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 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 static __read_mostly unsigned long *update_shares_data;
1568 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1571 * Calculate and set the cpu's group shares.
1573 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1574 unsigned long sd_shares,
1575 unsigned long sd_rq_weight,
1576 unsigned long *usd_rq_weight)
1578 unsigned long shares, rq_weight;
1581 rq_weight = usd_rq_weight[cpu];
1584 rq_weight = NICE_0_LOAD;
1588 * \Sum_j shares_j * rq_weight_i
1589 * shares_i = -----------------------------
1590 * \Sum_j rq_weight_j
1592 shares = (sd_shares * rq_weight) / sd_rq_weight;
1593 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1595 if (abs(shares - tg->se[cpu]->load.weight) >
1596 sysctl_sched_shares_thresh) {
1597 struct rq *rq = cpu_rq(cpu);
1598 unsigned long flags;
1600 spin_lock_irqsave(&rq->lock, flags);
1601 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1602 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1603 __set_se_shares(tg->se[cpu], shares);
1604 spin_unlock_irqrestore(&rq->lock, flags);
1609 * Re-compute the task group their per cpu shares over the given domain.
1610 * This needs to be done in a bottom-up fashion because the rq weight of a
1611 * parent group depends on the shares of its child groups.
1613 static int tg_shares_up(struct task_group *tg, void *data)
1615 unsigned long weight, rq_weight = 0, shares = 0;
1616 unsigned long *usd_rq_weight;
1617 struct sched_domain *sd = data;
1618 unsigned long flags;
1624 local_irq_save(flags);
1625 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1627 for_each_cpu(i, sched_domain_span(sd)) {
1628 weight = tg->cfs_rq[i]->load.weight;
1629 usd_rq_weight[i] = weight;
1632 * If there are currently no tasks on the cpu pretend there
1633 * is one of average load so that when a new task gets to
1634 * run here it will not get delayed by group starvation.
1637 weight = NICE_0_LOAD;
1639 rq_weight += weight;
1640 shares += tg->cfs_rq[i]->shares;
1643 if ((!shares && rq_weight) || shares > tg->shares)
1644 shares = tg->shares;
1646 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1647 shares = tg->shares;
1649 for_each_cpu(i, sched_domain_span(sd))
1650 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1652 local_irq_restore(flags);
1658 * Compute the cpu's hierarchical load factor for each task group.
1659 * This needs to be done in a top-down fashion because the load of a child
1660 * group is a fraction of its parents load.
1662 static int tg_load_down(struct task_group *tg, void *data)
1665 long cpu = (long)data;
1668 load = cpu_rq(cpu)->load.weight;
1670 load = tg->parent->cfs_rq[cpu]->h_load;
1671 load *= tg->cfs_rq[cpu]->shares;
1672 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1675 tg->cfs_rq[cpu]->h_load = load;
1680 static void update_shares(struct sched_domain *sd)
1685 if (root_task_group_empty())
1688 now = cpu_clock(raw_smp_processor_id());
1689 elapsed = now - sd->last_update;
1691 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1692 sd->last_update = now;
1693 walk_tg_tree(tg_nop, tg_shares_up, sd);
1697 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1699 if (root_task_group_empty())
1702 spin_unlock(&rq->lock);
1704 spin_lock(&rq->lock);
1707 static void update_h_load(long cpu)
1709 if (root_task_group_empty())
1712 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1717 static inline void update_shares(struct sched_domain *sd)
1721 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1727 #ifdef CONFIG_PREEMPT
1729 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1732 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1733 * way at the expense of forcing extra atomic operations in all
1734 * invocations. This assures that the double_lock is acquired using the
1735 * same underlying policy as the spinlock_t on this architecture, which
1736 * reduces latency compared to the unfair variant below. However, it
1737 * also adds more overhead and therefore may reduce throughput.
1739 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1740 __releases(this_rq->lock)
1741 __acquires(busiest->lock)
1742 __acquires(this_rq->lock)
1744 spin_unlock(&this_rq->lock);
1745 double_rq_lock(this_rq, busiest);
1752 * Unfair double_lock_balance: Optimizes throughput at the expense of
1753 * latency by eliminating extra atomic operations when the locks are
1754 * already in proper order on entry. This favors lower cpu-ids and will
1755 * grant the double lock to lower cpus over higher ids under contention,
1756 * regardless of entry order into the function.
1758 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1759 __releases(this_rq->lock)
1760 __acquires(busiest->lock)
1761 __acquires(this_rq->lock)
1765 if (unlikely(!spin_trylock(&busiest->lock))) {
1766 if (busiest < this_rq) {
1767 spin_unlock(&this_rq->lock);
1768 spin_lock(&busiest->lock);
1769 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1772 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1777 #endif /* CONFIG_PREEMPT */
1780 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1782 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1784 if (unlikely(!irqs_disabled())) {
1785 /* printk() doesn't work good under rq->lock */
1786 spin_unlock(&this_rq->lock);
1790 return _double_lock_balance(this_rq, busiest);
1793 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1794 __releases(busiest->lock)
1796 spin_unlock(&busiest->lock);
1797 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1801 #ifdef CONFIG_FAIR_GROUP_SCHED
1802 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1805 cfs_rq->shares = shares;
1810 static void calc_load_account_active(struct rq *this_rq);
1812 #include "sched_stats.h"
1813 #include "sched_idletask.c"
1814 #include "sched_fair.c"
1815 #include "sched_rt.c"
1816 #ifdef CONFIG_SCHED_DEBUG
1817 # include "sched_debug.c"
1820 #define sched_class_highest (&rt_sched_class)
1821 #define for_each_class(class) \
1822 for (class = sched_class_highest; class; class = class->next)
1824 static void inc_nr_running(struct rq *rq)
1829 static void dec_nr_running(struct rq *rq)
1834 static void set_load_weight(struct task_struct *p)
1836 if (task_has_rt_policy(p)) {
1837 p->se.load.weight = prio_to_weight[0] * 2;
1838 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1843 * SCHED_IDLE tasks get minimal weight:
1845 if (p->policy == SCHED_IDLE) {
1846 p->se.load.weight = WEIGHT_IDLEPRIO;
1847 p->se.load.inv_weight = WMULT_IDLEPRIO;
1851 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1852 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1855 static void update_avg(u64 *avg, u64 sample)
1857 s64 diff = sample - *avg;
1861 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1864 p->se.start_runtime = p->se.sum_exec_runtime;
1866 sched_info_queued(p);
1867 p->sched_class->enqueue_task(rq, p, wakeup);
1871 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1874 if (p->se.last_wakeup) {
1875 update_avg(&p->se.avg_overlap,
1876 p->se.sum_exec_runtime - p->se.last_wakeup);
1877 p->se.last_wakeup = 0;
1879 update_avg(&p->se.avg_wakeup,
1880 sysctl_sched_wakeup_granularity);
1884 sched_info_dequeued(p);
1885 p->sched_class->dequeue_task(rq, p, sleep);
1890 * __normal_prio - return the priority that is based on the static prio
1892 static inline int __normal_prio(struct task_struct *p)
1894 return p->static_prio;
1898 * Calculate the expected normal priority: i.e. priority
1899 * without taking RT-inheritance into account. Might be
1900 * boosted by interactivity modifiers. Changes upon fork,
1901 * setprio syscalls, and whenever the interactivity
1902 * estimator recalculates.
1904 static inline int normal_prio(struct task_struct *p)
1908 if (task_has_rt_policy(p))
1909 prio = MAX_RT_PRIO-1 - p->rt_priority;
1911 prio = __normal_prio(p);
1916 * Calculate the current priority, i.e. the priority
1917 * taken into account by the scheduler. This value might
1918 * be boosted by RT tasks, or might be boosted by
1919 * interactivity modifiers. Will be RT if the task got
1920 * RT-boosted. If not then it returns p->normal_prio.
1922 static int effective_prio(struct task_struct *p)
1924 p->normal_prio = normal_prio(p);
1926 * If we are RT tasks or we were boosted to RT priority,
1927 * keep the priority unchanged. Otherwise, update priority
1928 * to the normal priority:
1930 if (!rt_prio(p->prio))
1931 return p->normal_prio;
1936 * activate_task - move a task to the runqueue.
1938 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1940 if (task_contributes_to_load(p))
1941 rq->nr_uninterruptible--;
1943 enqueue_task(rq, p, wakeup);
1948 * deactivate_task - remove a task from the runqueue.
1950 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1952 if (task_contributes_to_load(p))
1953 rq->nr_uninterruptible++;
1955 dequeue_task(rq, p, sleep);
1960 * task_curr - is this task currently executing on a CPU?
1961 * @p: the task in question.
1963 inline int task_curr(const struct task_struct *p)
1965 return cpu_curr(task_cpu(p)) == p;
1968 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1970 set_task_rq(p, cpu);
1973 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1974 * successfuly executed on another CPU. We must ensure that updates of
1975 * per-task data have been completed by this moment.
1978 task_thread_info(p)->cpu = cpu;
1982 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1983 const struct sched_class *prev_class,
1984 int oldprio, int running)
1986 if (prev_class != p->sched_class) {
1987 if (prev_class->switched_from)
1988 prev_class->switched_from(rq, p, running);
1989 p->sched_class->switched_to(rq, p, running);
1991 p->sched_class->prio_changed(rq, p, oldprio, running);
1996 * Is this task likely cache-hot:
1999 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2004 * Buddy candidates are cache hot:
2006 if (sched_feat(CACHE_HOT_BUDDY) &&
2007 (&p->se == cfs_rq_of(&p->se)->next ||
2008 &p->se == cfs_rq_of(&p->se)->last))
2011 if (p->sched_class != &fair_sched_class)
2014 if (sysctl_sched_migration_cost == -1)
2016 if (sysctl_sched_migration_cost == 0)
2019 delta = now - p->se.exec_start;
2021 return delta < (s64)sysctl_sched_migration_cost;
2025 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2027 int old_cpu = task_cpu(p);
2028 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2029 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2030 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2033 clock_offset = old_rq->clock - new_rq->clock;
2035 trace_sched_migrate_task(p, new_cpu);
2037 #ifdef CONFIG_SCHEDSTATS
2038 if (p->se.wait_start)
2039 p->se.wait_start -= clock_offset;
2040 if (p->se.sleep_start)
2041 p->se.sleep_start -= clock_offset;
2042 if (p->se.block_start)
2043 p->se.block_start -= clock_offset;
2045 if (old_cpu != new_cpu) {
2046 p->se.nr_migrations++;
2047 new_rq->nr_migrations_in++;
2048 #ifdef CONFIG_SCHEDSTATS
2049 if (task_hot(p, old_rq->clock, NULL))
2050 schedstat_inc(p, se.nr_forced2_migrations);
2052 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2055 p->se.vruntime -= old_cfsrq->min_vruntime -
2056 new_cfsrq->min_vruntime;
2058 __set_task_cpu(p, new_cpu);
2061 struct migration_req {
2062 struct list_head list;
2064 struct task_struct *task;
2067 struct completion done;
2071 * The task's runqueue lock must be held.
2072 * Returns true if you have to wait for migration thread.
2075 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2077 struct rq *rq = task_rq(p);
2080 * If the task is not on a runqueue (and not running), then
2081 * it is sufficient to simply update the task's cpu field.
2083 if (!p->se.on_rq && !task_running(rq, p)) {
2084 set_task_cpu(p, dest_cpu);
2088 init_completion(&req->done);
2090 req->dest_cpu = dest_cpu;
2091 list_add(&req->list, &rq->migration_queue);
2097 * wait_task_context_switch - wait for a thread to complete at least one
2100 * @p must not be current.
2102 void wait_task_context_switch(struct task_struct *p)
2104 unsigned long nvcsw, nivcsw, flags;
2112 * The runqueue is assigned before the actual context
2113 * switch. We need to take the runqueue lock.
2115 * We could check initially without the lock but it is
2116 * very likely that we need to take the lock in every
2119 rq = task_rq_lock(p, &flags);
2120 running = task_running(rq, p);
2121 task_rq_unlock(rq, &flags);
2123 if (likely(!running))
2126 * The switch count is incremented before the actual
2127 * context switch. We thus wait for two switches to be
2128 * sure at least one completed.
2130 if ((p->nvcsw - nvcsw) > 1)
2132 if ((p->nivcsw - nivcsw) > 1)
2140 * wait_task_inactive - wait for a thread to unschedule.
2142 * If @match_state is nonzero, it's the @p->state value just checked and
2143 * not expected to change. If it changes, i.e. @p might have woken up,
2144 * then return zero. When we succeed in waiting for @p to be off its CPU,
2145 * we return a positive number (its total switch count). If a second call
2146 * a short while later returns the same number, the caller can be sure that
2147 * @p has remained unscheduled the whole time.
2149 * The caller must ensure that the task *will* unschedule sometime soon,
2150 * else this function might spin for a *long* time. This function can't
2151 * be called with interrupts off, or it may introduce deadlock with
2152 * smp_call_function() if an IPI is sent by the same process we are
2153 * waiting to become inactive.
2155 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2157 unsigned long flags;
2164 * We do the initial early heuristics without holding
2165 * any task-queue locks at all. We'll only try to get
2166 * the runqueue lock when things look like they will
2172 * If the task is actively running on another CPU
2173 * still, just relax and busy-wait without holding
2176 * NOTE! Since we don't hold any locks, it's not
2177 * even sure that "rq" stays as the right runqueue!
2178 * But we don't care, since "task_running()" will
2179 * return false if the runqueue has changed and p
2180 * is actually now running somewhere else!
2182 while (task_running(rq, p)) {
2183 if (match_state && unlikely(p->state != match_state))
2189 * Ok, time to look more closely! We need the rq
2190 * lock now, to be *sure*. If we're wrong, we'll
2191 * just go back and repeat.
2193 rq = task_rq_lock(p, &flags);
2194 trace_sched_wait_task(rq, p);
2195 running = task_running(rq, p);
2196 on_rq = p->se.on_rq;
2198 if (!match_state || p->state == match_state)
2199 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2200 task_rq_unlock(rq, &flags);
2203 * If it changed from the expected state, bail out now.
2205 if (unlikely(!ncsw))
2209 * Was it really running after all now that we
2210 * checked with the proper locks actually held?
2212 * Oops. Go back and try again..
2214 if (unlikely(running)) {
2220 * It's not enough that it's not actively running,
2221 * it must be off the runqueue _entirely_, and not
2224 * So if it was still runnable (but just not actively
2225 * running right now), it's preempted, and we should
2226 * yield - it could be a while.
2228 if (unlikely(on_rq)) {
2229 schedule_timeout_uninterruptible(1);
2234 * Ahh, all good. It wasn't running, and it wasn't
2235 * runnable, which means that it will never become
2236 * running in the future either. We're all done!
2245 * kick_process - kick a running thread to enter/exit the kernel
2246 * @p: the to-be-kicked thread
2248 * Cause a process which is running on another CPU to enter
2249 * kernel-mode, without any delay. (to get signals handled.)
2251 * NOTE: this function doesnt have to take the runqueue lock,
2252 * because all it wants to ensure is that the remote task enters
2253 * the kernel. If the IPI races and the task has been migrated
2254 * to another CPU then no harm is done and the purpose has been
2257 void kick_process(struct task_struct *p)
2263 if ((cpu != smp_processor_id()) && task_curr(p))
2264 smp_send_reschedule(cpu);
2267 EXPORT_SYMBOL_GPL(kick_process);
2268 #endif /* CONFIG_SMP */
2271 * task_oncpu_function_call - call a function on the cpu on which a task runs
2272 * @p: the task to evaluate
2273 * @func: the function to be called
2274 * @info: the function call argument
2276 * Calls the function @func when the task is currently running. This might
2277 * be on the current CPU, which just calls the function directly
2279 void task_oncpu_function_call(struct task_struct *p,
2280 void (*func) (void *info), void *info)
2287 smp_call_function_single(cpu, func, info, 1);
2292 * try_to_wake_up - wake up a thread
2293 * @p: the to-be-woken-up thread
2294 * @state: the mask of task states that can be woken
2295 * @sync: do a synchronous wakeup?
2297 * Put it on the run-queue if it's not already there. The "current"
2298 * thread is always on the run-queue (except when the actual
2299 * re-schedule is in progress), and as such you're allowed to do
2300 * the simpler "current->state = TASK_RUNNING" to mark yourself
2301 * runnable without the overhead of this.
2303 * returns failure only if the task is already active.
2305 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2308 int cpu, orig_cpu, this_cpu, success = 0;
2309 unsigned long flags;
2312 if (!sched_feat(SYNC_WAKEUPS))
2313 wake_flags &= ~WF_SYNC;
2315 this_cpu = get_cpu();
2318 rq = task_rq_lock(p, &flags);
2319 update_rq_clock(rq);
2320 if (!(p->state & state))
2330 if (unlikely(task_running(rq, p)))
2334 * In order to handle concurrent wakeups and release the rq->lock
2335 * we put the task in TASK_WAKING state.
2337 * First fix up the nr_uninterruptible count:
2339 if (task_contributes_to_load(p))
2340 rq->nr_uninterruptible--;
2341 p->state = TASK_WAKING;
2342 task_rq_unlock(rq, &flags);
2344 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2345 if (cpu != orig_cpu)
2346 set_task_cpu(p, cpu);
2348 rq = task_rq_lock(p, &flags);
2349 WARN_ON(p->state != TASK_WAKING);
2352 #ifdef CONFIG_SCHEDSTATS
2353 schedstat_inc(rq, ttwu_count);
2354 if (cpu == this_cpu)
2355 schedstat_inc(rq, ttwu_local);
2357 struct sched_domain *sd;
2358 for_each_domain(this_cpu, sd) {
2359 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2360 schedstat_inc(sd, ttwu_wake_remote);
2365 #endif /* CONFIG_SCHEDSTATS */
2368 #endif /* CONFIG_SMP */
2369 schedstat_inc(p, se.nr_wakeups);
2370 if (wake_flags & WF_SYNC)
2371 schedstat_inc(p, se.nr_wakeups_sync);
2372 if (orig_cpu != cpu)
2373 schedstat_inc(p, se.nr_wakeups_migrate);
2374 if (cpu == this_cpu)
2375 schedstat_inc(p, se.nr_wakeups_local);
2377 schedstat_inc(p, se.nr_wakeups_remote);
2378 activate_task(rq, p, 1);
2382 * Only attribute actual wakeups done by this task.
2384 if (!in_interrupt()) {
2385 struct sched_entity *se = ¤t->se;
2386 u64 sample = se->sum_exec_runtime;
2388 if (se->last_wakeup)
2389 sample -= se->last_wakeup;
2391 sample -= se->start_runtime;
2392 update_avg(&se->avg_wakeup, sample);
2394 se->last_wakeup = se->sum_exec_runtime;
2398 trace_sched_wakeup(rq, p, success);
2399 check_preempt_curr(rq, p, wake_flags);
2401 p->state = TASK_RUNNING;
2403 if (p->sched_class->task_wake_up)
2404 p->sched_class->task_wake_up(rq, p);
2407 task_rq_unlock(rq, &flags);
2414 * wake_up_process - Wake up a specific process
2415 * @p: The process to be woken up.
2417 * Attempt to wake up the nominated process and move it to the set of runnable
2418 * processes. Returns 1 if the process was woken up, 0 if it was already
2421 * It may be assumed that this function implies a write memory barrier before
2422 * changing the task state if and only if any tasks are woken up.
2424 int wake_up_process(struct task_struct *p)
2426 return try_to_wake_up(p, TASK_ALL, 0);
2428 EXPORT_SYMBOL(wake_up_process);
2430 int wake_up_state(struct task_struct *p, unsigned int state)
2432 return try_to_wake_up(p, state, 0);
2436 * Perform scheduler related setup for a newly forked process p.
2437 * p is forked by current.
2439 * __sched_fork() is basic setup used by init_idle() too:
2441 static void __sched_fork(struct task_struct *p)
2443 p->se.exec_start = 0;
2444 p->se.sum_exec_runtime = 0;
2445 p->se.prev_sum_exec_runtime = 0;
2446 p->se.nr_migrations = 0;
2447 p->se.last_wakeup = 0;
2448 p->se.avg_overlap = 0;
2449 p->se.start_runtime = 0;
2450 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2451 p->se.avg_running = 0;
2453 #ifdef CONFIG_SCHEDSTATS
2454 p->se.wait_start = 0;
2456 p->se.wait_count = 0;
2459 p->se.sleep_start = 0;
2460 p->se.sleep_max = 0;
2461 p->se.sum_sleep_runtime = 0;
2463 p->se.block_start = 0;
2464 p->se.block_max = 0;
2466 p->se.slice_max = 0;
2468 p->se.nr_migrations_cold = 0;
2469 p->se.nr_failed_migrations_affine = 0;
2470 p->se.nr_failed_migrations_running = 0;
2471 p->se.nr_failed_migrations_hot = 0;
2472 p->se.nr_forced_migrations = 0;
2473 p->se.nr_forced2_migrations = 0;
2475 p->se.nr_wakeups = 0;
2476 p->se.nr_wakeups_sync = 0;
2477 p->se.nr_wakeups_migrate = 0;
2478 p->se.nr_wakeups_local = 0;
2479 p->se.nr_wakeups_remote = 0;
2480 p->se.nr_wakeups_affine = 0;
2481 p->se.nr_wakeups_affine_attempts = 0;
2482 p->se.nr_wakeups_passive = 0;
2483 p->se.nr_wakeups_idle = 0;
2487 INIT_LIST_HEAD(&p->rt.run_list);
2489 INIT_LIST_HEAD(&p->se.group_node);
2491 #ifdef CONFIG_PREEMPT_NOTIFIERS
2492 INIT_HLIST_HEAD(&p->preempt_notifiers);
2496 * We mark the process as running here, but have not actually
2497 * inserted it onto the runqueue yet. This guarantees that
2498 * nobody will actually run it, and a signal or other external
2499 * event cannot wake it up and insert it on the runqueue either.
2501 p->state = TASK_RUNNING;
2505 * fork()/clone()-time setup:
2507 void sched_fork(struct task_struct *p, int clone_flags)
2509 int cpu = get_cpu();
2514 * Make sure we do not leak PI boosting priority to the child.
2516 p->prio = current->normal_prio;
2519 * Revert to default priority/policy on fork if requested.
2521 if (unlikely(p->sched_reset_on_fork)) {
2522 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR)
2523 p->policy = SCHED_NORMAL;
2525 if (p->normal_prio < DEFAULT_PRIO)
2526 p->prio = DEFAULT_PRIO;
2528 if (PRIO_TO_NICE(p->static_prio) < 0) {
2529 p->static_prio = NICE_TO_PRIO(0);
2534 * We don't need the reset flag anymore after the fork. It has
2535 * fulfilled its duty:
2537 p->sched_reset_on_fork = 0;
2540 if (!rt_prio(p->prio))
2541 p->sched_class = &fair_sched_class;
2544 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_FORK, 0);
2546 set_task_cpu(p, cpu);
2548 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2549 if (likely(sched_info_on()))
2550 memset(&p->sched_info, 0, sizeof(p->sched_info));
2552 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2555 #ifdef CONFIG_PREEMPT
2556 /* Want to start with kernel preemption disabled. */
2557 task_thread_info(p)->preempt_count = 1;
2559 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2565 * wake_up_new_task - wake up a newly created task for the first time.
2567 * This function will do some initial scheduler statistics housekeeping
2568 * that must be done for every newly created context, then puts the task
2569 * on the runqueue and wakes it.
2571 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2573 unsigned long flags;
2576 rq = task_rq_lock(p, &flags);
2577 BUG_ON(p->state != TASK_RUNNING);
2578 update_rq_clock(rq);
2580 p->prio = effective_prio(p);
2582 if (!p->sched_class->task_new || !current->se.on_rq) {
2583 activate_task(rq, p, 0);
2586 * Let the scheduling class do new task startup
2587 * management (if any):
2589 p->sched_class->task_new(rq, p);
2592 trace_sched_wakeup_new(rq, p, 1);
2593 check_preempt_curr(rq, p, WF_FORK);
2595 if (p->sched_class->task_wake_up)
2596 p->sched_class->task_wake_up(rq, p);
2598 task_rq_unlock(rq, &flags);
2601 #ifdef CONFIG_PREEMPT_NOTIFIERS
2604 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2605 * @notifier: notifier struct to register
2607 void preempt_notifier_register(struct preempt_notifier *notifier)
2609 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2611 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2614 * preempt_notifier_unregister - no longer interested in preemption notifications
2615 * @notifier: notifier struct to unregister
2617 * This is safe to call from within a preemption notifier.
2619 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2621 hlist_del(¬ifier->link);
2623 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2625 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2627 struct preempt_notifier *notifier;
2628 struct hlist_node *node;
2630 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2631 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2635 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2636 struct task_struct *next)
2638 struct preempt_notifier *notifier;
2639 struct hlist_node *node;
2641 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2642 notifier->ops->sched_out(notifier, next);
2645 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2647 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2652 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2653 struct task_struct *next)
2657 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2660 * prepare_task_switch - prepare to switch tasks
2661 * @rq: the runqueue preparing to switch
2662 * @prev: the current task that is being switched out
2663 * @next: the task we are going to switch to.
2665 * This is called with the rq lock held and interrupts off. It must
2666 * be paired with a subsequent finish_task_switch after the context
2669 * prepare_task_switch sets up locking and calls architecture specific
2673 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2674 struct task_struct *next)
2676 fire_sched_out_preempt_notifiers(prev, next);
2677 prepare_lock_switch(rq, next);
2678 prepare_arch_switch(next);
2682 * finish_task_switch - clean up after a task-switch
2683 * @rq: runqueue associated with task-switch
2684 * @prev: the thread we just switched away from.
2686 * finish_task_switch must be called after the context switch, paired
2687 * with a prepare_task_switch call before the context switch.
2688 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2689 * and do any other architecture-specific cleanup actions.
2691 * Note that we may have delayed dropping an mm in context_switch(). If
2692 * so, we finish that here outside of the runqueue lock. (Doing it
2693 * with the lock held can cause deadlocks; see schedule() for
2696 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2697 __releases(rq->lock)
2699 struct mm_struct *mm = rq->prev_mm;
2705 * A task struct has one reference for the use as "current".
2706 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2707 * schedule one last time. The schedule call will never return, and
2708 * the scheduled task must drop that reference.
2709 * The test for TASK_DEAD must occur while the runqueue locks are
2710 * still held, otherwise prev could be scheduled on another cpu, die
2711 * there before we look at prev->state, and then the reference would
2713 * Manfred Spraul <manfred@colorfullife.com>
2715 prev_state = prev->state;
2716 finish_arch_switch(prev);
2717 perf_event_task_sched_in(current, cpu_of(rq));
2718 finish_lock_switch(rq, prev);
2720 fire_sched_in_preempt_notifiers(current);
2723 if (unlikely(prev_state == TASK_DEAD)) {
2725 * Remove function-return probe instances associated with this
2726 * task and put them back on the free list.
2728 kprobe_flush_task(prev);
2729 put_task_struct(prev);
2735 /* assumes rq->lock is held */
2736 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2738 if (prev->sched_class->pre_schedule)
2739 prev->sched_class->pre_schedule(rq, prev);
2742 /* rq->lock is NOT held, but preemption is disabled */
2743 static inline void post_schedule(struct rq *rq)
2745 if (rq->post_schedule) {
2746 unsigned long flags;
2748 spin_lock_irqsave(&rq->lock, flags);
2749 if (rq->curr->sched_class->post_schedule)
2750 rq->curr->sched_class->post_schedule(rq);
2751 spin_unlock_irqrestore(&rq->lock, flags);
2753 rq->post_schedule = 0;
2759 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2763 static inline void post_schedule(struct rq *rq)
2770 * schedule_tail - first thing a freshly forked thread must call.
2771 * @prev: the thread we just switched away from.
2773 asmlinkage void schedule_tail(struct task_struct *prev)
2774 __releases(rq->lock)
2776 struct rq *rq = this_rq();
2778 finish_task_switch(rq, prev);
2781 * FIXME: do we need to worry about rq being invalidated by the
2786 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2787 /* In this case, finish_task_switch does not reenable preemption */
2790 if (current->set_child_tid)
2791 put_user(task_pid_vnr(current), current->set_child_tid);
2795 * context_switch - switch to the new MM and the new
2796 * thread's register state.
2799 context_switch(struct rq *rq, struct task_struct *prev,
2800 struct task_struct *next)
2802 struct mm_struct *mm, *oldmm;
2804 prepare_task_switch(rq, prev, next);
2805 trace_sched_switch(rq, prev, next);
2807 oldmm = prev->active_mm;
2809 * For paravirt, this is coupled with an exit in switch_to to
2810 * combine the page table reload and the switch backend into
2813 arch_start_context_switch(prev);
2815 if (unlikely(!mm)) {
2816 next->active_mm = oldmm;
2817 atomic_inc(&oldmm->mm_count);
2818 enter_lazy_tlb(oldmm, next);
2820 switch_mm(oldmm, mm, next);
2822 if (unlikely(!prev->mm)) {
2823 prev->active_mm = NULL;
2824 rq->prev_mm = oldmm;
2827 * Since the runqueue lock will be released by the next
2828 * task (which is an invalid locking op but in the case
2829 * of the scheduler it's an obvious special-case), so we
2830 * do an early lockdep release here:
2832 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2833 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2836 /* Here we just switch the register state and the stack. */
2837 switch_to(prev, next, prev);
2841 * this_rq must be evaluated again because prev may have moved
2842 * CPUs since it called schedule(), thus the 'rq' on its stack
2843 * frame will be invalid.
2845 finish_task_switch(this_rq(), prev);
2849 * nr_running, nr_uninterruptible and nr_context_switches:
2851 * externally visible scheduler statistics: current number of runnable
2852 * threads, current number of uninterruptible-sleeping threads, total
2853 * number of context switches performed since bootup.
2855 unsigned long nr_running(void)
2857 unsigned long i, sum = 0;
2859 for_each_online_cpu(i)
2860 sum += cpu_rq(i)->nr_running;
2865 unsigned long nr_uninterruptible(void)
2867 unsigned long i, sum = 0;
2869 for_each_possible_cpu(i)
2870 sum += cpu_rq(i)->nr_uninterruptible;
2873 * Since we read the counters lockless, it might be slightly
2874 * inaccurate. Do not allow it to go below zero though:
2876 if (unlikely((long)sum < 0))
2882 unsigned long long nr_context_switches(void)
2885 unsigned long long sum = 0;
2887 for_each_possible_cpu(i)
2888 sum += cpu_rq(i)->nr_switches;
2893 unsigned long nr_iowait(void)
2895 unsigned long i, sum = 0;
2897 for_each_possible_cpu(i)
2898 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2903 unsigned long nr_iowait_cpu(void)
2905 struct rq *this = this_rq();
2906 return atomic_read(&this->nr_iowait);
2909 unsigned long this_cpu_load(void)
2911 struct rq *this = this_rq();
2912 return this->cpu_load[0];
2916 /* Variables and functions for calc_load */
2917 static atomic_long_t calc_load_tasks;
2918 static unsigned long calc_load_update;
2919 unsigned long avenrun[3];
2920 EXPORT_SYMBOL(avenrun);
2923 * get_avenrun - get the load average array
2924 * @loads: pointer to dest load array
2925 * @offset: offset to add
2926 * @shift: shift count to shift the result left
2928 * These values are estimates at best, so no need for locking.
2930 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2932 loads[0] = (avenrun[0] + offset) << shift;
2933 loads[1] = (avenrun[1] + offset) << shift;
2934 loads[2] = (avenrun[2] + offset) << shift;
2937 static unsigned long
2938 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2941 load += active * (FIXED_1 - exp);
2942 return load >> FSHIFT;
2946 * calc_load - update the avenrun load estimates 10 ticks after the
2947 * CPUs have updated calc_load_tasks.
2949 void calc_global_load(void)
2951 unsigned long upd = calc_load_update + 10;
2954 if (time_before(jiffies, upd))
2957 active = atomic_long_read(&calc_load_tasks);
2958 active = active > 0 ? active * FIXED_1 : 0;
2960 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2961 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2962 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2964 calc_load_update += LOAD_FREQ;
2968 * Either called from update_cpu_load() or from a cpu going idle
2970 static void calc_load_account_active(struct rq *this_rq)
2972 long nr_active, delta;
2974 nr_active = this_rq->nr_running;
2975 nr_active += (long) this_rq->nr_uninterruptible;
2977 if (nr_active != this_rq->calc_load_active) {
2978 delta = nr_active - this_rq->calc_load_active;
2979 this_rq->calc_load_active = nr_active;
2980 atomic_long_add(delta, &calc_load_tasks);
2985 * Externally visible per-cpu scheduler statistics:
2986 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2988 u64 cpu_nr_migrations(int cpu)
2990 return cpu_rq(cpu)->nr_migrations_in;
2994 * Update rq->cpu_load[] statistics. This function is usually called every
2995 * scheduler tick (TICK_NSEC).
2997 static void update_cpu_load(struct rq *this_rq)
2999 unsigned long this_load = this_rq->load.weight;
3002 this_rq->nr_load_updates++;
3004 /* Update our load: */
3005 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3006 unsigned long old_load, new_load;
3008 /* scale is effectively 1 << i now, and >> i divides by scale */
3010 old_load = this_rq->cpu_load[i];
3011 new_load = this_load;
3013 * Round up the averaging division if load is increasing. This
3014 * prevents us from getting stuck on 9 if the load is 10, for
3017 if (new_load > old_load)
3018 new_load += scale-1;
3019 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3022 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3023 this_rq->calc_load_update += LOAD_FREQ;
3024 calc_load_account_active(this_rq);
3031 * double_rq_lock - safely lock two runqueues
3033 * Note this does not disable interrupts like task_rq_lock,
3034 * you need to do so manually before calling.
3036 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3037 __acquires(rq1->lock)
3038 __acquires(rq2->lock)
3040 BUG_ON(!irqs_disabled());
3042 spin_lock(&rq1->lock);
3043 __acquire(rq2->lock); /* Fake it out ;) */
3046 spin_lock(&rq1->lock);
3047 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3049 spin_lock(&rq2->lock);
3050 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3053 update_rq_clock(rq1);
3054 update_rq_clock(rq2);
3058 * double_rq_unlock - safely unlock two runqueues
3060 * Note this does not restore interrupts like task_rq_unlock,
3061 * you need to do so manually after calling.
3063 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3064 __releases(rq1->lock)
3065 __releases(rq2->lock)
3067 spin_unlock(&rq1->lock);
3069 spin_unlock(&rq2->lock);
3071 __release(rq2->lock);
3075 * If dest_cpu is allowed for this process, migrate the task to it.
3076 * This is accomplished by forcing the cpu_allowed mask to only
3077 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3078 * the cpu_allowed mask is restored.
3080 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3082 struct migration_req req;
3083 unsigned long flags;
3086 rq = task_rq_lock(p, &flags);
3087 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3088 || unlikely(!cpu_active(dest_cpu)))
3091 /* force the process onto the specified CPU */
3092 if (migrate_task(p, dest_cpu, &req)) {
3093 /* Need to wait for migration thread (might exit: take ref). */
3094 struct task_struct *mt = rq->migration_thread;
3096 get_task_struct(mt);
3097 task_rq_unlock(rq, &flags);
3098 wake_up_process(mt);
3099 put_task_struct(mt);
3100 wait_for_completion(&req.done);
3105 task_rq_unlock(rq, &flags);
3109 * sched_exec - execve() is a valuable balancing opportunity, because at
3110 * this point the task has the smallest effective memory and cache footprint.
3112 void sched_exec(void)
3114 int new_cpu, this_cpu = get_cpu();
3115 new_cpu = current->sched_class->select_task_rq(current, SD_BALANCE_EXEC, 0);
3117 if (new_cpu != this_cpu)
3118 sched_migrate_task(current, new_cpu);
3122 * pull_task - move a task from a remote runqueue to the local runqueue.
3123 * Both runqueues must be locked.
3125 static void pull_task(struct rq *src_rq, struct task_struct *p,
3126 struct rq *this_rq, int this_cpu)
3128 deactivate_task(src_rq, p, 0);
3129 set_task_cpu(p, this_cpu);
3130 activate_task(this_rq, p, 0);
3132 * Note that idle threads have a prio of MAX_PRIO, for this test
3133 * to be always true for them.
3135 check_preempt_curr(this_rq, p, 0);
3139 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3142 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3143 struct sched_domain *sd, enum cpu_idle_type idle,
3146 int tsk_cache_hot = 0;
3148 * We do not migrate tasks that are:
3149 * 1) running (obviously), or
3150 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3151 * 3) are cache-hot on their current CPU.
3153 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3154 schedstat_inc(p, se.nr_failed_migrations_affine);
3159 if (task_running(rq, p)) {
3160 schedstat_inc(p, se.nr_failed_migrations_running);
3165 * Aggressive migration if:
3166 * 1) task is cache cold, or
3167 * 2) too many balance attempts have failed.
3170 tsk_cache_hot = task_hot(p, rq->clock, sd);
3171 if (!tsk_cache_hot ||
3172 sd->nr_balance_failed > sd->cache_nice_tries) {
3173 #ifdef CONFIG_SCHEDSTATS
3174 if (tsk_cache_hot) {
3175 schedstat_inc(sd, lb_hot_gained[idle]);
3176 schedstat_inc(p, se.nr_forced_migrations);
3182 if (tsk_cache_hot) {
3183 schedstat_inc(p, se.nr_failed_migrations_hot);
3189 static unsigned long
3190 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3191 unsigned long max_load_move, struct sched_domain *sd,
3192 enum cpu_idle_type idle, int *all_pinned,
3193 int *this_best_prio, struct rq_iterator *iterator)
3195 int loops = 0, pulled = 0, pinned = 0;
3196 struct task_struct *p;
3197 long rem_load_move = max_load_move;
3199 if (max_load_move == 0)
3205 * Start the load-balancing iterator:
3207 p = iterator->start(iterator->arg);
3209 if (!p || loops++ > sysctl_sched_nr_migrate)
3212 if ((p->se.load.weight >> 1) > rem_load_move ||
3213 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3214 p = iterator->next(iterator->arg);
3218 pull_task(busiest, p, this_rq, this_cpu);
3220 rem_load_move -= p->se.load.weight;
3222 #ifdef CONFIG_PREEMPT
3224 * NEWIDLE balancing is a source of latency, so preemptible kernels
3225 * will stop after the first task is pulled to minimize the critical
3228 if (idle == CPU_NEWLY_IDLE)
3233 * We only want to steal up to the prescribed amount of weighted load.
3235 if (rem_load_move > 0) {
3236 if (p->prio < *this_best_prio)
3237 *this_best_prio = p->prio;
3238 p = iterator->next(iterator->arg);
3243 * Right now, this is one of only two places pull_task() is called,
3244 * so we can safely collect pull_task() stats here rather than
3245 * inside pull_task().
3247 schedstat_add(sd, lb_gained[idle], pulled);
3250 *all_pinned = pinned;
3252 return max_load_move - rem_load_move;
3256 * move_tasks tries to move up to max_load_move weighted load from busiest to
3257 * this_rq, as part of a balancing operation within domain "sd".
3258 * Returns 1 if successful and 0 otherwise.
3260 * Called with both runqueues locked.
3262 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3263 unsigned long max_load_move,
3264 struct sched_domain *sd, enum cpu_idle_type idle,
3267 const struct sched_class *class = sched_class_highest;
3268 unsigned long total_load_moved = 0;
3269 int this_best_prio = this_rq->curr->prio;
3273 class->load_balance(this_rq, this_cpu, busiest,
3274 max_load_move - total_load_moved,
3275 sd, idle, all_pinned, &this_best_prio);
3276 class = class->next;
3278 #ifdef CONFIG_PREEMPT
3280 * NEWIDLE balancing is a source of latency, so preemptible
3281 * kernels will stop after the first task is pulled to minimize
3282 * the critical section.
3284 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3287 } while (class && max_load_move > total_load_moved);
3289 return total_load_moved > 0;
3293 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3294 struct sched_domain *sd, enum cpu_idle_type idle,
3295 struct rq_iterator *iterator)
3297 struct task_struct *p = iterator->start(iterator->arg);
3301 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3302 pull_task(busiest, p, this_rq, this_cpu);
3304 * Right now, this is only the second place pull_task()
3305 * is called, so we can safely collect pull_task()
3306 * stats here rather than inside pull_task().
3308 schedstat_inc(sd, lb_gained[idle]);
3312 p = iterator->next(iterator->arg);
3319 * move_one_task tries to move exactly one task from busiest to this_rq, as
3320 * part of active balancing operations within "domain".
3321 * Returns 1 if successful and 0 otherwise.
3323 * Called with both runqueues locked.
3325 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3326 struct sched_domain *sd, enum cpu_idle_type idle)
3328 const struct sched_class *class;
3330 for_each_class(class) {
3331 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3337 /********** Helpers for find_busiest_group ************************/
3339 * sd_lb_stats - Structure to store the statistics of a sched_domain
3340 * during load balancing.
3342 struct sd_lb_stats {
3343 struct sched_group *busiest; /* Busiest group in this sd */
3344 struct sched_group *this; /* Local group in this sd */
3345 unsigned long total_load; /* Total load of all groups in sd */
3346 unsigned long total_pwr; /* Total power of all groups in sd */
3347 unsigned long avg_load; /* Average load across all groups in sd */
3349 /** Statistics of this group */
3350 unsigned long this_load;
3351 unsigned long this_load_per_task;
3352 unsigned long this_nr_running;
3354 /* Statistics of the busiest group */
3355 unsigned long max_load;
3356 unsigned long busiest_load_per_task;
3357 unsigned long busiest_nr_running;
3359 int group_imb; /* Is there imbalance in this sd */
3360 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3361 int power_savings_balance; /* Is powersave balance needed for this sd */
3362 struct sched_group *group_min; /* Least loaded group in sd */
3363 struct sched_group *group_leader; /* Group which relieves group_min */
3364 unsigned long min_load_per_task; /* load_per_task in group_min */
3365 unsigned long leader_nr_running; /* Nr running of group_leader */
3366 unsigned long min_nr_running; /* Nr running of group_min */
3371 * sg_lb_stats - stats of a sched_group required for load_balancing
3373 struct sg_lb_stats {
3374 unsigned long avg_load; /*Avg load across the CPUs of the group */
3375 unsigned long group_load; /* Total load over the CPUs of the group */
3376 unsigned long sum_nr_running; /* Nr tasks running in the group */
3377 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3378 unsigned long group_capacity;
3379 int group_imb; /* Is there an imbalance in the group ? */
3383 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3384 * @group: The group whose first cpu is to be returned.
3386 static inline unsigned int group_first_cpu(struct sched_group *group)
3388 return cpumask_first(sched_group_cpus(group));
3392 * get_sd_load_idx - Obtain the load index for a given sched domain.
3393 * @sd: The sched_domain whose load_idx is to be obtained.
3394 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3396 static inline int get_sd_load_idx(struct sched_domain *sd,
3397 enum cpu_idle_type idle)
3403 load_idx = sd->busy_idx;
3406 case CPU_NEWLY_IDLE:
3407 load_idx = sd->newidle_idx;
3410 load_idx = sd->idle_idx;
3418 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3420 * init_sd_power_savings_stats - Initialize power savings statistics for
3421 * the given sched_domain, during load balancing.
3423 * @sd: Sched domain whose power-savings statistics are to be initialized.
3424 * @sds: Variable containing the statistics for sd.
3425 * @idle: Idle status of the CPU at which we're performing load-balancing.
3427 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3428 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3431 * Busy processors will not participate in power savings
3434 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3435 sds->power_savings_balance = 0;
3437 sds->power_savings_balance = 1;
3438 sds->min_nr_running = ULONG_MAX;
3439 sds->leader_nr_running = 0;
3444 * update_sd_power_savings_stats - Update the power saving stats for a
3445 * sched_domain while performing load balancing.
3447 * @group: sched_group belonging to the sched_domain under consideration.
3448 * @sds: Variable containing the statistics of the sched_domain
3449 * @local_group: Does group contain the CPU for which we're performing
3451 * @sgs: Variable containing the statistics of the group.
3453 static inline void update_sd_power_savings_stats(struct sched_group *group,
3454 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3457 if (!sds->power_savings_balance)
3461 * If the local group is idle or completely loaded
3462 * no need to do power savings balance at this domain
3464 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3465 !sds->this_nr_running))
3466 sds->power_savings_balance = 0;
3469 * If a group is already running at full capacity or idle,
3470 * don't include that group in power savings calculations
3472 if (!sds->power_savings_balance ||
3473 sgs->sum_nr_running >= sgs->group_capacity ||
3474 !sgs->sum_nr_running)
3478 * Calculate the group which has the least non-idle load.
3479 * This is the group from where we need to pick up the load
3482 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3483 (sgs->sum_nr_running == sds->min_nr_running &&
3484 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3485 sds->group_min = group;
3486 sds->min_nr_running = sgs->sum_nr_running;
3487 sds->min_load_per_task = sgs->sum_weighted_load /
3488 sgs->sum_nr_running;
3492 * Calculate the group which is almost near its
3493 * capacity but still has some space to pick up some load
3494 * from other group and save more power
3496 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3499 if (sgs->sum_nr_running > sds->leader_nr_running ||
3500 (sgs->sum_nr_running == sds->leader_nr_running &&
3501 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3502 sds->group_leader = group;
3503 sds->leader_nr_running = sgs->sum_nr_running;
3508 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3509 * @sds: Variable containing the statistics of the sched_domain
3510 * under consideration.
3511 * @this_cpu: Cpu at which we're currently performing load-balancing.
3512 * @imbalance: Variable to store the imbalance.
3515 * Check if we have potential to perform some power-savings balance.
3516 * If yes, set the busiest group to be the least loaded group in the
3517 * sched_domain, so that it's CPUs can be put to idle.
3519 * Returns 1 if there is potential to perform power-savings balance.
3522 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3523 int this_cpu, unsigned long *imbalance)
3525 if (!sds->power_savings_balance)
3528 if (sds->this != sds->group_leader ||
3529 sds->group_leader == sds->group_min)
3532 *imbalance = sds->min_load_per_task;
3533 sds->busiest = sds->group_min;
3538 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3539 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3540 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3545 static inline void update_sd_power_savings_stats(struct sched_group *group,
3546 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3551 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3552 int this_cpu, unsigned long *imbalance)
3556 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3559 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3561 return SCHED_LOAD_SCALE;
3564 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3566 return default_scale_freq_power(sd, cpu);
3569 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3571 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3572 unsigned long smt_gain = sd->smt_gain;
3579 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3581 return default_scale_smt_power(sd, cpu);
3584 unsigned long scale_rt_power(int cpu)
3586 struct rq *rq = cpu_rq(cpu);
3587 u64 total, available;
3589 sched_avg_update(rq);
3591 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3592 available = total - rq->rt_avg;
3594 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3595 total = SCHED_LOAD_SCALE;
3597 total >>= SCHED_LOAD_SHIFT;
3599 return div_u64(available, total);
3602 static void update_cpu_power(struct sched_domain *sd, int cpu)
3604 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3605 unsigned long power = SCHED_LOAD_SCALE;
3606 struct sched_group *sdg = sd->groups;
3608 if (sched_feat(ARCH_POWER))
3609 power *= arch_scale_freq_power(sd, cpu);
3611 power *= default_scale_freq_power(sd, cpu);
3613 power >>= SCHED_LOAD_SHIFT;
3615 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3616 if (sched_feat(ARCH_POWER))
3617 power *= arch_scale_smt_power(sd, cpu);
3619 power *= default_scale_smt_power(sd, cpu);
3621 power >>= SCHED_LOAD_SHIFT;
3624 power *= scale_rt_power(cpu);
3625 power >>= SCHED_LOAD_SHIFT;
3630 sdg->cpu_power = power;
3633 static void update_group_power(struct sched_domain *sd, int cpu)
3635 struct sched_domain *child = sd->child;
3636 struct sched_group *group, *sdg = sd->groups;
3637 unsigned long power;
3640 update_cpu_power(sd, cpu);
3646 group = child->groups;
3648 power += group->cpu_power;
3649 group = group->next;
3650 } while (group != child->groups);
3652 sdg->cpu_power = power;
3656 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3657 * @group: sched_group whose statistics are to be updated.
3658 * @this_cpu: Cpu for which load balance is currently performed.
3659 * @idle: Idle status of this_cpu
3660 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3661 * @sd_idle: Idle status of the sched_domain containing group.
3662 * @local_group: Does group contain this_cpu.
3663 * @cpus: Set of cpus considered for load balancing.
3664 * @balance: Should we balance.
3665 * @sgs: variable to hold the statistics for this group.
3667 static inline void update_sg_lb_stats(struct sched_domain *sd,
3668 struct sched_group *group, int this_cpu,
3669 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3670 int local_group, const struct cpumask *cpus,
3671 int *balance, struct sg_lb_stats *sgs)
3673 unsigned long load, max_cpu_load, min_cpu_load;
3675 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3676 unsigned long sum_avg_load_per_task;
3677 unsigned long avg_load_per_task;
3680 balance_cpu = group_first_cpu(group);
3681 if (balance_cpu == this_cpu)
3682 update_group_power(sd, this_cpu);
3685 /* Tally up the load of all CPUs in the group */
3686 sum_avg_load_per_task = avg_load_per_task = 0;
3688 min_cpu_load = ~0UL;
3690 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3691 struct rq *rq = cpu_rq(i);
3693 if (*sd_idle && rq->nr_running)
3696 /* Bias balancing toward cpus of our domain */
3698 if (idle_cpu(i) && !first_idle_cpu) {
3703 load = target_load(i, load_idx);
3705 load = source_load(i, load_idx);
3706 if (load > max_cpu_load)
3707 max_cpu_load = load;
3708 if (min_cpu_load > load)
3709 min_cpu_load = load;
3712 sgs->group_load += load;
3713 sgs->sum_nr_running += rq->nr_running;
3714 sgs->sum_weighted_load += weighted_cpuload(i);
3716 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3720 * First idle cpu or the first cpu(busiest) in this sched group
3721 * is eligible for doing load balancing at this and above
3722 * domains. In the newly idle case, we will allow all the cpu's
3723 * to do the newly idle load balance.
3725 if (idle != CPU_NEWLY_IDLE && local_group &&
3726 balance_cpu != this_cpu && balance) {
3731 /* Adjust by relative CPU power of the group */
3732 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3736 * Consider the group unbalanced when the imbalance is larger
3737 * than the average weight of two tasks.
3739 * APZ: with cgroup the avg task weight can vary wildly and
3740 * might not be a suitable number - should we keep a
3741 * normalized nr_running number somewhere that negates
3744 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3747 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3750 sgs->group_capacity =
3751 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3755 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3756 * @sd: sched_domain whose statistics are to be updated.
3757 * @this_cpu: Cpu for which load balance is currently performed.
3758 * @idle: Idle status of this_cpu
3759 * @sd_idle: Idle status of the sched_domain containing group.
3760 * @cpus: Set of cpus considered for load balancing.
3761 * @balance: Should we balance.
3762 * @sds: variable to hold the statistics for this sched_domain.
3764 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3765 enum cpu_idle_type idle, int *sd_idle,
3766 const struct cpumask *cpus, int *balance,
3767 struct sd_lb_stats *sds)
3769 struct sched_domain *child = sd->child;
3770 struct sched_group *group = sd->groups;
3771 struct sg_lb_stats sgs;
3772 int load_idx, prefer_sibling = 0;
3774 if (child && child->flags & SD_PREFER_SIBLING)
3777 init_sd_power_savings_stats(sd, sds, idle);
3778 load_idx = get_sd_load_idx(sd, idle);
3783 local_group = cpumask_test_cpu(this_cpu,
3784 sched_group_cpus(group));
3785 memset(&sgs, 0, sizeof(sgs));
3786 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3787 local_group, cpus, balance, &sgs);
3789 if (local_group && balance && !(*balance))
3792 sds->total_load += sgs.group_load;
3793 sds->total_pwr += group->cpu_power;
3796 * In case the child domain prefers tasks go to siblings
3797 * first, lower the group capacity to one so that we'll try
3798 * and move all the excess tasks away.
3801 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3804 sds->this_load = sgs.avg_load;
3806 sds->this_nr_running = sgs.sum_nr_running;
3807 sds->this_load_per_task = sgs.sum_weighted_load;
3808 } else if (sgs.avg_load > sds->max_load &&
3809 (sgs.sum_nr_running > sgs.group_capacity ||
3811 sds->max_load = sgs.avg_load;
3812 sds->busiest = group;
3813 sds->busiest_nr_running = sgs.sum_nr_running;
3814 sds->busiest_load_per_task = sgs.sum_weighted_load;
3815 sds->group_imb = sgs.group_imb;
3818 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3819 group = group->next;
3820 } while (group != sd->groups);
3824 * fix_small_imbalance - Calculate the minor imbalance that exists
3825 * amongst the groups of a sched_domain, during
3827 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3828 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3829 * @imbalance: Variable to store the imbalance.
3831 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3832 int this_cpu, unsigned long *imbalance)
3834 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3835 unsigned int imbn = 2;
3837 if (sds->this_nr_running) {
3838 sds->this_load_per_task /= sds->this_nr_running;
3839 if (sds->busiest_load_per_task >
3840 sds->this_load_per_task)
3843 sds->this_load_per_task =
3844 cpu_avg_load_per_task(this_cpu);
3846 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3847 sds->busiest_load_per_task * imbn) {
3848 *imbalance = sds->busiest_load_per_task;
3853 * OK, we don't have enough imbalance to justify moving tasks,
3854 * however we may be able to increase total CPU power used by
3858 pwr_now += sds->busiest->cpu_power *
3859 min(sds->busiest_load_per_task, sds->max_load);
3860 pwr_now += sds->this->cpu_power *
3861 min(sds->this_load_per_task, sds->this_load);
3862 pwr_now /= SCHED_LOAD_SCALE;
3864 /* Amount of load we'd subtract */
3865 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3866 sds->busiest->cpu_power;
3867 if (sds->max_load > tmp)
3868 pwr_move += sds->busiest->cpu_power *
3869 min(sds->busiest_load_per_task, sds->max_load - tmp);
3871 /* Amount of load we'd add */
3872 if (sds->max_load * sds->busiest->cpu_power <
3873 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3874 tmp = (sds->max_load * sds->busiest->cpu_power) /
3875 sds->this->cpu_power;
3877 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3878 sds->this->cpu_power;
3879 pwr_move += sds->this->cpu_power *
3880 min(sds->this_load_per_task, sds->this_load + tmp);
3881 pwr_move /= SCHED_LOAD_SCALE;
3883 /* Move if we gain throughput */
3884 if (pwr_move > pwr_now)
3885 *imbalance = sds->busiest_load_per_task;
3889 * calculate_imbalance - Calculate the amount of imbalance present within the
3890 * groups of a given sched_domain during load balance.
3891 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3892 * @this_cpu: Cpu for which currently load balance is being performed.
3893 * @imbalance: The variable to store the imbalance.
3895 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3896 unsigned long *imbalance)
3898 unsigned long max_pull;
3900 * In the presence of smp nice balancing, certain scenarios can have
3901 * max load less than avg load(as we skip the groups at or below
3902 * its cpu_power, while calculating max_load..)
3904 if (sds->max_load < sds->avg_load) {
3906 return fix_small_imbalance(sds, this_cpu, imbalance);
3909 /* Don't want to pull so many tasks that a group would go idle */
3910 max_pull = min(sds->max_load - sds->avg_load,
3911 sds->max_load - sds->busiest_load_per_task);
3913 /* How much load to actually move to equalise the imbalance */
3914 *imbalance = min(max_pull * sds->busiest->cpu_power,
3915 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3919 * if *imbalance is less than the average load per runnable task
3920 * there is no gaurantee that any tasks will be moved so we'll have
3921 * a think about bumping its value to force at least one task to be
3924 if (*imbalance < sds->busiest_load_per_task)
3925 return fix_small_imbalance(sds, this_cpu, imbalance);
3928 /******* find_busiest_group() helpers end here *********************/
3931 * find_busiest_group - Returns the busiest group within the sched_domain
3932 * if there is an imbalance. If there isn't an imbalance, and
3933 * the user has opted for power-savings, it returns a group whose
3934 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3935 * such a group exists.
3937 * Also calculates the amount of weighted load which should be moved
3938 * to restore balance.
3940 * @sd: The sched_domain whose busiest group is to be returned.
3941 * @this_cpu: The cpu for which load balancing is currently being performed.
3942 * @imbalance: Variable which stores amount of weighted load which should
3943 * be moved to restore balance/put a group to idle.
3944 * @idle: The idle status of this_cpu.
3945 * @sd_idle: The idleness of sd
3946 * @cpus: The set of CPUs under consideration for load-balancing.
3947 * @balance: Pointer to a variable indicating if this_cpu
3948 * is the appropriate cpu to perform load balancing at this_level.
3950 * Returns: - the busiest group if imbalance exists.
3951 * - If no imbalance and user has opted for power-savings balance,
3952 * return the least loaded group whose CPUs can be
3953 * put to idle by rebalancing its tasks onto our group.
3955 static struct sched_group *
3956 find_busiest_group(struct sched_domain *sd, int this_cpu,
3957 unsigned long *imbalance, enum cpu_idle_type idle,
3958 int *sd_idle, const struct cpumask *cpus, int *balance)
3960 struct sd_lb_stats sds;
3962 memset(&sds, 0, sizeof(sds));
3965 * Compute the various statistics relavent for load balancing at
3968 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3971 /* Cases where imbalance does not exist from POV of this_cpu */
3972 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3974 * 2) There is no busy sibling group to pull from.
3975 * 3) This group is the busiest group.
3976 * 4) This group is more busy than the avg busieness at this
3978 * 5) The imbalance is within the specified limit.
3979 * 6) Any rebalance would lead to ping-pong
3981 if (balance && !(*balance))
3984 if (!sds.busiest || sds.busiest_nr_running == 0)
3987 if (sds.this_load >= sds.max_load)
3990 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3992 if (sds.this_load >= sds.avg_load)
3995 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3998 sds.busiest_load_per_task /= sds.busiest_nr_running;
4000 sds.busiest_load_per_task =
4001 min(sds.busiest_load_per_task, sds.avg_load);
4004 * We're trying to get all the cpus to the average_load, so we don't
4005 * want to push ourselves above the average load, nor do we wish to
4006 * reduce the max loaded cpu below the average load, as either of these
4007 * actions would just result in more rebalancing later, and ping-pong
4008 * tasks around. Thus we look for the minimum possible imbalance.
4009 * Negative imbalances (*we* are more loaded than anyone else) will
4010 * be counted as no imbalance for these purposes -- we can't fix that
4011 * by pulling tasks to us. Be careful of negative numbers as they'll
4012 * appear as very large values with unsigned longs.
4014 if (sds.max_load <= sds.busiest_load_per_task)
4017 /* Looks like there is an imbalance. Compute it */
4018 calculate_imbalance(&sds, this_cpu, imbalance);
4023 * There is no obvious imbalance. But check if we can do some balancing
4026 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4034 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4037 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4038 unsigned long imbalance, const struct cpumask *cpus)
4040 struct rq *busiest = NULL, *rq;
4041 unsigned long max_load = 0;
4044 for_each_cpu(i, sched_group_cpus(group)) {
4045 unsigned long power = power_of(i);
4046 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4049 if (!cpumask_test_cpu(i, cpus))
4053 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4056 if (capacity && rq->nr_running == 1 && wl > imbalance)
4059 if (wl > max_load) {
4069 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4070 * so long as it is large enough.
4072 #define MAX_PINNED_INTERVAL 512
4074 /* Working cpumask for load_balance and load_balance_newidle. */
4075 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4078 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4079 * tasks if there is an imbalance.
4081 static int load_balance(int this_cpu, struct rq *this_rq,
4082 struct sched_domain *sd, enum cpu_idle_type idle,
4085 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4086 struct sched_group *group;
4087 unsigned long imbalance;
4089 unsigned long flags;
4090 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4092 cpumask_setall(cpus);
4095 * When power savings policy is enabled for the parent domain, idle
4096 * sibling can pick up load irrespective of busy siblings. In this case,
4097 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4098 * portraying it as CPU_NOT_IDLE.
4100 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4101 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4104 schedstat_inc(sd, lb_count[idle]);
4108 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4115 schedstat_inc(sd, lb_nobusyg[idle]);
4119 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4121 schedstat_inc(sd, lb_nobusyq[idle]);
4125 BUG_ON(busiest == this_rq);
4127 schedstat_add(sd, lb_imbalance[idle], imbalance);
4130 if (busiest->nr_running > 1) {
4132 * Attempt to move tasks. If find_busiest_group has found
4133 * an imbalance but busiest->nr_running <= 1, the group is
4134 * still unbalanced. ld_moved simply stays zero, so it is
4135 * correctly treated as an imbalance.
4137 local_irq_save(flags);
4138 double_rq_lock(this_rq, busiest);
4139 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4140 imbalance, sd, idle, &all_pinned);
4141 double_rq_unlock(this_rq, busiest);
4142 local_irq_restore(flags);
4145 * some other cpu did the load balance for us.
4147 if (ld_moved && this_cpu != smp_processor_id())
4148 resched_cpu(this_cpu);
4150 /* All tasks on this runqueue were pinned by CPU affinity */
4151 if (unlikely(all_pinned)) {
4152 cpumask_clear_cpu(cpu_of(busiest), cpus);
4153 if (!cpumask_empty(cpus))
4160 schedstat_inc(sd, lb_failed[idle]);
4161 sd->nr_balance_failed++;
4163 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4165 spin_lock_irqsave(&busiest->lock, flags);
4167 /* don't kick the migration_thread, if the curr
4168 * task on busiest cpu can't be moved to this_cpu
4170 if (!cpumask_test_cpu(this_cpu,
4171 &busiest->curr->cpus_allowed)) {
4172 spin_unlock_irqrestore(&busiest->lock, flags);
4174 goto out_one_pinned;
4177 if (!busiest->active_balance) {
4178 busiest->active_balance = 1;
4179 busiest->push_cpu = this_cpu;
4182 spin_unlock_irqrestore(&busiest->lock, flags);
4184 wake_up_process(busiest->migration_thread);
4187 * We've kicked active balancing, reset the failure
4190 sd->nr_balance_failed = sd->cache_nice_tries+1;
4193 sd->nr_balance_failed = 0;
4195 if (likely(!active_balance)) {
4196 /* We were unbalanced, so reset the balancing interval */
4197 sd->balance_interval = sd->min_interval;
4200 * If we've begun active balancing, start to back off. This
4201 * case may not be covered by the all_pinned logic if there
4202 * is only 1 task on the busy runqueue (because we don't call
4205 if (sd->balance_interval < sd->max_interval)
4206 sd->balance_interval *= 2;
4209 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4210 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4216 schedstat_inc(sd, lb_balanced[idle]);
4218 sd->nr_balance_failed = 0;
4221 /* tune up the balancing interval */
4222 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4223 (sd->balance_interval < sd->max_interval))
4224 sd->balance_interval *= 2;
4226 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4227 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4238 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4239 * tasks if there is an imbalance.
4241 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4242 * this_rq is locked.
4245 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4247 struct sched_group *group;
4248 struct rq *busiest = NULL;
4249 unsigned long imbalance;
4253 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4255 cpumask_setall(cpus);
4258 * When power savings policy is enabled for the parent domain, idle
4259 * sibling can pick up load irrespective of busy siblings. In this case,
4260 * let the state of idle sibling percolate up as IDLE, instead of
4261 * portraying it as CPU_NOT_IDLE.
4263 if (sd->flags & SD_SHARE_CPUPOWER &&
4264 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4267 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4269 update_shares_locked(this_rq, sd);
4270 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4271 &sd_idle, cpus, NULL);
4273 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4277 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4279 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4283 BUG_ON(busiest == this_rq);
4285 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4288 if (busiest->nr_running > 1) {
4289 /* Attempt to move tasks */
4290 double_lock_balance(this_rq, busiest);
4291 /* this_rq->clock is already updated */
4292 update_rq_clock(busiest);
4293 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4294 imbalance, sd, CPU_NEWLY_IDLE,
4296 double_unlock_balance(this_rq, busiest);
4298 if (unlikely(all_pinned)) {
4299 cpumask_clear_cpu(cpu_of(busiest), cpus);
4300 if (!cpumask_empty(cpus))
4306 int active_balance = 0;
4308 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4309 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4310 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4313 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4316 if (sd->nr_balance_failed++ < 2)
4320 * The only task running in a non-idle cpu can be moved to this
4321 * cpu in an attempt to completely freeup the other CPU
4322 * package. The same method used to move task in load_balance()
4323 * have been extended for load_balance_newidle() to speedup
4324 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4326 * The package power saving logic comes from
4327 * find_busiest_group(). If there are no imbalance, then
4328 * f_b_g() will return NULL. However when sched_mc={1,2} then
4329 * f_b_g() will select a group from which a running task may be
4330 * pulled to this cpu in order to make the other package idle.
4331 * If there is no opportunity to make a package idle and if
4332 * there are no imbalance, then f_b_g() will return NULL and no
4333 * action will be taken in load_balance_newidle().
4335 * Under normal task pull operation due to imbalance, there
4336 * will be more than one task in the source run queue and
4337 * move_tasks() will succeed. ld_moved will be true and this
4338 * active balance code will not be triggered.
4341 /* Lock busiest in correct order while this_rq is held */
4342 double_lock_balance(this_rq, busiest);
4345 * don't kick the migration_thread, if the curr
4346 * task on busiest cpu can't be moved to this_cpu
4348 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4349 double_unlock_balance(this_rq, busiest);
4354 if (!busiest->active_balance) {
4355 busiest->active_balance = 1;
4356 busiest->push_cpu = this_cpu;
4360 double_unlock_balance(this_rq, busiest);
4362 * Should not call ttwu while holding a rq->lock
4364 spin_unlock(&this_rq->lock);
4366 wake_up_process(busiest->migration_thread);
4367 spin_lock(&this_rq->lock);
4370 sd->nr_balance_failed = 0;
4372 update_shares_locked(this_rq, sd);
4376 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4377 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4378 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4380 sd->nr_balance_failed = 0;
4386 * idle_balance is called by schedule() if this_cpu is about to become
4387 * idle. Attempts to pull tasks from other CPUs.
4389 static void idle_balance(int this_cpu, struct rq *this_rq)
4391 struct sched_domain *sd;
4392 int pulled_task = 0;
4393 unsigned long next_balance = jiffies + HZ;
4395 for_each_domain(this_cpu, sd) {
4396 unsigned long interval;
4398 if (!(sd->flags & SD_LOAD_BALANCE))
4401 if (sd->flags & SD_BALANCE_NEWIDLE)
4402 /* If we've pulled tasks over stop searching: */
4403 pulled_task = load_balance_newidle(this_cpu, this_rq,
4406 interval = msecs_to_jiffies(sd->balance_interval);
4407 if (time_after(next_balance, sd->last_balance + interval))
4408 next_balance = sd->last_balance + interval;
4412 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4414 * We are going idle. next_balance may be set based on
4415 * a busy processor. So reset next_balance.
4417 this_rq->next_balance = next_balance;
4422 * active_load_balance is run by migration threads. It pushes running tasks
4423 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4424 * running on each physical CPU where possible, and avoids physical /
4425 * logical imbalances.
4427 * Called with busiest_rq locked.
4429 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4431 int target_cpu = busiest_rq->push_cpu;
4432 struct sched_domain *sd;
4433 struct rq *target_rq;
4435 /* Is there any task to move? */
4436 if (busiest_rq->nr_running <= 1)
4439 target_rq = cpu_rq(target_cpu);
4442 * This condition is "impossible", if it occurs
4443 * we need to fix it. Originally reported by
4444 * Bjorn Helgaas on a 128-cpu setup.
4446 BUG_ON(busiest_rq == target_rq);
4448 /* move a task from busiest_rq to target_rq */
4449 double_lock_balance(busiest_rq, target_rq);
4450 update_rq_clock(busiest_rq);
4451 update_rq_clock(target_rq);
4453 /* Search for an sd spanning us and the target CPU. */
4454 for_each_domain(target_cpu, sd) {
4455 if ((sd->flags & SD_LOAD_BALANCE) &&
4456 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4461 schedstat_inc(sd, alb_count);
4463 if (move_one_task(target_rq, target_cpu, busiest_rq,
4465 schedstat_inc(sd, alb_pushed);
4467 schedstat_inc(sd, alb_failed);
4469 double_unlock_balance(busiest_rq, target_rq);
4474 atomic_t load_balancer;
4475 cpumask_var_t cpu_mask;
4476 cpumask_var_t ilb_grp_nohz_mask;
4477 } nohz ____cacheline_aligned = {
4478 .load_balancer = ATOMIC_INIT(-1),
4481 int get_nohz_load_balancer(void)
4483 return atomic_read(&nohz.load_balancer);
4486 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4488 * lowest_flag_domain - Return lowest sched_domain containing flag.
4489 * @cpu: The cpu whose lowest level of sched domain is to
4491 * @flag: The flag to check for the lowest sched_domain
4492 * for the given cpu.
4494 * Returns the lowest sched_domain of a cpu which contains the given flag.
4496 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4498 struct sched_domain *sd;
4500 for_each_domain(cpu, sd)
4501 if (sd && (sd->flags & flag))
4508 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4509 * @cpu: The cpu whose domains we're iterating over.
4510 * @sd: variable holding the value of the power_savings_sd
4512 * @flag: The flag to filter the sched_domains to be iterated.
4514 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4515 * set, starting from the lowest sched_domain to the highest.
4517 #define for_each_flag_domain(cpu, sd, flag) \
4518 for (sd = lowest_flag_domain(cpu, flag); \
4519 (sd && (sd->flags & flag)); sd = sd->parent)
4522 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4523 * @ilb_group: group to be checked for semi-idleness
4525 * Returns: 1 if the group is semi-idle. 0 otherwise.
4527 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4528 * and atleast one non-idle CPU. This helper function checks if the given
4529 * sched_group is semi-idle or not.
4531 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4533 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4534 sched_group_cpus(ilb_group));
4537 * A sched_group is semi-idle when it has atleast one busy cpu
4538 * and atleast one idle cpu.
4540 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4543 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4549 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4550 * @cpu: The cpu which is nominating a new idle_load_balancer.
4552 * Returns: Returns the id of the idle load balancer if it exists,
4553 * Else, returns >= nr_cpu_ids.
4555 * This algorithm picks the idle load balancer such that it belongs to a
4556 * semi-idle powersavings sched_domain. The idea is to try and avoid
4557 * completely idle packages/cores just for the purpose of idle load balancing
4558 * when there are other idle cpu's which are better suited for that job.
4560 static int find_new_ilb(int cpu)
4562 struct sched_domain *sd;
4563 struct sched_group *ilb_group;
4566 * Have idle load balancer selection from semi-idle packages only
4567 * when power-aware load balancing is enabled
4569 if (!(sched_smt_power_savings || sched_mc_power_savings))
4573 * Optimize for the case when we have no idle CPUs or only one
4574 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4576 if (cpumask_weight(nohz.cpu_mask) < 2)
4579 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4580 ilb_group = sd->groups;
4583 if (is_semi_idle_group(ilb_group))
4584 return cpumask_first(nohz.ilb_grp_nohz_mask);
4586 ilb_group = ilb_group->next;
4588 } while (ilb_group != sd->groups);
4592 return cpumask_first(nohz.cpu_mask);
4594 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4595 static inline int find_new_ilb(int call_cpu)
4597 return cpumask_first(nohz.cpu_mask);
4602 * This routine will try to nominate the ilb (idle load balancing)
4603 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4604 * load balancing on behalf of all those cpus. If all the cpus in the system
4605 * go into this tickless mode, then there will be no ilb owner (as there is
4606 * no need for one) and all the cpus will sleep till the next wakeup event
4609 * For the ilb owner, tick is not stopped. And this tick will be used
4610 * for idle load balancing. ilb owner will still be part of
4613 * While stopping the tick, this cpu will become the ilb owner if there
4614 * is no other owner. And will be the owner till that cpu becomes busy
4615 * or if all cpus in the system stop their ticks at which point
4616 * there is no need for ilb owner.
4618 * When the ilb owner becomes busy, it nominates another owner, during the
4619 * next busy scheduler_tick()
4621 int select_nohz_load_balancer(int stop_tick)
4623 int cpu = smp_processor_id();
4626 cpu_rq(cpu)->in_nohz_recently = 1;
4628 if (!cpu_active(cpu)) {
4629 if (atomic_read(&nohz.load_balancer) != cpu)
4633 * If we are going offline and still the leader,
4636 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4642 cpumask_set_cpu(cpu, nohz.cpu_mask);
4644 /* time for ilb owner also to sleep */
4645 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4646 if (atomic_read(&nohz.load_balancer) == cpu)
4647 atomic_set(&nohz.load_balancer, -1);
4651 if (atomic_read(&nohz.load_balancer) == -1) {
4652 /* make me the ilb owner */
4653 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4655 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4658 if (!(sched_smt_power_savings ||
4659 sched_mc_power_savings))
4662 * Check to see if there is a more power-efficient
4665 new_ilb = find_new_ilb(cpu);
4666 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4667 atomic_set(&nohz.load_balancer, -1);
4668 resched_cpu(new_ilb);
4674 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4677 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4679 if (atomic_read(&nohz.load_balancer) == cpu)
4680 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4687 static DEFINE_SPINLOCK(balancing);
4690 * It checks each scheduling domain to see if it is due to be balanced,
4691 * and initiates a balancing operation if so.
4693 * Balancing parameters are set up in arch_init_sched_domains.
4695 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4698 struct rq *rq = cpu_rq(cpu);
4699 unsigned long interval;
4700 struct sched_domain *sd;
4701 /* Earliest time when we have to do rebalance again */
4702 unsigned long next_balance = jiffies + 60*HZ;
4703 int update_next_balance = 0;
4706 for_each_domain(cpu, sd) {
4707 if (!(sd->flags & SD_LOAD_BALANCE))
4710 interval = sd->balance_interval;
4711 if (idle != CPU_IDLE)
4712 interval *= sd->busy_factor;
4714 /* scale ms to jiffies */
4715 interval = msecs_to_jiffies(interval);
4716 if (unlikely(!interval))
4718 if (interval > HZ*NR_CPUS/10)
4719 interval = HZ*NR_CPUS/10;
4721 need_serialize = sd->flags & SD_SERIALIZE;
4723 if (need_serialize) {
4724 if (!spin_trylock(&balancing))
4728 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4729 if (load_balance(cpu, rq, sd, idle, &balance)) {
4731 * We've pulled tasks over so either we're no
4732 * longer idle, or one of our SMT siblings is
4735 idle = CPU_NOT_IDLE;
4737 sd->last_balance = jiffies;
4740 spin_unlock(&balancing);
4742 if (time_after(next_balance, sd->last_balance + interval)) {
4743 next_balance = sd->last_balance + interval;
4744 update_next_balance = 1;
4748 * Stop the load balance at this level. There is another
4749 * CPU in our sched group which is doing load balancing more
4757 * next_balance will be updated only when there is a need.
4758 * When the cpu is attached to null domain for ex, it will not be
4761 if (likely(update_next_balance))
4762 rq->next_balance = next_balance;
4766 * run_rebalance_domains is triggered when needed from the scheduler tick.
4767 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4768 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4770 static void run_rebalance_domains(struct softirq_action *h)
4772 int this_cpu = smp_processor_id();
4773 struct rq *this_rq = cpu_rq(this_cpu);
4774 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4775 CPU_IDLE : CPU_NOT_IDLE;
4777 rebalance_domains(this_cpu, idle);
4781 * If this cpu is the owner for idle load balancing, then do the
4782 * balancing on behalf of the other idle cpus whose ticks are
4785 if (this_rq->idle_at_tick &&
4786 atomic_read(&nohz.load_balancer) == this_cpu) {
4790 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4791 if (balance_cpu == this_cpu)
4795 * If this cpu gets work to do, stop the load balancing
4796 * work being done for other cpus. Next load
4797 * balancing owner will pick it up.
4802 rebalance_domains(balance_cpu, CPU_IDLE);
4804 rq = cpu_rq(balance_cpu);
4805 if (time_after(this_rq->next_balance, rq->next_balance))
4806 this_rq->next_balance = rq->next_balance;
4812 static inline int on_null_domain(int cpu)
4814 return !rcu_dereference(cpu_rq(cpu)->sd);
4818 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4820 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4821 * idle load balancing owner or decide to stop the periodic load balancing,
4822 * if the whole system is idle.
4824 static inline void trigger_load_balance(struct rq *rq, int cpu)
4828 * If we were in the nohz mode recently and busy at the current
4829 * scheduler tick, then check if we need to nominate new idle
4832 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4833 rq->in_nohz_recently = 0;
4835 if (atomic_read(&nohz.load_balancer) == cpu) {
4836 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4837 atomic_set(&nohz.load_balancer, -1);
4840 if (atomic_read(&nohz.load_balancer) == -1) {
4841 int ilb = find_new_ilb(cpu);
4843 if (ilb < nr_cpu_ids)
4849 * If this cpu is idle and doing idle load balancing for all the
4850 * cpus with ticks stopped, is it time for that to stop?
4852 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4853 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4859 * If this cpu is idle and the idle load balancing is done by
4860 * someone else, then no need raise the SCHED_SOFTIRQ
4862 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4863 cpumask_test_cpu(cpu, nohz.cpu_mask))
4866 /* Don't need to rebalance while attached to NULL domain */
4867 if (time_after_eq(jiffies, rq->next_balance) &&
4868 likely(!on_null_domain(cpu)))
4869 raise_softirq(SCHED_SOFTIRQ);
4872 #else /* CONFIG_SMP */
4875 * on UP we do not need to balance between CPUs:
4877 static inline void idle_balance(int cpu, struct rq *rq)
4883 DEFINE_PER_CPU(struct kernel_stat, kstat);
4885 EXPORT_PER_CPU_SYMBOL(kstat);
4888 * Return any ns on the sched_clock that have not yet been accounted in
4889 * @p in case that task is currently running.
4891 * Called with task_rq_lock() held on @rq.
4893 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4897 if (task_current(rq, p)) {
4898 update_rq_clock(rq);
4899 ns = rq->clock - p->se.exec_start;
4907 unsigned long long task_delta_exec(struct task_struct *p)
4909 unsigned long flags;
4913 rq = task_rq_lock(p, &flags);
4914 ns = do_task_delta_exec(p, rq);
4915 task_rq_unlock(rq, &flags);
4921 * Return accounted runtime for the task.
4922 * In case the task is currently running, return the runtime plus current's
4923 * pending runtime that have not been accounted yet.
4925 unsigned long long task_sched_runtime(struct task_struct *p)
4927 unsigned long flags;
4931 rq = task_rq_lock(p, &flags);
4932 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4933 task_rq_unlock(rq, &flags);
4939 * Return sum_exec_runtime for the thread group.
4940 * In case the task is currently running, return the sum plus current's
4941 * pending runtime that have not been accounted yet.
4943 * Note that the thread group might have other running tasks as well,
4944 * so the return value not includes other pending runtime that other
4945 * running tasks might have.
4947 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4949 struct task_cputime totals;
4950 unsigned long flags;
4954 rq = task_rq_lock(p, &flags);
4955 thread_group_cputime(p, &totals);
4956 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4957 task_rq_unlock(rq, &flags);
4963 * Account user cpu time to a process.
4964 * @p: the process that the cpu time gets accounted to
4965 * @cputime: the cpu time spent in user space since the last update
4966 * @cputime_scaled: cputime scaled by cpu frequency
4968 void account_user_time(struct task_struct *p, cputime_t cputime,
4969 cputime_t cputime_scaled)
4971 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4974 /* Add user time to process. */
4975 p->utime = cputime_add(p->utime, cputime);
4976 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4977 account_group_user_time(p, cputime);
4979 /* Add user time to cpustat. */
4980 tmp = cputime_to_cputime64(cputime);
4981 if (TASK_NICE(p) > 0)
4982 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4984 cpustat->user = cputime64_add(cpustat->user, tmp);
4986 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4987 /* Account for user time used */
4988 acct_update_integrals(p);
4992 * Account guest cpu time to a process.
4993 * @p: the process that the cpu time gets accounted to
4994 * @cputime: the cpu time spent in virtual machine since the last update
4995 * @cputime_scaled: cputime scaled by cpu frequency
4997 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4998 cputime_t cputime_scaled)
5001 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5003 tmp = cputime_to_cputime64(cputime);
5005 /* Add guest time to process. */
5006 p->utime = cputime_add(p->utime, cputime);
5007 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5008 account_group_user_time(p, cputime);
5009 p->gtime = cputime_add(p->gtime, cputime);
5011 /* Add guest time to cpustat. */
5012 cpustat->user = cputime64_add(cpustat->user, tmp);
5013 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5017 * Account system cpu time to a process.
5018 * @p: the process that the cpu time gets accounted to
5019 * @hardirq_offset: the offset to subtract from hardirq_count()
5020 * @cputime: the cpu time spent in kernel space since the last update
5021 * @cputime_scaled: cputime scaled by cpu frequency
5023 void account_system_time(struct task_struct *p, int hardirq_offset,
5024 cputime_t cputime, cputime_t cputime_scaled)
5026 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5029 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5030 account_guest_time(p, cputime, cputime_scaled);
5034 /* Add system time to process. */
5035 p->stime = cputime_add(p->stime, cputime);
5036 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5037 account_group_system_time(p, cputime);
5039 /* Add system time to cpustat. */
5040 tmp = cputime_to_cputime64(cputime);
5041 if (hardirq_count() - hardirq_offset)
5042 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5043 else if (softirq_count())
5044 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5046 cpustat->system = cputime64_add(cpustat->system, tmp);
5048 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5050 /* Account for system time used */
5051 acct_update_integrals(p);
5055 * Account for involuntary wait time.
5056 * @steal: the cpu time spent in involuntary wait
5058 void account_steal_time(cputime_t cputime)
5060 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5061 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5063 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5067 * Account for idle time.
5068 * @cputime: the cpu time spent in idle wait
5070 void account_idle_time(cputime_t cputime)
5072 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5073 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5074 struct rq *rq = this_rq();
5076 if (atomic_read(&rq->nr_iowait) > 0)
5077 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5079 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5082 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5085 * Account a single tick of cpu time.
5086 * @p: the process that the cpu time gets accounted to
5087 * @user_tick: indicates if the tick is a user or a system tick
5089 void account_process_tick(struct task_struct *p, int user_tick)
5091 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5092 struct rq *rq = this_rq();
5095 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5096 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5097 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5100 account_idle_time(cputime_one_jiffy);
5104 * Account multiple ticks of steal time.
5105 * @p: the process from which the cpu time has been stolen
5106 * @ticks: number of stolen ticks
5108 void account_steal_ticks(unsigned long ticks)
5110 account_steal_time(jiffies_to_cputime(ticks));
5114 * Account multiple ticks of idle time.
5115 * @ticks: number of stolen ticks
5117 void account_idle_ticks(unsigned long ticks)
5119 account_idle_time(jiffies_to_cputime(ticks));
5125 * Use precise platform statistics if available:
5127 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5128 cputime_t task_utime(struct task_struct *p)
5133 cputime_t task_stime(struct task_struct *p)
5138 cputime_t task_utime(struct task_struct *p)
5140 clock_t utime = cputime_to_clock_t(p->utime),
5141 total = utime + cputime_to_clock_t(p->stime);
5145 * Use CFS's precise accounting:
5147 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5151 do_div(temp, total);
5153 utime = (clock_t)temp;
5155 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5156 return p->prev_utime;
5159 cputime_t task_stime(struct task_struct *p)
5164 * Use CFS's precise accounting. (we subtract utime from
5165 * the total, to make sure the total observed by userspace
5166 * grows monotonically - apps rely on that):
5168 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5169 cputime_to_clock_t(task_utime(p));
5172 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5174 return p->prev_stime;
5178 inline cputime_t task_gtime(struct task_struct *p)
5184 * This function gets called by the timer code, with HZ frequency.
5185 * We call it with interrupts disabled.
5187 * It also gets called by the fork code, when changing the parent's
5190 void scheduler_tick(void)
5192 int cpu = smp_processor_id();
5193 struct rq *rq = cpu_rq(cpu);
5194 struct task_struct *curr = rq->curr;
5198 spin_lock(&rq->lock);
5199 update_rq_clock(rq);
5200 update_cpu_load(rq);
5201 curr->sched_class->task_tick(rq, curr, 0);
5202 spin_unlock(&rq->lock);
5204 perf_event_task_tick(curr, cpu);
5207 rq->idle_at_tick = idle_cpu(cpu);
5208 trigger_load_balance(rq, cpu);
5212 notrace unsigned long get_parent_ip(unsigned long addr)
5214 if (in_lock_functions(addr)) {
5215 addr = CALLER_ADDR2;
5216 if (in_lock_functions(addr))
5217 addr = CALLER_ADDR3;
5222 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5223 defined(CONFIG_PREEMPT_TRACER))
5225 void __kprobes add_preempt_count(int val)
5227 #ifdef CONFIG_DEBUG_PREEMPT
5231 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5234 preempt_count() += val;
5235 #ifdef CONFIG_DEBUG_PREEMPT
5237 * Spinlock count overflowing soon?
5239 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5242 if (preempt_count() == val)
5243 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5245 EXPORT_SYMBOL(add_preempt_count);
5247 void __kprobes sub_preempt_count(int val)
5249 #ifdef CONFIG_DEBUG_PREEMPT
5253 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5256 * Is the spinlock portion underflowing?
5258 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5259 !(preempt_count() & PREEMPT_MASK)))
5263 if (preempt_count() == val)
5264 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5265 preempt_count() -= val;
5267 EXPORT_SYMBOL(sub_preempt_count);
5272 * Print scheduling while atomic bug:
5274 static noinline void __schedule_bug(struct task_struct *prev)
5276 struct pt_regs *regs = get_irq_regs();
5278 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5279 prev->comm, prev->pid, preempt_count());
5281 debug_show_held_locks(prev);
5283 if (irqs_disabled())
5284 print_irqtrace_events(prev);
5293 * Various schedule()-time debugging checks and statistics:
5295 static inline void schedule_debug(struct task_struct *prev)
5298 * Test if we are atomic. Since do_exit() needs to call into
5299 * schedule() atomically, we ignore that path for now.
5300 * Otherwise, whine if we are scheduling when we should not be.
5302 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5303 __schedule_bug(prev);
5305 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5307 schedstat_inc(this_rq(), sched_count);
5308 #ifdef CONFIG_SCHEDSTATS
5309 if (unlikely(prev->lock_depth >= 0)) {
5310 schedstat_inc(this_rq(), bkl_count);
5311 schedstat_inc(prev, sched_info.bkl_count);
5316 static void put_prev_task(struct rq *rq, struct task_struct *p)
5318 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
5320 update_avg(&p->se.avg_running, runtime);
5322 if (p->state == TASK_RUNNING) {
5324 * In order to avoid avg_overlap growing stale when we are
5325 * indeed overlapping and hence not getting put to sleep, grow
5326 * the avg_overlap on preemption.
5328 * We use the average preemption runtime because that
5329 * correlates to the amount of cache footprint a task can
5332 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5333 update_avg(&p->se.avg_overlap, runtime);
5335 update_avg(&p->se.avg_running, 0);
5337 p->sched_class->put_prev_task(rq, p);
5341 * Pick up the highest-prio task:
5343 static inline struct task_struct *
5344 pick_next_task(struct rq *rq)
5346 const struct sched_class *class;
5347 struct task_struct *p;
5350 * Optimization: we know that if all tasks are in
5351 * the fair class we can call that function directly:
5353 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5354 p = fair_sched_class.pick_next_task(rq);
5359 class = sched_class_highest;
5361 p = class->pick_next_task(rq);
5365 * Will never be NULL as the idle class always
5366 * returns a non-NULL p:
5368 class = class->next;
5373 * schedule() is the main scheduler function.
5375 asmlinkage void __sched schedule(void)
5377 struct task_struct *prev, *next;
5378 unsigned long *switch_count;
5384 cpu = smp_processor_id();
5388 switch_count = &prev->nivcsw;
5390 release_kernel_lock(prev);
5391 need_resched_nonpreemptible:
5393 schedule_debug(prev);
5395 if (sched_feat(HRTICK))
5398 spin_lock_irq(&rq->lock);
5399 update_rq_clock(rq);
5400 clear_tsk_need_resched(prev);
5402 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5403 if (unlikely(signal_pending_state(prev->state, prev)))
5404 prev->state = TASK_RUNNING;
5406 deactivate_task(rq, prev, 1);
5407 switch_count = &prev->nvcsw;
5410 pre_schedule(rq, prev);
5412 if (unlikely(!rq->nr_running))
5413 idle_balance(cpu, rq);
5415 put_prev_task(rq, prev);
5416 next = pick_next_task(rq);
5418 if (likely(prev != next)) {
5419 sched_info_switch(prev, next);
5420 perf_event_task_sched_out(prev, next, cpu);
5426 context_switch(rq, prev, next); /* unlocks the rq */
5428 * the context switch might have flipped the stack from under
5429 * us, hence refresh the local variables.
5431 cpu = smp_processor_id();
5434 spin_unlock_irq(&rq->lock);
5438 if (unlikely(reacquire_kernel_lock(current) < 0))
5439 goto need_resched_nonpreemptible;
5441 preempt_enable_no_resched();
5445 EXPORT_SYMBOL(schedule);
5449 * Look out! "owner" is an entirely speculative pointer
5450 * access and not reliable.
5452 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5457 if (!sched_feat(OWNER_SPIN))
5460 #ifdef CONFIG_DEBUG_PAGEALLOC
5462 * Need to access the cpu field knowing that
5463 * DEBUG_PAGEALLOC could have unmapped it if
5464 * the mutex owner just released it and exited.
5466 if (probe_kernel_address(&owner->cpu, cpu))
5473 * Even if the access succeeded (likely case),
5474 * the cpu field may no longer be valid.
5476 if (cpu >= nr_cpumask_bits)
5480 * We need to validate that we can do a
5481 * get_cpu() and that we have the percpu area.
5483 if (!cpu_online(cpu))
5490 * Owner changed, break to re-assess state.
5492 if (lock->owner != owner)
5496 * Is that owner really running on that cpu?
5498 if (task_thread_info(rq->curr) != owner || need_resched())
5508 #ifdef CONFIG_PREEMPT
5510 * this is the entry point to schedule() from in-kernel preemption
5511 * off of preempt_enable. Kernel preemptions off return from interrupt
5512 * occur there and call schedule directly.
5514 asmlinkage void __sched preempt_schedule(void)
5516 struct thread_info *ti = current_thread_info();
5519 * If there is a non-zero preempt_count or interrupts are disabled,
5520 * we do not want to preempt the current task. Just return..
5522 if (likely(ti->preempt_count || irqs_disabled()))
5526 add_preempt_count(PREEMPT_ACTIVE);
5528 sub_preempt_count(PREEMPT_ACTIVE);
5531 * Check again in case we missed a preemption opportunity
5532 * between schedule and now.
5535 } while (need_resched());
5537 EXPORT_SYMBOL(preempt_schedule);
5540 * this is the entry point to schedule() from kernel preemption
5541 * off of irq context.
5542 * Note, that this is called and return with irqs disabled. This will
5543 * protect us against recursive calling from irq.
5545 asmlinkage void __sched preempt_schedule_irq(void)
5547 struct thread_info *ti = current_thread_info();
5549 /* Catch callers which need to be fixed */
5550 BUG_ON(ti->preempt_count || !irqs_disabled());
5553 add_preempt_count(PREEMPT_ACTIVE);
5556 local_irq_disable();
5557 sub_preempt_count(PREEMPT_ACTIVE);
5560 * Check again in case we missed a preemption opportunity
5561 * between schedule and now.
5564 } while (need_resched());
5567 #endif /* CONFIG_PREEMPT */
5569 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5572 return try_to_wake_up(curr->private, mode, wake_flags);
5574 EXPORT_SYMBOL(default_wake_function);
5577 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5578 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5579 * number) then we wake all the non-exclusive tasks and one exclusive task.
5581 * There are circumstances in which we can try to wake a task which has already
5582 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5583 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5585 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5586 int nr_exclusive, int wake_flags, void *key)
5588 wait_queue_t *curr, *next;
5590 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5591 unsigned flags = curr->flags;
5593 if (curr->func(curr, mode, wake_flags, key) &&
5594 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5600 * __wake_up - wake up threads blocked on a waitqueue.
5602 * @mode: which threads
5603 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5604 * @key: is directly passed to the wakeup function
5606 * It may be assumed that this function implies a write memory barrier before
5607 * changing the task state if and only if any tasks are woken up.
5609 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5610 int nr_exclusive, void *key)
5612 unsigned long flags;
5614 spin_lock_irqsave(&q->lock, flags);
5615 __wake_up_common(q, mode, nr_exclusive, 0, key);
5616 spin_unlock_irqrestore(&q->lock, flags);
5618 EXPORT_SYMBOL(__wake_up);
5621 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5623 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5625 __wake_up_common(q, mode, 1, 0, NULL);
5628 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5630 __wake_up_common(q, mode, 1, 0, key);
5634 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5636 * @mode: which threads
5637 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5638 * @key: opaque value to be passed to wakeup targets
5640 * The sync wakeup differs that the waker knows that it will schedule
5641 * away soon, so while the target thread will be woken up, it will not
5642 * be migrated to another CPU - ie. the two threads are 'synchronized'
5643 * with each other. This can prevent needless bouncing between CPUs.
5645 * On UP it can prevent extra preemption.
5647 * It may be assumed that this function implies a write memory barrier before
5648 * changing the task state if and only if any tasks are woken up.
5650 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5651 int nr_exclusive, void *key)
5653 unsigned long flags;
5654 int wake_flags = WF_SYNC;
5659 if (unlikely(!nr_exclusive))
5662 spin_lock_irqsave(&q->lock, flags);
5663 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5664 spin_unlock_irqrestore(&q->lock, flags);
5666 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5669 * __wake_up_sync - see __wake_up_sync_key()
5671 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5673 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5675 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5678 * complete: - signals a single thread waiting on this completion
5679 * @x: holds the state of this particular completion
5681 * This will wake up a single thread waiting on this completion. Threads will be
5682 * awakened in the same order in which they were queued.
5684 * See also complete_all(), wait_for_completion() and related routines.
5686 * It may be assumed that this function implies a write memory barrier before
5687 * changing the task state if and only if any tasks are woken up.
5689 void complete(struct completion *x)
5691 unsigned long flags;
5693 spin_lock_irqsave(&x->wait.lock, flags);
5695 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5696 spin_unlock_irqrestore(&x->wait.lock, flags);
5698 EXPORT_SYMBOL(complete);
5701 * complete_all: - signals all threads waiting on this completion
5702 * @x: holds the state of this particular completion
5704 * This will wake up all threads waiting on this particular completion event.
5706 * It may be assumed that this function implies a write memory barrier before
5707 * changing the task state if and only if any tasks are woken up.
5709 void complete_all(struct completion *x)
5711 unsigned long flags;
5713 spin_lock_irqsave(&x->wait.lock, flags);
5714 x->done += UINT_MAX/2;
5715 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5716 spin_unlock_irqrestore(&x->wait.lock, flags);
5718 EXPORT_SYMBOL(complete_all);
5720 static inline long __sched
5721 do_wait_for_common(struct completion *x, long timeout, int state)
5724 DECLARE_WAITQUEUE(wait, current);
5726 wait.flags |= WQ_FLAG_EXCLUSIVE;
5727 __add_wait_queue_tail(&x->wait, &wait);
5729 if (signal_pending_state(state, current)) {
5730 timeout = -ERESTARTSYS;
5733 __set_current_state(state);
5734 spin_unlock_irq(&x->wait.lock);
5735 timeout = schedule_timeout(timeout);
5736 spin_lock_irq(&x->wait.lock);
5737 } while (!x->done && timeout);
5738 __remove_wait_queue(&x->wait, &wait);
5743 return timeout ?: 1;
5747 wait_for_common(struct completion *x, long timeout, int state)
5751 spin_lock_irq(&x->wait.lock);
5752 timeout = do_wait_for_common(x, timeout, state);
5753 spin_unlock_irq(&x->wait.lock);
5758 * wait_for_completion: - waits for completion of a task
5759 * @x: holds the state of this particular completion
5761 * This waits to be signaled for completion of a specific task. It is NOT
5762 * interruptible and there is no timeout.
5764 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5765 * and interrupt capability. Also see complete().
5767 void __sched wait_for_completion(struct completion *x)
5769 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5771 EXPORT_SYMBOL(wait_for_completion);
5774 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5775 * @x: holds the state of this particular completion
5776 * @timeout: timeout value in jiffies
5778 * This waits for either a completion of a specific task to be signaled or for a
5779 * specified timeout to expire. The timeout is in jiffies. It is not
5782 unsigned long __sched
5783 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5785 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5787 EXPORT_SYMBOL(wait_for_completion_timeout);
5790 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5791 * @x: holds the state of this particular completion
5793 * This waits for completion of a specific task to be signaled. It is
5796 int __sched wait_for_completion_interruptible(struct completion *x)
5798 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5799 if (t == -ERESTARTSYS)
5803 EXPORT_SYMBOL(wait_for_completion_interruptible);
5806 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5807 * @x: holds the state of this particular completion
5808 * @timeout: timeout value in jiffies
5810 * This waits for either a completion of a specific task to be signaled or for a
5811 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5813 unsigned long __sched
5814 wait_for_completion_interruptible_timeout(struct completion *x,
5815 unsigned long timeout)
5817 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5819 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5822 * wait_for_completion_killable: - waits for completion of a task (killable)
5823 * @x: holds the state of this particular completion
5825 * This waits to be signaled for completion of a specific task. It can be
5826 * interrupted by a kill signal.
5828 int __sched wait_for_completion_killable(struct completion *x)
5830 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5831 if (t == -ERESTARTSYS)
5835 EXPORT_SYMBOL(wait_for_completion_killable);
5838 * try_wait_for_completion - try to decrement a completion without blocking
5839 * @x: completion structure
5841 * Returns: 0 if a decrement cannot be done without blocking
5842 * 1 if a decrement succeeded.
5844 * If a completion is being used as a counting completion,
5845 * attempt to decrement the counter without blocking. This
5846 * enables us to avoid waiting if the resource the completion
5847 * is protecting is not available.
5849 bool try_wait_for_completion(struct completion *x)
5853 spin_lock_irq(&x->wait.lock);
5858 spin_unlock_irq(&x->wait.lock);
5861 EXPORT_SYMBOL(try_wait_for_completion);
5864 * completion_done - Test to see if a completion has any waiters
5865 * @x: completion structure
5867 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5868 * 1 if there are no waiters.
5871 bool completion_done(struct completion *x)
5875 spin_lock_irq(&x->wait.lock);
5878 spin_unlock_irq(&x->wait.lock);
5881 EXPORT_SYMBOL(completion_done);
5884 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5886 unsigned long flags;
5889 init_waitqueue_entry(&wait, current);
5891 __set_current_state(state);
5893 spin_lock_irqsave(&q->lock, flags);
5894 __add_wait_queue(q, &wait);
5895 spin_unlock(&q->lock);
5896 timeout = schedule_timeout(timeout);
5897 spin_lock_irq(&q->lock);
5898 __remove_wait_queue(q, &wait);
5899 spin_unlock_irqrestore(&q->lock, flags);
5904 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5906 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5908 EXPORT_SYMBOL(interruptible_sleep_on);
5911 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5913 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5915 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5917 void __sched sleep_on(wait_queue_head_t *q)
5919 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5921 EXPORT_SYMBOL(sleep_on);
5923 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5925 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5927 EXPORT_SYMBOL(sleep_on_timeout);
5929 #ifdef CONFIG_RT_MUTEXES
5932 * rt_mutex_setprio - set the current priority of a task
5934 * @prio: prio value (kernel-internal form)
5936 * This function changes the 'effective' priority of a task. It does
5937 * not touch ->normal_prio like __setscheduler().
5939 * Used by the rt_mutex code to implement priority inheritance logic.
5941 void rt_mutex_setprio(struct task_struct *p, int prio)
5943 unsigned long flags;
5944 int oldprio, on_rq, running;
5946 const struct sched_class *prev_class = p->sched_class;
5948 BUG_ON(prio < 0 || prio > MAX_PRIO);
5950 rq = task_rq_lock(p, &flags);
5951 update_rq_clock(rq);
5954 on_rq = p->se.on_rq;
5955 running = task_current(rq, p);
5957 dequeue_task(rq, p, 0);
5959 p->sched_class->put_prev_task(rq, p);
5962 p->sched_class = &rt_sched_class;
5964 p->sched_class = &fair_sched_class;
5969 p->sched_class->set_curr_task(rq);
5971 enqueue_task(rq, p, 0);
5973 check_class_changed(rq, p, prev_class, oldprio, running);
5975 task_rq_unlock(rq, &flags);
5980 void set_user_nice(struct task_struct *p, long nice)
5982 int old_prio, delta, on_rq;
5983 unsigned long flags;
5986 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5989 * We have to be careful, if called from sys_setpriority(),
5990 * the task might be in the middle of scheduling on another CPU.
5992 rq = task_rq_lock(p, &flags);
5993 update_rq_clock(rq);
5995 * The RT priorities are set via sched_setscheduler(), but we still
5996 * allow the 'normal' nice value to be set - but as expected
5997 * it wont have any effect on scheduling until the task is
5998 * SCHED_FIFO/SCHED_RR:
6000 if (task_has_rt_policy(p)) {
6001 p->static_prio = NICE_TO_PRIO(nice);
6004 on_rq = p->se.on_rq;
6006 dequeue_task(rq, p, 0);
6008 p->static_prio = NICE_TO_PRIO(nice);
6011 p->prio = effective_prio(p);
6012 delta = p->prio - old_prio;
6015 enqueue_task(rq, p, 0);
6017 * If the task increased its priority or is running and
6018 * lowered its priority, then reschedule its CPU:
6020 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6021 resched_task(rq->curr);
6024 task_rq_unlock(rq, &flags);
6026 EXPORT_SYMBOL(set_user_nice);
6029 * can_nice - check if a task can reduce its nice value
6033 int can_nice(const struct task_struct *p, const int nice)
6035 /* convert nice value [19,-20] to rlimit style value [1,40] */
6036 int nice_rlim = 20 - nice;
6038 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6039 capable(CAP_SYS_NICE));
6042 #ifdef __ARCH_WANT_SYS_NICE
6045 * sys_nice - change the priority of the current process.
6046 * @increment: priority increment
6048 * sys_setpriority is a more generic, but much slower function that
6049 * does similar things.
6051 SYSCALL_DEFINE1(nice, int, increment)
6056 * Setpriority might change our priority at the same moment.
6057 * We don't have to worry. Conceptually one call occurs first
6058 * and we have a single winner.
6060 if (increment < -40)
6065 nice = TASK_NICE(current) + increment;
6071 if (increment < 0 && !can_nice(current, nice))
6074 retval = security_task_setnice(current, nice);
6078 set_user_nice(current, nice);
6085 * task_prio - return the priority value of a given task.
6086 * @p: the task in question.
6088 * This is the priority value as seen by users in /proc.
6089 * RT tasks are offset by -200. Normal tasks are centered
6090 * around 0, value goes from -16 to +15.
6092 int task_prio(const struct task_struct *p)
6094 return p->prio - MAX_RT_PRIO;
6098 * task_nice - return the nice value of a given task.
6099 * @p: the task in question.
6101 int task_nice(const struct task_struct *p)
6103 return TASK_NICE(p);
6105 EXPORT_SYMBOL(task_nice);
6108 * idle_cpu - is a given cpu idle currently?
6109 * @cpu: the processor in question.
6111 int idle_cpu(int cpu)
6113 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6117 * idle_task - return the idle task for a given cpu.
6118 * @cpu: the processor in question.
6120 struct task_struct *idle_task(int cpu)
6122 return cpu_rq(cpu)->idle;
6126 * find_process_by_pid - find a process with a matching PID value.
6127 * @pid: the pid in question.
6129 static struct task_struct *find_process_by_pid(pid_t pid)
6131 return pid ? find_task_by_vpid(pid) : current;
6134 /* Actually do priority change: must hold rq lock. */
6136 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6138 BUG_ON(p->se.on_rq);
6141 switch (p->policy) {
6145 p->sched_class = &fair_sched_class;
6149 p->sched_class = &rt_sched_class;
6153 p->rt_priority = prio;
6154 p->normal_prio = normal_prio(p);
6155 /* we are holding p->pi_lock already */
6156 p->prio = rt_mutex_getprio(p);
6161 * check the target process has a UID that matches the current process's
6163 static bool check_same_owner(struct task_struct *p)
6165 const struct cred *cred = current_cred(), *pcred;
6169 pcred = __task_cred(p);
6170 match = (cred->euid == pcred->euid ||
6171 cred->euid == pcred->uid);
6176 static int __sched_setscheduler(struct task_struct *p, int policy,
6177 struct sched_param *param, bool user)
6179 int retval, oldprio, oldpolicy = -1, on_rq, running;
6180 unsigned long flags;
6181 const struct sched_class *prev_class = p->sched_class;
6185 /* may grab non-irq protected spin_locks */
6186 BUG_ON(in_interrupt());
6188 /* double check policy once rq lock held */
6190 reset_on_fork = p->sched_reset_on_fork;
6191 policy = oldpolicy = p->policy;
6193 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6194 policy &= ~SCHED_RESET_ON_FORK;
6196 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6197 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6198 policy != SCHED_IDLE)
6203 * Valid priorities for SCHED_FIFO and SCHED_RR are
6204 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6205 * SCHED_BATCH and SCHED_IDLE is 0.
6207 if (param->sched_priority < 0 ||
6208 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6209 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6211 if (rt_policy(policy) != (param->sched_priority != 0))
6215 * Allow unprivileged RT tasks to decrease priority:
6217 if (user && !capable(CAP_SYS_NICE)) {
6218 if (rt_policy(policy)) {
6219 unsigned long rlim_rtprio;
6221 if (!lock_task_sighand(p, &flags))
6223 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6224 unlock_task_sighand(p, &flags);
6226 /* can't set/change the rt policy */
6227 if (policy != p->policy && !rlim_rtprio)
6230 /* can't increase priority */
6231 if (param->sched_priority > p->rt_priority &&
6232 param->sched_priority > rlim_rtprio)
6236 * Like positive nice levels, dont allow tasks to
6237 * move out of SCHED_IDLE either:
6239 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6242 /* can't change other user's priorities */
6243 if (!check_same_owner(p))
6246 /* Normal users shall not reset the sched_reset_on_fork flag */
6247 if (p->sched_reset_on_fork && !reset_on_fork)
6252 #ifdef CONFIG_RT_GROUP_SCHED
6254 * Do not allow realtime tasks into groups that have no runtime
6257 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6258 task_group(p)->rt_bandwidth.rt_runtime == 0)
6262 retval = security_task_setscheduler(p, policy, param);
6268 * make sure no PI-waiters arrive (or leave) while we are
6269 * changing the priority of the task:
6271 spin_lock_irqsave(&p->pi_lock, flags);
6273 * To be able to change p->policy safely, the apropriate
6274 * runqueue lock must be held.
6276 rq = __task_rq_lock(p);
6277 /* recheck policy now with rq lock held */
6278 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6279 policy = oldpolicy = -1;
6280 __task_rq_unlock(rq);
6281 spin_unlock_irqrestore(&p->pi_lock, flags);
6284 update_rq_clock(rq);
6285 on_rq = p->se.on_rq;
6286 running = task_current(rq, p);
6288 deactivate_task(rq, p, 0);
6290 p->sched_class->put_prev_task(rq, p);
6292 p->sched_reset_on_fork = reset_on_fork;
6295 __setscheduler(rq, p, policy, param->sched_priority);
6298 p->sched_class->set_curr_task(rq);
6300 activate_task(rq, p, 0);
6302 check_class_changed(rq, p, prev_class, oldprio, running);
6304 __task_rq_unlock(rq);
6305 spin_unlock_irqrestore(&p->pi_lock, flags);
6307 rt_mutex_adjust_pi(p);
6313 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6314 * @p: the task in question.
6315 * @policy: new policy.
6316 * @param: structure containing the new RT priority.
6318 * NOTE that the task may be already dead.
6320 int sched_setscheduler(struct task_struct *p, int policy,
6321 struct sched_param *param)
6323 return __sched_setscheduler(p, policy, param, true);
6325 EXPORT_SYMBOL_GPL(sched_setscheduler);
6328 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6329 * @p: the task in question.
6330 * @policy: new policy.
6331 * @param: structure containing the new RT priority.
6333 * Just like sched_setscheduler, only don't bother checking if the
6334 * current context has permission. For example, this is needed in
6335 * stop_machine(): we create temporary high priority worker threads,
6336 * but our caller might not have that capability.
6338 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6339 struct sched_param *param)
6341 return __sched_setscheduler(p, policy, param, false);
6345 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6347 struct sched_param lparam;
6348 struct task_struct *p;
6351 if (!param || pid < 0)
6353 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6358 p = find_process_by_pid(pid);
6360 retval = sched_setscheduler(p, policy, &lparam);
6367 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6368 * @pid: the pid in question.
6369 * @policy: new policy.
6370 * @param: structure containing the new RT priority.
6372 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6373 struct sched_param __user *, param)
6375 /* negative values for policy are not valid */
6379 return do_sched_setscheduler(pid, policy, param);
6383 * sys_sched_setparam - set/change the RT priority of a thread
6384 * @pid: the pid in question.
6385 * @param: structure containing the new RT priority.
6387 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6389 return do_sched_setscheduler(pid, -1, param);
6393 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6394 * @pid: the pid in question.
6396 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6398 struct task_struct *p;
6405 read_lock(&tasklist_lock);
6406 p = find_process_by_pid(pid);
6408 retval = security_task_getscheduler(p);
6411 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6413 read_unlock(&tasklist_lock);
6418 * sys_sched_getparam - get the RT priority of a thread
6419 * @pid: the pid in question.
6420 * @param: structure containing the RT priority.
6422 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6424 struct sched_param lp;
6425 struct task_struct *p;
6428 if (!param || pid < 0)
6431 read_lock(&tasklist_lock);
6432 p = find_process_by_pid(pid);
6437 retval = security_task_getscheduler(p);
6441 lp.sched_priority = p->rt_priority;
6442 read_unlock(&tasklist_lock);
6445 * This one might sleep, we cannot do it with a spinlock held ...
6447 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6452 read_unlock(&tasklist_lock);
6456 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6458 cpumask_var_t cpus_allowed, new_mask;
6459 struct task_struct *p;
6463 read_lock(&tasklist_lock);
6465 p = find_process_by_pid(pid);
6467 read_unlock(&tasklist_lock);
6473 * It is not safe to call set_cpus_allowed with the
6474 * tasklist_lock held. We will bump the task_struct's
6475 * usage count and then drop tasklist_lock.
6478 read_unlock(&tasklist_lock);
6480 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6484 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6486 goto out_free_cpus_allowed;
6489 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6492 retval = security_task_setscheduler(p, 0, NULL);
6496 cpuset_cpus_allowed(p, cpus_allowed);
6497 cpumask_and(new_mask, in_mask, cpus_allowed);
6499 retval = set_cpus_allowed_ptr(p, new_mask);
6502 cpuset_cpus_allowed(p, cpus_allowed);
6503 if (!cpumask_subset(new_mask, cpus_allowed)) {
6505 * We must have raced with a concurrent cpuset
6506 * update. Just reset the cpus_allowed to the
6507 * cpuset's cpus_allowed
6509 cpumask_copy(new_mask, cpus_allowed);
6514 free_cpumask_var(new_mask);
6515 out_free_cpus_allowed:
6516 free_cpumask_var(cpus_allowed);
6523 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6524 struct cpumask *new_mask)
6526 if (len < cpumask_size())
6527 cpumask_clear(new_mask);
6528 else if (len > cpumask_size())
6529 len = cpumask_size();
6531 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6535 * sys_sched_setaffinity - set the cpu affinity of a process
6536 * @pid: pid of the process
6537 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6538 * @user_mask_ptr: user-space pointer to the new cpu mask
6540 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6541 unsigned long __user *, user_mask_ptr)
6543 cpumask_var_t new_mask;
6546 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6549 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6551 retval = sched_setaffinity(pid, new_mask);
6552 free_cpumask_var(new_mask);
6556 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6558 struct task_struct *p;
6562 read_lock(&tasklist_lock);
6565 p = find_process_by_pid(pid);
6569 retval = security_task_getscheduler(p);
6573 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6576 read_unlock(&tasklist_lock);
6583 * sys_sched_getaffinity - get the cpu affinity of a process
6584 * @pid: pid of the process
6585 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6586 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6588 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6589 unsigned long __user *, user_mask_ptr)
6594 if (len < cpumask_size())
6597 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6600 ret = sched_getaffinity(pid, mask);
6602 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6605 ret = cpumask_size();
6607 free_cpumask_var(mask);
6613 * sys_sched_yield - yield the current processor to other threads.
6615 * This function yields the current CPU to other tasks. If there are no
6616 * other threads running on this CPU then this function will return.
6618 SYSCALL_DEFINE0(sched_yield)
6620 struct rq *rq = this_rq_lock();
6622 schedstat_inc(rq, yld_count);
6623 current->sched_class->yield_task(rq);
6626 * Since we are going to call schedule() anyway, there's
6627 * no need to preempt or enable interrupts:
6629 __release(rq->lock);
6630 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6631 _raw_spin_unlock(&rq->lock);
6632 preempt_enable_no_resched();
6639 static inline int should_resched(void)
6641 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6644 static void __cond_resched(void)
6646 add_preempt_count(PREEMPT_ACTIVE);
6648 sub_preempt_count(PREEMPT_ACTIVE);
6651 int __sched _cond_resched(void)
6653 if (should_resched()) {
6659 EXPORT_SYMBOL(_cond_resched);
6662 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6663 * call schedule, and on return reacquire the lock.
6665 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6666 * operations here to prevent schedule() from being called twice (once via
6667 * spin_unlock(), once by hand).
6669 int __cond_resched_lock(spinlock_t *lock)
6671 int resched = should_resched();
6674 lockdep_assert_held(lock);
6676 if (spin_needbreak(lock) || resched) {
6687 EXPORT_SYMBOL(__cond_resched_lock);
6689 int __sched __cond_resched_softirq(void)
6691 BUG_ON(!in_softirq());
6693 if (should_resched()) {
6701 EXPORT_SYMBOL(__cond_resched_softirq);
6704 * yield - yield the current processor to other threads.
6706 * This is a shortcut for kernel-space yielding - it marks the
6707 * thread runnable and calls sys_sched_yield().
6709 void __sched yield(void)
6711 set_current_state(TASK_RUNNING);
6714 EXPORT_SYMBOL(yield);
6717 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6718 * that process accounting knows that this is a task in IO wait state.
6720 * But don't do that if it is a deliberate, throttling IO wait (this task
6721 * has set its backing_dev_info: the queue against which it should throttle)
6723 void __sched io_schedule(void)
6725 struct rq *rq = raw_rq();
6727 delayacct_blkio_start();
6728 atomic_inc(&rq->nr_iowait);
6729 current->in_iowait = 1;
6731 current->in_iowait = 0;
6732 atomic_dec(&rq->nr_iowait);
6733 delayacct_blkio_end();
6735 EXPORT_SYMBOL(io_schedule);
6737 long __sched io_schedule_timeout(long timeout)
6739 struct rq *rq = raw_rq();
6742 delayacct_blkio_start();
6743 atomic_inc(&rq->nr_iowait);
6744 current->in_iowait = 1;
6745 ret = schedule_timeout(timeout);
6746 current->in_iowait = 0;
6747 atomic_dec(&rq->nr_iowait);
6748 delayacct_blkio_end();
6753 * sys_sched_get_priority_max - return maximum RT priority.
6754 * @policy: scheduling class.
6756 * this syscall returns the maximum rt_priority that can be used
6757 * by a given scheduling class.
6759 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6766 ret = MAX_USER_RT_PRIO-1;
6778 * sys_sched_get_priority_min - return minimum RT priority.
6779 * @policy: scheduling class.
6781 * this syscall returns the minimum rt_priority that can be used
6782 * by a given scheduling class.
6784 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6802 * sys_sched_rr_get_interval - return the default timeslice of a process.
6803 * @pid: pid of the process.
6804 * @interval: userspace pointer to the timeslice value.
6806 * this syscall writes the default timeslice value of a given process
6807 * into the user-space timespec buffer. A value of '0' means infinity.
6809 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6810 struct timespec __user *, interval)
6812 struct task_struct *p;
6813 unsigned int time_slice;
6821 read_lock(&tasklist_lock);
6822 p = find_process_by_pid(pid);
6826 retval = security_task_getscheduler(p);
6830 time_slice = p->sched_class->get_rr_interval(p);
6832 read_unlock(&tasklist_lock);
6833 jiffies_to_timespec(time_slice, &t);
6834 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6838 read_unlock(&tasklist_lock);
6842 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6844 void sched_show_task(struct task_struct *p)
6846 unsigned long free = 0;
6849 state = p->state ? __ffs(p->state) + 1 : 0;
6850 printk(KERN_INFO "%-13.13s %c", p->comm,
6851 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6852 #if BITS_PER_LONG == 32
6853 if (state == TASK_RUNNING)
6854 printk(KERN_CONT " running ");
6856 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6858 if (state == TASK_RUNNING)
6859 printk(KERN_CONT " running task ");
6861 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6863 #ifdef CONFIG_DEBUG_STACK_USAGE
6864 free = stack_not_used(p);
6866 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6867 task_pid_nr(p), task_pid_nr(p->real_parent),
6868 (unsigned long)task_thread_info(p)->flags);
6870 show_stack(p, NULL);
6873 void show_state_filter(unsigned long state_filter)
6875 struct task_struct *g, *p;
6877 #if BITS_PER_LONG == 32
6879 " task PC stack pid father\n");
6882 " task PC stack pid father\n");
6884 read_lock(&tasklist_lock);
6885 do_each_thread(g, p) {
6887 * reset the NMI-timeout, listing all files on a slow
6888 * console might take alot of time:
6890 touch_nmi_watchdog();
6891 if (!state_filter || (p->state & state_filter))
6893 } while_each_thread(g, p);
6895 touch_all_softlockup_watchdogs();
6897 #ifdef CONFIG_SCHED_DEBUG
6898 sysrq_sched_debug_show();
6900 read_unlock(&tasklist_lock);
6902 * Only show locks if all tasks are dumped:
6904 if (state_filter == -1)
6905 debug_show_all_locks();
6908 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6910 idle->sched_class = &idle_sched_class;
6914 * init_idle - set up an idle thread for a given CPU
6915 * @idle: task in question
6916 * @cpu: cpu the idle task belongs to
6918 * NOTE: this function does not set the idle thread's NEED_RESCHED
6919 * flag, to make booting more robust.
6921 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6923 struct rq *rq = cpu_rq(cpu);
6924 unsigned long flags;
6926 spin_lock_irqsave(&rq->lock, flags);
6929 idle->se.exec_start = sched_clock();
6931 idle->prio = idle->normal_prio = MAX_PRIO;
6932 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6933 __set_task_cpu(idle, cpu);
6935 rq->curr = rq->idle = idle;
6936 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6939 spin_unlock_irqrestore(&rq->lock, flags);
6941 /* Set the preempt count _outside_ the spinlocks! */
6942 #if defined(CONFIG_PREEMPT)
6943 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6945 task_thread_info(idle)->preempt_count = 0;
6948 * The idle tasks have their own, simple scheduling class:
6950 idle->sched_class = &idle_sched_class;
6951 ftrace_graph_init_task(idle);
6955 * In a system that switches off the HZ timer nohz_cpu_mask
6956 * indicates which cpus entered this state. This is used
6957 * in the rcu update to wait only for active cpus. For system
6958 * which do not switch off the HZ timer nohz_cpu_mask should
6959 * always be CPU_BITS_NONE.
6961 cpumask_var_t nohz_cpu_mask;
6964 * Increase the granularity value when there are more CPUs,
6965 * because with more CPUs the 'effective latency' as visible
6966 * to users decreases. But the relationship is not linear,
6967 * so pick a second-best guess by going with the log2 of the
6970 * This idea comes from the SD scheduler of Con Kolivas:
6972 static inline void sched_init_granularity(void)
6974 unsigned int factor = 1 + ilog2(num_online_cpus());
6975 const unsigned long limit = 200000000;
6977 sysctl_sched_min_granularity *= factor;
6978 if (sysctl_sched_min_granularity > limit)
6979 sysctl_sched_min_granularity = limit;
6981 sysctl_sched_latency *= factor;
6982 if (sysctl_sched_latency > limit)
6983 sysctl_sched_latency = limit;
6985 sysctl_sched_wakeup_granularity *= factor;
6987 sysctl_sched_shares_ratelimit *= factor;
6992 * This is how migration works:
6994 * 1) we queue a struct migration_req structure in the source CPU's
6995 * runqueue and wake up that CPU's migration thread.
6996 * 2) we down() the locked semaphore => thread blocks.
6997 * 3) migration thread wakes up (implicitly it forces the migrated
6998 * thread off the CPU)
6999 * 4) it gets the migration request and checks whether the migrated
7000 * task is still in the wrong runqueue.
7001 * 5) if it's in the wrong runqueue then the migration thread removes
7002 * it and puts it into the right queue.
7003 * 6) migration thread up()s the semaphore.
7004 * 7) we wake up and the migration is done.
7008 * Change a given task's CPU affinity. Migrate the thread to a
7009 * proper CPU and schedule it away if the CPU it's executing on
7010 * is removed from the allowed bitmask.
7012 * NOTE: the caller must have a valid reference to the task, the
7013 * task must not exit() & deallocate itself prematurely. The
7014 * call is not atomic; no spinlocks may be held.
7016 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7018 struct migration_req req;
7019 unsigned long flags;
7023 rq = task_rq_lock(p, &flags);
7024 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7029 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7030 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7035 if (p->sched_class->set_cpus_allowed)
7036 p->sched_class->set_cpus_allowed(p, new_mask);
7038 cpumask_copy(&p->cpus_allowed, new_mask);
7039 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7042 /* Can the task run on the task's current CPU? If so, we're done */
7043 if (cpumask_test_cpu(task_cpu(p), new_mask))
7046 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7047 /* Need help from migration thread: drop lock and wait. */
7048 struct task_struct *mt = rq->migration_thread;
7050 get_task_struct(mt);
7051 task_rq_unlock(rq, &flags);
7052 wake_up_process(rq->migration_thread);
7053 put_task_struct(mt);
7054 wait_for_completion(&req.done);
7055 tlb_migrate_finish(p->mm);
7059 task_rq_unlock(rq, &flags);
7063 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7066 * Move (not current) task off this cpu, onto dest cpu. We're doing
7067 * this because either it can't run here any more (set_cpus_allowed()
7068 * away from this CPU, or CPU going down), or because we're
7069 * attempting to rebalance this task on exec (sched_exec).
7071 * So we race with normal scheduler movements, but that's OK, as long
7072 * as the task is no longer on this CPU.
7074 * Returns non-zero if task was successfully migrated.
7076 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7078 struct rq *rq_dest, *rq_src;
7081 if (unlikely(!cpu_active(dest_cpu)))
7084 rq_src = cpu_rq(src_cpu);
7085 rq_dest = cpu_rq(dest_cpu);
7087 double_rq_lock(rq_src, rq_dest);
7088 /* Already moved. */
7089 if (task_cpu(p) != src_cpu)
7091 /* Affinity changed (again). */
7092 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7095 on_rq = p->se.on_rq;
7097 deactivate_task(rq_src, p, 0);
7099 set_task_cpu(p, dest_cpu);
7101 activate_task(rq_dest, p, 0);
7102 check_preempt_curr(rq_dest, p, 0);
7107 double_rq_unlock(rq_src, rq_dest);
7111 #define RCU_MIGRATION_IDLE 0
7112 #define RCU_MIGRATION_NEED_QS 1
7113 #define RCU_MIGRATION_GOT_QS 2
7114 #define RCU_MIGRATION_MUST_SYNC 3
7117 * migration_thread - this is a highprio system thread that performs
7118 * thread migration by bumping thread off CPU then 'pushing' onto
7121 static int migration_thread(void *data)
7124 int cpu = (long)data;
7128 BUG_ON(rq->migration_thread != current);
7130 set_current_state(TASK_INTERRUPTIBLE);
7131 while (!kthread_should_stop()) {
7132 struct migration_req *req;
7133 struct list_head *head;
7135 spin_lock_irq(&rq->lock);
7137 if (cpu_is_offline(cpu)) {
7138 spin_unlock_irq(&rq->lock);
7142 if (rq->active_balance) {
7143 active_load_balance(rq, cpu);
7144 rq->active_balance = 0;
7147 head = &rq->migration_queue;
7149 if (list_empty(head)) {
7150 spin_unlock_irq(&rq->lock);
7152 set_current_state(TASK_INTERRUPTIBLE);
7155 req = list_entry(head->next, struct migration_req, list);
7156 list_del_init(head->next);
7158 if (req->task != NULL) {
7159 spin_unlock(&rq->lock);
7160 __migrate_task(req->task, cpu, req->dest_cpu);
7161 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7162 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7163 spin_unlock(&rq->lock);
7165 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7166 spin_unlock(&rq->lock);
7167 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7171 complete(&req->done);
7173 __set_current_state(TASK_RUNNING);
7178 #ifdef CONFIG_HOTPLUG_CPU
7180 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7184 local_irq_disable();
7185 ret = __migrate_task(p, src_cpu, dest_cpu);
7191 * Figure out where task on dead CPU should go, use force if necessary.
7193 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7196 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7199 /* Look for allowed, online CPU in same node. */
7200 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7201 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7204 /* Any allowed, online CPU? */
7205 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7206 if (dest_cpu < nr_cpu_ids)
7209 /* No more Mr. Nice Guy. */
7210 if (dest_cpu >= nr_cpu_ids) {
7211 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7212 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7215 * Don't tell them about moving exiting tasks or
7216 * kernel threads (both mm NULL), since they never
7219 if (p->mm && printk_ratelimit()) {
7220 printk(KERN_INFO "process %d (%s) no "
7221 "longer affine to cpu%d\n",
7222 task_pid_nr(p), p->comm, dead_cpu);
7227 /* It can have affinity changed while we were choosing. */
7228 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7233 * While a dead CPU has no uninterruptible tasks queued at this point,
7234 * it might still have a nonzero ->nr_uninterruptible counter, because
7235 * for performance reasons the counter is not stricly tracking tasks to
7236 * their home CPUs. So we just add the counter to another CPU's counter,
7237 * to keep the global sum constant after CPU-down:
7239 static void migrate_nr_uninterruptible(struct rq *rq_src)
7241 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7242 unsigned long flags;
7244 local_irq_save(flags);
7245 double_rq_lock(rq_src, rq_dest);
7246 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7247 rq_src->nr_uninterruptible = 0;
7248 double_rq_unlock(rq_src, rq_dest);
7249 local_irq_restore(flags);
7252 /* Run through task list and migrate tasks from the dead cpu. */
7253 static void migrate_live_tasks(int src_cpu)
7255 struct task_struct *p, *t;
7257 read_lock(&tasklist_lock);
7259 do_each_thread(t, p) {
7263 if (task_cpu(p) == src_cpu)
7264 move_task_off_dead_cpu(src_cpu, p);
7265 } while_each_thread(t, p);
7267 read_unlock(&tasklist_lock);
7271 * Schedules idle task to be the next runnable task on current CPU.
7272 * It does so by boosting its priority to highest possible.
7273 * Used by CPU offline code.
7275 void sched_idle_next(void)
7277 int this_cpu = smp_processor_id();
7278 struct rq *rq = cpu_rq(this_cpu);
7279 struct task_struct *p = rq->idle;
7280 unsigned long flags;
7282 /* cpu has to be offline */
7283 BUG_ON(cpu_online(this_cpu));
7286 * Strictly not necessary since rest of the CPUs are stopped by now
7287 * and interrupts disabled on the current cpu.
7289 spin_lock_irqsave(&rq->lock, flags);
7291 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7293 update_rq_clock(rq);
7294 activate_task(rq, p, 0);
7296 spin_unlock_irqrestore(&rq->lock, flags);
7300 * Ensures that the idle task is using init_mm right before its cpu goes
7303 void idle_task_exit(void)
7305 struct mm_struct *mm = current->active_mm;
7307 BUG_ON(cpu_online(smp_processor_id()));
7310 switch_mm(mm, &init_mm, current);
7314 /* called under rq->lock with disabled interrupts */
7315 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7317 struct rq *rq = cpu_rq(dead_cpu);
7319 /* Must be exiting, otherwise would be on tasklist. */
7320 BUG_ON(!p->exit_state);
7322 /* Cannot have done final schedule yet: would have vanished. */
7323 BUG_ON(p->state == TASK_DEAD);
7328 * Drop lock around migration; if someone else moves it,
7329 * that's OK. No task can be added to this CPU, so iteration is
7332 spin_unlock_irq(&rq->lock);
7333 move_task_off_dead_cpu(dead_cpu, p);
7334 spin_lock_irq(&rq->lock);
7339 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7340 static void migrate_dead_tasks(unsigned int dead_cpu)
7342 struct rq *rq = cpu_rq(dead_cpu);
7343 struct task_struct *next;
7346 if (!rq->nr_running)
7348 update_rq_clock(rq);
7349 next = pick_next_task(rq);
7352 next->sched_class->put_prev_task(rq, next);
7353 migrate_dead(dead_cpu, next);
7359 * remove the tasks which were accounted by rq from calc_load_tasks.
7361 static void calc_global_load_remove(struct rq *rq)
7363 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7364 rq->calc_load_active = 0;
7366 #endif /* CONFIG_HOTPLUG_CPU */
7368 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7370 static struct ctl_table sd_ctl_dir[] = {
7372 .procname = "sched_domain",
7378 static struct ctl_table sd_ctl_root[] = {
7380 .ctl_name = CTL_KERN,
7381 .procname = "kernel",
7383 .child = sd_ctl_dir,
7388 static struct ctl_table *sd_alloc_ctl_entry(int n)
7390 struct ctl_table *entry =
7391 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7396 static void sd_free_ctl_entry(struct ctl_table **tablep)
7398 struct ctl_table *entry;
7401 * In the intermediate directories, both the child directory and
7402 * procname are dynamically allocated and could fail but the mode
7403 * will always be set. In the lowest directory the names are
7404 * static strings and all have proc handlers.
7406 for (entry = *tablep; entry->mode; entry++) {
7408 sd_free_ctl_entry(&entry->child);
7409 if (entry->proc_handler == NULL)
7410 kfree(entry->procname);
7418 set_table_entry(struct ctl_table *entry,
7419 const char *procname, void *data, int maxlen,
7420 mode_t mode, proc_handler *proc_handler)
7422 entry->procname = procname;
7424 entry->maxlen = maxlen;
7426 entry->proc_handler = proc_handler;
7429 static struct ctl_table *
7430 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7432 struct ctl_table *table = sd_alloc_ctl_entry(13);
7437 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7438 sizeof(long), 0644, proc_doulongvec_minmax);
7439 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7440 sizeof(long), 0644, proc_doulongvec_minmax);
7441 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7442 sizeof(int), 0644, proc_dointvec_minmax);
7443 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7444 sizeof(int), 0644, proc_dointvec_minmax);
7445 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7446 sizeof(int), 0644, proc_dointvec_minmax);
7447 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7448 sizeof(int), 0644, proc_dointvec_minmax);
7449 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7450 sizeof(int), 0644, proc_dointvec_minmax);
7451 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7452 sizeof(int), 0644, proc_dointvec_minmax);
7453 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7454 sizeof(int), 0644, proc_dointvec_minmax);
7455 set_table_entry(&table[9], "cache_nice_tries",
7456 &sd->cache_nice_tries,
7457 sizeof(int), 0644, proc_dointvec_minmax);
7458 set_table_entry(&table[10], "flags", &sd->flags,
7459 sizeof(int), 0644, proc_dointvec_minmax);
7460 set_table_entry(&table[11], "name", sd->name,
7461 CORENAME_MAX_SIZE, 0444, proc_dostring);
7462 /* &table[12] is terminator */
7467 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7469 struct ctl_table *entry, *table;
7470 struct sched_domain *sd;
7471 int domain_num = 0, i;
7474 for_each_domain(cpu, sd)
7476 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7481 for_each_domain(cpu, sd) {
7482 snprintf(buf, 32, "domain%d", i);
7483 entry->procname = kstrdup(buf, GFP_KERNEL);
7485 entry->child = sd_alloc_ctl_domain_table(sd);
7492 static struct ctl_table_header *sd_sysctl_header;
7493 static void register_sched_domain_sysctl(void)
7495 int i, cpu_num = num_online_cpus();
7496 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7499 WARN_ON(sd_ctl_dir[0].child);
7500 sd_ctl_dir[0].child = entry;
7505 for_each_online_cpu(i) {
7506 snprintf(buf, 32, "cpu%d", i);
7507 entry->procname = kstrdup(buf, GFP_KERNEL);
7509 entry->child = sd_alloc_ctl_cpu_table(i);
7513 WARN_ON(sd_sysctl_header);
7514 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7517 /* may be called multiple times per register */
7518 static void unregister_sched_domain_sysctl(void)
7520 if (sd_sysctl_header)
7521 unregister_sysctl_table(sd_sysctl_header);
7522 sd_sysctl_header = NULL;
7523 if (sd_ctl_dir[0].child)
7524 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7527 static void register_sched_domain_sysctl(void)
7530 static void unregister_sched_domain_sysctl(void)
7535 static void set_rq_online(struct rq *rq)
7538 const struct sched_class *class;
7540 cpumask_set_cpu(rq->cpu, rq->rd->online);
7543 for_each_class(class) {
7544 if (class->rq_online)
7545 class->rq_online(rq);
7550 static void set_rq_offline(struct rq *rq)
7553 const struct sched_class *class;
7555 for_each_class(class) {
7556 if (class->rq_offline)
7557 class->rq_offline(rq);
7560 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7566 * migration_call - callback that gets triggered when a CPU is added.
7567 * Here we can start up the necessary migration thread for the new CPU.
7569 static int __cpuinit
7570 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7572 struct task_struct *p;
7573 int cpu = (long)hcpu;
7574 unsigned long flags;
7579 case CPU_UP_PREPARE:
7580 case CPU_UP_PREPARE_FROZEN:
7581 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7584 kthread_bind(p, cpu);
7585 /* Must be high prio: stop_machine expects to yield to it. */
7586 rq = task_rq_lock(p, &flags);
7587 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7588 task_rq_unlock(rq, &flags);
7590 cpu_rq(cpu)->migration_thread = p;
7591 rq->calc_load_update = calc_load_update;
7595 case CPU_ONLINE_FROZEN:
7596 /* Strictly unnecessary, as first user will wake it. */
7597 wake_up_process(cpu_rq(cpu)->migration_thread);
7599 /* Update our root-domain */
7601 spin_lock_irqsave(&rq->lock, flags);
7603 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7607 spin_unlock_irqrestore(&rq->lock, flags);
7610 #ifdef CONFIG_HOTPLUG_CPU
7611 case CPU_UP_CANCELED:
7612 case CPU_UP_CANCELED_FROZEN:
7613 if (!cpu_rq(cpu)->migration_thread)
7615 /* Unbind it from offline cpu so it can run. Fall thru. */
7616 kthread_bind(cpu_rq(cpu)->migration_thread,
7617 cpumask_any(cpu_online_mask));
7618 kthread_stop(cpu_rq(cpu)->migration_thread);
7619 put_task_struct(cpu_rq(cpu)->migration_thread);
7620 cpu_rq(cpu)->migration_thread = NULL;
7624 case CPU_DEAD_FROZEN:
7625 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7626 migrate_live_tasks(cpu);
7628 kthread_stop(rq->migration_thread);
7629 put_task_struct(rq->migration_thread);
7630 rq->migration_thread = NULL;
7631 /* Idle task back to normal (off runqueue, low prio) */
7632 spin_lock_irq(&rq->lock);
7633 update_rq_clock(rq);
7634 deactivate_task(rq, rq->idle, 0);
7635 rq->idle->static_prio = MAX_PRIO;
7636 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7637 rq->idle->sched_class = &idle_sched_class;
7638 migrate_dead_tasks(cpu);
7639 spin_unlock_irq(&rq->lock);
7641 migrate_nr_uninterruptible(rq);
7642 BUG_ON(rq->nr_running != 0);
7643 calc_global_load_remove(rq);
7645 * No need to migrate the tasks: it was best-effort if
7646 * they didn't take sched_hotcpu_mutex. Just wake up
7649 spin_lock_irq(&rq->lock);
7650 while (!list_empty(&rq->migration_queue)) {
7651 struct migration_req *req;
7653 req = list_entry(rq->migration_queue.next,
7654 struct migration_req, list);
7655 list_del_init(&req->list);
7656 spin_unlock_irq(&rq->lock);
7657 complete(&req->done);
7658 spin_lock_irq(&rq->lock);
7660 spin_unlock_irq(&rq->lock);
7664 case CPU_DYING_FROZEN:
7665 /* Update our root-domain */
7667 spin_lock_irqsave(&rq->lock, flags);
7669 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7672 spin_unlock_irqrestore(&rq->lock, flags);
7680 * Register at high priority so that task migration (migrate_all_tasks)
7681 * happens before everything else. This has to be lower priority than
7682 * the notifier in the perf_event subsystem, though.
7684 static struct notifier_block __cpuinitdata migration_notifier = {
7685 .notifier_call = migration_call,
7689 static int __init migration_init(void)
7691 void *cpu = (void *)(long)smp_processor_id();
7694 /* Start one for the boot CPU: */
7695 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7696 BUG_ON(err == NOTIFY_BAD);
7697 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7698 register_cpu_notifier(&migration_notifier);
7702 early_initcall(migration_init);
7707 #ifdef CONFIG_SCHED_DEBUG
7709 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7710 struct cpumask *groupmask)
7712 struct sched_group *group = sd->groups;
7715 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7716 cpumask_clear(groupmask);
7718 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7720 if (!(sd->flags & SD_LOAD_BALANCE)) {
7721 printk("does not load-balance\n");
7723 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7728 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7730 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7731 printk(KERN_ERR "ERROR: domain->span does not contain "
7734 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7735 printk(KERN_ERR "ERROR: domain->groups does not contain"
7739 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7743 printk(KERN_ERR "ERROR: group is NULL\n");
7747 if (!group->cpu_power) {
7748 printk(KERN_CONT "\n");
7749 printk(KERN_ERR "ERROR: domain->cpu_power not "
7754 if (!cpumask_weight(sched_group_cpus(group))) {
7755 printk(KERN_CONT "\n");
7756 printk(KERN_ERR "ERROR: empty group\n");
7760 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7761 printk(KERN_CONT "\n");
7762 printk(KERN_ERR "ERROR: repeated CPUs\n");
7766 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7768 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7770 printk(KERN_CONT " %s", str);
7771 if (group->cpu_power != SCHED_LOAD_SCALE) {
7772 printk(KERN_CONT " (cpu_power = %d)",
7776 group = group->next;
7777 } while (group != sd->groups);
7778 printk(KERN_CONT "\n");
7780 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7781 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7784 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7785 printk(KERN_ERR "ERROR: parent span is not a superset "
7786 "of domain->span\n");
7790 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7792 cpumask_var_t groupmask;
7796 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7800 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7802 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7803 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7808 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7815 free_cpumask_var(groupmask);
7817 #else /* !CONFIG_SCHED_DEBUG */
7818 # define sched_domain_debug(sd, cpu) do { } while (0)
7819 #endif /* CONFIG_SCHED_DEBUG */
7821 static int sd_degenerate(struct sched_domain *sd)
7823 if (cpumask_weight(sched_domain_span(sd)) == 1)
7826 /* Following flags need at least 2 groups */
7827 if (sd->flags & (SD_LOAD_BALANCE |
7828 SD_BALANCE_NEWIDLE |
7832 SD_SHARE_PKG_RESOURCES)) {
7833 if (sd->groups != sd->groups->next)
7837 /* Following flags don't use groups */
7838 if (sd->flags & (SD_WAKE_AFFINE))
7845 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7847 unsigned long cflags = sd->flags, pflags = parent->flags;
7849 if (sd_degenerate(parent))
7852 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7855 /* Flags needing groups don't count if only 1 group in parent */
7856 if (parent->groups == parent->groups->next) {
7857 pflags &= ~(SD_LOAD_BALANCE |
7858 SD_BALANCE_NEWIDLE |
7862 SD_SHARE_PKG_RESOURCES);
7863 if (nr_node_ids == 1)
7864 pflags &= ~SD_SERIALIZE;
7866 if (~cflags & pflags)
7872 static void free_rootdomain(struct root_domain *rd)
7874 cpupri_cleanup(&rd->cpupri);
7876 free_cpumask_var(rd->rto_mask);
7877 free_cpumask_var(rd->online);
7878 free_cpumask_var(rd->span);
7882 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7884 struct root_domain *old_rd = NULL;
7885 unsigned long flags;
7887 spin_lock_irqsave(&rq->lock, flags);
7892 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7895 cpumask_clear_cpu(rq->cpu, old_rd->span);
7898 * If we dont want to free the old_rt yet then
7899 * set old_rd to NULL to skip the freeing later
7902 if (!atomic_dec_and_test(&old_rd->refcount))
7906 atomic_inc(&rd->refcount);
7909 cpumask_set_cpu(rq->cpu, rd->span);
7910 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7913 spin_unlock_irqrestore(&rq->lock, flags);
7916 free_rootdomain(old_rd);
7919 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7921 gfp_t gfp = GFP_KERNEL;
7923 memset(rd, 0, sizeof(*rd));
7928 if (!alloc_cpumask_var(&rd->span, gfp))
7930 if (!alloc_cpumask_var(&rd->online, gfp))
7932 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7935 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7940 free_cpumask_var(rd->rto_mask);
7942 free_cpumask_var(rd->online);
7944 free_cpumask_var(rd->span);
7949 static void init_defrootdomain(void)
7951 init_rootdomain(&def_root_domain, true);
7953 atomic_set(&def_root_domain.refcount, 1);
7956 static struct root_domain *alloc_rootdomain(void)
7958 struct root_domain *rd;
7960 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7964 if (init_rootdomain(rd, false) != 0) {
7973 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7974 * hold the hotplug lock.
7977 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7979 struct rq *rq = cpu_rq(cpu);
7980 struct sched_domain *tmp;
7982 /* Remove the sched domains which do not contribute to scheduling. */
7983 for (tmp = sd; tmp; ) {
7984 struct sched_domain *parent = tmp->parent;
7988 if (sd_parent_degenerate(tmp, parent)) {
7989 tmp->parent = parent->parent;
7991 parent->parent->child = tmp;
7996 if (sd && sd_degenerate(sd)) {
8002 sched_domain_debug(sd, cpu);
8004 rq_attach_root(rq, rd);
8005 rcu_assign_pointer(rq->sd, sd);
8008 /* cpus with isolated domains */
8009 static cpumask_var_t cpu_isolated_map;
8011 /* Setup the mask of cpus configured for isolated domains */
8012 static int __init isolated_cpu_setup(char *str)
8014 cpulist_parse(str, cpu_isolated_map);
8018 __setup("isolcpus=", isolated_cpu_setup);
8021 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8022 * to a function which identifies what group(along with sched group) a CPU
8023 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8024 * (due to the fact that we keep track of groups covered with a struct cpumask).
8026 * init_sched_build_groups will build a circular linked list of the groups
8027 * covered by the given span, and will set each group's ->cpumask correctly,
8028 * and ->cpu_power to 0.
8031 init_sched_build_groups(const struct cpumask *span,
8032 const struct cpumask *cpu_map,
8033 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8034 struct sched_group **sg,
8035 struct cpumask *tmpmask),
8036 struct cpumask *covered, struct cpumask *tmpmask)
8038 struct sched_group *first = NULL, *last = NULL;
8041 cpumask_clear(covered);
8043 for_each_cpu(i, span) {
8044 struct sched_group *sg;
8045 int group = group_fn(i, cpu_map, &sg, tmpmask);
8048 if (cpumask_test_cpu(i, covered))
8051 cpumask_clear(sched_group_cpus(sg));
8054 for_each_cpu(j, span) {
8055 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8058 cpumask_set_cpu(j, covered);
8059 cpumask_set_cpu(j, sched_group_cpus(sg));
8070 #define SD_NODES_PER_DOMAIN 16
8075 * find_next_best_node - find the next node to include in a sched_domain
8076 * @node: node whose sched_domain we're building
8077 * @used_nodes: nodes already in the sched_domain
8079 * Find the next node to include in a given scheduling domain. Simply
8080 * finds the closest node not already in the @used_nodes map.
8082 * Should use nodemask_t.
8084 static int find_next_best_node(int node, nodemask_t *used_nodes)
8086 int i, n, val, min_val, best_node = 0;
8090 for (i = 0; i < nr_node_ids; i++) {
8091 /* Start at @node */
8092 n = (node + i) % nr_node_ids;
8094 if (!nr_cpus_node(n))
8097 /* Skip already used nodes */
8098 if (node_isset(n, *used_nodes))
8101 /* Simple min distance search */
8102 val = node_distance(node, n);
8104 if (val < min_val) {
8110 node_set(best_node, *used_nodes);
8115 * sched_domain_node_span - get a cpumask for a node's sched_domain
8116 * @node: node whose cpumask we're constructing
8117 * @span: resulting cpumask
8119 * Given a node, construct a good cpumask for its sched_domain to span. It
8120 * should be one that prevents unnecessary balancing, but also spreads tasks
8123 static void sched_domain_node_span(int node, struct cpumask *span)
8125 nodemask_t used_nodes;
8128 cpumask_clear(span);
8129 nodes_clear(used_nodes);
8131 cpumask_or(span, span, cpumask_of_node(node));
8132 node_set(node, used_nodes);
8134 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8135 int next_node = find_next_best_node(node, &used_nodes);
8137 cpumask_or(span, span, cpumask_of_node(next_node));
8140 #endif /* CONFIG_NUMA */
8142 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8145 * The cpus mask in sched_group and sched_domain hangs off the end.
8147 * ( See the the comments in include/linux/sched.h:struct sched_group
8148 * and struct sched_domain. )
8150 struct static_sched_group {
8151 struct sched_group sg;
8152 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8155 struct static_sched_domain {
8156 struct sched_domain sd;
8157 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8163 cpumask_var_t domainspan;
8164 cpumask_var_t covered;
8165 cpumask_var_t notcovered;
8167 cpumask_var_t nodemask;
8168 cpumask_var_t this_sibling_map;
8169 cpumask_var_t this_core_map;
8170 cpumask_var_t send_covered;
8171 cpumask_var_t tmpmask;
8172 struct sched_group **sched_group_nodes;
8173 struct root_domain *rd;
8177 sa_sched_groups = 0,
8182 sa_this_sibling_map,
8184 sa_sched_group_nodes,
8194 * SMT sched-domains:
8196 #ifdef CONFIG_SCHED_SMT
8197 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8198 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8201 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8202 struct sched_group **sg, struct cpumask *unused)
8205 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8208 #endif /* CONFIG_SCHED_SMT */
8211 * multi-core sched-domains:
8213 #ifdef CONFIG_SCHED_MC
8214 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8215 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8216 #endif /* CONFIG_SCHED_MC */
8218 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8220 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8221 struct sched_group **sg, struct cpumask *mask)
8225 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8226 group = cpumask_first(mask);
8228 *sg = &per_cpu(sched_group_core, group).sg;
8231 #elif defined(CONFIG_SCHED_MC)
8233 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8234 struct sched_group **sg, struct cpumask *unused)
8237 *sg = &per_cpu(sched_group_core, cpu).sg;
8242 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8243 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8246 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8247 struct sched_group **sg, struct cpumask *mask)
8250 #ifdef CONFIG_SCHED_MC
8251 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8252 group = cpumask_first(mask);
8253 #elif defined(CONFIG_SCHED_SMT)
8254 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8255 group = cpumask_first(mask);
8260 *sg = &per_cpu(sched_group_phys, group).sg;
8266 * The init_sched_build_groups can't handle what we want to do with node
8267 * groups, so roll our own. Now each node has its own list of groups which
8268 * gets dynamically allocated.
8270 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8271 static struct sched_group ***sched_group_nodes_bycpu;
8273 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8274 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8276 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8277 struct sched_group **sg,
8278 struct cpumask *nodemask)
8282 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8283 group = cpumask_first(nodemask);
8286 *sg = &per_cpu(sched_group_allnodes, group).sg;
8290 static void init_numa_sched_groups_power(struct sched_group *group_head)
8292 struct sched_group *sg = group_head;
8298 for_each_cpu(j, sched_group_cpus(sg)) {
8299 struct sched_domain *sd;
8301 sd = &per_cpu(phys_domains, j).sd;
8302 if (j != group_first_cpu(sd->groups)) {
8304 * Only add "power" once for each
8310 sg->cpu_power += sd->groups->cpu_power;
8313 } while (sg != group_head);
8316 static int build_numa_sched_groups(struct s_data *d,
8317 const struct cpumask *cpu_map, int num)
8319 struct sched_domain *sd;
8320 struct sched_group *sg, *prev;
8323 cpumask_clear(d->covered);
8324 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8325 if (cpumask_empty(d->nodemask)) {
8326 d->sched_group_nodes[num] = NULL;
8330 sched_domain_node_span(num, d->domainspan);
8331 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8333 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8336 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8340 d->sched_group_nodes[num] = sg;
8342 for_each_cpu(j, d->nodemask) {
8343 sd = &per_cpu(node_domains, j).sd;
8348 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8350 cpumask_or(d->covered, d->covered, d->nodemask);
8353 for (j = 0; j < nr_node_ids; j++) {
8354 n = (num + j) % nr_node_ids;
8355 cpumask_complement(d->notcovered, d->covered);
8356 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8357 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8358 if (cpumask_empty(d->tmpmask))
8360 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8361 if (cpumask_empty(d->tmpmask))
8363 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8367 "Can not alloc domain group for node %d\n", j);
8371 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8372 sg->next = prev->next;
8373 cpumask_or(d->covered, d->covered, d->tmpmask);
8380 #endif /* CONFIG_NUMA */
8383 /* Free memory allocated for various sched_group structures */
8384 static void free_sched_groups(const struct cpumask *cpu_map,
8385 struct cpumask *nodemask)
8389 for_each_cpu(cpu, cpu_map) {
8390 struct sched_group **sched_group_nodes
8391 = sched_group_nodes_bycpu[cpu];
8393 if (!sched_group_nodes)
8396 for (i = 0; i < nr_node_ids; i++) {
8397 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8399 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8400 if (cpumask_empty(nodemask))
8410 if (oldsg != sched_group_nodes[i])
8413 kfree(sched_group_nodes);
8414 sched_group_nodes_bycpu[cpu] = NULL;
8417 #else /* !CONFIG_NUMA */
8418 static void free_sched_groups(const struct cpumask *cpu_map,
8419 struct cpumask *nodemask)
8422 #endif /* CONFIG_NUMA */
8425 * Initialize sched groups cpu_power.
8427 * cpu_power indicates the capacity of sched group, which is used while
8428 * distributing the load between different sched groups in a sched domain.
8429 * Typically cpu_power for all the groups in a sched domain will be same unless
8430 * there are asymmetries in the topology. If there are asymmetries, group
8431 * having more cpu_power will pickup more load compared to the group having
8434 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8436 struct sched_domain *child;
8437 struct sched_group *group;
8441 WARN_ON(!sd || !sd->groups);
8443 if (cpu != group_first_cpu(sd->groups))
8448 sd->groups->cpu_power = 0;
8451 power = SCHED_LOAD_SCALE;
8452 weight = cpumask_weight(sched_domain_span(sd));
8454 * SMT siblings share the power of a single core.
8455 * Usually multiple threads get a better yield out of
8456 * that one core than a single thread would have,
8457 * reflect that in sd->smt_gain.
8459 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8460 power *= sd->smt_gain;
8462 power >>= SCHED_LOAD_SHIFT;
8464 sd->groups->cpu_power += power;
8469 * Add cpu_power of each child group to this groups cpu_power.
8471 group = child->groups;
8473 sd->groups->cpu_power += group->cpu_power;
8474 group = group->next;
8475 } while (group != child->groups);
8479 * Initializers for schedule domains
8480 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8483 #ifdef CONFIG_SCHED_DEBUG
8484 # define SD_INIT_NAME(sd, type) sd->name = #type
8486 # define SD_INIT_NAME(sd, type) do { } while (0)
8489 #define SD_INIT(sd, type) sd_init_##type(sd)
8491 #define SD_INIT_FUNC(type) \
8492 static noinline void sd_init_##type(struct sched_domain *sd) \
8494 memset(sd, 0, sizeof(*sd)); \
8495 *sd = SD_##type##_INIT; \
8496 sd->level = SD_LV_##type; \
8497 SD_INIT_NAME(sd, type); \
8502 SD_INIT_FUNC(ALLNODES)
8505 #ifdef CONFIG_SCHED_SMT
8506 SD_INIT_FUNC(SIBLING)
8508 #ifdef CONFIG_SCHED_MC
8512 static int default_relax_domain_level = -1;
8514 static int __init setup_relax_domain_level(char *str)
8518 val = simple_strtoul(str, NULL, 0);
8519 if (val < SD_LV_MAX)
8520 default_relax_domain_level = val;
8524 __setup("relax_domain_level=", setup_relax_domain_level);
8526 static void set_domain_attribute(struct sched_domain *sd,
8527 struct sched_domain_attr *attr)
8531 if (!attr || attr->relax_domain_level < 0) {
8532 if (default_relax_domain_level < 0)
8535 request = default_relax_domain_level;
8537 request = attr->relax_domain_level;
8538 if (request < sd->level) {
8539 /* turn off idle balance on this domain */
8540 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8542 /* turn on idle balance on this domain */
8543 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8547 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8548 const struct cpumask *cpu_map)
8551 case sa_sched_groups:
8552 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8553 d->sched_group_nodes = NULL;
8555 free_rootdomain(d->rd); /* fall through */
8557 free_cpumask_var(d->tmpmask); /* fall through */
8558 case sa_send_covered:
8559 free_cpumask_var(d->send_covered); /* fall through */
8560 case sa_this_core_map:
8561 free_cpumask_var(d->this_core_map); /* fall through */
8562 case sa_this_sibling_map:
8563 free_cpumask_var(d->this_sibling_map); /* fall through */
8565 free_cpumask_var(d->nodemask); /* fall through */
8566 case sa_sched_group_nodes:
8568 kfree(d->sched_group_nodes); /* fall through */
8570 free_cpumask_var(d->notcovered); /* fall through */
8572 free_cpumask_var(d->covered); /* fall through */
8574 free_cpumask_var(d->domainspan); /* fall through */
8581 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8582 const struct cpumask *cpu_map)
8585 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8587 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8588 return sa_domainspan;
8589 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8591 /* Allocate the per-node list of sched groups */
8592 d->sched_group_nodes = kcalloc(nr_node_ids,
8593 sizeof(struct sched_group *), GFP_KERNEL);
8594 if (!d->sched_group_nodes) {
8595 printk(KERN_WARNING "Can not alloc sched group node list\n");
8596 return sa_notcovered;
8598 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8600 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8601 return sa_sched_group_nodes;
8602 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8604 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8605 return sa_this_sibling_map;
8606 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8607 return sa_this_core_map;
8608 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8609 return sa_send_covered;
8610 d->rd = alloc_rootdomain();
8612 printk(KERN_WARNING "Cannot alloc root domain\n");
8615 return sa_rootdomain;
8618 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8619 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8621 struct sched_domain *sd = NULL;
8623 struct sched_domain *parent;
8626 if (cpumask_weight(cpu_map) >
8627 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8628 sd = &per_cpu(allnodes_domains, i).sd;
8629 SD_INIT(sd, ALLNODES);
8630 set_domain_attribute(sd, attr);
8631 cpumask_copy(sched_domain_span(sd), cpu_map);
8632 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8637 sd = &per_cpu(node_domains, i).sd;
8639 set_domain_attribute(sd, attr);
8640 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8641 sd->parent = parent;
8644 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8649 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8650 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8651 struct sched_domain *parent, int i)
8653 struct sched_domain *sd;
8654 sd = &per_cpu(phys_domains, i).sd;
8656 set_domain_attribute(sd, attr);
8657 cpumask_copy(sched_domain_span(sd), d->nodemask);
8658 sd->parent = parent;
8661 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8665 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8666 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8667 struct sched_domain *parent, int i)
8669 struct sched_domain *sd = parent;
8670 #ifdef CONFIG_SCHED_MC
8671 sd = &per_cpu(core_domains, i).sd;
8673 set_domain_attribute(sd, attr);
8674 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8675 sd->parent = parent;
8677 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8682 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8683 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8684 struct sched_domain *parent, int i)
8686 struct sched_domain *sd = parent;
8687 #ifdef CONFIG_SCHED_SMT
8688 sd = &per_cpu(cpu_domains, i).sd;
8689 SD_INIT(sd, SIBLING);
8690 set_domain_attribute(sd, attr);
8691 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8692 sd->parent = parent;
8694 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8699 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8700 const struct cpumask *cpu_map, int cpu)
8703 #ifdef CONFIG_SCHED_SMT
8704 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8705 cpumask_and(d->this_sibling_map, cpu_map,
8706 topology_thread_cpumask(cpu));
8707 if (cpu == cpumask_first(d->this_sibling_map))
8708 init_sched_build_groups(d->this_sibling_map, cpu_map,
8710 d->send_covered, d->tmpmask);
8713 #ifdef CONFIG_SCHED_MC
8714 case SD_LV_MC: /* set up multi-core groups */
8715 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8716 if (cpu == cpumask_first(d->this_core_map))
8717 init_sched_build_groups(d->this_core_map, cpu_map,
8719 d->send_covered, d->tmpmask);
8722 case SD_LV_CPU: /* set up physical groups */
8723 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8724 if (!cpumask_empty(d->nodemask))
8725 init_sched_build_groups(d->nodemask, cpu_map,
8727 d->send_covered, d->tmpmask);
8730 case SD_LV_ALLNODES:
8731 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8732 d->send_covered, d->tmpmask);
8741 * Build sched domains for a given set of cpus and attach the sched domains
8742 * to the individual cpus
8744 static int __build_sched_domains(const struct cpumask *cpu_map,
8745 struct sched_domain_attr *attr)
8747 enum s_alloc alloc_state = sa_none;
8749 struct sched_domain *sd;
8755 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8756 if (alloc_state != sa_rootdomain)
8758 alloc_state = sa_sched_groups;
8761 * Set up domains for cpus specified by the cpu_map.
8763 for_each_cpu(i, cpu_map) {
8764 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8767 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8768 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8769 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8770 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8773 for_each_cpu(i, cpu_map) {
8774 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8775 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8778 /* Set up physical groups */
8779 for (i = 0; i < nr_node_ids; i++)
8780 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8783 /* Set up node groups */
8785 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8787 for (i = 0; i < nr_node_ids; i++)
8788 if (build_numa_sched_groups(&d, cpu_map, i))
8792 /* Calculate CPU power for physical packages and nodes */
8793 #ifdef CONFIG_SCHED_SMT
8794 for_each_cpu(i, cpu_map) {
8795 sd = &per_cpu(cpu_domains, i).sd;
8796 init_sched_groups_power(i, sd);
8799 #ifdef CONFIG_SCHED_MC
8800 for_each_cpu(i, cpu_map) {
8801 sd = &per_cpu(core_domains, i).sd;
8802 init_sched_groups_power(i, sd);
8806 for_each_cpu(i, cpu_map) {
8807 sd = &per_cpu(phys_domains, i).sd;
8808 init_sched_groups_power(i, sd);
8812 for (i = 0; i < nr_node_ids; i++)
8813 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8815 if (d.sd_allnodes) {
8816 struct sched_group *sg;
8818 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8820 init_numa_sched_groups_power(sg);
8824 /* Attach the domains */
8825 for_each_cpu(i, cpu_map) {
8826 #ifdef CONFIG_SCHED_SMT
8827 sd = &per_cpu(cpu_domains, i).sd;
8828 #elif defined(CONFIG_SCHED_MC)
8829 sd = &per_cpu(core_domains, i).sd;
8831 sd = &per_cpu(phys_domains, i).sd;
8833 cpu_attach_domain(sd, d.rd, i);
8836 d.sched_group_nodes = NULL; /* don't free this we still need it */
8837 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8841 __free_domain_allocs(&d, alloc_state, cpu_map);
8845 static int build_sched_domains(const struct cpumask *cpu_map)
8847 return __build_sched_domains(cpu_map, NULL);
8850 static struct cpumask *doms_cur; /* current sched domains */
8851 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8852 static struct sched_domain_attr *dattr_cur;
8853 /* attribues of custom domains in 'doms_cur' */
8856 * Special case: If a kmalloc of a doms_cur partition (array of
8857 * cpumask) fails, then fallback to a single sched domain,
8858 * as determined by the single cpumask fallback_doms.
8860 static cpumask_var_t fallback_doms;
8863 * arch_update_cpu_topology lets virtualized architectures update the
8864 * cpu core maps. It is supposed to return 1 if the topology changed
8865 * or 0 if it stayed the same.
8867 int __attribute__((weak)) arch_update_cpu_topology(void)
8873 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8874 * For now this just excludes isolated cpus, but could be used to
8875 * exclude other special cases in the future.
8877 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8881 arch_update_cpu_topology();
8883 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8885 doms_cur = fallback_doms;
8886 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8888 err = build_sched_domains(doms_cur);
8889 register_sched_domain_sysctl();
8894 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8895 struct cpumask *tmpmask)
8897 free_sched_groups(cpu_map, tmpmask);
8901 * Detach sched domains from a group of cpus specified in cpu_map
8902 * These cpus will now be attached to the NULL domain
8904 static void detach_destroy_domains(const struct cpumask *cpu_map)
8906 /* Save because hotplug lock held. */
8907 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8910 for_each_cpu(i, cpu_map)
8911 cpu_attach_domain(NULL, &def_root_domain, i);
8912 synchronize_sched();
8913 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8916 /* handle null as "default" */
8917 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8918 struct sched_domain_attr *new, int idx_new)
8920 struct sched_domain_attr tmp;
8927 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8928 new ? (new + idx_new) : &tmp,
8929 sizeof(struct sched_domain_attr));
8933 * Partition sched domains as specified by the 'ndoms_new'
8934 * cpumasks in the array doms_new[] of cpumasks. This compares
8935 * doms_new[] to the current sched domain partitioning, doms_cur[].
8936 * It destroys each deleted domain and builds each new domain.
8938 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8939 * The masks don't intersect (don't overlap.) We should setup one
8940 * sched domain for each mask. CPUs not in any of the cpumasks will
8941 * not be load balanced. If the same cpumask appears both in the
8942 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8945 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8946 * ownership of it and will kfree it when done with it. If the caller
8947 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8948 * ndoms_new == 1, and partition_sched_domains() will fallback to
8949 * the single partition 'fallback_doms', it also forces the domains
8952 * If doms_new == NULL it will be replaced with cpu_online_mask.
8953 * ndoms_new == 0 is a special case for destroying existing domains,
8954 * and it will not create the default domain.
8956 * Call with hotplug lock held
8958 /* FIXME: Change to struct cpumask *doms_new[] */
8959 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8960 struct sched_domain_attr *dattr_new)
8965 mutex_lock(&sched_domains_mutex);
8967 /* always unregister in case we don't destroy any domains */
8968 unregister_sched_domain_sysctl();
8970 /* Let architecture update cpu core mappings. */
8971 new_topology = arch_update_cpu_topology();
8973 n = doms_new ? ndoms_new : 0;
8975 /* Destroy deleted domains */
8976 for (i = 0; i < ndoms_cur; i++) {
8977 for (j = 0; j < n && !new_topology; j++) {
8978 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8979 && dattrs_equal(dattr_cur, i, dattr_new, j))
8982 /* no match - a current sched domain not in new doms_new[] */
8983 detach_destroy_domains(doms_cur + i);
8988 if (doms_new == NULL) {
8990 doms_new = fallback_doms;
8991 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8992 WARN_ON_ONCE(dattr_new);
8995 /* Build new domains */
8996 for (i = 0; i < ndoms_new; i++) {
8997 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8998 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8999 && dattrs_equal(dattr_new, i, dattr_cur, j))
9002 /* no match - add a new doms_new */
9003 __build_sched_domains(doms_new + i,
9004 dattr_new ? dattr_new + i : NULL);
9009 /* Remember the new sched domains */
9010 if (doms_cur != fallback_doms)
9012 kfree(dattr_cur); /* kfree(NULL) is safe */
9013 doms_cur = doms_new;
9014 dattr_cur = dattr_new;
9015 ndoms_cur = ndoms_new;
9017 register_sched_domain_sysctl();
9019 mutex_unlock(&sched_domains_mutex);
9022 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9023 static void arch_reinit_sched_domains(void)
9027 /* Destroy domains first to force the rebuild */
9028 partition_sched_domains(0, NULL, NULL);
9030 rebuild_sched_domains();
9034 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9036 unsigned int level = 0;
9038 if (sscanf(buf, "%u", &level) != 1)
9042 * level is always be positive so don't check for
9043 * level < POWERSAVINGS_BALANCE_NONE which is 0
9044 * What happens on 0 or 1 byte write,
9045 * need to check for count as well?
9048 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9052 sched_smt_power_savings = level;
9054 sched_mc_power_savings = level;
9056 arch_reinit_sched_domains();
9061 #ifdef CONFIG_SCHED_MC
9062 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9065 return sprintf(page, "%u\n", sched_mc_power_savings);
9067 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9068 const char *buf, size_t count)
9070 return sched_power_savings_store(buf, count, 0);
9072 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9073 sched_mc_power_savings_show,
9074 sched_mc_power_savings_store);
9077 #ifdef CONFIG_SCHED_SMT
9078 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9081 return sprintf(page, "%u\n", sched_smt_power_savings);
9083 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9084 const char *buf, size_t count)
9086 return sched_power_savings_store(buf, count, 1);
9088 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9089 sched_smt_power_savings_show,
9090 sched_smt_power_savings_store);
9093 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9097 #ifdef CONFIG_SCHED_SMT
9099 err = sysfs_create_file(&cls->kset.kobj,
9100 &attr_sched_smt_power_savings.attr);
9102 #ifdef CONFIG_SCHED_MC
9103 if (!err && mc_capable())
9104 err = sysfs_create_file(&cls->kset.kobj,
9105 &attr_sched_mc_power_savings.attr);
9109 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9111 #ifndef CONFIG_CPUSETS
9113 * Add online and remove offline CPUs from the scheduler domains.
9114 * When cpusets are enabled they take over this function.
9116 static int update_sched_domains(struct notifier_block *nfb,
9117 unsigned long action, void *hcpu)
9121 case CPU_ONLINE_FROZEN:
9123 case CPU_DEAD_FROZEN:
9124 partition_sched_domains(1, NULL, NULL);
9133 static int update_runtime(struct notifier_block *nfb,
9134 unsigned long action, void *hcpu)
9136 int cpu = (int)(long)hcpu;
9139 case CPU_DOWN_PREPARE:
9140 case CPU_DOWN_PREPARE_FROZEN:
9141 disable_runtime(cpu_rq(cpu));
9144 case CPU_DOWN_FAILED:
9145 case CPU_DOWN_FAILED_FROZEN:
9147 case CPU_ONLINE_FROZEN:
9148 enable_runtime(cpu_rq(cpu));
9156 void __init sched_init_smp(void)
9158 cpumask_var_t non_isolated_cpus;
9160 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9161 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9163 #if defined(CONFIG_NUMA)
9164 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9166 BUG_ON(sched_group_nodes_bycpu == NULL);
9169 mutex_lock(&sched_domains_mutex);
9170 arch_init_sched_domains(cpu_online_mask);
9171 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9172 if (cpumask_empty(non_isolated_cpus))
9173 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9174 mutex_unlock(&sched_domains_mutex);
9177 #ifndef CONFIG_CPUSETS
9178 /* XXX: Theoretical race here - CPU may be hotplugged now */
9179 hotcpu_notifier(update_sched_domains, 0);
9182 /* RT runtime code needs to handle some hotplug events */
9183 hotcpu_notifier(update_runtime, 0);
9187 /* Move init over to a non-isolated CPU */
9188 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9190 sched_init_granularity();
9191 free_cpumask_var(non_isolated_cpus);
9193 init_sched_rt_class();
9196 void __init sched_init_smp(void)
9198 sched_init_granularity();
9200 #endif /* CONFIG_SMP */
9202 const_debug unsigned int sysctl_timer_migration = 1;
9204 int in_sched_functions(unsigned long addr)
9206 return in_lock_functions(addr) ||
9207 (addr >= (unsigned long)__sched_text_start
9208 && addr < (unsigned long)__sched_text_end);
9211 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9213 cfs_rq->tasks_timeline = RB_ROOT;
9214 INIT_LIST_HEAD(&cfs_rq->tasks);
9215 #ifdef CONFIG_FAIR_GROUP_SCHED
9218 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9221 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9223 struct rt_prio_array *array;
9226 array = &rt_rq->active;
9227 for (i = 0; i < MAX_RT_PRIO; i++) {
9228 INIT_LIST_HEAD(array->queue + i);
9229 __clear_bit(i, array->bitmap);
9231 /* delimiter for bitsearch: */
9232 __set_bit(MAX_RT_PRIO, array->bitmap);
9234 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9235 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9237 rt_rq->highest_prio.next = MAX_RT_PRIO;
9241 rt_rq->rt_nr_migratory = 0;
9242 rt_rq->overloaded = 0;
9243 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9247 rt_rq->rt_throttled = 0;
9248 rt_rq->rt_runtime = 0;
9249 spin_lock_init(&rt_rq->rt_runtime_lock);
9251 #ifdef CONFIG_RT_GROUP_SCHED
9252 rt_rq->rt_nr_boosted = 0;
9257 #ifdef CONFIG_FAIR_GROUP_SCHED
9258 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9259 struct sched_entity *se, int cpu, int add,
9260 struct sched_entity *parent)
9262 struct rq *rq = cpu_rq(cpu);
9263 tg->cfs_rq[cpu] = cfs_rq;
9264 init_cfs_rq(cfs_rq, rq);
9267 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9270 /* se could be NULL for init_task_group */
9275 se->cfs_rq = &rq->cfs;
9277 se->cfs_rq = parent->my_q;
9280 se->load.weight = tg->shares;
9281 se->load.inv_weight = 0;
9282 se->parent = parent;
9286 #ifdef CONFIG_RT_GROUP_SCHED
9287 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9288 struct sched_rt_entity *rt_se, int cpu, int add,
9289 struct sched_rt_entity *parent)
9291 struct rq *rq = cpu_rq(cpu);
9293 tg->rt_rq[cpu] = rt_rq;
9294 init_rt_rq(rt_rq, rq);
9296 rt_rq->rt_se = rt_se;
9297 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9299 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9301 tg->rt_se[cpu] = rt_se;
9306 rt_se->rt_rq = &rq->rt;
9308 rt_se->rt_rq = parent->my_q;
9310 rt_se->my_q = rt_rq;
9311 rt_se->parent = parent;
9312 INIT_LIST_HEAD(&rt_se->run_list);
9316 void __init sched_init(void)
9319 unsigned long alloc_size = 0, ptr;
9321 #ifdef CONFIG_FAIR_GROUP_SCHED
9322 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9324 #ifdef CONFIG_RT_GROUP_SCHED
9325 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9327 #ifdef CONFIG_USER_SCHED
9330 #ifdef CONFIG_CPUMASK_OFFSTACK
9331 alloc_size += num_possible_cpus() * cpumask_size();
9334 * As sched_init() is called before page_alloc is setup,
9335 * we use alloc_bootmem().
9338 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9340 #ifdef CONFIG_FAIR_GROUP_SCHED
9341 init_task_group.se = (struct sched_entity **)ptr;
9342 ptr += nr_cpu_ids * sizeof(void **);
9344 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9345 ptr += nr_cpu_ids * sizeof(void **);
9347 #ifdef CONFIG_USER_SCHED
9348 root_task_group.se = (struct sched_entity **)ptr;
9349 ptr += nr_cpu_ids * sizeof(void **);
9351 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9352 ptr += nr_cpu_ids * sizeof(void **);
9353 #endif /* CONFIG_USER_SCHED */
9354 #endif /* CONFIG_FAIR_GROUP_SCHED */
9355 #ifdef CONFIG_RT_GROUP_SCHED
9356 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9357 ptr += nr_cpu_ids * sizeof(void **);
9359 init_task_group.rt_rq = (struct rt_rq **)ptr;
9360 ptr += nr_cpu_ids * sizeof(void **);
9362 #ifdef CONFIG_USER_SCHED
9363 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9364 ptr += nr_cpu_ids * sizeof(void **);
9366 root_task_group.rt_rq = (struct rt_rq **)ptr;
9367 ptr += nr_cpu_ids * sizeof(void **);
9368 #endif /* CONFIG_USER_SCHED */
9369 #endif /* CONFIG_RT_GROUP_SCHED */
9370 #ifdef CONFIG_CPUMASK_OFFSTACK
9371 for_each_possible_cpu(i) {
9372 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9373 ptr += cpumask_size();
9375 #endif /* CONFIG_CPUMASK_OFFSTACK */
9379 init_defrootdomain();
9382 init_rt_bandwidth(&def_rt_bandwidth,
9383 global_rt_period(), global_rt_runtime());
9385 #ifdef CONFIG_RT_GROUP_SCHED
9386 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9387 global_rt_period(), global_rt_runtime());
9388 #ifdef CONFIG_USER_SCHED
9389 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9390 global_rt_period(), RUNTIME_INF);
9391 #endif /* CONFIG_USER_SCHED */
9392 #endif /* CONFIG_RT_GROUP_SCHED */
9394 #ifdef CONFIG_GROUP_SCHED
9395 list_add(&init_task_group.list, &task_groups);
9396 INIT_LIST_HEAD(&init_task_group.children);
9398 #ifdef CONFIG_USER_SCHED
9399 INIT_LIST_HEAD(&root_task_group.children);
9400 init_task_group.parent = &root_task_group;
9401 list_add(&init_task_group.siblings, &root_task_group.children);
9402 #endif /* CONFIG_USER_SCHED */
9403 #endif /* CONFIG_GROUP_SCHED */
9405 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9406 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9407 __alignof__(unsigned long));
9409 for_each_possible_cpu(i) {
9413 spin_lock_init(&rq->lock);
9415 rq->calc_load_active = 0;
9416 rq->calc_load_update = jiffies + LOAD_FREQ;
9417 init_cfs_rq(&rq->cfs, rq);
9418 init_rt_rq(&rq->rt, rq);
9419 #ifdef CONFIG_FAIR_GROUP_SCHED
9420 init_task_group.shares = init_task_group_load;
9421 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9422 #ifdef CONFIG_CGROUP_SCHED
9424 * How much cpu bandwidth does init_task_group get?
9426 * In case of task-groups formed thr' the cgroup filesystem, it
9427 * gets 100% of the cpu resources in the system. This overall
9428 * system cpu resource is divided among the tasks of
9429 * init_task_group and its child task-groups in a fair manner,
9430 * based on each entity's (task or task-group's) weight
9431 * (se->load.weight).
9433 * In other words, if init_task_group has 10 tasks of weight
9434 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9435 * then A0's share of the cpu resource is:
9437 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9439 * We achieve this by letting init_task_group's tasks sit
9440 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9442 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9443 #elif defined CONFIG_USER_SCHED
9444 root_task_group.shares = NICE_0_LOAD;
9445 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9447 * In case of task-groups formed thr' the user id of tasks,
9448 * init_task_group represents tasks belonging to root user.
9449 * Hence it forms a sibling of all subsequent groups formed.
9450 * In this case, init_task_group gets only a fraction of overall
9451 * system cpu resource, based on the weight assigned to root
9452 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9453 * by letting tasks of init_task_group sit in a separate cfs_rq
9454 * (init_tg_cfs_rq) and having one entity represent this group of
9455 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9457 init_tg_cfs_entry(&init_task_group,
9458 &per_cpu(init_tg_cfs_rq, i),
9459 &per_cpu(init_sched_entity, i), i, 1,
9460 root_task_group.se[i]);
9463 #endif /* CONFIG_FAIR_GROUP_SCHED */
9465 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9466 #ifdef CONFIG_RT_GROUP_SCHED
9467 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9468 #ifdef CONFIG_CGROUP_SCHED
9469 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9470 #elif defined CONFIG_USER_SCHED
9471 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9472 init_tg_rt_entry(&init_task_group,
9473 &per_cpu(init_rt_rq, i),
9474 &per_cpu(init_sched_rt_entity, i), i, 1,
9475 root_task_group.rt_se[i]);
9479 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9480 rq->cpu_load[j] = 0;
9484 rq->post_schedule = 0;
9485 rq->active_balance = 0;
9486 rq->next_balance = jiffies;
9490 rq->migration_thread = NULL;
9491 INIT_LIST_HEAD(&rq->migration_queue);
9492 rq_attach_root(rq, &def_root_domain);
9495 atomic_set(&rq->nr_iowait, 0);
9498 set_load_weight(&init_task);
9500 #ifdef CONFIG_PREEMPT_NOTIFIERS
9501 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9505 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9508 #ifdef CONFIG_RT_MUTEXES
9509 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9513 * The boot idle thread does lazy MMU switching as well:
9515 atomic_inc(&init_mm.mm_count);
9516 enter_lazy_tlb(&init_mm, current);
9519 * Make us the idle thread. Technically, schedule() should not be
9520 * called from this thread, however somewhere below it might be,
9521 * but because we are the idle thread, we just pick up running again
9522 * when this runqueue becomes "idle".
9524 init_idle(current, smp_processor_id());
9526 calc_load_update = jiffies + LOAD_FREQ;
9529 * During early bootup we pretend to be a normal task:
9531 current->sched_class = &fair_sched_class;
9533 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9534 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9537 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9538 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9540 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9545 scheduler_running = 1;
9548 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9549 static inline int preempt_count_equals(int preempt_offset)
9551 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9553 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9556 void __might_sleep(char *file, int line, int preempt_offset)
9559 static unsigned long prev_jiffy; /* ratelimiting */
9561 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9562 system_state != SYSTEM_RUNNING || oops_in_progress)
9564 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9566 prev_jiffy = jiffies;
9569 "BUG: sleeping function called from invalid context at %s:%d\n",
9572 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9573 in_atomic(), irqs_disabled(),
9574 current->pid, current->comm);
9576 debug_show_held_locks(current);
9577 if (irqs_disabled())
9578 print_irqtrace_events(current);
9582 EXPORT_SYMBOL(__might_sleep);
9585 #ifdef CONFIG_MAGIC_SYSRQ
9586 static void normalize_task(struct rq *rq, struct task_struct *p)
9590 update_rq_clock(rq);
9591 on_rq = p->se.on_rq;
9593 deactivate_task(rq, p, 0);
9594 __setscheduler(rq, p, SCHED_NORMAL, 0);
9596 activate_task(rq, p, 0);
9597 resched_task(rq->curr);
9601 void normalize_rt_tasks(void)
9603 struct task_struct *g, *p;
9604 unsigned long flags;
9607 read_lock_irqsave(&tasklist_lock, flags);
9608 do_each_thread(g, p) {
9610 * Only normalize user tasks:
9615 p->se.exec_start = 0;
9616 #ifdef CONFIG_SCHEDSTATS
9617 p->se.wait_start = 0;
9618 p->se.sleep_start = 0;
9619 p->se.block_start = 0;
9624 * Renice negative nice level userspace
9627 if (TASK_NICE(p) < 0 && p->mm)
9628 set_user_nice(p, 0);
9632 spin_lock(&p->pi_lock);
9633 rq = __task_rq_lock(p);
9635 normalize_task(rq, p);
9637 __task_rq_unlock(rq);
9638 spin_unlock(&p->pi_lock);
9639 } while_each_thread(g, p);
9641 read_unlock_irqrestore(&tasklist_lock, flags);
9644 #endif /* CONFIG_MAGIC_SYSRQ */
9648 * These functions are only useful for the IA64 MCA handling.
9650 * They can only be called when the whole system has been
9651 * stopped - every CPU needs to be quiescent, and no scheduling
9652 * activity can take place. Using them for anything else would
9653 * be a serious bug, and as a result, they aren't even visible
9654 * under any other configuration.
9658 * curr_task - return the current task for a given cpu.
9659 * @cpu: the processor in question.
9661 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9663 struct task_struct *curr_task(int cpu)
9665 return cpu_curr(cpu);
9669 * set_curr_task - set the current task for a given cpu.
9670 * @cpu: the processor in question.
9671 * @p: the task pointer to set.
9673 * Description: This function must only be used when non-maskable interrupts
9674 * are serviced on a separate stack. It allows the architecture to switch the
9675 * notion of the current task on a cpu in a non-blocking manner. This function
9676 * must be called with all CPU's synchronized, and interrupts disabled, the
9677 * and caller must save the original value of the current task (see
9678 * curr_task() above) and restore that value before reenabling interrupts and
9679 * re-starting the system.
9681 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9683 void set_curr_task(int cpu, struct task_struct *p)
9690 #ifdef CONFIG_FAIR_GROUP_SCHED
9691 static void free_fair_sched_group(struct task_group *tg)
9695 for_each_possible_cpu(i) {
9697 kfree(tg->cfs_rq[i]);
9707 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9709 struct cfs_rq *cfs_rq;
9710 struct sched_entity *se;
9714 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9717 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9721 tg->shares = NICE_0_LOAD;
9723 for_each_possible_cpu(i) {
9726 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9727 GFP_KERNEL, cpu_to_node(i));
9731 se = kzalloc_node(sizeof(struct sched_entity),
9732 GFP_KERNEL, cpu_to_node(i));
9736 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9745 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9747 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9748 &cpu_rq(cpu)->leaf_cfs_rq_list);
9751 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9753 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9755 #else /* !CONFG_FAIR_GROUP_SCHED */
9756 static inline void free_fair_sched_group(struct task_group *tg)
9761 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9766 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9770 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9773 #endif /* CONFIG_FAIR_GROUP_SCHED */
9775 #ifdef CONFIG_RT_GROUP_SCHED
9776 static void free_rt_sched_group(struct task_group *tg)
9780 destroy_rt_bandwidth(&tg->rt_bandwidth);
9782 for_each_possible_cpu(i) {
9784 kfree(tg->rt_rq[i]);
9786 kfree(tg->rt_se[i]);
9794 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9796 struct rt_rq *rt_rq;
9797 struct sched_rt_entity *rt_se;
9801 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9804 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9808 init_rt_bandwidth(&tg->rt_bandwidth,
9809 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9811 for_each_possible_cpu(i) {
9814 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9815 GFP_KERNEL, cpu_to_node(i));
9819 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9820 GFP_KERNEL, cpu_to_node(i));
9824 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9833 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9835 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9836 &cpu_rq(cpu)->leaf_rt_rq_list);
9839 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9841 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9843 #else /* !CONFIG_RT_GROUP_SCHED */
9844 static inline void free_rt_sched_group(struct task_group *tg)
9849 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9854 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9858 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9861 #endif /* CONFIG_RT_GROUP_SCHED */
9863 #ifdef CONFIG_GROUP_SCHED
9864 static void free_sched_group(struct task_group *tg)
9866 free_fair_sched_group(tg);
9867 free_rt_sched_group(tg);
9871 /* allocate runqueue etc for a new task group */
9872 struct task_group *sched_create_group(struct task_group *parent)
9874 struct task_group *tg;
9875 unsigned long flags;
9878 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9880 return ERR_PTR(-ENOMEM);
9882 if (!alloc_fair_sched_group(tg, parent))
9885 if (!alloc_rt_sched_group(tg, parent))
9888 spin_lock_irqsave(&task_group_lock, flags);
9889 for_each_possible_cpu(i) {
9890 register_fair_sched_group(tg, i);
9891 register_rt_sched_group(tg, i);
9893 list_add_rcu(&tg->list, &task_groups);
9895 WARN_ON(!parent); /* root should already exist */
9897 tg->parent = parent;
9898 INIT_LIST_HEAD(&tg->children);
9899 list_add_rcu(&tg->siblings, &parent->children);
9900 spin_unlock_irqrestore(&task_group_lock, flags);
9905 free_sched_group(tg);
9906 return ERR_PTR(-ENOMEM);
9909 /* rcu callback to free various structures associated with a task group */
9910 static void free_sched_group_rcu(struct rcu_head *rhp)
9912 /* now it should be safe to free those cfs_rqs */
9913 free_sched_group(container_of(rhp, struct task_group, rcu));
9916 /* Destroy runqueue etc associated with a task group */
9917 void sched_destroy_group(struct task_group *tg)
9919 unsigned long flags;
9922 spin_lock_irqsave(&task_group_lock, flags);
9923 for_each_possible_cpu(i) {
9924 unregister_fair_sched_group(tg, i);
9925 unregister_rt_sched_group(tg, i);
9927 list_del_rcu(&tg->list);
9928 list_del_rcu(&tg->siblings);
9929 spin_unlock_irqrestore(&task_group_lock, flags);
9931 /* wait for possible concurrent references to cfs_rqs complete */
9932 call_rcu(&tg->rcu, free_sched_group_rcu);
9935 /* change task's runqueue when it moves between groups.
9936 * The caller of this function should have put the task in its new group
9937 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9938 * reflect its new group.
9940 void sched_move_task(struct task_struct *tsk)
9943 unsigned long flags;
9946 rq = task_rq_lock(tsk, &flags);
9948 update_rq_clock(rq);
9950 running = task_current(rq, tsk);
9951 on_rq = tsk->se.on_rq;
9954 dequeue_task(rq, tsk, 0);
9955 if (unlikely(running))
9956 tsk->sched_class->put_prev_task(rq, tsk);
9958 set_task_rq(tsk, task_cpu(tsk));
9960 #ifdef CONFIG_FAIR_GROUP_SCHED
9961 if (tsk->sched_class->moved_group)
9962 tsk->sched_class->moved_group(tsk);
9965 if (unlikely(running))
9966 tsk->sched_class->set_curr_task(rq);
9968 enqueue_task(rq, tsk, 0);
9970 task_rq_unlock(rq, &flags);
9972 #endif /* CONFIG_GROUP_SCHED */
9974 #ifdef CONFIG_FAIR_GROUP_SCHED
9975 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9977 struct cfs_rq *cfs_rq = se->cfs_rq;
9982 dequeue_entity(cfs_rq, se, 0);
9984 se->load.weight = shares;
9985 se->load.inv_weight = 0;
9988 enqueue_entity(cfs_rq, se, 0);
9991 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9993 struct cfs_rq *cfs_rq = se->cfs_rq;
9994 struct rq *rq = cfs_rq->rq;
9995 unsigned long flags;
9997 spin_lock_irqsave(&rq->lock, flags);
9998 __set_se_shares(se, shares);
9999 spin_unlock_irqrestore(&rq->lock, flags);
10002 static DEFINE_MUTEX(shares_mutex);
10004 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10007 unsigned long flags;
10010 * We can't change the weight of the root cgroup.
10015 if (shares < MIN_SHARES)
10016 shares = MIN_SHARES;
10017 else if (shares > MAX_SHARES)
10018 shares = MAX_SHARES;
10020 mutex_lock(&shares_mutex);
10021 if (tg->shares == shares)
10024 spin_lock_irqsave(&task_group_lock, flags);
10025 for_each_possible_cpu(i)
10026 unregister_fair_sched_group(tg, i);
10027 list_del_rcu(&tg->siblings);
10028 spin_unlock_irqrestore(&task_group_lock, flags);
10030 /* wait for any ongoing reference to this group to finish */
10031 synchronize_sched();
10034 * Now we are free to modify the group's share on each cpu
10035 * w/o tripping rebalance_share or load_balance_fair.
10037 tg->shares = shares;
10038 for_each_possible_cpu(i) {
10040 * force a rebalance
10042 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10043 set_se_shares(tg->se[i], shares);
10047 * Enable load balance activity on this group, by inserting it back on
10048 * each cpu's rq->leaf_cfs_rq_list.
10050 spin_lock_irqsave(&task_group_lock, flags);
10051 for_each_possible_cpu(i)
10052 register_fair_sched_group(tg, i);
10053 list_add_rcu(&tg->siblings, &tg->parent->children);
10054 spin_unlock_irqrestore(&task_group_lock, flags);
10056 mutex_unlock(&shares_mutex);
10060 unsigned long sched_group_shares(struct task_group *tg)
10066 #ifdef CONFIG_RT_GROUP_SCHED
10068 * Ensure that the real time constraints are schedulable.
10070 static DEFINE_MUTEX(rt_constraints_mutex);
10072 static unsigned long to_ratio(u64 period, u64 runtime)
10074 if (runtime == RUNTIME_INF)
10077 return div64_u64(runtime << 20, period);
10080 /* Must be called with tasklist_lock held */
10081 static inline int tg_has_rt_tasks(struct task_group *tg)
10083 struct task_struct *g, *p;
10085 do_each_thread(g, p) {
10086 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10088 } while_each_thread(g, p);
10093 struct rt_schedulable_data {
10094 struct task_group *tg;
10099 static int tg_schedulable(struct task_group *tg, void *data)
10101 struct rt_schedulable_data *d = data;
10102 struct task_group *child;
10103 unsigned long total, sum = 0;
10104 u64 period, runtime;
10106 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10107 runtime = tg->rt_bandwidth.rt_runtime;
10110 period = d->rt_period;
10111 runtime = d->rt_runtime;
10114 #ifdef CONFIG_USER_SCHED
10115 if (tg == &root_task_group) {
10116 period = global_rt_period();
10117 runtime = global_rt_runtime();
10122 * Cannot have more runtime than the period.
10124 if (runtime > period && runtime != RUNTIME_INF)
10128 * Ensure we don't starve existing RT tasks.
10130 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10133 total = to_ratio(period, runtime);
10136 * Nobody can have more than the global setting allows.
10138 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10142 * The sum of our children's runtime should not exceed our own.
10144 list_for_each_entry_rcu(child, &tg->children, siblings) {
10145 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10146 runtime = child->rt_bandwidth.rt_runtime;
10148 if (child == d->tg) {
10149 period = d->rt_period;
10150 runtime = d->rt_runtime;
10153 sum += to_ratio(period, runtime);
10162 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10164 struct rt_schedulable_data data = {
10166 .rt_period = period,
10167 .rt_runtime = runtime,
10170 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10173 static int tg_set_bandwidth(struct task_group *tg,
10174 u64 rt_period, u64 rt_runtime)
10178 mutex_lock(&rt_constraints_mutex);
10179 read_lock(&tasklist_lock);
10180 err = __rt_schedulable(tg, rt_period, rt_runtime);
10184 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10185 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10186 tg->rt_bandwidth.rt_runtime = rt_runtime;
10188 for_each_possible_cpu(i) {
10189 struct rt_rq *rt_rq = tg->rt_rq[i];
10191 spin_lock(&rt_rq->rt_runtime_lock);
10192 rt_rq->rt_runtime = rt_runtime;
10193 spin_unlock(&rt_rq->rt_runtime_lock);
10195 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10197 read_unlock(&tasklist_lock);
10198 mutex_unlock(&rt_constraints_mutex);
10203 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10205 u64 rt_runtime, rt_period;
10207 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10208 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10209 if (rt_runtime_us < 0)
10210 rt_runtime = RUNTIME_INF;
10212 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10215 long sched_group_rt_runtime(struct task_group *tg)
10219 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10222 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10223 do_div(rt_runtime_us, NSEC_PER_USEC);
10224 return rt_runtime_us;
10227 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10229 u64 rt_runtime, rt_period;
10231 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10232 rt_runtime = tg->rt_bandwidth.rt_runtime;
10234 if (rt_period == 0)
10237 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10240 long sched_group_rt_period(struct task_group *tg)
10244 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10245 do_div(rt_period_us, NSEC_PER_USEC);
10246 return rt_period_us;
10249 static int sched_rt_global_constraints(void)
10251 u64 runtime, period;
10254 if (sysctl_sched_rt_period <= 0)
10257 runtime = global_rt_runtime();
10258 period = global_rt_period();
10261 * Sanity check on the sysctl variables.
10263 if (runtime > period && runtime != RUNTIME_INF)
10266 mutex_lock(&rt_constraints_mutex);
10267 read_lock(&tasklist_lock);
10268 ret = __rt_schedulable(NULL, 0, 0);
10269 read_unlock(&tasklist_lock);
10270 mutex_unlock(&rt_constraints_mutex);
10275 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10277 /* Don't accept realtime tasks when there is no way for them to run */
10278 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10284 #else /* !CONFIG_RT_GROUP_SCHED */
10285 static int sched_rt_global_constraints(void)
10287 unsigned long flags;
10290 if (sysctl_sched_rt_period <= 0)
10294 * There's always some RT tasks in the root group
10295 * -- migration, kstopmachine etc..
10297 if (sysctl_sched_rt_runtime == 0)
10300 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10301 for_each_possible_cpu(i) {
10302 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10304 spin_lock(&rt_rq->rt_runtime_lock);
10305 rt_rq->rt_runtime = global_rt_runtime();
10306 spin_unlock(&rt_rq->rt_runtime_lock);
10308 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10312 #endif /* CONFIG_RT_GROUP_SCHED */
10314 int sched_rt_handler(struct ctl_table *table, int write,
10315 void __user *buffer, size_t *lenp,
10319 int old_period, old_runtime;
10320 static DEFINE_MUTEX(mutex);
10322 mutex_lock(&mutex);
10323 old_period = sysctl_sched_rt_period;
10324 old_runtime = sysctl_sched_rt_runtime;
10326 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10328 if (!ret && write) {
10329 ret = sched_rt_global_constraints();
10331 sysctl_sched_rt_period = old_period;
10332 sysctl_sched_rt_runtime = old_runtime;
10334 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10335 def_rt_bandwidth.rt_period =
10336 ns_to_ktime(global_rt_period());
10339 mutex_unlock(&mutex);
10344 #ifdef CONFIG_CGROUP_SCHED
10346 /* return corresponding task_group object of a cgroup */
10347 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10349 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10350 struct task_group, css);
10353 static struct cgroup_subsys_state *
10354 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10356 struct task_group *tg, *parent;
10358 if (!cgrp->parent) {
10359 /* This is early initialization for the top cgroup */
10360 return &init_task_group.css;
10363 parent = cgroup_tg(cgrp->parent);
10364 tg = sched_create_group(parent);
10366 return ERR_PTR(-ENOMEM);
10372 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10374 struct task_group *tg = cgroup_tg(cgrp);
10376 sched_destroy_group(tg);
10380 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10382 #ifdef CONFIG_RT_GROUP_SCHED
10383 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10386 /* We don't support RT-tasks being in separate groups */
10387 if (tsk->sched_class != &fair_sched_class)
10394 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10395 struct task_struct *tsk, bool threadgroup)
10397 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10401 struct task_struct *c;
10403 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10404 retval = cpu_cgroup_can_attach_task(cgrp, c);
10416 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10417 struct cgroup *old_cont, struct task_struct *tsk,
10420 sched_move_task(tsk);
10422 struct task_struct *c;
10424 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10425 sched_move_task(c);
10431 #ifdef CONFIG_FAIR_GROUP_SCHED
10432 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10435 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10438 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10440 struct task_group *tg = cgroup_tg(cgrp);
10442 return (u64) tg->shares;
10444 #endif /* CONFIG_FAIR_GROUP_SCHED */
10446 #ifdef CONFIG_RT_GROUP_SCHED
10447 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10450 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10453 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10455 return sched_group_rt_runtime(cgroup_tg(cgrp));
10458 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10461 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10464 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10466 return sched_group_rt_period(cgroup_tg(cgrp));
10468 #endif /* CONFIG_RT_GROUP_SCHED */
10470 static struct cftype cpu_files[] = {
10471 #ifdef CONFIG_FAIR_GROUP_SCHED
10474 .read_u64 = cpu_shares_read_u64,
10475 .write_u64 = cpu_shares_write_u64,
10478 #ifdef CONFIG_RT_GROUP_SCHED
10480 .name = "rt_runtime_us",
10481 .read_s64 = cpu_rt_runtime_read,
10482 .write_s64 = cpu_rt_runtime_write,
10485 .name = "rt_period_us",
10486 .read_u64 = cpu_rt_period_read_uint,
10487 .write_u64 = cpu_rt_period_write_uint,
10492 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10494 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10497 struct cgroup_subsys cpu_cgroup_subsys = {
10499 .create = cpu_cgroup_create,
10500 .destroy = cpu_cgroup_destroy,
10501 .can_attach = cpu_cgroup_can_attach,
10502 .attach = cpu_cgroup_attach,
10503 .populate = cpu_cgroup_populate,
10504 .subsys_id = cpu_cgroup_subsys_id,
10508 #endif /* CONFIG_CGROUP_SCHED */
10510 #ifdef CONFIG_CGROUP_CPUACCT
10513 * CPU accounting code for task groups.
10515 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10516 * (balbir@in.ibm.com).
10519 /* track cpu usage of a group of tasks and its child groups */
10521 struct cgroup_subsys_state css;
10522 /* cpuusage holds pointer to a u64-type object on every cpu */
10524 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10525 struct cpuacct *parent;
10528 struct cgroup_subsys cpuacct_subsys;
10530 /* return cpu accounting group corresponding to this container */
10531 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10533 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10534 struct cpuacct, css);
10537 /* return cpu accounting group to which this task belongs */
10538 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10540 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10541 struct cpuacct, css);
10544 /* create a new cpu accounting group */
10545 static struct cgroup_subsys_state *cpuacct_create(
10546 struct cgroup_subsys *ss, struct cgroup *cgrp)
10548 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10554 ca->cpuusage = alloc_percpu(u64);
10558 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10559 if (percpu_counter_init(&ca->cpustat[i], 0))
10560 goto out_free_counters;
10563 ca->parent = cgroup_ca(cgrp->parent);
10569 percpu_counter_destroy(&ca->cpustat[i]);
10570 free_percpu(ca->cpuusage);
10574 return ERR_PTR(-ENOMEM);
10577 /* destroy an existing cpu accounting group */
10579 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10581 struct cpuacct *ca = cgroup_ca(cgrp);
10584 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10585 percpu_counter_destroy(&ca->cpustat[i]);
10586 free_percpu(ca->cpuusage);
10590 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10592 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10595 #ifndef CONFIG_64BIT
10597 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10599 spin_lock_irq(&cpu_rq(cpu)->lock);
10601 spin_unlock_irq(&cpu_rq(cpu)->lock);
10609 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10611 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10613 #ifndef CONFIG_64BIT
10615 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10617 spin_lock_irq(&cpu_rq(cpu)->lock);
10619 spin_unlock_irq(&cpu_rq(cpu)->lock);
10625 /* return total cpu usage (in nanoseconds) of a group */
10626 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10628 struct cpuacct *ca = cgroup_ca(cgrp);
10629 u64 totalcpuusage = 0;
10632 for_each_present_cpu(i)
10633 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10635 return totalcpuusage;
10638 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10641 struct cpuacct *ca = cgroup_ca(cgrp);
10650 for_each_present_cpu(i)
10651 cpuacct_cpuusage_write(ca, i, 0);
10657 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10658 struct seq_file *m)
10660 struct cpuacct *ca = cgroup_ca(cgroup);
10664 for_each_present_cpu(i) {
10665 percpu = cpuacct_cpuusage_read(ca, i);
10666 seq_printf(m, "%llu ", (unsigned long long) percpu);
10668 seq_printf(m, "\n");
10672 static const char *cpuacct_stat_desc[] = {
10673 [CPUACCT_STAT_USER] = "user",
10674 [CPUACCT_STAT_SYSTEM] = "system",
10677 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10678 struct cgroup_map_cb *cb)
10680 struct cpuacct *ca = cgroup_ca(cgrp);
10683 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10684 s64 val = percpu_counter_read(&ca->cpustat[i]);
10685 val = cputime64_to_clock_t(val);
10686 cb->fill(cb, cpuacct_stat_desc[i], val);
10691 static struct cftype files[] = {
10694 .read_u64 = cpuusage_read,
10695 .write_u64 = cpuusage_write,
10698 .name = "usage_percpu",
10699 .read_seq_string = cpuacct_percpu_seq_read,
10703 .read_map = cpuacct_stats_show,
10707 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10709 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10713 * charge this task's execution time to its accounting group.
10715 * called with rq->lock held.
10717 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10719 struct cpuacct *ca;
10722 if (unlikely(!cpuacct_subsys.active))
10725 cpu = task_cpu(tsk);
10731 for (; ca; ca = ca->parent) {
10732 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10733 *cpuusage += cputime;
10740 * Charge the system/user time to the task's accounting group.
10742 static void cpuacct_update_stats(struct task_struct *tsk,
10743 enum cpuacct_stat_index idx, cputime_t val)
10745 struct cpuacct *ca;
10747 if (unlikely(!cpuacct_subsys.active))
10754 percpu_counter_add(&ca->cpustat[idx], val);
10760 struct cgroup_subsys cpuacct_subsys = {
10762 .create = cpuacct_create,
10763 .destroy = cpuacct_destroy,
10764 .populate = cpuacct_populate,
10765 .subsys_id = cpuacct_subsys_id,
10767 #endif /* CONFIG_CGROUP_CPUACCT */
10771 int rcu_expedited_torture_stats(char *page)
10775 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10777 void synchronize_sched_expedited(void)
10780 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10782 #else /* #ifndef CONFIG_SMP */
10784 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10785 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10787 #define RCU_EXPEDITED_STATE_POST -2
10788 #define RCU_EXPEDITED_STATE_IDLE -1
10790 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10792 int rcu_expedited_torture_stats(char *page)
10797 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10798 for_each_online_cpu(cpu) {
10799 cnt += sprintf(&page[cnt], " %d:%d",
10800 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10802 cnt += sprintf(&page[cnt], "\n");
10805 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10807 static long synchronize_sched_expedited_count;
10810 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10811 * approach to force grace period to end quickly. This consumes
10812 * significant time on all CPUs, and is thus not recommended for
10813 * any sort of common-case code.
10815 * Note that it is illegal to call this function while holding any
10816 * lock that is acquired by a CPU-hotplug notifier. Failing to
10817 * observe this restriction will result in deadlock.
10819 void synchronize_sched_expedited(void)
10822 unsigned long flags;
10823 bool need_full_sync = 0;
10825 struct migration_req *req;
10829 smp_mb(); /* ensure prior mod happens before capturing snap. */
10830 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10832 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10834 if (trycount++ < 10)
10835 udelay(trycount * num_online_cpus());
10837 synchronize_sched();
10840 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10841 smp_mb(); /* ensure test happens before caller kfree */
10846 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10847 for_each_online_cpu(cpu) {
10849 req = &per_cpu(rcu_migration_req, cpu);
10850 init_completion(&req->done);
10852 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10853 spin_lock_irqsave(&rq->lock, flags);
10854 list_add(&req->list, &rq->migration_queue);
10855 spin_unlock_irqrestore(&rq->lock, flags);
10856 wake_up_process(rq->migration_thread);
10858 for_each_online_cpu(cpu) {
10859 rcu_expedited_state = cpu;
10860 req = &per_cpu(rcu_migration_req, cpu);
10862 wait_for_completion(&req->done);
10863 spin_lock_irqsave(&rq->lock, flags);
10864 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10865 need_full_sync = 1;
10866 req->dest_cpu = RCU_MIGRATION_IDLE;
10867 spin_unlock_irqrestore(&rq->lock, flags);
10869 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10870 mutex_unlock(&rcu_sched_expedited_mutex);
10872 if (need_full_sync)
10873 synchronize_sched();
10875 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10877 #endif /* #else #ifndef CONFIG_SMP */