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>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
123 static inline int rt_policy(int policy)
125 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
130 static inline int task_has_rt_policy(struct task_struct *p)
132 return rt_policy(p->policy);
136 * This is the priority-queue data structure of the RT scheduling class:
138 struct rt_prio_array {
139 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
140 struct list_head queue[MAX_RT_PRIO];
143 struct rt_bandwidth {
144 /* nests inside the rq lock: */
145 raw_spinlock_t rt_runtime_lock;
148 struct hrtimer rt_period_timer;
151 static struct rt_bandwidth def_rt_bandwidth;
153 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
155 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
157 struct rt_bandwidth *rt_b =
158 container_of(timer, struct rt_bandwidth, rt_period_timer);
164 now = hrtimer_cb_get_time(timer);
165 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
170 idle = do_sched_rt_period_timer(rt_b, overrun);
173 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
177 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
179 rt_b->rt_period = ns_to_ktime(period);
180 rt_b->rt_runtime = runtime;
182 raw_spin_lock_init(&rt_b->rt_runtime_lock);
184 hrtimer_init(&rt_b->rt_period_timer,
185 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
186 rt_b->rt_period_timer.function = sched_rt_period_timer;
189 static inline int rt_bandwidth_enabled(void)
191 return sysctl_sched_rt_runtime >= 0;
194 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
198 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
201 if (hrtimer_active(&rt_b->rt_period_timer))
204 raw_spin_lock(&rt_b->rt_runtime_lock);
209 if (hrtimer_active(&rt_b->rt_period_timer))
212 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
213 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
215 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
216 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
217 delta = ktime_to_ns(ktime_sub(hard, soft));
218 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
219 HRTIMER_MODE_ABS_PINNED, 0);
221 raw_spin_unlock(&rt_b->rt_runtime_lock);
224 #ifdef CONFIG_RT_GROUP_SCHED
225 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
227 hrtimer_cancel(&rt_b->rt_period_timer);
232 * sched_domains_mutex serializes calls to arch_init_sched_domains,
233 * detach_destroy_domains and partition_sched_domains.
235 static DEFINE_MUTEX(sched_domains_mutex);
237 #ifdef CONFIG_CGROUP_SCHED
239 #include <linux/cgroup.h>
243 static LIST_HEAD(task_groups);
245 /* task group related information */
247 struct cgroup_subsys_state css;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 /* schedulable entities of this group on each cpu */
251 struct sched_entity **se;
252 /* runqueue "owned" by this group on each cpu */
253 struct cfs_rq **cfs_rq;
254 unsigned long shares;
257 #ifdef CONFIG_RT_GROUP_SCHED
258 struct sched_rt_entity **rt_se;
259 struct rt_rq **rt_rq;
261 struct rt_bandwidth rt_bandwidth;
265 struct list_head list;
267 struct task_group *parent;
268 struct list_head siblings;
269 struct list_head children;
272 #define root_task_group init_task_group
274 /* task_group_lock serializes add/remove of task groups and also changes to
275 * a task group's cpu shares.
277 static DEFINE_SPINLOCK(task_group_lock);
279 #ifdef CONFIG_FAIR_GROUP_SCHED
282 static int root_task_group_empty(void)
284 return list_empty(&root_task_group.children);
288 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
299 #define MAX_SHARES (1UL << 18)
301 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group init_task_group;
309 /* return group to which a task belongs */
310 static inline struct task_group *task_group(struct task_struct *p)
312 struct task_group *tg;
314 #ifdef CONFIG_CGROUP_SCHED
315 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
316 struct task_group, css);
318 tg = &init_task_group;
323 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
324 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
326 #ifdef CONFIG_FAIR_GROUP_SCHED
327 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
328 p->se.parent = task_group(p)->se[cpu];
331 #ifdef CONFIG_RT_GROUP_SCHED
332 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
333 p->rt.parent = task_group(p)->rt_se[cpu];
339 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
340 static inline struct task_group *task_group(struct task_struct *p)
345 #endif /* CONFIG_CGROUP_SCHED */
347 /* CFS-related fields in a runqueue */
349 struct load_weight load;
350 unsigned long nr_running;
355 struct rb_root tasks_timeline;
356 struct rb_node *rb_leftmost;
358 struct list_head tasks;
359 struct list_head *balance_iterator;
362 * 'curr' points to currently running entity on this cfs_rq.
363 * It is set to NULL otherwise (i.e when none are currently running).
365 struct sched_entity *curr, *next, *last;
367 unsigned int nr_spread_over;
369 #ifdef CONFIG_FAIR_GROUP_SCHED
370 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
373 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
374 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
375 * (like users, containers etc.)
377 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
378 * list is used during load balance.
380 struct list_head leaf_cfs_rq_list;
381 struct task_group *tg; /* group that "owns" this runqueue */
385 * the part of load.weight contributed by tasks
387 unsigned long task_weight;
390 * h_load = weight * f(tg)
392 * Where f(tg) is the recursive weight fraction assigned to
395 unsigned long h_load;
398 * this cpu's part of tg->shares
400 unsigned long shares;
403 * load.weight at the time we set shares
405 unsigned long rq_weight;
410 /* Real-Time classes' related field in a runqueue: */
412 struct rt_prio_array active;
413 unsigned long rt_nr_running;
414 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
416 int curr; /* highest queued rt task prio */
418 int next; /* next highest */
423 unsigned long rt_nr_migratory;
424 unsigned long rt_nr_total;
426 struct plist_head pushable_tasks;
431 /* Nests inside the rq lock: */
432 raw_spinlock_t rt_runtime_lock;
434 #ifdef CONFIG_RT_GROUP_SCHED
435 unsigned long rt_nr_boosted;
438 struct list_head leaf_rt_rq_list;
439 struct task_group *tg;
446 * We add the notion of a root-domain which will be used to define per-domain
447 * variables. Each exclusive cpuset essentially defines an island domain by
448 * fully partitioning the member cpus from any other cpuset. Whenever a new
449 * exclusive cpuset is created, we also create and attach a new root-domain
456 cpumask_var_t online;
459 * The "RT overload" flag: it gets set if a CPU has more than
460 * one runnable RT task.
462 cpumask_var_t rto_mask;
465 struct cpupri cpupri;
470 * By default the system creates a single root-domain with all cpus as
471 * members (mimicking the global state we have today).
473 static struct root_domain def_root_domain;
478 * This is the main, per-CPU runqueue data structure.
480 * Locking rule: those places that want to lock multiple runqueues
481 * (such as the load balancing or the thread migration code), lock
482 * acquire operations must be ordered by ascending &runqueue.
489 * nr_running and cpu_load should be in the same cacheline because
490 * remote CPUs use both these fields when doing load calculation.
492 unsigned long nr_running;
493 #define CPU_LOAD_IDX_MAX 5
494 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
497 unsigned char in_nohz_recently;
499 unsigned int skip_clock_update;
501 /* capture load from *all* tasks on this cpu: */
502 struct load_weight load;
503 unsigned long nr_load_updates;
509 #ifdef CONFIG_FAIR_GROUP_SCHED
510 /* list of leaf cfs_rq on this cpu: */
511 struct list_head leaf_cfs_rq_list;
513 #ifdef CONFIG_RT_GROUP_SCHED
514 struct list_head leaf_rt_rq_list;
518 * This is part of a global counter where only the total sum
519 * over all CPUs matters. A task can increase this counter on
520 * one CPU and if it got migrated afterwards it may decrease
521 * it on another CPU. Always updated under the runqueue lock:
523 unsigned long nr_uninterruptible;
525 struct task_struct *curr, *idle;
526 unsigned long next_balance;
527 struct mm_struct *prev_mm;
534 struct root_domain *rd;
535 struct sched_domain *sd;
537 unsigned char idle_at_tick;
538 /* For active balancing */
542 /* cpu of this runqueue: */
546 unsigned long avg_load_per_task;
548 struct task_struct *migration_thread;
549 struct list_head migration_queue;
557 /* calc_load related fields */
558 unsigned long calc_load_update;
559 long calc_load_active;
561 #ifdef CONFIG_SCHED_HRTICK
563 int hrtick_csd_pending;
564 struct call_single_data hrtick_csd;
566 struct hrtimer hrtick_timer;
569 #ifdef CONFIG_SCHEDSTATS
571 struct sched_info rq_sched_info;
572 unsigned long long rq_cpu_time;
573 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
575 /* sys_sched_yield() stats */
576 unsigned int yld_count;
578 /* schedule() stats */
579 unsigned int sched_switch;
580 unsigned int sched_count;
581 unsigned int sched_goidle;
583 /* try_to_wake_up() stats */
584 unsigned int ttwu_count;
585 unsigned int ttwu_local;
588 unsigned int bkl_count;
592 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
595 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
597 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
600 * A queue event has occurred, and we're going to schedule. In
601 * this case, we can save a useless back to back clock update.
603 if (test_tsk_need_resched(p))
604 rq->skip_clock_update = 1;
607 static inline int cpu_of(struct rq *rq)
616 #define rcu_dereference_check_sched_domain(p) \
617 rcu_dereference_check((p), \
618 rcu_read_lock_sched_held() || \
619 lockdep_is_held(&sched_domains_mutex))
622 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
623 * See detach_destroy_domains: synchronize_sched for details.
625 * The domain tree of any CPU may only be accessed from within
626 * preempt-disabled sections.
628 #define for_each_domain(cpu, __sd) \
629 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
631 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
632 #define this_rq() (&__get_cpu_var(runqueues))
633 #define task_rq(p) cpu_rq(task_cpu(p))
634 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
635 #define raw_rq() (&__raw_get_cpu_var(runqueues))
637 inline void update_rq_clock(struct rq *rq)
639 if (!rq->skip_clock_update)
640 rq->clock = sched_clock_cpu(cpu_of(rq));
644 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
646 #ifdef CONFIG_SCHED_DEBUG
647 # define const_debug __read_mostly
649 # define const_debug static const
654 * @cpu: the processor in question.
656 * Returns true if the current cpu runqueue is locked.
657 * This interface allows printk to be called with the runqueue lock
658 * held and know whether or not it is OK to wake up the klogd.
660 int runqueue_is_locked(int cpu)
662 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
666 * Debugging: various feature bits
669 #define SCHED_FEAT(name, enabled) \
670 __SCHED_FEAT_##name ,
673 #include "sched_features.h"
678 #define SCHED_FEAT(name, enabled) \
679 (1UL << __SCHED_FEAT_##name) * enabled |
681 const_debug unsigned int sysctl_sched_features =
682 #include "sched_features.h"
687 #ifdef CONFIG_SCHED_DEBUG
688 #define SCHED_FEAT(name, enabled) \
691 static __read_mostly char *sched_feat_names[] = {
692 #include "sched_features.h"
698 static int sched_feat_show(struct seq_file *m, void *v)
702 for (i = 0; sched_feat_names[i]; i++) {
703 if (!(sysctl_sched_features & (1UL << i)))
705 seq_printf(m, "%s ", sched_feat_names[i]);
713 sched_feat_write(struct file *filp, const char __user *ubuf,
714 size_t cnt, loff_t *ppos)
724 if (copy_from_user(&buf, ubuf, cnt))
729 if (strncmp(buf, "NO_", 3) == 0) {
734 for (i = 0; sched_feat_names[i]; i++) {
735 int len = strlen(sched_feat_names[i]);
737 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
739 sysctl_sched_features &= ~(1UL << i);
741 sysctl_sched_features |= (1UL << i);
746 if (!sched_feat_names[i])
754 static int sched_feat_open(struct inode *inode, struct file *filp)
756 return single_open(filp, sched_feat_show, NULL);
759 static const struct file_operations sched_feat_fops = {
760 .open = sched_feat_open,
761 .write = sched_feat_write,
764 .release = single_release,
767 static __init int sched_init_debug(void)
769 debugfs_create_file("sched_features", 0644, NULL, NULL,
774 late_initcall(sched_init_debug);
778 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
781 * Number of tasks to iterate in a single balance run.
782 * Limited because this is done with IRQs disabled.
784 const_debug unsigned int sysctl_sched_nr_migrate = 32;
787 * ratelimit for updating the group shares.
790 unsigned int sysctl_sched_shares_ratelimit = 250000;
791 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
794 * Inject some fuzzyness into changing the per-cpu group shares
795 * this avoids remote rq-locks at the expense of fairness.
798 unsigned int sysctl_sched_shares_thresh = 4;
801 * period over which we average the RT time consumption, measured
806 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
809 * period over which we measure -rt task cpu usage in us.
812 unsigned int sysctl_sched_rt_period = 1000000;
814 static __read_mostly int scheduler_running;
817 * part of the period that we allow rt tasks to run in us.
820 int sysctl_sched_rt_runtime = 950000;
822 static inline u64 global_rt_period(void)
824 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
827 static inline u64 global_rt_runtime(void)
829 if (sysctl_sched_rt_runtime < 0)
832 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
835 #ifndef prepare_arch_switch
836 # define prepare_arch_switch(next) do { } while (0)
838 #ifndef finish_arch_switch
839 # define finish_arch_switch(prev) do { } while (0)
842 static inline int task_current(struct rq *rq, struct task_struct *p)
844 return rq->curr == p;
847 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
848 static inline int task_running(struct rq *rq, struct task_struct *p)
850 return task_current(rq, p);
853 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
857 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
859 #ifdef CONFIG_DEBUG_SPINLOCK
860 /* this is a valid case when another task releases the spinlock */
861 rq->lock.owner = current;
864 * If we are tracking spinlock dependencies then we have to
865 * fix up the runqueue lock - which gets 'carried over' from
868 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
870 raw_spin_unlock_irq(&rq->lock);
873 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
874 static inline int task_running(struct rq *rq, struct task_struct *p)
879 return task_current(rq, p);
883 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
887 * We can optimise this out completely for !SMP, because the
888 * SMP rebalancing from interrupt is the only thing that cares
893 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
894 raw_spin_unlock_irq(&rq->lock);
896 raw_spin_unlock(&rq->lock);
900 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
904 * After ->oncpu is cleared, the task can be moved to a different CPU.
905 * We must ensure this doesn't happen until the switch is completely
911 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
915 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
918 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
921 static inline int task_is_waking(struct task_struct *p)
923 return unlikely(p->state == TASK_WAKING);
927 * __task_rq_lock - lock the runqueue a given task resides on.
928 * Must be called interrupts disabled.
930 static inline struct rq *__task_rq_lock(struct task_struct *p)
937 raw_spin_lock(&rq->lock);
938 if (likely(rq == task_rq(p)))
940 raw_spin_unlock(&rq->lock);
945 * task_rq_lock - lock the runqueue a given task resides on and disable
946 * interrupts. Note the ordering: we can safely lookup the task_rq without
947 * explicitly disabling preemption.
949 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
955 local_irq_save(*flags);
957 raw_spin_lock(&rq->lock);
958 if (likely(rq == task_rq(p)))
960 raw_spin_unlock_irqrestore(&rq->lock, *flags);
964 void task_rq_unlock_wait(struct task_struct *p)
966 struct rq *rq = task_rq(p);
968 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
969 raw_spin_unlock_wait(&rq->lock);
972 static void __task_rq_unlock(struct rq *rq)
975 raw_spin_unlock(&rq->lock);
978 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
981 raw_spin_unlock_irqrestore(&rq->lock, *flags);
985 * this_rq_lock - lock this runqueue and disable interrupts.
987 static struct rq *this_rq_lock(void)
994 raw_spin_lock(&rq->lock);
999 #ifdef CONFIG_SCHED_HRTICK
1001 * Use HR-timers to deliver accurate preemption points.
1003 * Its all a bit involved since we cannot program an hrt while holding the
1004 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1007 * When we get rescheduled we reprogram the hrtick_timer outside of the
1013 * - enabled by features
1014 * - hrtimer is actually high res
1016 static inline int hrtick_enabled(struct rq *rq)
1018 if (!sched_feat(HRTICK))
1020 if (!cpu_active(cpu_of(rq)))
1022 return hrtimer_is_hres_active(&rq->hrtick_timer);
1025 static void hrtick_clear(struct rq *rq)
1027 if (hrtimer_active(&rq->hrtick_timer))
1028 hrtimer_cancel(&rq->hrtick_timer);
1032 * High-resolution timer tick.
1033 * Runs from hardirq context with interrupts disabled.
1035 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1037 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1039 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1041 raw_spin_lock(&rq->lock);
1042 update_rq_clock(rq);
1043 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1044 raw_spin_unlock(&rq->lock);
1046 return HRTIMER_NORESTART;
1051 * called from hardirq (IPI) context
1053 static void __hrtick_start(void *arg)
1055 struct rq *rq = arg;
1057 raw_spin_lock(&rq->lock);
1058 hrtimer_restart(&rq->hrtick_timer);
1059 rq->hrtick_csd_pending = 0;
1060 raw_spin_unlock(&rq->lock);
1064 * Called to set the hrtick timer state.
1066 * called with rq->lock held and irqs disabled
1068 static void hrtick_start(struct rq *rq, u64 delay)
1070 struct hrtimer *timer = &rq->hrtick_timer;
1071 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1073 hrtimer_set_expires(timer, time);
1075 if (rq == this_rq()) {
1076 hrtimer_restart(timer);
1077 } else if (!rq->hrtick_csd_pending) {
1078 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1079 rq->hrtick_csd_pending = 1;
1084 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1086 int cpu = (int)(long)hcpu;
1089 case CPU_UP_CANCELED:
1090 case CPU_UP_CANCELED_FROZEN:
1091 case CPU_DOWN_PREPARE:
1092 case CPU_DOWN_PREPARE_FROZEN:
1094 case CPU_DEAD_FROZEN:
1095 hrtick_clear(cpu_rq(cpu));
1102 static __init void init_hrtick(void)
1104 hotcpu_notifier(hotplug_hrtick, 0);
1108 * Called to set the hrtick timer state.
1110 * called with rq->lock held and irqs disabled
1112 static void hrtick_start(struct rq *rq, u64 delay)
1114 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1115 HRTIMER_MODE_REL_PINNED, 0);
1118 static inline void init_hrtick(void)
1121 #endif /* CONFIG_SMP */
1123 static void init_rq_hrtick(struct rq *rq)
1126 rq->hrtick_csd_pending = 0;
1128 rq->hrtick_csd.flags = 0;
1129 rq->hrtick_csd.func = __hrtick_start;
1130 rq->hrtick_csd.info = rq;
1133 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1134 rq->hrtick_timer.function = hrtick;
1136 #else /* CONFIG_SCHED_HRTICK */
1137 static inline void hrtick_clear(struct rq *rq)
1141 static inline void init_rq_hrtick(struct rq *rq)
1145 static inline void init_hrtick(void)
1148 #endif /* CONFIG_SCHED_HRTICK */
1151 * resched_task - mark a task 'to be rescheduled now'.
1153 * On UP this means the setting of the need_resched flag, on SMP it
1154 * might also involve a cross-CPU call to trigger the scheduler on
1159 #ifndef tsk_is_polling
1160 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1163 static void resched_task(struct task_struct *p)
1167 assert_raw_spin_locked(&task_rq(p)->lock);
1169 if (test_tsk_need_resched(p))
1172 set_tsk_need_resched(p);
1175 if (cpu == smp_processor_id())
1178 /* NEED_RESCHED must be visible before we test polling */
1180 if (!tsk_is_polling(p))
1181 smp_send_reschedule(cpu);
1184 static void resched_cpu(int cpu)
1186 struct rq *rq = cpu_rq(cpu);
1187 unsigned long flags;
1189 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1191 resched_task(cpu_curr(cpu));
1192 raw_spin_unlock_irqrestore(&rq->lock, flags);
1197 * When add_timer_on() enqueues a timer into the timer wheel of an
1198 * idle CPU then this timer might expire before the next timer event
1199 * which is scheduled to wake up that CPU. In case of a completely
1200 * idle system the next event might even be infinite time into the
1201 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1202 * leaves the inner idle loop so the newly added timer is taken into
1203 * account when the CPU goes back to idle and evaluates the timer
1204 * wheel for the next timer event.
1206 void wake_up_idle_cpu(int cpu)
1208 struct rq *rq = cpu_rq(cpu);
1210 if (cpu == smp_processor_id())
1214 * This is safe, as this function is called with the timer
1215 * wheel base lock of (cpu) held. When the CPU is on the way
1216 * to idle and has not yet set rq->curr to idle then it will
1217 * be serialized on the timer wheel base lock and take the new
1218 * timer into account automatically.
1220 if (rq->curr != rq->idle)
1224 * We can set TIF_RESCHED on the idle task of the other CPU
1225 * lockless. The worst case is that the other CPU runs the
1226 * idle task through an additional NOOP schedule()
1228 set_tsk_need_resched(rq->idle);
1230 /* NEED_RESCHED must be visible before we test polling */
1232 if (!tsk_is_polling(rq->idle))
1233 smp_send_reschedule(cpu);
1236 int nohz_ratelimit(int cpu)
1238 struct rq *rq = cpu_rq(cpu);
1239 u64 diff = rq->clock - rq->nohz_stamp;
1241 rq->nohz_stamp = rq->clock;
1243 return diff < (NSEC_PER_SEC / HZ) >> 1;
1246 #endif /* CONFIG_NO_HZ */
1248 static u64 sched_avg_period(void)
1250 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1253 static void sched_avg_update(struct rq *rq)
1255 s64 period = sched_avg_period();
1257 while ((s64)(rq->clock - rq->age_stamp) > period) {
1258 rq->age_stamp += period;
1263 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1265 rq->rt_avg += rt_delta;
1266 sched_avg_update(rq);
1269 #else /* !CONFIG_SMP */
1270 static void resched_task(struct task_struct *p)
1272 assert_raw_spin_locked(&task_rq(p)->lock);
1273 set_tsk_need_resched(p);
1276 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1279 #endif /* CONFIG_SMP */
1281 #if BITS_PER_LONG == 32
1282 # define WMULT_CONST (~0UL)
1284 # define WMULT_CONST (1UL << 32)
1287 #define WMULT_SHIFT 32
1290 * Shift right and round:
1292 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1295 * delta *= weight / lw
1297 static unsigned long
1298 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1299 struct load_weight *lw)
1303 if (!lw->inv_weight) {
1304 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1307 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1311 tmp = (u64)delta_exec * weight;
1313 * Check whether we'd overflow the 64-bit multiplication:
1315 if (unlikely(tmp > WMULT_CONST))
1316 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1319 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1321 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1324 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1330 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1337 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1338 * of tasks with abnormal "nice" values across CPUs the contribution that
1339 * each task makes to its run queue's load is weighted according to its
1340 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1341 * scaled version of the new time slice allocation that they receive on time
1345 #define WEIGHT_IDLEPRIO 3
1346 #define WMULT_IDLEPRIO 1431655765
1349 * Nice levels are multiplicative, with a gentle 10% change for every
1350 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1351 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1352 * that remained on nice 0.
1354 * The "10% effect" is relative and cumulative: from _any_ nice level,
1355 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1356 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1357 * If a task goes up by ~10% and another task goes down by ~10% then
1358 * the relative distance between them is ~25%.)
1360 static const int prio_to_weight[40] = {
1361 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1362 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1363 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1364 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1365 /* 0 */ 1024, 820, 655, 526, 423,
1366 /* 5 */ 335, 272, 215, 172, 137,
1367 /* 10 */ 110, 87, 70, 56, 45,
1368 /* 15 */ 36, 29, 23, 18, 15,
1372 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1374 * In cases where the weight does not change often, we can use the
1375 * precalculated inverse to speed up arithmetics by turning divisions
1376 * into multiplications:
1378 static const u32 prio_to_wmult[40] = {
1379 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1380 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1381 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1382 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1383 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1384 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1385 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1386 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1389 /* Time spent by the tasks of the cpu accounting group executing in ... */
1390 enum cpuacct_stat_index {
1391 CPUACCT_STAT_USER, /* ... user mode */
1392 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1394 CPUACCT_STAT_NSTATS,
1397 #ifdef CONFIG_CGROUP_CPUACCT
1398 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1399 static void cpuacct_update_stats(struct task_struct *tsk,
1400 enum cpuacct_stat_index idx, cputime_t val);
1402 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1403 static inline void cpuacct_update_stats(struct task_struct *tsk,
1404 enum cpuacct_stat_index idx, cputime_t val) {}
1407 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1409 update_load_add(&rq->load, load);
1412 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1414 update_load_sub(&rq->load, load);
1417 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1418 typedef int (*tg_visitor)(struct task_group *, void *);
1421 * Iterate the full tree, calling @down when first entering a node and @up when
1422 * leaving it for the final time.
1424 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1426 struct task_group *parent, *child;
1430 parent = &root_task_group;
1432 ret = (*down)(parent, data);
1435 list_for_each_entry_rcu(child, &parent->children, siblings) {
1442 ret = (*up)(parent, data);
1447 parent = parent->parent;
1456 static int tg_nop(struct task_group *tg, void *data)
1463 /* Used instead of source_load when we know the type == 0 */
1464 static unsigned long weighted_cpuload(const int cpu)
1466 return cpu_rq(cpu)->load.weight;
1470 * Return a low guess at the load of a migration-source cpu weighted
1471 * according to the scheduling class and "nice" value.
1473 * We want to under-estimate the load of migration sources, to
1474 * balance conservatively.
1476 static unsigned long source_load(int cpu, int type)
1478 struct rq *rq = cpu_rq(cpu);
1479 unsigned long total = weighted_cpuload(cpu);
1481 if (type == 0 || !sched_feat(LB_BIAS))
1484 return min(rq->cpu_load[type-1], total);
1488 * Return a high guess at the load of a migration-target cpu weighted
1489 * according to the scheduling class and "nice" value.
1491 static unsigned long target_load(int cpu, int type)
1493 struct rq *rq = cpu_rq(cpu);
1494 unsigned long total = weighted_cpuload(cpu);
1496 if (type == 0 || !sched_feat(LB_BIAS))
1499 return max(rq->cpu_load[type-1], total);
1502 static struct sched_group *group_of(int cpu)
1504 struct sched_domain *sd = rcu_dereference_sched(cpu_rq(cpu)->sd);
1512 static unsigned long power_of(int cpu)
1514 struct sched_group *group = group_of(cpu);
1517 return SCHED_LOAD_SCALE;
1519 return group->cpu_power;
1522 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1524 static unsigned long cpu_avg_load_per_task(int cpu)
1526 struct rq *rq = cpu_rq(cpu);
1527 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1530 rq->avg_load_per_task = rq->load.weight / nr_running;
1532 rq->avg_load_per_task = 0;
1534 return rq->avg_load_per_task;
1537 #ifdef CONFIG_FAIR_GROUP_SCHED
1539 static __read_mostly unsigned long __percpu *update_shares_data;
1541 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1544 * Calculate and set the cpu's group shares.
1546 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1547 unsigned long sd_shares,
1548 unsigned long sd_rq_weight,
1549 unsigned long *usd_rq_weight)
1551 unsigned long shares, rq_weight;
1554 rq_weight = usd_rq_weight[cpu];
1557 rq_weight = NICE_0_LOAD;
1561 * \Sum_j shares_j * rq_weight_i
1562 * shares_i = -----------------------------
1563 * \Sum_j rq_weight_j
1565 shares = (sd_shares * rq_weight) / sd_rq_weight;
1566 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1568 if (abs(shares - tg->se[cpu]->load.weight) >
1569 sysctl_sched_shares_thresh) {
1570 struct rq *rq = cpu_rq(cpu);
1571 unsigned long flags;
1573 raw_spin_lock_irqsave(&rq->lock, flags);
1574 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1575 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1576 __set_se_shares(tg->se[cpu], shares);
1577 raw_spin_unlock_irqrestore(&rq->lock, flags);
1582 * Re-compute the task group their per cpu shares over the given domain.
1583 * This needs to be done in a bottom-up fashion because the rq weight of a
1584 * parent group depends on the shares of its child groups.
1586 static int tg_shares_up(struct task_group *tg, void *data)
1588 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1589 unsigned long *usd_rq_weight;
1590 struct sched_domain *sd = data;
1591 unsigned long flags;
1597 local_irq_save(flags);
1598 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1600 for_each_cpu(i, sched_domain_span(sd)) {
1601 weight = tg->cfs_rq[i]->load.weight;
1602 usd_rq_weight[i] = weight;
1604 rq_weight += weight;
1606 * If there are currently no tasks on the cpu pretend there
1607 * is one of average load so that when a new task gets to
1608 * run here it will not get delayed by group starvation.
1611 weight = NICE_0_LOAD;
1613 sum_weight += weight;
1614 shares += tg->cfs_rq[i]->shares;
1618 rq_weight = sum_weight;
1620 if ((!shares && rq_weight) || shares > tg->shares)
1621 shares = tg->shares;
1623 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1624 shares = tg->shares;
1626 for_each_cpu(i, sched_domain_span(sd))
1627 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1629 local_irq_restore(flags);
1635 * Compute the cpu's hierarchical load factor for each task group.
1636 * This needs to be done in a top-down fashion because the load of a child
1637 * group is a fraction of its parents load.
1639 static int tg_load_down(struct task_group *tg, void *data)
1642 long cpu = (long)data;
1645 load = cpu_rq(cpu)->load.weight;
1647 load = tg->parent->cfs_rq[cpu]->h_load;
1648 load *= tg->cfs_rq[cpu]->shares;
1649 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1652 tg->cfs_rq[cpu]->h_load = load;
1657 static void update_shares(struct sched_domain *sd)
1662 if (root_task_group_empty())
1665 now = cpu_clock(raw_smp_processor_id());
1666 elapsed = now - sd->last_update;
1668 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1669 sd->last_update = now;
1670 walk_tg_tree(tg_nop, tg_shares_up, sd);
1674 static void update_h_load(long cpu)
1676 if (root_task_group_empty())
1679 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1684 static inline void update_shares(struct sched_domain *sd)
1690 #ifdef CONFIG_PREEMPT
1692 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1695 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1696 * way at the expense of forcing extra atomic operations in all
1697 * invocations. This assures that the double_lock is acquired using the
1698 * same underlying policy as the spinlock_t on this architecture, which
1699 * reduces latency compared to the unfair variant below. However, it
1700 * also adds more overhead and therefore may reduce throughput.
1702 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1703 __releases(this_rq->lock)
1704 __acquires(busiest->lock)
1705 __acquires(this_rq->lock)
1707 raw_spin_unlock(&this_rq->lock);
1708 double_rq_lock(this_rq, busiest);
1715 * Unfair double_lock_balance: Optimizes throughput at the expense of
1716 * latency by eliminating extra atomic operations when the locks are
1717 * already in proper order on entry. This favors lower cpu-ids and will
1718 * grant the double lock to lower cpus over higher ids under contention,
1719 * regardless of entry order into the function.
1721 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1722 __releases(this_rq->lock)
1723 __acquires(busiest->lock)
1724 __acquires(this_rq->lock)
1728 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1729 if (busiest < this_rq) {
1730 raw_spin_unlock(&this_rq->lock);
1731 raw_spin_lock(&busiest->lock);
1732 raw_spin_lock_nested(&this_rq->lock,
1733 SINGLE_DEPTH_NESTING);
1736 raw_spin_lock_nested(&busiest->lock,
1737 SINGLE_DEPTH_NESTING);
1742 #endif /* CONFIG_PREEMPT */
1745 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1747 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1749 if (unlikely(!irqs_disabled())) {
1750 /* printk() doesn't work good under rq->lock */
1751 raw_spin_unlock(&this_rq->lock);
1755 return _double_lock_balance(this_rq, busiest);
1758 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1759 __releases(busiest->lock)
1761 raw_spin_unlock(&busiest->lock);
1762 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1766 * double_rq_lock - safely lock two runqueues
1768 * Note this does not disable interrupts like task_rq_lock,
1769 * you need to do so manually before calling.
1771 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1772 __acquires(rq1->lock)
1773 __acquires(rq2->lock)
1775 BUG_ON(!irqs_disabled());
1777 raw_spin_lock(&rq1->lock);
1778 __acquire(rq2->lock); /* Fake it out ;) */
1781 raw_spin_lock(&rq1->lock);
1782 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1784 raw_spin_lock(&rq2->lock);
1785 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1791 * double_rq_unlock - safely unlock two runqueues
1793 * Note this does not restore interrupts like task_rq_unlock,
1794 * you need to do so manually after calling.
1796 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1797 __releases(rq1->lock)
1798 __releases(rq2->lock)
1800 raw_spin_unlock(&rq1->lock);
1802 raw_spin_unlock(&rq2->lock);
1804 __release(rq2->lock);
1809 #ifdef CONFIG_FAIR_GROUP_SCHED
1810 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1813 cfs_rq->shares = shares;
1818 static void calc_load_account_idle(struct rq *this_rq);
1819 static void update_sysctl(void);
1820 static int get_update_sysctl_factor(void);
1822 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1824 set_task_rq(p, cpu);
1827 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1828 * successfuly executed on another CPU. We must ensure that updates of
1829 * per-task data have been completed by this moment.
1832 task_thread_info(p)->cpu = cpu;
1836 static const struct sched_class rt_sched_class;
1838 #define sched_class_highest (&rt_sched_class)
1839 #define for_each_class(class) \
1840 for (class = sched_class_highest; class; class = class->next)
1842 #include "sched_stats.h"
1844 static void inc_nr_running(struct rq *rq)
1849 static void dec_nr_running(struct rq *rq)
1854 static void set_load_weight(struct task_struct *p)
1856 if (task_has_rt_policy(p)) {
1857 p->se.load.weight = prio_to_weight[0] * 2;
1858 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1863 * SCHED_IDLE tasks get minimal weight:
1865 if (p->policy == SCHED_IDLE) {
1866 p->se.load.weight = WEIGHT_IDLEPRIO;
1867 p->se.load.inv_weight = WMULT_IDLEPRIO;
1871 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1872 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1875 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1877 update_rq_clock(rq);
1878 sched_info_queued(p);
1879 p->sched_class->enqueue_task(rq, p, flags);
1883 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1885 update_rq_clock(rq);
1886 sched_info_dequeued(p);
1887 p->sched_class->dequeue_task(rq, p, flags);
1892 * activate_task - move a task to the runqueue.
1894 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1896 if (task_contributes_to_load(p))
1897 rq->nr_uninterruptible--;
1899 enqueue_task(rq, p, flags);
1904 * deactivate_task - remove a task from the runqueue.
1906 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1908 if (task_contributes_to_load(p))
1909 rq->nr_uninterruptible++;
1911 dequeue_task(rq, p, flags);
1915 #include "sched_idletask.c"
1916 #include "sched_fair.c"
1917 #include "sched_rt.c"
1918 #ifdef CONFIG_SCHED_DEBUG
1919 # include "sched_debug.c"
1923 * __normal_prio - return the priority that is based on the static prio
1925 static inline int __normal_prio(struct task_struct *p)
1927 return p->static_prio;
1931 * Calculate the expected normal priority: i.e. priority
1932 * without taking RT-inheritance into account. Might be
1933 * boosted by interactivity modifiers. Changes upon fork,
1934 * setprio syscalls, and whenever the interactivity
1935 * estimator recalculates.
1937 static inline int normal_prio(struct task_struct *p)
1941 if (task_has_rt_policy(p))
1942 prio = MAX_RT_PRIO-1 - p->rt_priority;
1944 prio = __normal_prio(p);
1949 * Calculate the current priority, i.e. the priority
1950 * taken into account by the scheduler. This value might
1951 * be boosted by RT tasks, or might be boosted by
1952 * interactivity modifiers. Will be RT if the task got
1953 * RT-boosted. If not then it returns p->normal_prio.
1955 static int effective_prio(struct task_struct *p)
1957 p->normal_prio = normal_prio(p);
1959 * If we are RT tasks or we were boosted to RT priority,
1960 * keep the priority unchanged. Otherwise, update priority
1961 * to the normal priority:
1963 if (!rt_prio(p->prio))
1964 return p->normal_prio;
1969 * task_curr - is this task currently executing on a CPU?
1970 * @p: the task in question.
1972 inline int task_curr(const struct task_struct *p)
1974 return cpu_curr(task_cpu(p)) == p;
1977 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1978 const struct sched_class *prev_class,
1979 int oldprio, int running)
1981 if (prev_class != p->sched_class) {
1982 if (prev_class->switched_from)
1983 prev_class->switched_from(rq, p, running);
1984 p->sched_class->switched_to(rq, p, running);
1986 p->sched_class->prio_changed(rq, p, oldprio, running);
1991 * Is this task likely cache-hot:
1994 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1998 if (p->sched_class != &fair_sched_class)
2002 * Buddy candidates are cache hot:
2004 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2005 (&p->se == cfs_rq_of(&p->se)->next ||
2006 &p->se == cfs_rq_of(&p->se)->last))
2009 if (sysctl_sched_migration_cost == -1)
2011 if (sysctl_sched_migration_cost == 0)
2014 delta = now - p->se.exec_start;
2016 return delta < (s64)sysctl_sched_migration_cost;
2019 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2021 #ifdef CONFIG_SCHED_DEBUG
2023 * We should never call set_task_cpu() on a blocked task,
2024 * ttwu() will sort out the placement.
2026 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2027 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2030 trace_sched_migrate_task(p, new_cpu);
2032 if (task_cpu(p) != new_cpu) {
2033 p->se.nr_migrations++;
2034 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2037 __set_task_cpu(p, new_cpu);
2040 struct migration_req {
2041 struct list_head list;
2043 struct task_struct *task;
2046 struct completion done;
2050 * The task's runqueue lock must be held.
2051 * Returns true if you have to wait for migration thread.
2054 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2056 struct rq *rq = task_rq(p);
2059 * If the task is not on a runqueue (and not running), then
2060 * the next wake-up will properly place the task.
2062 if (!p->se.on_rq && !task_running(rq, p))
2065 init_completion(&req->done);
2067 req->dest_cpu = dest_cpu;
2068 list_add(&req->list, &rq->migration_queue);
2074 * wait_task_context_switch - wait for a thread to complete at least one
2077 * @p must not be current.
2079 void wait_task_context_switch(struct task_struct *p)
2081 unsigned long nvcsw, nivcsw, flags;
2089 * The runqueue is assigned before the actual context
2090 * switch. We need to take the runqueue lock.
2092 * We could check initially without the lock but it is
2093 * very likely that we need to take the lock in every
2096 rq = task_rq_lock(p, &flags);
2097 running = task_running(rq, p);
2098 task_rq_unlock(rq, &flags);
2100 if (likely(!running))
2103 * The switch count is incremented before the actual
2104 * context switch. We thus wait for two switches to be
2105 * sure at least one completed.
2107 if ((p->nvcsw - nvcsw) > 1)
2109 if ((p->nivcsw - nivcsw) > 1)
2117 * wait_task_inactive - wait for a thread to unschedule.
2119 * If @match_state is nonzero, it's the @p->state value just checked and
2120 * not expected to change. If it changes, i.e. @p might have woken up,
2121 * then return zero. When we succeed in waiting for @p to be off its CPU,
2122 * we return a positive number (its total switch count). If a second call
2123 * a short while later returns the same number, the caller can be sure that
2124 * @p has remained unscheduled the whole time.
2126 * The caller must ensure that the task *will* unschedule sometime soon,
2127 * else this function might spin for a *long* time. This function can't
2128 * be called with interrupts off, or it may introduce deadlock with
2129 * smp_call_function() if an IPI is sent by the same process we are
2130 * waiting to become inactive.
2132 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2134 unsigned long flags;
2141 * We do the initial early heuristics without holding
2142 * any task-queue locks at all. We'll only try to get
2143 * the runqueue lock when things look like they will
2149 * If the task is actively running on another CPU
2150 * still, just relax and busy-wait without holding
2153 * NOTE! Since we don't hold any locks, it's not
2154 * even sure that "rq" stays as the right runqueue!
2155 * But we don't care, since "task_running()" will
2156 * return false if the runqueue has changed and p
2157 * is actually now running somewhere else!
2159 while (task_running(rq, p)) {
2160 if (match_state && unlikely(p->state != match_state))
2166 * Ok, time to look more closely! We need the rq
2167 * lock now, to be *sure*. If we're wrong, we'll
2168 * just go back and repeat.
2170 rq = task_rq_lock(p, &flags);
2171 trace_sched_wait_task(rq, p);
2172 running = task_running(rq, p);
2173 on_rq = p->se.on_rq;
2175 if (!match_state || p->state == match_state)
2176 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2177 task_rq_unlock(rq, &flags);
2180 * If it changed from the expected state, bail out now.
2182 if (unlikely(!ncsw))
2186 * Was it really running after all now that we
2187 * checked with the proper locks actually held?
2189 * Oops. Go back and try again..
2191 if (unlikely(running)) {
2197 * It's not enough that it's not actively running,
2198 * it must be off the runqueue _entirely_, and not
2201 * So if it was still runnable (but just not actively
2202 * running right now), it's preempted, and we should
2203 * yield - it could be a while.
2205 if (unlikely(on_rq)) {
2206 schedule_timeout_uninterruptible(1);
2211 * Ahh, all good. It wasn't running, and it wasn't
2212 * runnable, which means that it will never become
2213 * running in the future either. We're all done!
2222 * kick_process - kick a running thread to enter/exit the kernel
2223 * @p: the to-be-kicked thread
2225 * Cause a process which is running on another CPU to enter
2226 * kernel-mode, without any delay. (to get signals handled.)
2228 * NOTE: this function doesnt have to take the runqueue lock,
2229 * because all it wants to ensure is that the remote task enters
2230 * the kernel. If the IPI races and the task has been migrated
2231 * to another CPU then no harm is done and the purpose has been
2234 void kick_process(struct task_struct *p)
2240 if ((cpu != smp_processor_id()) && task_curr(p))
2241 smp_send_reschedule(cpu);
2244 EXPORT_SYMBOL_GPL(kick_process);
2245 #endif /* CONFIG_SMP */
2248 * task_oncpu_function_call - call a function on the cpu on which a task runs
2249 * @p: the task to evaluate
2250 * @func: the function to be called
2251 * @info: the function call argument
2253 * Calls the function @func when the task is currently running. This might
2254 * be on the current CPU, which just calls the function directly
2256 void task_oncpu_function_call(struct task_struct *p,
2257 void (*func) (void *info), void *info)
2264 smp_call_function_single(cpu, func, info, 1);
2270 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2272 static int select_fallback_rq(int cpu, struct task_struct *p)
2275 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2277 /* Look for allowed, online CPU in same node. */
2278 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2279 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2282 /* Any allowed, online CPU? */
2283 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2284 if (dest_cpu < nr_cpu_ids)
2287 /* No more Mr. Nice Guy. */
2288 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2289 dest_cpu = cpuset_cpus_allowed_fallback(p);
2291 * Don't tell them about moving exiting tasks or
2292 * kernel threads (both mm NULL), since they never
2295 if (p->mm && printk_ratelimit()) {
2296 printk(KERN_INFO "process %d (%s) no "
2297 "longer affine to cpu%d\n",
2298 task_pid_nr(p), p->comm, cpu);
2306 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2309 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2311 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2314 * In order not to call set_task_cpu() on a blocking task we need
2315 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2318 * Since this is common to all placement strategies, this lives here.
2320 * [ this allows ->select_task() to simply return task_cpu(p) and
2321 * not worry about this generic constraint ]
2323 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2325 cpu = select_fallback_rq(task_cpu(p), p);
2330 static void update_avg(u64 *avg, u64 sample)
2332 s64 diff = sample - *avg;
2338 * try_to_wake_up - wake up a thread
2339 * @p: the to-be-woken-up thread
2340 * @state: the mask of task states that can be woken
2341 * @sync: do a synchronous wakeup?
2343 * Put it on the run-queue if it's not already there. The "current"
2344 * thread is always on the run-queue (except when the actual
2345 * re-schedule is in progress), and as such you're allowed to do
2346 * the simpler "current->state = TASK_RUNNING" to mark yourself
2347 * runnable without the overhead of this.
2349 * returns failure only if the task is already active.
2351 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2354 int cpu, orig_cpu, this_cpu, success = 0;
2355 unsigned long flags;
2356 unsigned long en_flags = ENQUEUE_WAKEUP;
2359 this_cpu = get_cpu();
2362 rq = task_rq_lock(p, &flags);
2363 if (!(p->state & state))
2373 if (unlikely(task_running(rq, p)))
2377 * In order to handle concurrent wakeups and release the rq->lock
2378 * we put the task in TASK_WAKING state.
2380 * First fix up the nr_uninterruptible count:
2382 if (task_contributes_to_load(p)) {
2383 if (likely(cpu_online(orig_cpu)))
2384 rq->nr_uninterruptible--;
2386 this_rq()->nr_uninterruptible--;
2388 p->state = TASK_WAKING;
2390 if (p->sched_class->task_waking) {
2391 p->sched_class->task_waking(rq, p);
2392 en_flags |= ENQUEUE_WAKING;
2395 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2396 if (cpu != orig_cpu)
2397 set_task_cpu(p, cpu);
2398 __task_rq_unlock(rq);
2401 raw_spin_lock(&rq->lock);
2404 * We migrated the task without holding either rq->lock, however
2405 * since the task is not on the task list itself, nobody else
2406 * will try and migrate the task, hence the rq should match the
2407 * cpu we just moved it to.
2409 WARN_ON(task_cpu(p) != cpu);
2410 WARN_ON(p->state != TASK_WAKING);
2412 #ifdef CONFIG_SCHEDSTATS
2413 schedstat_inc(rq, ttwu_count);
2414 if (cpu == this_cpu)
2415 schedstat_inc(rq, ttwu_local);
2417 struct sched_domain *sd;
2418 for_each_domain(this_cpu, sd) {
2419 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2420 schedstat_inc(sd, ttwu_wake_remote);
2425 #endif /* CONFIG_SCHEDSTATS */
2428 #endif /* CONFIG_SMP */
2429 schedstat_inc(p, se.statistics.nr_wakeups);
2430 if (wake_flags & WF_SYNC)
2431 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2432 if (orig_cpu != cpu)
2433 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2434 if (cpu == this_cpu)
2435 schedstat_inc(p, se.statistics.nr_wakeups_local);
2437 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2438 activate_task(rq, p, en_flags);
2442 trace_sched_wakeup(rq, p, success);
2443 check_preempt_curr(rq, p, wake_flags);
2445 p->state = TASK_RUNNING;
2447 if (p->sched_class->task_woken)
2448 p->sched_class->task_woken(rq, p);
2450 if (unlikely(rq->idle_stamp)) {
2451 u64 delta = rq->clock - rq->idle_stamp;
2452 u64 max = 2*sysctl_sched_migration_cost;
2457 update_avg(&rq->avg_idle, delta);
2462 task_rq_unlock(rq, &flags);
2469 * wake_up_process - Wake up a specific process
2470 * @p: The process to be woken up.
2472 * Attempt to wake up the nominated process and move it to the set of runnable
2473 * processes. Returns 1 if the process was woken up, 0 if it was already
2476 * It may be assumed that this function implies a write memory barrier before
2477 * changing the task state if and only if any tasks are woken up.
2479 int wake_up_process(struct task_struct *p)
2481 return try_to_wake_up(p, TASK_ALL, 0);
2483 EXPORT_SYMBOL(wake_up_process);
2485 int wake_up_state(struct task_struct *p, unsigned int state)
2487 return try_to_wake_up(p, state, 0);
2491 * Perform scheduler related setup for a newly forked process p.
2492 * p is forked by current.
2494 * __sched_fork() is basic setup used by init_idle() too:
2496 static void __sched_fork(struct task_struct *p)
2498 p->se.exec_start = 0;
2499 p->se.sum_exec_runtime = 0;
2500 p->se.prev_sum_exec_runtime = 0;
2501 p->se.nr_migrations = 0;
2503 #ifdef CONFIG_SCHEDSTATS
2504 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2507 INIT_LIST_HEAD(&p->rt.run_list);
2509 INIT_LIST_HEAD(&p->se.group_node);
2511 #ifdef CONFIG_PREEMPT_NOTIFIERS
2512 INIT_HLIST_HEAD(&p->preempt_notifiers);
2517 * fork()/clone()-time setup:
2519 void sched_fork(struct task_struct *p, int clone_flags)
2521 int cpu = get_cpu();
2525 * We mark the process as running here. This guarantees that
2526 * nobody will actually run it, and a signal or other external
2527 * event cannot wake it up and insert it on the runqueue either.
2529 p->state = TASK_RUNNING;
2532 * Revert to default priority/policy on fork if requested.
2534 if (unlikely(p->sched_reset_on_fork)) {
2535 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2536 p->policy = SCHED_NORMAL;
2537 p->normal_prio = p->static_prio;
2540 if (PRIO_TO_NICE(p->static_prio) < 0) {
2541 p->static_prio = NICE_TO_PRIO(0);
2542 p->normal_prio = p->static_prio;
2547 * We don't need the reset flag anymore after the fork. It has
2548 * fulfilled its duty:
2550 p->sched_reset_on_fork = 0;
2554 * Make sure we do not leak PI boosting priority to the child.
2556 p->prio = current->normal_prio;
2558 if (!rt_prio(p->prio))
2559 p->sched_class = &fair_sched_class;
2561 if (p->sched_class->task_fork)
2562 p->sched_class->task_fork(p);
2564 set_task_cpu(p, cpu);
2566 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2567 if (likely(sched_info_on()))
2568 memset(&p->sched_info, 0, sizeof(p->sched_info));
2570 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2573 #ifdef CONFIG_PREEMPT
2574 /* Want to start with kernel preemption disabled. */
2575 task_thread_info(p)->preempt_count = 1;
2577 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2583 * wake_up_new_task - wake up a newly created task for the first time.
2585 * This function will do some initial scheduler statistics housekeeping
2586 * that must be done for every newly created context, then puts the task
2587 * on the runqueue and wakes it.
2589 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2591 unsigned long flags;
2593 int cpu __maybe_unused = get_cpu();
2596 rq = task_rq_lock(p, &flags);
2597 p->state = TASK_WAKING;
2600 * Fork balancing, do it here and not earlier because:
2601 * - cpus_allowed can change in the fork path
2602 * - any previously selected cpu might disappear through hotplug
2604 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2605 * without people poking at ->cpus_allowed.
2607 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2608 set_task_cpu(p, cpu);
2610 p->state = TASK_RUNNING;
2611 task_rq_unlock(rq, &flags);
2614 rq = task_rq_lock(p, &flags);
2615 activate_task(rq, p, 0);
2616 trace_sched_wakeup_new(rq, p, 1);
2617 check_preempt_curr(rq, p, WF_FORK);
2619 if (p->sched_class->task_woken)
2620 p->sched_class->task_woken(rq, p);
2622 task_rq_unlock(rq, &flags);
2626 #ifdef CONFIG_PREEMPT_NOTIFIERS
2629 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2630 * @notifier: notifier struct to register
2632 void preempt_notifier_register(struct preempt_notifier *notifier)
2634 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2636 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2639 * preempt_notifier_unregister - no longer interested in preemption notifications
2640 * @notifier: notifier struct to unregister
2642 * This is safe to call from within a preemption notifier.
2644 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2646 hlist_del(¬ifier->link);
2648 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2650 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2652 struct preempt_notifier *notifier;
2653 struct hlist_node *node;
2655 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2656 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2660 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2661 struct task_struct *next)
2663 struct preempt_notifier *notifier;
2664 struct hlist_node *node;
2666 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2667 notifier->ops->sched_out(notifier, next);
2670 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2672 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2677 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2678 struct task_struct *next)
2682 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2685 * prepare_task_switch - prepare to switch tasks
2686 * @rq: the runqueue preparing to switch
2687 * @prev: the current task that is being switched out
2688 * @next: the task we are going to switch to.
2690 * This is called with the rq lock held and interrupts off. It must
2691 * be paired with a subsequent finish_task_switch after the context
2694 * prepare_task_switch sets up locking and calls architecture specific
2698 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2699 struct task_struct *next)
2701 fire_sched_out_preempt_notifiers(prev, next);
2702 prepare_lock_switch(rq, next);
2703 prepare_arch_switch(next);
2707 * finish_task_switch - clean up after a task-switch
2708 * @rq: runqueue associated with task-switch
2709 * @prev: the thread we just switched away from.
2711 * finish_task_switch must be called after the context switch, paired
2712 * with a prepare_task_switch call before the context switch.
2713 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2714 * and do any other architecture-specific cleanup actions.
2716 * Note that we may have delayed dropping an mm in context_switch(). If
2717 * so, we finish that here outside of the runqueue lock. (Doing it
2718 * with the lock held can cause deadlocks; see schedule() for
2721 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2722 __releases(rq->lock)
2724 struct mm_struct *mm = rq->prev_mm;
2730 * A task struct has one reference for the use as "current".
2731 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2732 * schedule one last time. The schedule call will never return, and
2733 * the scheduled task must drop that reference.
2734 * The test for TASK_DEAD must occur while the runqueue locks are
2735 * still held, otherwise prev could be scheduled on another cpu, die
2736 * there before we look at prev->state, and then the reference would
2738 * Manfred Spraul <manfred@colorfullife.com>
2740 prev_state = prev->state;
2741 finish_arch_switch(prev);
2742 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2743 local_irq_disable();
2744 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2745 perf_event_task_sched_in(current);
2746 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2748 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2749 finish_lock_switch(rq, prev);
2751 fire_sched_in_preempt_notifiers(current);
2754 if (unlikely(prev_state == TASK_DEAD)) {
2756 * Remove function-return probe instances associated with this
2757 * task and put them back on the free list.
2759 kprobe_flush_task(prev);
2760 put_task_struct(prev);
2766 /* assumes rq->lock is held */
2767 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2769 if (prev->sched_class->pre_schedule)
2770 prev->sched_class->pre_schedule(rq, prev);
2773 /* rq->lock is NOT held, but preemption is disabled */
2774 static inline void post_schedule(struct rq *rq)
2776 if (rq->post_schedule) {
2777 unsigned long flags;
2779 raw_spin_lock_irqsave(&rq->lock, flags);
2780 if (rq->curr->sched_class->post_schedule)
2781 rq->curr->sched_class->post_schedule(rq);
2782 raw_spin_unlock_irqrestore(&rq->lock, flags);
2784 rq->post_schedule = 0;
2790 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2794 static inline void post_schedule(struct rq *rq)
2801 * schedule_tail - first thing a freshly forked thread must call.
2802 * @prev: the thread we just switched away from.
2804 asmlinkage void schedule_tail(struct task_struct *prev)
2805 __releases(rq->lock)
2807 struct rq *rq = this_rq();
2809 finish_task_switch(rq, prev);
2812 * FIXME: do we need to worry about rq being invalidated by the
2817 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2818 /* In this case, finish_task_switch does not reenable preemption */
2821 if (current->set_child_tid)
2822 put_user(task_pid_vnr(current), current->set_child_tid);
2826 * context_switch - switch to the new MM and the new
2827 * thread's register state.
2830 context_switch(struct rq *rq, struct task_struct *prev,
2831 struct task_struct *next)
2833 struct mm_struct *mm, *oldmm;
2835 prepare_task_switch(rq, prev, next);
2836 trace_sched_switch(rq, prev, next);
2838 oldmm = prev->active_mm;
2840 * For paravirt, this is coupled with an exit in switch_to to
2841 * combine the page table reload and the switch backend into
2844 arch_start_context_switch(prev);
2847 next->active_mm = oldmm;
2848 atomic_inc(&oldmm->mm_count);
2849 enter_lazy_tlb(oldmm, next);
2851 switch_mm(oldmm, mm, next);
2853 if (likely(!prev->mm)) {
2854 prev->active_mm = NULL;
2855 rq->prev_mm = oldmm;
2858 * Since the runqueue lock will be released by the next
2859 * task (which is an invalid locking op but in the case
2860 * of the scheduler it's an obvious special-case), so we
2861 * do an early lockdep release here:
2863 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2864 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2867 /* Here we just switch the register state and the stack. */
2868 switch_to(prev, next, prev);
2872 * this_rq must be evaluated again because prev may have moved
2873 * CPUs since it called schedule(), thus the 'rq' on its stack
2874 * frame will be invalid.
2876 finish_task_switch(this_rq(), prev);
2880 * nr_running, nr_uninterruptible and nr_context_switches:
2882 * externally visible scheduler statistics: current number of runnable
2883 * threads, current number of uninterruptible-sleeping threads, total
2884 * number of context switches performed since bootup.
2886 unsigned long nr_running(void)
2888 unsigned long i, sum = 0;
2890 for_each_online_cpu(i)
2891 sum += cpu_rq(i)->nr_running;
2896 unsigned long nr_uninterruptible(void)
2898 unsigned long i, sum = 0;
2900 for_each_possible_cpu(i)
2901 sum += cpu_rq(i)->nr_uninterruptible;
2904 * Since we read the counters lockless, it might be slightly
2905 * inaccurate. Do not allow it to go below zero though:
2907 if (unlikely((long)sum < 0))
2913 unsigned long long nr_context_switches(void)
2916 unsigned long long sum = 0;
2918 for_each_possible_cpu(i)
2919 sum += cpu_rq(i)->nr_switches;
2924 unsigned long nr_iowait(void)
2926 unsigned long i, sum = 0;
2928 for_each_possible_cpu(i)
2929 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2934 unsigned long nr_iowait_cpu(void)
2936 struct rq *this = this_rq();
2937 return atomic_read(&this->nr_iowait);
2940 unsigned long this_cpu_load(void)
2942 struct rq *this = this_rq();
2943 return this->cpu_load[0];
2947 /* Variables and functions for calc_load */
2948 static atomic_long_t calc_load_tasks;
2949 static unsigned long calc_load_update;
2950 unsigned long avenrun[3];
2951 EXPORT_SYMBOL(avenrun);
2953 static long calc_load_fold_active(struct rq *this_rq)
2955 long nr_active, delta = 0;
2957 nr_active = this_rq->nr_running;
2958 nr_active += (long) this_rq->nr_uninterruptible;
2960 if (nr_active != this_rq->calc_load_active) {
2961 delta = nr_active - this_rq->calc_load_active;
2962 this_rq->calc_load_active = nr_active;
2970 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2972 * When making the ILB scale, we should try to pull this in as well.
2974 static atomic_long_t calc_load_tasks_idle;
2976 static void calc_load_account_idle(struct rq *this_rq)
2980 delta = calc_load_fold_active(this_rq);
2982 atomic_long_add(delta, &calc_load_tasks_idle);
2985 static long calc_load_fold_idle(void)
2990 * Its got a race, we don't care...
2992 if (atomic_long_read(&calc_load_tasks_idle))
2993 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2998 static void calc_load_account_idle(struct rq *this_rq)
3002 static inline long calc_load_fold_idle(void)
3009 * get_avenrun - get the load average array
3010 * @loads: pointer to dest load array
3011 * @offset: offset to add
3012 * @shift: shift count to shift the result left
3014 * These values are estimates at best, so no need for locking.
3016 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3018 loads[0] = (avenrun[0] + offset) << shift;
3019 loads[1] = (avenrun[1] + offset) << shift;
3020 loads[2] = (avenrun[2] + offset) << shift;
3023 static unsigned long
3024 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3027 load += active * (FIXED_1 - exp);
3028 return load >> FSHIFT;
3032 * calc_load - update the avenrun load estimates 10 ticks after the
3033 * CPUs have updated calc_load_tasks.
3035 void calc_global_load(void)
3037 unsigned long upd = calc_load_update + 10;
3040 if (time_before(jiffies, upd))
3043 active = atomic_long_read(&calc_load_tasks);
3044 active = active > 0 ? active * FIXED_1 : 0;
3046 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3047 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3048 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3050 calc_load_update += LOAD_FREQ;
3054 * Called from update_cpu_load() to periodically update this CPU's
3057 static void calc_load_account_active(struct rq *this_rq)
3061 if (time_before(jiffies, this_rq->calc_load_update))
3064 delta = calc_load_fold_active(this_rq);
3065 delta += calc_load_fold_idle();
3067 atomic_long_add(delta, &calc_load_tasks);
3069 this_rq->calc_load_update += LOAD_FREQ;
3073 * Update rq->cpu_load[] statistics. This function is usually called every
3074 * scheduler tick (TICK_NSEC).
3076 static void update_cpu_load(struct rq *this_rq)
3078 unsigned long this_load = this_rq->load.weight;
3081 this_rq->nr_load_updates++;
3083 /* Update our load: */
3084 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3085 unsigned long old_load, new_load;
3087 /* scale is effectively 1 << i now, and >> i divides by scale */
3089 old_load = this_rq->cpu_load[i];
3090 new_load = this_load;
3092 * Round up the averaging division if load is increasing. This
3093 * prevents us from getting stuck on 9 if the load is 10, for
3096 if (new_load > old_load)
3097 new_load += scale-1;
3098 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3101 calc_load_account_active(this_rq);
3107 * sched_exec - execve() is a valuable balancing opportunity, because at
3108 * this point the task has the smallest effective memory and cache footprint.
3110 void sched_exec(void)
3112 struct task_struct *p = current;
3113 struct migration_req req;
3114 unsigned long flags;
3118 rq = task_rq_lock(p, &flags);
3119 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3120 if (dest_cpu == smp_processor_id())
3124 * select_task_rq() can race against ->cpus_allowed
3126 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3127 likely(cpu_active(dest_cpu)) &&
3128 migrate_task(p, dest_cpu, &req)) {
3129 /* Need to wait for migration thread (might exit: take ref). */
3130 struct task_struct *mt = rq->migration_thread;
3132 get_task_struct(mt);
3133 task_rq_unlock(rq, &flags);
3134 wake_up_process(mt);
3135 put_task_struct(mt);
3136 wait_for_completion(&req.done);
3141 task_rq_unlock(rq, &flags);
3146 DEFINE_PER_CPU(struct kernel_stat, kstat);
3148 EXPORT_PER_CPU_SYMBOL(kstat);
3151 * Return any ns on the sched_clock that have not yet been accounted in
3152 * @p in case that task is currently running.
3154 * Called with task_rq_lock() held on @rq.
3156 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3160 if (task_current(rq, p)) {
3161 update_rq_clock(rq);
3162 ns = rq->clock - p->se.exec_start;
3170 unsigned long long task_delta_exec(struct task_struct *p)
3172 unsigned long flags;
3176 rq = task_rq_lock(p, &flags);
3177 ns = do_task_delta_exec(p, rq);
3178 task_rq_unlock(rq, &flags);
3184 * Return accounted runtime for the task.
3185 * In case the task is currently running, return the runtime plus current's
3186 * pending runtime that have not been accounted yet.
3188 unsigned long long task_sched_runtime(struct task_struct *p)
3190 unsigned long flags;
3194 rq = task_rq_lock(p, &flags);
3195 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3196 task_rq_unlock(rq, &flags);
3202 * Return sum_exec_runtime for the thread group.
3203 * In case the task is currently running, return the sum plus current's
3204 * pending runtime that have not been accounted yet.
3206 * Note that the thread group might have other running tasks as well,
3207 * so the return value not includes other pending runtime that other
3208 * running tasks might have.
3210 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3212 struct task_cputime totals;
3213 unsigned long flags;
3217 rq = task_rq_lock(p, &flags);
3218 thread_group_cputime(p, &totals);
3219 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3220 task_rq_unlock(rq, &flags);
3226 * Account user cpu time to a process.
3227 * @p: the process that the cpu time gets accounted to
3228 * @cputime: the cpu time spent in user space since the last update
3229 * @cputime_scaled: cputime scaled by cpu frequency
3231 void account_user_time(struct task_struct *p, cputime_t cputime,
3232 cputime_t cputime_scaled)
3234 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3237 /* Add user time to process. */
3238 p->utime = cputime_add(p->utime, cputime);
3239 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3240 account_group_user_time(p, cputime);
3242 /* Add user time to cpustat. */
3243 tmp = cputime_to_cputime64(cputime);
3244 if (TASK_NICE(p) > 0)
3245 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3247 cpustat->user = cputime64_add(cpustat->user, tmp);
3249 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3250 /* Account for user time used */
3251 acct_update_integrals(p);
3255 * Account guest cpu time to a process.
3256 * @p: the process that the cpu time gets accounted to
3257 * @cputime: the cpu time spent in virtual machine since the last update
3258 * @cputime_scaled: cputime scaled by cpu frequency
3260 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3261 cputime_t cputime_scaled)
3264 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3266 tmp = cputime_to_cputime64(cputime);
3268 /* Add guest time to process. */
3269 p->utime = cputime_add(p->utime, cputime);
3270 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3271 account_group_user_time(p, cputime);
3272 p->gtime = cputime_add(p->gtime, cputime);
3274 /* Add guest time to cpustat. */
3275 if (TASK_NICE(p) > 0) {
3276 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3277 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3279 cpustat->user = cputime64_add(cpustat->user, tmp);
3280 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3285 * Account system cpu time to a process.
3286 * @p: the process that the cpu time gets accounted to
3287 * @hardirq_offset: the offset to subtract from hardirq_count()
3288 * @cputime: the cpu time spent in kernel space since the last update
3289 * @cputime_scaled: cputime scaled by cpu frequency
3291 void account_system_time(struct task_struct *p, int hardirq_offset,
3292 cputime_t cputime, cputime_t cputime_scaled)
3294 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3297 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3298 account_guest_time(p, cputime, cputime_scaled);
3302 /* Add system time to process. */
3303 p->stime = cputime_add(p->stime, cputime);
3304 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3305 account_group_system_time(p, cputime);
3307 /* Add system time to cpustat. */
3308 tmp = cputime_to_cputime64(cputime);
3309 if (hardirq_count() - hardirq_offset)
3310 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3311 else if (softirq_count())
3312 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3314 cpustat->system = cputime64_add(cpustat->system, tmp);
3316 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3318 /* Account for system time used */
3319 acct_update_integrals(p);
3323 * Account for involuntary wait time.
3324 * @steal: the cpu time spent in involuntary wait
3326 void account_steal_time(cputime_t cputime)
3328 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3329 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3331 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3335 * Account for idle time.
3336 * @cputime: the cpu time spent in idle wait
3338 void account_idle_time(cputime_t cputime)
3340 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3341 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3342 struct rq *rq = this_rq();
3344 if (atomic_read(&rq->nr_iowait) > 0)
3345 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3347 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3350 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3353 * Account a single tick of cpu time.
3354 * @p: the process that the cpu time gets accounted to
3355 * @user_tick: indicates if the tick is a user or a system tick
3357 void account_process_tick(struct task_struct *p, int user_tick)
3359 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3360 struct rq *rq = this_rq();
3363 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3364 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3365 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3368 account_idle_time(cputime_one_jiffy);
3372 * Account multiple ticks of steal time.
3373 * @p: the process from which the cpu time has been stolen
3374 * @ticks: number of stolen ticks
3376 void account_steal_ticks(unsigned long ticks)
3378 account_steal_time(jiffies_to_cputime(ticks));
3382 * Account multiple ticks of idle time.
3383 * @ticks: number of stolen ticks
3385 void account_idle_ticks(unsigned long ticks)
3387 account_idle_time(jiffies_to_cputime(ticks));
3393 * Use precise platform statistics if available:
3395 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3396 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3402 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3404 struct task_cputime cputime;
3406 thread_group_cputime(p, &cputime);
3408 *ut = cputime.utime;
3409 *st = cputime.stime;
3413 #ifndef nsecs_to_cputime
3414 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3417 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3419 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3422 * Use CFS's precise accounting:
3424 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3429 temp = (u64)(rtime * utime);
3430 do_div(temp, total);
3431 utime = (cputime_t)temp;
3436 * Compare with previous values, to keep monotonicity:
3438 p->prev_utime = max(p->prev_utime, utime);
3439 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3441 *ut = p->prev_utime;
3442 *st = p->prev_stime;
3446 * Must be called with siglock held.
3448 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3450 struct signal_struct *sig = p->signal;
3451 struct task_cputime cputime;
3452 cputime_t rtime, utime, total;
3454 thread_group_cputime(p, &cputime);
3456 total = cputime_add(cputime.utime, cputime.stime);
3457 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3462 temp = (u64)(rtime * cputime.utime);
3463 do_div(temp, total);
3464 utime = (cputime_t)temp;
3468 sig->prev_utime = max(sig->prev_utime, utime);
3469 sig->prev_stime = max(sig->prev_stime,
3470 cputime_sub(rtime, sig->prev_utime));
3472 *ut = sig->prev_utime;
3473 *st = sig->prev_stime;
3478 * This function gets called by the timer code, with HZ frequency.
3479 * We call it with interrupts disabled.
3481 * It also gets called by the fork code, when changing the parent's
3484 void scheduler_tick(void)
3486 int cpu = smp_processor_id();
3487 struct rq *rq = cpu_rq(cpu);
3488 struct task_struct *curr = rq->curr;
3492 raw_spin_lock(&rq->lock);
3493 update_rq_clock(rq);
3494 update_cpu_load(rq);
3495 curr->sched_class->task_tick(rq, curr, 0);
3496 raw_spin_unlock(&rq->lock);
3498 perf_event_task_tick(curr);
3501 rq->idle_at_tick = idle_cpu(cpu);
3502 trigger_load_balance(rq, cpu);
3506 notrace unsigned long get_parent_ip(unsigned long addr)
3508 if (in_lock_functions(addr)) {
3509 addr = CALLER_ADDR2;
3510 if (in_lock_functions(addr))
3511 addr = CALLER_ADDR3;
3516 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3517 defined(CONFIG_PREEMPT_TRACER))
3519 void __kprobes add_preempt_count(int val)
3521 #ifdef CONFIG_DEBUG_PREEMPT
3525 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3528 preempt_count() += val;
3529 #ifdef CONFIG_DEBUG_PREEMPT
3531 * Spinlock count overflowing soon?
3533 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3536 if (preempt_count() == val)
3537 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3539 EXPORT_SYMBOL(add_preempt_count);
3541 void __kprobes sub_preempt_count(int val)
3543 #ifdef CONFIG_DEBUG_PREEMPT
3547 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3550 * Is the spinlock portion underflowing?
3552 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3553 !(preempt_count() & PREEMPT_MASK)))
3557 if (preempt_count() == val)
3558 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3559 preempt_count() -= val;
3561 EXPORT_SYMBOL(sub_preempt_count);
3566 * Print scheduling while atomic bug:
3568 static noinline void __schedule_bug(struct task_struct *prev)
3570 struct pt_regs *regs = get_irq_regs();
3572 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3573 prev->comm, prev->pid, preempt_count());
3575 debug_show_held_locks(prev);
3577 if (irqs_disabled())
3578 print_irqtrace_events(prev);
3587 * Various schedule()-time debugging checks and statistics:
3589 static inline void schedule_debug(struct task_struct *prev)
3592 * Test if we are atomic. Since do_exit() needs to call into
3593 * schedule() atomically, we ignore that path for now.
3594 * Otherwise, whine if we are scheduling when we should not be.
3596 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3597 __schedule_bug(prev);
3599 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3601 schedstat_inc(this_rq(), sched_count);
3602 #ifdef CONFIG_SCHEDSTATS
3603 if (unlikely(prev->lock_depth >= 0)) {
3604 schedstat_inc(this_rq(), bkl_count);
3605 schedstat_inc(prev, sched_info.bkl_count);
3610 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3613 update_rq_clock(rq);
3614 rq->skip_clock_update = 0;
3615 prev->sched_class->put_prev_task(rq, prev);
3619 * Pick up the highest-prio task:
3621 static inline struct task_struct *
3622 pick_next_task(struct rq *rq)
3624 const struct sched_class *class;
3625 struct task_struct *p;
3628 * Optimization: we know that if all tasks are in
3629 * the fair class we can call that function directly:
3631 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3632 p = fair_sched_class.pick_next_task(rq);
3637 class = sched_class_highest;
3639 p = class->pick_next_task(rq);
3643 * Will never be NULL as the idle class always
3644 * returns a non-NULL p:
3646 class = class->next;
3651 * schedule() is the main scheduler function.
3653 asmlinkage void __sched schedule(void)
3655 struct task_struct *prev, *next;
3656 unsigned long *switch_count;
3662 cpu = smp_processor_id();
3666 switch_count = &prev->nivcsw;
3668 release_kernel_lock(prev);
3669 need_resched_nonpreemptible:
3671 schedule_debug(prev);
3673 if (sched_feat(HRTICK))
3676 raw_spin_lock_irq(&rq->lock);
3677 clear_tsk_need_resched(prev);
3679 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3680 if (unlikely(signal_pending_state(prev->state, prev)))
3681 prev->state = TASK_RUNNING;
3683 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3684 switch_count = &prev->nvcsw;
3687 pre_schedule(rq, prev);
3689 if (unlikely(!rq->nr_running))
3690 idle_balance(cpu, rq);
3692 put_prev_task(rq, prev);
3693 next = pick_next_task(rq);
3695 if (likely(prev != next)) {
3696 sched_info_switch(prev, next);
3697 perf_event_task_sched_out(prev, next);
3703 context_switch(rq, prev, next); /* unlocks the rq */
3705 * the context switch might have flipped the stack from under
3706 * us, hence refresh the local variables.
3708 cpu = smp_processor_id();
3711 raw_spin_unlock_irq(&rq->lock);
3715 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3717 switch_count = &prev->nivcsw;
3718 goto need_resched_nonpreemptible;
3721 preempt_enable_no_resched();
3725 EXPORT_SYMBOL(schedule);
3727 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3729 * Look out! "owner" is an entirely speculative pointer
3730 * access and not reliable.
3732 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3737 if (!sched_feat(OWNER_SPIN))
3740 #ifdef CONFIG_DEBUG_PAGEALLOC
3742 * Need to access the cpu field knowing that
3743 * DEBUG_PAGEALLOC could have unmapped it if
3744 * the mutex owner just released it and exited.
3746 if (probe_kernel_address(&owner->cpu, cpu))
3753 * Even if the access succeeded (likely case),
3754 * the cpu field may no longer be valid.
3756 if (cpu >= nr_cpumask_bits)
3760 * We need to validate that we can do a
3761 * get_cpu() and that we have the percpu area.
3763 if (!cpu_online(cpu))
3770 * Owner changed, break to re-assess state.
3772 if (lock->owner != owner)
3776 * Is that owner really running on that cpu?
3778 if (task_thread_info(rq->curr) != owner || need_resched())
3788 #ifdef CONFIG_PREEMPT
3790 * this is the entry point to schedule() from in-kernel preemption
3791 * off of preempt_enable. Kernel preemptions off return from interrupt
3792 * occur there and call schedule directly.
3794 asmlinkage void __sched preempt_schedule(void)
3796 struct thread_info *ti = current_thread_info();
3799 * If there is a non-zero preempt_count or interrupts are disabled,
3800 * we do not want to preempt the current task. Just return..
3802 if (likely(ti->preempt_count || irqs_disabled()))
3806 add_preempt_count(PREEMPT_ACTIVE);
3808 sub_preempt_count(PREEMPT_ACTIVE);
3811 * Check again in case we missed a preemption opportunity
3812 * between schedule and now.
3815 } while (need_resched());
3817 EXPORT_SYMBOL(preempt_schedule);
3820 * this is the entry point to schedule() from kernel preemption
3821 * off of irq context.
3822 * Note, that this is called and return with irqs disabled. This will
3823 * protect us against recursive calling from irq.
3825 asmlinkage void __sched preempt_schedule_irq(void)
3827 struct thread_info *ti = current_thread_info();
3829 /* Catch callers which need to be fixed */
3830 BUG_ON(ti->preempt_count || !irqs_disabled());
3833 add_preempt_count(PREEMPT_ACTIVE);
3836 local_irq_disable();
3837 sub_preempt_count(PREEMPT_ACTIVE);
3840 * Check again in case we missed a preemption opportunity
3841 * between schedule and now.
3844 } while (need_resched());
3847 #endif /* CONFIG_PREEMPT */
3849 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3852 return try_to_wake_up(curr->private, mode, wake_flags);
3854 EXPORT_SYMBOL(default_wake_function);
3857 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3858 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3859 * number) then we wake all the non-exclusive tasks and one exclusive task.
3861 * There are circumstances in which we can try to wake a task which has already
3862 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3863 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3865 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3866 int nr_exclusive, int wake_flags, void *key)
3868 wait_queue_t *curr, *next;
3870 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3871 unsigned flags = curr->flags;
3873 if (curr->func(curr, mode, wake_flags, key) &&
3874 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3880 * __wake_up - wake up threads blocked on a waitqueue.
3882 * @mode: which threads
3883 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3884 * @key: is directly passed to the wakeup function
3886 * It may be assumed that this function implies a write memory barrier before
3887 * changing the task state if and only if any tasks are woken up.
3889 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3890 int nr_exclusive, void *key)
3892 unsigned long flags;
3894 spin_lock_irqsave(&q->lock, flags);
3895 __wake_up_common(q, mode, nr_exclusive, 0, key);
3896 spin_unlock_irqrestore(&q->lock, flags);
3898 EXPORT_SYMBOL(__wake_up);
3901 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3903 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3905 __wake_up_common(q, mode, 1, 0, NULL);
3908 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3910 __wake_up_common(q, mode, 1, 0, key);
3914 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3916 * @mode: which threads
3917 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3918 * @key: opaque value to be passed to wakeup targets
3920 * The sync wakeup differs that the waker knows that it will schedule
3921 * away soon, so while the target thread will be woken up, it will not
3922 * be migrated to another CPU - ie. the two threads are 'synchronized'
3923 * with each other. This can prevent needless bouncing between CPUs.
3925 * On UP it can prevent extra preemption.
3927 * It may be assumed that this function implies a write memory barrier before
3928 * changing the task state if and only if any tasks are woken up.
3930 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3931 int nr_exclusive, void *key)
3933 unsigned long flags;
3934 int wake_flags = WF_SYNC;
3939 if (unlikely(!nr_exclusive))
3942 spin_lock_irqsave(&q->lock, flags);
3943 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3944 spin_unlock_irqrestore(&q->lock, flags);
3946 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3949 * __wake_up_sync - see __wake_up_sync_key()
3951 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3953 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3955 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3958 * complete: - signals a single thread waiting on this completion
3959 * @x: holds the state of this particular completion
3961 * This will wake up a single thread waiting on this completion. Threads will be
3962 * awakened in the same order in which they were queued.
3964 * See also complete_all(), wait_for_completion() and related routines.
3966 * It may be assumed that this function implies a write memory barrier before
3967 * changing the task state if and only if any tasks are woken up.
3969 void complete(struct completion *x)
3971 unsigned long flags;
3973 spin_lock_irqsave(&x->wait.lock, flags);
3975 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3976 spin_unlock_irqrestore(&x->wait.lock, flags);
3978 EXPORT_SYMBOL(complete);
3981 * complete_all: - signals all threads waiting on this completion
3982 * @x: holds the state of this particular completion
3984 * This will wake up all threads waiting on this particular completion event.
3986 * It may be assumed that this function implies a write memory barrier before
3987 * changing the task state if and only if any tasks are woken up.
3989 void complete_all(struct completion *x)
3991 unsigned long flags;
3993 spin_lock_irqsave(&x->wait.lock, flags);
3994 x->done += UINT_MAX/2;
3995 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3996 spin_unlock_irqrestore(&x->wait.lock, flags);
3998 EXPORT_SYMBOL(complete_all);
4000 static inline long __sched
4001 do_wait_for_common(struct completion *x, long timeout, int state)
4004 DECLARE_WAITQUEUE(wait, current);
4006 wait.flags |= WQ_FLAG_EXCLUSIVE;
4007 __add_wait_queue_tail(&x->wait, &wait);
4009 if (signal_pending_state(state, current)) {
4010 timeout = -ERESTARTSYS;
4013 __set_current_state(state);
4014 spin_unlock_irq(&x->wait.lock);
4015 timeout = schedule_timeout(timeout);
4016 spin_lock_irq(&x->wait.lock);
4017 } while (!x->done && timeout);
4018 __remove_wait_queue(&x->wait, &wait);
4023 return timeout ?: 1;
4027 wait_for_common(struct completion *x, long timeout, int state)
4031 spin_lock_irq(&x->wait.lock);
4032 timeout = do_wait_for_common(x, timeout, state);
4033 spin_unlock_irq(&x->wait.lock);
4038 * wait_for_completion: - waits for completion of a task
4039 * @x: holds the state of this particular completion
4041 * This waits to be signaled for completion of a specific task. It is NOT
4042 * interruptible and there is no timeout.
4044 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4045 * and interrupt capability. Also see complete().
4047 void __sched wait_for_completion(struct completion *x)
4049 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4051 EXPORT_SYMBOL(wait_for_completion);
4054 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4055 * @x: holds the state of this particular completion
4056 * @timeout: timeout value in jiffies
4058 * This waits for either a completion of a specific task to be signaled or for a
4059 * specified timeout to expire. The timeout is in jiffies. It is not
4062 unsigned long __sched
4063 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4065 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4067 EXPORT_SYMBOL(wait_for_completion_timeout);
4070 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4071 * @x: holds the state of this particular completion
4073 * This waits for completion of a specific task to be signaled. It is
4076 int __sched wait_for_completion_interruptible(struct completion *x)
4078 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4079 if (t == -ERESTARTSYS)
4083 EXPORT_SYMBOL(wait_for_completion_interruptible);
4086 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4087 * @x: holds the state of this particular completion
4088 * @timeout: timeout value in jiffies
4090 * This waits for either a completion of a specific task to be signaled or for a
4091 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4093 unsigned long __sched
4094 wait_for_completion_interruptible_timeout(struct completion *x,
4095 unsigned long timeout)
4097 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4099 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4102 * wait_for_completion_killable: - waits for completion of a task (killable)
4103 * @x: holds the state of this particular completion
4105 * This waits to be signaled for completion of a specific task. It can be
4106 * interrupted by a kill signal.
4108 int __sched wait_for_completion_killable(struct completion *x)
4110 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4111 if (t == -ERESTARTSYS)
4115 EXPORT_SYMBOL(wait_for_completion_killable);
4118 * try_wait_for_completion - try to decrement a completion without blocking
4119 * @x: completion structure
4121 * Returns: 0 if a decrement cannot be done without blocking
4122 * 1 if a decrement succeeded.
4124 * If a completion is being used as a counting completion,
4125 * attempt to decrement the counter without blocking. This
4126 * enables us to avoid waiting if the resource the completion
4127 * is protecting is not available.
4129 bool try_wait_for_completion(struct completion *x)
4131 unsigned long flags;
4134 spin_lock_irqsave(&x->wait.lock, flags);
4139 spin_unlock_irqrestore(&x->wait.lock, flags);
4142 EXPORT_SYMBOL(try_wait_for_completion);
4145 * completion_done - Test to see if a completion has any waiters
4146 * @x: completion structure
4148 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4149 * 1 if there are no waiters.
4152 bool completion_done(struct completion *x)
4154 unsigned long flags;
4157 spin_lock_irqsave(&x->wait.lock, flags);
4160 spin_unlock_irqrestore(&x->wait.lock, flags);
4163 EXPORT_SYMBOL(completion_done);
4166 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4168 unsigned long flags;
4171 init_waitqueue_entry(&wait, current);
4173 __set_current_state(state);
4175 spin_lock_irqsave(&q->lock, flags);
4176 __add_wait_queue(q, &wait);
4177 spin_unlock(&q->lock);
4178 timeout = schedule_timeout(timeout);
4179 spin_lock_irq(&q->lock);
4180 __remove_wait_queue(q, &wait);
4181 spin_unlock_irqrestore(&q->lock, flags);
4186 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4188 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4190 EXPORT_SYMBOL(interruptible_sleep_on);
4193 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4195 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4197 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4199 void __sched sleep_on(wait_queue_head_t *q)
4201 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4203 EXPORT_SYMBOL(sleep_on);
4205 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4207 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4209 EXPORT_SYMBOL(sleep_on_timeout);
4211 #ifdef CONFIG_RT_MUTEXES
4214 * rt_mutex_setprio - set the current priority of a task
4216 * @prio: prio value (kernel-internal form)
4218 * This function changes the 'effective' priority of a task. It does
4219 * not touch ->normal_prio like __setscheduler().
4221 * Used by the rt_mutex code to implement priority inheritance logic.
4223 void rt_mutex_setprio(struct task_struct *p, int prio)
4225 unsigned long flags;
4226 int oldprio, on_rq, running;
4228 const struct sched_class *prev_class;
4230 BUG_ON(prio < 0 || prio > MAX_PRIO);
4232 rq = task_rq_lock(p, &flags);
4235 prev_class = p->sched_class;
4236 on_rq = p->se.on_rq;
4237 running = task_current(rq, p);
4239 dequeue_task(rq, p, 0);
4241 p->sched_class->put_prev_task(rq, p);
4244 p->sched_class = &rt_sched_class;
4246 p->sched_class = &fair_sched_class;
4251 p->sched_class->set_curr_task(rq);
4253 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4255 check_class_changed(rq, p, prev_class, oldprio, running);
4257 task_rq_unlock(rq, &flags);
4262 void set_user_nice(struct task_struct *p, long nice)
4264 int old_prio, delta, on_rq;
4265 unsigned long flags;
4268 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4271 * We have to be careful, if called from sys_setpriority(),
4272 * the task might be in the middle of scheduling on another CPU.
4274 rq = task_rq_lock(p, &flags);
4276 * The RT priorities are set via sched_setscheduler(), but we still
4277 * allow the 'normal' nice value to be set - but as expected
4278 * it wont have any effect on scheduling until the task is
4279 * SCHED_FIFO/SCHED_RR:
4281 if (task_has_rt_policy(p)) {
4282 p->static_prio = NICE_TO_PRIO(nice);
4285 on_rq = p->se.on_rq;
4287 dequeue_task(rq, p, 0);
4289 p->static_prio = NICE_TO_PRIO(nice);
4292 p->prio = effective_prio(p);
4293 delta = p->prio - old_prio;
4296 enqueue_task(rq, p, 0);
4298 * If the task increased its priority or is running and
4299 * lowered its priority, then reschedule its CPU:
4301 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4302 resched_task(rq->curr);
4305 task_rq_unlock(rq, &flags);
4307 EXPORT_SYMBOL(set_user_nice);
4310 * can_nice - check if a task can reduce its nice value
4314 int can_nice(const struct task_struct *p, const int nice)
4316 /* convert nice value [19,-20] to rlimit style value [1,40] */
4317 int nice_rlim = 20 - nice;
4319 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4320 capable(CAP_SYS_NICE));
4323 #ifdef __ARCH_WANT_SYS_NICE
4326 * sys_nice - change the priority of the current process.
4327 * @increment: priority increment
4329 * sys_setpriority is a more generic, but much slower function that
4330 * does similar things.
4332 SYSCALL_DEFINE1(nice, int, increment)
4337 * Setpriority might change our priority at the same moment.
4338 * We don't have to worry. Conceptually one call occurs first
4339 * and we have a single winner.
4341 if (increment < -40)
4346 nice = TASK_NICE(current) + increment;
4352 if (increment < 0 && !can_nice(current, nice))
4355 retval = security_task_setnice(current, nice);
4359 set_user_nice(current, nice);
4366 * task_prio - return the priority value of a given task.
4367 * @p: the task in question.
4369 * This is the priority value as seen by users in /proc.
4370 * RT tasks are offset by -200. Normal tasks are centered
4371 * around 0, value goes from -16 to +15.
4373 int task_prio(const struct task_struct *p)
4375 return p->prio - MAX_RT_PRIO;
4379 * task_nice - return the nice value of a given task.
4380 * @p: the task in question.
4382 int task_nice(const struct task_struct *p)
4384 return TASK_NICE(p);
4386 EXPORT_SYMBOL(task_nice);
4389 * idle_cpu - is a given cpu idle currently?
4390 * @cpu: the processor in question.
4392 int idle_cpu(int cpu)
4394 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4398 * idle_task - return the idle task for a given cpu.
4399 * @cpu: the processor in question.
4401 struct task_struct *idle_task(int cpu)
4403 return cpu_rq(cpu)->idle;
4407 * find_process_by_pid - find a process with a matching PID value.
4408 * @pid: the pid in question.
4410 static struct task_struct *find_process_by_pid(pid_t pid)
4412 return pid ? find_task_by_vpid(pid) : current;
4415 /* Actually do priority change: must hold rq lock. */
4417 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4419 BUG_ON(p->se.on_rq);
4422 p->rt_priority = prio;
4423 p->normal_prio = normal_prio(p);
4424 /* we are holding p->pi_lock already */
4425 p->prio = rt_mutex_getprio(p);
4426 if (rt_prio(p->prio))
4427 p->sched_class = &rt_sched_class;
4429 p->sched_class = &fair_sched_class;
4434 * check the target process has a UID that matches the current process's
4436 static bool check_same_owner(struct task_struct *p)
4438 const struct cred *cred = current_cred(), *pcred;
4442 pcred = __task_cred(p);
4443 match = (cred->euid == pcred->euid ||
4444 cred->euid == pcred->uid);
4449 static int __sched_setscheduler(struct task_struct *p, int policy,
4450 struct sched_param *param, bool user)
4452 int retval, oldprio, oldpolicy = -1, on_rq, running;
4453 unsigned long flags;
4454 const struct sched_class *prev_class;
4458 /* may grab non-irq protected spin_locks */
4459 BUG_ON(in_interrupt());
4461 /* double check policy once rq lock held */
4463 reset_on_fork = p->sched_reset_on_fork;
4464 policy = oldpolicy = p->policy;
4466 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4467 policy &= ~SCHED_RESET_ON_FORK;
4469 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4470 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4471 policy != SCHED_IDLE)
4476 * Valid priorities for SCHED_FIFO and SCHED_RR are
4477 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4478 * SCHED_BATCH and SCHED_IDLE is 0.
4480 if (param->sched_priority < 0 ||
4481 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4482 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4484 if (rt_policy(policy) != (param->sched_priority != 0))
4488 * Allow unprivileged RT tasks to decrease priority:
4490 if (user && !capable(CAP_SYS_NICE)) {
4491 if (rt_policy(policy)) {
4492 unsigned long rlim_rtprio;
4494 if (!lock_task_sighand(p, &flags))
4496 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4497 unlock_task_sighand(p, &flags);
4499 /* can't set/change the rt policy */
4500 if (policy != p->policy && !rlim_rtprio)
4503 /* can't increase priority */
4504 if (param->sched_priority > p->rt_priority &&
4505 param->sched_priority > rlim_rtprio)
4509 * Like positive nice levels, dont allow tasks to
4510 * move out of SCHED_IDLE either:
4512 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4515 /* can't change other user's priorities */
4516 if (!check_same_owner(p))
4519 /* Normal users shall not reset the sched_reset_on_fork flag */
4520 if (p->sched_reset_on_fork && !reset_on_fork)
4525 #ifdef CONFIG_RT_GROUP_SCHED
4527 * Do not allow realtime tasks into groups that have no runtime
4530 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4531 task_group(p)->rt_bandwidth.rt_runtime == 0)
4535 retval = security_task_setscheduler(p, policy, param);
4541 * make sure no PI-waiters arrive (or leave) while we are
4542 * changing the priority of the task:
4544 raw_spin_lock_irqsave(&p->pi_lock, flags);
4546 * To be able to change p->policy safely, the apropriate
4547 * runqueue lock must be held.
4549 rq = __task_rq_lock(p);
4550 /* recheck policy now with rq lock held */
4551 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4552 policy = oldpolicy = -1;
4553 __task_rq_unlock(rq);
4554 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4557 on_rq = p->se.on_rq;
4558 running = task_current(rq, p);
4560 deactivate_task(rq, p, 0);
4562 p->sched_class->put_prev_task(rq, p);
4564 p->sched_reset_on_fork = reset_on_fork;
4567 prev_class = p->sched_class;
4568 __setscheduler(rq, p, policy, param->sched_priority);
4571 p->sched_class->set_curr_task(rq);
4573 activate_task(rq, p, 0);
4575 check_class_changed(rq, p, prev_class, oldprio, running);
4577 __task_rq_unlock(rq);
4578 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4580 rt_mutex_adjust_pi(p);
4586 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4587 * @p: the task in question.
4588 * @policy: new policy.
4589 * @param: structure containing the new RT priority.
4591 * NOTE that the task may be already dead.
4593 int sched_setscheduler(struct task_struct *p, int policy,
4594 struct sched_param *param)
4596 return __sched_setscheduler(p, policy, param, true);
4598 EXPORT_SYMBOL_GPL(sched_setscheduler);
4601 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4602 * @p: the task in question.
4603 * @policy: new policy.
4604 * @param: structure containing the new RT priority.
4606 * Just like sched_setscheduler, only don't bother checking if the
4607 * current context has permission. For example, this is needed in
4608 * stop_machine(): we create temporary high priority worker threads,
4609 * but our caller might not have that capability.
4611 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4612 struct sched_param *param)
4614 return __sched_setscheduler(p, policy, param, false);
4618 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4620 struct sched_param lparam;
4621 struct task_struct *p;
4624 if (!param || pid < 0)
4626 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4631 p = find_process_by_pid(pid);
4633 retval = sched_setscheduler(p, policy, &lparam);
4640 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4641 * @pid: the pid in question.
4642 * @policy: new policy.
4643 * @param: structure containing the new RT priority.
4645 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4646 struct sched_param __user *, param)
4648 /* negative values for policy are not valid */
4652 return do_sched_setscheduler(pid, policy, param);
4656 * sys_sched_setparam - set/change the RT priority of a thread
4657 * @pid: the pid in question.
4658 * @param: structure containing the new RT priority.
4660 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4662 return do_sched_setscheduler(pid, -1, param);
4666 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4667 * @pid: the pid in question.
4669 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4671 struct task_struct *p;
4679 p = find_process_by_pid(pid);
4681 retval = security_task_getscheduler(p);
4684 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4691 * sys_sched_getparam - get the RT priority of a thread
4692 * @pid: the pid in question.
4693 * @param: structure containing the RT priority.
4695 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4697 struct sched_param lp;
4698 struct task_struct *p;
4701 if (!param || pid < 0)
4705 p = find_process_by_pid(pid);
4710 retval = security_task_getscheduler(p);
4714 lp.sched_priority = p->rt_priority;
4718 * This one might sleep, we cannot do it with a spinlock held ...
4720 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4729 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4731 cpumask_var_t cpus_allowed, new_mask;
4732 struct task_struct *p;
4738 p = find_process_by_pid(pid);
4745 /* Prevent p going away */
4749 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4753 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4755 goto out_free_cpus_allowed;
4758 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4761 retval = security_task_setscheduler(p, 0, NULL);
4765 cpuset_cpus_allowed(p, cpus_allowed);
4766 cpumask_and(new_mask, in_mask, cpus_allowed);
4768 retval = set_cpus_allowed_ptr(p, new_mask);
4771 cpuset_cpus_allowed(p, cpus_allowed);
4772 if (!cpumask_subset(new_mask, cpus_allowed)) {
4774 * We must have raced with a concurrent cpuset
4775 * update. Just reset the cpus_allowed to the
4776 * cpuset's cpus_allowed
4778 cpumask_copy(new_mask, cpus_allowed);
4783 free_cpumask_var(new_mask);
4784 out_free_cpus_allowed:
4785 free_cpumask_var(cpus_allowed);
4792 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4793 struct cpumask *new_mask)
4795 if (len < cpumask_size())
4796 cpumask_clear(new_mask);
4797 else if (len > cpumask_size())
4798 len = cpumask_size();
4800 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4804 * sys_sched_setaffinity - set the cpu affinity of a process
4805 * @pid: pid of the process
4806 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4807 * @user_mask_ptr: user-space pointer to the new cpu mask
4809 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4810 unsigned long __user *, user_mask_ptr)
4812 cpumask_var_t new_mask;
4815 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4818 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4820 retval = sched_setaffinity(pid, new_mask);
4821 free_cpumask_var(new_mask);
4825 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4827 struct task_struct *p;
4828 unsigned long flags;
4836 p = find_process_by_pid(pid);
4840 retval = security_task_getscheduler(p);
4844 rq = task_rq_lock(p, &flags);
4845 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4846 task_rq_unlock(rq, &flags);
4856 * sys_sched_getaffinity - get the cpu affinity of a process
4857 * @pid: pid of the process
4858 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4859 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4861 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4862 unsigned long __user *, user_mask_ptr)
4867 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4869 if (len & (sizeof(unsigned long)-1))
4872 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4875 ret = sched_getaffinity(pid, mask);
4877 size_t retlen = min_t(size_t, len, cpumask_size());
4879 if (copy_to_user(user_mask_ptr, mask, retlen))
4884 free_cpumask_var(mask);
4890 * sys_sched_yield - yield the current processor to other threads.
4892 * This function yields the current CPU to other tasks. If there are no
4893 * other threads running on this CPU then this function will return.
4895 SYSCALL_DEFINE0(sched_yield)
4897 struct rq *rq = this_rq_lock();
4899 schedstat_inc(rq, yld_count);
4900 current->sched_class->yield_task(rq);
4903 * Since we are going to call schedule() anyway, there's
4904 * no need to preempt or enable interrupts:
4906 __release(rq->lock);
4907 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4908 do_raw_spin_unlock(&rq->lock);
4909 preempt_enable_no_resched();
4916 static inline int should_resched(void)
4918 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4921 static void __cond_resched(void)
4923 add_preempt_count(PREEMPT_ACTIVE);
4925 sub_preempt_count(PREEMPT_ACTIVE);
4928 int __sched _cond_resched(void)
4930 if (should_resched()) {
4936 EXPORT_SYMBOL(_cond_resched);
4939 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4940 * call schedule, and on return reacquire the lock.
4942 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4943 * operations here to prevent schedule() from being called twice (once via
4944 * spin_unlock(), once by hand).
4946 int __cond_resched_lock(spinlock_t *lock)
4948 int resched = should_resched();
4951 lockdep_assert_held(lock);
4953 if (spin_needbreak(lock) || resched) {
4964 EXPORT_SYMBOL(__cond_resched_lock);
4966 int __sched __cond_resched_softirq(void)
4968 BUG_ON(!in_softirq());
4970 if (should_resched()) {
4978 EXPORT_SYMBOL(__cond_resched_softirq);
4981 * yield - yield the current processor to other threads.
4983 * This is a shortcut for kernel-space yielding - it marks the
4984 * thread runnable and calls sys_sched_yield().
4986 void __sched yield(void)
4988 set_current_state(TASK_RUNNING);
4991 EXPORT_SYMBOL(yield);
4994 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4995 * that process accounting knows that this is a task in IO wait state.
4997 void __sched io_schedule(void)
4999 struct rq *rq = raw_rq();
5001 delayacct_blkio_start();
5002 atomic_inc(&rq->nr_iowait);
5003 current->in_iowait = 1;
5005 current->in_iowait = 0;
5006 atomic_dec(&rq->nr_iowait);
5007 delayacct_blkio_end();
5009 EXPORT_SYMBOL(io_schedule);
5011 long __sched io_schedule_timeout(long timeout)
5013 struct rq *rq = raw_rq();
5016 delayacct_blkio_start();
5017 atomic_inc(&rq->nr_iowait);
5018 current->in_iowait = 1;
5019 ret = schedule_timeout(timeout);
5020 current->in_iowait = 0;
5021 atomic_dec(&rq->nr_iowait);
5022 delayacct_blkio_end();
5027 * sys_sched_get_priority_max - return maximum RT priority.
5028 * @policy: scheduling class.
5030 * this syscall returns the maximum rt_priority that can be used
5031 * by a given scheduling class.
5033 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5040 ret = MAX_USER_RT_PRIO-1;
5052 * sys_sched_get_priority_min - return minimum RT priority.
5053 * @policy: scheduling class.
5055 * this syscall returns the minimum rt_priority that can be used
5056 * by a given scheduling class.
5058 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5076 * sys_sched_rr_get_interval - return the default timeslice of a process.
5077 * @pid: pid of the process.
5078 * @interval: userspace pointer to the timeslice value.
5080 * this syscall writes the default timeslice value of a given process
5081 * into the user-space timespec buffer. A value of '0' means infinity.
5083 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5084 struct timespec __user *, interval)
5086 struct task_struct *p;
5087 unsigned int time_slice;
5088 unsigned long flags;
5098 p = find_process_by_pid(pid);
5102 retval = security_task_getscheduler(p);
5106 rq = task_rq_lock(p, &flags);
5107 time_slice = p->sched_class->get_rr_interval(rq, p);
5108 task_rq_unlock(rq, &flags);
5111 jiffies_to_timespec(time_slice, &t);
5112 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5120 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5122 void sched_show_task(struct task_struct *p)
5124 unsigned long free = 0;
5127 state = p->state ? __ffs(p->state) + 1 : 0;
5128 printk(KERN_INFO "%-13.13s %c", p->comm,
5129 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5130 #if BITS_PER_LONG == 32
5131 if (state == TASK_RUNNING)
5132 printk(KERN_CONT " running ");
5134 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5136 if (state == TASK_RUNNING)
5137 printk(KERN_CONT " running task ");
5139 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5141 #ifdef CONFIG_DEBUG_STACK_USAGE
5142 free = stack_not_used(p);
5144 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5145 task_pid_nr(p), task_pid_nr(p->real_parent),
5146 (unsigned long)task_thread_info(p)->flags);
5148 show_stack(p, NULL);
5151 void show_state_filter(unsigned long state_filter)
5153 struct task_struct *g, *p;
5155 #if BITS_PER_LONG == 32
5157 " task PC stack pid father\n");
5160 " task PC stack pid father\n");
5162 read_lock(&tasklist_lock);
5163 do_each_thread(g, p) {
5165 * reset the NMI-timeout, listing all files on a slow
5166 * console might take alot of time:
5168 touch_nmi_watchdog();
5169 if (!state_filter || (p->state & state_filter))
5171 } while_each_thread(g, p);
5173 touch_all_softlockup_watchdogs();
5175 #ifdef CONFIG_SCHED_DEBUG
5176 sysrq_sched_debug_show();
5178 read_unlock(&tasklist_lock);
5180 * Only show locks if all tasks are dumped:
5183 debug_show_all_locks();
5186 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5188 idle->sched_class = &idle_sched_class;
5192 * init_idle - set up an idle thread for a given CPU
5193 * @idle: task in question
5194 * @cpu: cpu the idle task belongs to
5196 * NOTE: this function does not set the idle thread's NEED_RESCHED
5197 * flag, to make booting more robust.
5199 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5201 struct rq *rq = cpu_rq(cpu);
5202 unsigned long flags;
5204 raw_spin_lock_irqsave(&rq->lock, flags);
5207 idle->state = TASK_RUNNING;
5208 idle->se.exec_start = sched_clock();
5210 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5211 __set_task_cpu(idle, cpu);
5213 rq->curr = rq->idle = idle;
5214 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5217 raw_spin_unlock_irqrestore(&rq->lock, flags);
5219 /* Set the preempt count _outside_ the spinlocks! */
5220 #if defined(CONFIG_PREEMPT)
5221 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5223 task_thread_info(idle)->preempt_count = 0;
5226 * The idle tasks have their own, simple scheduling class:
5228 idle->sched_class = &idle_sched_class;
5229 ftrace_graph_init_task(idle);
5233 * In a system that switches off the HZ timer nohz_cpu_mask
5234 * indicates which cpus entered this state. This is used
5235 * in the rcu update to wait only for active cpus. For system
5236 * which do not switch off the HZ timer nohz_cpu_mask should
5237 * always be CPU_BITS_NONE.
5239 cpumask_var_t nohz_cpu_mask;
5242 * Increase the granularity value when there are more CPUs,
5243 * because with more CPUs the 'effective latency' as visible
5244 * to users decreases. But the relationship is not linear,
5245 * so pick a second-best guess by going with the log2 of the
5248 * This idea comes from the SD scheduler of Con Kolivas:
5250 static int get_update_sysctl_factor(void)
5252 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5253 unsigned int factor;
5255 switch (sysctl_sched_tunable_scaling) {
5256 case SCHED_TUNABLESCALING_NONE:
5259 case SCHED_TUNABLESCALING_LINEAR:
5262 case SCHED_TUNABLESCALING_LOG:
5264 factor = 1 + ilog2(cpus);
5271 static void update_sysctl(void)
5273 unsigned int factor = get_update_sysctl_factor();
5275 #define SET_SYSCTL(name) \
5276 (sysctl_##name = (factor) * normalized_sysctl_##name)
5277 SET_SYSCTL(sched_min_granularity);
5278 SET_SYSCTL(sched_latency);
5279 SET_SYSCTL(sched_wakeup_granularity);
5280 SET_SYSCTL(sched_shares_ratelimit);
5284 static inline void sched_init_granularity(void)
5291 * This is how migration works:
5293 * 1) we queue a struct migration_req structure in the source CPU's
5294 * runqueue and wake up that CPU's migration thread.
5295 * 2) we down() the locked semaphore => thread blocks.
5296 * 3) migration thread wakes up (implicitly it forces the migrated
5297 * thread off the CPU)
5298 * 4) it gets the migration request and checks whether the migrated
5299 * task is still in the wrong runqueue.
5300 * 5) if it's in the wrong runqueue then the migration thread removes
5301 * it and puts it into the right queue.
5302 * 6) migration thread up()s the semaphore.
5303 * 7) we wake up and the migration is done.
5307 * Change a given task's CPU affinity. Migrate the thread to a
5308 * proper CPU and schedule it away if the CPU it's executing on
5309 * is removed from the allowed bitmask.
5311 * NOTE: the caller must have a valid reference to the task, the
5312 * task must not exit() & deallocate itself prematurely. The
5313 * call is not atomic; no spinlocks may be held.
5315 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5317 struct migration_req req;
5318 unsigned long flags;
5323 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5324 * drop the rq->lock and still rely on ->cpus_allowed.
5327 while (task_is_waking(p))
5329 rq = task_rq_lock(p, &flags);
5330 if (task_is_waking(p)) {
5331 task_rq_unlock(rq, &flags);
5335 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5340 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5341 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5346 if (p->sched_class->set_cpus_allowed)
5347 p->sched_class->set_cpus_allowed(p, new_mask);
5349 cpumask_copy(&p->cpus_allowed, new_mask);
5350 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5353 /* Can the task run on the task's current CPU? If so, we're done */
5354 if (cpumask_test_cpu(task_cpu(p), new_mask))
5357 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5358 /* Need help from migration thread: drop lock and wait. */
5359 struct task_struct *mt = rq->migration_thread;
5361 get_task_struct(mt);
5362 task_rq_unlock(rq, &flags);
5363 wake_up_process(mt);
5364 put_task_struct(mt);
5365 wait_for_completion(&req.done);
5366 tlb_migrate_finish(p->mm);
5370 task_rq_unlock(rq, &flags);
5374 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5377 * Move (not current) task off this cpu, onto dest cpu. We're doing
5378 * this because either it can't run here any more (set_cpus_allowed()
5379 * away from this CPU, or CPU going down), or because we're
5380 * attempting to rebalance this task on exec (sched_exec).
5382 * So we race with normal scheduler movements, but that's OK, as long
5383 * as the task is no longer on this CPU.
5385 * Returns non-zero if task was successfully migrated.
5387 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5389 struct rq *rq_dest, *rq_src;
5392 if (unlikely(!cpu_active(dest_cpu)))
5395 rq_src = cpu_rq(src_cpu);
5396 rq_dest = cpu_rq(dest_cpu);
5398 double_rq_lock(rq_src, rq_dest);
5399 /* Already moved. */
5400 if (task_cpu(p) != src_cpu)
5402 /* Affinity changed (again). */
5403 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5407 * If we're not on a rq, the next wake-up will ensure we're
5411 deactivate_task(rq_src, p, 0);
5412 set_task_cpu(p, dest_cpu);
5413 activate_task(rq_dest, p, 0);
5414 check_preempt_curr(rq_dest, p, 0);
5419 double_rq_unlock(rq_src, rq_dest);
5423 #define RCU_MIGRATION_IDLE 0
5424 #define RCU_MIGRATION_NEED_QS 1
5425 #define RCU_MIGRATION_GOT_QS 2
5426 #define RCU_MIGRATION_MUST_SYNC 3
5429 * migration_thread - this is a highprio system thread that performs
5430 * thread migration by bumping thread off CPU then 'pushing' onto
5433 static int migration_thread(void *data)
5436 int cpu = (long)data;
5440 BUG_ON(rq->migration_thread != current);
5442 set_current_state(TASK_INTERRUPTIBLE);
5443 while (!kthread_should_stop()) {
5444 struct migration_req *req;
5445 struct list_head *head;
5447 raw_spin_lock_irq(&rq->lock);
5449 if (cpu_is_offline(cpu)) {
5450 raw_spin_unlock_irq(&rq->lock);
5454 if (rq->active_balance) {
5455 active_load_balance(rq, cpu);
5456 rq->active_balance = 0;
5459 head = &rq->migration_queue;
5461 if (list_empty(head)) {
5462 raw_spin_unlock_irq(&rq->lock);
5464 set_current_state(TASK_INTERRUPTIBLE);
5467 req = list_entry(head->next, struct migration_req, list);
5468 list_del_init(head->next);
5470 if (req->task != NULL) {
5471 raw_spin_unlock(&rq->lock);
5472 __migrate_task(req->task, cpu, req->dest_cpu);
5473 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5474 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5475 raw_spin_unlock(&rq->lock);
5477 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5478 raw_spin_unlock(&rq->lock);
5479 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5483 complete(&req->done);
5485 __set_current_state(TASK_RUNNING);
5490 #ifdef CONFIG_HOTPLUG_CPU
5492 * Figure out where task on dead CPU should go, use force if necessary.
5494 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5496 struct rq *rq = cpu_rq(dead_cpu);
5497 int needs_cpu, uninitialized_var(dest_cpu);
5498 unsigned long flags;
5500 local_irq_save(flags);
5502 raw_spin_lock(&rq->lock);
5503 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5505 dest_cpu = select_fallback_rq(dead_cpu, p);
5506 raw_spin_unlock(&rq->lock);
5508 * It can only fail if we race with set_cpus_allowed(),
5509 * in the racer should migrate the task anyway.
5512 __migrate_task(p, dead_cpu, dest_cpu);
5513 local_irq_restore(flags);
5517 * While a dead CPU has no uninterruptible tasks queued at this point,
5518 * it might still have a nonzero ->nr_uninterruptible counter, because
5519 * for performance reasons the counter is not stricly tracking tasks to
5520 * their home CPUs. So we just add the counter to another CPU's counter,
5521 * to keep the global sum constant after CPU-down:
5523 static void migrate_nr_uninterruptible(struct rq *rq_src)
5525 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5526 unsigned long flags;
5528 local_irq_save(flags);
5529 double_rq_lock(rq_src, rq_dest);
5530 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5531 rq_src->nr_uninterruptible = 0;
5532 double_rq_unlock(rq_src, rq_dest);
5533 local_irq_restore(flags);
5536 /* Run through task list and migrate tasks from the dead cpu. */
5537 static void migrate_live_tasks(int src_cpu)
5539 struct task_struct *p, *t;
5541 read_lock(&tasklist_lock);
5543 do_each_thread(t, p) {
5547 if (task_cpu(p) == src_cpu)
5548 move_task_off_dead_cpu(src_cpu, p);
5549 } while_each_thread(t, p);
5551 read_unlock(&tasklist_lock);
5555 * Schedules idle task to be the next runnable task on current CPU.
5556 * It does so by boosting its priority to highest possible.
5557 * Used by CPU offline code.
5559 void sched_idle_next(void)
5561 int this_cpu = smp_processor_id();
5562 struct rq *rq = cpu_rq(this_cpu);
5563 struct task_struct *p = rq->idle;
5564 unsigned long flags;
5566 /* cpu has to be offline */
5567 BUG_ON(cpu_online(this_cpu));
5570 * Strictly not necessary since rest of the CPUs are stopped by now
5571 * and interrupts disabled on the current cpu.
5573 raw_spin_lock_irqsave(&rq->lock, flags);
5575 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5577 activate_task(rq, p, 0);
5579 raw_spin_unlock_irqrestore(&rq->lock, flags);
5583 * Ensures that the idle task is using init_mm right before its cpu goes
5586 void idle_task_exit(void)
5588 struct mm_struct *mm = current->active_mm;
5590 BUG_ON(cpu_online(smp_processor_id()));
5593 switch_mm(mm, &init_mm, current);
5597 /* called under rq->lock with disabled interrupts */
5598 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5600 struct rq *rq = cpu_rq(dead_cpu);
5602 /* Must be exiting, otherwise would be on tasklist. */
5603 BUG_ON(!p->exit_state);
5605 /* Cannot have done final schedule yet: would have vanished. */
5606 BUG_ON(p->state == TASK_DEAD);
5611 * Drop lock around migration; if someone else moves it,
5612 * that's OK. No task can be added to this CPU, so iteration is
5615 raw_spin_unlock_irq(&rq->lock);
5616 move_task_off_dead_cpu(dead_cpu, p);
5617 raw_spin_lock_irq(&rq->lock);
5622 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5623 static void migrate_dead_tasks(unsigned int dead_cpu)
5625 struct rq *rq = cpu_rq(dead_cpu);
5626 struct task_struct *next;
5629 if (!rq->nr_running)
5631 next = pick_next_task(rq);
5634 next->sched_class->put_prev_task(rq, next);
5635 migrate_dead(dead_cpu, next);
5641 * remove the tasks which were accounted by rq from calc_load_tasks.
5643 static void calc_global_load_remove(struct rq *rq)
5645 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5646 rq->calc_load_active = 0;
5648 #endif /* CONFIG_HOTPLUG_CPU */
5650 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5652 static struct ctl_table sd_ctl_dir[] = {
5654 .procname = "sched_domain",
5660 static struct ctl_table sd_ctl_root[] = {
5662 .procname = "kernel",
5664 .child = sd_ctl_dir,
5669 static struct ctl_table *sd_alloc_ctl_entry(int n)
5671 struct ctl_table *entry =
5672 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5677 static void sd_free_ctl_entry(struct ctl_table **tablep)
5679 struct ctl_table *entry;
5682 * In the intermediate directories, both the child directory and
5683 * procname are dynamically allocated and could fail but the mode
5684 * will always be set. In the lowest directory the names are
5685 * static strings and all have proc handlers.
5687 for (entry = *tablep; entry->mode; entry++) {
5689 sd_free_ctl_entry(&entry->child);
5690 if (entry->proc_handler == NULL)
5691 kfree(entry->procname);
5699 set_table_entry(struct ctl_table *entry,
5700 const char *procname, void *data, int maxlen,
5701 mode_t mode, proc_handler *proc_handler)
5703 entry->procname = procname;
5705 entry->maxlen = maxlen;
5707 entry->proc_handler = proc_handler;
5710 static struct ctl_table *
5711 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5713 struct ctl_table *table = sd_alloc_ctl_entry(13);
5718 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5719 sizeof(long), 0644, proc_doulongvec_minmax);
5720 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5721 sizeof(long), 0644, proc_doulongvec_minmax);
5722 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5723 sizeof(int), 0644, proc_dointvec_minmax);
5724 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5725 sizeof(int), 0644, proc_dointvec_minmax);
5726 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5727 sizeof(int), 0644, proc_dointvec_minmax);
5728 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5729 sizeof(int), 0644, proc_dointvec_minmax);
5730 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5731 sizeof(int), 0644, proc_dointvec_minmax);
5732 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5733 sizeof(int), 0644, proc_dointvec_minmax);
5734 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5735 sizeof(int), 0644, proc_dointvec_minmax);
5736 set_table_entry(&table[9], "cache_nice_tries",
5737 &sd->cache_nice_tries,
5738 sizeof(int), 0644, proc_dointvec_minmax);
5739 set_table_entry(&table[10], "flags", &sd->flags,
5740 sizeof(int), 0644, proc_dointvec_minmax);
5741 set_table_entry(&table[11], "name", sd->name,
5742 CORENAME_MAX_SIZE, 0444, proc_dostring);
5743 /* &table[12] is terminator */
5748 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5750 struct ctl_table *entry, *table;
5751 struct sched_domain *sd;
5752 int domain_num = 0, i;
5755 for_each_domain(cpu, sd)
5757 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5762 for_each_domain(cpu, sd) {
5763 snprintf(buf, 32, "domain%d", i);
5764 entry->procname = kstrdup(buf, GFP_KERNEL);
5766 entry->child = sd_alloc_ctl_domain_table(sd);
5773 static struct ctl_table_header *sd_sysctl_header;
5774 static void register_sched_domain_sysctl(void)
5776 int i, cpu_num = num_possible_cpus();
5777 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5780 WARN_ON(sd_ctl_dir[0].child);
5781 sd_ctl_dir[0].child = entry;
5786 for_each_possible_cpu(i) {
5787 snprintf(buf, 32, "cpu%d", i);
5788 entry->procname = kstrdup(buf, GFP_KERNEL);
5790 entry->child = sd_alloc_ctl_cpu_table(i);
5794 WARN_ON(sd_sysctl_header);
5795 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5798 /* may be called multiple times per register */
5799 static void unregister_sched_domain_sysctl(void)
5801 if (sd_sysctl_header)
5802 unregister_sysctl_table(sd_sysctl_header);
5803 sd_sysctl_header = NULL;
5804 if (sd_ctl_dir[0].child)
5805 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5808 static void register_sched_domain_sysctl(void)
5811 static void unregister_sched_domain_sysctl(void)
5816 static void set_rq_online(struct rq *rq)
5819 const struct sched_class *class;
5821 cpumask_set_cpu(rq->cpu, rq->rd->online);
5824 for_each_class(class) {
5825 if (class->rq_online)
5826 class->rq_online(rq);
5831 static void set_rq_offline(struct rq *rq)
5834 const struct sched_class *class;
5836 for_each_class(class) {
5837 if (class->rq_offline)
5838 class->rq_offline(rq);
5841 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5847 * migration_call - callback that gets triggered when a CPU is added.
5848 * Here we can start up the necessary migration thread for the new CPU.
5850 static int __cpuinit
5851 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5853 struct task_struct *p;
5854 int cpu = (long)hcpu;
5855 unsigned long flags;
5860 case CPU_UP_PREPARE:
5861 case CPU_UP_PREPARE_FROZEN:
5862 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5865 kthread_bind(p, cpu);
5866 /* Must be high prio: stop_machine expects to yield to it. */
5867 rq = task_rq_lock(p, &flags);
5868 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5869 task_rq_unlock(rq, &flags);
5871 cpu_rq(cpu)->migration_thread = p;
5872 rq->calc_load_update = calc_load_update;
5876 case CPU_ONLINE_FROZEN:
5877 /* Strictly unnecessary, as first user will wake it. */
5878 wake_up_process(cpu_rq(cpu)->migration_thread);
5880 /* Update our root-domain */
5882 raw_spin_lock_irqsave(&rq->lock, flags);
5884 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5888 raw_spin_unlock_irqrestore(&rq->lock, flags);
5891 #ifdef CONFIG_HOTPLUG_CPU
5892 case CPU_UP_CANCELED:
5893 case CPU_UP_CANCELED_FROZEN:
5894 if (!cpu_rq(cpu)->migration_thread)
5896 /* Unbind it from offline cpu so it can run. Fall thru. */
5897 kthread_bind(cpu_rq(cpu)->migration_thread,
5898 cpumask_any(cpu_online_mask));
5899 kthread_stop(cpu_rq(cpu)->migration_thread);
5900 put_task_struct(cpu_rq(cpu)->migration_thread);
5901 cpu_rq(cpu)->migration_thread = NULL;
5905 case CPU_DEAD_FROZEN:
5906 migrate_live_tasks(cpu);
5908 kthread_stop(rq->migration_thread);
5909 put_task_struct(rq->migration_thread);
5910 rq->migration_thread = NULL;
5911 /* Idle task back to normal (off runqueue, low prio) */
5912 raw_spin_lock_irq(&rq->lock);
5913 deactivate_task(rq, rq->idle, 0);
5914 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5915 rq->idle->sched_class = &idle_sched_class;
5916 migrate_dead_tasks(cpu);
5917 raw_spin_unlock_irq(&rq->lock);
5918 migrate_nr_uninterruptible(rq);
5919 BUG_ON(rq->nr_running != 0);
5920 calc_global_load_remove(rq);
5922 * No need to migrate the tasks: it was best-effort if
5923 * they didn't take sched_hotcpu_mutex. Just wake up
5926 raw_spin_lock_irq(&rq->lock);
5927 while (!list_empty(&rq->migration_queue)) {
5928 struct migration_req *req;
5930 req = list_entry(rq->migration_queue.next,
5931 struct migration_req, list);
5932 list_del_init(&req->list);
5933 raw_spin_unlock_irq(&rq->lock);
5934 complete(&req->done);
5935 raw_spin_lock_irq(&rq->lock);
5937 raw_spin_unlock_irq(&rq->lock);
5941 case CPU_DYING_FROZEN:
5942 /* Update our root-domain */
5944 raw_spin_lock_irqsave(&rq->lock, flags);
5946 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5949 raw_spin_unlock_irqrestore(&rq->lock, flags);
5957 * Register at high priority so that task migration (migrate_all_tasks)
5958 * happens before everything else. This has to be lower priority than
5959 * the notifier in the perf_event subsystem, though.
5961 static struct notifier_block __cpuinitdata migration_notifier = {
5962 .notifier_call = migration_call,
5966 static int __init migration_init(void)
5968 void *cpu = (void *)(long)smp_processor_id();
5971 /* Start one for the boot CPU: */
5972 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5973 BUG_ON(err == NOTIFY_BAD);
5974 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5975 register_cpu_notifier(&migration_notifier);
5979 early_initcall(migration_init);
5984 #ifdef CONFIG_SCHED_DEBUG
5986 static __read_mostly int sched_domain_debug_enabled;
5988 static int __init sched_domain_debug_setup(char *str)
5990 sched_domain_debug_enabled = 1;
5994 early_param("sched_debug", sched_domain_debug_setup);
5996 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5997 struct cpumask *groupmask)
5999 struct sched_group *group = sd->groups;
6002 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6003 cpumask_clear(groupmask);
6005 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6007 if (!(sd->flags & SD_LOAD_BALANCE)) {
6008 printk("does not load-balance\n");
6010 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6015 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6017 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6018 printk(KERN_ERR "ERROR: domain->span does not contain "
6021 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6022 printk(KERN_ERR "ERROR: domain->groups does not contain"
6026 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6030 printk(KERN_ERR "ERROR: group is NULL\n");
6034 if (!group->cpu_power) {
6035 printk(KERN_CONT "\n");
6036 printk(KERN_ERR "ERROR: domain->cpu_power not "
6041 if (!cpumask_weight(sched_group_cpus(group))) {
6042 printk(KERN_CONT "\n");
6043 printk(KERN_ERR "ERROR: empty group\n");
6047 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6048 printk(KERN_CONT "\n");
6049 printk(KERN_ERR "ERROR: repeated CPUs\n");
6053 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6055 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6057 printk(KERN_CONT " %s", str);
6058 if (group->cpu_power != SCHED_LOAD_SCALE) {
6059 printk(KERN_CONT " (cpu_power = %d)",
6063 group = group->next;
6064 } while (group != sd->groups);
6065 printk(KERN_CONT "\n");
6067 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6068 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6071 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6072 printk(KERN_ERR "ERROR: parent span is not a superset "
6073 "of domain->span\n");
6077 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6079 cpumask_var_t groupmask;
6082 if (!sched_domain_debug_enabled)
6086 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6090 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6092 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6093 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6098 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6105 free_cpumask_var(groupmask);
6107 #else /* !CONFIG_SCHED_DEBUG */
6108 # define sched_domain_debug(sd, cpu) do { } while (0)
6109 #endif /* CONFIG_SCHED_DEBUG */
6111 static int sd_degenerate(struct sched_domain *sd)
6113 if (cpumask_weight(sched_domain_span(sd)) == 1)
6116 /* Following flags need at least 2 groups */
6117 if (sd->flags & (SD_LOAD_BALANCE |
6118 SD_BALANCE_NEWIDLE |
6122 SD_SHARE_PKG_RESOURCES)) {
6123 if (sd->groups != sd->groups->next)
6127 /* Following flags don't use groups */
6128 if (sd->flags & (SD_WAKE_AFFINE))
6135 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6137 unsigned long cflags = sd->flags, pflags = parent->flags;
6139 if (sd_degenerate(parent))
6142 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6145 /* Flags needing groups don't count if only 1 group in parent */
6146 if (parent->groups == parent->groups->next) {
6147 pflags &= ~(SD_LOAD_BALANCE |
6148 SD_BALANCE_NEWIDLE |
6152 SD_SHARE_PKG_RESOURCES);
6153 if (nr_node_ids == 1)
6154 pflags &= ~SD_SERIALIZE;
6156 if (~cflags & pflags)
6162 static void free_rootdomain(struct root_domain *rd)
6164 synchronize_sched();
6166 cpupri_cleanup(&rd->cpupri);
6168 free_cpumask_var(rd->rto_mask);
6169 free_cpumask_var(rd->online);
6170 free_cpumask_var(rd->span);
6174 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6176 struct root_domain *old_rd = NULL;
6177 unsigned long flags;
6179 raw_spin_lock_irqsave(&rq->lock, flags);
6184 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6187 cpumask_clear_cpu(rq->cpu, old_rd->span);
6190 * If we dont want to free the old_rt yet then
6191 * set old_rd to NULL to skip the freeing later
6194 if (!atomic_dec_and_test(&old_rd->refcount))
6198 atomic_inc(&rd->refcount);
6201 cpumask_set_cpu(rq->cpu, rd->span);
6202 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6205 raw_spin_unlock_irqrestore(&rq->lock, flags);
6208 free_rootdomain(old_rd);
6211 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6213 gfp_t gfp = GFP_KERNEL;
6215 memset(rd, 0, sizeof(*rd));
6220 if (!alloc_cpumask_var(&rd->span, gfp))
6222 if (!alloc_cpumask_var(&rd->online, gfp))
6224 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6227 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6232 free_cpumask_var(rd->rto_mask);
6234 free_cpumask_var(rd->online);
6236 free_cpumask_var(rd->span);
6241 static void init_defrootdomain(void)
6243 init_rootdomain(&def_root_domain, true);
6245 atomic_set(&def_root_domain.refcount, 1);
6248 static struct root_domain *alloc_rootdomain(void)
6250 struct root_domain *rd;
6252 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6256 if (init_rootdomain(rd, false) != 0) {
6265 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6266 * hold the hotplug lock.
6269 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6271 struct rq *rq = cpu_rq(cpu);
6272 struct sched_domain *tmp;
6274 /* Remove the sched domains which do not contribute to scheduling. */
6275 for (tmp = sd; tmp; ) {
6276 struct sched_domain *parent = tmp->parent;
6280 if (sd_parent_degenerate(tmp, parent)) {
6281 tmp->parent = parent->parent;
6283 parent->parent->child = tmp;
6288 if (sd && sd_degenerate(sd)) {
6294 sched_domain_debug(sd, cpu);
6296 rq_attach_root(rq, rd);
6297 rcu_assign_pointer(rq->sd, sd);
6300 /* cpus with isolated domains */
6301 static cpumask_var_t cpu_isolated_map;
6303 /* Setup the mask of cpus configured for isolated domains */
6304 static int __init isolated_cpu_setup(char *str)
6306 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6307 cpulist_parse(str, cpu_isolated_map);
6311 __setup("isolcpus=", isolated_cpu_setup);
6314 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6315 * to a function which identifies what group(along with sched group) a CPU
6316 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6317 * (due to the fact that we keep track of groups covered with a struct cpumask).
6319 * init_sched_build_groups will build a circular linked list of the groups
6320 * covered by the given span, and will set each group's ->cpumask correctly,
6321 * and ->cpu_power to 0.
6324 init_sched_build_groups(const struct cpumask *span,
6325 const struct cpumask *cpu_map,
6326 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6327 struct sched_group **sg,
6328 struct cpumask *tmpmask),
6329 struct cpumask *covered, struct cpumask *tmpmask)
6331 struct sched_group *first = NULL, *last = NULL;
6334 cpumask_clear(covered);
6336 for_each_cpu(i, span) {
6337 struct sched_group *sg;
6338 int group = group_fn(i, cpu_map, &sg, tmpmask);
6341 if (cpumask_test_cpu(i, covered))
6344 cpumask_clear(sched_group_cpus(sg));
6347 for_each_cpu(j, span) {
6348 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6351 cpumask_set_cpu(j, covered);
6352 cpumask_set_cpu(j, sched_group_cpus(sg));
6363 #define SD_NODES_PER_DOMAIN 16
6368 * find_next_best_node - find the next node to include in a sched_domain
6369 * @node: node whose sched_domain we're building
6370 * @used_nodes: nodes already in the sched_domain
6372 * Find the next node to include in a given scheduling domain. Simply
6373 * finds the closest node not already in the @used_nodes map.
6375 * Should use nodemask_t.
6377 static int find_next_best_node(int node, nodemask_t *used_nodes)
6379 int i, n, val, min_val, best_node = 0;
6383 for (i = 0; i < nr_node_ids; i++) {
6384 /* Start at @node */
6385 n = (node + i) % nr_node_ids;
6387 if (!nr_cpus_node(n))
6390 /* Skip already used nodes */
6391 if (node_isset(n, *used_nodes))
6394 /* Simple min distance search */
6395 val = node_distance(node, n);
6397 if (val < min_val) {
6403 node_set(best_node, *used_nodes);
6408 * sched_domain_node_span - get a cpumask for a node's sched_domain
6409 * @node: node whose cpumask we're constructing
6410 * @span: resulting cpumask
6412 * Given a node, construct a good cpumask for its sched_domain to span. It
6413 * should be one that prevents unnecessary balancing, but also spreads tasks
6416 static void sched_domain_node_span(int node, struct cpumask *span)
6418 nodemask_t used_nodes;
6421 cpumask_clear(span);
6422 nodes_clear(used_nodes);
6424 cpumask_or(span, span, cpumask_of_node(node));
6425 node_set(node, used_nodes);
6427 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6428 int next_node = find_next_best_node(node, &used_nodes);
6430 cpumask_or(span, span, cpumask_of_node(next_node));
6433 #endif /* CONFIG_NUMA */
6435 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6438 * The cpus mask in sched_group and sched_domain hangs off the end.
6440 * ( See the the comments in include/linux/sched.h:struct sched_group
6441 * and struct sched_domain. )
6443 struct static_sched_group {
6444 struct sched_group sg;
6445 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6448 struct static_sched_domain {
6449 struct sched_domain sd;
6450 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6456 cpumask_var_t domainspan;
6457 cpumask_var_t covered;
6458 cpumask_var_t notcovered;
6460 cpumask_var_t nodemask;
6461 cpumask_var_t this_sibling_map;
6462 cpumask_var_t this_core_map;
6463 cpumask_var_t send_covered;
6464 cpumask_var_t tmpmask;
6465 struct sched_group **sched_group_nodes;
6466 struct root_domain *rd;
6470 sa_sched_groups = 0,
6475 sa_this_sibling_map,
6477 sa_sched_group_nodes,
6487 * SMT sched-domains:
6489 #ifdef CONFIG_SCHED_SMT
6490 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6491 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6494 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6495 struct sched_group **sg, struct cpumask *unused)
6498 *sg = &per_cpu(sched_groups, cpu).sg;
6501 #endif /* CONFIG_SCHED_SMT */
6504 * multi-core sched-domains:
6506 #ifdef CONFIG_SCHED_MC
6507 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6508 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6509 #endif /* CONFIG_SCHED_MC */
6511 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6513 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6514 struct sched_group **sg, struct cpumask *mask)
6518 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6519 group = cpumask_first(mask);
6521 *sg = &per_cpu(sched_group_core, group).sg;
6524 #elif defined(CONFIG_SCHED_MC)
6526 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6527 struct sched_group **sg, struct cpumask *unused)
6530 *sg = &per_cpu(sched_group_core, cpu).sg;
6535 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6536 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6539 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6540 struct sched_group **sg, struct cpumask *mask)
6543 #ifdef CONFIG_SCHED_MC
6544 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6545 group = cpumask_first(mask);
6546 #elif defined(CONFIG_SCHED_SMT)
6547 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6548 group = cpumask_first(mask);
6553 *sg = &per_cpu(sched_group_phys, group).sg;
6559 * The init_sched_build_groups can't handle what we want to do with node
6560 * groups, so roll our own. Now each node has its own list of groups which
6561 * gets dynamically allocated.
6563 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6564 static struct sched_group ***sched_group_nodes_bycpu;
6566 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6567 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6569 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6570 struct sched_group **sg,
6571 struct cpumask *nodemask)
6575 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6576 group = cpumask_first(nodemask);
6579 *sg = &per_cpu(sched_group_allnodes, group).sg;
6583 static void init_numa_sched_groups_power(struct sched_group *group_head)
6585 struct sched_group *sg = group_head;
6591 for_each_cpu(j, sched_group_cpus(sg)) {
6592 struct sched_domain *sd;
6594 sd = &per_cpu(phys_domains, j).sd;
6595 if (j != group_first_cpu(sd->groups)) {
6597 * Only add "power" once for each
6603 sg->cpu_power += sd->groups->cpu_power;
6606 } while (sg != group_head);
6609 static int build_numa_sched_groups(struct s_data *d,
6610 const struct cpumask *cpu_map, int num)
6612 struct sched_domain *sd;
6613 struct sched_group *sg, *prev;
6616 cpumask_clear(d->covered);
6617 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6618 if (cpumask_empty(d->nodemask)) {
6619 d->sched_group_nodes[num] = NULL;
6623 sched_domain_node_span(num, d->domainspan);
6624 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6626 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6629 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6633 d->sched_group_nodes[num] = sg;
6635 for_each_cpu(j, d->nodemask) {
6636 sd = &per_cpu(node_domains, j).sd;
6641 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6643 cpumask_or(d->covered, d->covered, d->nodemask);
6646 for (j = 0; j < nr_node_ids; j++) {
6647 n = (num + j) % nr_node_ids;
6648 cpumask_complement(d->notcovered, d->covered);
6649 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6650 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6651 if (cpumask_empty(d->tmpmask))
6653 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6654 if (cpumask_empty(d->tmpmask))
6656 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6660 "Can not alloc domain group for node %d\n", j);
6664 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6665 sg->next = prev->next;
6666 cpumask_or(d->covered, d->covered, d->tmpmask);
6673 #endif /* CONFIG_NUMA */
6676 /* Free memory allocated for various sched_group structures */
6677 static void free_sched_groups(const struct cpumask *cpu_map,
6678 struct cpumask *nodemask)
6682 for_each_cpu(cpu, cpu_map) {
6683 struct sched_group **sched_group_nodes
6684 = sched_group_nodes_bycpu[cpu];
6686 if (!sched_group_nodes)
6689 for (i = 0; i < nr_node_ids; i++) {
6690 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6692 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6693 if (cpumask_empty(nodemask))
6703 if (oldsg != sched_group_nodes[i])
6706 kfree(sched_group_nodes);
6707 sched_group_nodes_bycpu[cpu] = NULL;
6710 #else /* !CONFIG_NUMA */
6711 static void free_sched_groups(const struct cpumask *cpu_map,
6712 struct cpumask *nodemask)
6715 #endif /* CONFIG_NUMA */
6718 * Initialize sched groups cpu_power.
6720 * cpu_power indicates the capacity of sched group, which is used while
6721 * distributing the load between different sched groups in a sched domain.
6722 * Typically cpu_power for all the groups in a sched domain will be same unless
6723 * there are asymmetries in the topology. If there are asymmetries, group
6724 * having more cpu_power will pickup more load compared to the group having
6727 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6729 struct sched_domain *child;
6730 struct sched_group *group;
6734 WARN_ON(!sd || !sd->groups);
6736 if (cpu != group_first_cpu(sd->groups))
6741 sd->groups->cpu_power = 0;
6744 power = SCHED_LOAD_SCALE;
6745 weight = cpumask_weight(sched_domain_span(sd));
6747 * SMT siblings share the power of a single core.
6748 * Usually multiple threads get a better yield out of
6749 * that one core than a single thread would have,
6750 * reflect that in sd->smt_gain.
6752 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6753 power *= sd->smt_gain;
6755 power >>= SCHED_LOAD_SHIFT;
6757 sd->groups->cpu_power += power;
6762 * Add cpu_power of each child group to this groups cpu_power.
6764 group = child->groups;
6766 sd->groups->cpu_power += group->cpu_power;
6767 group = group->next;
6768 } while (group != child->groups);
6772 * Initializers for schedule domains
6773 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6776 #ifdef CONFIG_SCHED_DEBUG
6777 # define SD_INIT_NAME(sd, type) sd->name = #type
6779 # define SD_INIT_NAME(sd, type) do { } while (0)
6782 #define SD_INIT(sd, type) sd_init_##type(sd)
6784 #define SD_INIT_FUNC(type) \
6785 static noinline void sd_init_##type(struct sched_domain *sd) \
6787 memset(sd, 0, sizeof(*sd)); \
6788 *sd = SD_##type##_INIT; \
6789 sd->level = SD_LV_##type; \
6790 SD_INIT_NAME(sd, type); \
6795 SD_INIT_FUNC(ALLNODES)
6798 #ifdef CONFIG_SCHED_SMT
6799 SD_INIT_FUNC(SIBLING)
6801 #ifdef CONFIG_SCHED_MC
6805 static int default_relax_domain_level = -1;
6807 static int __init setup_relax_domain_level(char *str)
6811 val = simple_strtoul(str, NULL, 0);
6812 if (val < SD_LV_MAX)
6813 default_relax_domain_level = val;
6817 __setup("relax_domain_level=", setup_relax_domain_level);
6819 static void set_domain_attribute(struct sched_domain *sd,
6820 struct sched_domain_attr *attr)
6824 if (!attr || attr->relax_domain_level < 0) {
6825 if (default_relax_domain_level < 0)
6828 request = default_relax_domain_level;
6830 request = attr->relax_domain_level;
6831 if (request < sd->level) {
6832 /* turn off idle balance on this domain */
6833 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6835 /* turn on idle balance on this domain */
6836 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6840 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6841 const struct cpumask *cpu_map)
6844 case sa_sched_groups:
6845 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6846 d->sched_group_nodes = NULL;
6848 free_rootdomain(d->rd); /* fall through */
6850 free_cpumask_var(d->tmpmask); /* fall through */
6851 case sa_send_covered:
6852 free_cpumask_var(d->send_covered); /* fall through */
6853 case sa_this_core_map:
6854 free_cpumask_var(d->this_core_map); /* fall through */
6855 case sa_this_sibling_map:
6856 free_cpumask_var(d->this_sibling_map); /* fall through */
6858 free_cpumask_var(d->nodemask); /* fall through */
6859 case sa_sched_group_nodes:
6861 kfree(d->sched_group_nodes); /* fall through */
6863 free_cpumask_var(d->notcovered); /* fall through */
6865 free_cpumask_var(d->covered); /* fall through */
6867 free_cpumask_var(d->domainspan); /* fall through */
6874 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6875 const struct cpumask *cpu_map)
6878 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6880 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6881 return sa_domainspan;
6882 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6884 /* Allocate the per-node list of sched groups */
6885 d->sched_group_nodes = kcalloc(nr_node_ids,
6886 sizeof(struct sched_group *), GFP_KERNEL);
6887 if (!d->sched_group_nodes) {
6888 printk(KERN_WARNING "Can not alloc sched group node list\n");
6889 return sa_notcovered;
6891 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6893 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6894 return sa_sched_group_nodes;
6895 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6897 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6898 return sa_this_sibling_map;
6899 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6900 return sa_this_core_map;
6901 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6902 return sa_send_covered;
6903 d->rd = alloc_rootdomain();
6905 printk(KERN_WARNING "Cannot alloc root domain\n");
6908 return sa_rootdomain;
6911 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6912 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6914 struct sched_domain *sd = NULL;
6916 struct sched_domain *parent;
6919 if (cpumask_weight(cpu_map) >
6920 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6921 sd = &per_cpu(allnodes_domains, i).sd;
6922 SD_INIT(sd, ALLNODES);
6923 set_domain_attribute(sd, attr);
6924 cpumask_copy(sched_domain_span(sd), cpu_map);
6925 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6930 sd = &per_cpu(node_domains, i).sd;
6932 set_domain_attribute(sd, attr);
6933 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6934 sd->parent = parent;
6937 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6942 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6943 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6944 struct sched_domain *parent, int i)
6946 struct sched_domain *sd;
6947 sd = &per_cpu(phys_domains, i).sd;
6949 set_domain_attribute(sd, attr);
6950 cpumask_copy(sched_domain_span(sd), d->nodemask);
6951 sd->parent = parent;
6954 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6958 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6959 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6960 struct sched_domain *parent, int i)
6962 struct sched_domain *sd = parent;
6963 #ifdef CONFIG_SCHED_MC
6964 sd = &per_cpu(core_domains, i).sd;
6966 set_domain_attribute(sd, attr);
6967 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6968 sd->parent = parent;
6970 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6975 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6976 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6977 struct sched_domain *parent, int i)
6979 struct sched_domain *sd = parent;
6980 #ifdef CONFIG_SCHED_SMT
6981 sd = &per_cpu(cpu_domains, i).sd;
6982 SD_INIT(sd, SIBLING);
6983 set_domain_attribute(sd, attr);
6984 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6985 sd->parent = parent;
6987 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6992 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6993 const struct cpumask *cpu_map, int cpu)
6996 #ifdef CONFIG_SCHED_SMT
6997 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6998 cpumask_and(d->this_sibling_map, cpu_map,
6999 topology_thread_cpumask(cpu));
7000 if (cpu == cpumask_first(d->this_sibling_map))
7001 init_sched_build_groups(d->this_sibling_map, cpu_map,
7003 d->send_covered, d->tmpmask);
7006 #ifdef CONFIG_SCHED_MC
7007 case SD_LV_MC: /* set up multi-core groups */
7008 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7009 if (cpu == cpumask_first(d->this_core_map))
7010 init_sched_build_groups(d->this_core_map, cpu_map,
7012 d->send_covered, d->tmpmask);
7015 case SD_LV_CPU: /* set up physical groups */
7016 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7017 if (!cpumask_empty(d->nodemask))
7018 init_sched_build_groups(d->nodemask, cpu_map,
7020 d->send_covered, d->tmpmask);
7023 case SD_LV_ALLNODES:
7024 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7025 d->send_covered, d->tmpmask);
7034 * Build sched domains for a given set of cpus and attach the sched domains
7035 * to the individual cpus
7037 static int __build_sched_domains(const struct cpumask *cpu_map,
7038 struct sched_domain_attr *attr)
7040 enum s_alloc alloc_state = sa_none;
7042 struct sched_domain *sd;
7048 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7049 if (alloc_state != sa_rootdomain)
7051 alloc_state = sa_sched_groups;
7054 * Set up domains for cpus specified by the cpu_map.
7056 for_each_cpu(i, cpu_map) {
7057 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7060 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7061 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7062 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7063 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7066 for_each_cpu(i, cpu_map) {
7067 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7068 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7071 /* Set up physical groups */
7072 for (i = 0; i < nr_node_ids; i++)
7073 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7076 /* Set up node groups */
7078 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7080 for (i = 0; i < nr_node_ids; i++)
7081 if (build_numa_sched_groups(&d, cpu_map, i))
7085 /* Calculate CPU power for physical packages and nodes */
7086 #ifdef CONFIG_SCHED_SMT
7087 for_each_cpu(i, cpu_map) {
7088 sd = &per_cpu(cpu_domains, i).sd;
7089 init_sched_groups_power(i, sd);
7092 #ifdef CONFIG_SCHED_MC
7093 for_each_cpu(i, cpu_map) {
7094 sd = &per_cpu(core_domains, i).sd;
7095 init_sched_groups_power(i, sd);
7099 for_each_cpu(i, cpu_map) {
7100 sd = &per_cpu(phys_domains, i).sd;
7101 init_sched_groups_power(i, sd);
7105 for (i = 0; i < nr_node_ids; i++)
7106 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7108 if (d.sd_allnodes) {
7109 struct sched_group *sg;
7111 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7113 init_numa_sched_groups_power(sg);
7117 /* Attach the domains */
7118 for_each_cpu(i, cpu_map) {
7119 #ifdef CONFIG_SCHED_SMT
7120 sd = &per_cpu(cpu_domains, i).sd;
7121 #elif defined(CONFIG_SCHED_MC)
7122 sd = &per_cpu(core_domains, i).sd;
7124 sd = &per_cpu(phys_domains, i).sd;
7126 cpu_attach_domain(sd, d.rd, i);
7129 d.sched_group_nodes = NULL; /* don't free this we still need it */
7130 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7134 __free_domain_allocs(&d, alloc_state, cpu_map);
7138 static int build_sched_domains(const struct cpumask *cpu_map)
7140 return __build_sched_domains(cpu_map, NULL);
7143 static cpumask_var_t *doms_cur; /* current sched domains */
7144 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7145 static struct sched_domain_attr *dattr_cur;
7146 /* attribues of custom domains in 'doms_cur' */
7149 * Special case: If a kmalloc of a doms_cur partition (array of
7150 * cpumask) fails, then fallback to a single sched domain,
7151 * as determined by the single cpumask fallback_doms.
7153 static cpumask_var_t fallback_doms;
7156 * arch_update_cpu_topology lets virtualized architectures update the
7157 * cpu core maps. It is supposed to return 1 if the topology changed
7158 * or 0 if it stayed the same.
7160 int __attribute__((weak)) arch_update_cpu_topology(void)
7165 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7168 cpumask_var_t *doms;
7170 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7173 for (i = 0; i < ndoms; i++) {
7174 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7175 free_sched_domains(doms, i);
7182 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7185 for (i = 0; i < ndoms; i++)
7186 free_cpumask_var(doms[i]);
7191 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7192 * For now this just excludes isolated cpus, but could be used to
7193 * exclude other special cases in the future.
7195 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7199 arch_update_cpu_topology();
7201 doms_cur = alloc_sched_domains(ndoms_cur);
7203 doms_cur = &fallback_doms;
7204 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7206 err = build_sched_domains(doms_cur[0]);
7207 register_sched_domain_sysctl();
7212 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7213 struct cpumask *tmpmask)
7215 free_sched_groups(cpu_map, tmpmask);
7219 * Detach sched domains from a group of cpus specified in cpu_map
7220 * These cpus will now be attached to the NULL domain
7222 static void detach_destroy_domains(const struct cpumask *cpu_map)
7224 /* Save because hotplug lock held. */
7225 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7228 for_each_cpu(i, cpu_map)
7229 cpu_attach_domain(NULL, &def_root_domain, i);
7230 synchronize_sched();
7231 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7234 /* handle null as "default" */
7235 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7236 struct sched_domain_attr *new, int idx_new)
7238 struct sched_domain_attr tmp;
7245 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7246 new ? (new + idx_new) : &tmp,
7247 sizeof(struct sched_domain_attr));
7251 * Partition sched domains as specified by the 'ndoms_new'
7252 * cpumasks in the array doms_new[] of cpumasks. This compares
7253 * doms_new[] to the current sched domain partitioning, doms_cur[].
7254 * It destroys each deleted domain and builds each new domain.
7256 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7257 * The masks don't intersect (don't overlap.) We should setup one
7258 * sched domain for each mask. CPUs not in any of the cpumasks will
7259 * not be load balanced. If the same cpumask appears both in the
7260 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7263 * The passed in 'doms_new' should be allocated using
7264 * alloc_sched_domains. This routine takes ownership of it and will
7265 * free_sched_domains it when done with it. If the caller failed the
7266 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7267 * and partition_sched_domains() will fallback to the single partition
7268 * 'fallback_doms', it also forces the domains to be rebuilt.
7270 * If doms_new == NULL it will be replaced with cpu_online_mask.
7271 * ndoms_new == 0 is a special case for destroying existing domains,
7272 * and it will not create the default domain.
7274 * Call with hotplug lock held
7276 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7277 struct sched_domain_attr *dattr_new)
7282 mutex_lock(&sched_domains_mutex);
7284 /* always unregister in case we don't destroy any domains */
7285 unregister_sched_domain_sysctl();
7287 /* Let architecture update cpu core mappings. */
7288 new_topology = arch_update_cpu_topology();
7290 n = doms_new ? ndoms_new : 0;
7292 /* Destroy deleted domains */
7293 for (i = 0; i < ndoms_cur; i++) {
7294 for (j = 0; j < n && !new_topology; j++) {
7295 if (cpumask_equal(doms_cur[i], doms_new[j])
7296 && dattrs_equal(dattr_cur, i, dattr_new, j))
7299 /* no match - a current sched domain not in new doms_new[] */
7300 detach_destroy_domains(doms_cur[i]);
7305 if (doms_new == NULL) {
7307 doms_new = &fallback_doms;
7308 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7309 WARN_ON_ONCE(dattr_new);
7312 /* Build new domains */
7313 for (i = 0; i < ndoms_new; i++) {
7314 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7315 if (cpumask_equal(doms_new[i], doms_cur[j])
7316 && dattrs_equal(dattr_new, i, dattr_cur, j))
7319 /* no match - add a new doms_new */
7320 __build_sched_domains(doms_new[i],
7321 dattr_new ? dattr_new + i : NULL);
7326 /* Remember the new sched domains */
7327 if (doms_cur != &fallback_doms)
7328 free_sched_domains(doms_cur, ndoms_cur);
7329 kfree(dattr_cur); /* kfree(NULL) is safe */
7330 doms_cur = doms_new;
7331 dattr_cur = dattr_new;
7332 ndoms_cur = ndoms_new;
7334 register_sched_domain_sysctl();
7336 mutex_unlock(&sched_domains_mutex);
7339 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7340 static void arch_reinit_sched_domains(void)
7344 /* Destroy domains first to force the rebuild */
7345 partition_sched_domains(0, NULL, NULL);
7347 rebuild_sched_domains();
7351 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7353 unsigned int level = 0;
7355 if (sscanf(buf, "%u", &level) != 1)
7359 * level is always be positive so don't check for
7360 * level < POWERSAVINGS_BALANCE_NONE which is 0
7361 * What happens on 0 or 1 byte write,
7362 * need to check for count as well?
7365 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7369 sched_smt_power_savings = level;
7371 sched_mc_power_savings = level;
7373 arch_reinit_sched_domains();
7378 #ifdef CONFIG_SCHED_MC
7379 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7380 struct sysdev_class_attribute *attr,
7383 return sprintf(page, "%u\n", sched_mc_power_savings);
7385 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7386 struct sysdev_class_attribute *attr,
7387 const char *buf, size_t count)
7389 return sched_power_savings_store(buf, count, 0);
7391 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7392 sched_mc_power_savings_show,
7393 sched_mc_power_savings_store);
7396 #ifdef CONFIG_SCHED_SMT
7397 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7398 struct sysdev_class_attribute *attr,
7401 return sprintf(page, "%u\n", sched_smt_power_savings);
7403 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7404 struct sysdev_class_attribute *attr,
7405 const char *buf, size_t count)
7407 return sched_power_savings_store(buf, count, 1);
7409 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7410 sched_smt_power_savings_show,
7411 sched_smt_power_savings_store);
7414 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7418 #ifdef CONFIG_SCHED_SMT
7420 err = sysfs_create_file(&cls->kset.kobj,
7421 &attr_sched_smt_power_savings.attr);
7423 #ifdef CONFIG_SCHED_MC
7424 if (!err && mc_capable())
7425 err = sysfs_create_file(&cls->kset.kobj,
7426 &attr_sched_mc_power_savings.attr);
7430 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7432 #ifndef CONFIG_CPUSETS
7434 * Add online and remove offline CPUs from the scheduler domains.
7435 * When cpusets are enabled they take over this function.
7437 static int update_sched_domains(struct notifier_block *nfb,
7438 unsigned long action, void *hcpu)
7442 case CPU_ONLINE_FROZEN:
7443 case CPU_DOWN_PREPARE:
7444 case CPU_DOWN_PREPARE_FROZEN:
7445 case CPU_DOWN_FAILED:
7446 case CPU_DOWN_FAILED_FROZEN:
7447 partition_sched_domains(1, NULL, NULL);
7456 static int update_runtime(struct notifier_block *nfb,
7457 unsigned long action, void *hcpu)
7459 int cpu = (int)(long)hcpu;
7462 case CPU_DOWN_PREPARE:
7463 case CPU_DOWN_PREPARE_FROZEN:
7464 disable_runtime(cpu_rq(cpu));
7467 case CPU_DOWN_FAILED:
7468 case CPU_DOWN_FAILED_FROZEN:
7470 case CPU_ONLINE_FROZEN:
7471 enable_runtime(cpu_rq(cpu));
7479 void __init sched_init_smp(void)
7481 cpumask_var_t non_isolated_cpus;
7483 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7484 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7486 #if defined(CONFIG_NUMA)
7487 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7489 BUG_ON(sched_group_nodes_bycpu == NULL);
7492 mutex_lock(&sched_domains_mutex);
7493 arch_init_sched_domains(cpu_active_mask);
7494 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7495 if (cpumask_empty(non_isolated_cpus))
7496 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7497 mutex_unlock(&sched_domains_mutex);
7500 #ifndef CONFIG_CPUSETS
7501 /* XXX: Theoretical race here - CPU may be hotplugged now */
7502 hotcpu_notifier(update_sched_domains, 0);
7505 /* RT runtime code needs to handle some hotplug events */
7506 hotcpu_notifier(update_runtime, 0);
7510 /* Move init over to a non-isolated CPU */
7511 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7513 sched_init_granularity();
7514 free_cpumask_var(non_isolated_cpus);
7516 init_sched_rt_class();
7519 void __init sched_init_smp(void)
7521 sched_init_granularity();
7523 #endif /* CONFIG_SMP */
7525 const_debug unsigned int sysctl_timer_migration = 1;
7527 int in_sched_functions(unsigned long addr)
7529 return in_lock_functions(addr) ||
7530 (addr >= (unsigned long)__sched_text_start
7531 && addr < (unsigned long)__sched_text_end);
7534 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7536 cfs_rq->tasks_timeline = RB_ROOT;
7537 INIT_LIST_HEAD(&cfs_rq->tasks);
7538 #ifdef CONFIG_FAIR_GROUP_SCHED
7541 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7544 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7546 struct rt_prio_array *array;
7549 array = &rt_rq->active;
7550 for (i = 0; i < MAX_RT_PRIO; i++) {
7551 INIT_LIST_HEAD(array->queue + i);
7552 __clear_bit(i, array->bitmap);
7554 /* delimiter for bitsearch: */
7555 __set_bit(MAX_RT_PRIO, array->bitmap);
7557 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7558 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7560 rt_rq->highest_prio.next = MAX_RT_PRIO;
7564 rt_rq->rt_nr_migratory = 0;
7565 rt_rq->overloaded = 0;
7566 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7570 rt_rq->rt_throttled = 0;
7571 rt_rq->rt_runtime = 0;
7572 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7574 #ifdef CONFIG_RT_GROUP_SCHED
7575 rt_rq->rt_nr_boosted = 0;
7580 #ifdef CONFIG_FAIR_GROUP_SCHED
7581 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7582 struct sched_entity *se, int cpu, int add,
7583 struct sched_entity *parent)
7585 struct rq *rq = cpu_rq(cpu);
7586 tg->cfs_rq[cpu] = cfs_rq;
7587 init_cfs_rq(cfs_rq, rq);
7590 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7593 /* se could be NULL for init_task_group */
7598 se->cfs_rq = &rq->cfs;
7600 se->cfs_rq = parent->my_q;
7603 se->load.weight = tg->shares;
7604 se->load.inv_weight = 0;
7605 se->parent = parent;
7609 #ifdef CONFIG_RT_GROUP_SCHED
7610 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7611 struct sched_rt_entity *rt_se, int cpu, int add,
7612 struct sched_rt_entity *parent)
7614 struct rq *rq = cpu_rq(cpu);
7616 tg->rt_rq[cpu] = rt_rq;
7617 init_rt_rq(rt_rq, rq);
7619 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7621 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7623 tg->rt_se[cpu] = rt_se;
7628 rt_se->rt_rq = &rq->rt;
7630 rt_se->rt_rq = parent->my_q;
7632 rt_se->my_q = rt_rq;
7633 rt_se->parent = parent;
7634 INIT_LIST_HEAD(&rt_se->run_list);
7638 void __init sched_init(void)
7641 unsigned long alloc_size = 0, ptr;
7643 #ifdef CONFIG_FAIR_GROUP_SCHED
7644 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7646 #ifdef CONFIG_RT_GROUP_SCHED
7647 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7649 #ifdef CONFIG_CPUMASK_OFFSTACK
7650 alloc_size += num_possible_cpus() * cpumask_size();
7653 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7655 #ifdef CONFIG_FAIR_GROUP_SCHED
7656 init_task_group.se = (struct sched_entity **)ptr;
7657 ptr += nr_cpu_ids * sizeof(void **);
7659 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7660 ptr += nr_cpu_ids * sizeof(void **);
7662 #endif /* CONFIG_FAIR_GROUP_SCHED */
7663 #ifdef CONFIG_RT_GROUP_SCHED
7664 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7665 ptr += nr_cpu_ids * sizeof(void **);
7667 init_task_group.rt_rq = (struct rt_rq **)ptr;
7668 ptr += nr_cpu_ids * sizeof(void **);
7670 #endif /* CONFIG_RT_GROUP_SCHED */
7671 #ifdef CONFIG_CPUMASK_OFFSTACK
7672 for_each_possible_cpu(i) {
7673 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7674 ptr += cpumask_size();
7676 #endif /* CONFIG_CPUMASK_OFFSTACK */
7680 init_defrootdomain();
7683 init_rt_bandwidth(&def_rt_bandwidth,
7684 global_rt_period(), global_rt_runtime());
7686 #ifdef CONFIG_RT_GROUP_SCHED
7687 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7688 global_rt_period(), global_rt_runtime());
7689 #endif /* CONFIG_RT_GROUP_SCHED */
7691 #ifdef CONFIG_CGROUP_SCHED
7692 list_add(&init_task_group.list, &task_groups);
7693 INIT_LIST_HEAD(&init_task_group.children);
7695 #endif /* CONFIG_CGROUP_SCHED */
7697 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7698 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7699 __alignof__(unsigned long));
7701 for_each_possible_cpu(i) {
7705 raw_spin_lock_init(&rq->lock);
7707 rq->calc_load_active = 0;
7708 rq->calc_load_update = jiffies + LOAD_FREQ;
7709 init_cfs_rq(&rq->cfs, rq);
7710 init_rt_rq(&rq->rt, rq);
7711 #ifdef CONFIG_FAIR_GROUP_SCHED
7712 init_task_group.shares = init_task_group_load;
7713 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7714 #ifdef CONFIG_CGROUP_SCHED
7716 * How much cpu bandwidth does init_task_group get?
7718 * In case of task-groups formed thr' the cgroup filesystem, it
7719 * gets 100% of the cpu resources in the system. This overall
7720 * system cpu resource is divided among the tasks of
7721 * init_task_group and its child task-groups in a fair manner,
7722 * based on each entity's (task or task-group's) weight
7723 * (se->load.weight).
7725 * In other words, if init_task_group has 10 tasks of weight
7726 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7727 * then A0's share of the cpu resource is:
7729 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7731 * We achieve this by letting init_task_group's tasks sit
7732 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7734 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7736 #endif /* CONFIG_FAIR_GROUP_SCHED */
7738 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7739 #ifdef CONFIG_RT_GROUP_SCHED
7740 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7741 #ifdef CONFIG_CGROUP_SCHED
7742 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7746 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7747 rq->cpu_load[j] = 0;
7751 rq->post_schedule = 0;
7752 rq->active_balance = 0;
7753 rq->next_balance = jiffies;
7757 rq->migration_thread = NULL;
7759 rq->avg_idle = 2*sysctl_sched_migration_cost;
7760 INIT_LIST_HEAD(&rq->migration_queue);
7761 rq_attach_root(rq, &def_root_domain);
7764 atomic_set(&rq->nr_iowait, 0);
7767 set_load_weight(&init_task);
7769 #ifdef CONFIG_PREEMPT_NOTIFIERS
7770 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7774 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7777 #ifdef CONFIG_RT_MUTEXES
7778 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7782 * The boot idle thread does lazy MMU switching as well:
7784 atomic_inc(&init_mm.mm_count);
7785 enter_lazy_tlb(&init_mm, current);
7788 * Make us the idle thread. Technically, schedule() should not be
7789 * called from this thread, however somewhere below it might be,
7790 * but because we are the idle thread, we just pick up running again
7791 * when this runqueue becomes "idle".
7793 init_idle(current, smp_processor_id());
7795 calc_load_update = jiffies + LOAD_FREQ;
7798 * During early bootup we pretend to be a normal task:
7800 current->sched_class = &fair_sched_class;
7802 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7803 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7806 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7807 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7809 /* May be allocated at isolcpus cmdline parse time */
7810 if (cpu_isolated_map == NULL)
7811 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7816 scheduler_running = 1;
7819 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7820 static inline int preempt_count_equals(int preempt_offset)
7822 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7824 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7827 void __might_sleep(const char *file, int line, int preempt_offset)
7830 static unsigned long prev_jiffy; /* ratelimiting */
7832 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7833 system_state != SYSTEM_RUNNING || oops_in_progress)
7835 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7837 prev_jiffy = jiffies;
7840 "BUG: sleeping function called from invalid context at %s:%d\n",
7843 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7844 in_atomic(), irqs_disabled(),
7845 current->pid, current->comm);
7847 debug_show_held_locks(current);
7848 if (irqs_disabled())
7849 print_irqtrace_events(current);
7853 EXPORT_SYMBOL(__might_sleep);
7856 #ifdef CONFIG_MAGIC_SYSRQ
7857 static void normalize_task(struct rq *rq, struct task_struct *p)
7861 on_rq = p->se.on_rq;
7863 deactivate_task(rq, p, 0);
7864 __setscheduler(rq, p, SCHED_NORMAL, 0);
7866 activate_task(rq, p, 0);
7867 resched_task(rq->curr);
7871 void normalize_rt_tasks(void)
7873 struct task_struct *g, *p;
7874 unsigned long flags;
7877 read_lock_irqsave(&tasklist_lock, flags);
7878 do_each_thread(g, p) {
7880 * Only normalize user tasks:
7885 p->se.exec_start = 0;
7886 #ifdef CONFIG_SCHEDSTATS
7887 p->se.statistics.wait_start = 0;
7888 p->se.statistics.sleep_start = 0;
7889 p->se.statistics.block_start = 0;
7894 * Renice negative nice level userspace
7897 if (TASK_NICE(p) < 0 && p->mm)
7898 set_user_nice(p, 0);
7902 raw_spin_lock(&p->pi_lock);
7903 rq = __task_rq_lock(p);
7905 normalize_task(rq, p);
7907 __task_rq_unlock(rq);
7908 raw_spin_unlock(&p->pi_lock);
7909 } while_each_thread(g, p);
7911 read_unlock_irqrestore(&tasklist_lock, flags);
7914 #endif /* CONFIG_MAGIC_SYSRQ */
7918 * These functions are only useful for the IA64 MCA handling.
7920 * They can only be called when the whole system has been
7921 * stopped - every CPU needs to be quiescent, and no scheduling
7922 * activity can take place. Using them for anything else would
7923 * be a serious bug, and as a result, they aren't even visible
7924 * under any other configuration.
7928 * curr_task - return the current task for a given cpu.
7929 * @cpu: the processor in question.
7931 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7933 struct task_struct *curr_task(int cpu)
7935 return cpu_curr(cpu);
7939 * set_curr_task - set the current task for a given cpu.
7940 * @cpu: the processor in question.
7941 * @p: the task pointer to set.
7943 * Description: This function must only be used when non-maskable interrupts
7944 * are serviced on a separate stack. It allows the architecture to switch the
7945 * notion of the current task on a cpu in a non-blocking manner. This function
7946 * must be called with all CPU's synchronized, and interrupts disabled, the
7947 * and caller must save the original value of the current task (see
7948 * curr_task() above) and restore that value before reenabling interrupts and
7949 * re-starting the system.
7951 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7953 void set_curr_task(int cpu, struct task_struct *p)
7960 #ifdef CONFIG_FAIR_GROUP_SCHED
7961 static void free_fair_sched_group(struct task_group *tg)
7965 for_each_possible_cpu(i) {
7967 kfree(tg->cfs_rq[i]);
7977 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7979 struct cfs_rq *cfs_rq;
7980 struct sched_entity *se;
7984 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7987 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7991 tg->shares = NICE_0_LOAD;
7993 for_each_possible_cpu(i) {
7996 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7997 GFP_KERNEL, cpu_to_node(i));
8001 se = kzalloc_node(sizeof(struct sched_entity),
8002 GFP_KERNEL, cpu_to_node(i));
8006 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8017 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8019 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8020 &cpu_rq(cpu)->leaf_cfs_rq_list);
8023 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8025 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8027 #else /* !CONFG_FAIR_GROUP_SCHED */
8028 static inline void free_fair_sched_group(struct task_group *tg)
8033 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8038 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8042 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8045 #endif /* CONFIG_FAIR_GROUP_SCHED */
8047 #ifdef CONFIG_RT_GROUP_SCHED
8048 static void free_rt_sched_group(struct task_group *tg)
8052 destroy_rt_bandwidth(&tg->rt_bandwidth);
8054 for_each_possible_cpu(i) {
8056 kfree(tg->rt_rq[i]);
8058 kfree(tg->rt_se[i]);
8066 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8068 struct rt_rq *rt_rq;
8069 struct sched_rt_entity *rt_se;
8073 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8076 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8080 init_rt_bandwidth(&tg->rt_bandwidth,
8081 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8083 for_each_possible_cpu(i) {
8086 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8087 GFP_KERNEL, cpu_to_node(i));
8091 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8092 GFP_KERNEL, cpu_to_node(i));
8096 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8107 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8109 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8110 &cpu_rq(cpu)->leaf_rt_rq_list);
8113 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8115 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8117 #else /* !CONFIG_RT_GROUP_SCHED */
8118 static inline void free_rt_sched_group(struct task_group *tg)
8123 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8128 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8132 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8135 #endif /* CONFIG_RT_GROUP_SCHED */
8137 #ifdef CONFIG_CGROUP_SCHED
8138 static void free_sched_group(struct task_group *tg)
8140 free_fair_sched_group(tg);
8141 free_rt_sched_group(tg);
8145 /* allocate runqueue etc for a new task group */
8146 struct task_group *sched_create_group(struct task_group *parent)
8148 struct task_group *tg;
8149 unsigned long flags;
8152 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8154 return ERR_PTR(-ENOMEM);
8156 if (!alloc_fair_sched_group(tg, parent))
8159 if (!alloc_rt_sched_group(tg, parent))
8162 spin_lock_irqsave(&task_group_lock, flags);
8163 for_each_possible_cpu(i) {
8164 register_fair_sched_group(tg, i);
8165 register_rt_sched_group(tg, i);
8167 list_add_rcu(&tg->list, &task_groups);
8169 WARN_ON(!parent); /* root should already exist */
8171 tg->parent = parent;
8172 INIT_LIST_HEAD(&tg->children);
8173 list_add_rcu(&tg->siblings, &parent->children);
8174 spin_unlock_irqrestore(&task_group_lock, flags);
8179 free_sched_group(tg);
8180 return ERR_PTR(-ENOMEM);
8183 /* rcu callback to free various structures associated with a task group */
8184 static void free_sched_group_rcu(struct rcu_head *rhp)
8186 /* now it should be safe to free those cfs_rqs */
8187 free_sched_group(container_of(rhp, struct task_group, rcu));
8190 /* Destroy runqueue etc associated with a task group */
8191 void sched_destroy_group(struct task_group *tg)
8193 unsigned long flags;
8196 spin_lock_irqsave(&task_group_lock, flags);
8197 for_each_possible_cpu(i) {
8198 unregister_fair_sched_group(tg, i);
8199 unregister_rt_sched_group(tg, i);
8201 list_del_rcu(&tg->list);
8202 list_del_rcu(&tg->siblings);
8203 spin_unlock_irqrestore(&task_group_lock, flags);
8205 /* wait for possible concurrent references to cfs_rqs complete */
8206 call_rcu(&tg->rcu, free_sched_group_rcu);
8209 /* change task's runqueue when it moves between groups.
8210 * The caller of this function should have put the task in its new group
8211 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8212 * reflect its new group.
8214 void sched_move_task(struct task_struct *tsk)
8217 unsigned long flags;
8220 rq = task_rq_lock(tsk, &flags);
8222 running = task_current(rq, tsk);
8223 on_rq = tsk->se.on_rq;
8226 dequeue_task(rq, tsk, 0);
8227 if (unlikely(running))
8228 tsk->sched_class->put_prev_task(rq, tsk);
8230 set_task_rq(tsk, task_cpu(tsk));
8232 #ifdef CONFIG_FAIR_GROUP_SCHED
8233 if (tsk->sched_class->moved_group)
8234 tsk->sched_class->moved_group(tsk, on_rq);
8237 if (unlikely(running))
8238 tsk->sched_class->set_curr_task(rq);
8240 enqueue_task(rq, tsk, 0);
8242 task_rq_unlock(rq, &flags);
8244 #endif /* CONFIG_CGROUP_SCHED */
8246 #ifdef CONFIG_FAIR_GROUP_SCHED
8247 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8249 struct cfs_rq *cfs_rq = se->cfs_rq;
8254 dequeue_entity(cfs_rq, se, 0);
8256 se->load.weight = shares;
8257 se->load.inv_weight = 0;
8260 enqueue_entity(cfs_rq, se, 0);
8263 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8265 struct cfs_rq *cfs_rq = se->cfs_rq;
8266 struct rq *rq = cfs_rq->rq;
8267 unsigned long flags;
8269 raw_spin_lock_irqsave(&rq->lock, flags);
8270 __set_se_shares(se, shares);
8271 raw_spin_unlock_irqrestore(&rq->lock, flags);
8274 static DEFINE_MUTEX(shares_mutex);
8276 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8279 unsigned long flags;
8282 * We can't change the weight of the root cgroup.
8287 if (shares < MIN_SHARES)
8288 shares = MIN_SHARES;
8289 else if (shares > MAX_SHARES)
8290 shares = MAX_SHARES;
8292 mutex_lock(&shares_mutex);
8293 if (tg->shares == shares)
8296 spin_lock_irqsave(&task_group_lock, flags);
8297 for_each_possible_cpu(i)
8298 unregister_fair_sched_group(tg, i);
8299 list_del_rcu(&tg->siblings);
8300 spin_unlock_irqrestore(&task_group_lock, flags);
8302 /* wait for any ongoing reference to this group to finish */
8303 synchronize_sched();
8306 * Now we are free to modify the group's share on each cpu
8307 * w/o tripping rebalance_share or load_balance_fair.
8309 tg->shares = shares;
8310 for_each_possible_cpu(i) {
8314 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8315 set_se_shares(tg->se[i], shares);
8319 * Enable load balance activity on this group, by inserting it back on
8320 * each cpu's rq->leaf_cfs_rq_list.
8322 spin_lock_irqsave(&task_group_lock, flags);
8323 for_each_possible_cpu(i)
8324 register_fair_sched_group(tg, i);
8325 list_add_rcu(&tg->siblings, &tg->parent->children);
8326 spin_unlock_irqrestore(&task_group_lock, flags);
8328 mutex_unlock(&shares_mutex);
8332 unsigned long sched_group_shares(struct task_group *tg)
8338 #ifdef CONFIG_RT_GROUP_SCHED
8340 * Ensure that the real time constraints are schedulable.
8342 static DEFINE_MUTEX(rt_constraints_mutex);
8344 static unsigned long to_ratio(u64 period, u64 runtime)
8346 if (runtime == RUNTIME_INF)
8349 return div64_u64(runtime << 20, period);
8352 /* Must be called with tasklist_lock held */
8353 static inline int tg_has_rt_tasks(struct task_group *tg)
8355 struct task_struct *g, *p;
8357 do_each_thread(g, p) {
8358 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8360 } while_each_thread(g, p);
8365 struct rt_schedulable_data {
8366 struct task_group *tg;
8371 static int tg_schedulable(struct task_group *tg, void *data)
8373 struct rt_schedulable_data *d = data;
8374 struct task_group *child;
8375 unsigned long total, sum = 0;
8376 u64 period, runtime;
8378 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8379 runtime = tg->rt_bandwidth.rt_runtime;
8382 period = d->rt_period;
8383 runtime = d->rt_runtime;
8387 * Cannot have more runtime than the period.
8389 if (runtime > period && runtime != RUNTIME_INF)
8393 * Ensure we don't starve existing RT tasks.
8395 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8398 total = to_ratio(period, runtime);
8401 * Nobody can have more than the global setting allows.
8403 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8407 * The sum of our children's runtime should not exceed our own.
8409 list_for_each_entry_rcu(child, &tg->children, siblings) {
8410 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8411 runtime = child->rt_bandwidth.rt_runtime;
8413 if (child == d->tg) {
8414 period = d->rt_period;
8415 runtime = d->rt_runtime;
8418 sum += to_ratio(period, runtime);
8427 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8429 struct rt_schedulable_data data = {
8431 .rt_period = period,
8432 .rt_runtime = runtime,
8435 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8438 static int tg_set_bandwidth(struct task_group *tg,
8439 u64 rt_period, u64 rt_runtime)
8443 mutex_lock(&rt_constraints_mutex);
8444 read_lock(&tasklist_lock);
8445 err = __rt_schedulable(tg, rt_period, rt_runtime);
8449 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8450 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8451 tg->rt_bandwidth.rt_runtime = rt_runtime;
8453 for_each_possible_cpu(i) {
8454 struct rt_rq *rt_rq = tg->rt_rq[i];
8456 raw_spin_lock(&rt_rq->rt_runtime_lock);
8457 rt_rq->rt_runtime = rt_runtime;
8458 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8460 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8462 read_unlock(&tasklist_lock);
8463 mutex_unlock(&rt_constraints_mutex);
8468 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8470 u64 rt_runtime, rt_period;
8472 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8473 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8474 if (rt_runtime_us < 0)
8475 rt_runtime = RUNTIME_INF;
8477 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8480 long sched_group_rt_runtime(struct task_group *tg)
8484 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8487 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8488 do_div(rt_runtime_us, NSEC_PER_USEC);
8489 return rt_runtime_us;
8492 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8494 u64 rt_runtime, rt_period;
8496 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8497 rt_runtime = tg->rt_bandwidth.rt_runtime;
8502 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8505 long sched_group_rt_period(struct task_group *tg)
8509 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8510 do_div(rt_period_us, NSEC_PER_USEC);
8511 return rt_period_us;
8514 static int sched_rt_global_constraints(void)
8516 u64 runtime, period;
8519 if (sysctl_sched_rt_period <= 0)
8522 runtime = global_rt_runtime();
8523 period = global_rt_period();
8526 * Sanity check on the sysctl variables.
8528 if (runtime > period && runtime != RUNTIME_INF)
8531 mutex_lock(&rt_constraints_mutex);
8532 read_lock(&tasklist_lock);
8533 ret = __rt_schedulable(NULL, 0, 0);
8534 read_unlock(&tasklist_lock);
8535 mutex_unlock(&rt_constraints_mutex);
8540 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8542 /* Don't accept realtime tasks when there is no way for them to run */
8543 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8549 #else /* !CONFIG_RT_GROUP_SCHED */
8550 static int sched_rt_global_constraints(void)
8552 unsigned long flags;
8555 if (sysctl_sched_rt_period <= 0)
8559 * There's always some RT tasks in the root group
8560 * -- migration, kstopmachine etc..
8562 if (sysctl_sched_rt_runtime == 0)
8565 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8566 for_each_possible_cpu(i) {
8567 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8569 raw_spin_lock(&rt_rq->rt_runtime_lock);
8570 rt_rq->rt_runtime = global_rt_runtime();
8571 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8573 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8577 #endif /* CONFIG_RT_GROUP_SCHED */
8579 int sched_rt_handler(struct ctl_table *table, int write,
8580 void __user *buffer, size_t *lenp,
8584 int old_period, old_runtime;
8585 static DEFINE_MUTEX(mutex);
8588 old_period = sysctl_sched_rt_period;
8589 old_runtime = sysctl_sched_rt_runtime;
8591 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8593 if (!ret && write) {
8594 ret = sched_rt_global_constraints();
8596 sysctl_sched_rt_period = old_period;
8597 sysctl_sched_rt_runtime = old_runtime;
8599 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8600 def_rt_bandwidth.rt_period =
8601 ns_to_ktime(global_rt_period());
8604 mutex_unlock(&mutex);
8609 #ifdef CONFIG_CGROUP_SCHED
8611 /* return corresponding task_group object of a cgroup */
8612 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8614 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8615 struct task_group, css);
8618 static struct cgroup_subsys_state *
8619 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8621 struct task_group *tg, *parent;
8623 if (!cgrp->parent) {
8624 /* This is early initialization for the top cgroup */
8625 return &init_task_group.css;
8628 parent = cgroup_tg(cgrp->parent);
8629 tg = sched_create_group(parent);
8631 return ERR_PTR(-ENOMEM);
8637 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8639 struct task_group *tg = cgroup_tg(cgrp);
8641 sched_destroy_group(tg);
8645 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8647 #ifdef CONFIG_RT_GROUP_SCHED
8648 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8651 /* We don't support RT-tasks being in separate groups */
8652 if (tsk->sched_class != &fair_sched_class)
8659 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8660 struct task_struct *tsk, bool threadgroup)
8662 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8666 struct task_struct *c;
8668 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8669 retval = cpu_cgroup_can_attach_task(cgrp, c);
8681 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8682 struct cgroup *old_cont, struct task_struct *tsk,
8685 sched_move_task(tsk);
8687 struct task_struct *c;
8689 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8696 #ifdef CONFIG_FAIR_GROUP_SCHED
8697 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8700 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8703 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8705 struct task_group *tg = cgroup_tg(cgrp);
8707 return (u64) tg->shares;
8709 #endif /* CONFIG_FAIR_GROUP_SCHED */
8711 #ifdef CONFIG_RT_GROUP_SCHED
8712 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8715 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8718 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8720 return sched_group_rt_runtime(cgroup_tg(cgrp));
8723 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8726 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8729 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8731 return sched_group_rt_period(cgroup_tg(cgrp));
8733 #endif /* CONFIG_RT_GROUP_SCHED */
8735 static struct cftype cpu_files[] = {
8736 #ifdef CONFIG_FAIR_GROUP_SCHED
8739 .read_u64 = cpu_shares_read_u64,
8740 .write_u64 = cpu_shares_write_u64,
8743 #ifdef CONFIG_RT_GROUP_SCHED
8745 .name = "rt_runtime_us",
8746 .read_s64 = cpu_rt_runtime_read,
8747 .write_s64 = cpu_rt_runtime_write,
8750 .name = "rt_period_us",
8751 .read_u64 = cpu_rt_period_read_uint,
8752 .write_u64 = cpu_rt_period_write_uint,
8757 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8759 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8762 struct cgroup_subsys cpu_cgroup_subsys = {
8764 .create = cpu_cgroup_create,
8765 .destroy = cpu_cgroup_destroy,
8766 .can_attach = cpu_cgroup_can_attach,
8767 .attach = cpu_cgroup_attach,
8768 .populate = cpu_cgroup_populate,
8769 .subsys_id = cpu_cgroup_subsys_id,
8773 #endif /* CONFIG_CGROUP_SCHED */
8775 #ifdef CONFIG_CGROUP_CPUACCT
8778 * CPU accounting code for task groups.
8780 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8781 * (balbir@in.ibm.com).
8784 /* track cpu usage of a group of tasks and its child groups */
8786 struct cgroup_subsys_state css;
8787 /* cpuusage holds pointer to a u64-type object on every cpu */
8788 u64 __percpu *cpuusage;
8789 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8790 struct cpuacct *parent;
8793 struct cgroup_subsys cpuacct_subsys;
8795 /* return cpu accounting group corresponding to this container */
8796 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8798 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8799 struct cpuacct, css);
8802 /* return cpu accounting group to which this task belongs */
8803 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8805 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8806 struct cpuacct, css);
8809 /* create a new cpu accounting group */
8810 static struct cgroup_subsys_state *cpuacct_create(
8811 struct cgroup_subsys *ss, struct cgroup *cgrp)
8813 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8819 ca->cpuusage = alloc_percpu(u64);
8823 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8824 if (percpu_counter_init(&ca->cpustat[i], 0))
8825 goto out_free_counters;
8828 ca->parent = cgroup_ca(cgrp->parent);
8834 percpu_counter_destroy(&ca->cpustat[i]);
8835 free_percpu(ca->cpuusage);
8839 return ERR_PTR(-ENOMEM);
8842 /* destroy an existing cpu accounting group */
8844 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8846 struct cpuacct *ca = cgroup_ca(cgrp);
8849 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8850 percpu_counter_destroy(&ca->cpustat[i]);
8851 free_percpu(ca->cpuusage);
8855 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8857 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8860 #ifndef CONFIG_64BIT
8862 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8864 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8866 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8874 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8876 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8878 #ifndef CONFIG_64BIT
8880 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8882 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8884 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8890 /* return total cpu usage (in nanoseconds) of a group */
8891 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8893 struct cpuacct *ca = cgroup_ca(cgrp);
8894 u64 totalcpuusage = 0;
8897 for_each_present_cpu(i)
8898 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8900 return totalcpuusage;
8903 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8906 struct cpuacct *ca = cgroup_ca(cgrp);
8915 for_each_present_cpu(i)
8916 cpuacct_cpuusage_write(ca, i, 0);
8922 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8925 struct cpuacct *ca = cgroup_ca(cgroup);
8929 for_each_present_cpu(i) {
8930 percpu = cpuacct_cpuusage_read(ca, i);
8931 seq_printf(m, "%llu ", (unsigned long long) percpu);
8933 seq_printf(m, "\n");
8937 static const char *cpuacct_stat_desc[] = {
8938 [CPUACCT_STAT_USER] = "user",
8939 [CPUACCT_STAT_SYSTEM] = "system",
8942 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8943 struct cgroup_map_cb *cb)
8945 struct cpuacct *ca = cgroup_ca(cgrp);
8948 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8949 s64 val = percpu_counter_read(&ca->cpustat[i]);
8950 val = cputime64_to_clock_t(val);
8951 cb->fill(cb, cpuacct_stat_desc[i], val);
8956 static struct cftype files[] = {
8959 .read_u64 = cpuusage_read,
8960 .write_u64 = cpuusage_write,
8963 .name = "usage_percpu",
8964 .read_seq_string = cpuacct_percpu_seq_read,
8968 .read_map = cpuacct_stats_show,
8972 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8974 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8978 * charge this task's execution time to its accounting group.
8980 * called with rq->lock held.
8982 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8987 if (unlikely(!cpuacct_subsys.active))
8990 cpu = task_cpu(tsk);
8996 for (; ca; ca = ca->parent) {
8997 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8998 *cpuusage += cputime;
9005 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9006 * in cputime_t units. As a result, cpuacct_update_stats calls
9007 * percpu_counter_add with values large enough to always overflow the
9008 * per cpu batch limit causing bad SMP scalability.
9010 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9011 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9012 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9015 #define CPUACCT_BATCH \
9016 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9018 #define CPUACCT_BATCH 0
9022 * Charge the system/user time to the task's accounting group.
9024 static void cpuacct_update_stats(struct task_struct *tsk,
9025 enum cpuacct_stat_index idx, cputime_t val)
9028 int batch = CPUACCT_BATCH;
9030 if (unlikely(!cpuacct_subsys.active))
9037 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9043 struct cgroup_subsys cpuacct_subsys = {
9045 .create = cpuacct_create,
9046 .destroy = cpuacct_destroy,
9047 .populate = cpuacct_populate,
9048 .subsys_id = cpuacct_subsys_id,
9050 #endif /* CONFIG_CGROUP_CPUACCT */
9054 int rcu_expedited_torture_stats(char *page)
9058 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9060 void synchronize_sched_expedited(void)
9063 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9065 #else /* #ifndef CONFIG_SMP */
9067 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9068 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9070 #define RCU_EXPEDITED_STATE_POST -2
9071 #define RCU_EXPEDITED_STATE_IDLE -1
9073 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9075 int rcu_expedited_torture_stats(char *page)
9080 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9081 for_each_online_cpu(cpu) {
9082 cnt += sprintf(&page[cnt], " %d:%d",
9083 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9085 cnt += sprintf(&page[cnt], "\n");
9088 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9090 static long synchronize_sched_expedited_count;
9093 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9094 * approach to force grace period to end quickly. This consumes
9095 * significant time on all CPUs, and is thus not recommended for
9096 * any sort of common-case code.
9098 * Note that it is illegal to call this function while holding any
9099 * lock that is acquired by a CPU-hotplug notifier. Failing to
9100 * observe this restriction will result in deadlock.
9102 void synchronize_sched_expedited(void)
9105 unsigned long flags;
9106 bool need_full_sync = 0;
9108 struct migration_req *req;
9112 smp_mb(); /* ensure prior mod happens before capturing snap. */
9113 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9115 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9117 if (trycount++ < 10)
9118 udelay(trycount * num_online_cpus());
9120 synchronize_sched();
9123 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9124 smp_mb(); /* ensure test happens before caller kfree */
9129 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9130 for_each_online_cpu(cpu) {
9132 req = &per_cpu(rcu_migration_req, cpu);
9133 init_completion(&req->done);
9135 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9136 raw_spin_lock_irqsave(&rq->lock, flags);
9137 list_add(&req->list, &rq->migration_queue);
9138 raw_spin_unlock_irqrestore(&rq->lock, flags);
9139 wake_up_process(rq->migration_thread);
9141 for_each_online_cpu(cpu) {
9142 rcu_expedited_state = cpu;
9143 req = &per_cpu(rcu_migration_req, cpu);
9145 wait_for_completion(&req->done);
9146 raw_spin_lock_irqsave(&rq->lock, flags);
9147 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9149 req->dest_cpu = RCU_MIGRATION_IDLE;
9150 raw_spin_unlock_irqrestore(&rq->lock, flags);
9152 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9153 synchronize_sched_expedited_count++;
9154 mutex_unlock(&rcu_sched_expedited_mutex);
9157 synchronize_sched();
9159 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9161 #endif /* #else #ifndef CONFIG_SMP */