds->this_load_per_task)
3849 sds->this_load_per_task =
3850 cpu_avg_load_per_task(this_cpu);
3852 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3853 sds->busiest_load_per_task * imbn) {
3854 *imbalance = sds->busiest_load_per_task;
3859 * OK, we don't have enough imbalance to justify moving tasks,
3860 * however we may be able to increase total CPU power used by
3864 pwr_now += sds->busiest->cpu_power *
3865 min(sds->busiest_load_per_task, sds->max_load);
3866 pwr_now += sds->this->cpu_power *
3867 min(sds->this_load_per_task, sds->this_load);
3868 pwr_now /= SCHED_LOAD_SCALE;
3870 /* Amount of load we'd subtract */
3871 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3872 sds->busiest->cpu_power;
3873 if (sds->max_load > tmp)
3874 pwr_move += sds->busiest->cpu_power *
3875 min(sds->busiest_load_per_task, sds->max_load - tmp);
3877 /* Amount of load we'd add */
3878 if (sds->max_load * sds->busiest->cpu_power <
3879 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3880 tmp = (sds->max_load * sds->busiest->cpu_power) /
3881 sds->this->cpu_power;
3883 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3884 sds->this->cpu_power;
3885 pwr_move += sds->this->cpu_power *
3886 min(sds->this_load_per_task, sds->this_load + tmp);
3887 pwr_move /= SCHED_LOAD_SCALE;
3889 /* Move if we gain throughput */
3890 if (pwr_move > pwr_now)
3891 *imbalance = sds->busiest_load_per_task;
3895 * calculate_imbalance - Calculate the amount of imbalance present within the
3896 * groups of a given sched_domain during load balance.
3897 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3898 * @this_cpu: Cpu for which currently load balance is being performed.
3899 * @imbalance: The variable to store the imbalance.
3901 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3902 unsigned long *imbalance)
3904 unsigned long max_pull;
3906 * In the presence of smp nice balancing, certain scenarios can have
3907 * max load less than avg load(as we skip the groups at or below
3908 * its cpu_power, while calculating max_load..)
3910 if (sds->max_load < sds->avg_load) {
3912 return fix_small_imbalance(sds, this_cpu, imbalance);
3915 /* Don't want to pull so many tasks that a group would go idle */
3916 max_pull = min(sds->max_load - sds->avg_load,
3917 sds->max_load - sds->busiest_load_per_task);
3919 /* How much load to actually move to equalise the imbalance */
3920 *imbalance = min(max_pull * sds->busiest->cpu_power,
3921 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3925 * if *imbalance is less than the average load per runnable task
3926 * there is no gaurantee that any tasks will be moved so we'll have
3927 * a think about bumping its value to force at least one task to be
3930 if (*imbalance < sds->busiest_load_per_task)
3931 return fix_small_imbalance(sds, this_cpu, imbalance);
3934 /******* find_busiest_group() helpers end here *********************/
3937 * find_busiest_group - Returns the busiest group within the sched_domain
3938 * if there is an imbalance. If there isn't an imbalance, and
3939 * the user has opted for power-savings, it returns a group whose
3940 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3941 * such a group exists.
3943 * Also calculates the amount of weighted load which should be moved
3944 * to restore balance.
3946 * @sd: The sched_domain whose busiest group is to be returned.
3947 * @this_cpu: The cpu for which load balancing is currently being performed.
3948 * @imbalance: Variable which stores amount of weighted load which should
3949 * be moved to restore balance/put a group to idle.
3950 * @idle: The idle status of this_cpu.
3951 * @sd_idle: The idleness of sd
3952 * @cpus: The set of CPUs under consideration for load-balancing.
3953 * @balance: Pointer to a variable indicating if this_cpu
3954 * is the appropriate cpu to perform load balancing at this_level.
3956 * Returns: - the busiest group if imbalance exists.
3957 * - If no imbalance and user has opted for power-savings balance,
3958 * return the least loaded group whose CPUs can be
3959 * put to idle by rebalancing its tasks onto our group.
3961 static struct sched_group *
3962 find_busiest_group(struct sched_domain *sd, int this_cpu,
3963 unsigned long *imbalance, enum cpu_idle_type idle,
3964 int *sd_idle, const struct cpumask *cpus, int *balance)
3966 struct sd_lb_stats sds;
3968 memset(&sds, 0, sizeof(sds));
3971 * Compute the various statistics relavent for load balancing at
3974 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3977 /* Cases where imbalance does not exist from POV of this_cpu */
3978 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3980 * 2) There is no busy sibling group to pull from.
3981 * 3) This group is the busiest group.
3982 * 4) This group is more busy than the avg busieness at this
3984 * 5) The imbalance is within the specified limit.
3985 * 6) Any rebalance would lead to ping-pong
3987 if (balance && !(*balance))
3990 if (!sds.busiest || sds.busiest_nr_running == 0)
3993 if (sds.this_load >= sds.max_load)
3996 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3998 if (sds.this_load >= sds.avg_load)
4001 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4004 sds.busiest_load_per_task /= sds.busiest_nr_running;
4006 sds.busiest_load_per_task =
4007 min(sds.busiest_load_per_task, sds.avg_load);
4010 * We're trying to get all the cpus to the average_load, so we don't
4011 * want to push ourselves above the average load, nor do we wish to
4012 * reduce the max loaded cpu below the average load, as either of these
4013 * actions would just result in more rebalancing later, and ping-pong
4014 * tasks around. Thus we look for the minimum possible imbalance.
4015 * Negative imbalances (*we* are more loaded than anyone else) will
4016 * be counted as no imbalance for these purposes -- we can't fix that
4017 * by pulling tasks to us. Be careful of negative numbers as they'll
4018 * appear as very large values with unsigned longs.
4020 if (sds.max_load <= sds.busiest_load_per_task)
4023 /* Looks like there is an imbalance. Compute it */
4024 calculate_imbalance(&sds, this_cpu, imbalance);
4029 * There is no obvious imbalance. But check if we can do some balancing
4032 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4040 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4043 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4044 unsigned long imbalance, const struct cpumask *cpus)
4046 struct rq *busiest = NULL, *rq;
4047 unsigned long max_load = 0;
4050 for_each_cpu(i, sched_group_cpus(group)) {
4051 unsigned long power = power_of(i);
4052 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4055 if (!cpumask_test_cpu(i, cpus))
4059 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4062 if (capacity && rq->nr_running == 1 && wl > imbalance)
4065 if (wl > max_load) {
4075 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4076 * so long as it is large enough.
4078 #define MAX_PINNED_INTERVAL 512
4080 /* Working cpumask for load_balance and load_balance_newidle. */
4081 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4084 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4085 * tasks if there is an imbalance.
4087 static int load_balance(int this_cpu, struct rq *this_rq,
4088 struct sched_domain *sd, enum cpu_idle_type idle,
4091 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4092 struct sched_group *group;
4093 unsigned long imbalance;
4095 unsigned long flags;
4096 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4098 cpumask_setall(cpus);
4101 * When power savings policy is enabled for the parent domain, idle
4102 * sibling can pick up load irrespective of busy siblings. In this case,
4103 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4104 * portraying it as CPU_NOT_IDLE.
4106 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4107 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4110 schedstat_inc(sd, lb_count[idle]);
4114 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4121 schedstat_inc(sd, lb_nobusyg[idle]);
4125 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4127 schedstat_inc(sd, lb_nobusyq[idle]);
4131 BUG_ON(busiest == this_rq);
4133 schedstat_add(sd, lb_imbalance[idle], imbalance);
4136 if (busiest->nr_running > 1) {
4138 * Attempt to move tasks. If find_busiest_group has found
4139 * an imbalance but busiest->nr_running <= 1, the group is
4140 * still unbalanced. ld_moved simply stays zero, so it is
4141 * correctly treated as an imbalance.
4143 local_irq_save(flags);
4144 double_rq_lock(this_rq, busiest);
4145 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4146 imbalance, sd, idle, &all_pinned);
4147 double_rq_unlock(this_rq, busiest);
4148 local_irq_restore(flags);
4151 * some other cpu did the load balance for us.
4153 if (ld_moved && this_cpu != smp_processor_id())
4154 resched_cpu(this_cpu);
4156 /* All tasks on this runqueue were pinned by CPU affinity */
4157 if (unlikely(all_pinned)) {
4158 cpumask_clear_cpu(cpu_of(busiest), cpus);
4159 if (!cpumask_empty(cpus))
4166 schedstat_inc(sd, lb_failed[idle]);
4167 sd->nr_balance_failed++;
4169 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4171 spin_lock_irqsave(&busiest->lock, flags);
4173 /* don't kick the migration_thread, if the curr
4174 * task on busiest cpu can't be moved to this_cpu
4176 if (!cpumask_test_cpu(this_cpu,
4177 &busiest->curr->cpus_allowed)) {
4178 spin_unlock_irqrestore(&busiest->lock, flags);
4180 goto out_one_pinned;
4183 if (!busiest->active_balance) {
4184 busiest->active_balance = 1;
4185 busiest->push_cpu = this_cpu;
4188 spin_unlock_irqrestore(&busiest->lock, flags);
4190 wake_up_process(busiest->migration_thread);
4193 * We've kicked active balancing, reset the failure
4196 sd->nr_balance_failed = sd->cache_nice_tries+1;
4199 sd->nr_balance_failed = 0;
4201 if (likely(!active_balance)) {
4202 /* We were unbalanced, so reset the balancing interval */
4203 sd->balance_interval = sd->min_interval;
4206 * If we've begun active balancing, start to back off. This
4207 * case may not be covered by the all_pinned logic if there
4208 * is only 1 task on the busy runqueue (because we don't call
4211 if (sd->balance_interval < sd->max_interval)
4212 sd->balance_interval *= 2;
4215 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4216 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4222 schedstat_inc(sd, lb_balanced[idle]);
4224 sd->nr_balance_failed = 0;
4227 /* tune up the balancing interval */
4228 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4229 (sd->balance_interval < sd->max_interval))
4230 sd->balance_interval *= 2;
4232 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4233 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4244 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4245 * tasks if there is an imbalance.
4247 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4248 * this_rq is locked.
4251 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4253 struct sched_group *group;
4254 struct rq *busiest = NULL;
4255 unsigned long imbalance;
4259 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4261 cpumask_setall(cpus);
4264 * When power savings policy is enabled for the parent domain, idle
4265 * sibling can pick up load irrespective of busy siblings. In this case,
4266 * let the state of idle sibling percolate up as IDLE, instead of
4267 * portraying it as CPU_NOT_IDLE.
4269 if (sd->flags & SD_SHARE_CPUPOWER &&
4270 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4273 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4275 update_shares_locked(this_rq, sd);
4276 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4277 &sd_idle, cpus, NULL);
4279 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4283 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4285 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4289 BUG_ON(busiest == this_rq);
4291 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4294 if (busiest->nr_running > 1) {
4295 /* Attempt to move tasks */
4296 double_lock_balance(this_rq, busiest);
4297 /* this_rq->clock is already updated */
4298 update_rq_clock(busiest);
4299 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4300 imbalance, sd, CPU_NEWLY_IDLE,
4302 double_unlock_balance(this_rq, busiest);
4304 if (unlikely(all_pinned)) {
4305 cpumask_clear_cpu(cpu_of(busiest), cpus);
4306 if (!cpumask_empty(cpus))
4312 int active_balance = 0;
4314 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4315 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4316 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4319 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4322 if (sd->nr_balance_failed++ < 2)
4326 * The only task running in a non-idle cpu can be moved to this
4327 * cpu in an attempt to completely freeup the other CPU
4328 * package. The same method used to move task in load_balance()
4329 * have been extended for load_balance_newidle() to speedup
4330 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4332 * The package power saving logic comes from
4333 * find_busiest_group(). If there are no imbalance, then
4334 * f_b_g() will return NULL. However when sched_mc={1,2} then
4335 * f_b_g() will select a group from which a running task may be
4336 * pulled to this cpu in order to make the other package idle.
4337 * If there is no opportunity to make a package idle and if
4338 * there are no imbalance, then f_b_g() will return NULL and no
4339 * action will be taken in load_balance_newidle().
4341 * Under normal task pull operation due to imbalance, there
4342 * will be more than one task in the source run queue and
4343 * move_tasks() will succeed. ld_moved will be true and this
4344 * active balance code will not be triggered.
4347 /* Lock busiest in correct order while this_rq is held */
4348 double_lock_balance(this_rq, busiest);
4351 * don't kick the migration_thread, if the curr
4352 * task on busiest cpu can't be moved to this_cpu
4354 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4355 double_unlock_balance(this_rq, busiest);
4360 if (!busiest->active_balance) {
4361 busiest->active_balance = 1;
4362 busiest->push_cpu = this_cpu;
4366 double_unlock_balance(this_rq, busiest);
4368 * Should not call ttwu while holding a rq->lock
4370 spin_unlock(&this_rq->lock);
4372 wake_up_process(busiest->migration_thread);
4373 spin_lock(&this_rq->lock);
4376 sd->nr_balance_failed = 0;
4378 update_shares_locked(this_rq, sd);
4382 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4383 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4384 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4386 sd->nr_balance_failed = 0;
4392 * idle_balance is called by schedule() if this_cpu is about to become
4393 * idle. Attempts to pull tasks from other CPUs.
4395 static void idle_balance(int this_cpu, struct rq *this_rq)
4397 struct sched_domain *sd;
4398 int pulled_task = 0;
4399 unsigned long next_balance = jiffies + HZ;
4401 for_each_domain(this_cpu, sd) {
4402 unsigned long interval;
4404 if (!(sd->flags & SD_LOAD_BALANCE))
4407 if (sd->flags & SD_BALANCE_NEWIDLE)
4408 /* If we've pulled tasks over stop searching: */
4409 pulled_task = load_balance_newidle(this_cpu, this_rq,
4412 interval = msecs_to_jiffies(sd->balance_interval);
4413 if (time_after(next_balance, sd->last_balance + interval))
4414 next_balance = sd->last_balance + interval;
4418 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4420 * We are going idle. next_balance may be set based on
4421 * a busy processor. So reset next_balance.
4423 this_rq->next_balance = next_balance;
4428 * active_load_balance is run by migration threads. It pushes running tasks
4429 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4430 * running on each physical CPU where possible, and avoids physical /
4431 * logical imbalances.
4433 * Called with busiest_rq locked.
4435 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4437 int target_cpu = busiest_rq->push_cpu;
4438 struct sched_domain *sd;
4439 struct rq *target_rq;
4441 /* Is there any task to move? */
4442 if (busiest_rq->nr_running <= 1)
4445 target_rq = cpu_rq(target_cpu);
4448 * This condition is "impossible", if it occurs
4449 * we need to fix it. Originally reported by
4450 * Bjorn Helgaas on a 128-cpu setup.
4452 BUG_ON(busiest_rq == target_rq);
4454 /* move a task from busiest_rq to target_rq */
4455 double_lock_balance(busiest_rq, target_rq);
4456 update_rq_clock(busiest_rq);
4457 update_rq_clock(target_rq);
4459 /* Search for an sd spanning us and the target CPU. */
4460 for_each_domain(target_cpu, sd) {
4461 if ((sd->flags & SD_LOAD_BALANCE) &&
4462 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4467 schedstat_inc(sd, alb_count);
4469 if (move_one_task(target_rq, target_cpu, busiest_rq,
4471 schedstat_inc(sd, alb_pushed);
4473 schedstat_inc(sd, alb_failed);
4475 double_unlock_balance(busiest_rq, target_rq);
4480 atomic_t load_balancer;
4481 cpumask_var_t cpu_mask;
4482 cpumask_var_t ilb_grp_nohz_mask;
4483 } nohz ____cacheline_aligned = {
4484 .load_balancer = ATOMIC_INIT(-1),
4487 int get_nohz_load_balancer(void)
4489 return atomic_read(&nohz.load_balancer);
4492 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4494 * lowest_flag_domain - Return lowest sched_domain containing flag.
4495 * @cpu: The cpu whose lowest level of sched domain is to
4497 * @flag: The flag to check for the lowest sched_domain
4498 * for the given cpu.
4500 * Returns the lowest sched_domain of a cpu which contains the given flag.
4502 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4504 struct sched_domain *sd;
4506 for_each_domain(cpu, sd)
4507 if (sd && (sd->flags & flag))
4514 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4515 * @cpu: The cpu whose domains we're iterating over.
4516 * @sd: variable holding the value of the power_savings_sd
4518 * @flag: The flag to filter the sched_domains to be iterated.
4520 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4521 * set, starting from the lowest sched_domain to the highest.
4523 #define for_each_flag_domain(cpu, sd, flag) \
4524 for (sd = lowest_flag_domain(cpu, flag); \
4525 (sd && (sd->flags & flag)); sd = sd->parent)
4528 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4529 * @ilb_group: group to be checked for semi-idleness
4531 * Returns: 1 if the group is semi-idle. 0 otherwise.
4533 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4534 * and atleast one non-idle CPU. This helper function checks if the given
4535 * sched_group is semi-idle or not.
4537 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4539 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4540 sched_group_cpus(ilb_group));
4543 * A sched_group is semi-idle when it has atleast one busy cpu
4544 * and atleast one idle cpu.
4546 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4549 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4555 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4556 * @cpu: The cpu which is nominating a new idle_load_balancer.
4558 * Returns: Returns the id of the idle load balancer if it exists,
4559 * Else, returns >= nr_cpu_ids.
4561 * This algorithm picks the idle load balancer such that it belongs to a
4562 * semi-idle powersavings sched_domain. The idea is to try and avoid
4563 * completely idle packages/cores just for the purpose of idle load balancing
4564 * when there are other idle cpu's which are better suited for that job.
4566 static int find_new_ilb(int cpu)
4568 struct sched_domain *sd;
4569 struct sched_group *ilb_group;
4572 * Have idle load balancer selection from semi-idle packages only
4573 * when power-aware load balancing is enabled
4575 if (!(sched_smt_power_savings || sched_mc_power_savings))
4579 * Optimize for the case when we have no idle CPUs or only one
4580 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4582 if (cpumask_weight(nohz.cpu_mask) < 2)
4585 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4586 ilb_group = sd->groups;
4589 if (is_semi_idle_group(ilb_group))
4590 return cpumask_first(nohz.ilb_grp_nohz_mask);
4592 ilb_group = ilb_group->next;
4594 } while (ilb_group != sd->groups);
4598 return cpumask_first(nohz.cpu_mask);
4600 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4601 static inline int find_new_ilb(int call_cpu)
4603 return cpumask_first(nohz.cpu_mask);
4608 * This routine will try to nominate the ilb (idle load balancing)
4609 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4610 * load balancing on behalf of all those cpus. If all the cpus in the system
4611 * go into this tickless mode, then there will be no ilb owner (as there is
4612 * no need for one) and all the cpus will sleep till the next wakeup event
4615 * For the ilb owner, tick is not stopped. And this tick will be used
4616 * for idle load balancing. ilb owner will still be part of
4619 * While stopping the tick, this cpu will become the ilb owner if there
4620 * is no other owner. And will be the owner till that cpu becomes busy
4621 * or if all cpus in the system stop their ticks at which point
4622 * there is no need for ilb owner.
4624 * When the ilb owner becomes busy, it nominates another owner, during the
4625 * next busy scheduler_tick()
4627 int select_nohz_load_balancer(int stop_tick)
4629 int cpu = smp_processor_id();
4632 cpu_rq(cpu)->in_nohz_recently = 1;
4634 if (!cpu_active(cpu)) {
4635 if (atomic_read(&nohz.load_balancer) != cpu)
4639 * If we are going offline and still the leader,
4642 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4648 cpumask_set_cpu(cpu, nohz.cpu_mask);
4650 /* time for ilb owner also to sleep */
4651 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4652 if (atomic_read(&nohz.load_balancer) == cpu)
4653 atomic_set(&nohz.load_balancer, -1);
4657 if (atomic_read(&nohz.load_balancer) == -1) {
4658 /* make me the ilb owner */
4659 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4661 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4664 if (!(sched_smt_power_savings ||
4665 sched_mc_power_savings))
4668 * Check to see if there is a more power-efficient
4671 new_ilb = find_new_ilb(cpu);
4672 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4673 atomic_set(&nohz.load_balancer, -1);
4674 resched_cpu(new_ilb);
4680 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4683 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4685 if (atomic_read(&nohz.load_balancer) == cpu)
4686 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4693 static DEFINE_SPINLOCK(balancing);
4696 * It checks each scheduling domain to see if it is due to be balanced,
4697 * and initiates a balancing operation if so.
4699 * Balancing parameters are set up in arch_init_sched_domains.
4701 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4704 struct rq *rq = cpu_rq(cpu);
4705 unsigned long interval;
4706 struct sched_domain *sd;
4707 /* Earliest time when we have to do rebalance again */
4708 unsigned long next_balance = jiffies + 60*HZ;
4709 int update_next_balance = 0;
4712 for_each_domain(cpu, sd) {
4713 if (!(sd->flags & SD_LOAD_BALANCE))
4716 interval = sd->balance_interval;
4717 if (idle != CPU_IDLE)
4718 interval *= sd->busy_factor;
4720 /* scale ms to jiffies */
4721 interval = msecs_to_jiffies(interval);
4722 if (unlikely(!interval))
4724 if (interval > HZ*NR_CPUS/10)
4725 interval = HZ*NR_CPUS/10;
4727 need_serialize = sd->flags & SD_SERIALIZE;
4729 if (need_serialize) {
4730 if (!spin_trylock(&balancing))
4734 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4735 if (load_balance(cpu, rq, sd, idle, &balance)) {
4737 * We've pulled tasks over so either we're no
4738 * longer idle, or one of our SMT siblings is
4741 idle = CPU_NOT_IDLE;
4743 sd->last_balance = jiffies;
4746 spin_unlock(&balancing);
4748 if (time_after(next_balance, sd->last_balance + interval)) {
4749 next_balance = sd->last_balance + interval;
4750 update_next_balance = 1;
4754 * Stop the load balance at this level. There is another
4755 * CPU in our sched group which is doing load balancing more
4763 * next_balance will be updated only when there is a need.
4764 * When the cpu is attached to null domain for ex, it will not be
4767 if (likely(update_next_balance))
4768 rq->next_balance = next_balance;
4772 * run_rebalance_domains is triggered when needed from the scheduler tick.
4773 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4774 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4776 static void run_rebalance_domains(struct softirq_action *h)
4778 int this_cpu = smp_processor_id();
4779 struct rq *this_rq = cpu_rq(this_cpu);
4780 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4781 CPU_IDLE : CPU_NOT_IDLE;
4783 rebalance_domains(this_cpu, idle);
4787 * If this cpu is the owner for idle load balancing, then do the
4788 * balancing on behalf of the other idle cpus whose ticks are
4791 if (this_rq->idle_at_tick &&
4792 atomic_read(&nohz.load_balancer) == this_cpu) {
4796 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4797 if (balance_cpu == this_cpu)
4801 * If this cpu gets work to do, stop the load balancing
4802 * work being done for other cpus. Next load
4803 * balancing owner will pick it up.
4808 rebalance_domains(balance_cpu, CPU_IDLE);
4810 rq = cpu_rq(balance_cpu);
4811 if (time_after(this_rq->next_balance, rq->next_balance))
4812 this_rq->next_balance = rq->next_balance;
4818 static inline int on_null_domain(int cpu)
4820 return !rcu_dereference(cpu_rq(cpu)->sd);
4824 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4826 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4827 * idle load balancing owner or decide to stop the periodic load balancing,
4828 * if the whole system is idle.
4830 static inline void trigger_load_balance(struct rq *rq, int cpu)
4834 * If we were in the nohz mode recently and busy at the current
4835 * scheduler tick, then check if we need to nominate new idle
4838 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4839 rq->in_nohz_recently = 0;
4841 if (atomic_read(&nohz.load_balancer) == cpu) {
4842 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4843 atomic_set(&nohz.load_balancer, -1);
4846 if (atomic_read(&nohz.load_balancer) == -1) {
4847 int ilb = find_new_ilb(cpu);
4849 if (ilb < nr_cpu_ids)
4855 * If this cpu is idle and doing idle load balancing for all the
4856 * cpus with ticks stopped, is it time for that to stop?
4858 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4859 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4865 * If this cpu is idle and the idle load balancing is done by
4866 * someone else, then no need raise the SCHED_SOFTIRQ
4868 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4869 cpumask_test_cpu(cpu, nohz.cpu_mask))
4872 /* Don't need to rebalance while attached to NULL domain */
4873 if (time_after_eq(jiffies, rq->next_balance) &&
4874 likely(!on_null_domain(cpu)))
4875 raise_softirq(SCHED_SOFTIRQ);
4878 #else /* CONFIG_SMP */
4881 * on UP we do not need to balance between CPUs:
4883 static inline void idle_balance(int cpu, struct rq *rq)
4889 DEFINE_PER_CPU(struct kernel_stat, kstat);
4891 EXPORT_PER_CPU_SYMBOL(kstat);
4894 * Return any ns on the sched_clock that have not yet been accounted in
4895 * @p in case that task is currently running.
4897 * Called with task_rq_lock() held on @rq.
4899 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4903 if (task_current(rq, p)) {
4904 update_rq_clock(rq);
4905 ns = rq->clock - p->se.exec_start;
4913 unsigned long long task_delta_exec(struct task_struct *p)
4915 unsigned long flags;
4919 rq = task_rq_lock(p, &flags);
4920 ns = do_task_delta_exec(p, rq);
4921 task_rq_unlock(rq, &flags);
4927 * Return accounted runtime for the task.
4928 * In case the task is currently running, return the runtime plus current's
4929 * pending runtime that have not been accounted yet.
4931 unsigned long long task_sched_runtime(struct task_struct *p)
4933 unsigned long flags;
4937 rq = task_rq_lock(p, &flags);
4938 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4939 task_rq_unlock(rq, &flags);
4945 * Return sum_exec_runtime for the thread group.
4946 * In case the task is currently running, return the sum plus current's
4947 * pending runtime that have not been accounted yet.
4949 * Note that the thread group might have other running tasks as well,
4950 * so the return value not includes other pending runtime that other
4951 * running tasks might have.
4953 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4955 struct task_cputime totals;
4956 unsigned long flags;
4960 rq = task_rq_lock(p, &flags);
4961 thread_group_cputime(p, &totals);
4962 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4963 task_rq_unlock(rq, &flags);
4969 * Account user cpu time to a process.
4970 * @p: the process that the cpu time gets accounted to
4971 * @cputime: the cpu time spent in user space since the last update
4972 * @cputime_scaled: cputime scaled by cpu frequency
4974 void account_user_time(struct task_struct *p, cputime_t cputime,
4975 cputime_t cputime_scaled)
4977 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4980 /* Add user time to process. */
4981 p->utime = cputime_add(p->utime, cputime);
4982 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4983 account_group_user_time(p, cputime);
4985 /* Add user time to cpustat. */
4986 tmp = cputime_to_cputime64(cputime);
4987 if (TASK_NICE(p) > 0)
4988 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4990 cpustat->user = cputime64_add(cpustat->user, tmp);
4992 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4993 /* Account for user time used */
4994 acct_update_integrals(p);
4998 * Account guest cpu time to a process.
4999 * @p: the process that the cpu time gets accounted to
5000 * @cputime: the cpu time spent in virtual machine since the last update
5001 * @cputime_scaled: cputime scaled by cpu frequency
5003 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5004 cputime_t cputime_scaled)
5007 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5009 tmp = cputime_to_cputime64(cputime);
5011 /* Add guest time to process. */
5012 p->utime = cputime_add(p->utime, cputime);
5013 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5014 account_group_user_time(p, cputime);
5015 p->gtime = cputime_add(p->gtime, cputime);
5017 /* Add guest time to cpustat. */
5018 cpustat->user = cputime64_add(cpustat->user, tmp);
5019 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5023 * Account system cpu time to a process.
5024 * @p: the process that the cpu time gets accounted to
5025 * @hardirq_offset: the offset to subtract from hardirq_count()
5026 * @cputime: the cpu time spent in kernel space since the last update
5027 * @cputime_scaled: cputime scaled by cpu frequency
5029 void account_system_time(struct task_struct *p, int hardirq_offset,
5030 cputime_t cputime, cputime_t cputime_scaled)
5032 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5035 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5036 account_guest_time(p, cputime, cputime_scaled);
5040 /* Add system time to process. */
5041 p->stime = cputime_add(p->stime, cputime);
5042 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5043 account_group_system_time(p, cputime);
5045 /* Add system time to cpustat. */
5046 tmp = cputime_to_cputime64(cputime);
5047 if (hardirq_count() - hardirq_offset)
5048 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5049 else if (softirq_count())
5050 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5052 cpustat->system = cputime64_add(cpustat->system, tmp);
5054 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5056 /* Account for system time used */
5057 acct_update_integrals(p);
5061 * Account for involuntary wait time.
5062 * @steal: the cpu time spent in involuntary wait
5064 void account_steal_time(cputime_t cputime)
5066 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5067 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5069 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5073 * Account for idle time.
5074 * @cputime: the cpu time spent in idle wait
5076 void account_idle_time(cputime_t cputime)
5078 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5079 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5080 struct rq *rq = this_rq();
5082 if (atomic_read(&rq->nr_iowait) > 0)
5083 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5085 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5088 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5091 * Account a single tick of cpu time.
5092 * @p: the process that the cpu time gets accounted to
5093 * @user_tick: indicates if the tick is a user or a system tick
5095 void account_process_tick(struct task_struct *p, int user_tick)
5097 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5098 struct rq *rq = this_rq();
5101 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5102 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5103 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5106 account_idle_time(cputime_one_jiffy);
5110 * Account multiple ticks of steal time.
5111 * @p: the process from which the cpu time has been stolen
5112 * @ticks: number of stolen ticks
5114 void account_steal_ticks(unsigned long ticks)
5116 account_steal_time(jiffies_to_cputime(ticks));
5120 * Account multiple ticks of idle time.
5121 * @ticks: number of stolen ticks
5123 void account_idle_ticks(unsigned long ticks)
5125 account_idle_time(jiffies_to_cputime(ticks));
5131 * Use precise platform statistics if available:
5133 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5134 cputime_t task_utime(struct task_struct *p)
5139 cputime_t task_stime(struct task_struct *p)
5144 cputime_t task_utime(struct task_struct *p)
5146 clock_t utime = cputime_to_clock_t(p->utime),
5147 total = utime + cputime_to_clock_t(p->stime);
5151 * Use CFS's precise accounting:
5153 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5157 do_div(temp, total);
5159 utime = (clock_t)temp;
5161 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5162 return p->prev_utime;
5165 cputime_t task_stime(struct task_struct *p)
5170 * Use CFS's precise accounting. (we subtract utime from
5171 * the total, to make sure the total observed by userspace
5172 * grows monotonically - apps rely on that):
5174 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5175 cputime_to_clock_t(task_utime(p));
5178 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5180 return p->prev_stime;
5184 inline cputime_t task_gtime(struct task_struct *p)
5190 * This function gets called by the timer code, with HZ frequency.
5191 * We call it with interrupts disabled.
5193 * It also gets called by the fork code, when changing the parent's
5196 void scheduler_tick(void)
5198 int cpu = smp_processor_id();
5199 struct rq *rq = cpu_rq(cpu);
5200 struct task_struct *curr = rq->curr;
5204 spin_lock(&rq->lock);
5205 update_rq_clock(rq);
5206 update_cpu_load(rq);
5207 curr->sched_class->task_tick(rq, curr, 0);
5208 spin_unlock(&rq->lock);
5210 perf_event_task_tick(curr, cpu);
5213 rq->idle_at_tick = idle_cpu(cpu);
5214 trigger_load_balance(rq, cpu);
5218 notrace unsigned long get_parent_ip(unsigned long addr)
5220 if (in_lock_functions(addr)) {
5221 addr = CALLER_ADDR2;
5222 if (in_lock_functions(addr))
5223 addr = CALLER_ADDR3;
5228 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5229 defined(CONFIG_PREEMPT_TRACER))
5231 void __kprobes add_preempt_count(int val)
5233 #ifdef CONFIG_DEBUG_PREEMPT
5237 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5240 preempt_count() += val;
5241 #ifdef CONFIG_DEBUG_PREEMPT
5243 * Spinlock count overflowing soon?
5245 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5248 if (preempt_count() == val)
5249 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5251 EXPORT_SYMBOL(add_preempt_count);
5253 void __kprobes sub_preempt_count(int val)
5255 #ifdef CONFIG_DEBUG_PREEMPT
5259 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5262 * Is the spinlock portion underflowing?
5264 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5265 !(preempt_count() & PREEMPT_MASK)))
5269 if (preempt_count() == val)
5270 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5271 preempt_count() -= val;
5273 EXPORT_SYMBOL(sub_preempt_count);
5278 * Print scheduling while atomic bug:
5280 static noinline void __schedule_bug(struct task_struct *prev)
5282 struct pt_regs *regs = get_irq_regs();
5284 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5285 prev->comm, prev->pid, preempt_count());
5287 debug_show_held_locks(prev);
5289 if (irqs_disabled())
5290 print_irqtrace_events(prev);
5299 * Various schedule()-time debugging checks and statistics:
5301 static inline void schedule_debug(struct task_struct *prev)
5304 * Test if we are atomic. Since do_exit() needs to call into
5305 * schedule() atomically, we ignore that path for now.
5306 * Otherwise, whine if we are scheduling when we should not be.
5308 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5309 __schedule_bug(prev);
5311 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5313 schedstat_inc(this_rq(), sched_count);
5314 #ifdef CONFIG_SCHEDSTATS
5315 if (unlikely(prev->lock_depth >= 0)) {
5316 schedstat_inc(this_rq(), bkl_count);
5317 schedstat_inc(prev, sched_info.bkl_count);
5322 static void put_prev_task(struct rq *rq, struct task_struct *p)
5324 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
5326 update_avg(&p->se.avg_running, runtime);
5328 if (p->state == TASK_RUNNING) {
5330 * In order to avoid avg_overlap growing stale when we are
5331 * indeed overlapping and hence not getting put to sleep, grow
5332 * the avg_overlap on preemption.
5334 * We use the average preemption runtime because that
5335 * correlates to the amount of cache footprint a task can
5338 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5339 update_avg(&p->se.avg_overlap, runtime);
5341 update_avg(&p->se.avg_running, 0);
5343 p->sched_class->put_prev_task(rq, p);
5347 * Pick up the highest-prio task:
5349 static inline struct task_struct *
5350 pick_next_task(struct rq *rq)
5352 const struct sched_class *class;
5353 struct task_struct *p;
5356 * Optimization: we know that if all tasks are in
5357 * the fair class we can call that function directly:
5359 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5360 p = fair_sched_class.pick_next_task(rq);
5365 class = sched_class_highest;
5367 p = class->pick_next_task(rq);
5371 * Will never be NULL as the idle class always
5372 * returns a non-NULL p:
5374 class = class->next;
5379 * schedule() is the main scheduler function.
5381 asmlinkage void __sched schedule(void)
5383 struct task_struct *prev, *next;
5384 unsigned long *switch_count;
5390 cpu = smp_processor_id();
5394 switch_count = &prev->nivcsw;
5396 release_kernel_lock(prev);
5397 need_resched_nonpreemptible:
5399 schedule_debug(prev);
5401 if (sched_feat(HRTICK))
5404 spin_lock_irq(&rq->lock);
5405 update_rq_clock(rq);
5406 clear_tsk_need_resched(prev);
5408 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5409 if (unlikely(signal_pending_state(prev->state, prev)))
5410 prev->state = TASK_RUNNING;
5412 deactivate_task(rq, prev, 1);
5413 switch_count = &prev->nvcsw;
5416 pre_schedule(rq, prev);
5418 if (unlikely(!rq->nr_running))
5419 idle_balance(cpu, rq);
5421 put_prev_task(rq, prev);
5422 next = pick_next_task(rq);
5424 if (likely(prev != next)) {
5425 sched_info_switch(prev, next);
5426 perf_event_task_sched_out(prev, next, cpu);
5432 context_switch(rq, prev, next); /* unlocks the rq */
5434 * the context switch might have flipped the stack from under
5435 * us, hence refresh the local variables.
5437 cpu = smp_processor_id();
5440 spin_unlock_irq(&rq->lock);
5444 if (unlikely(reacquire_kernel_lock(current) < 0))
5445 goto need_resched_nonpreemptible;
5447 preempt_enable_no_resched();
5451 EXPORT_SYMBOL(schedule);
5455 * Look out! "owner" is an entirely speculative pointer
5456 * access and not reliable.
5458 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5463 if (!sched_feat(OWNER_SPIN))
5466 #ifdef CONFIG_DEBUG_PAGEALLOC
5468 * Need to access the cpu field knowing that
5469 * DEBUG_PAGEALLOC could have unmapped it if
5470 * the mutex owner just released it and exited.
5472 if (probe_kernel_address(&owner->cpu, cpu))
5479 * Even if the access succeeded (likely case),
5480 * the cpu field may no longer be valid.
5482 if (cpu >= nr_cpumask_bits)
5486 * We need to validate that we can do a
5487 * get_cpu() and that we have the percpu area.
5489 if (!cpu_online(cpu))
5496 * Owner changed, break to re-assess state.
5498 if (lock->owner != owner)
5502 * Is that owner really running on that cpu?
5504 if (task_thread_info(rq->curr) != owner || need_resched())
5514 #ifdef CONFIG_PREEMPT
5516 * this is the entry point to schedule() from in-kernel preemption
5517 * off of preempt_enable. Kernel preemptions off return from interrupt
5518 * occur there and call schedule directly.
5520 asmlinkage void __sched preempt_schedule(void)
5522 struct thread_info *ti = current_thread_info();
5525 * If there is a non-zero preempt_count or interrupts are disabled,
5526 * we do not want to preempt the current task. Just return..
5528 if (likely(ti->preempt_count || irqs_disabled()))
5532 add_preempt_count(PREEMPT_ACTIVE);
5534 sub_preempt_count(PREEMPT_ACTIVE);
5537 * Check again in case we missed a preemption opportunity
5538 * between schedule and now.
5541 } while (need_resched());
5543 EXPORT_SYMBOL(preempt_schedule);
5546 * this is the entry point to schedule() from kernel preemption
5547 * off of irq context.
5548 * Note, that this is called and return with irqs disabled. This will
5549 * protect us against recursive calling from irq.
5551 asmlinkage void __sched preempt_schedule_irq(void)
5553 struct thread_info *ti = current_thread_info();
5555 /* Catch callers which need to be fixed */
5556 BUG_ON(ti->preempt_count || !irqs_disabled());
5559 add_preempt_count(PREEMPT_ACTIVE);
5562 local_irq_disable();
5563 sub_preempt_count(PREEMPT_ACTIVE);
5566 * Check again in case we missed a preemption opportunity
5567 * between schedule and now.
5570 } while (need_resched());
5573 #endif /* CONFIG_PREEMPT */
5575 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5578 return try_to_wake_up(curr->private, mode, wake_flags);
5580 EXPORT_SYMBOL(default_wake_function);
5583 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5584 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5585 * number) then we wake all the non-exclusive tasks and one exclusive task.
5587 * There are circumstances in which we can try to wake a task which has already
5588 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5589 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5591 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5592 int nr_exclusive, int wake_flags, void *key)
5594 wait_queue_t *curr, *next;
5596 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5597 unsigned flags = curr->flags;
5599 if (curr->func(curr, mode, wake_flags, key) &&
5600 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5606 * __wake_up - wake up threads blocked on a waitqueue.
5608 * @mode: which threads
5609 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5610 * @key: is directly passed to the wakeup function
5612 * It may be assumed that this function implies a write memory barrier before
5613 * changing the task state if and only if any tasks are woken up.
5615 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5616 int nr_exclusive, void *key)
5618 unsigned long flags;
5620 spin_lock_irqsave(&q->lock, flags);
5621 __wake_up_common(q, mode, nr_exclusive, 0, key);
5622 spin_unlock_irqrestore(&q->lock, flags);
5624 EXPORT_SYMBOL(__wake_up);
5627 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5629 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5631 __wake_up_common(q, mode, 1, 0, NULL);
5634 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5636 __wake_up_common(q, mode, 1, 0, key);
5640 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5642 * @mode: which threads
5643 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5644 * @key: opaque value to be passed to wakeup targets
5646 * The sync wakeup differs that the waker knows that it will schedule
5647 * away soon, so while the target thread will be woken up, it will not
5648 * be migrated to another CPU - ie. the two threads are 'synchronized'
5649 * with each other. This can prevent needless bouncing between CPUs.
5651 * On UP it can prevent extra preemption.
5653 * It may be assumed that this function implies a write memory barrier before
5654 * changing the task state if and only if any tasks are woken up.
5656 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5657 int nr_exclusive, void *key)
5659 unsigned long flags;
5660 int wake_flags = WF_SYNC;
5665 if (unlikely(!nr_exclusive))
5668 spin_lock_irqsave(&q->lock, flags);
5669 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5670 spin_unlock_irqrestore(&q->lock, flags);
5672 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5675 * __wake_up_sync - see __wake_up_sync_key()
5677 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5679 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5681 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5684 * complete: - signals a single thread waiting on this completion
5685 * @x: holds the state of this particular completion
5687 * This will wake up a single thread waiting on this completion. Threads will be
5688 * awakened in the same order in which they were queued.
5690 * See also complete_all(), wait_for_completion() and related routines.
5692 * It may be assumed that this function implies a write memory barrier before
5693 * changing the task state if and only if any tasks are woken up.
5695 void complete(struct completion *x)
5697 unsigned long flags;
5699 spin_lock_irqsave(&x->wait.lock, flags);
5701 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5702 spin_unlock_irqrestore(&x->wait.lock, flags);
5704 EXPORT_SYMBOL(complete);
5707 * complete_all: - signals all threads waiting on this completion
5708 * @x: holds the state of this particular completion
5710 * This will wake up all threads waiting on this particular completion event.
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 complete_all(struct completion *x)
5717 unsigned long flags;
5719 spin_lock_irqsave(&x->wait.lock, flags);
5720 x->done += UINT_MAX/2;
5721 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5722 spin_unlock_irqrestore(&x->wait.lock, flags);
5724 EXPORT_SYMBOL(complete_all);
5726 static inline long __sched
5727 do_wait_for_common(struct completion *x, long timeout, int state)
5730 DECLARE_WAITQUEUE(wait, current);
5732 wait.flags |= WQ_FLAG_EXCLUSIVE;
5733 __add_wait_queue_tail(&x->wait, &wait);
5735 if (signal_pending_state(state, current)) {
5736 timeout = -ERESTARTSYS;
5739 __set_current_state(state);
5740 spin_unlock_irq(&x->wait.lock);
5741 timeout = schedule_timeout(timeout);
5742 spin_lock_irq(&x->wait.lock);
5743 } while (!x->done && timeout);
5744 __remove_wait_queue(&x->wait, &wait);
5749 return timeout ?: 1;
5753 wait_for_common(struct completion *x, long timeout, int state)
5757 spin_lock_irq(&x->wait.lock);
5758 timeout = do_wait_for_common(x, timeout, state);
5759 spin_unlock_irq(&x->wait.lock);
5764 * wait_for_completion: - waits for completion of a task
5765 * @x: holds the state of this particular completion
5767 * This waits to be signaled for completion of a specific task. It is NOT
5768 * interruptible and there is no timeout.
5770 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5771 * and interrupt capability. Also see complete().
5773 void __sched wait_for_completion(struct completion *x)
5775 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5777 EXPORT_SYMBOL(wait_for_completion);
5780 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5781 * @x: holds the state of this particular completion
5782 * @timeout: timeout value in jiffies
5784 * This waits for either a completion of a specific task to be signaled or for a
5785 * specified timeout to expire. The timeout is in jiffies. It is not
5788 unsigned long __sched
5789 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5791 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5793 EXPORT_SYMBOL(wait_for_completion_timeout);
5796 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5797 * @x: holds the state of this particular completion
5799 * This waits for completion of a specific task to be signaled. It is
5802 int __sched wait_for_completion_interruptible(struct completion *x)
5804 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5805 if (t == -ERESTARTSYS)
5809 EXPORT_SYMBOL(wait_for_completion_interruptible);
5812 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5813 * @x: holds the state of this particular completion
5814 * @timeout: timeout value in jiffies
5816 * This waits for either a completion of a specific task to be signaled or for a
5817 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5819 unsigned long __sched
5820 wait_for_completion_interruptible_timeout(struct completion *x,
5821 unsigned long timeout)
5823 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5825 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5828 * wait_for_completion_killable: - waits for completion of a task (killable)
5829 * @x: holds the state of this particular completion
5831 * This waits to be signaled for completion of a specific task. It can be
5832 * interrupted by a kill signal.
5834 int __sched wait_for_completion_killable(struct completion *x)
5836 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5837 if (t == -ERESTARTSYS)
5841 EXPORT_SYMBOL(wait_for_completion_killable);
5844 * try_wait_for_completion - try to decrement a completion without blocking
5845 * @x: completion structure
5847 * Returns: 0 if a decrement cannot be done without blocking
5848 * 1 if a decrement succeeded.
5850 * If a completion is being used as a counting completion,
5851 * attempt to decrement the counter without blocking. This
5852 * enables us to avoid waiting if the resource the completion
5853 * is protecting is not available.
5855 bool try_wait_for_completion(struct completion *x)
5859 spin_lock_irq(&x->wait.lock);
5864 spin_unlock_irq(&x->wait.lock);
5867 EXPORT_SYMBOL(try_wait_for_completion);
5870 * completion_done - Test to see if a completion has any waiters
5871 * @x: completion structure
5873 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5874 * 1 if there are no waiters.
5877 bool completion_done(struct completion *x)
5881 spin_lock_irq(&x->wait.lock);
5884 spin_unlock_irq(&x->wait.lock);
5887 EXPORT_SYMBOL(completion_done);
5890 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5892 unsigned long flags;
5895 init_waitqueue_entry(&wait, current);
5897 __set_current_state(state);
5899 spin_lock_irqsave(&q->lock, flags);
5900 __add_wait_queue(q, &wait);
5901 spin_unlock(&q->lock);
5902 timeout = schedule_timeout(timeout);
5903 spin_lock_irq(&q->lock);
5904 __remove_wait_queue(q, &wait);
5905 spin_unlock_irqrestore(&q->lock, flags);
5910 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5912 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5914 EXPORT_SYMBOL(interruptible_sleep_on);
5917 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5919 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5921 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5923 void __sched sleep_on(wait_queue_head_t *q)
5925 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5927 EXPORT_SYMBOL(sleep_on);
5929 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5931 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5933 EXPORT_SYMBOL(sleep_on_timeout);
5935 #ifdef CONFIG_RT_MUTEXES
5938 * rt_mutex_setprio - set the current priority of a task
5940 * @prio: prio value (kernel-internal form)
5942 * This function changes the 'effective' priority of a task. It does
5943 * not touch ->normal_prio like __setscheduler().
5945 * Used by the rt_mutex code to implement priority inheritance logic.
5947 void rt_mutex_setprio(struct task_struct *p, int prio)
5949 unsigned long flags;
5950 int oldprio, on_rq, running;
5952 const struct sched_class *prev_class = p->sched_class;
5954 BUG_ON(prio < 0 || prio > MAX_PRIO);
5956 rq = task_rq_lock(p, &flags);
5957 update_rq_clock(rq);
5960 on_rq = p->se.on_rq;
5961 running = task_current(rq, p);
5963 dequeue_task(rq, p, 0);
5965 p->sched_class->put_prev_task(rq, p);
5968 p->sched_class = &rt_sched_class;
5970 p->sched_class = &fair_sched_class;
5975 p->sched_class->set_curr_task(rq);
5977 enqueue_task(rq, p, 0);
5979 check_class_changed(rq, p, prev_class, oldprio, running);
5981 task_rq_unlock(rq, &flags);
5986 void set_user_nice(struct task_struct *p, long nice)
5988 int old_prio, delta, on_rq;
5989 unsigned long flags;
5992 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5995 * We have to be careful, if called from sys_setpriority(),
5996 * the task might be in the middle of scheduling on another CPU.
5998 rq = task_rq_lock(p, &flags);
5999 update_rq_clock(rq);
6001 * The RT priorities are set via sched_setscheduler(), but we still
6002 * allow the 'normal' nice value to be set - but as expected
6003 * it wont have any effect on scheduling until the task is
6004 * SCHED_FIFO/SCHED_RR:
6006 if (task_has_rt_policy(p)) {
6007 p->static_prio = NICE_TO_PRIO(nice);
6010 on_rq = p->se.on_rq;
6012 dequeue_task(rq, p, 0);
6014 p->static_prio = NICE_TO_PRIO(nice);
6017 p->prio = effective_prio(p);
6018 delta = p->prio - old_prio;
6021 enqueue_task(rq, p, 0);
6023 * If the task increased its priority or is running and
6024 * lowered its priority, then reschedule its CPU:
6026 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6027 resched_task(rq->curr);
6030 task_rq_unlock(rq, &flags);
6032 EXPORT_SYMBOL(set_user_nice);
6035 * can_nice - check if a task can reduce its nice value
6039 int can_nice(const struct task_struct *p, const int nice)
6041 /* convert nice value [19,-20] to rlimit style value [1,40] */
6042 int nice_rlim = 20 - nice;
6044 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6045 capable(CAP_SYS_NICE));
6048 #ifdef __ARCH_WANT_SYS_NICE
6051 * sys_nice - change the priority of the current process.
6052 * @increment: priority increment
6054 * sys_setpriority is a more generic, but much slower function that
6055 * does similar things.
6057 SYSCALL_DEFINE1(nice, int, increment)
6062 * Setpriority might change our priority at the same moment.
6063 * We don't have to worry. Conceptually one call occurs first
6064 * and we have a single winner.
6066 if (increment < -40)
6071 nice = TASK_NICE(current) + increment;
6077 if (increment < 0 && !can_nice(current, nice))
6080 retval = security_task_setnice(current, nice);
6084 set_user_nice(current, nice);
6091 * task_prio - return the priority value of a given task.
6092 * @p: the task in question.
6094 * This is the priority value as seen by users in /proc.
6095 * RT tasks are offset by -200. Normal tasks are centered
6096 * around 0, value goes from -16 to +15.
6098 int task_prio(const struct task_struct *p)
6100 return p->prio - MAX_RT_PRIO;
6104 * task_nice - return the nice value of a given task.
6105 * @p: the task in question.
6107 int task_nice(const struct task_struct *p)
6109 return TASK_NICE(p);
6111 EXPORT_SYMBOL(task_nice);
6114 * idle_cpu - is a given cpu idle currently?
6115 * @cpu: the processor in question.
6117 int idle_cpu(int cpu)
6119 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6123 * idle_task - return the idle task for a given cpu.
6124 * @cpu: the processor in question.
6126 struct task_struct *idle_task(int cpu)
6128 return cpu_rq(cpu)->idle;
6132 * find_process_by_pid - find a process with a matching PID value.
6133 * @pid: the pid in question.
6135 static struct task_struct *find_process_by_pid(pid_t pid)
6137 return pid ? find_task_by_vpid(pid) : current;
6140 /* Actually do priority change: must hold rq lock. */
6142 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6144 BUG_ON(p->se.on_rq);
6147 switch (p->policy) {
6151 p->sched_class = &fair_sched_class;
6155 p->sched_class = &rt_sched_class;
6159 p->rt_priority = prio;
6160 p->normal_prio = normal_prio(p);
6161 /* we are holding p->pi_lock already */
6162 p->prio = rt_mutex_getprio(p);
6167 * check the target process has a UID that matches the current process's
6169 static bool check_same_owner(struct task_struct *p)
6171 const struct cred *cred = current_cred(), *pcred;
6175 pcred = __task_cred(p);
6176 match = (cred->euid == pcred->euid ||
6177 cred->euid == pcred->uid);
6182 static int __sched_setscheduler(struct task_struct *p, int policy,
6183 struct sched_param *param, bool user)
6185 int retval, oldprio, oldpolicy = -1, on_rq, running;
6186 unsigned long flags;
6187 const struct sched_class *prev_class = p->sched_class;
6191 /* may grab non-irq protected spin_locks */
6192 BUG_ON(in_interrupt());
6194 /* double check policy once rq lock held */
6196 reset_on_fork = p->sched_reset_on_fork;
6197 policy = oldpolicy = p->policy;
6199 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6200 policy &= ~SCHED_RESET_ON_FORK;
6202 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6203 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6204 policy != SCHED_IDLE)
6209 * Valid priorities for SCHED_FIFO and SCHED_RR are
6210 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6211 * SCHED_BATCH and SCHED_IDLE is 0.
6213 if (param->sched_priority < 0 ||
6214 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6215 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6217 if (rt_policy(policy) != (param->sched_priority != 0))
6221 * Allow unprivileged RT tasks to decrease priority:
6223 if (user && !capable(CAP_SYS_NICE)) {
6224 if (rt_policy(policy)) {
6225 unsigned long rlim_rtprio;
6227 if (!lock_task_sighand(p, &flags))
6229 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6230 unlock_task_sighand(p, &flags);
6232 /* can't set/change the rt policy */
6233 if (policy != p->policy && !rlim_rtprio)
6236 /* can't increase priority */
6237 if (param->sched_priority > p->rt_priority &&
6238 param->sched_priority > rlim_rtprio)
6242 * Like positive nice levels, dont allow tasks to
6243 * move out of SCHED_IDLE either:
6245 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6248 /* can't change other user's priorities */
6249 if (!check_same_owner(p))
6252 /* Normal users shall not reset the sched_reset_on_fork flag */
6253 if (p->sched_reset_on_fork && !reset_on_fork)
6258 #ifdef CONFIG_RT_GROUP_SCHED
6260 * Do not allow realtime tasks into groups that have no runtime
6263 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6264 task_group(p)->rt_bandwidth.rt_runtime == 0)
6268 retval = security_task_setscheduler(p, policy, param);
6274 * make sure no PI-waiters arrive (or leave) while we are
6275 * changing the priority of the task:
6277 spin_lock_irqsave(&p->pi_lock, flags);
6279 * To be able to change p->policy safely, the apropriate
6280 * runqueue lock must be held.
6282 rq = __task_rq_lock(p);
6283 /* recheck policy now with rq lock held */
6284 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6285 policy = oldpolicy = -1;
6286 __task_rq_unlock(rq);
6287 spin_unlock_irqrestore(&p->pi_lock, flags);
6290 update_rq_clock(rq);
6291 on_rq = p->se.on_rq;
6292 running = task_current(rq, p);
6294 deactivate_task(rq, p, 0);
6296 p->sched_class->put_prev_task(rq, p);
6298 p->sched_reset_on_fork = reset_on_fork;
6301 __setscheduler(rq, p, policy, param->sched_priority);
6304 p->sched_class->set_curr_task(rq);
6306 activate_task(rq, p, 0);
6308 check_class_changed(rq, p, prev_class, oldprio, running);
6310 __task_rq_unlock(rq);
6311 spin_unlock_irqrestore(&p->pi_lock, flags);
6313 rt_mutex_adjust_pi(p);
6319 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6320 * @p: the task in question.
6321 * @policy: new policy.
6322 * @param: structure containing the new RT priority.
6324 * NOTE that the task may be already dead.
6326 int sched_setscheduler(struct task_struct *p, int policy,
6327 struct sched_param *param)
6329 return __sched_setscheduler(p, policy, param, true);
6331 EXPORT_SYMBOL_GPL(sched_setscheduler);
6334 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6335 * @p: the task in question.
6336 * @policy: new policy.
6337 * @param: structure containing the new RT priority.
6339 * Just like sched_setscheduler, only don't bother checking if the
6340 * current context has permission. For example, this is needed in
6341 * stop_machine(): we create temporary high priority worker threads,
6342 * but our caller might not have that capability.
6344 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6345 struct sched_param *param)
6347 return __sched_setscheduler(p, policy, param, false);
6351 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6353 struct sched_param lparam;
6354 struct task_struct *p;
6357 if (!param || pid < 0)
6359 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6364 p = find_process_by_pid(pid);
6366 retval = sched_setscheduler(p, policy, &lparam);
6373 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6374 * @pid: the pid in question.
6375 * @policy: new policy.
6376 * @param: structure containing the new RT priority.
6378 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6379 struct sched_param __user *, param)
6381 /* negative values for policy are not valid */
6385 return do_sched_setscheduler(pid, policy, param);
6389 * sys_sched_setparam - set/change the RT priority of a thread
6390 * @pid: the pid in question.
6391 * @param: structure containing the new RT priority.
6393 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6395 return do_sched_setscheduler(pid, -1, param);
6399 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6400 * @pid: the pid in question.
6402 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6404 struct task_struct *p;
6411 read_lock(&tasklist_lock);
6412 p = find_process_by_pid(pid);
6414 retval = security_task_getscheduler(p);
6417 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6419 read_unlock(&tasklist_lock);
6424 * sys_sched_getparam - get the RT priority of a thread
6425 * @pid: the pid in question.
6426 * @param: structure containing the RT priority.
6428 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6430 struct sched_param lp;
6431 struct task_struct *p;
6434 if (!param || pid < 0)
6437 read_lock(&tasklist_lock);
6438 p = find_process_by_pid(pid);
6443 retval = security_task_getscheduler(p);
6447 lp.sched_priority = p->rt_priority;
6448 read_unlock(&tasklist_lock);
6451 * This one might sleep, we cannot do it with a spinlock held ...
6453 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6458 read_unlock(&tasklist_lock);
6462 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6464 cpumask_var_t cpus_allowed, new_mask;
6465 struct task_struct *p;
6469 read_lock(&tasklist_lock);
6471 p = find_process_by_pid(pid);
6473 read_unlock(&tasklist_lock);
6479 * It is not safe to call set_cpus_allowed with the
6480 * tasklist_lock held. We will bump the task_struct's
6481 * usage count and then drop tasklist_lock.
6484 read_unlock(&tasklist_lock);
6486 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6490 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6492 goto out_free_cpus_allowed;
6495 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6498 retval = security_task_setscheduler(p, 0, NULL);
6502 cpuset_cpus_allowed(p, cpus_allowed);
6503 cpumask_and(new_mask, in_mask, cpus_allowed);
6505 retval = set_cpus_allowed_ptr(p, new_mask);
6508 cpuset_cpus_allowed(p, cpus_allowed);
6509 if (!cpumask_subset(new_mask, cpus_allowed)) {
6511 * We must have raced with a concurrent cpuset
6512 * update. Just reset the cpus_allowed to the
6513 * cpuset's cpus_allowed
6515 cpumask_copy(new_mask, cpus_allowed);
6520 free_cpumask_var(new_mask);
6521 out_free_cpus_allowed:
6522 free_cpumask_var(cpus_allowed);
6529 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6530 struct cpumask *new_mask)
6532 if (len < cpumask_size())
6533 cpumask_clear(new_mask);
6534 else if (len > cpumask_size())
6535 len = cpumask_size();
6537 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6541 * sys_sched_setaffinity - set the cpu affinity of a process
6542 * @pid: pid of the process
6543 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6544 * @user_mask_ptr: user-space pointer to the new cpu mask
6546 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6547 unsigned long __user *, user_mask_ptr)
6549 cpumask_var_t new_mask;
6552 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6555 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6557 retval = sched_setaffinity(pid, new_mask);
6558 free_cpumask_var(new_mask);
6562 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6564 struct task_struct *p;
6568 read_lock(&tasklist_lock);
6571 p = find_process_by_pid(pid);
6575 retval = security_task_getscheduler(p);
6579 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6582 read_unlock(&tasklist_lock);
6589 * sys_sched_getaffinity - get the cpu affinity of a process
6590 * @pid: pid of the process
6591 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6592 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6594 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6595 unsigned long __user *, user_mask_ptr)
6600 if (len < cpumask_size())
6603 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6606 ret = sched_getaffinity(pid, mask);
6608 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6611 ret = cpumask_size();
6613 free_cpumask_var(mask);
6619 * sys_sched_yield - yield the current processor to other threads.
6621 * This function yields the current CPU to other tasks. If there are no
6622 * other threads running on this CPU then this function will return.
6624 SYSCALL_DEFINE0(sched_yield)
6626 struct rq *rq = this_rq_lock();
6628 schedstat_inc(rq, yld_count);
6629 current->sched_class->yield_task(rq);
6632 * Since we are going to call schedule() anyway, there's
6633 * no need to preempt or enable interrupts:
6635 __release(rq->lock);
6636 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6637 _raw_spin_unlock(&rq->lock);
6638 preempt_enable_no_resched();
6645 static inline int should_resched(void)
6647 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6650 static void __cond_resched(void)
6652 add_preempt_count(PREEMPT_ACTIVE);
6654 sub_preempt_count(PREEMPT_ACTIVE);
6657 int __sched _cond_resched(void)
6659 if (should_resched()) {
6665 EXPORT_SYMBOL(_cond_resched);
6668 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6669 * call schedule, and on return reacquire the lock.
6671 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6672 * operations here to prevent schedule() from being called twice (once via
6673 * spin_unlock(), once by hand).
6675 int __cond_resched_lock(spinlock_t *lock)
6677 int resched = should_resched();
6680 lockdep_assert_held(lock);
6682 if (spin_needbreak(lock) || resched) {
6693 EXPORT_SYMBOL(__cond_resched_lock);
6695 int __sched __cond_resched_softirq(void)
6697 BUG_ON(!in_softirq());
6699 if (should_resched()) {
6707 EXPORT_SYMBOL(__cond_resched_softirq);
6710 * yield - yield the current processor to other threads.
6712 * This is a shortcut for kernel-space yielding - it marks the
6713 * thread runnable and calls sys_sched_yield().
6715 void __sched yield(void)
6717 set_current_state(TASK_RUNNING);
6720 EXPORT_SYMBOL(yield);
6723 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6724 * that process accounting knows that this is a task in IO wait state.
6726 * But don't do that if it is a deliberate, throttling IO wait (this task
6727 * has set its backing_dev_info: the queue against which it should throttle)
6729 void __sched io_schedule(void)
6731 struct rq *rq = raw_rq();
6733 delayacct_blkio_start();
6734 atomic_inc(&rq->nr_iowait);
6735 current->in_iowait = 1;
6737 current->in_iowait = 0;
6738 atomic_dec(&rq->nr_iowait);
6739 delayacct_blkio_end();
6741 EXPORT_SYMBOL(io_schedule);
6743 long __sched io_schedule_timeout(long timeout)
6745 struct rq *rq = raw_rq();
6748 delayacct_blkio_start();
6749 atomic_inc(&rq->nr_iowait);
6750 current->in_iowait = 1;
6751 ret = schedule_timeout(timeout);
6752 current->in_iowait = 0;
6753 atomic_dec(&rq->nr_iowait);
6754 delayacct_blkio_end();
6759 * sys_sched_get_priority_max - return maximum RT priority.
6760 * @policy: scheduling class.
6762 * this syscall returns the maximum rt_priority that can be used
6763 * by a given scheduling class.
6765 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6772 ret = MAX_USER_RT_PRIO-1;
6784 * sys_sched_get_priority_min - return minimum RT priority.
6785 * @policy: scheduling class.
6787 * this syscall returns the minimum rt_priority that can be used
6788 * by a given scheduling class.
6790 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6808 * sys_sched_rr_get_interval - return the default timeslice of a process.
6809 * @pid: pid of the process.
6810 * @interval: userspace pointer to the timeslice value.
6812 * this syscall writes the default timeslice value of a given process
6813 * into the user-space timespec buffer. A value of '0' means infinity.
6815 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6816 struct timespec __user *, interval)
6818 struct task_struct *p;
6819 unsigned int time_slice;
6827 read_lock(&tasklist_lock);
6828 p = find_process_by_pid(pid);
6832 retval = security_task_getscheduler(p);
6836 time_slice = p->sched_class->get_rr_interval(p);
6838 read_unlock(&tasklist_lock);
6839 jiffies_to_timespec(time_slice, &t);
6840 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6844 read_unlock(&tasklist_lock);
6848 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6850 void sched_show_task(struct task_struct *p)
6852 unsigned long free = 0;
6855 state = p->state ? __ffs(p->state) + 1 : 0;
6856 printk(KERN_INFO "%-13.13s %c", p->comm,
6857 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6858 #if BITS_PER_LONG == 32
6859 if (state == TASK_RUNNING)
6860 printk(KERN_CONT " running ");
6862 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6864 if (state == TASK_RUNNING)
6865 printk(KERN_CONT " running task ");
6867 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6869 #ifdef CONFIG_DEBUG_STACK_USAGE
6870 free = stack_not_used(p);
6872 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6873 task_pid_nr(p), task_pid_nr(p->real_parent),
6874 (unsigned long)task_thread_info(p)->flags);
6876 show_stack(p, NULL);
6879 void show_state_filter(unsigned long state_filter)
6881 struct task_struct *g, *p;
6883 #if BITS_PER_LONG == 32
6885 " task PC stack pid father\n");
6888 " task PC stack pid father\n");
6890 read_lock(&tasklist_lock);
6891 do_each_thread(g, p) {
6893 * reset the NMI-timeout, listing all files on a slow
6894 * console might take alot of time:
6896 touch_nmi_watchdog();
6897 if (!state_filter || (p->state & state_filter))
6899 } while_each_thread(g, p);
6901 touch_all_softlockup_watchdogs();
6903 #ifdef CONFIG_SCHED_DEBUG
6904 sysrq_sched_debug_show();
6906 read_unlock(&tasklist_lock);
6908 * Only show locks if all tasks are dumped:
6910 if (state_filter == -1)
6911 debug_show_all_locks();
6914 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6916 idle->sched_class = &idle_sched_class;
6920 * init_idle - set up an idle thread for a given CPU
6921 * @idle: task in question
6922 * @cpu: cpu the idle task belongs to
6924 * NOTE: this function does not set the idle thread's NEED_RESCHED
6925 * flag, to make booting more robust.
6927 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6929 struct rq *rq = cpu_rq(cpu);
6930 unsigned long flags;
6932 spin_lock_irqsave(&rq->lock, flags);
6935 idle->se.exec_start = sched_clock();
6937 idle->prio = idle->normal_prio = MAX_PRIO;
6938 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6939 __set_task_cpu(idle, cpu);
6941 rq->curr = rq->idle = idle;
6942 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6945 spin_unlock_irqrestore(&rq->lock, flags);
6947 /* Set the preempt count _outside_ the spinlocks! */
6948 #if defined(CONFIG_PREEMPT)
6949 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6951 task_thread_info(idle)->preempt_count = 0;
6954 * The idle tasks have their own, simple scheduling class:
6956 idle->sched_class = &idle_sched_class;
6957 ftrace_graph_init_task(idle);
6961 * In a system that switches off the HZ timer nohz_cpu_mask
6962 * indicates which cpus entered this state. This is used
6963 * in the rcu update to wait only for active cpus. For system
6964 * which do not switch off the HZ timer nohz_cpu_mask should
6965 * always be CPU_BITS_NONE.
6967 cpumask_var_t nohz_cpu_mask;
6970 * Increase the granularity value when there are more CPUs,
6971 * because with more CPUs the 'effective latency' as visible
6972 * to users decreases. But the relationship is not linear,
6973 * so pick a second-best guess by going with the log2 of the
6976 * This idea comes from the SD scheduler of Con Kolivas:
6978 static inline void sched_init_granularity(void)
6980 unsigned int factor = 1 + ilog2(num_online_cpus());
6981 const unsigned long limit = 200000000;
6983 sysctl_sched_min_granularity *= factor;
6984 if (sysctl_sched_min_granularity > limit)
6985 sysctl_sched_min_granularity = limit;
6987 sysctl_sched_latency *= factor;
6988 if (sysctl_sched_latency > limit)
6989 sysctl_sched_latency = limit;
6991 sysctl_sched_wakeup_granularity *= factor;
6993 sysctl_sched_shares_ratelimit *= factor;
6998 * This is how migration works:
7000 * 1) we queue a struct migration_req structure in the source CPU's
7001 * runqueue and wake up that CPU's migration thread.
7002 * 2) we down() the locked semaphore => thread blocks.
7003 * 3) migration thread wakes up (implicitly it forces the migrated
7004 * thread off the CPU)
7005 * 4) it gets the migration request and checks whether the migrated
7006 * task is still in the wrong runqueue.
7007 * 5) if it's in the wrong runqueue then the migration thread removes
7008 * it and puts it into the right queue.
7009 * 6) migration thread up()s the semaphore.
7010 * 7) we wake up and the migration is done.
7014 * Change a given task's CPU affinity. Migrate the thread to a
7015 * proper CPU and schedule it away if the CPU it's executing on
7016 * is removed from the allowed bitmask.
7018 * NOTE: the caller must have a valid reference to the task, the
7019 * task must not exit() & deallocate itself prematurely. The
7020 * call is not atomic; no spinlocks may be held.
7022 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7024 struct migration_req req;
7025 unsigned long flags;
7029 rq = task_rq_lock(p, &flags);
7030 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7035 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7036 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7041 if (p->sched_class->set_cpus_allowed)
7042 p->sched_class->set_cpus_allowed(p, new_mask);
7044 cpumask_copy(&p->cpus_allowed, new_mask);
7045 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7048 /* Can the task run on the task's current CPU? If so, we're done */
7049 if (cpumask_test_cpu(task_cpu(p), new_mask))
7052 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7053 /* Need help from migration thread: drop lock and wait. */
7054 struct task_struct *mt = rq->migration_thread;
7056 get_task_struct(mt);
7057 task_rq_unlock(rq, &flags);
7058 wake_up_process(rq->migration_thread);
7059 put_task_struct(mt);
7060 wait_for_completion(&req.done);
7061 tlb_migrate_finish(p->mm);
7065 task_rq_unlock(rq, &flags);
7069 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7072 * Move (not current) task off this cpu, onto dest cpu. We're doing
7073 * this because either it can't run here any more (set_cpus_allowed()
7074 * away from this CPU, or CPU going down), or because we're
7075 * attempting to rebalance this task on exec (sched_exec).
7077 * So we race with normal scheduler movements, but that's OK, as long
7078 * as the task is no longer on this CPU.
7080 * Returns non-zero if task was successfully migrated.
7082 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7084 struct rq *rq_dest, *rq_src;
7087 if (unlikely(!cpu_active(dest_cpu)))
7090 rq_src = cpu_rq(src_cpu);
7091 rq_dest = cpu_rq(dest_cpu);
7093 double_rq_lock(rq_src, rq_dest);
7094 /* Already moved. */
7095 if (task_cpu(p) != src_cpu)
7097 /* Affinity changed (again). */
7098 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7101 on_rq = p->se.on_rq;
7103 deactivate_task(rq_src, p, 0);
7105 set_task_cpu(p, dest_cpu);
7107 activate_task(rq_dest, p, 0);
7108 check_preempt_curr(rq_dest, p, 0);
7113 double_rq_unlock(rq_src, rq_dest);
7117 #define RCU_MIGRATION_IDLE 0
7118 #define RCU_MIGRATION_NEED_QS 1
7119 #define RCU_MIGRATION_GOT_QS 2
7120 #define RCU_MIGRATION_MUST_SYNC 3
7123 * migration_thread - this is a highprio system thread that performs
7124 * thread migration by bumping thread off CPU then 'pushing' onto
7127 static int migration_thread(void *data)
7130 int cpu = (long)data;
7134 BUG_ON(rq->migration_thread != current);
7136 set_current_state(TASK_INTERRUPTIBLE);
7137 while (!kthread_should_stop()) {
7138 struct migration_req *req;
7139 struct list_head *head;
7141 spin_lock_irq(&rq->lock);
7143 if (cpu_is_offline(cpu)) {
7144 spin_unlock_irq(&rq->lock);
7148 if (rq->active_balance) {
7149 active_load_balance(rq, cpu);
7150 rq->active_balance = 0;
7153 head = &rq->migration_queue;
7155 if (list_empty(head)) {
7156 spin_unlock_irq(&rq->lock);
7158 set_current_state(TASK_INTERRUPTIBLE);
7161 req = list_entry(head->next, struct migration_req, list);
7162 list_del_init(head->next);
7164 if (req->task != NULL) {
7165 spin_unlock(&rq->lock);
7166 __migrate_task(req->task, cpu, req->dest_cpu);
7167 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7168 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7169 spin_unlock(&rq->lock);
7171 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7172 spin_unlock(&rq->lock);
7173 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7177 complete(&req->done);
7179 __set_current_state(TASK_RUNNING);
7184 #ifdef CONFIG_HOTPLUG_CPU
7186 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7190 local_irq_disable();
7191 ret = __migrate_task(p, src_cpu, dest_cpu);
7197 * Figure out where task on dead CPU should go, use force if necessary.
7199 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7202 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7205 /* Look for allowed, online CPU in same node. */
7206 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7207 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7210 /* Any allowed, online CPU? */
7211 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7212 if (dest_cpu < nr_cpu_ids)
7215 /* No more Mr. Nice Guy. */
7216 if (dest_cpu >= nr_cpu_ids) {
7217 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7218 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7221 * Don't tell them about moving exiting tasks or
7222 * kernel threads (both mm NULL), since they never
7225 if (p->mm && printk_ratelimit()) {
7226 printk(KERN_INFO "process %d (%s) no "
7227 "longer affine to cpu%d\n",
7228 task_pid_nr(p), p->comm, dead_cpu);
7233 /* It can have affinity changed while we were choosing. */
7234 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7239 * While a dead CPU has no uninterruptible tasks queued at this point,
7240 * it might still have a nonzero ->nr_uninterruptible counter, because
7241 * for performance reasons the counter is not stricly tracking tasks to
7242 * their home CPUs. So we just add the counter to another CPU's counter,
7243 * to keep the global sum constant after CPU-down:
7245 static void migrate_nr_uninterruptible(struct rq *rq_src)
7247 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7248 unsigned long flags;
7250 local_irq_save(flags);
7251 double_rq_lock(rq_src, rq_dest);
7252 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7253 rq_src->nr_uninterruptible = 0;
7254 double_rq_unlock(rq_src, rq_dest);
7255 local_irq_restore(flags);
7258 /* Run through task list and migrate tasks from the dead cpu. */
7259 static void migrate_live_tasks(int src_cpu)
7261 struct task_struct *p, *t;
7263 read_lock(&tasklist_lock);
7265 do_each_thread(t, p) {
7269 if (task_cpu(p) == src_cpu)
7270 move_task_off_dead_cpu(src_cpu, p);
7271 } while_each_thread(t, p);
7273 read_unlock(&tasklist_lock);
7277 * Schedules idle task to be the next runnable task on current CPU.
7278 * It does so by boosting its priority to highest possible.
7279 * Used by CPU offline code.
7281 void sched_idle_next(void)
7283 int this_cpu = smp_processor_id();
7284 struct rq *rq = cpu_rq(this_cpu);
7285 struct task_struct *p = rq->idle;
7286 unsigned long flags;
7288 /* cpu has to be offline */
7289 BUG_ON(cpu_online(this_cpu));
7292 * Strictly not necessary since rest of the CPUs are stopped by now
7293 * and interrupts disabled on the current cpu.
7295 spin_lock_irqsave(&rq->lock, flags);
7297 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7299 update_rq_clock(rq);
7300 activate_task(rq, p, 0);
7302 spin_unlock_irqrestore(&rq->lock, flags);
7306 * Ensures that the idle task is using init_mm right before its cpu goes
7309 void idle_task_exit(void)
7311 struct mm_struct *mm = current->active_mm;
7313 BUG_ON(cpu_online(smp_processor_id()));
7316 switch_mm(mm, &init_mm, current);
7320 /* called under rq->lock with disabled interrupts */
7321 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7323 struct rq *rq = cpu_rq(dead_cpu);
7325 /* Must be exiting, otherwise would be on tasklist. */
7326 BUG_ON(!p->exit_state);
7328 /* Cannot have done final schedule yet: would have vanished. */
7329 BUG_ON(p->state == TASK_DEAD);
7334 * Drop lock around migration; if someone else moves it,
7335 * that's OK. No task can be added to this CPU, so iteration is
7338 spin_unlock_irq(&rq->lock);
7339 move_task_off_dead_cpu(dead_cpu, p);
7340 spin_lock_irq(&rq->lock);
7345 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7346 static void migrate_dead_tasks(unsigned int dead_cpu)
7348 struct rq *rq = cpu_rq(dead_cpu);
7349 struct task_struct *next;
7352 if (!rq->nr_running)
7354 update_rq_clock(rq);
7355 next = pick_next_task(rq);
7358 next->sched_class->put_prev_task(rq, next);
7359 migrate_dead(dead_cpu, next);
7365 * remove the tasks which were accounted by rq from calc_load_tasks.
7367 static void calc_global_load_remove(struct rq *rq)
7369 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7370 rq->calc_load_active = 0;
7372 #endif /* CONFIG_HOTPLUG_CPU */
7374 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7376 static struct ctl_table sd_ctl_dir[] = {
7378 .procname = "sched_domain",
7384 static struct ctl_table sd_ctl_root[] = {
7386 .ctl_name = CTL_KERN,
7387 .procname = "kernel",
7389 .child = sd_ctl_dir,
7394 static struct ctl_table *sd_alloc_ctl_entry(int n)
7396 struct ctl_table *entry =
7397 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7402 static void sd_free_ctl_entry(struct ctl_table **tablep)
7404 struct ctl_table *entry;
7407 * In the intermediate directories, both the child directory and
7408 * procname are dynamically allocated and could fail but the mode
7409 * will always be set. In the lowest directory the names are
7410 * static strings and all have proc handlers.
7412 for (entry = *tablep; entry->mode; entry++) {
7414 sd_free_ctl_entry(&entry->child);
7415 if (entry->proc_handler == NULL)
7416 kfree(entry->procname);
7424 set_table_entry(struct ctl_table *entry,
7425 const char *procname, void *data, int maxlen,
7426 mode_t mode, proc_handler *proc_handler)
7428 entry->procname = procname;
7430 entry->maxlen = maxlen;
7432 entry->proc_handler = proc_handler;
7435 static struct ctl_table *
7436 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7438 struct ctl_table *table = sd_alloc_ctl_entry(13);
7443 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7444 sizeof(long), 0644, proc_doulongvec_minmax);
7445 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7446 sizeof(long), 0644, proc_doulongvec_minmax);
7447 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7448 sizeof(int), 0644, proc_dointvec_minmax);
7449 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7450 sizeof(int), 0644, proc_dointvec_minmax);
7451 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7452 sizeof(int), 0644, proc_dointvec_minmax);
7453 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7454 sizeof(int), 0644, proc_dointvec_minmax);
7455 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7456 sizeof(int), 0644, proc_dointvec_minmax);
7457 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7458 sizeof(int), 0644, proc_dointvec_minmax);
7459 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7460 sizeof(int), 0644, proc_dointvec_minmax);
7461 set_table_entry(&table[9], "cache_nice_tries",
7462 &sd->cache_nice_tries,
7463 sizeof(int), 0644, proc_dointvec_minmax);
7464 set_table_entry(&table[10], "flags", &sd->flags,
7465 sizeof(int), 0644, proc_dointvec_minmax);
7466 set_table_entry(&table[11], "name", sd->name,
7467 CORENAME_MAX_SIZE, 0444, proc_dostring);
7468 /* &table[12] is terminator */
7473 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7475 struct ctl_table *entry, *table;
7476 struct sched_domain *sd;
7477 int domain_num = 0, i;
7480 for_each_domain(cpu, sd)
7482 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7487 for_each_domain(cpu, sd) {
7488 snprintf(buf, 32, "domain%d", i);
7489 entry->procname = kstrdup(buf, GFP_KERNEL);
7491 entry->child = sd_alloc_ctl_domain_table(sd);
7498 static struct ctl_table_header *sd_sysctl_header;
7499 static void register_sched_domain_sysctl(void)
7501 int i, cpu_num = num_online_cpus();
7502 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7505 WARN_ON(sd_ctl_dir[0].child);
7506 sd_ctl_dir[0].child = entry;
7511 for_each_online_cpu(i) {
7512 snprintf(buf, 32, "cpu%d", i);
7513 entry->procname = kstrdup(buf, GFP_KERNEL);
7515 entry->child = sd_alloc_ctl_cpu_table(i);
7519 WARN_ON(sd_sysctl_header);
7520 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7523 /* may be called multiple times per register */
7524 static void unregister_sched_domain_sysctl(void)
7526 if (sd_sysctl_header)
7527 unregister_sysctl_table(sd_sysctl_header);
7528 sd_sysctl_header = NULL;
7529 if (sd_ctl_dir[0].child)
7530 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7533 static void register_sched_domain_sysctl(void)
7536 static void unregister_sched_domain_sysctl(void)
7541 static void set_rq_online(struct rq *rq)
7544 const struct sched_class *class;
7546 cpumask_set_cpu(rq->cpu, rq->rd->online);
7549 for_each_class(class) {
7550 if (class->rq_online)
7551 class->rq_online(rq);
7556 static void set_rq_offline(struct rq *rq)
7559 const struct sched_class *class;
7561 for_each_class(class) {
7562 if (class->rq_offline)
7563 class->rq_offline(rq);
7566 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7572 * migration_call - callback that gets triggered when a CPU is added.
7573 * Here we can start up the necessary migration thread for the new CPU.
7575 static int __cpuinit
7576 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7578 struct task_struct *p;
7579 int cpu = (long)hcpu;
7580 unsigned long flags;
7585 case CPU_UP_PREPARE:
7586 case CPU_UP_PREPARE_FROZEN:
7587 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7590 kthread_bind(p, cpu);
7591 /* Must be high prio: stop_machine expects to yield to it. */
7592 rq = task_rq_lock(p, &flags);
7593 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7594 task_rq_unlock(rq, &flags);
7596 cpu_rq(cpu)->migration_thread = p;
7597 rq->calc_load_update = calc_load_update;
7601 case CPU_ONLINE_FROZEN:
7602 /* Strictly unnecessary, as first user will wake it. */
7603 wake_up_process(cpu_rq(cpu)->migration_thread);
7605 /* Update our root-domain */
7607 spin_lock_irqsave(&rq->lock, flags);
7609 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7613 spin_unlock_irqrestore(&rq->lock, flags);
7616 #ifdef CONFIG_HOTPLUG_CPU
7617 case CPU_UP_CANCELED:
7618 case CPU_UP_CANCELED_FROZEN:
7619 if (!cpu_rq(cpu)->migration_thread)
7621 /* Unbind it from offline cpu so it can run. Fall thru. */
7622 kthread_bind(cpu_rq(cpu)->migration_thread,
7623 cpumask_any(cpu_online_mask));
7624 kthread_stop(cpu_rq(cpu)->migration_thread);
7625 put_task_struct(cpu_rq(cpu)->migration_thread);
7626 cpu_rq(cpu)->migration_thread = NULL;
7630 case CPU_DEAD_FROZEN:
7631 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7632 migrate_live_tasks(cpu);
7634 kthread_stop(rq->migration_thread);
7635 put_task_struct(rq->migration_thread);
7636 rq->migration_thread = NULL;
7637 /* Idle task back to normal (off runqueue, low prio) */
7638 spin_lock_irq(&rq->lock);
7639 update_rq_clock(rq);
7640 deactivate_task(rq, rq->idle, 0);
7641 rq->idle->static_prio = MAX_PRIO;
7642 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7643 rq->idle->sched_class = &idle_sched_class;
7644 migrate_dead_tasks(cpu);
7645 spin_unlock_irq(&rq->lock);
7647 migrate_nr_uninterruptible(rq);
7648 BUG_ON(rq->nr_running != 0);
7649 calc_global_load_remove(rq);
7651 * No need to migrate the tasks: it was best-effort if
7652 * they didn't take sched_hotcpu_mutex. Just wake up
7655 spin_lock_irq(&rq->lock);
7656 while (!list_empty(&rq->migration_queue)) {
7657 struct migration_req *req;
7659 req = list_entry(rq->migration_queue.next,
7660 struct migration_req, list);
7661 list_del_init(&req->list);
7662 spin_unlock_irq(&rq->lock);
7663 complete(&req->done);
7664 spin_lock_irq(&rq->lock);
7666 spin_unlock_irq(&rq->lock);
7670 case CPU_DYING_FROZEN:
7671 /* Update our root-domain */
7673 spin_lock_irqsave(&rq->lock, flags);
7675 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7678 spin_unlock_irqrestore(&rq->lock, flags);
7686 * Register at high priority so that task migration (migrate_all_tasks)
7687 * happens before everything else. This has to be lower priority than
7688 * the notifier in the perf_event subsystem, though.
7690 static struct notifier_block __cpuinitdata migration_notifier = {
7691 .notifier_call = migration_call,
7695 static int __init migration_init(void)
7697 void *cpu = (void *)(long)smp_processor_id();
7700 /* Start one for the boot CPU: */
7701 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7702 BUG_ON(err == NOTIFY_BAD);
7703 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7704 register_cpu_notifier(&migration_notifier);
7708 early_initcall(migration_init);
7713 #ifdef CONFIG_SCHED_DEBUG
7715 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7716 struct cpumask *groupmask)
7718 struct sched_group *group = sd->groups;
7721 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7722 cpumask_clear(groupmask);
7724 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7726 if (!(sd->flags & SD_LOAD_BALANCE)) {
7727 printk("does not load-balance\n");
7729 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7734 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7736 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7737 printk(KERN_ERR "ERROR: domain->span does not contain "
7740 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7741 printk(KERN_ERR "ERROR: domain->groups does not contain"
7745 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7749 printk(KERN_ERR "ERROR: group is NULL\n");
7753 if (!group->cpu_power) {
7754 printk(KERN_CONT "\n");
7755 printk(KERN_ERR "ERROR: domain->cpu_power not "
7760 if (!cpumask_weight(sched_group_cpus(group))) {
7761 printk(KERN_CONT "\n");
7762 printk(KERN_ERR "ERROR: empty group\n");
7766 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7767 printk(KERN_CONT "\n");
7768 printk(KERN_ERR "ERROR: repeated CPUs\n");
7772 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7774 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7776 printk(KERN_CONT " %s", str);
7777 if (group->cpu_power != SCHED_LOAD_SCALE) {
7778 printk(KERN_CONT " (cpu_power = %d)",
7782 group = group->next;
7783 } while (group != sd->groups);
7784 printk(KERN_CONT "\n");
7786 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7787 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7790 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7791 printk(KERN_ERR "ERROR: parent span is not a superset "
7792 "of domain->span\n");
7796 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7798 cpumask_var_t groupmask;
7802 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7806 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7808 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7809 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7814 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7821 free_cpumask_var(groupmask);
7823 #else /* !CONFIG_SCHED_DEBUG */
7824 # define sched_domain_debug(sd, cpu) do { } while (0)
7825 #endif /* CONFIG_SCHED_DEBUG */
7827 static int sd_degenerate(struct sched_domain *sd)
7829 if (cpumask_weight(sched_domain_span(sd)) == 1)
7832 /* Following flags need at least 2 groups */
7833 if (sd->flags & (SD_LOAD_BALANCE |
7834 SD_BALANCE_NEWIDLE |
7838 SD_SHARE_PKG_RESOURCES)) {
7839 if (sd->groups != sd->groups->next)
7843 /* Following flags don't use groups */
7844 if (sd->flags & (SD_WAKE_AFFINE))
7851 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7853 unsigned long cflags = sd->flags, pflags = parent->flags;
7855 if (sd_degenerate(parent))
7858 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7861 /* Flags needing groups don't count if only 1 group in parent */
7862 if (parent->groups == parent->groups->next) {
7863 pflags &= ~(SD_LOAD_BALANCE |
7864 SD_BALANCE_NEWIDLE |
7868 SD_SHARE_PKG_RESOURCES);
7869 if (nr_node_ids == 1)
7870 pflags &= ~SD_SERIALIZE;
7872 if (~cflags & pflags)
7878 static void free_rootdomain(struct root_domain *rd)
7880 cpupri_cleanup(&rd->cpupri);
7882 free_cpumask_var(rd->rto_mask);
7883 free_cpumask_var(rd->online);
7884 free_cpumask_var(rd->span);
7888 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7890 struct root_domain *old_rd = NULL;
7891 unsigned long flags;
7893 spin_lock_irqsave(&rq->lock, flags);
7898 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7901 cpumask_clear_cpu(rq->cpu, old_rd->span);
7904 * If we dont want to free the old_rt yet then
7905 * set old_rd to NULL to skip the freeing later
7908 if (!atomic_dec_and_test(&old_rd->refcount))
7912 atomic_inc(&rd->refcount);
7915 cpumask_set_cpu(rq->cpu, rd->span);
7916 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7919 spin_unlock_irqrestore(&rq->lock, flags);
7922 free_rootdomain(old_rd);
7925 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7927 gfp_t gfp = GFP_KERNEL;
7929 memset(rd, 0, sizeof(*rd));
7934 if (!alloc_cpumask_var(&rd->span, gfp))
7936 if (!alloc_cpumask_var(&rd->online, gfp))
7938 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7941 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7946 free_cpumask_var(rd->rto_mask);
7948 free_cpumask_var(rd->online);
7950 free_cpumask_var(rd->span);
7955 static void init_defrootdomain(void)
7957 init_rootdomain(&def_root_domain, true);
7959 atomic_set(&def_root_domain.refcount, 1);
7962 static struct root_domain *alloc_rootdomain(void)
7964 struct root_domain *rd;
7966 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7970 if (init_rootdomain(rd, false) != 0) {
7979 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7980 * hold the hotplug lock.
7983 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7985 struct rq *rq = cpu_rq(cpu);
7986 struct sched_domain *tmp;
7988 /* Remove the sched domains which do not contribute to scheduling. */
7989 for (tmp = sd; tmp; ) {
7990 struct sched_domain *parent = tmp->parent;
7994 if (sd_parent_degenerate(tmp, parent)) {
7995 tmp->parent = parent->parent;
7997 parent->parent->child = tmp;
8002 if (sd && sd_degenerate(sd)) {
8008 sched_domain_debug(sd, cpu);
8010 rq_attach_root(rq, rd);
8011 rcu_assign_pointer(rq->sd, sd);
8014 /* cpus with isolated domains */
8015 static cpumask_var_t cpu_isolated_map;
8017 /* Setup the mask of cpus configured for isolated domains */
8018 static int __init isolated_cpu_setup(char *str)
8020 cpulist_parse(str, cpu_isolated_map);
8024 __setup("isolcpus=", isolated_cpu_setup);
8027 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8028 * to a function which identifies what group(along with sched group) a CPU
8029 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8030 * (due to the fact that we keep track of groups covered with a struct cpumask).
8032 * init_sched_build_groups will build a circular linked list of the groups
8033 * covered by the given span, and will set each group's ->cpumask correctly,
8034 * and ->cpu_power to 0.
8037 init_sched_build_groups(const struct cpumask *span,
8038 const struct cpumask *cpu_map,
8039 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8040 struct sched_group **sg,
8041 struct cpumask *tmpmask),
8042 struct cpumask *covered, struct cpumask *tmpmask)
8044 struct sched_group *first = NULL, *last = NULL;
8047 cpumask_clear(covered);
8049 for_each_cpu(i, span) {
8050 struct sched_group *sg;
8051 int group = group_fn(i, cpu_map, &sg, tmpmask);
8054 if (cpumask_test_cpu(i, covered))
8057 cpumask_clear(sched_group_cpus(sg));
8060 for_each_cpu(j, span) {
8061 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8064 cpumask_set_cpu(j, covered);
8065 cpumask_set_cpu(j, sched_group_cpus(sg));
8076 #define SD_NODES_PER_DOMAIN 16
8081 * find_next_best_node - find the next node to include in a sched_domain
8082 * @node: node whose sched_domain we're building
8083 * @used_nodes: nodes already in the sched_domain
8085 * Find the next node to include in a given scheduling domain. Simply
8086 * finds the closest node not already in the @used_nodes map.
8088 * Should use nodemask_t.
8090 static int find_next_best_node(int node, nodemask_t *used_nodes)
8092 int i, n, val, min_val, best_node = 0;
8096 for (i = 0; i < nr_node_ids; i++) {
8097 /* Start at @node */
8098 n = (node + i) % nr_node_ids;
8100 if (!nr_cpus_node(n))
8103 /* Skip already used nodes */
8104 if (node_isset(n, *used_nodes))
8107 /* Simple min distance search */
8108 val = node_distance(node, n);
8110 if (val < min_val) {
8116 node_set(best_node, *used_nodes);
8121 * sched_domain_node_span - get a cpumask for a node's sched_domain
8122 * @node: node whose cpumask we're constructing
8123 * @span: resulting cpumask
8125 * Given a node, construct a good cpumask for its sched_domain to span. It
8126 * should be one that prevents unnecessary balancing, but also spreads tasks
8129 static void sched_domain_node_span(int node, struct cpumask *span)
8131 nodemask_t used_nodes;
8134 cpumask_clear(span);
8135 nodes_clear(used_nodes);
8137 cpumask_or(span, span, cpumask_of_node(node));
8138 node_set(node, used_nodes);
8140 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8141 int next_node = find_next_best_node(node, &used_nodes);
8143 cpumask_or(span, span, cpumask_of_node(next_node));
8146 #endif /* CONFIG_NUMA */
8148 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8151 * The cpus mask in sched_group and sched_domain hangs off the end.
8153 * ( See the the comments in include/linux/sched.h:struct sched_group
8154 * and struct sched_domain. )
8156 struct static_sched_group {
8157 struct sched_group sg;
8158 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8161 struct static_sched_domain {
8162 struct sched_domain sd;
8163 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8169 cpumask_var_t domainspan;
8170 cpumask_var_t covered;
8171 cpumask_var_t notcovered;
8173 cpumask_var_t nodemask;
8174 cpumask_var_t this_sibling_map;
8175 cpumask_var_t this_core_map;
8176 cpumask_var_t send_covered;
8177 cpumask_var_t tmpmask;
8178 struct sched_group **sched_group_nodes;
8179 struct root_domain *rd;
8183 sa_sched_groups = 0,
8188 sa_this_sibling_map,
8190 sa_sched_group_nodes,
8200 * SMT sched-domains:
8202 #ifdef CONFIG_SCHED_SMT
8203 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8204 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8207 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8208 struct sched_group **sg, struct cpumask *unused)
8211 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8214 #endif /* CONFIG_SCHED_SMT */
8217 * multi-core sched-domains:
8219 #ifdef CONFIG_SCHED_MC
8220 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8221 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8222 #endif /* CONFIG_SCHED_MC */
8224 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8226 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8227 struct sched_group **sg, struct cpumask *mask)
8231 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8232 group = cpumask_first(mask);
8234 *sg = &per_cpu(sched_group_core, group).sg;
8237 #elif defined(CONFIG_SCHED_MC)
8239 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8240 struct sched_group **sg, struct cpumask *unused)
8243 *sg = &per_cpu(sched_group_core, cpu).sg;
8248 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8249 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8252 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8253 struct sched_group **sg, struct cpumask *mask)
8256 #ifdef CONFIG_SCHED_MC
8257 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8258 group = cpumask_first(mask);
8259 #elif defined(CONFIG_SCHED_SMT)
8260 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8261 group = cpumask_first(mask);
8266 *sg = &per_cpu(sched_group_phys, group).sg;
8272 * The init_sched_build_groups can't handle what we want to do with node
8273 * groups, so roll our own. Now each node has its own list of groups which
8274 * gets dynamically allocated.
8276 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8277 static struct sched_group ***sched_group_nodes_bycpu;
8279 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8280 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8282 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8283 struct sched_group **sg,
8284 struct cpumask *nodemask)
8288 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8289 group = cpumask_first(nodemask);
8292 *sg = &per_cpu(sched_group_allnodes, group).sg;
8296 static void init_numa_sched_groups_power(struct sched_group *group_head)
8298 struct sched_group *sg = group_head;
8304 for_each_cpu(j, sched_group_cpus(sg)) {
8305 struct sched_domain *sd;
8307 sd = &per_cpu(phys_domains, j).sd;
8308 if (j != group_first_cpu(sd->groups)) {
8310 * Only add "power" once for each
8316 sg->cpu_power += sd->groups->cpu_power;
8319 } while (sg != group_head);
8322 static int build_numa_sched_groups(struct s_data *d,
8323 const struct cpumask *cpu_map, int num)
8325 struct sched_domain *sd;
8326 struct sched_group *sg, *prev;
8329 cpumask_clear(d->covered);
8330 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8331 if (cpumask_empty(d->nodemask)) {
8332 d->sched_group_nodes[num] = NULL;
8336 sched_domain_node_span(num, d->domainspan);
8337 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8339 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8342 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8346 d->sched_group_nodes[num] = sg;
8348 for_each_cpu(j, d->nodemask) {
8349 sd = &per_cpu(node_domains, j).sd;
8354 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8356 cpumask_or(d->covered, d->covered, d->nodemask);
8359 for (j = 0; j < nr_node_ids; j++) {
8360 n = (num + j) % nr_node_ids;
8361 cpumask_complement(d->notcovered, d->covered);
8362 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8363 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8364 if (cpumask_empty(d->tmpmask))
8366 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8367 if (cpumask_empty(d->tmpmask))
8369 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8373 "Can not alloc domain group for node %d\n", j);
8377 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8378 sg->next = prev->next;
8379 cpumask_or(d->covered, d->covered, d->tmpmask);
8386 #endif /* CONFIG_NUMA */
8389 /* Free memory allocated for various sched_group structures */
8390 static void free_sched_groups(const struct cpumask *cpu_map,
8391 struct cpumask *nodemask)
8395 for_each_cpu(cpu, cpu_map) {
8396 struct sched_group **sched_group_nodes
8397 = sched_group_nodes_bycpu[cpu];
8399 if (!sched_group_nodes)
8402 for (i = 0; i < nr_node_ids; i++) {
8403 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8405 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8406 if (cpumask_empty(nodemask))
8416 if (oldsg != sched_group_nodes[i])
8419 kfree(sched_group_nodes);
8420 sched_group_nodes_bycpu[cpu] = NULL;
8423 #else /* !CONFIG_NUMA */
8424 static void free_sched_groups(const struct cpumask *cpu_map,
8425 struct cpumask *nodemask)
8428 #endif /* CONFIG_NUMA */
8431 * Initialize sched groups cpu_power.
8433 * cpu_power indicates the capacity of sched group, which is used while
8434 * distributing the load between different sched groups in a sched domain.
8435 * Typically cpu_power for all the groups in a sched domain will be same unless
8436 * there are asymmetries in the topology. If there are asymmetries, group
8437 * having more cpu_power will pickup more load compared to the group having
8440 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8442 struct sched_domain *child;
8443 struct sched_group *group;
8447 WARN_ON(!sd || !sd->groups);
8449 if (cpu != group_first_cpu(sd->groups))
8454 sd->groups->cpu_power = 0;
8457 power = SCHED_LOAD_SCALE;
8458 weight = cpumask_weight(sched_domain_span(sd));
8460 * SMT siblings share the power of a single core.
8461 * Usually multiple threads get a better yield out of
8462 * that one core than a single thread would have,
8463 * reflect that in sd->smt_gain.
8465 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8466 power *= sd->smt_gain;
8468 power >>= SCHED_LOAD_SHIFT;
8470 sd->groups->cpu_power += power;
8475 * Add cpu_power of each child group to this groups cpu_power.
8477 group = child->groups;
8479 sd->groups->cpu_power += group->cpu_power;
8480 group = group->next;
8481 } while (group != child->groups);
8485 * Initializers for schedule domains
8486 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8489 #ifdef CONFIG_SCHED_DEBUG
8490 # define SD_INIT_NAME(sd, type) sd->name = #type
8492 # define SD_INIT_NAME(sd, type) do { } while (0)
8495 #define SD_INIT(sd, type) sd_init_##type(sd)
8497 #define SD_INIT_FUNC(type) \
8498 static noinline void sd_init_##type(struct sched_domain *sd) \
8500 memset(sd, 0, sizeof(*sd)); \
8501 *sd = SD_##type##_INIT; \
8502 sd->level = SD_LV_##type; \
8503 SD_INIT_NAME(sd, type); \
8508 SD_INIT_FUNC(ALLNODES)
8511 #ifdef CONFIG_SCHED_SMT
8512 SD_INIT_FUNC(SIBLING)
8514 #ifdef CONFIG_SCHED_MC
8518 static int default_relax_domain_level = -1;
8520 static int __init setup_relax_domain_level(char *str)
8524 val = simple_strtoul(str, NULL, 0);
8525 if (val < SD_LV_MAX)
8526 default_relax_domain_level = val;
8530 __setup("relax_domain_level=", setup_relax_domain_level);
8532 static void set_domain_attribute(struct sched_domain *sd,
8533 struct sched_domain_attr *attr)
8537 if (!attr || attr->relax_domain_level < 0) {
8538 if (default_relax_domain_level < 0)
8541 request = default_relax_domain_level;
8543 request = attr->relax_domain_level;
8544 if (request < sd->level) {
8545 /* turn off idle balance on this domain */
8546 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8548 /* turn on idle balance on this domain */
8549 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8553 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8554 const struct cpumask *cpu_map)
8557 case sa_sched_groups:
8558 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8559 d->sched_group_nodes = NULL;
8561 free_rootdomain(d->rd); /* fall through */
8563 free_cpumask_var(d->tmpmask); /* fall through */
8564 case sa_send_covered:
8565 free_cpumask_var(d->send_covered); /* fall through */
8566 case sa_this_core_map:
8567 free_cpumask_var(d->this_core_map); /* fall through */
8568 case sa_this_sibling_map:
8569 free_cpumask_var(d->this_sibling_map); /* fall through */
8571 free_cpumask_var(d->nodemask); /* fall through */
8572 case sa_sched_group_nodes:
8574 kfree(d->sched_group_nodes); /* fall through */
8576 free_cpumask_var(d->notcovered); /* fall through */
8578 free_cpumask_var(d->covered); /* fall through */
8580 free_cpumask_var(d->domainspan); /* fall through */
8587 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8588 const struct cpumask *cpu_map)
8591 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8593 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8594 return sa_domainspan;
8595 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8597 /* Allocate the per-node list of sched groups */
8598 d->sched_group_nodes = kcalloc(nr_node_ids,
8599 sizeof(struct sched_group *), GFP_KERNEL);
8600 if (!d->sched_group_nodes) {
8601 printk(KERN_WARNING "Can not alloc sched group node list\n");
8602 return sa_notcovered;
8604 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8606 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8607 return sa_sched_group_nodes;
8608 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8610 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8611 return sa_this_sibling_map;
8612 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8613 return sa_this_core_map;
8614 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8615 return sa_send_covered;
8616 d->rd = alloc_rootdomain();
8618 printk(KERN_WARNING "Cannot alloc root domain\n");
8621 return sa_rootdomain;
8624 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8625 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8627 struct sched_domain *sd = NULL;
8629 struct sched_domain *parent;
8632 if (cpumask_weight(cpu_map) >
8633 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8634 sd = &per_cpu(allnodes_domains, i).sd;
8635 SD_INIT(sd, ALLNODES);
8636 set_domain_attribute(sd, attr);
8637 cpumask_copy(sched_domain_span(sd), cpu_map);
8638 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8643 sd = &per_cpu(node_domains, i).sd;
8645 set_domain_attribute(sd, attr);
8646 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8647 sd->parent = parent;
8650 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8655 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8656 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8657 struct sched_domain *parent, int i)
8659 struct sched_domain *sd;
8660 sd = &per_cpu(phys_domains, i).sd;
8662 set_domain_attribute(sd, attr);
8663 cpumask_copy(sched_domain_span(sd), d->nodemask);
8664 sd->parent = parent;
8667 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8671 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8672 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8673 struct sched_domain *parent, int i)
8675 struct sched_domain *sd = parent;
8676 #ifdef CONFIG_SCHED_MC
8677 sd = &per_cpu(core_domains, i).sd;
8679 set_domain_attribute(sd, attr);
8680 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8681 sd->parent = parent;
8683 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8688 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8689 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8690 struct sched_domain *parent, int i)
8692 struct sched_domain *sd = parent;
8693 #ifdef CONFIG_SCHED_SMT
8694 sd = &per_cpu(cpu_domains, i).sd;
8695 SD_INIT(sd, SIBLING);
8696 set_domain_attribute(sd, attr);
8697 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8698 sd->parent = parent;
8700 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8705 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8706 const struct cpumask *cpu_map, int cpu)
8709 #ifdef CONFIG_SCHED_SMT
8710 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8711 cpumask_and(d->this_sibling_map, cpu_map,
8712 topology_thread_cpumask(cpu));
8713 if (cpu == cpumask_first(d->this_sibling_map))
8714 init_sched_build_groups(d->this_sibling_map, cpu_map,
8716 d->send_covered, d->tmpmask);
8719 #ifdef CONFIG_SCHED_MC
8720 case SD_LV_MC: /* set up multi-core groups */
8721 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8722 if (cpu == cpumask_first(d->this_core_map))
8723 init_sched_build_groups(d->this_core_map, cpu_map,
8725 d->send_covered, d->tmpmask);
8728 case SD_LV_CPU: /* set up physical groups */
8729 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8730 if (!cpumask_empty(d->nodemask))
8731 init_sched_build_groups(d->nodemask, cpu_map,
8733 d->send_covered, d->tmpmask);
8736 case SD_LV_ALLNODES:
8737 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8738 d->send_covered, d->tmpmask);
8747 * Build sched domains for a given set of cpus and attach the sched domains
8748 * to the individual cpus
8750 static int __build_sched_domains(const struct cpumask *cpu_map,
8751 struct sched_domain_attr *attr)
8753 enum s_alloc alloc_state = sa_none;
8755 struct sched_domain *sd;
8761 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8762 if (alloc_state != sa_rootdomain)
8764 alloc_state = sa_sched_groups;
8767 * Set up domains for cpus specified by the cpu_map.
8769 for_each_cpu(i, cpu_map) {
8770 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8773 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8774 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8775 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8776 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8779 for_each_cpu(i, cpu_map) {
8780 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8781 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8784 /* Set up physical groups */
8785 for (i = 0; i < nr_node_ids; i++)
8786 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8789 /* Set up node groups */
8791 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8793 for (i = 0; i < nr_node_ids; i++)
8794 if (build_numa_sched_groups(&d, cpu_map, i))
8798 /* Calculate CPU power for physical packages and nodes */
8799 #ifdef CONFIG_SCHED_SMT
8800 for_each_cpu(i, cpu_map) {
8801 sd = &per_cpu(cpu_domains, i).sd;
8802 init_sched_groups_power(i, sd);
8805 #ifdef CONFIG_SCHED_MC
8806 for_each_cpu(i, cpu_map) {
8807 sd = &per_cpu(core_domains, i).sd;
8808 init_sched_groups_power(i, sd);
8812 for_each_cpu(i, cpu_map) {
8813 sd = &per_cpu(phys_domains, i).sd;
8814 init_sched_groups_power(i, sd);
8818 for (i = 0; i < nr_node_ids; i++)
8819 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8821 if (d.sd_allnodes) {
8822 struct sched_group *sg;
8824 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8826 init_numa_sched_groups_power(sg);
8830 /* Attach the domains */
8831 for_each_cpu(i, cpu_map) {
8832 #ifdef CONFIG_SCHED_SMT
8833 sd = &per_cpu(cpu_domains, i).sd;
8834 #elif defined(CONFIG_SCHED_MC)
8835 sd = &per_cpu(core_domains, i).sd;
8837 sd = &per_cpu(phys_domains, i).sd;
8839 cpu_attach_domain(sd, d.rd, i);
8842 d.sched_group_nodes = NULL; /* don't free this we still need it */
8843 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8847 __free_domain_allocs(&d, alloc_state, cpu_map);
8851 static int build_sched_domains(const struct cpumask *cpu_map)
8853 return __build_sched_domains(cpu_map, NULL);
8856 static struct cpumask *doms_cur; /* current sched domains */
8857 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8858 static struct sched_domain_attr *dattr_cur;
8859 /* attribues of custom domains in 'doms_cur' */
8862 * Special case: If a kmalloc of a doms_cur partition (array of
8863 * cpumask) fails, then fallback to a single sched domain,
8864 * as determined by the single cpumask fallback_doms.
8866 static cpumask_var_t fallback_doms;
8869 * arch_update_cpu_topology lets virtualized architectures update the
8870 * cpu core maps. It is supposed to return 1 if the topology changed
8871 * or 0 if it stayed the same.
8873 int __attribute__((weak)) arch_update_cpu_topology(void)
8879 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8880 * For now this just excludes isolated cpus, but could be used to
8881 * exclude other special cases in the future.
8883 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8887 arch_update_cpu_topology();
8889 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8891 doms_cur = fallback_doms;
8892 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8894 err = build_sched_domains(doms_cur);
8895 register_sched_domain_sysctl();
8900 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8901 struct cpumask *tmpmask)
8903 free_sched_groups(cpu_map, tmpmask);
8907 * Detach sched domains from a group of cpus specified in cpu_map
8908 * These cpus will now be attached to the NULL domain
8910 static void detach_destroy_domains(const struct cpumask *cpu_map)
8912 /* Save because hotplug lock held. */
8913 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8916 for_each_cpu(i, cpu_map)
8917 cpu_attach_domain(NULL, &def_root_domain, i);
8918 synchronize_sched();
8919 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8922 /* handle null as "default" */
8923 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8924 struct sched_domain_attr *new, int idx_new)
8926 struct sched_domain_attr tmp;
8933 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8934 new ? (new + idx_new) : &tmp,
8935 sizeof(struct sched_domain_attr));
8939 * Partition sched domains as specified by the 'ndoms_new'
8940 * cpumasks in the array doms_new[] of cpumasks. This compares
8941 * doms_new[] to the current sched domain partitioning, doms_cur[].
8942 * It destroys each deleted domain and builds each new domain.
8944 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8945 * The masks don't intersect (don't overlap.) We should setup one
8946 * sched domain for each mask. CPUs not in any of the cpumasks will
8947 * not be load balanced. If the same cpumask appears both in the
8948 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8951 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8952 * ownership of it and will kfree it when done with it. If the caller
8953 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8954 * ndoms_new == 1, and partition_sched_domains() will fallback to
8955 * the single partition 'fallback_doms', it also forces the domains
8958 * If doms_new == NULL it will be replaced with cpu_online_mask.
8959 * ndoms_new == 0 is a special case for destroying existing domains,
8960 * and it will not create the default domain.
8962 * Call with hotplug lock held
8964 /* FIXME: Change to struct cpumask *doms_new[] */
8965 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8966 struct sched_domain_attr *dattr_new)
8971 mutex_lock(&sched_domains_mutex);
8973 /* always unregister in case we don't destroy any domains */
8974 unregister_sched_domain_sysctl();
8976 /* Let architecture update cpu core mappings. */
8977 new_topology = arch_update_cpu_topology();
8979 n = doms_new ? ndoms_new : 0;
8981 /* Destroy deleted domains */
8982 for (i = 0; i < ndoms_cur; i++) {
8983 for (j = 0; j < n && !new_topology; j++) {
8984 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8985 && dattrs_equal(dattr_cur, i, dattr_new, j))
8988 /* no match - a current sched domain not in new doms_new[] */
8989 detach_destroy_domains(doms_cur + i);
8994 if (doms_new == NULL) {
8996 doms_new = fallback_doms;
8997 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8998 WARN_ON_ONCE(dattr_new);
9001 /* Build new domains */
9002 for (i = 0; i < ndoms_new; i++) {
9003 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9004 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9005 && dattrs_equal(dattr_new, i, dattr_cur, j))
9008 /* no match - add a new doms_new */
9009 __build_sched_domains(doms_new + i,
9010 dattr_new ? dattr_new + i : NULL);
9015 /* Remember the new sched domains */
9016 if (doms_cur != fallback_doms)
9018 kfree(dattr_cur); /* kfree(NULL) is safe */
9019 doms_cur = doms_new;
9020 dattr_cur = dattr_new;
9021 ndoms_cur = ndoms_new;
9023 register_sched_domain_sysctl();
9025 mutex_unlock(&sched_domains_mutex);
9028 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9029 static void arch_reinit_sched_domains(void)
9033 /* Destroy domains first to force the rebuild */
9034 partition_sched_domains(0, NULL, NULL);
9036 rebuild_sched_domains();
9040 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9042 unsigned int level = 0;
9044 if (sscanf(buf, "%u", &level) != 1)
9048 * level is always be positive so don't check for
9049 * level < POWERSAVINGS_BALANCE_NONE which is 0
9050 * What happens on 0 or 1 byte write,
9051 * need to check for count as well?
9054 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9058 sched_smt_power_savings = level;
9060 sched_mc_power_savings = level;
9062 arch_reinit_sched_domains();
9067 #ifdef CONFIG_SCHED_MC
9068 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9071 return sprintf(page, "%u\n", sched_mc_power_savings);
9073 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9074 const char *buf, size_t count)
9076 return sched_power_savings_store(buf, count, 0);
9078 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9079 sched_mc_power_savings_show,
9080 sched_mc_power_savings_store);
9083 #ifdef CONFIG_SCHED_SMT
9084 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9087 return sprintf(page, "%u\n", sched_smt_power_savings);
9089 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9090 const char *buf, size_t count)
9092 return sched_power_savings_store(buf, count, 1);
9094 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9095 sched_smt_power_savings_show,
9096 sched_smt_power_savings_store);
9099 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9103 #ifdef CONFIG_SCHED_SMT
9105 err = sysfs_create_file(&cls->kset.kobj,
9106 &attr_sched_smt_power_savings.attr);
9108 #ifdef CONFIG_SCHED_MC
9109 if (!err && mc_capable())
9110 err = sysfs_create_file(&cls->kset.kobj,
9111 &attr_sched_mc_power_savings.attr);
9115 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9117 #ifndef CONFIG_CPUSETS
9119 * Add online and remove offline CPUs from the scheduler domains.
9120 * When cpusets are enabled they take over this function.
9122 static int update_sched_domains(struct notifier_block *nfb,
9123 unsigned long action, void *hcpu)
9127 case CPU_ONLINE_FROZEN:
9129 case CPU_DEAD_FROZEN:
9130 partition_sched_domains(1, NULL, NULL);
9139 static int update_runtime(struct notifier_block *nfb,
9140 unsigned long action, void *hcpu)
9142 int cpu = (int)(long)hcpu;
9145 case CPU_DOWN_PREPARE:
9146 case CPU_DOWN_PREPARE_FROZEN:
9147 disable_runtime(cpu_rq(cpu));
9150 case CPU_DOWN_FAILED:
9151 case CPU_DOWN_FAILED_FROZEN:
9153 case CPU_ONLINE_FROZEN:
9154 enable_runtime(cpu_rq(cpu));
9162 void __init sched_init_smp(void)
9164 cpumask_var_t non_isolated_cpus;
9166 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9167 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9169 #if defined(CONFIG_NUMA)
9170 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9172 BUG_ON(sched_group_nodes_bycpu == NULL);
9175 mutex_lock(&sched_domains_mutex);
9176 arch_init_sched_domains(cpu_online_mask);
9177 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9178 if (cpumask_empty(non_isolated_cpus))
9179 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9180 mutex_unlock(&sched_domains_mutex);
9183 #ifndef CONFIG_CPUSETS
9184 /* XXX: Theoretical race here - CPU may be hotplugged now */
9185 hotcpu_notifier(update_sched_domains, 0);
9188 /* RT runtime code needs to handle some hotplug events */
9189 hotcpu_notifier(update_runtime, 0);
9193 /* Move init over to a non-isolated CPU */
9194 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9196 sched_init_granularity();
9197 free_cpumask_var(non_isolated_cpus);
9199 init_sched_rt_class();
9202 void __init sched_init_smp(void)
9204 sched_init_granularity();
9206 #endif /* CONFIG_SMP */
9208 const_debug unsigned int sysctl_timer_migration = 1;
9210 int in_sched_functions(unsigned long addr)
9212 return in_lock_functions(addr) ||
9213 (addr >= (unsigned long)__sched_text_start
9214 && addr < (unsigned long)__sched_text_end);
9217 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9219 cfs_rq->tasks_timeline = RB_ROOT;
9220 INIT_LIST_HEAD(&cfs_rq->tasks);
9221 #ifdef CONFIG_FAIR_GROUP_SCHED
9224 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9227 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9229 struct rt_prio_array *array;
9232 array = &rt_rq->active;
9233 for (i = 0; i < MAX_RT_PRIO; i++) {
9234 INIT_LIST_HEAD(array->queue + i);
9235 __clear_bit(i, array->bitmap);
9237 /* delimiter for bitsearch: */
9238 __set_bit(MAX_RT_PRIO, array->bitmap);
9240 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9241 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9243 rt_rq->highest_prio.next = MAX_RT_PRIO;
9247 rt_rq->rt_nr_migratory = 0;
9248 rt_rq->overloaded = 0;
9249 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9253 rt_rq->rt_throttled = 0;
9254 rt_rq->rt_runtime = 0;
9255 spin_lock_init(&rt_rq->rt_runtime_lock);
9257 #ifdef CONFIG_RT_GROUP_SCHED
9258 rt_rq->rt_nr_boosted = 0;
9263 #ifdef CONFIG_FAIR_GROUP_SCHED
9264 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9265 struct sched_entity *se, int cpu, int add,
9266 struct sched_entity *parent)
9268 struct rq *rq = cpu_rq(cpu);
9269 tg->cfs_rq[cpu] = cfs_rq;
9270 init_cfs_rq(cfs_rq, rq);
9273 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9276 /* se could be NULL for init_task_group */
9281 se->cfs_rq = &rq->cfs;
9283 se->cfs_rq = parent->my_q;
9286 se->load.weight = tg->shares;
9287 se->load.inv_weight = 0;
9288 se->parent = parent;
9292 #ifdef CONFIG_RT_GROUP_SCHED
9293 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9294 struct sched_rt_entity *rt_se, int cpu, int add,
9295 struct sched_rt_entity *parent)
9297 struct rq *rq = cpu_rq(cpu);
9299 tg->rt_rq[cpu] = rt_rq;
9300 init_rt_rq(rt_rq, rq);
9302 rt_rq->rt_se = rt_se;
9303 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9305 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9307 tg->rt_se[cpu] = rt_se;
9312 rt_se->rt_rq = &rq->rt;
9314 rt_se->rt_rq = parent->my_q;
9316 rt_se->my_q = rt_rq;
9317 rt_se->parent = parent;
9318 INIT_LIST_HEAD(&rt_se->run_list);
9322 void __init sched_init(void)
9325 unsigned long alloc_size = 0, ptr;
9327 #ifdef CONFIG_FAIR_GROUP_SCHED
9328 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9330 #ifdef CONFIG_RT_GROUP_SCHED
9331 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9333 #ifdef CONFIG_USER_SCHED
9336 #ifdef CONFIG_CPUMASK_OFFSTACK
9337 alloc_size += num_possible_cpus() * cpumask_size();
9340 * As sched_init() is called before page_alloc is setup,
9341 * we use alloc_bootmem().
9344 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9346 #ifdef CONFIG_FAIR_GROUP_SCHED
9347 init_task_group.se = (struct sched_entity **)ptr;
9348 ptr += nr_cpu_ids * sizeof(void **);
9350 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9351 ptr += nr_cpu_ids * sizeof(void **);
9353 #ifdef CONFIG_USER_SCHED
9354 root_task_group.se = (struct sched_entity **)ptr;
9355 ptr += nr_cpu_ids * sizeof(void **);
9357 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9358 ptr += nr_cpu_ids * sizeof(void **);
9359 #endif /* CONFIG_USER_SCHED */
9360 #endif /* CONFIG_FAIR_GROUP_SCHED */
9361 #ifdef CONFIG_RT_GROUP_SCHED
9362 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9363 ptr += nr_cpu_ids * sizeof(void **);
9365 init_task_group.rt_rq = (struct rt_rq **)ptr;
9366 ptr += nr_cpu_ids * sizeof(void **);
9368 #ifdef CONFIG_USER_SCHED
9369 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9370 ptr += nr_cpu_ids * sizeof(void **);
9372 root_task_group.rt_rq = (struct rt_rq **)ptr;
9373 ptr += nr_cpu_ids * sizeof(void **);
9374 #endif /* CONFIG_USER_SCHED */
9375 #endif /* CONFIG_RT_GROUP_SCHED */
9376 #ifdef CONFIG_CPUMASK_OFFSTACK
9377 for_each_possible_cpu(i) {
9378 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9379 ptr += cpumask_size();
9381 #endif /* CONFIG_CPUMASK_OFFSTACK */
9385 init_defrootdomain();
9388 init_rt_bandwidth(&def_rt_bandwidth,
9389 global_rt_period(), global_rt_runtime());
9391 #ifdef CONFIG_RT_GROUP_SCHED
9392 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9393 global_rt_period(), global_rt_runtime());
9394 #ifdef CONFIG_USER_SCHED
9395 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9396 global_rt_period(), RUNTIME_INF);
9397 #endif /* CONFIG_USER_SCHED */
9398 #endif /* CONFIG_RT_GROUP_SCHED */
9400 #ifdef CONFIG_GROUP_SCHED
9401 list_add(&init_task_group.list, &task_groups);
9402 INIT_LIST_HEAD(&init_task_group.children);
9404 #ifdef CONFIG_USER_SCHED
9405 INIT_LIST_HEAD(&root_task_group.children);
9406 init_task_group.parent = &root_task_group;
9407 list_add(&init_task_group.siblings, &root_task_group.children);
9408 #endif /* CONFIG_USER_SCHED */
9409 #endif /* CONFIG_GROUP_SCHED */
9411 for_each_possible_cpu(i) {
9415 spin_lock_init(&rq->lock);
9417 rq->calc_load_active = 0;
9418 rq->calc_load_update = jiffies + LOAD_FREQ;
9419 init_cfs_rq(&rq->cfs, rq);
9420 init_rt_rq(&rq->rt, rq);
9421 #ifdef CONFIG_FAIR_GROUP_SCHED
9422 init_task_group.shares = init_task_group_load;
9423 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9424 #ifdef CONFIG_CGROUP_SCHED
9426 * How much cpu bandwidth does init_task_group get?
9428 * In case of task-groups formed thr' the cgroup filesystem, it
9429 * gets 100% of the cpu resources in the system. This overall
9430 * system cpu resource is divided among the tasks of
9431 * init_task_group and its child task-groups in a fair manner,
9432 * based on each entity's (task or task-group's) weight
9433 * (se->load.weight).
9435 * In other words, if init_task_group has 10 tasks of weight
9436 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9437 * then A0's share of the cpu resource is:
9439 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9441 * We achieve this by letting init_task_group's tasks sit
9442 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9444 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9445 #elif defined CONFIG_USER_SCHED
9446 root_task_group.shares = NICE_0_LOAD;
9447 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9449 * In case of task-groups formed thr' the user id of tasks,
9450 * init_task_group represents tasks belonging to root user.
9451 * Hence it forms a sibling of all subsequent groups formed.
9452 * In this case, init_task_group gets only a fraction of overall
9453 * system cpu resource, based on the weight assigned to root
9454 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9455 * by letting tasks of init_task_group sit in a separate cfs_rq
9456 * (init_tg_cfs_rq) and having one entity represent this group of
9457 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9459 init_tg_cfs_entry(&init_task_group,
9460 &per_cpu(init_tg_cfs_rq, i),
9461 &per_cpu(init_sched_entity, i), i, 1,
9462 root_task_group.se[i]);
9465 #endif /* CONFIG_FAIR_GROUP_SCHED */
9467 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9468 #ifdef CONFIG_RT_GROUP_SCHED
9469 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9470 #ifdef CONFIG_CGROUP_SCHED
9471 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9472 #elif defined CONFIG_USER_SCHED
9473 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9474 init_tg_rt_entry(&init_task_group,
9475 &per_cpu(init_rt_rq, i),
9476 &per_cpu(init_sched_rt_entity, i), i, 1,
9477 root_task_group.rt_se[i]);
9481 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9482 rq->cpu_load[j] = 0;
9486 rq->post_schedule = 0;
9487 rq->active_balance = 0;
9488 rq->next_balance = jiffies;
9492 rq->migration_thread = NULL;
9493 INIT_LIST_HEAD(&rq->migration_queue);
9494 rq_attach_root(rq, &def_root_domain);
9497 atomic_set(&rq->nr_iowait, 0);
9500 set_load_weight(&init_task);
9502 #ifdef CONFIG_PREEMPT_NOTIFIERS
9503 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9507 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9510 #ifdef CONFIG_RT_MUTEXES
9511 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9515 * The boot idle thread does lazy MMU switching as well:
9517 atomic_inc(&init_mm.mm_count);
9518 enter_lazy_tlb(&init_mm, current);
9521 * Make us the idle thread. Technically, schedule() should not be
9522 * called from this thread, however somewhere below it might be,
9523 * but because we are the idle thread, we just pick up running again
9524 * when this runqueue becomes "idle".
9526 init_idle(current, smp_processor_id());
9528 calc_load_update = jiffies + LOAD_FREQ;
9531 * During early bootup we pretend to be a normal task:
9533 current->sched_class = &fair_sched_class;
9535 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9536 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9539 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9540 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9542 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9547 scheduler_running = 1;
9550 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9551 static inline int preempt_count_equals(int preempt_offset)
9553 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9555 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9558 void __might_sleep(char *file, int line, int preempt_offset)
9561 static unsigned long prev_jiffy; /* ratelimiting */
9563 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9564 system_state != SYSTEM_RUNNING || oops_in_progress)
9566 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9568 prev_jiffy = jiffies;
9571 "BUG: sleeping function called from invalid context at %s:%d\n",
9574 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9575 in_atomic(), irqs_disabled(),
9576 current->pid, current->comm);
9578 debug_show_held_locks(current);
9579 if (irqs_disabled())
9580 print_irqtrace_events(current);
9584 EXPORT_SYMBOL(__might_sleep);
9587 #ifdef CONFIG_MAGIC_SYSRQ
9588 static void normalize_task(struct rq *rq, struct task_struct *p)
9592 update_rq_clock(rq);
9593 on_rq = p->se.on_rq;
9595 deactivate_task(rq, p, 0);
9596 __setscheduler(rq, p, SCHED_NORMAL, 0);
9598 activate_task(rq, p, 0);
9599 resched_task(rq->curr);
9603 void normalize_rt_tasks(void)
9605 struct task_struct *g, *p;
9606 unsigned long flags;
9609 read_lock_irqsave(&tasklist_lock, flags);
9610 do_each_thread(g, p) {
9612 * Only normalize user tasks:
9617 p->se.exec_start = 0;
9618 #ifdef CONFIG_SCHEDSTATS
9619 p->se.wait_start = 0;
9620 p->se.sleep_start = 0;
9621 p->se.block_start = 0;
9626 * Renice negative nice level userspace
9629 if (TASK_NICE(p) < 0 && p->mm)
9630 set_user_nice(p, 0);
9634 spin_lock(&p->pi_lock);
9635 rq = __task_rq_lock(p);
9637 normalize_task(rq, p);
9639 __task_rq_unlock(rq);
9640 spin_unlock(&p->pi_lock);
9641 } while_each_thread(g, p);
9643 read_unlock_irqrestore(&tasklist_lock, flags);
9646 #endif /* CONFIG_MAGIC_SYSRQ */
9650 * These functions are only useful for the IA64 MCA handling.
9652 * They can only be called when the whole system has been
9653 * stopped - every CPU needs to be quiescent, and no scheduling
9654 * activity can take place. Using them for anything else would
9655 * be a serious bug, and as a result, they aren't even visible
9656 * under any other configuration.
9660 * curr_task - return the current task for a given cpu.
9661 * @cpu: the processor in question.
9663 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9665 struct task_struct *curr_task(int cpu)
9667 return cpu_curr(cpu);
9671 * set_curr_task - set the current task for a given cpu.
9672 * @cpu: the processor in question.
9673 * @p: the task pointer to set.
9675 * Description: This function must only be used when non-maskable interrupts
9676 * are serviced on a separate stack. It allows the architecture to switch the
9677 * notion of the current task on a cpu in a non-blocking manner. This function
9678 * must be called with all CPU's synchronized, and interrupts disabled, the
9679 * and caller must save the original value of the current task (see
9680 * curr_task() above) and restore that value before reenabling interrupts and
9681 * re-starting the system.
9683 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9685 void set_curr_task(int cpu, struct task_struct *p)
9692 #ifdef CONFIG_FAIR_GROUP_SCHED
9693 static void free_fair_sched_group(struct task_group *tg)
9697 for_each_possible_cpu(i) {
9699 kfree(tg->cfs_rq[i]);
9709 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9711 struct cfs_rq *cfs_rq;
9712 struct sched_entity *se;
9716 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9719 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9723 tg->shares = NICE_0_LOAD;
9725 for_each_possible_cpu(i) {
9728 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9729 GFP_KERNEL, cpu_to_node(i));
9733 se = kzalloc_node(sizeof(struct sched_entity),
9734 GFP_KERNEL, cpu_to_node(i));
9738 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9747 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9749 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9750 &cpu_rq(cpu)->leaf_cfs_rq_list);
9753 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9755 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9757 #else /* !CONFG_FAIR_GROUP_SCHED */
9758 static inline void free_fair_sched_group(struct task_group *tg)
9763 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9768 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9772 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9775 #endif /* CONFIG_FAIR_GROUP_SCHED */
9777 #ifdef CONFIG_RT_GROUP_SCHED
9778 static void free_rt_sched_group(struct task_group *tg)
9782 destroy_rt_bandwidth(&tg->rt_bandwidth);
9784 for_each_possible_cpu(i) {
9786 kfree(tg->rt_rq[i]);
9788 kfree(tg->rt_se[i]);
9796 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9798 struct rt_rq *rt_rq;
9799 struct sched_rt_entity *rt_se;
9803 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9806 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9810 init_rt_bandwidth(&tg->rt_bandwidth,
9811 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9813 for_each_possible_cpu(i) {
9816 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9817 GFP_KERNEL, cpu_to_node(i));
9821 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9822 GFP_KERNEL, cpu_to_node(i));
9826 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9835 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9837 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9838 &cpu_rq(cpu)->leaf_rt_rq_list);
9841 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9843 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9845 #else /* !CONFIG_RT_GROUP_SCHED */
9846 static inline void free_rt_sched_group(struct task_group *tg)
9851 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9856 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9860 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9863 #endif /* CONFIG_RT_GROUP_SCHED */
9865 #ifdef CONFIG_GROUP_SCHED
9866 static void free_sched_group(struct task_group *tg)
9868 free_fair_sched_group(tg);
9869 free_rt_sched_group(tg);
9873 /* allocate runqueue etc for a new task group */
9874 struct task_group *sched_create_group(struct task_group *parent)
9876 struct task_group *tg;
9877 unsigned long flags;
9880 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9882 return ERR_PTR(-ENOMEM);
9884 if (!alloc_fair_sched_group(tg, parent))
9887 if (!alloc_rt_sched_group(tg, parent))
9890 spin_lock_irqsave(&task_group_lock, flags);
9891 for_each_possible_cpu(i) {
9892 register_fair_sched_group(tg, i);
9893 register_rt_sched_group(tg, i);
9895 list_add_rcu(&tg->list, &task_groups);
9897 WARN_ON(!parent); /* root should already exist */
9899 tg->parent = parent;
9900 INIT_LIST_HEAD(&tg->children);
9901 list_add_rcu(&tg->siblings, &parent->children);
9902 spin_unlock_irqrestore(&task_group_lock, flags);
9907 free_sched_group(tg);
9908 return ERR_PTR(-ENOMEM);
9911 /* rcu callback to free various structures associated with a task group */
9912 static void free_sched_group_rcu(struct rcu_head *rhp)
9914 /* now it should be safe to free those cfs_rqs */
9915 free_sched_group(container_of(rhp, struct task_group, rcu));
9918 /* Destroy runqueue etc associated with a task group */
9919 void sched_destroy_group(struct task_group *tg)
9921 unsigned long flags;
9924 spin_lock_irqsave(&task_group_lock, flags);
9925 for_each_possible_cpu(i) {
9926 unregister_fair_sched_group(tg, i);
9927 unregister_rt_sched_group(tg, i);
9929 list_del_rcu(&tg->list);
9930 list_del_rcu(&tg->siblings);
9931 spin_unlock_irqrestore(&task_group_lock, flags);
9933 /* wait for possible concurrent references to cfs_rqs complete */
9934 call_rcu(&tg->rcu, free_sched_group_rcu);
9937 /* change task's runqueue when it moves between groups.
9938 * The caller of this function should have put the task in its new group
9939 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9940 * reflect its new group.
9942 void sched_move_task(struct task_struct *tsk)
9945 unsigned long flags;
9948 rq = task_rq_lock(tsk, &flags);
9950 update_rq_clock(rq);
9952 running = task_current(rq, tsk);
9953 on_rq = tsk->se.on_rq;
9956 dequeue_task(rq, tsk, 0);
9957 if (unlikely(running))
9958 tsk->sched_class->put_prev_task(rq, tsk);
9960 set_task_rq(tsk, task_cpu(tsk));
9962 #ifdef CONFIG_FAIR_GROUP_SCHED
9963 if (tsk->sched_class->moved_group)
9964 tsk->sched_class->moved_group(tsk);
9967 if (unlikely(running))
9968 tsk->sched_class->set_curr_task(rq);
9970 enqueue_task(rq, tsk, 0);
9972 task_rq_unlock(rq, &flags);
9974 #endif /* CONFIG_GROUP_SCHED */
9976 #ifdef CONFIG_FAIR_GROUP_SCHED
9977 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9979 struct cfs_rq *cfs_rq = se->cfs_rq;
9984 dequeue_entity(cfs_rq, se, 0);
9986 se->load.weight = shares;
9987 se->load.inv_weight = 0;
9990 enqueue_entity(cfs_rq, se, 0);
9993 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9995 struct cfs_rq *cfs_rq = se->cfs_rq;
9996 struct rq *rq = cfs_rq->rq;
9997 unsigned long flags;
9999 spin_lock_irqsave(&rq->lock, flags);
10000 __set_se_shares(se, shares);
10001 spin_unlock_irqrestore(&rq->lock, flags);
10004 static DEFINE_MUTEX(shares_mutex);
10006 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10009 unsigned long flags;
10012 * We can't change the weight of the root cgroup.
10017 if (shares < MIN_SHARES)
10018 shares = MIN_SHARES;
10019 else if (shares > MAX_SHARES)
10020 shares = MAX_SHARES;
10022 mutex_lock(&shares_mutex);
10023 if (tg->shares == shares)
10026 spin_lock_irqsave(&task_group_lock, flags);
10027 for_each_possible_cpu(i)
10028 unregister_fair_sched_group(tg, i);
10029 list_del_rcu(&tg->siblings);
10030 spin_unlock_irqrestore(&task_group_lock, flags);
10032 /* wait for any ongoing reference to this group to finish */
10033 synchronize_sched();
10036 * Now we are free to modify the group's share on each cpu
10037 * w/o tripping rebalance_share or load_balance_fair.
10039 tg->shares = shares;
10040 for_each_possible_cpu(i) {
10042 * force a rebalance
10044 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10045 set_se_shares(tg->se[i], shares);
10049 * Enable load balance activity on this group, by inserting it back on
10050 * each cpu's rq->leaf_cfs_rq_list.
10052 spin_lock_irqsave(&task_group_lock, flags);
10053 for_each_possible_cpu(i)
10054 register_fair_sched_group(tg, i);
10055 list_add_rcu(&tg->siblings, &tg->parent->children);
10056 spin_unlock_irqrestore(&task_group_lock, flags);
10058 mutex_unlock(&shares_mutex);
10062 unsigned long sched_group_shares(struct task_group *tg)
10068 #ifdef CONFIG_RT_GROUP_SCHED
10070 * Ensure that the real time constraints are schedulable.
10072 static DEFINE_MUTEX(rt_constraints_mutex);
10074 static unsigned long to_ratio(u64 period, u64 runtime)
10076 if (runtime == RUNTIME_INF)
10079 return div64_u64(runtime << 20, period);
10082 /* Must be called with tasklist_lock held */
10083 static inline int tg_has_rt_tasks(struct task_group *tg)
10085 struct task_struct *g, *p;
10087 do_each_thread(g, p) {
10088 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10090 } while_each_thread(g, p);
10095 struct rt_schedulable_data {
10096 struct task_group *tg;
10101 static int tg_schedulable(struct task_group *tg, void *data)
10103 struct rt_schedulable_data *d = data;
10104 struct task_group *child;
10105 unsigned long total, sum = 0;
10106 u64 period, runtime;
10108 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10109 runtime = tg->rt_bandwidth.rt_runtime;
10112 period = d->rt_period;
10113 runtime = d->rt_runtime;
10116 #ifdef CONFIG_USER_SCHED
10117 if (tg == &root_task_group) {
10118 period = global_rt_period();
10119 runtime = global_rt_runtime();
10124 * Cannot have more runtime than the period.
10126 if (runtime > period && runtime != RUNTIME_INF)
10130 * Ensure we don't starve existing RT tasks.
10132 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10135 total = to_ratio(period, runtime);
10138 * Nobody can have more than the global setting allows.
10140 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10144 * The sum of our children's runtime should not exceed our own.
10146 list_for_each_entry_rcu(child, &tg->children, siblings) {
10147 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10148 runtime = child->rt_bandwidth.rt_runtime;
10150 if (child == d->tg) {
10151 period = d->rt_period;
10152 runtime = d->rt_runtime;
10155 sum += to_ratio(period, runtime);
10164 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10166 struct rt_schedulable_data data = {
10168 .rt_period = period,
10169 .rt_runtime = runtime,
10172 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10175 static int tg_set_bandwidth(struct task_group *tg,
10176 u64 rt_period, u64 rt_runtime)
10180 mutex_lock(&rt_constraints_mutex);
10181 read_lock(&tasklist_lock);
10182 err = __rt_schedulable(tg, rt_period, rt_runtime);
10186 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10187 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10188 tg->rt_bandwidth.rt_runtime = rt_runtime;
10190 for_each_possible_cpu(i) {
10191 struct rt_rq *rt_rq = tg->rt_rq[i];
10193 spin_lock(&rt_rq->rt_runtime_lock);
10194 rt_rq->rt_runtime = rt_runtime;
10195 spin_unlock(&rt_rq->rt_runtime_lock);
10197 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10199 read_unlock(&tasklist_lock);
10200 mutex_unlock(&rt_constraints_mutex);
10205 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10207 u64 rt_runtime, rt_period;
10209 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10210 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10211 if (rt_runtime_us < 0)
10212 rt_runtime = RUNTIME_INF;
10214 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10217 long sched_group_rt_runtime(struct task_group *tg)
10221 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10224 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10225 do_div(rt_runtime_us, NSEC_PER_USEC);
10226 return rt_runtime_us;
10229 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10231 u64 rt_runtime, rt_period;
10233 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10234 rt_runtime = tg->rt_bandwidth.rt_runtime;
10236 if (rt_period == 0)
10239 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10242 long sched_group_rt_period(struct task_group *tg)
10246 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10247 do_div(rt_period_us, NSEC_PER_USEC);
10248 return rt_period_us;
10251 static int sched_rt_global_constraints(void)
10253 u64 runtime, period;
10256 if (sysctl_sched_rt_period <= 0)
10259 runtime = global_rt_runtime();
10260 period = global_rt_period();
10263 * Sanity check on the sysctl variables.
10265 if (runtime > period && runtime != RUNTIME_INF)
10268 mutex_lock(&rt_constraints_mutex);
10269 read_lock(&tasklist_lock);
10270 ret = __rt_schedulable(NULL, 0, 0);
10271 read_unlock(&tasklist_lock);
10272 mutex_unlock(&rt_constraints_mutex);
10277 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10279 /* Don't accept realtime tasks when there is no way for them to run */
10280 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10286 #else /* !CONFIG_RT_GROUP_SCHED */
10287 static int sched_rt_global_constraints(void)
10289 unsigned long flags;
10292 if (sysctl_sched_rt_period <= 0)
10296 * There's always some RT tasks in the root group
10297 * -- migration, kstopmachine etc..
10299 if (sysctl_sched_rt_runtime == 0)
10302 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10303 for_each_possible_cpu(i) {
10304 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10306 spin_lock(&rt_rq->rt_runtime_lock);
10307 rt_rq->rt_runtime = global_rt_runtime();
10308 spin_unlock(&rt_rq->rt_runtime_lock);
10310 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10314 #endif /* CONFIG_RT_GROUP_SCHED */
10316 int sched_rt_handler(struct ctl_table *table, int write,
10317 void __user *buffer, size_t *lenp,
10321 int old_period, old_runtime;
10322 static DEFINE_MUTEX(mutex);
10324 mutex_lock(&mutex);
10325 old_period = sysctl_sched_rt_period;
10326 old_runtime = sysctl_sched_rt_runtime;
10328 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10330 if (!ret && write) {
10331 ret = sched_rt_global_constraints();
10333 sysctl_sched_rt_period = old_period;
10334 sysctl_sched_rt_runtime = old_runtime;
10336 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10337 def_rt_bandwidth.rt_period =
10338 ns_to_ktime(global_rt_period());
10341 mutex_unlock(&mutex);
10346 #ifdef CONFIG_CGROUP_SCHED
10348 /* return corresponding task_group object of a cgroup */
10349 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10351 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10352 struct task_group, css);
10355 static struct cgroup_subsys_state *
10356 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10358 struct task_group *tg, *parent;
10360 if (!cgrp->parent) {
10361 /* This is early initialization for the top cgroup */
10362 return &init_task_group.css;
10365 parent = cgroup_tg(cgrp->parent);
10366 tg = sched_create_group(parent);
10368 return ERR_PTR(-ENOMEM);
10374 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10376 struct task_group *tg = cgroup_tg(cgrp);
10378 sched_destroy_group(tg);
10382 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10384 #ifdef CONFIG_RT_GROUP_SCHED
10385 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10388 /* We don't support RT-tasks being in separate groups */
10389 if (tsk->sched_class != &fair_sched_class)
10396 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10397 struct task_struct *tsk, bool threadgroup)
10399 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10403 struct task_struct *c;
10405 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10406 retval = cpu_cgroup_can_attach_task(cgrp, c);
10418 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10419 struct cgroup *old_cont, struct task_struct *tsk,
10422 sched_move_task(tsk);
10424 struct task_struct *c;
10426 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10427 sched_move_task(c);
10433 #ifdef CONFIG_FAIR_GROUP_SCHED
10434 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10437 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10440 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10442 struct task_group *tg = cgroup_tg(cgrp);
10444 return (u64) tg->shares;
10446 #endif /* CONFIG_FAIR_GROUP_SCHED */
10448 #ifdef CONFIG_RT_GROUP_SCHED
10449 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10452 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10455 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10457 return sched_group_rt_runtime(cgroup_tg(cgrp));
10460 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10463 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10466 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10468 return sched_group_rt_period(cgroup_tg(cgrp));
10470 #endif /* CONFIG_RT_GROUP_SCHED */
10472 static struct cftype cpu_files[] = {
10473 #ifdef CONFIG_FAIR_GROUP_SCHED
10476 .read_u64 = cpu_shares_read_u64,
10477 .write_u64 = cpu_shares_write_u64,
10480 #ifdef CONFIG_RT_GROUP_SCHED
10482 .name = "rt_runtime_us",
10483 .read_s64 = cpu_rt_runtime_read,
10484 .write_s64 = cpu_rt_runtime_write,
10487 .name = "rt_period_us",
10488 .read_u64 = cpu_rt_period_read_uint,
10489 .write_u64 = cpu_rt_period_write_uint,
10494 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10496 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10499 struct cgroup_subsys cpu_cgroup_subsys = {
10501 .create = cpu_cgroup_create,
10502 .destroy = cpu_cgroup_destroy,
10503 .can_attach = cpu_cgroup_can_attach,
10504 .attach = cpu_cgroup_attach,
10505 .populate = cpu_cgroup_populate,
10506 .subsys_id = cpu_cgroup_subsys_id,
10510 #endif /* CONFIG_CGROUP_SCHED */
10512 #ifdef CONFIG_CGROUP_CPUACCT
10515 * CPU accounting code for task groups.
10517 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10518 * (balbir@in.ibm.com).
10521 /* track cpu usage of a group of tasks and its child groups */
10523 struct cgroup_subsys_state css;
10524 /* cpuusage holds pointer to a u64-type object on every cpu */
10526 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10527 struct cpuacct *parent;
10530 struct cgroup_subsys cpuacct_subsys;
10532 /* return cpu accounting group corresponding to this container */
10533 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10535 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10536 struct cpuacct, css);
10539 /* return cpu accounting group to which this task belongs */
10540 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10542 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10543 struct cpuacct, css);
10546 /* create a new cpu accounting group */
10547 static struct cgroup_subsys_state *cpuacct_create(
10548 struct cgroup_subsys *ss, struct cgroup *cgrp)
10550 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10556 ca->cpuusage = alloc_percpu(u64);
10560 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10561 if (percpu_counter_init(&ca->cpustat[i], 0))
10562 goto out_free_counters;
10565 ca->parent = cgroup_ca(cgrp->parent);
10571 percpu_counter_destroy(&ca->cpustat[i]);
10572 free_percpu(ca->cpuusage);
10576 return ERR_PTR(-ENOMEM);
10579 /* destroy an existing cpu accounting group */
10581 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10583 struct cpuacct *ca = cgroup_ca(cgrp);
10586 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10587 percpu_counter_destroy(&ca->cpustat[i]);
10588 free_percpu(ca->cpuusage);
10592 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10594 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10597 #ifndef CONFIG_64BIT
10599 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10601 spin_lock_irq(&cpu_rq(cpu)->lock);
10603 spin_unlock_irq(&cpu_rq(cpu)->lock);
10611 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10613 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10615 #ifndef CONFIG_64BIT
10617 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10619 spin_lock_irq(&cpu_rq(cpu)->lock);
10621 spin_unlock_irq(&cpu_rq(cpu)->lock);
10627 /* return total cpu usage (in nanoseconds) of a group */
10628 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10630 struct cpuacct *ca = cgroup_ca(cgrp);
10631 u64 totalcpuusage = 0;
10634 for_each_present_cpu(i)
10635 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10637 return totalcpuusage;
10640 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10643 struct cpuacct *ca = cgroup_ca(cgrp);
10652 for_each_present_cpu(i)
10653 cpuacct_cpuusage_write(ca, i, 0);
10659 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10660 struct seq_file *m)
10662 struct cpuacct *ca = cgroup_ca(cgroup);
10666 for_each_present_cpu(i) {
10667 percpu = cpuacct_cpuusage_read(ca, i);
10668 seq_printf(m, "%llu ", (unsigned long long) percpu);
10670 seq_printf(m, "\n");
10674 static const char *cpuacct_stat_desc[] = {
10675 [CPUACCT_STAT_USER] = "user",
10676 [CPUACCT_STAT_SYSTEM] = "system",
10679 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10680 struct cgroup_map_cb *cb)
10682 struct cpuacct *ca = cgroup_ca(cgrp);
10685 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10686 s64 val = percpu_counter_read(&ca->cpustat[i]);
10687 val = cputime64_to_clock_t(val);
10688 cb->fill(cb, cpuacct_stat_desc[i], val);
10693 static struct cftype files[] = {
10696 .read_u64 = cpuusage_read,
10697 .write_u64 = cpuusage_write,
10700 .name = "usage_percpu",
10701 .read_seq_string = cpuacct_percpu_seq_read,
10705 .read_map = cpuacct_stats_show,
10709 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10711 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10715 * charge this task's execution time to its accounting group.
10717 * called with rq->lock held.
10719 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10721 struct cpuacct *ca;
10724 if (unlikely(!cpuacct_subsys.active))
10727 cpu = task_cpu(tsk);
10733 for (; ca; ca = ca->parent) {
10734 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10735 *cpuusage += cputime;
10742 * Charge the system/user time to the task's accounting group.
10744 static void cpuacct_update_stats(struct task_struct *tsk,
10745 enum cpuacct_stat_index idx, cputime_t val)
10747 struct cpuacct *ca;
10749 if (unlikely(!cpuacct_subsys.active))
10756 percpu_counter_add(&ca->cpustat[idx], val);
10762 struct cgroup_subsys cpuacct_subsys = {
10764 .create = cpuacct_create,
10765 .destroy = cpuacct_destroy,
10766 .populate = cpuacct_populate,
10767 .subsys_id = cpuacct_subsys_id,
10769 #endif /* CONFIG_CGROUP_CPUACCT */
10773 int rcu_expedited_torture_stats(char *page)
10777 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10779 void synchronize_sched_expedited(void)
10782 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10784 #else /* #ifndef CONFIG_SMP */
10786 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10787 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10789 #define RCU_EXPEDITED_STATE_POST -2
10790 #define RCU_EXPEDITED_STATE_IDLE -1
10792 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10794 int rcu_expedited_torture_stats(char *page)
10799 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10800 for_each_online_cpu(cpu) {
10801 cnt += sprintf(&page[cnt], " %d:%d",
10802 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10804 cnt += sprintf(&page[cnt], "\n");
10807 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10809 static long synchronize_sched_expedited_count;
10812 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10813 * approach to force grace period to end quickly. This consumes
10814 * significant time on all CPUs, and is thus not recommended for
10815 * any sort of common-case code.
10817 * Note that it is illegal to call this function while holding any
10818 * lock that is acquired by a CPU-hotplug notifier. Failing to
10819 * observe this restriction will result in deadlock.
10821 void synchronize_sched_expedited(void)
10824 unsigned long flags;
10825 bool need_full_sync = 0;
10827 struct migration_req *req;
10831 smp_mb(); /* ensure prior mod happens before capturing snap. */
10832 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10834 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10836 if (trycount++ < 10)
10837 udelay(trycount * num_online_cpus());
10839 synchronize_sched();
10842 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10843 smp_mb(); /* ensure test happens before caller kfree */
10848 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10849 for_each_online_cpu(cpu) {
10851 req = &per_cpu(rcu_migration_req, cpu);
10852 init_completion(&req->done);
10854 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10855 spin_lock_irqsave(&rq->lock, flags);
10856 list_add(&req->list, &rq->migration_queue);
10857 spin_unlock_irqrestore(&rq->lock, flags);
10858 wake_up_process(rq->migration_thread);
10860 for_each_online_cpu(cpu) {
10861 rcu_expedited_state = cpu;
10862 req = &per_cpu(rcu_migration_req, cpu);
10864 wait_for_completion(&req->done);
10865 spin_lock_irqsave(&rq->lock, flags);
10866 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10867 need_full_sync = 1;
10868 req->dest_cpu = RCU_MIGRATION_IDLE;
10869 spin_unlock_irqrestore(&rq->lock, flags);
10871 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10872 mutex_unlock(&rcu_sched_expedited_mutex);
10874 if (need_full_sync)
10875 synchronize_sched();
10877 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10879 #endif /* #else #ifndef CONFIG_SMP */