4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 raw_spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 raw_spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 raw_spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 raw_spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_CGROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 struct cgroup_subsys_state css;
248 #ifdef CONFIG_FAIR_GROUP_SCHED
249 /* schedulable entities of this group on each cpu */
250 struct sched_entity **se;
251 /* runqueue "owned" by this group on each cpu */
252 struct cfs_rq **cfs_rq;
253 unsigned long shares;
256 #ifdef CONFIG_RT_GROUP_SCHED
257 struct sched_rt_entity **rt_se;
258 struct rt_rq **rt_rq;
260 struct rt_bandwidth rt_bandwidth;
264 struct list_head list;
266 struct task_group *parent;
267 struct list_head siblings;
268 struct list_head children;
271 #define root_task_group init_task_group
273 /* task_group_lock serializes add/remove of task groups and also changes to
274 * a task group's cpu shares.
276 static DEFINE_SPINLOCK(task_group_lock);
278 #ifdef CONFIG_FAIR_GROUP_SCHED
281 static int root_task_group_empty(void)
283 return list_empty(&root_task_group.children);
287 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
290 * A weight of 0 or 1 can cause arithmetics problems.
291 * A weight of a cfs_rq is the sum of weights of which entities
292 * are queued on this cfs_rq, so a weight of a entity should not be
293 * too large, so as the shares value of a task group.
294 * (The default weight is 1024 - so there's no practical
295 * limitation from this.)
298 #define MAX_SHARES (1UL << 18)
300 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
303 /* Default task group.
304 * Every task in system belong to this group at bootup.
306 struct task_group init_task_group;
308 /* return group to which a task belongs */
309 static inline struct task_group *task_group(struct task_struct *p)
311 struct task_group *tg;
313 #ifdef CONFIG_CGROUP_SCHED
314 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
315 struct task_group, css);
317 tg = &init_task_group;
322 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
323 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
327 p->se.parent = task_group(p)->se[cpu];
330 #ifdef CONFIG_RT_GROUP_SCHED
331 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
332 p->rt.parent = task_group(p)->rt_se[cpu];
338 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
339 static inline struct task_group *task_group(struct task_struct *p)
344 #endif /* CONFIG_CGROUP_SCHED */
346 /* CFS-related fields in a runqueue */
348 struct load_weight load;
349 unsigned long nr_running;
354 struct rb_root tasks_timeline;
355 struct rb_node *rb_leftmost;
357 struct list_head tasks;
358 struct list_head *balance_iterator;
361 * 'curr' points to currently running entity on this cfs_rq.
362 * It is set to NULL otherwise (i.e when none are currently running).
364 struct sched_entity *curr, *next, *last;
366 unsigned int nr_spread_over;
368 #ifdef CONFIG_FAIR_GROUP_SCHED
369 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
372 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
373 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
374 * (like users, containers etc.)
376 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
377 * list is used during load balance.
379 struct list_head leaf_cfs_rq_list;
380 struct task_group *tg; /* group that "owns" this runqueue */
384 * the part of load.weight contributed by tasks
386 unsigned long task_weight;
389 * h_load = weight * f(tg)
391 * Where f(tg) is the recursive weight fraction assigned to
394 unsigned long h_load;
397 * this cpu's part of tg->shares
399 unsigned long shares;
402 * load.weight at the time we set shares
404 unsigned long rq_weight;
409 /* Real-Time classes' related field in a runqueue: */
411 struct rt_prio_array active;
412 unsigned long rt_nr_running;
413 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
415 int curr; /* highest queued rt task prio */
417 int next; /* next highest */
422 unsigned long rt_nr_migratory;
423 unsigned long rt_nr_total;
425 struct plist_head pushable_tasks;
430 /* Nests inside the rq lock: */
431 raw_spinlock_t rt_runtime_lock;
433 #ifdef CONFIG_RT_GROUP_SCHED
434 unsigned long rt_nr_boosted;
437 struct list_head leaf_rt_rq_list;
438 struct task_group *tg;
445 * We add the notion of a root-domain which will be used to define per-domain
446 * variables. Each exclusive cpuset essentially defines an island domain by
447 * fully partitioning the member cpus from any other cpuset. Whenever a new
448 * exclusive cpuset is created, we also create and attach a new root-domain
455 cpumask_var_t online;
458 * The "RT overload" flag: it gets set if a CPU has more than
459 * one runnable RT task.
461 cpumask_var_t rto_mask;
464 struct cpupri cpupri;
469 * By default the system creates a single root-domain with all cpus as
470 * members (mimicking the global state we have today).
472 static struct root_domain def_root_domain;
477 * This is the main, per-CPU runqueue data structure.
479 * Locking rule: those places that want to lock multiple runqueues
480 * (such as the load balancing or the thread migration code), lock
481 * acquire operations must be ordered by ascending &runqueue.
488 * nr_running and cpu_load should be in the same cacheline because
489 * remote CPUs use both these fields when doing load calculation.
491 unsigned long nr_running;
492 #define CPU_LOAD_IDX_MAX 5
493 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
496 unsigned char in_nohz_recently;
498 /* capture load from *all* tasks on this cpu: */
499 struct load_weight load;
500 unsigned long nr_load_updates;
506 #ifdef CONFIG_FAIR_GROUP_SCHED
507 /* list of leaf cfs_rq on this cpu: */
508 struct list_head leaf_cfs_rq_list;
510 #ifdef CONFIG_RT_GROUP_SCHED
511 struct list_head leaf_rt_rq_list;
515 * This is part of a global counter where only the total sum
516 * over all CPUs matters. A task can increase this counter on
517 * one CPU and if it got migrated afterwards it may decrease
518 * it on another CPU. Always updated under the runqueue lock:
520 unsigned long nr_uninterruptible;
522 struct task_struct *curr, *idle;
523 unsigned long next_balance;
524 struct mm_struct *prev_mm;
531 struct root_domain *rd;
532 struct sched_domain *sd;
534 unsigned char idle_at_tick;
535 /* For active balancing */
539 /* cpu of this runqueue: */
543 unsigned long avg_load_per_task;
545 struct task_struct *migration_thread;
546 struct list_head migration_queue;
554 /* calc_load related fields */
555 unsigned long calc_load_update;
556 long calc_load_active;
558 #ifdef CONFIG_SCHED_HRTICK
560 int hrtick_csd_pending;
561 struct call_single_data hrtick_csd;
563 struct hrtimer hrtick_timer;
566 #ifdef CONFIG_SCHEDSTATS
568 struct sched_info rq_sched_info;
569 unsigned long long rq_cpu_time;
570 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
572 /* sys_sched_yield() stats */
573 unsigned int yld_count;
575 /* schedule() stats */
576 unsigned int sched_switch;
577 unsigned int sched_count;
578 unsigned int sched_goidle;
580 /* try_to_wake_up() stats */
581 unsigned int ttwu_count;
582 unsigned int ttwu_local;
585 unsigned int bkl_count;
589 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
592 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
594 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
597 static inline int cpu_of(struct rq *rq)
606 #define rcu_dereference_check_sched_domain(p) \
607 rcu_dereference_check((p), \
608 rcu_read_lock_sched_held() || \
609 lockdep_is_held(&sched_domains_mutex))
612 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
613 * See detach_destroy_domains: synchronize_sched for details.
615 * The domain tree of any CPU may only be accessed from within
616 * preempt-disabled sections.
618 #define for_each_domain(cpu, __sd) \
619 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
621 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
622 #define this_rq() (&__get_cpu_var(runqueues))
623 #define task_rq(p) cpu_rq(task_cpu(p))
624 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
625 #define raw_rq() (&__raw_get_cpu_var(runqueues))
627 inline void update_rq_clock(struct rq *rq)
629 rq->clock = sched_clock_cpu(cpu_of(rq));
633 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
635 #ifdef CONFIG_SCHED_DEBUG
636 # define const_debug __read_mostly
638 # define const_debug static const
643 * @cpu: the processor in question.
645 * Returns true if the current cpu runqueue is locked.
646 * This interface allows printk to be called with the runqueue lock
647 * held and know whether or not it is OK to wake up the klogd.
649 int runqueue_is_locked(int cpu)
651 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
655 * Debugging: various feature bits
658 #define SCHED_FEAT(name, enabled) \
659 __SCHED_FEAT_##name ,
662 #include "sched_features.h"
667 #define SCHED_FEAT(name, enabled) \
668 (1UL << __SCHED_FEAT_##name) * enabled |
670 const_debug unsigned int sysctl_sched_features =
671 #include "sched_features.h"
676 #ifdef CONFIG_SCHED_DEBUG
677 #define SCHED_FEAT(name, enabled) \
680 static __read_mostly char *sched_feat_names[] = {
681 #include "sched_features.h"
687 static int sched_feat_show(struct seq_file *m, void *v)
691 for (i = 0; sched_feat_names[i]; i++) {
692 if (!(sysctl_sched_features & (1UL << i)))
694 seq_printf(m, "%s ", sched_feat_names[i]);
702 sched_feat_write(struct file *filp, const char __user *ubuf,
703 size_t cnt, loff_t *ppos)
713 if (copy_from_user(&buf, ubuf, cnt))
718 if (strncmp(buf, "NO_", 3) == 0) {
723 for (i = 0; sched_feat_names[i]; i++) {
724 int len = strlen(sched_feat_names[i]);
726 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
728 sysctl_sched_features &= ~(1UL << i);
730 sysctl_sched_features |= (1UL << i);
735 if (!sched_feat_names[i])
743 static int sched_feat_open(struct inode *inode, struct file *filp)
745 return single_open(filp, sched_feat_show, NULL);
748 static const struct file_operations sched_feat_fops = {
749 .open = sched_feat_open,
750 .write = sched_feat_write,
753 .release = single_release,
756 static __init int sched_init_debug(void)
758 debugfs_create_file("sched_features", 0644, NULL, NULL,
763 late_initcall(sched_init_debug);
767 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
770 * Number of tasks to iterate in a single balance run.
771 * Limited because this is done with IRQs disabled.
773 const_debug unsigned int sysctl_sched_nr_migrate = 32;
776 * ratelimit for updating the group shares.
779 unsigned int sysctl_sched_shares_ratelimit = 250000;
780 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
783 * Inject some fuzzyness into changing the per-cpu group shares
784 * this avoids remote rq-locks at the expense of fairness.
787 unsigned int sysctl_sched_shares_thresh = 4;
790 * period over which we average the RT time consumption, measured
795 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
798 * period over which we measure -rt task cpu usage in us.
801 unsigned int sysctl_sched_rt_period = 1000000;
803 static __read_mostly int scheduler_running;
806 * part of the period that we allow rt tasks to run in us.
809 int sysctl_sched_rt_runtime = 950000;
811 static inline u64 global_rt_period(void)
813 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
816 static inline u64 global_rt_runtime(void)
818 if (sysctl_sched_rt_runtime < 0)
821 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
824 #ifndef prepare_arch_switch
825 # define prepare_arch_switch(next) do { } while (0)
827 #ifndef finish_arch_switch
828 # define finish_arch_switch(prev) do { } while (0)
831 static inline int task_current(struct rq *rq, struct task_struct *p)
833 return rq->curr == p;
836 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
837 static inline int task_running(struct rq *rq, struct task_struct *p)
839 return task_current(rq, p);
842 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
846 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
848 #ifdef CONFIG_DEBUG_SPINLOCK
849 /* this is a valid case when another task releases the spinlock */
850 rq->lock.owner = current;
853 * If we are tracking spinlock dependencies then we have to
854 * fix up the runqueue lock - which gets 'carried over' from
857 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
859 raw_spin_unlock_irq(&rq->lock);
862 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
863 static inline int task_running(struct rq *rq, struct task_struct *p)
868 return task_current(rq, p);
872 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
876 * We can optimise this out completely for !SMP, because the
877 * SMP rebalancing from interrupt is the only thing that cares
882 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
883 raw_spin_unlock_irq(&rq->lock);
885 raw_spin_unlock(&rq->lock);
889 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
893 * After ->oncpu is cleared, the task can be moved to a different CPU.
894 * We must ensure this doesn't happen until the switch is completely
900 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
904 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
907 * Check whether the task is waking, we use this to synchronize against
908 * ttwu() so that task_cpu() reports a stable number.
910 * We need to make an exception for PF_STARTING tasks because the fork
911 * path might require task_rq_lock() to work, eg. it can call
912 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
914 static inline int task_is_waking(struct task_struct *p)
916 return unlikely((p->state == TASK_WAKING) && !(p->flags & PF_STARTING));
920 * __task_rq_lock - lock the runqueue a given task resides on.
921 * Must be called interrupts disabled.
923 static inline struct rq *__task_rq_lock(struct task_struct *p)
929 while (task_is_waking(p))
932 raw_spin_lock(&rq->lock);
933 if (likely(rq == task_rq(p) && !task_is_waking(p)))
935 raw_spin_unlock(&rq->lock);
940 * task_rq_lock - lock the runqueue a given task resides on and disable
941 * interrupts. Note the ordering: we can safely lookup the task_rq without
942 * explicitly disabling preemption.
944 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
950 while (task_is_waking(p))
952 local_irq_save(*flags);
954 raw_spin_lock(&rq->lock);
955 if (likely(rq == task_rq(p) && !task_is_waking(p)))
957 raw_spin_unlock_irqrestore(&rq->lock, *flags);
961 void task_rq_unlock_wait(struct task_struct *p)
963 struct rq *rq = task_rq(p);
965 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
966 raw_spin_unlock_wait(&rq->lock);
969 static void __task_rq_unlock(struct rq *rq)
972 raw_spin_unlock(&rq->lock);
975 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
978 raw_spin_unlock_irqrestore(&rq->lock, *flags);
982 * this_rq_lock - lock this runqueue and disable interrupts.
984 static struct rq *this_rq_lock(void)
991 raw_spin_lock(&rq->lock);
996 #ifdef CONFIG_SCHED_HRTICK
998 * Use HR-timers to deliver accurate preemption points.
1000 * Its all a bit involved since we cannot program an hrt while holding the
1001 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1004 * When we get rescheduled we reprogram the hrtick_timer outside of the
1010 * - enabled by features
1011 * - hrtimer is actually high res
1013 static inline int hrtick_enabled(struct rq *rq)
1015 if (!sched_feat(HRTICK))
1017 if (!cpu_active(cpu_of(rq)))
1019 return hrtimer_is_hres_active(&rq->hrtick_timer);
1022 static void hrtick_clear(struct rq *rq)
1024 if (hrtimer_active(&rq->hrtick_timer))
1025 hrtimer_cancel(&rq->hrtick_timer);
1029 * High-resolution timer tick.
1030 * Runs from hardirq context with interrupts disabled.
1032 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1034 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1036 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1038 raw_spin_lock(&rq->lock);
1039 update_rq_clock(rq);
1040 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1041 raw_spin_unlock(&rq->lock);
1043 return HRTIMER_NORESTART;
1048 * called from hardirq (IPI) context
1050 static void __hrtick_start(void *arg)
1052 struct rq *rq = arg;
1054 raw_spin_lock(&rq->lock);
1055 hrtimer_restart(&rq->hrtick_timer);
1056 rq->hrtick_csd_pending = 0;
1057 raw_spin_unlock(&rq->lock);
1061 * Called to set the hrtick timer state.
1063 * called with rq->lock held and irqs disabled
1065 static void hrtick_start(struct rq *rq, u64 delay)
1067 struct hrtimer *timer = &rq->hrtick_timer;
1068 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1070 hrtimer_set_expires(timer, time);
1072 if (rq == this_rq()) {
1073 hrtimer_restart(timer);
1074 } else if (!rq->hrtick_csd_pending) {
1075 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1076 rq->hrtick_csd_pending = 1;
1081 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1083 int cpu = (int)(long)hcpu;
1086 case CPU_UP_CANCELED:
1087 case CPU_UP_CANCELED_FROZEN:
1088 case CPU_DOWN_PREPARE:
1089 case CPU_DOWN_PREPARE_FROZEN:
1091 case CPU_DEAD_FROZEN:
1092 hrtick_clear(cpu_rq(cpu));
1099 static __init void init_hrtick(void)
1101 hotcpu_notifier(hotplug_hrtick, 0);
1105 * Called to set the hrtick timer state.
1107 * called with rq->lock held and irqs disabled
1109 static void hrtick_start(struct rq *rq, u64 delay)
1111 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1112 HRTIMER_MODE_REL_PINNED, 0);
1115 static inline void init_hrtick(void)
1118 #endif /* CONFIG_SMP */
1120 static void init_rq_hrtick(struct rq *rq)
1123 rq->hrtick_csd_pending = 0;
1125 rq->hrtick_csd.flags = 0;
1126 rq->hrtick_csd.func = __hrtick_start;
1127 rq->hrtick_csd.info = rq;
1130 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1131 rq->hrtick_timer.function = hrtick;
1133 #else /* CONFIG_SCHED_HRTICK */
1134 static inline void hrtick_clear(struct rq *rq)
1138 static inline void init_rq_hrtick(struct rq *rq)
1142 static inline void init_hrtick(void)
1145 #endif /* CONFIG_SCHED_HRTICK */
1148 * resched_task - mark a task 'to be rescheduled now'.
1150 * On UP this means the setting of the need_resched flag, on SMP it
1151 * might also involve a cross-CPU call to trigger the scheduler on
1156 #ifndef tsk_is_polling
1157 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1160 static void resched_task(struct task_struct *p)
1164 assert_raw_spin_locked(&task_rq(p)->lock);
1166 if (test_tsk_need_resched(p))
1169 set_tsk_need_resched(p);
1172 if (cpu == smp_processor_id())
1175 /* NEED_RESCHED must be visible before we test polling */
1177 if (!tsk_is_polling(p))
1178 smp_send_reschedule(cpu);
1181 static void resched_cpu(int cpu)
1183 struct rq *rq = cpu_rq(cpu);
1184 unsigned long flags;
1186 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1188 resched_task(cpu_curr(cpu));
1189 raw_spin_unlock_irqrestore(&rq->lock, flags);
1194 * When add_timer_on() enqueues a timer into the timer wheel of an
1195 * idle CPU then this timer might expire before the next timer event
1196 * which is scheduled to wake up that CPU. In case of a completely
1197 * idle system the next event might even be infinite time into the
1198 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1199 * leaves the inner idle loop so the newly added timer is taken into
1200 * account when the CPU goes back to idle and evaluates the timer
1201 * wheel for the next timer event.
1203 void wake_up_idle_cpu(int cpu)
1205 struct rq *rq = cpu_rq(cpu);
1207 if (cpu == smp_processor_id())
1211 * This is safe, as this function is called with the timer
1212 * wheel base lock of (cpu) held. When the CPU is on the way
1213 * to idle and has not yet set rq->curr to idle then it will
1214 * be serialized on the timer wheel base lock and take the new
1215 * timer into account automatically.
1217 if (rq->curr != rq->idle)
1221 * We can set TIF_RESCHED on the idle task of the other CPU
1222 * lockless. The worst case is that the other CPU runs the
1223 * idle task through an additional NOOP schedule()
1225 set_tsk_need_resched(rq->idle);
1227 /* NEED_RESCHED must be visible before we test polling */
1229 if (!tsk_is_polling(rq->idle))
1230 smp_send_reschedule(cpu);
1233 int nohz_ratelimit(int cpu)
1235 struct rq *rq = cpu_rq(cpu);
1236 u64 diff = rq->clock - rq->nohz_stamp;
1238 rq->nohz_stamp = rq->clock;
1240 return diff < (NSEC_PER_SEC / HZ) >> 1;
1243 #endif /* CONFIG_NO_HZ */
1245 static u64 sched_avg_period(void)
1247 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1250 static void sched_avg_update(struct rq *rq)
1252 s64 period = sched_avg_period();
1254 while ((s64)(rq->clock - rq->age_stamp) > period) {
1255 rq->age_stamp += period;
1260 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1262 rq->rt_avg += rt_delta;
1263 sched_avg_update(rq);
1266 #else /* !CONFIG_SMP */
1267 static void resched_task(struct task_struct *p)
1269 assert_raw_spin_locked(&task_rq(p)->lock);
1270 set_tsk_need_resched(p);
1273 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1276 #endif /* CONFIG_SMP */
1278 #if BITS_PER_LONG == 32
1279 # define WMULT_CONST (~0UL)
1281 # define WMULT_CONST (1UL << 32)
1284 #define WMULT_SHIFT 32
1287 * Shift right and round:
1289 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1292 * delta *= weight / lw
1294 static unsigned long
1295 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1296 struct load_weight *lw)
1300 if (!lw->inv_weight) {
1301 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1304 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1308 tmp = (u64)delta_exec * weight;
1310 * Check whether we'd overflow the 64-bit multiplication:
1312 if (unlikely(tmp > WMULT_CONST))
1313 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1316 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1318 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1321 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1327 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1334 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1335 * of tasks with abnormal "nice" values across CPUs the contribution that
1336 * each task makes to its run queue's load is weighted according to its
1337 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1338 * scaled version of the new time slice allocation that they receive on time
1342 #define WEIGHT_IDLEPRIO 3
1343 #define WMULT_IDLEPRIO 1431655765
1346 * Nice levels are multiplicative, with a gentle 10% change for every
1347 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1348 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1349 * that remained on nice 0.
1351 * The "10% effect" is relative and cumulative: from _any_ nice level,
1352 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1353 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1354 * If a task goes up by ~10% and another task goes down by ~10% then
1355 * the relative distance between them is ~25%.)
1357 static const int prio_to_weight[40] = {
1358 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1359 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1360 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1361 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1362 /* 0 */ 1024, 820, 655, 526, 423,
1363 /* 5 */ 335, 272, 215, 172, 137,
1364 /* 10 */ 110, 87, 70, 56, 45,
1365 /* 15 */ 36, 29, 23, 18, 15,
1369 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1371 * In cases where the weight does not change often, we can use the
1372 * precalculated inverse to speed up arithmetics by turning divisions
1373 * into multiplications:
1375 static const u32 prio_to_wmult[40] = {
1376 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1377 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1378 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1379 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1380 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1381 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1382 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1383 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1386 /* Time spent by the tasks of the cpu accounting group executing in ... */
1387 enum cpuacct_stat_index {
1388 CPUACCT_STAT_USER, /* ... user mode */
1389 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1391 CPUACCT_STAT_NSTATS,
1394 #ifdef CONFIG_CGROUP_CPUACCT
1395 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1396 static void cpuacct_update_stats(struct task_struct *tsk,
1397 enum cpuacct_stat_index idx, cputime_t val);
1399 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1400 static inline void cpuacct_update_stats(struct task_struct *tsk,
1401 enum cpuacct_stat_index idx, cputime_t val) {}
1404 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1406 update_load_add(&rq->load, load);
1409 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1411 update_load_sub(&rq->load, load);
1414 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1415 typedef int (*tg_visitor)(struct task_group *, void *);
1418 * Iterate the full tree, calling @down when first entering a node and @up when
1419 * leaving it for the final time.
1421 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1423 struct task_group *parent, *child;
1427 parent = &root_task_group;
1429 ret = (*down)(parent, data);
1432 list_for_each_entry_rcu(child, &parent->children, siblings) {
1439 ret = (*up)(parent, data);
1444 parent = parent->parent;
1453 static int tg_nop(struct task_group *tg, void *data)
1460 /* Used instead of source_load when we know the type == 0 */
1461 static unsigned long weighted_cpuload(const int cpu)
1463 return cpu_rq(cpu)->load.weight;
1467 * Return a low guess at the load of a migration-source cpu weighted
1468 * according to the scheduling class and "nice" value.
1470 * We want to under-estimate the load of migration sources, to
1471 * balance conservatively.
1473 static unsigned long source_load(int cpu, int type)
1475 struct rq *rq = cpu_rq(cpu);
1476 unsigned long total = weighted_cpuload(cpu);
1478 if (type == 0 || !sched_feat(LB_BIAS))
1481 return min(rq->cpu_load[type-1], total);
1485 * Return a high guess at the load of a migration-target cpu weighted
1486 * according to the scheduling class and "nice" value.
1488 static unsigned long target_load(int cpu, int type)
1490 struct rq *rq = cpu_rq(cpu);
1491 unsigned long total = weighted_cpuload(cpu);
1493 if (type == 0 || !sched_feat(LB_BIAS))
1496 return max(rq->cpu_load[type-1], total);
1499 static struct sched_group *group_of(int cpu)
1501 struct sched_domain *sd = rcu_dereference_sched(cpu_rq(cpu)->sd);
1509 static unsigned long power_of(int cpu)
1511 struct sched_group *group = group_of(cpu);
1514 return SCHED_LOAD_SCALE;
1516 return group->cpu_power;
1519 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1521 static unsigned long cpu_avg_load_per_task(int cpu)
1523 struct rq *rq = cpu_rq(cpu);
1524 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1527 rq->avg_load_per_task = rq->load.weight / nr_running;
1529 rq->avg_load_per_task = 0;
1531 return rq->avg_load_per_task;
1534 #ifdef CONFIG_FAIR_GROUP_SCHED
1536 static __read_mostly unsigned long *update_shares_data;
1538 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1541 * Calculate and set the cpu's group shares.
1543 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1544 unsigned long sd_shares,
1545 unsigned long sd_rq_weight,
1546 unsigned long *usd_rq_weight)
1548 unsigned long shares, rq_weight;
1551 rq_weight = usd_rq_weight[cpu];
1554 rq_weight = NICE_0_LOAD;
1558 * \Sum_j shares_j * rq_weight_i
1559 * shares_i = -----------------------------
1560 * \Sum_j rq_weight_j
1562 shares = (sd_shares * rq_weight) / sd_rq_weight;
1563 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1565 if (abs(shares - tg->se[cpu]->load.weight) >
1566 sysctl_sched_shares_thresh) {
1567 struct rq *rq = cpu_rq(cpu);
1568 unsigned long flags;
1570 raw_spin_lock_irqsave(&rq->lock, flags);
1571 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1572 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1573 __set_se_shares(tg->se[cpu], shares);
1574 raw_spin_unlock_irqrestore(&rq->lock, flags);
1579 * Re-compute the task group their per cpu shares over the given domain.
1580 * This needs to be done in a bottom-up fashion because the rq weight of a
1581 * parent group depends on the shares of its child groups.
1583 static int tg_shares_up(struct task_group *tg, void *data)
1585 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1586 unsigned long *usd_rq_weight;
1587 struct sched_domain *sd = data;
1588 unsigned long flags;
1594 local_irq_save(flags);
1595 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1597 for_each_cpu(i, sched_domain_span(sd)) {
1598 weight = tg->cfs_rq[i]->load.weight;
1599 usd_rq_weight[i] = weight;
1601 rq_weight += weight;
1603 * If there are currently no tasks on the cpu pretend there
1604 * is one of average load so that when a new task gets to
1605 * run here it will not get delayed by group starvation.
1608 weight = NICE_0_LOAD;
1610 sum_weight += weight;
1611 shares += tg->cfs_rq[i]->shares;
1615 rq_weight = sum_weight;
1617 if ((!shares && rq_weight) || shares > tg->shares)
1618 shares = tg->shares;
1620 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1621 shares = tg->shares;
1623 for_each_cpu(i, sched_domain_span(sd))
1624 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1626 local_irq_restore(flags);
1632 * Compute the cpu's hierarchical load factor for each task group.
1633 * This needs to be done in a top-down fashion because the load of a child
1634 * group is a fraction of its parents load.
1636 static int tg_load_down(struct task_group *tg, void *data)
1639 long cpu = (long)data;
1642 load = cpu_rq(cpu)->load.weight;
1644 load = tg->parent->cfs_rq[cpu]->h_load;
1645 load *= tg->cfs_rq[cpu]->shares;
1646 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1649 tg->cfs_rq[cpu]->h_load = load;
1654 static void update_shares(struct sched_domain *sd)
1659 if (root_task_group_empty())
1662 now = cpu_clock(raw_smp_processor_id());
1663 elapsed = now - sd->last_update;
1665 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1666 sd->last_update = now;
1667 walk_tg_tree(tg_nop, tg_shares_up, sd);
1671 static void update_h_load(long cpu)
1673 if (root_task_group_empty())
1676 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1681 static inline void update_shares(struct sched_domain *sd)
1687 #ifdef CONFIG_PREEMPT
1689 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1692 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1693 * way at the expense of forcing extra atomic operations in all
1694 * invocations. This assures that the double_lock is acquired using the
1695 * same underlying policy as the spinlock_t on this architecture, which
1696 * reduces latency compared to the unfair variant below. However, it
1697 * also adds more overhead and therefore may reduce throughput.
1699 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1700 __releases(this_rq->lock)
1701 __acquires(busiest->lock)
1702 __acquires(this_rq->lock)
1704 raw_spin_unlock(&this_rq->lock);
1705 double_rq_lock(this_rq, busiest);
1712 * Unfair double_lock_balance: Optimizes throughput at the expense of
1713 * latency by eliminating extra atomic operations when the locks are
1714 * already in proper order on entry. This favors lower cpu-ids and will
1715 * grant the double lock to lower cpus over higher ids under contention,
1716 * regardless of entry order into the function.
1718 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1719 __releases(this_rq->lock)
1720 __acquires(busiest->lock)
1721 __acquires(this_rq->lock)
1725 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1726 if (busiest < this_rq) {
1727 raw_spin_unlock(&this_rq->lock);
1728 raw_spin_lock(&busiest->lock);
1729 raw_spin_lock_nested(&this_rq->lock,
1730 SINGLE_DEPTH_NESTING);
1733 raw_spin_lock_nested(&busiest->lock,
1734 SINGLE_DEPTH_NESTING);
1739 #endif /* CONFIG_PREEMPT */
1742 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1744 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1746 if (unlikely(!irqs_disabled())) {
1747 /* printk() doesn't work good under rq->lock */
1748 raw_spin_unlock(&this_rq->lock);
1752 return _double_lock_balance(this_rq, busiest);
1755 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1756 __releases(busiest->lock)
1758 raw_spin_unlock(&busiest->lock);
1759 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1763 * double_rq_lock - safely lock two runqueues
1765 * Note this does not disable interrupts like task_rq_lock,
1766 * you need to do so manually before calling.
1768 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1769 __acquires(rq1->lock)
1770 __acquires(rq2->lock)
1772 BUG_ON(!irqs_disabled());
1774 raw_spin_lock(&rq1->lock);
1775 __acquire(rq2->lock); /* Fake it out ;) */
1778 raw_spin_lock(&rq1->lock);
1779 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1781 raw_spin_lock(&rq2->lock);
1782 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1785 update_rq_clock(rq1);
1786 update_rq_clock(rq2);
1790 * double_rq_unlock - safely unlock two runqueues
1792 * Note this does not restore interrupts like task_rq_unlock,
1793 * you need to do so manually after calling.
1795 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1796 __releases(rq1->lock)
1797 __releases(rq2->lock)
1799 raw_spin_unlock(&rq1->lock);
1801 raw_spin_unlock(&rq2->lock);
1803 __release(rq2->lock);
1808 #ifdef CONFIG_FAIR_GROUP_SCHED
1809 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1812 cfs_rq->shares = shares;
1817 static void calc_load_account_active(struct rq *this_rq);
1818 static void update_sysctl(void);
1819 static int get_update_sysctl_factor(void);
1821 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1823 set_task_rq(p, cpu);
1826 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1827 * successfuly executed on another CPU. We must ensure that updates of
1828 * per-task data have been completed by this moment.
1831 task_thread_info(p)->cpu = cpu;
1835 static const struct sched_class rt_sched_class;
1837 #define sched_class_highest (&rt_sched_class)
1838 #define for_each_class(class) \
1839 for (class = sched_class_highest; class; class = class->next)
1841 #include "sched_stats.h"
1843 static void inc_nr_running(struct rq *rq)
1848 static void dec_nr_running(struct rq *rq)
1853 static void set_load_weight(struct task_struct *p)
1855 if (task_has_rt_policy(p)) {
1856 p->se.load.weight = prio_to_weight[0] * 2;
1857 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1862 * SCHED_IDLE tasks get minimal weight:
1864 if (p->policy == SCHED_IDLE) {
1865 p->se.load.weight = WEIGHT_IDLEPRIO;
1866 p->se.load.inv_weight = WMULT_IDLEPRIO;
1870 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1871 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1874 static void update_avg(u64 *avg, u64 sample)
1876 s64 diff = sample - *avg;
1881 enqueue_task(struct rq *rq, struct task_struct *p, int wakeup, bool head)
1883 sched_info_queued(p);
1884 p->sched_class->enqueue_task(rq, p, wakeup, head);
1888 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1890 sched_info_dequeued(p);
1891 p->sched_class->dequeue_task(rq, p, sleep);
1896 * activate_task - move a task to the runqueue.
1898 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1900 if (task_contributes_to_load(p))
1901 rq->nr_uninterruptible--;
1903 enqueue_task(rq, p, wakeup, false);
1908 * deactivate_task - remove a task from the runqueue.
1910 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1912 if (task_contributes_to_load(p))
1913 rq->nr_uninterruptible++;
1915 dequeue_task(rq, p, sleep);
1919 #include "sched_idletask.c"
1920 #include "sched_fair.c"
1921 #include "sched_rt.c"
1922 #ifdef CONFIG_SCHED_DEBUG
1923 # include "sched_debug.c"
1927 * __normal_prio - return the priority that is based on the static prio
1929 static inline int __normal_prio(struct task_struct *p)
1931 return p->static_prio;
1935 * Calculate the expected normal priority: i.e. priority
1936 * without taking RT-inheritance into account. Might be
1937 * boosted by interactivity modifiers. Changes upon fork,
1938 * setprio syscalls, and whenever the interactivity
1939 * estimator recalculates.
1941 static inline int normal_prio(struct task_struct *p)
1945 if (task_has_rt_policy(p))
1946 prio = MAX_RT_PRIO-1 - p->rt_priority;
1948 prio = __normal_prio(p);
1953 * Calculate the current priority, i.e. the priority
1954 * taken into account by the scheduler. This value might
1955 * be boosted by RT tasks, or might be boosted by
1956 * interactivity modifiers. Will be RT if the task got
1957 * RT-boosted. If not then it returns p->normal_prio.
1959 static int effective_prio(struct task_struct *p)
1961 p->normal_prio = normal_prio(p);
1963 * If we are RT tasks or we were boosted to RT priority,
1964 * keep the priority unchanged. Otherwise, update priority
1965 * to the normal priority:
1967 if (!rt_prio(p->prio))
1968 return p->normal_prio;
1973 * task_curr - is this task currently executing on a CPU?
1974 * @p: the task in question.
1976 inline int task_curr(const struct task_struct *p)
1978 return cpu_curr(task_cpu(p)) == p;
1981 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1982 const struct sched_class *prev_class,
1983 int oldprio, int running)
1985 if (prev_class != p->sched_class) {
1986 if (prev_class->switched_from)
1987 prev_class->switched_from(rq, p, running);
1988 p->sched_class->switched_to(rq, p, running);
1990 p->sched_class->prio_changed(rq, p, oldprio, running);
1995 * Is this task likely cache-hot:
1998 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2002 if (p->sched_class != &fair_sched_class)
2006 * Buddy candidates are cache hot:
2008 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2009 (&p->se == cfs_rq_of(&p->se)->next ||
2010 &p->se == cfs_rq_of(&p->se)->last))
2013 if (sysctl_sched_migration_cost == -1)
2015 if (sysctl_sched_migration_cost == 0)
2018 delta = now - p->se.exec_start;
2020 return delta < (s64)sysctl_sched_migration_cost;
2023 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2025 #ifdef CONFIG_SCHED_DEBUG
2027 * We should never call set_task_cpu() on a blocked task,
2028 * ttwu() will sort out the placement.
2030 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2031 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2034 trace_sched_migrate_task(p, new_cpu);
2036 if (task_cpu(p) != new_cpu) {
2037 p->se.nr_migrations++;
2038 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2041 __set_task_cpu(p, new_cpu);
2044 struct migration_req {
2045 struct list_head list;
2047 struct task_struct *task;
2050 struct completion done;
2054 * The task's runqueue lock must be held.
2055 * Returns true if you have to wait for migration thread.
2058 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2060 struct rq *rq = task_rq(p);
2063 * If the task is not on a runqueue (and not running), then
2064 * the next wake-up will properly place the task.
2066 if (!p->se.on_rq && !task_running(rq, p))
2069 init_completion(&req->done);
2071 req->dest_cpu = dest_cpu;
2072 list_add(&req->list, &rq->migration_queue);
2078 * wait_task_context_switch - wait for a thread to complete at least one
2081 * @p must not be current.
2083 void wait_task_context_switch(struct task_struct *p)
2085 unsigned long nvcsw, nivcsw, flags;
2093 * The runqueue is assigned before the actual context
2094 * switch. We need to take the runqueue lock.
2096 * We could check initially without the lock but it is
2097 * very likely that we need to take the lock in every
2100 rq = task_rq_lock(p, &flags);
2101 running = task_running(rq, p);
2102 task_rq_unlock(rq, &flags);
2104 if (likely(!running))
2107 * The switch count is incremented before the actual
2108 * context switch. We thus wait for two switches to be
2109 * sure at least one completed.
2111 if ((p->nvcsw - nvcsw) > 1)
2113 if ((p->nivcsw - nivcsw) > 1)
2121 * wait_task_inactive - wait for a thread to unschedule.
2123 * If @match_state is nonzero, it's the @p->state value just checked and
2124 * not expected to change. If it changes, i.e. @p might have woken up,
2125 * then return zero. When we succeed in waiting for @p to be off its CPU,
2126 * we return a positive number (its total switch count). If a second call
2127 * a short while later returns the same number, the caller can be sure that
2128 * @p has remained unscheduled the whole time.
2130 * The caller must ensure that the task *will* unschedule sometime soon,
2131 * else this function might spin for a *long* time. This function can't
2132 * be called with interrupts off, or it may introduce deadlock with
2133 * smp_call_function() if an IPI is sent by the same process we are
2134 * waiting to become inactive.
2136 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2138 unsigned long flags;
2145 * We do the initial early heuristics without holding
2146 * any task-queue locks at all. We'll only try to get
2147 * the runqueue lock when things look like they will
2153 * If the task is actively running on another CPU
2154 * still, just relax and busy-wait without holding
2157 * NOTE! Since we don't hold any locks, it's not
2158 * even sure that "rq" stays as the right runqueue!
2159 * But we don't care, since "task_running()" will
2160 * return false if the runqueue has changed and p
2161 * is actually now running somewhere else!
2163 while (task_running(rq, p)) {
2164 if (match_state && unlikely(p->state != match_state))
2170 * Ok, time to look more closely! We need the rq
2171 * lock now, to be *sure*. If we're wrong, we'll
2172 * just go back and repeat.
2174 rq = task_rq_lock(p, &flags);
2175 trace_sched_wait_task(rq, p);
2176 running = task_running(rq, p);
2177 on_rq = p->se.on_rq;
2179 if (!match_state || p->state == match_state)
2180 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2181 task_rq_unlock(rq, &flags);
2184 * If it changed from the expected state, bail out now.
2186 if (unlikely(!ncsw))
2190 * Was it really running after all now that we
2191 * checked with the proper locks actually held?
2193 * Oops. Go back and try again..
2195 if (unlikely(running)) {
2201 * It's not enough that it's not actively running,
2202 * it must be off the runqueue _entirely_, and not
2205 * So if it was still runnable (but just not actively
2206 * running right now), it's preempted, and we should
2207 * yield - it could be a while.
2209 if (unlikely(on_rq)) {
2210 schedule_timeout_uninterruptible(1);
2215 * Ahh, all good. It wasn't running, and it wasn't
2216 * runnable, which means that it will never become
2217 * running in the future either. We're all done!
2226 * kick_process - kick a running thread to enter/exit the kernel
2227 * @p: the to-be-kicked thread
2229 * Cause a process which is running on another CPU to enter
2230 * kernel-mode, without any delay. (to get signals handled.)
2232 * NOTE: this function doesnt have to take the runqueue lock,
2233 * because all it wants to ensure is that the remote task enters
2234 * the kernel. If the IPI races and the task has been migrated
2235 * to another CPU then no harm is done and the purpose has been
2238 void kick_process(struct task_struct *p)
2244 if ((cpu != smp_processor_id()) && task_curr(p))
2245 smp_send_reschedule(cpu);
2248 EXPORT_SYMBOL_GPL(kick_process);
2249 #endif /* CONFIG_SMP */
2252 * task_oncpu_function_call - call a function on the cpu on which a task runs
2253 * @p: the task to evaluate
2254 * @func: the function to be called
2255 * @info: the function call argument
2257 * Calls the function @func when the task is currently running. This might
2258 * be on the current CPU, which just calls the function directly
2260 void task_oncpu_function_call(struct task_struct *p,
2261 void (*func) (void *info), void *info)
2268 smp_call_function_single(cpu, func, info, 1);
2273 static int select_fallback_rq(int cpu, struct task_struct *p)
2276 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2278 /* Look for allowed, online CPU in same node. */
2279 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2280 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2283 /* Any allowed, online CPU? */
2284 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2285 if (dest_cpu < nr_cpu_ids)
2288 /* No more Mr. Nice Guy. */
2289 if (dest_cpu >= nr_cpu_ids) {
2291 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2293 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2296 * Don't tell them about moving exiting tasks or
2297 * kernel threads (both mm NULL), since they never
2300 if (p->mm && printk_ratelimit()) {
2301 printk(KERN_INFO "process %d (%s) no "
2302 "longer affine to cpu%d\n",
2303 task_pid_nr(p), p->comm, cpu);
2311 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2312 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2315 * exec: is unstable, retry loop
2316 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2319 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2321 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2324 * In order not to call set_task_cpu() on a blocking task we need
2325 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2328 * Since this is common to all placement strategies, this lives here.
2330 * [ this allows ->select_task() to simply return task_cpu(p) and
2331 * not worry about this generic constraint ]
2333 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2335 cpu = select_fallback_rq(task_cpu(p), p);
2342 * try_to_wake_up - wake up a thread
2343 * @p: the to-be-woken-up thread
2344 * @state: the mask of task states that can be woken
2345 * @sync: do a synchronous wakeup?
2347 * Put it on the run-queue if it's not already there. The "current"
2348 * thread is always on the run-queue (except when the actual
2349 * re-schedule is in progress), and as such you're allowed to do
2350 * the simpler "current->state = TASK_RUNNING" to mark yourself
2351 * runnable without the overhead of this.
2353 * returns failure only if the task is already active.
2355 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2358 int cpu, orig_cpu, this_cpu, success = 0;
2359 unsigned long flags;
2362 if (!sched_feat(SYNC_WAKEUPS))
2363 wake_flags &= ~WF_SYNC;
2365 this_cpu = get_cpu();
2368 rq = task_rq_lock(p, &flags);
2369 update_rq_clock(rq);
2370 if (!(p->state & state))
2380 if (unlikely(task_running(rq, p)))
2384 * In order to handle concurrent wakeups and release the rq->lock
2385 * we put the task in TASK_WAKING state.
2387 * First fix up the nr_uninterruptible count:
2389 if (task_contributes_to_load(p))
2390 rq->nr_uninterruptible--;
2391 p->state = TASK_WAKING;
2393 if (p->sched_class->task_waking)
2394 p->sched_class->task_waking(rq, p);
2396 __task_rq_unlock(rq);
2398 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2399 if (cpu != orig_cpu) {
2401 * Since we migrate the task without holding any rq->lock,
2402 * we need to be careful with task_rq_lock(), since that
2403 * might end up locking an invalid rq.
2405 set_task_cpu(p, cpu);
2409 raw_spin_lock(&rq->lock);
2410 update_rq_clock(rq);
2413 * We migrated the task without holding either rq->lock, however
2414 * since the task is not on the task list itself, nobody else
2415 * will try and migrate the task, hence the rq should match the
2416 * cpu we just moved it to.
2418 WARN_ON(task_cpu(p) != cpu);
2419 WARN_ON(p->state != TASK_WAKING);
2421 #ifdef CONFIG_SCHEDSTATS
2422 schedstat_inc(rq, ttwu_count);
2423 if (cpu == this_cpu)
2424 schedstat_inc(rq, ttwu_local);
2426 struct sched_domain *sd;
2427 for_each_domain(this_cpu, sd) {
2428 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2429 schedstat_inc(sd, ttwu_wake_remote);
2434 #endif /* CONFIG_SCHEDSTATS */
2437 #endif /* CONFIG_SMP */
2438 schedstat_inc(p, se.statistics.nr_wakeups);
2439 if (wake_flags & WF_SYNC)
2440 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2441 if (orig_cpu != cpu)
2442 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2443 if (cpu == this_cpu)
2444 schedstat_inc(p, se.statistics.nr_wakeups_local);
2446 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2447 activate_task(rq, p, 1);
2451 trace_sched_wakeup(rq, p, success);
2452 check_preempt_curr(rq, p, wake_flags);
2454 p->state = TASK_RUNNING;
2456 if (p->sched_class->task_woken)
2457 p->sched_class->task_woken(rq, p);
2459 if (unlikely(rq->idle_stamp)) {
2460 u64 delta = rq->clock - rq->idle_stamp;
2461 u64 max = 2*sysctl_sched_migration_cost;
2466 update_avg(&rq->avg_idle, delta);
2471 task_rq_unlock(rq, &flags);
2478 * wake_up_process - Wake up a specific process
2479 * @p: The process to be woken up.
2481 * Attempt to wake up the nominated process and move it to the set of runnable
2482 * processes. Returns 1 if the process was woken up, 0 if it was already
2485 * It may be assumed that this function implies a write memory barrier before
2486 * changing the task state if and only if any tasks are woken up.
2488 int wake_up_process(struct task_struct *p)
2490 return try_to_wake_up(p, TASK_ALL, 0);
2492 EXPORT_SYMBOL(wake_up_process);
2494 int wake_up_state(struct task_struct *p, unsigned int state)
2496 return try_to_wake_up(p, state, 0);
2500 * Perform scheduler related setup for a newly forked process p.
2501 * p is forked by current.
2503 * __sched_fork() is basic setup used by init_idle() too:
2505 static void __sched_fork(struct task_struct *p)
2507 p->se.exec_start = 0;
2508 p->se.sum_exec_runtime = 0;
2509 p->se.prev_sum_exec_runtime = 0;
2510 p->se.nr_migrations = 0;
2512 #ifdef CONFIG_SCHEDSTATS
2513 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2516 INIT_LIST_HEAD(&p->rt.run_list);
2518 INIT_LIST_HEAD(&p->se.group_node);
2520 #ifdef CONFIG_PREEMPT_NOTIFIERS
2521 INIT_HLIST_HEAD(&p->preempt_notifiers);
2526 * fork()/clone()-time setup:
2528 void sched_fork(struct task_struct *p, int clone_flags)
2530 int cpu = get_cpu();
2534 * We mark the process as waking here. This guarantees that
2535 * nobody will actually run it, and a signal or other external
2536 * event cannot wake it up and insert it on the runqueue either.
2538 p->state = TASK_WAKING;
2541 * Revert to default priority/policy on fork if requested.
2543 if (unlikely(p->sched_reset_on_fork)) {
2544 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2545 p->policy = SCHED_NORMAL;
2546 p->normal_prio = p->static_prio;
2549 if (PRIO_TO_NICE(p->static_prio) < 0) {
2550 p->static_prio = NICE_TO_PRIO(0);
2551 p->normal_prio = p->static_prio;
2556 * We don't need the reset flag anymore after the fork. It has
2557 * fulfilled its duty:
2559 p->sched_reset_on_fork = 0;
2563 * Make sure we do not leak PI boosting priority to the child.
2565 p->prio = current->normal_prio;
2567 if (!rt_prio(p->prio))
2568 p->sched_class = &fair_sched_class;
2570 if (p->sched_class->task_fork)
2571 p->sched_class->task_fork(p);
2573 set_task_cpu(p, cpu);
2575 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2576 if (likely(sched_info_on()))
2577 memset(&p->sched_info, 0, sizeof(p->sched_info));
2579 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2582 #ifdef CONFIG_PREEMPT
2583 /* Want to start with kernel preemption disabled. */
2584 task_thread_info(p)->preempt_count = 1;
2586 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2592 * wake_up_new_task - wake up a newly created task for the first time.
2594 * This function will do some initial scheduler statistics housekeeping
2595 * that must be done for every newly created context, then puts the task
2596 * on the runqueue and wakes it.
2598 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2600 unsigned long flags;
2602 int cpu = get_cpu();
2606 * Fork balancing, do it here and not earlier because:
2607 * - cpus_allowed can change in the fork path
2608 * - any previously selected cpu might disappear through hotplug
2610 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2611 * ->cpus_allowed is stable, we have preemption disabled, meaning
2612 * cpu_online_mask is stable.
2614 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2615 set_task_cpu(p, cpu);
2619 * Since the task is not on the rq and we still have TASK_WAKING set
2620 * nobody else will migrate this task.
2623 raw_spin_lock_irqsave(&rq->lock, flags);
2625 BUG_ON(p->state != TASK_WAKING);
2626 p->state = TASK_RUNNING;
2627 update_rq_clock(rq);
2628 activate_task(rq, p, 0);
2629 trace_sched_wakeup_new(rq, p, 1);
2630 check_preempt_curr(rq, p, WF_FORK);
2632 if (p->sched_class->task_woken)
2633 p->sched_class->task_woken(rq, p);
2635 task_rq_unlock(rq, &flags);
2639 #ifdef CONFIG_PREEMPT_NOTIFIERS
2642 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2643 * @notifier: notifier struct to register
2645 void preempt_notifier_register(struct preempt_notifier *notifier)
2647 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2649 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2652 * preempt_notifier_unregister - no longer interested in preemption notifications
2653 * @notifier: notifier struct to unregister
2655 * This is safe to call from within a preemption notifier.
2657 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2659 hlist_del(¬ifier->link);
2661 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2663 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2665 struct preempt_notifier *notifier;
2666 struct hlist_node *node;
2668 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2669 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2673 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2674 struct task_struct *next)
2676 struct preempt_notifier *notifier;
2677 struct hlist_node *node;
2679 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2680 notifier->ops->sched_out(notifier, next);
2683 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2685 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2690 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2691 struct task_struct *next)
2695 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2698 * prepare_task_switch - prepare to switch tasks
2699 * @rq: the runqueue preparing to switch
2700 * @prev: the current task that is being switched out
2701 * @next: the task we are going to switch to.
2703 * This is called with the rq lock held and interrupts off. It must
2704 * be paired with a subsequent finish_task_switch after the context
2707 * prepare_task_switch sets up locking and calls architecture specific
2711 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2712 struct task_struct *next)
2714 fire_sched_out_preempt_notifiers(prev, next);
2715 prepare_lock_switch(rq, next);
2716 prepare_arch_switch(next);
2720 * finish_task_switch - clean up after a task-switch
2721 * @rq: runqueue associated with task-switch
2722 * @prev: the thread we just switched away from.
2724 * finish_task_switch must be called after the context switch, paired
2725 * with a prepare_task_switch call before the context switch.
2726 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2727 * and do any other architecture-specific cleanup actions.
2729 * Note that we may have delayed dropping an mm in context_switch(). If
2730 * so, we finish that here outside of the runqueue lock. (Doing it
2731 * with the lock held can cause deadlocks; see schedule() for
2734 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2735 __releases(rq->lock)
2737 struct mm_struct *mm = rq->prev_mm;
2743 * A task struct has one reference for the use as "current".
2744 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2745 * schedule one last time. The schedule call will never return, and
2746 * the scheduled task must drop that reference.
2747 * The test for TASK_DEAD must occur while the runqueue locks are
2748 * still held, otherwise prev could be scheduled on another cpu, die
2749 * there before we look at prev->state, and then the reference would
2751 * Manfred Spraul <manfred@colorfullife.com>
2753 prev_state = prev->state;
2754 finish_arch_switch(prev);
2755 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2756 local_irq_disable();
2757 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2758 perf_event_task_sched_in(current);
2759 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2761 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2762 finish_lock_switch(rq, prev);
2764 fire_sched_in_preempt_notifiers(current);
2767 if (unlikely(prev_state == TASK_DEAD)) {
2769 * Remove function-return probe instances associated with this
2770 * task and put them back on the free list.
2772 kprobe_flush_task(prev);
2773 put_task_struct(prev);
2779 /* assumes rq->lock is held */
2780 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2782 if (prev->sched_class->pre_schedule)
2783 prev->sched_class->pre_schedule(rq, prev);
2786 /* rq->lock is NOT held, but preemption is disabled */
2787 static inline void post_schedule(struct rq *rq)
2789 if (rq->post_schedule) {
2790 unsigned long flags;
2792 raw_spin_lock_irqsave(&rq->lock, flags);
2793 if (rq->curr->sched_class->post_schedule)
2794 rq->curr->sched_class->post_schedule(rq);
2795 raw_spin_unlock_irqrestore(&rq->lock, flags);
2797 rq->post_schedule = 0;
2803 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2807 static inline void post_schedule(struct rq *rq)
2814 * schedule_tail - first thing a freshly forked thread must call.
2815 * @prev: the thread we just switched away from.
2817 asmlinkage void schedule_tail(struct task_struct *prev)
2818 __releases(rq->lock)
2820 struct rq *rq = this_rq();
2822 finish_task_switch(rq, prev);
2825 * FIXME: do we need to worry about rq being invalidated by the
2830 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2831 /* In this case, finish_task_switch does not reenable preemption */
2834 if (current->set_child_tid)
2835 put_user(task_pid_vnr(current), current->set_child_tid);
2839 * context_switch - switch to the new MM and the new
2840 * thread's register state.
2843 context_switch(struct rq *rq, struct task_struct *prev,
2844 struct task_struct *next)
2846 struct mm_struct *mm, *oldmm;
2848 prepare_task_switch(rq, prev, next);
2849 trace_sched_switch(rq, prev, next);
2851 oldmm = prev->active_mm;
2853 * For paravirt, this is coupled with an exit in switch_to to
2854 * combine the page table reload and the switch backend into
2857 arch_start_context_switch(prev);
2860 next->active_mm = oldmm;
2861 atomic_inc(&oldmm->mm_count);
2862 enter_lazy_tlb(oldmm, next);
2864 switch_mm(oldmm, mm, next);
2866 if (likely(!prev->mm)) {
2867 prev->active_mm = NULL;
2868 rq->prev_mm = oldmm;
2871 * Since the runqueue lock will be released by the next
2872 * task (which is an invalid locking op but in the case
2873 * of the scheduler it's an obvious special-case), so we
2874 * do an early lockdep release here:
2876 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2877 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2880 /* Here we just switch the register state and the stack. */
2881 switch_to(prev, next, prev);
2885 * this_rq must be evaluated again because prev may have moved
2886 * CPUs since it called schedule(), thus the 'rq' on its stack
2887 * frame will be invalid.
2889 finish_task_switch(this_rq(), prev);
2893 * nr_running, nr_uninterruptible and nr_context_switches:
2895 * externally visible scheduler statistics: current number of runnable
2896 * threads, current number of uninterruptible-sleeping threads, total
2897 * number of context switches performed since bootup.
2899 unsigned long nr_running(void)
2901 unsigned long i, sum = 0;
2903 for_each_online_cpu(i)
2904 sum += cpu_rq(i)->nr_running;
2909 unsigned long nr_uninterruptible(void)
2911 unsigned long i, sum = 0;
2913 for_each_possible_cpu(i)
2914 sum += cpu_rq(i)->nr_uninterruptible;
2917 * Since we read the counters lockless, it might be slightly
2918 * inaccurate. Do not allow it to go below zero though:
2920 if (unlikely((long)sum < 0))
2926 unsigned long long nr_context_switches(void)
2929 unsigned long long sum = 0;
2931 for_each_possible_cpu(i)
2932 sum += cpu_rq(i)->nr_switches;
2937 unsigned long nr_iowait(void)
2939 unsigned long i, sum = 0;
2941 for_each_possible_cpu(i)
2942 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2947 unsigned long nr_iowait_cpu(void)
2949 struct rq *this = this_rq();
2950 return atomic_read(&this->nr_iowait);
2953 unsigned long this_cpu_load(void)
2955 struct rq *this = this_rq();
2956 return this->cpu_load[0];
2960 /* Variables and functions for calc_load */
2961 static atomic_long_t calc_load_tasks;
2962 static unsigned long calc_load_update;
2963 unsigned long avenrun[3];
2964 EXPORT_SYMBOL(avenrun);
2967 * get_avenrun - get the load average array
2968 * @loads: pointer to dest load array
2969 * @offset: offset to add
2970 * @shift: shift count to shift the result left
2972 * These values are estimates at best, so no need for locking.
2974 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2976 loads[0] = (avenrun[0] + offset) << shift;
2977 loads[1] = (avenrun[1] + offset) << shift;
2978 loads[2] = (avenrun[2] + offset) << shift;
2981 static unsigned long
2982 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2985 load += active * (FIXED_1 - exp);
2986 return load >> FSHIFT;
2990 * calc_load - update the avenrun load estimates 10 ticks after the
2991 * CPUs have updated calc_load_tasks.
2993 void calc_global_load(void)
2995 unsigned long upd = calc_load_update + 10;
2998 if (time_before(jiffies, upd))
3001 active = atomic_long_read(&calc_load_tasks);
3002 active = active > 0 ? active * FIXED_1 : 0;
3004 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3005 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3006 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3008 calc_load_update += LOAD_FREQ;
3012 * Either called from update_cpu_load() or from a cpu going idle
3014 static void calc_load_account_active(struct rq *this_rq)
3016 long nr_active, delta;
3018 nr_active = this_rq->nr_running;
3019 nr_active += (long) this_rq->nr_uninterruptible;
3021 if (nr_active != this_rq->calc_load_active) {
3022 delta = nr_active - this_rq->calc_load_active;
3023 this_rq->calc_load_active = nr_active;
3024 atomic_long_add(delta, &calc_load_tasks);
3029 * Update rq->cpu_load[] statistics. This function is usually called every
3030 * scheduler tick (TICK_NSEC).
3032 static void update_cpu_load(struct rq *this_rq)
3034 unsigned long this_load = this_rq->load.weight;
3037 this_rq->nr_load_updates++;
3039 /* Update our load: */
3040 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3041 unsigned long old_load, new_load;
3043 /* scale is effectively 1 << i now, and >> i divides by scale */
3045 old_load = this_rq->cpu_load[i];
3046 new_load = this_load;
3048 * Round up the averaging division if load is increasing. This
3049 * prevents us from getting stuck on 9 if the load is 10, for
3052 if (new_load > old_load)
3053 new_load += scale-1;
3054 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3057 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3058 this_rq->calc_load_update += LOAD_FREQ;
3059 calc_load_account_active(this_rq);
3066 * sched_exec - execve() is a valuable balancing opportunity, because at
3067 * this point the task has the smallest effective memory and cache footprint.
3069 void sched_exec(void)
3071 struct task_struct *p = current;
3072 struct migration_req req;
3073 int dest_cpu, this_cpu;
3074 unsigned long flags;
3078 this_cpu = get_cpu();
3079 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3080 if (dest_cpu == this_cpu) {
3085 rq = task_rq_lock(p, &flags);
3089 * select_task_rq() can race against ->cpus_allowed
3091 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3092 || unlikely(!cpu_active(dest_cpu))) {
3093 task_rq_unlock(rq, &flags);
3097 /* force the process onto the specified CPU */
3098 if (migrate_task(p, dest_cpu, &req)) {
3099 /* Need to wait for migration thread (might exit: take ref). */
3100 struct task_struct *mt = rq->migration_thread;
3102 get_task_struct(mt);
3103 task_rq_unlock(rq, &flags);
3104 wake_up_process(mt);
3105 put_task_struct(mt);
3106 wait_for_completion(&req.done);
3110 task_rq_unlock(rq, &flags);
3115 DEFINE_PER_CPU(struct kernel_stat, kstat);
3117 EXPORT_PER_CPU_SYMBOL(kstat);
3120 * Return any ns on the sched_clock that have not yet been accounted in
3121 * @p in case that task is currently running.
3123 * Called with task_rq_lock() held on @rq.
3125 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3129 if (task_current(rq, p)) {
3130 update_rq_clock(rq);
3131 ns = rq->clock - p->se.exec_start;
3139 unsigned long long task_delta_exec(struct task_struct *p)
3141 unsigned long flags;
3145 rq = task_rq_lock(p, &flags);
3146 ns = do_task_delta_exec(p, rq);
3147 task_rq_unlock(rq, &flags);
3153 * Return accounted runtime for the task.
3154 * In case the task is currently running, return the runtime plus current's
3155 * pending runtime that have not been accounted yet.
3157 unsigned long long task_sched_runtime(struct task_struct *p)
3159 unsigned long flags;
3163 rq = task_rq_lock(p, &flags);
3164 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3165 task_rq_unlock(rq, &flags);
3171 * Return sum_exec_runtime for the thread group.
3172 * In case the task is currently running, return the sum plus current's
3173 * pending runtime that have not been accounted yet.
3175 * Note that the thread group might have other running tasks as well,
3176 * so the return value not includes other pending runtime that other
3177 * running tasks might have.
3179 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3181 struct task_cputime totals;
3182 unsigned long flags;
3186 rq = task_rq_lock(p, &flags);
3187 thread_group_cputime(p, &totals);
3188 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3189 task_rq_unlock(rq, &flags);
3195 * Account user cpu time to a process.
3196 * @p: the process that the cpu time gets accounted to
3197 * @cputime: the cpu time spent in user space since the last update
3198 * @cputime_scaled: cputime scaled by cpu frequency
3200 void account_user_time(struct task_struct *p, cputime_t cputime,
3201 cputime_t cputime_scaled)
3203 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3206 /* Add user time to process. */
3207 p->utime = cputime_add(p->utime, cputime);
3208 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3209 account_group_user_time(p, cputime);
3211 /* Add user time to cpustat. */
3212 tmp = cputime_to_cputime64(cputime);
3213 if (TASK_NICE(p) > 0)
3214 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3216 cpustat->user = cputime64_add(cpustat->user, tmp);
3218 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3219 /* Account for user time used */
3220 acct_update_integrals(p);
3224 * Account guest cpu time to a process.
3225 * @p: the process that the cpu time gets accounted to
3226 * @cputime: the cpu time spent in virtual machine since the last update
3227 * @cputime_scaled: cputime scaled by cpu frequency
3229 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3230 cputime_t cputime_scaled)
3233 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3235 tmp = cputime_to_cputime64(cputime);
3237 /* Add guest 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);
3241 p->gtime = cputime_add(p->gtime, cputime);
3243 /* Add guest time to cpustat. */
3244 if (TASK_NICE(p) > 0) {
3245 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3246 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3248 cpustat->user = cputime64_add(cpustat->user, tmp);
3249 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3254 * Account system cpu time to a process.
3255 * @p: the process that the cpu time gets accounted to
3256 * @hardirq_offset: the offset to subtract from hardirq_count()
3257 * @cputime: the cpu time spent in kernel space since the last update
3258 * @cputime_scaled: cputime scaled by cpu frequency
3260 void account_system_time(struct task_struct *p, int hardirq_offset,
3261 cputime_t cputime, cputime_t cputime_scaled)
3263 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3266 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3267 account_guest_time(p, cputime, cputime_scaled);
3271 /* Add system time to process. */
3272 p->stime = cputime_add(p->stime, cputime);
3273 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3274 account_group_system_time(p, cputime);
3276 /* Add system time to cpustat. */
3277 tmp = cputime_to_cputime64(cputime);
3278 if (hardirq_count() - hardirq_offset)
3279 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3280 else if (softirq_count())
3281 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3283 cpustat->system = cputime64_add(cpustat->system, tmp);
3285 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3287 /* Account for system time used */
3288 acct_update_integrals(p);
3292 * Account for involuntary wait time.
3293 * @steal: the cpu time spent in involuntary wait
3295 void account_steal_time(cputime_t cputime)
3297 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3298 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3300 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3304 * Account for idle time.
3305 * @cputime: the cpu time spent in idle wait
3307 void account_idle_time(cputime_t cputime)
3309 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3310 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3311 struct rq *rq = this_rq();
3313 if (atomic_read(&rq->nr_iowait) > 0)
3314 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3316 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3319 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3322 * Account a single tick of cpu time.
3323 * @p: the process that the cpu time gets accounted to
3324 * @user_tick: indicates if the tick is a user or a system tick
3326 void account_process_tick(struct task_struct *p, int user_tick)
3328 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3329 struct rq *rq = this_rq();
3332 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3333 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3334 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3337 account_idle_time(cputime_one_jiffy);
3341 * Account multiple ticks of steal time.
3342 * @p: the process from which the cpu time has been stolen
3343 * @ticks: number of stolen ticks
3345 void account_steal_ticks(unsigned long ticks)
3347 account_steal_time(jiffies_to_cputime(ticks));
3351 * Account multiple ticks of idle time.
3352 * @ticks: number of stolen ticks
3354 void account_idle_ticks(unsigned long ticks)
3356 account_idle_time(jiffies_to_cputime(ticks));
3362 * Use precise platform statistics if available:
3364 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3365 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3371 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3373 struct task_cputime cputime;
3375 thread_group_cputime(p, &cputime);
3377 *ut = cputime.utime;
3378 *st = cputime.stime;
3382 #ifndef nsecs_to_cputime
3383 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3386 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3388 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3391 * Use CFS's precise accounting:
3393 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3398 temp = (u64)(rtime * utime);
3399 do_div(temp, total);
3400 utime = (cputime_t)temp;
3405 * Compare with previous values, to keep monotonicity:
3407 p->prev_utime = max(p->prev_utime, utime);
3408 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3410 *ut = p->prev_utime;
3411 *st = p->prev_stime;
3415 * Must be called with siglock held.
3417 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3419 struct signal_struct *sig = p->signal;
3420 struct task_cputime cputime;
3421 cputime_t rtime, utime, total;
3423 thread_group_cputime(p, &cputime);
3425 total = cputime_add(cputime.utime, cputime.stime);
3426 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3431 temp = (u64)(rtime * cputime.utime);
3432 do_div(temp, total);
3433 utime = (cputime_t)temp;
3437 sig->prev_utime = max(sig->prev_utime, utime);
3438 sig->prev_stime = max(sig->prev_stime,
3439 cputime_sub(rtime, sig->prev_utime));
3441 *ut = sig->prev_utime;
3442 *st = sig->prev_stime;
3447 * This function gets called by the timer code, with HZ frequency.
3448 * We call it with interrupts disabled.
3450 * It also gets called by the fork code, when changing the parent's
3453 void scheduler_tick(void)
3455 int cpu = smp_processor_id();
3456 struct rq *rq = cpu_rq(cpu);
3457 struct task_struct *curr = rq->curr;
3461 raw_spin_lock(&rq->lock);
3462 update_rq_clock(rq);
3463 update_cpu_load(rq);
3464 curr->sched_class->task_tick(rq, curr, 0);
3465 raw_spin_unlock(&rq->lock);
3467 perf_event_task_tick(curr);
3470 rq->idle_at_tick = idle_cpu(cpu);
3471 trigger_load_balance(rq, cpu);
3475 notrace unsigned long get_parent_ip(unsigned long addr)
3477 if (in_lock_functions(addr)) {
3478 addr = CALLER_ADDR2;
3479 if (in_lock_functions(addr))
3480 addr = CALLER_ADDR3;
3485 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3486 defined(CONFIG_PREEMPT_TRACER))
3488 void __kprobes add_preempt_count(int val)
3490 #ifdef CONFIG_DEBUG_PREEMPT
3494 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3497 preempt_count() += val;
3498 #ifdef CONFIG_DEBUG_PREEMPT
3500 * Spinlock count overflowing soon?
3502 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3505 if (preempt_count() == val)
3506 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3508 EXPORT_SYMBOL(add_preempt_count);
3510 void __kprobes sub_preempt_count(int val)
3512 #ifdef CONFIG_DEBUG_PREEMPT
3516 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3519 * Is the spinlock portion underflowing?
3521 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3522 !(preempt_count() & PREEMPT_MASK)))
3526 if (preempt_count() == val)
3527 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3528 preempt_count() -= val;
3530 EXPORT_SYMBOL(sub_preempt_count);
3535 * Print scheduling while atomic bug:
3537 static noinline void __schedule_bug(struct task_struct *prev)
3539 struct pt_regs *regs = get_irq_regs();
3541 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3542 prev->comm, prev->pid, preempt_count());
3544 debug_show_held_locks(prev);
3546 if (irqs_disabled())
3547 print_irqtrace_events(prev);
3556 * Various schedule()-time debugging checks and statistics:
3558 static inline void schedule_debug(struct task_struct *prev)
3561 * Test if we are atomic. Since do_exit() needs to call into
3562 * schedule() atomically, we ignore that path for now.
3563 * Otherwise, whine if we are scheduling when we should not be.
3565 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3566 __schedule_bug(prev);
3568 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3570 schedstat_inc(this_rq(), sched_count);
3571 #ifdef CONFIG_SCHEDSTATS
3572 if (unlikely(prev->lock_depth >= 0)) {
3573 schedstat_inc(this_rq(), bkl_count);
3574 schedstat_inc(prev, sched_info.bkl_count);
3579 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3581 prev->sched_class->put_prev_task(rq, prev);
3585 * Pick up the highest-prio task:
3587 static inline struct task_struct *
3588 pick_next_task(struct rq *rq)
3590 const struct sched_class *class;
3591 struct task_struct *p;
3594 * Optimization: we know that if all tasks are in
3595 * the fair class we can call that function directly:
3597 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3598 p = fair_sched_class.pick_next_task(rq);
3603 class = sched_class_highest;
3605 p = class->pick_next_task(rq);
3609 * Will never be NULL as the idle class always
3610 * returns a non-NULL p:
3612 class = class->next;
3617 * schedule() is the main scheduler function.
3619 asmlinkage void __sched schedule(void)
3621 struct task_struct *prev, *next;
3622 unsigned long *switch_count;
3628 cpu = smp_processor_id();
3632 switch_count = &prev->nivcsw;
3634 release_kernel_lock(prev);
3635 need_resched_nonpreemptible:
3637 schedule_debug(prev);
3639 if (sched_feat(HRTICK))
3642 raw_spin_lock_irq(&rq->lock);
3643 update_rq_clock(rq);
3644 clear_tsk_need_resched(prev);
3646 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3647 if (unlikely(signal_pending_state(prev->state, prev)))
3648 prev->state = TASK_RUNNING;
3650 deactivate_task(rq, prev, 1);
3651 switch_count = &prev->nvcsw;
3654 pre_schedule(rq, prev);
3656 if (unlikely(!rq->nr_running))
3657 idle_balance(cpu, rq);
3659 put_prev_task(rq, prev);
3660 next = pick_next_task(rq);
3662 if (likely(prev != next)) {
3663 sched_info_switch(prev, next);
3664 perf_event_task_sched_out(prev, next);
3670 context_switch(rq, prev, next); /* unlocks the rq */
3672 * the context switch might have flipped the stack from under
3673 * us, hence refresh the local variables.
3675 cpu = smp_processor_id();
3678 raw_spin_unlock_irq(&rq->lock);
3682 if (unlikely(reacquire_kernel_lock(current) < 0)) {
3684 switch_count = &prev->nivcsw;
3685 goto need_resched_nonpreemptible;
3688 preempt_enable_no_resched();
3692 EXPORT_SYMBOL(schedule);
3694 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3696 * Look out! "owner" is an entirely speculative pointer
3697 * access and not reliable.
3699 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3704 if (!sched_feat(OWNER_SPIN))
3707 #ifdef CONFIG_DEBUG_PAGEALLOC
3709 * Need to access the cpu field knowing that
3710 * DEBUG_PAGEALLOC could have unmapped it if
3711 * the mutex owner just released it and exited.
3713 if (probe_kernel_address(&owner->cpu, cpu))
3720 * Even if the access succeeded (likely case),
3721 * the cpu field may no longer be valid.
3723 if (cpu >= nr_cpumask_bits)
3727 * We need to validate that we can do a
3728 * get_cpu() and that we have the percpu area.
3730 if (!cpu_online(cpu))
3737 * Owner changed, break to re-assess state.
3739 if (lock->owner != owner)
3743 * Is that owner really running on that cpu?
3745 if (task_thread_info(rq->curr) != owner || need_resched())
3755 #ifdef CONFIG_PREEMPT
3757 * this is the entry point to schedule() from in-kernel preemption
3758 * off of preempt_enable. Kernel preemptions off return from interrupt
3759 * occur there and call schedule directly.
3761 asmlinkage void __sched preempt_schedule(void)
3763 struct thread_info *ti = current_thread_info();
3766 * If there is a non-zero preempt_count or interrupts are disabled,
3767 * we do not want to preempt the current task. Just return..
3769 if (likely(ti->preempt_count || irqs_disabled()))
3773 add_preempt_count(PREEMPT_ACTIVE);
3775 sub_preempt_count(PREEMPT_ACTIVE);
3778 * Check again in case we missed a preemption opportunity
3779 * between schedule and now.
3782 } while (need_resched());
3784 EXPORT_SYMBOL(preempt_schedule);
3787 * this is the entry point to schedule() from kernel preemption
3788 * off of irq context.
3789 * Note, that this is called and return with irqs disabled. This will
3790 * protect us against recursive calling from irq.
3792 asmlinkage void __sched preempt_schedule_irq(void)
3794 struct thread_info *ti = current_thread_info();
3796 /* Catch callers which need to be fixed */
3797 BUG_ON(ti->preempt_count || !irqs_disabled());
3800 add_preempt_count(PREEMPT_ACTIVE);
3803 local_irq_disable();
3804 sub_preempt_count(PREEMPT_ACTIVE);
3807 * Check again in case we missed a preemption opportunity
3808 * between schedule and now.
3811 } while (need_resched());
3814 #endif /* CONFIG_PREEMPT */
3816 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3819 return try_to_wake_up(curr->private, mode, wake_flags);
3821 EXPORT_SYMBOL(default_wake_function);
3824 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3825 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3826 * number) then we wake all the non-exclusive tasks and one exclusive task.
3828 * There are circumstances in which we can try to wake a task which has already
3829 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3830 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3832 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3833 int nr_exclusive, int wake_flags, void *key)
3835 wait_queue_t *curr, *next;
3837 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3838 unsigned flags = curr->flags;
3840 if (curr->func(curr, mode, wake_flags, key) &&
3841 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3847 * __wake_up - wake up threads blocked on a waitqueue.
3849 * @mode: which threads
3850 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3851 * @key: is directly passed to the wakeup function
3853 * It may be assumed that this function implies a write memory barrier before
3854 * changing the task state if and only if any tasks are woken up.
3856 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3857 int nr_exclusive, void *key)
3859 unsigned long flags;
3861 spin_lock_irqsave(&q->lock, flags);
3862 __wake_up_common(q, mode, nr_exclusive, 0, key);
3863 spin_unlock_irqrestore(&q->lock, flags);
3865 EXPORT_SYMBOL(__wake_up);
3868 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3870 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3872 __wake_up_common(q, mode, 1, 0, NULL);
3875 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3877 __wake_up_common(q, mode, 1, 0, key);
3881 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3883 * @mode: which threads
3884 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3885 * @key: opaque value to be passed to wakeup targets
3887 * The sync wakeup differs that the waker knows that it will schedule
3888 * away soon, so while the target thread will be woken up, it will not
3889 * be migrated to another CPU - ie. the two threads are 'synchronized'
3890 * with each other. This can prevent needless bouncing between CPUs.
3892 * On UP it can prevent extra preemption.
3894 * It may be assumed that this function implies a write memory barrier before
3895 * changing the task state if and only if any tasks are woken up.
3897 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
3898 int nr_exclusive, void *key)
3900 unsigned long flags;
3901 int wake_flags = WF_SYNC;
3906 if (unlikely(!nr_exclusive))
3909 spin_lock_irqsave(&q->lock, flags);
3910 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
3911 spin_unlock_irqrestore(&q->lock, flags);
3913 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
3916 * __wake_up_sync - see __wake_up_sync_key()
3918 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
3920 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
3922 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
3925 * complete: - signals a single thread waiting on this completion
3926 * @x: holds the state of this particular completion
3928 * This will wake up a single thread waiting on this completion. Threads will be
3929 * awakened in the same order in which they were queued.
3931 * See also complete_all(), wait_for_completion() and related routines.
3933 * It may be assumed that this function implies a write memory barrier before
3934 * changing the task state if and only if any tasks are woken up.
3936 void complete(struct completion *x)
3938 unsigned long flags;
3940 spin_lock_irqsave(&x->wait.lock, flags);
3942 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
3943 spin_unlock_irqrestore(&x->wait.lock, flags);
3945 EXPORT_SYMBOL(complete);
3948 * complete_all: - signals all threads waiting on this completion
3949 * @x: holds the state of this particular completion
3951 * This will wake up all threads waiting on this particular completion event.
3953 * It may be assumed that this function implies a write memory barrier before
3954 * changing the task state if and only if any tasks are woken up.
3956 void complete_all(struct completion *x)
3958 unsigned long flags;
3960 spin_lock_irqsave(&x->wait.lock, flags);
3961 x->done += UINT_MAX/2;
3962 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
3963 spin_unlock_irqrestore(&x->wait.lock, flags);
3965 EXPORT_SYMBOL(complete_all);
3967 static inline long __sched
3968 do_wait_for_common(struct completion *x, long timeout, int state)
3971 DECLARE_WAITQUEUE(wait, current);
3973 wait.flags |= WQ_FLAG_EXCLUSIVE;
3974 __add_wait_queue_tail(&x->wait, &wait);
3976 if (signal_pending_state(state, current)) {
3977 timeout = -ERESTARTSYS;
3980 __set_current_state(state);
3981 spin_unlock_irq(&x->wait.lock);
3982 timeout = schedule_timeout(timeout);
3983 spin_lock_irq(&x->wait.lock);
3984 } while (!x->done && timeout);
3985 __remove_wait_queue(&x->wait, &wait);
3990 return timeout ?: 1;
3994 wait_for_common(struct completion *x, long timeout, int state)
3998 spin_lock_irq(&x->wait.lock);
3999 timeout = do_wait_for_common(x, timeout, state);
4000 spin_unlock_irq(&x->wait.lock);
4005 * wait_for_completion: - waits for completion of a task
4006 * @x: holds the state of this particular completion
4008 * This waits to be signaled for completion of a specific task. It is NOT
4009 * interruptible and there is no timeout.
4011 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4012 * and interrupt capability. Also see complete().
4014 void __sched wait_for_completion(struct completion *x)
4016 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4018 EXPORT_SYMBOL(wait_for_completion);
4021 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4022 * @x: holds the state of this particular completion
4023 * @timeout: timeout value in jiffies
4025 * This waits for either a completion of a specific task to be signaled or for a
4026 * specified timeout to expire. The timeout is in jiffies. It is not
4029 unsigned long __sched
4030 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4032 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4034 EXPORT_SYMBOL(wait_for_completion_timeout);
4037 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4038 * @x: holds the state of this particular completion
4040 * This waits for completion of a specific task to be signaled. It is
4043 int __sched wait_for_completion_interruptible(struct completion *x)
4045 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4046 if (t == -ERESTARTSYS)
4050 EXPORT_SYMBOL(wait_for_completion_interruptible);
4053 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4054 * @x: holds the state of this particular completion
4055 * @timeout: timeout value in jiffies
4057 * This waits for either a completion of a specific task to be signaled or for a
4058 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4060 unsigned long __sched
4061 wait_for_completion_interruptible_timeout(struct completion *x,
4062 unsigned long timeout)
4064 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4066 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4069 * wait_for_completion_killable: - waits for completion of a task (killable)
4070 * @x: holds the state of this particular completion
4072 * This waits to be signaled for completion of a specific task. It can be
4073 * interrupted by a kill signal.
4075 int __sched wait_for_completion_killable(struct completion *x)
4077 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4078 if (t == -ERESTARTSYS)
4082 EXPORT_SYMBOL(wait_for_completion_killable);
4085 * try_wait_for_completion - try to decrement a completion without blocking
4086 * @x: completion structure
4088 * Returns: 0 if a decrement cannot be done without blocking
4089 * 1 if a decrement succeeded.
4091 * If a completion is being used as a counting completion,
4092 * attempt to decrement the counter without blocking. This
4093 * enables us to avoid waiting if the resource the completion
4094 * is protecting is not available.
4096 bool try_wait_for_completion(struct completion *x)
4098 unsigned long flags;
4101 spin_lock_irqsave(&x->wait.lock, flags);
4106 spin_unlock_irqrestore(&x->wait.lock, flags);
4109 EXPORT_SYMBOL(try_wait_for_completion);
4112 * completion_done - Test to see if a completion has any waiters
4113 * @x: completion structure
4115 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4116 * 1 if there are no waiters.
4119 bool completion_done(struct completion *x)
4121 unsigned long flags;
4124 spin_lock_irqsave(&x->wait.lock, flags);
4127 spin_unlock_irqrestore(&x->wait.lock, flags);
4130 EXPORT_SYMBOL(completion_done);
4133 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4135 unsigned long flags;
4138 init_waitqueue_entry(&wait, current);
4140 __set_current_state(state);
4142 spin_lock_irqsave(&q->lock, flags);
4143 __add_wait_queue(q, &wait);
4144 spin_unlock(&q->lock);
4145 timeout = schedule_timeout(timeout);
4146 spin_lock_irq(&q->lock);
4147 __remove_wait_queue(q, &wait);
4148 spin_unlock_irqrestore(&q->lock, flags);
4153 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4155 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4157 EXPORT_SYMBOL(interruptible_sleep_on);
4160 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4162 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4164 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4166 void __sched sleep_on(wait_queue_head_t *q)
4168 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4170 EXPORT_SYMBOL(sleep_on);
4172 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4174 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4176 EXPORT_SYMBOL(sleep_on_timeout);
4178 #ifdef CONFIG_RT_MUTEXES
4181 * rt_mutex_setprio - set the current priority of a task
4183 * @prio: prio value (kernel-internal form)
4185 * This function changes the 'effective' priority of a task. It does
4186 * not touch ->normal_prio like __setscheduler().
4188 * Used by the rt_mutex code to implement priority inheritance logic.
4190 void rt_mutex_setprio(struct task_struct *p, int prio)
4192 unsigned long flags;
4193 int oldprio, on_rq, running;
4195 const struct sched_class *prev_class;
4197 BUG_ON(prio < 0 || prio > MAX_PRIO);
4199 rq = task_rq_lock(p, &flags);
4200 update_rq_clock(rq);
4203 prev_class = p->sched_class;
4204 on_rq = p->se.on_rq;
4205 running = task_current(rq, p);
4207 dequeue_task(rq, p, 0);
4209 p->sched_class->put_prev_task(rq, p);
4212 p->sched_class = &rt_sched_class;
4214 p->sched_class = &fair_sched_class;
4219 p->sched_class->set_curr_task(rq);
4221 enqueue_task(rq, p, 0, oldprio < prio);
4223 check_class_changed(rq, p, prev_class, oldprio, running);
4225 task_rq_unlock(rq, &flags);
4230 void set_user_nice(struct task_struct *p, long nice)
4232 int old_prio, delta, on_rq;
4233 unsigned long flags;
4236 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4239 * We have to be careful, if called from sys_setpriority(),
4240 * the task might be in the middle of scheduling on another CPU.
4242 rq = task_rq_lock(p, &flags);
4243 update_rq_clock(rq);
4245 * The RT priorities are set via sched_setscheduler(), but we still
4246 * allow the 'normal' nice value to be set - but as expected
4247 * it wont have any effect on scheduling until the task is
4248 * SCHED_FIFO/SCHED_RR:
4250 if (task_has_rt_policy(p)) {
4251 p->static_prio = NICE_TO_PRIO(nice);
4254 on_rq = p->se.on_rq;
4256 dequeue_task(rq, p, 0);
4258 p->static_prio = NICE_TO_PRIO(nice);
4261 p->prio = effective_prio(p);
4262 delta = p->prio - old_prio;
4265 enqueue_task(rq, p, 0, false);
4267 * If the task increased its priority or is running and
4268 * lowered its priority, then reschedule its CPU:
4270 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4271 resched_task(rq->curr);
4274 task_rq_unlock(rq, &flags);
4276 EXPORT_SYMBOL(set_user_nice);
4279 * can_nice - check if a task can reduce its nice value
4283 int can_nice(const struct task_struct *p, const int nice)
4285 /* convert nice value [19,-20] to rlimit style value [1,40] */
4286 int nice_rlim = 20 - nice;
4288 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4289 capable(CAP_SYS_NICE));
4292 #ifdef __ARCH_WANT_SYS_NICE
4295 * sys_nice - change the priority of the current process.
4296 * @increment: priority increment
4298 * sys_setpriority is a more generic, but much slower function that
4299 * does similar things.
4301 SYSCALL_DEFINE1(nice, int, increment)
4306 * Setpriority might change our priority at the same moment.
4307 * We don't have to worry. Conceptually one call occurs first
4308 * and we have a single winner.
4310 if (increment < -40)
4315 nice = TASK_NICE(current) + increment;
4321 if (increment < 0 && !can_nice(current, nice))
4324 retval = security_task_setnice(current, nice);
4328 set_user_nice(current, nice);
4335 * task_prio - return the priority value of a given task.
4336 * @p: the task in question.
4338 * This is the priority value as seen by users in /proc.
4339 * RT tasks are offset by -200. Normal tasks are centered
4340 * around 0, value goes from -16 to +15.
4342 int task_prio(const struct task_struct *p)
4344 return p->prio - MAX_RT_PRIO;
4348 * task_nice - return the nice value of a given task.
4349 * @p: the task in question.
4351 int task_nice(const struct task_struct *p)
4353 return TASK_NICE(p);
4355 EXPORT_SYMBOL(task_nice);
4358 * idle_cpu - is a given cpu idle currently?
4359 * @cpu: the processor in question.
4361 int idle_cpu(int cpu)
4363 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4367 * idle_task - return the idle task for a given cpu.
4368 * @cpu: the processor in question.
4370 struct task_struct *idle_task(int cpu)
4372 return cpu_rq(cpu)->idle;
4376 * find_process_by_pid - find a process with a matching PID value.
4377 * @pid: the pid in question.
4379 static struct task_struct *find_process_by_pid(pid_t pid)
4381 return pid ? find_task_by_vpid(pid) : current;
4384 /* Actually do priority change: must hold rq lock. */
4386 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4388 BUG_ON(p->se.on_rq);
4391 p->rt_priority = prio;
4392 p->normal_prio = normal_prio(p);
4393 /* we are holding p->pi_lock already */
4394 p->prio = rt_mutex_getprio(p);
4395 if (rt_prio(p->prio))
4396 p->sched_class = &rt_sched_class;
4398 p->sched_class = &fair_sched_class;
4403 * check the target process has a UID that matches the current process's
4405 static bool check_same_owner(struct task_struct *p)
4407 const struct cred *cred = current_cred(), *pcred;
4411 pcred = __task_cred(p);
4412 match = (cred->euid == pcred->euid ||
4413 cred->euid == pcred->uid);
4418 static int __sched_setscheduler(struct task_struct *p, int policy,
4419 struct sched_param *param, bool user)
4421 int retval, oldprio, oldpolicy = -1, on_rq, running;
4422 unsigned long flags;
4423 const struct sched_class *prev_class;
4427 /* may grab non-irq protected spin_locks */
4428 BUG_ON(in_interrupt());
4430 /* double check policy once rq lock held */
4432 reset_on_fork = p->sched_reset_on_fork;
4433 policy = oldpolicy = p->policy;
4435 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4436 policy &= ~SCHED_RESET_ON_FORK;
4438 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4439 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4440 policy != SCHED_IDLE)
4445 * Valid priorities for SCHED_FIFO and SCHED_RR are
4446 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4447 * SCHED_BATCH and SCHED_IDLE is 0.
4449 if (param->sched_priority < 0 ||
4450 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4451 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4453 if (rt_policy(policy) != (param->sched_priority != 0))
4457 * Allow unprivileged RT tasks to decrease priority:
4459 if (user && !capable(CAP_SYS_NICE)) {
4460 if (rt_policy(policy)) {
4461 unsigned long rlim_rtprio;
4463 if (!lock_task_sighand(p, &flags))
4465 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
4466 unlock_task_sighand(p, &flags);
4468 /* can't set/change the rt policy */
4469 if (policy != p->policy && !rlim_rtprio)
4472 /* can't increase priority */
4473 if (param->sched_priority > p->rt_priority &&
4474 param->sched_priority > rlim_rtprio)
4478 * Like positive nice levels, dont allow tasks to
4479 * move out of SCHED_IDLE either:
4481 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4484 /* can't change other user's priorities */
4485 if (!check_same_owner(p))
4488 /* Normal users shall not reset the sched_reset_on_fork flag */
4489 if (p->sched_reset_on_fork && !reset_on_fork)
4494 #ifdef CONFIG_RT_GROUP_SCHED
4496 * Do not allow realtime tasks into groups that have no runtime
4499 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4500 task_group(p)->rt_bandwidth.rt_runtime == 0)
4504 retval = security_task_setscheduler(p, policy, param);
4510 * make sure no PI-waiters arrive (or leave) while we are
4511 * changing the priority of the task:
4513 raw_spin_lock_irqsave(&p->pi_lock, flags);
4515 * To be able to change p->policy safely, the apropriate
4516 * runqueue lock must be held.
4518 rq = __task_rq_lock(p);
4519 /* recheck policy now with rq lock held */
4520 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4521 policy = oldpolicy = -1;
4522 __task_rq_unlock(rq);
4523 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4526 update_rq_clock(rq);
4527 on_rq = p->se.on_rq;
4528 running = task_current(rq, p);
4530 deactivate_task(rq, p, 0);
4532 p->sched_class->put_prev_task(rq, p);
4534 p->sched_reset_on_fork = reset_on_fork;
4537 prev_class = p->sched_class;
4538 __setscheduler(rq, p, policy, param->sched_priority);
4541 p->sched_class->set_curr_task(rq);
4543 activate_task(rq, p, 0);
4545 check_class_changed(rq, p, prev_class, oldprio, running);
4547 __task_rq_unlock(rq);
4548 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4550 rt_mutex_adjust_pi(p);
4556 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4557 * @p: the task in question.
4558 * @policy: new policy.
4559 * @param: structure containing the new RT priority.
4561 * NOTE that the task may be already dead.
4563 int sched_setscheduler(struct task_struct *p, int policy,
4564 struct sched_param *param)
4566 return __sched_setscheduler(p, policy, param, true);
4568 EXPORT_SYMBOL_GPL(sched_setscheduler);
4571 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4572 * @p: the task in question.
4573 * @policy: new policy.
4574 * @param: structure containing the new RT priority.
4576 * Just like sched_setscheduler, only don't bother checking if the
4577 * current context has permission. For example, this is needed in
4578 * stop_machine(): we create temporary high priority worker threads,
4579 * but our caller might not have that capability.
4581 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4582 struct sched_param *param)
4584 return __sched_setscheduler(p, policy, param, false);
4588 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4590 struct sched_param lparam;
4591 struct task_struct *p;
4594 if (!param || pid < 0)
4596 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4601 p = find_process_by_pid(pid);
4603 retval = sched_setscheduler(p, policy, &lparam);
4610 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4611 * @pid: the pid in question.
4612 * @policy: new policy.
4613 * @param: structure containing the new RT priority.
4615 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4616 struct sched_param __user *, param)
4618 /* negative values for policy are not valid */
4622 return do_sched_setscheduler(pid, policy, param);
4626 * sys_sched_setparam - set/change the RT priority of a thread
4627 * @pid: the pid in question.
4628 * @param: structure containing the new RT priority.
4630 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4632 return do_sched_setscheduler(pid, -1, param);
4636 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4637 * @pid: the pid in question.
4639 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4641 struct task_struct *p;
4649 p = find_process_by_pid(pid);
4651 retval = security_task_getscheduler(p);
4654 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4661 * sys_sched_getparam - get the RT priority of a thread
4662 * @pid: the pid in question.
4663 * @param: structure containing the RT priority.
4665 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4667 struct sched_param lp;
4668 struct task_struct *p;
4671 if (!param || pid < 0)
4675 p = find_process_by_pid(pid);
4680 retval = security_task_getscheduler(p);
4684 lp.sched_priority = p->rt_priority;
4688 * This one might sleep, we cannot do it with a spinlock held ...
4690 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4699 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4701 cpumask_var_t cpus_allowed, new_mask;
4702 struct task_struct *p;
4708 p = find_process_by_pid(pid);
4715 /* Prevent p going away */
4719 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4723 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4725 goto out_free_cpus_allowed;
4728 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4731 retval = security_task_setscheduler(p, 0, NULL);
4735 cpuset_cpus_allowed(p, cpus_allowed);
4736 cpumask_and(new_mask, in_mask, cpus_allowed);
4738 retval = set_cpus_allowed_ptr(p, new_mask);
4741 cpuset_cpus_allowed(p, cpus_allowed);
4742 if (!cpumask_subset(new_mask, cpus_allowed)) {
4744 * We must have raced with a concurrent cpuset
4745 * update. Just reset the cpus_allowed to the
4746 * cpuset's cpus_allowed
4748 cpumask_copy(new_mask, cpus_allowed);
4753 free_cpumask_var(new_mask);
4754 out_free_cpus_allowed:
4755 free_cpumask_var(cpus_allowed);
4762 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4763 struct cpumask *new_mask)
4765 if (len < cpumask_size())
4766 cpumask_clear(new_mask);
4767 else if (len > cpumask_size())
4768 len = cpumask_size();
4770 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4774 * sys_sched_setaffinity - set the cpu affinity of a process
4775 * @pid: pid of the process
4776 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4777 * @user_mask_ptr: user-space pointer to the new cpu mask
4779 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4780 unsigned long __user *, user_mask_ptr)
4782 cpumask_var_t new_mask;
4785 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4788 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4790 retval = sched_setaffinity(pid, new_mask);
4791 free_cpumask_var(new_mask);
4795 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4797 struct task_struct *p;
4798 unsigned long flags;
4806 p = find_process_by_pid(pid);
4810 retval = security_task_getscheduler(p);
4814 rq = task_rq_lock(p, &flags);
4815 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4816 task_rq_unlock(rq, &flags);
4826 * sys_sched_getaffinity - get the cpu affinity of a process
4827 * @pid: pid of the process
4828 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4829 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4831 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4832 unsigned long __user *, user_mask_ptr)
4837 if (len < cpumask_size())
4840 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4843 ret = sched_getaffinity(pid, mask);
4845 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
4848 ret = cpumask_size();
4850 free_cpumask_var(mask);
4856 * sys_sched_yield - yield the current processor to other threads.
4858 * This function yields the current CPU to other tasks. If there are no
4859 * other threads running on this CPU then this function will return.
4861 SYSCALL_DEFINE0(sched_yield)
4863 struct rq *rq = this_rq_lock();
4865 schedstat_inc(rq, yld_count);
4866 current->sched_class->yield_task(rq);
4869 * Since we are going to call schedule() anyway, there's
4870 * no need to preempt or enable interrupts:
4872 __release(rq->lock);
4873 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
4874 do_raw_spin_unlock(&rq->lock);
4875 preempt_enable_no_resched();
4882 static inline int should_resched(void)
4884 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
4887 static void __cond_resched(void)
4889 add_preempt_count(PREEMPT_ACTIVE);
4891 sub_preempt_count(PREEMPT_ACTIVE);
4894 int __sched _cond_resched(void)
4896 if (should_resched()) {
4902 EXPORT_SYMBOL(_cond_resched);
4905 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4906 * call schedule, and on return reacquire the lock.
4908 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4909 * operations here to prevent schedule() from being called twice (once via
4910 * spin_unlock(), once by hand).
4912 int __cond_resched_lock(spinlock_t *lock)
4914 int resched = should_resched();
4917 lockdep_assert_held(lock);
4919 if (spin_needbreak(lock) || resched) {
4930 EXPORT_SYMBOL(__cond_resched_lock);
4932 int __sched __cond_resched_softirq(void)
4934 BUG_ON(!in_softirq());
4936 if (should_resched()) {
4944 EXPORT_SYMBOL(__cond_resched_softirq);
4947 * yield - yield the current processor to other threads.
4949 * This is a shortcut for kernel-space yielding - it marks the
4950 * thread runnable and calls sys_sched_yield().
4952 void __sched yield(void)
4954 set_current_state(TASK_RUNNING);
4957 EXPORT_SYMBOL(yield);
4960 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4961 * that process accounting knows that this is a task in IO wait state.
4963 void __sched io_schedule(void)
4965 struct rq *rq = raw_rq();
4967 delayacct_blkio_start();
4968 atomic_inc(&rq->nr_iowait);
4969 current->in_iowait = 1;
4971 current->in_iowait = 0;
4972 atomic_dec(&rq->nr_iowait);
4973 delayacct_blkio_end();
4975 EXPORT_SYMBOL(io_schedule);
4977 long __sched io_schedule_timeout(long timeout)
4979 struct rq *rq = raw_rq();
4982 delayacct_blkio_start();
4983 atomic_inc(&rq->nr_iowait);
4984 current->in_iowait = 1;
4985 ret = schedule_timeout(timeout);
4986 current->in_iowait = 0;
4987 atomic_dec(&rq->nr_iowait);
4988 delayacct_blkio_end();
4993 * sys_sched_get_priority_max - return maximum RT priority.
4994 * @policy: scheduling class.
4996 * this syscall returns the maximum rt_priority that can be used
4997 * by a given scheduling class.
4999 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5006 ret = MAX_USER_RT_PRIO-1;
5018 * sys_sched_get_priority_min - return minimum RT priority.
5019 * @policy: scheduling class.
5021 * this syscall returns the minimum rt_priority that can be used
5022 * by a given scheduling class.
5024 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5042 * sys_sched_rr_get_interval - return the default timeslice of a process.
5043 * @pid: pid of the process.
5044 * @interval: userspace pointer to the timeslice value.
5046 * this syscall writes the default timeslice value of a given process
5047 * into the user-space timespec buffer. A value of '0' means infinity.
5049 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5050 struct timespec __user *, interval)
5052 struct task_struct *p;
5053 unsigned int time_slice;
5054 unsigned long flags;
5064 p = find_process_by_pid(pid);
5068 retval = security_task_getscheduler(p);
5072 rq = task_rq_lock(p, &flags);
5073 time_slice = p->sched_class->get_rr_interval(rq, p);
5074 task_rq_unlock(rq, &flags);
5077 jiffies_to_timespec(time_slice, &t);
5078 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5086 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5088 void sched_show_task(struct task_struct *p)
5090 unsigned long free = 0;
5093 state = p->state ? __ffs(p->state) + 1 : 0;
5094 printk(KERN_INFO "%-13.13s %c", p->comm,
5095 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5096 #if BITS_PER_LONG == 32
5097 if (state == TASK_RUNNING)
5098 printk(KERN_CONT " running ");
5100 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5102 if (state == TASK_RUNNING)
5103 printk(KERN_CONT " running task ");
5105 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5107 #ifdef CONFIG_DEBUG_STACK_USAGE
5108 free = stack_not_used(p);
5110 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5111 task_pid_nr(p), task_pid_nr(p->real_parent),
5112 (unsigned long)task_thread_info(p)->flags);
5114 show_stack(p, NULL);
5117 void show_state_filter(unsigned long state_filter)
5119 struct task_struct *g, *p;
5121 #if BITS_PER_LONG == 32
5123 " task PC stack pid father\n");
5126 " task PC stack pid father\n");
5128 read_lock(&tasklist_lock);
5129 do_each_thread(g, p) {
5131 * reset the NMI-timeout, listing all files on a slow
5132 * console might take alot of time:
5134 touch_nmi_watchdog();
5135 if (!state_filter || (p->state & state_filter))
5137 } while_each_thread(g, p);
5139 touch_all_softlockup_watchdogs();
5141 #ifdef CONFIG_SCHED_DEBUG
5142 sysrq_sched_debug_show();
5144 read_unlock(&tasklist_lock);
5146 * Only show locks if all tasks are dumped:
5149 debug_show_all_locks();
5152 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5154 idle->sched_class = &idle_sched_class;
5158 * init_idle - set up an idle thread for a given CPU
5159 * @idle: task in question
5160 * @cpu: cpu the idle task belongs to
5162 * NOTE: this function does not set the idle thread's NEED_RESCHED
5163 * flag, to make booting more robust.
5165 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5167 struct rq *rq = cpu_rq(cpu);
5168 unsigned long flags;
5170 raw_spin_lock_irqsave(&rq->lock, flags);
5173 idle->state = TASK_RUNNING;
5174 idle->se.exec_start = sched_clock();
5176 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5177 __set_task_cpu(idle, cpu);
5179 rq->curr = rq->idle = idle;
5180 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5183 raw_spin_unlock_irqrestore(&rq->lock, flags);
5185 /* Set the preempt count _outside_ the spinlocks! */
5186 #if defined(CONFIG_PREEMPT)
5187 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5189 task_thread_info(idle)->preempt_count = 0;
5192 * The idle tasks have their own, simple scheduling class:
5194 idle->sched_class = &idle_sched_class;
5195 ftrace_graph_init_task(idle);
5199 * In a system that switches off the HZ timer nohz_cpu_mask
5200 * indicates which cpus entered this state. This is used
5201 * in the rcu update to wait only for active cpus. For system
5202 * which do not switch off the HZ timer nohz_cpu_mask should
5203 * always be CPU_BITS_NONE.
5205 cpumask_var_t nohz_cpu_mask;
5208 * Increase the granularity value when there are more CPUs,
5209 * because with more CPUs the 'effective latency' as visible
5210 * to users decreases. But the relationship is not linear,
5211 * so pick a second-best guess by going with the log2 of the
5214 * This idea comes from the SD scheduler of Con Kolivas:
5216 static int get_update_sysctl_factor(void)
5218 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5219 unsigned int factor;
5221 switch (sysctl_sched_tunable_scaling) {
5222 case SCHED_TUNABLESCALING_NONE:
5225 case SCHED_TUNABLESCALING_LINEAR:
5228 case SCHED_TUNABLESCALING_LOG:
5230 factor = 1 + ilog2(cpus);
5237 static void update_sysctl(void)
5239 unsigned int factor = get_update_sysctl_factor();
5241 #define SET_SYSCTL(name) \
5242 (sysctl_##name = (factor) * normalized_sysctl_##name)
5243 SET_SYSCTL(sched_min_granularity);
5244 SET_SYSCTL(sched_latency);
5245 SET_SYSCTL(sched_wakeup_granularity);
5246 SET_SYSCTL(sched_shares_ratelimit);
5250 static inline void sched_init_granularity(void)
5257 * This is how migration works:
5259 * 1) we queue a struct migration_req structure in the source CPU's
5260 * runqueue and wake up that CPU's migration thread.
5261 * 2) we down() the locked semaphore => thread blocks.
5262 * 3) migration thread wakes up (implicitly it forces the migrated
5263 * thread off the CPU)
5264 * 4) it gets the migration request and checks whether the migrated
5265 * task is still in the wrong runqueue.
5266 * 5) if it's in the wrong runqueue then the migration thread removes
5267 * it and puts it into the right queue.
5268 * 6) migration thread up()s the semaphore.
5269 * 7) we wake up and the migration is done.
5273 * Change a given task's CPU affinity. Migrate the thread to a
5274 * proper CPU and schedule it away if the CPU it's executing on
5275 * is removed from the allowed bitmask.
5277 * NOTE: the caller must have a valid reference to the task, the
5278 * task must not exit() & deallocate itself prematurely. The
5279 * call is not atomic; no spinlocks may be held.
5281 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5283 struct migration_req req;
5284 unsigned long flags;
5288 rq = task_rq_lock(p, &flags);
5290 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5295 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5296 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5301 if (p->sched_class->set_cpus_allowed)
5302 p->sched_class->set_cpus_allowed(p, new_mask);
5304 cpumask_copy(&p->cpus_allowed, new_mask);
5305 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5308 /* Can the task run on the task's current CPU? If so, we're done */
5309 if (cpumask_test_cpu(task_cpu(p), new_mask))
5312 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
5313 /* Need help from migration thread: drop lock and wait. */
5314 struct task_struct *mt = rq->migration_thread;
5316 get_task_struct(mt);
5317 task_rq_unlock(rq, &flags);
5318 wake_up_process(rq->migration_thread);
5319 put_task_struct(mt);
5320 wait_for_completion(&req.done);
5321 tlb_migrate_finish(p->mm);
5325 task_rq_unlock(rq, &flags);
5329 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5332 * Move (not current) task off this cpu, onto dest cpu. We're doing
5333 * this because either it can't run here any more (set_cpus_allowed()
5334 * away from this CPU, or CPU going down), or because we're
5335 * attempting to rebalance this task on exec (sched_exec).
5337 * So we race with normal scheduler movements, but that's OK, as long
5338 * as the task is no longer on this CPU.
5340 * Returns non-zero if task was successfully migrated.
5342 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5344 struct rq *rq_dest, *rq_src;
5347 if (unlikely(!cpu_active(dest_cpu)))
5350 rq_src = cpu_rq(src_cpu);
5351 rq_dest = cpu_rq(dest_cpu);
5353 double_rq_lock(rq_src, rq_dest);
5354 /* Already moved. */
5355 if (task_cpu(p) != src_cpu)
5357 /* Affinity changed (again). */
5358 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5362 * If we're not on a rq, the next wake-up will ensure we're
5366 deactivate_task(rq_src, p, 0);
5367 set_task_cpu(p, dest_cpu);
5368 activate_task(rq_dest, p, 0);
5369 check_preempt_curr(rq_dest, p, 0);
5374 double_rq_unlock(rq_src, rq_dest);
5378 #define RCU_MIGRATION_IDLE 0
5379 #define RCU_MIGRATION_NEED_QS 1
5380 #define RCU_MIGRATION_GOT_QS 2
5381 #define RCU_MIGRATION_MUST_SYNC 3
5384 * migration_thread - this is a highprio system thread that performs
5385 * thread migration by bumping thread off CPU then 'pushing' onto
5388 static int migration_thread(void *data)
5391 int cpu = (long)data;
5395 BUG_ON(rq->migration_thread != current);
5397 set_current_state(TASK_INTERRUPTIBLE);
5398 while (!kthread_should_stop()) {
5399 struct migration_req *req;
5400 struct list_head *head;
5402 raw_spin_lock_irq(&rq->lock);
5404 if (cpu_is_offline(cpu)) {
5405 raw_spin_unlock_irq(&rq->lock);
5409 if (rq->active_balance) {
5410 active_load_balance(rq, cpu);
5411 rq->active_balance = 0;
5414 head = &rq->migration_queue;
5416 if (list_empty(head)) {
5417 raw_spin_unlock_irq(&rq->lock);
5419 set_current_state(TASK_INTERRUPTIBLE);
5422 req = list_entry(head->next, struct migration_req, list);
5423 list_del_init(head->next);
5425 if (req->task != NULL) {
5426 raw_spin_unlock(&rq->lock);
5427 __migrate_task(req->task, cpu, req->dest_cpu);
5428 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
5429 req->dest_cpu = RCU_MIGRATION_GOT_QS;
5430 raw_spin_unlock(&rq->lock);
5432 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
5433 raw_spin_unlock(&rq->lock);
5434 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
5438 complete(&req->done);
5440 __set_current_state(TASK_RUNNING);
5445 #ifdef CONFIG_HOTPLUG_CPU
5447 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
5451 local_irq_disable();
5452 ret = __migrate_task(p, src_cpu, dest_cpu);
5458 * Figure out where task on dead CPU should go, use force if necessary.
5460 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5465 dest_cpu = select_fallback_rq(dead_cpu, p);
5467 /* It can have affinity changed while we were choosing. */
5468 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
5473 * While a dead CPU has no uninterruptible tasks queued at this point,
5474 * it might still have a nonzero ->nr_uninterruptible counter, because
5475 * for performance reasons the counter is not stricly tracking tasks to
5476 * their home CPUs. So we just add the counter to another CPU's counter,
5477 * to keep the global sum constant after CPU-down:
5479 static void migrate_nr_uninterruptible(struct rq *rq_src)
5481 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5482 unsigned long flags;
5484 local_irq_save(flags);
5485 double_rq_lock(rq_src, rq_dest);
5486 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5487 rq_src->nr_uninterruptible = 0;
5488 double_rq_unlock(rq_src, rq_dest);
5489 local_irq_restore(flags);
5492 /* Run through task list and migrate tasks from the dead cpu. */
5493 static void migrate_live_tasks(int src_cpu)
5495 struct task_struct *p, *t;
5497 read_lock(&tasklist_lock);
5499 do_each_thread(t, p) {
5503 if (task_cpu(p) == src_cpu)
5504 move_task_off_dead_cpu(src_cpu, p);
5505 } while_each_thread(t, p);
5507 read_unlock(&tasklist_lock);
5511 * Schedules idle task to be the next runnable task on current CPU.
5512 * It does so by boosting its priority to highest possible.
5513 * Used by CPU offline code.
5515 void sched_idle_next(void)
5517 int this_cpu = smp_processor_id();
5518 struct rq *rq = cpu_rq(this_cpu);
5519 struct task_struct *p = rq->idle;
5520 unsigned long flags;
5522 /* cpu has to be offline */
5523 BUG_ON(cpu_online(this_cpu));
5526 * Strictly not necessary since rest of the CPUs are stopped by now
5527 * and interrupts disabled on the current cpu.
5529 raw_spin_lock_irqsave(&rq->lock, flags);
5531 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5533 update_rq_clock(rq);
5534 activate_task(rq, p, 0);
5536 raw_spin_unlock_irqrestore(&rq->lock, flags);
5540 * Ensures that the idle task is using init_mm right before its cpu goes
5543 void idle_task_exit(void)
5545 struct mm_struct *mm = current->active_mm;
5547 BUG_ON(cpu_online(smp_processor_id()));
5550 switch_mm(mm, &init_mm, current);
5554 /* called under rq->lock with disabled interrupts */
5555 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5557 struct rq *rq = cpu_rq(dead_cpu);
5559 /* Must be exiting, otherwise would be on tasklist. */
5560 BUG_ON(!p->exit_state);
5562 /* Cannot have done final schedule yet: would have vanished. */
5563 BUG_ON(p->state == TASK_DEAD);
5568 * Drop lock around migration; if someone else moves it,
5569 * that's OK. No task can be added to this CPU, so iteration is
5572 raw_spin_unlock_irq(&rq->lock);
5573 move_task_off_dead_cpu(dead_cpu, p);
5574 raw_spin_lock_irq(&rq->lock);
5579 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5580 static void migrate_dead_tasks(unsigned int dead_cpu)
5582 struct rq *rq = cpu_rq(dead_cpu);
5583 struct task_struct *next;
5586 if (!rq->nr_running)
5588 update_rq_clock(rq);
5589 next = pick_next_task(rq);
5592 next->sched_class->put_prev_task(rq, next);
5593 migrate_dead(dead_cpu, next);
5599 * remove the tasks which were accounted by rq from calc_load_tasks.
5601 static void calc_global_load_remove(struct rq *rq)
5603 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5604 rq->calc_load_active = 0;
5606 #endif /* CONFIG_HOTPLUG_CPU */
5608 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5610 static struct ctl_table sd_ctl_dir[] = {
5612 .procname = "sched_domain",
5618 static struct ctl_table sd_ctl_root[] = {
5620 .procname = "kernel",
5622 .child = sd_ctl_dir,
5627 static struct ctl_table *sd_alloc_ctl_entry(int n)
5629 struct ctl_table *entry =
5630 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5635 static void sd_free_ctl_entry(struct ctl_table **tablep)
5637 struct ctl_table *entry;
5640 * In the intermediate directories, both the child directory and
5641 * procname are dynamically allocated and could fail but the mode
5642 * will always be set. In the lowest directory the names are
5643 * static strings and all have proc handlers.
5645 for (entry = *tablep; entry->mode; entry++) {
5647 sd_free_ctl_entry(&entry->child);
5648 if (entry->proc_handler == NULL)
5649 kfree(entry->procname);
5657 set_table_entry(struct ctl_table *entry,
5658 const char *procname, void *data, int maxlen,
5659 mode_t mode, proc_handler *proc_handler)
5661 entry->procname = procname;
5663 entry->maxlen = maxlen;
5665 entry->proc_handler = proc_handler;
5668 static struct ctl_table *
5669 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5671 struct ctl_table *table = sd_alloc_ctl_entry(13);
5676 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5677 sizeof(long), 0644, proc_doulongvec_minmax);
5678 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5679 sizeof(long), 0644, proc_doulongvec_minmax);
5680 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5681 sizeof(int), 0644, proc_dointvec_minmax);
5682 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5683 sizeof(int), 0644, proc_dointvec_minmax);
5684 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5685 sizeof(int), 0644, proc_dointvec_minmax);
5686 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5687 sizeof(int), 0644, proc_dointvec_minmax);
5688 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5689 sizeof(int), 0644, proc_dointvec_minmax);
5690 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5691 sizeof(int), 0644, proc_dointvec_minmax);
5692 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5693 sizeof(int), 0644, proc_dointvec_minmax);
5694 set_table_entry(&table[9], "cache_nice_tries",
5695 &sd->cache_nice_tries,
5696 sizeof(int), 0644, proc_dointvec_minmax);
5697 set_table_entry(&table[10], "flags", &sd->flags,
5698 sizeof(int), 0644, proc_dointvec_minmax);
5699 set_table_entry(&table[11], "name", sd->name,
5700 CORENAME_MAX_SIZE, 0444, proc_dostring);
5701 /* &table[12] is terminator */
5706 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5708 struct ctl_table *entry, *table;
5709 struct sched_domain *sd;
5710 int domain_num = 0, i;
5713 for_each_domain(cpu, sd)
5715 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5720 for_each_domain(cpu, sd) {
5721 snprintf(buf, 32, "domain%d", i);
5722 entry->procname = kstrdup(buf, GFP_KERNEL);
5724 entry->child = sd_alloc_ctl_domain_table(sd);
5731 static struct ctl_table_header *sd_sysctl_header;
5732 static void register_sched_domain_sysctl(void)
5734 int i, cpu_num = num_possible_cpus();
5735 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5738 WARN_ON(sd_ctl_dir[0].child);
5739 sd_ctl_dir[0].child = entry;
5744 for_each_possible_cpu(i) {
5745 snprintf(buf, 32, "cpu%d", i);
5746 entry->procname = kstrdup(buf, GFP_KERNEL);
5748 entry->child = sd_alloc_ctl_cpu_table(i);
5752 WARN_ON(sd_sysctl_header);
5753 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5756 /* may be called multiple times per register */
5757 static void unregister_sched_domain_sysctl(void)
5759 if (sd_sysctl_header)
5760 unregister_sysctl_table(sd_sysctl_header);
5761 sd_sysctl_header = NULL;
5762 if (sd_ctl_dir[0].child)
5763 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5766 static void register_sched_domain_sysctl(void)
5769 static void unregister_sched_domain_sysctl(void)
5774 static void set_rq_online(struct rq *rq)
5777 const struct sched_class *class;
5779 cpumask_set_cpu(rq->cpu, rq->rd->online);
5782 for_each_class(class) {
5783 if (class->rq_online)
5784 class->rq_online(rq);
5789 static void set_rq_offline(struct rq *rq)
5792 const struct sched_class *class;
5794 for_each_class(class) {
5795 if (class->rq_offline)
5796 class->rq_offline(rq);
5799 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5805 * migration_call - callback that gets triggered when a CPU is added.
5806 * Here we can start up the necessary migration thread for the new CPU.
5808 static int __cpuinit
5809 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5811 struct task_struct *p;
5812 int cpu = (long)hcpu;
5813 unsigned long flags;
5818 case CPU_UP_PREPARE:
5819 case CPU_UP_PREPARE_FROZEN:
5820 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
5823 kthread_bind(p, cpu);
5824 /* Must be high prio: stop_machine expects to yield to it. */
5825 rq = task_rq_lock(p, &flags);
5826 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5827 task_rq_unlock(rq, &flags);
5829 cpu_rq(cpu)->migration_thread = p;
5830 rq->calc_load_update = calc_load_update;
5834 case CPU_ONLINE_FROZEN:
5835 /* Strictly unnecessary, as first user will wake it. */
5836 wake_up_process(cpu_rq(cpu)->migration_thread);
5838 /* Update our root-domain */
5840 raw_spin_lock_irqsave(&rq->lock, flags);
5842 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5846 raw_spin_unlock_irqrestore(&rq->lock, flags);
5849 #ifdef CONFIG_HOTPLUG_CPU
5850 case CPU_UP_CANCELED:
5851 case CPU_UP_CANCELED_FROZEN:
5852 if (!cpu_rq(cpu)->migration_thread)
5854 /* Unbind it from offline cpu so it can run. Fall thru. */
5855 kthread_bind(cpu_rq(cpu)->migration_thread,
5856 cpumask_any(cpu_online_mask));
5857 kthread_stop(cpu_rq(cpu)->migration_thread);
5858 put_task_struct(cpu_rq(cpu)->migration_thread);
5859 cpu_rq(cpu)->migration_thread = NULL;
5863 case CPU_DEAD_FROZEN:
5864 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5865 migrate_live_tasks(cpu);
5867 kthread_stop(rq->migration_thread);
5868 put_task_struct(rq->migration_thread);
5869 rq->migration_thread = NULL;
5870 /* Idle task back to normal (off runqueue, low prio) */
5871 raw_spin_lock_irq(&rq->lock);
5872 update_rq_clock(rq);
5873 deactivate_task(rq, rq->idle, 0);
5874 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5875 rq->idle->sched_class = &idle_sched_class;
5876 migrate_dead_tasks(cpu);
5877 raw_spin_unlock_irq(&rq->lock);
5879 migrate_nr_uninterruptible(rq);
5880 BUG_ON(rq->nr_running != 0);
5881 calc_global_load_remove(rq);
5883 * No need to migrate the tasks: it was best-effort if
5884 * they didn't take sched_hotcpu_mutex. Just wake up
5887 raw_spin_lock_irq(&rq->lock);
5888 while (!list_empty(&rq->migration_queue)) {
5889 struct migration_req *req;
5891 req = list_entry(rq->migration_queue.next,
5892 struct migration_req, list);
5893 list_del_init(&req->list);
5894 raw_spin_unlock_irq(&rq->lock);
5895 complete(&req->done);
5896 raw_spin_lock_irq(&rq->lock);
5898 raw_spin_unlock_irq(&rq->lock);
5902 case CPU_DYING_FROZEN:
5903 /* Update our root-domain */
5905 raw_spin_lock_irqsave(&rq->lock, flags);
5907 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5910 raw_spin_unlock_irqrestore(&rq->lock, flags);
5918 * Register at high priority so that task migration (migrate_all_tasks)
5919 * happens before everything else. This has to be lower priority than
5920 * the notifier in the perf_event subsystem, though.
5922 static struct notifier_block __cpuinitdata migration_notifier = {
5923 .notifier_call = migration_call,
5927 static int __init migration_init(void)
5929 void *cpu = (void *)(long)smp_processor_id();
5932 /* Start one for the boot CPU: */
5933 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5934 BUG_ON(err == NOTIFY_BAD);
5935 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5936 register_cpu_notifier(&migration_notifier);
5940 early_initcall(migration_init);
5945 #ifdef CONFIG_SCHED_DEBUG
5947 static __read_mostly int sched_domain_debug_enabled;
5949 static int __init sched_domain_debug_setup(char *str)
5951 sched_domain_debug_enabled = 1;
5955 early_param("sched_debug", sched_domain_debug_setup);
5957 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5958 struct cpumask *groupmask)
5960 struct sched_group *group = sd->groups;
5963 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5964 cpumask_clear(groupmask);
5966 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5968 if (!(sd->flags & SD_LOAD_BALANCE)) {
5969 printk("does not load-balance\n");
5971 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5976 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5978 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5979 printk(KERN_ERR "ERROR: domain->span does not contain "
5982 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
5983 printk(KERN_ERR "ERROR: domain->groups does not contain"
5987 printk(KERN_DEBUG "%*s groups:", level + 1, "");
5991 printk(KERN_ERR "ERROR: group is NULL\n");
5995 if (!group->cpu_power) {
5996 printk(KERN_CONT "\n");
5997 printk(KERN_ERR "ERROR: domain->cpu_power not "
6002 if (!cpumask_weight(sched_group_cpus(group))) {
6003 printk(KERN_CONT "\n");
6004 printk(KERN_ERR "ERROR: empty group\n");
6008 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6009 printk(KERN_CONT "\n");
6010 printk(KERN_ERR "ERROR: repeated CPUs\n");
6014 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6016 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6018 printk(KERN_CONT " %s", str);
6019 if (group->cpu_power != SCHED_LOAD_SCALE) {
6020 printk(KERN_CONT " (cpu_power = %d)",
6024 group = group->next;
6025 } while (group != sd->groups);
6026 printk(KERN_CONT "\n");
6028 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6029 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6032 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6033 printk(KERN_ERR "ERROR: parent span is not a superset "
6034 "of domain->span\n");
6038 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6040 cpumask_var_t groupmask;
6043 if (!sched_domain_debug_enabled)
6047 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6051 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6053 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6054 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6059 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6066 free_cpumask_var(groupmask);
6068 #else /* !CONFIG_SCHED_DEBUG */
6069 # define sched_domain_debug(sd, cpu) do { } while (0)
6070 #endif /* CONFIG_SCHED_DEBUG */
6072 static int sd_degenerate(struct sched_domain *sd)
6074 if (cpumask_weight(sched_domain_span(sd)) == 1)
6077 /* Following flags need at least 2 groups */
6078 if (sd->flags & (SD_LOAD_BALANCE |
6079 SD_BALANCE_NEWIDLE |
6083 SD_SHARE_PKG_RESOURCES)) {
6084 if (sd->groups != sd->groups->next)
6088 /* Following flags don't use groups */
6089 if (sd->flags & (SD_WAKE_AFFINE))
6096 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6098 unsigned long cflags = sd->flags, pflags = parent->flags;
6100 if (sd_degenerate(parent))
6103 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6106 /* Flags needing groups don't count if only 1 group in parent */
6107 if (parent->groups == parent->groups->next) {
6108 pflags &= ~(SD_LOAD_BALANCE |
6109 SD_BALANCE_NEWIDLE |
6113 SD_SHARE_PKG_RESOURCES);
6114 if (nr_node_ids == 1)
6115 pflags &= ~SD_SERIALIZE;
6117 if (~cflags & pflags)
6123 static void free_rootdomain(struct root_domain *rd)
6125 synchronize_sched();
6127 cpupri_cleanup(&rd->cpupri);
6129 free_cpumask_var(rd->rto_mask);
6130 free_cpumask_var(rd->online);
6131 free_cpumask_var(rd->span);
6135 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6137 struct root_domain *old_rd = NULL;
6138 unsigned long flags;
6140 raw_spin_lock_irqsave(&rq->lock, flags);
6145 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6148 cpumask_clear_cpu(rq->cpu, old_rd->span);
6151 * If we dont want to free the old_rt yet then
6152 * set old_rd to NULL to skip the freeing later
6155 if (!atomic_dec_and_test(&old_rd->refcount))
6159 atomic_inc(&rd->refcount);
6162 cpumask_set_cpu(rq->cpu, rd->span);
6163 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6166 raw_spin_unlock_irqrestore(&rq->lock, flags);
6169 free_rootdomain(old_rd);
6172 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6174 gfp_t gfp = GFP_KERNEL;
6176 memset(rd, 0, sizeof(*rd));
6181 if (!alloc_cpumask_var(&rd->span, gfp))
6183 if (!alloc_cpumask_var(&rd->online, gfp))
6185 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6188 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6193 free_cpumask_var(rd->rto_mask);
6195 free_cpumask_var(rd->online);
6197 free_cpumask_var(rd->span);
6202 static void init_defrootdomain(void)
6204 init_rootdomain(&def_root_domain, true);
6206 atomic_set(&def_root_domain.refcount, 1);
6209 static struct root_domain *alloc_rootdomain(void)
6211 struct root_domain *rd;
6213 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6217 if (init_rootdomain(rd, false) != 0) {
6226 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6227 * hold the hotplug lock.
6230 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6232 struct rq *rq = cpu_rq(cpu);
6233 struct sched_domain *tmp;
6235 /* Remove the sched domains which do not contribute to scheduling. */
6236 for (tmp = sd; tmp; ) {
6237 struct sched_domain *parent = tmp->parent;
6241 if (sd_parent_degenerate(tmp, parent)) {
6242 tmp->parent = parent->parent;
6244 parent->parent->child = tmp;
6249 if (sd && sd_degenerate(sd)) {
6255 sched_domain_debug(sd, cpu);
6257 rq_attach_root(rq, rd);
6258 rcu_assign_pointer(rq->sd, sd);
6261 /* cpus with isolated domains */
6262 static cpumask_var_t cpu_isolated_map;
6264 /* Setup the mask of cpus configured for isolated domains */
6265 static int __init isolated_cpu_setup(char *str)
6267 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6268 cpulist_parse(str, cpu_isolated_map);
6272 __setup("isolcpus=", isolated_cpu_setup);
6275 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6276 * to a function which identifies what group(along with sched group) a CPU
6277 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6278 * (due to the fact that we keep track of groups covered with a struct cpumask).
6280 * init_sched_build_groups will build a circular linked list of the groups
6281 * covered by the given span, and will set each group's ->cpumask correctly,
6282 * and ->cpu_power to 0.
6285 init_sched_build_groups(const struct cpumask *span,
6286 const struct cpumask *cpu_map,
6287 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6288 struct sched_group **sg,
6289 struct cpumask *tmpmask),
6290 struct cpumask *covered, struct cpumask *tmpmask)
6292 struct sched_group *first = NULL, *last = NULL;
6295 cpumask_clear(covered);
6297 for_each_cpu(i, span) {
6298 struct sched_group *sg;
6299 int group = group_fn(i, cpu_map, &sg, tmpmask);
6302 if (cpumask_test_cpu(i, covered))
6305 cpumask_clear(sched_group_cpus(sg));
6308 for_each_cpu(j, span) {
6309 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6312 cpumask_set_cpu(j, covered);
6313 cpumask_set_cpu(j, sched_group_cpus(sg));
6324 #define SD_NODES_PER_DOMAIN 16
6329 * find_next_best_node - find the next node to include in a sched_domain
6330 * @node: node whose sched_domain we're building
6331 * @used_nodes: nodes already in the sched_domain
6333 * Find the next node to include in a given scheduling domain. Simply
6334 * finds the closest node not already in the @used_nodes map.
6336 * Should use nodemask_t.
6338 static int find_next_best_node(int node, nodemask_t *used_nodes)
6340 int i, n, val, min_val, best_node = 0;
6344 for (i = 0; i < nr_node_ids; i++) {
6345 /* Start at @node */
6346 n = (node + i) % nr_node_ids;
6348 if (!nr_cpus_node(n))
6351 /* Skip already used nodes */
6352 if (node_isset(n, *used_nodes))
6355 /* Simple min distance search */
6356 val = node_distance(node, n);
6358 if (val < min_val) {
6364 node_set(best_node, *used_nodes);
6369 * sched_domain_node_span - get a cpumask for a node's sched_domain
6370 * @node: node whose cpumask we're constructing
6371 * @span: resulting cpumask
6373 * Given a node, construct a good cpumask for its sched_domain to span. It
6374 * should be one that prevents unnecessary balancing, but also spreads tasks
6377 static void sched_domain_node_span(int node, struct cpumask *span)
6379 nodemask_t used_nodes;
6382 cpumask_clear(span);
6383 nodes_clear(used_nodes);
6385 cpumask_or(span, span, cpumask_of_node(node));
6386 node_set(node, used_nodes);
6388 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6389 int next_node = find_next_best_node(node, &used_nodes);
6391 cpumask_or(span, span, cpumask_of_node(next_node));
6394 #endif /* CONFIG_NUMA */
6396 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6399 * The cpus mask in sched_group and sched_domain hangs off the end.
6401 * ( See the the comments in include/linux/sched.h:struct sched_group
6402 * and struct sched_domain. )
6404 struct static_sched_group {
6405 struct sched_group sg;
6406 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6409 struct static_sched_domain {
6410 struct sched_domain sd;
6411 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6417 cpumask_var_t domainspan;
6418 cpumask_var_t covered;
6419 cpumask_var_t notcovered;
6421 cpumask_var_t nodemask;
6422 cpumask_var_t this_sibling_map;
6423 cpumask_var_t this_core_map;
6424 cpumask_var_t send_covered;
6425 cpumask_var_t tmpmask;
6426 struct sched_group **sched_group_nodes;
6427 struct root_domain *rd;
6431 sa_sched_groups = 0,
6436 sa_this_sibling_map,
6438 sa_sched_group_nodes,
6448 * SMT sched-domains:
6450 #ifdef CONFIG_SCHED_SMT
6451 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6452 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6455 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6456 struct sched_group **sg, struct cpumask *unused)
6459 *sg = &per_cpu(sched_groups, cpu).sg;
6462 #endif /* CONFIG_SCHED_SMT */
6465 * multi-core sched-domains:
6467 #ifdef CONFIG_SCHED_MC
6468 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6469 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6470 #endif /* CONFIG_SCHED_MC */
6472 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6474 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6475 struct sched_group **sg, struct cpumask *mask)
6479 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6480 group = cpumask_first(mask);
6482 *sg = &per_cpu(sched_group_core, group).sg;
6485 #elif defined(CONFIG_SCHED_MC)
6487 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6488 struct sched_group **sg, struct cpumask *unused)
6491 *sg = &per_cpu(sched_group_core, cpu).sg;
6496 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6497 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6500 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6501 struct sched_group **sg, struct cpumask *mask)
6504 #ifdef CONFIG_SCHED_MC
6505 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6506 group = cpumask_first(mask);
6507 #elif defined(CONFIG_SCHED_SMT)
6508 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6509 group = cpumask_first(mask);
6514 *sg = &per_cpu(sched_group_phys, group).sg;
6520 * The init_sched_build_groups can't handle what we want to do with node
6521 * groups, so roll our own. Now each node has its own list of groups which
6522 * gets dynamically allocated.
6524 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6525 static struct sched_group ***sched_group_nodes_bycpu;
6527 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6528 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6530 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6531 struct sched_group **sg,
6532 struct cpumask *nodemask)
6536 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6537 group = cpumask_first(nodemask);
6540 *sg = &per_cpu(sched_group_allnodes, group).sg;
6544 static void init_numa_sched_groups_power(struct sched_group *group_head)
6546 struct sched_group *sg = group_head;
6552 for_each_cpu(j, sched_group_cpus(sg)) {
6553 struct sched_domain *sd;
6555 sd = &per_cpu(phys_domains, j).sd;
6556 if (j != group_first_cpu(sd->groups)) {
6558 * Only add "power" once for each
6564 sg->cpu_power += sd->groups->cpu_power;
6567 } while (sg != group_head);
6570 static int build_numa_sched_groups(struct s_data *d,
6571 const struct cpumask *cpu_map, int num)
6573 struct sched_domain *sd;
6574 struct sched_group *sg, *prev;
6577 cpumask_clear(d->covered);
6578 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6579 if (cpumask_empty(d->nodemask)) {
6580 d->sched_group_nodes[num] = NULL;
6584 sched_domain_node_span(num, d->domainspan);
6585 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6587 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6590 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6594 d->sched_group_nodes[num] = sg;
6596 for_each_cpu(j, d->nodemask) {
6597 sd = &per_cpu(node_domains, j).sd;
6602 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6604 cpumask_or(d->covered, d->covered, d->nodemask);
6607 for (j = 0; j < nr_node_ids; j++) {
6608 n = (num + j) % nr_node_ids;
6609 cpumask_complement(d->notcovered, d->covered);
6610 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6611 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6612 if (cpumask_empty(d->tmpmask))
6614 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6615 if (cpumask_empty(d->tmpmask))
6617 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6621 "Can not alloc domain group for node %d\n", j);
6625 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6626 sg->next = prev->next;
6627 cpumask_or(d->covered, d->covered, d->tmpmask);
6634 #endif /* CONFIG_NUMA */
6637 /* Free memory allocated for various sched_group structures */
6638 static void free_sched_groups(const struct cpumask *cpu_map,
6639 struct cpumask *nodemask)
6643 for_each_cpu(cpu, cpu_map) {
6644 struct sched_group **sched_group_nodes
6645 = sched_group_nodes_bycpu[cpu];
6647 if (!sched_group_nodes)
6650 for (i = 0; i < nr_node_ids; i++) {
6651 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6653 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6654 if (cpumask_empty(nodemask))
6664 if (oldsg != sched_group_nodes[i])
6667 kfree(sched_group_nodes);
6668 sched_group_nodes_bycpu[cpu] = NULL;
6671 #else /* !CONFIG_NUMA */
6672 static void free_sched_groups(const struct cpumask *cpu_map,
6673 struct cpumask *nodemask)
6676 #endif /* CONFIG_NUMA */
6679 * Initialize sched groups cpu_power.
6681 * cpu_power indicates the capacity of sched group, which is used while
6682 * distributing the load between different sched groups in a sched domain.
6683 * Typically cpu_power for all the groups in a sched domain will be same unless
6684 * there are asymmetries in the topology. If there are asymmetries, group
6685 * having more cpu_power will pickup more load compared to the group having
6688 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6690 struct sched_domain *child;
6691 struct sched_group *group;
6695 WARN_ON(!sd || !sd->groups);
6697 if (cpu != group_first_cpu(sd->groups))
6702 sd->groups->cpu_power = 0;
6705 power = SCHED_LOAD_SCALE;
6706 weight = cpumask_weight(sched_domain_span(sd));
6708 * SMT siblings share the power of a single core.
6709 * Usually multiple threads get a better yield out of
6710 * that one core than a single thread would have,
6711 * reflect that in sd->smt_gain.
6713 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6714 power *= sd->smt_gain;
6716 power >>= SCHED_LOAD_SHIFT;
6718 sd->groups->cpu_power += power;
6723 * Add cpu_power of each child group to this groups cpu_power.
6725 group = child->groups;
6727 sd->groups->cpu_power += group->cpu_power;
6728 group = group->next;
6729 } while (group != child->groups);
6733 * Initializers for schedule domains
6734 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6737 #ifdef CONFIG_SCHED_DEBUG
6738 # define SD_INIT_NAME(sd, type) sd->name = #type
6740 # define SD_INIT_NAME(sd, type) do { } while (0)
6743 #define SD_INIT(sd, type) sd_init_##type(sd)
6745 #define SD_INIT_FUNC(type) \
6746 static noinline void sd_init_##type(struct sched_domain *sd) \
6748 memset(sd, 0, sizeof(*sd)); \
6749 *sd = SD_##type##_INIT; \
6750 sd->level = SD_LV_##type; \
6751 SD_INIT_NAME(sd, type); \
6756 SD_INIT_FUNC(ALLNODES)
6759 #ifdef CONFIG_SCHED_SMT
6760 SD_INIT_FUNC(SIBLING)
6762 #ifdef CONFIG_SCHED_MC
6766 static int default_relax_domain_level = -1;
6768 static int __init setup_relax_domain_level(char *str)
6772 val = simple_strtoul(str, NULL, 0);
6773 if (val < SD_LV_MAX)
6774 default_relax_domain_level = val;
6778 __setup("relax_domain_level=", setup_relax_domain_level);
6780 static void set_domain_attribute(struct sched_domain *sd,
6781 struct sched_domain_attr *attr)
6785 if (!attr || attr->relax_domain_level < 0) {
6786 if (default_relax_domain_level < 0)
6789 request = default_relax_domain_level;
6791 request = attr->relax_domain_level;
6792 if (request < sd->level) {
6793 /* turn off idle balance on this domain */
6794 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6796 /* turn on idle balance on this domain */
6797 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6801 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6802 const struct cpumask *cpu_map)
6805 case sa_sched_groups:
6806 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6807 d->sched_group_nodes = NULL;
6809 free_rootdomain(d->rd); /* fall through */
6811 free_cpumask_var(d->tmpmask); /* fall through */
6812 case sa_send_covered:
6813 free_cpumask_var(d->send_covered); /* fall through */
6814 case sa_this_core_map:
6815 free_cpumask_var(d->this_core_map); /* fall through */
6816 case sa_this_sibling_map:
6817 free_cpumask_var(d->this_sibling_map); /* fall through */
6819 free_cpumask_var(d->nodemask); /* fall through */
6820 case sa_sched_group_nodes:
6822 kfree(d->sched_group_nodes); /* fall through */
6824 free_cpumask_var(d->notcovered); /* fall through */
6826 free_cpumask_var(d->covered); /* fall through */
6828 free_cpumask_var(d->domainspan); /* fall through */
6835 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6836 const struct cpumask *cpu_map)
6839 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6841 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6842 return sa_domainspan;
6843 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6845 /* Allocate the per-node list of sched groups */
6846 d->sched_group_nodes = kcalloc(nr_node_ids,
6847 sizeof(struct sched_group *), GFP_KERNEL);
6848 if (!d->sched_group_nodes) {
6849 printk(KERN_WARNING "Can not alloc sched group node list\n");
6850 return sa_notcovered;
6852 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6854 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6855 return sa_sched_group_nodes;
6856 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6858 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6859 return sa_this_sibling_map;
6860 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6861 return sa_this_core_map;
6862 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6863 return sa_send_covered;
6864 d->rd = alloc_rootdomain();
6866 printk(KERN_WARNING "Cannot alloc root domain\n");
6869 return sa_rootdomain;
6872 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6873 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6875 struct sched_domain *sd = NULL;
6877 struct sched_domain *parent;
6880 if (cpumask_weight(cpu_map) >
6881 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6882 sd = &per_cpu(allnodes_domains, i).sd;
6883 SD_INIT(sd, ALLNODES);
6884 set_domain_attribute(sd, attr);
6885 cpumask_copy(sched_domain_span(sd), cpu_map);
6886 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6891 sd = &per_cpu(node_domains, i).sd;
6893 set_domain_attribute(sd, attr);
6894 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6895 sd->parent = parent;
6898 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6903 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6904 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6905 struct sched_domain *parent, int i)
6907 struct sched_domain *sd;
6908 sd = &per_cpu(phys_domains, i).sd;
6910 set_domain_attribute(sd, attr);
6911 cpumask_copy(sched_domain_span(sd), d->nodemask);
6912 sd->parent = parent;
6915 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6919 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6920 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6921 struct sched_domain *parent, int i)
6923 struct sched_domain *sd = parent;
6924 #ifdef CONFIG_SCHED_MC
6925 sd = &per_cpu(core_domains, i).sd;
6927 set_domain_attribute(sd, attr);
6928 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6929 sd->parent = parent;
6931 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
6936 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
6937 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6938 struct sched_domain *parent, int i)
6940 struct sched_domain *sd = parent;
6941 #ifdef CONFIG_SCHED_SMT
6942 sd = &per_cpu(cpu_domains, i).sd;
6943 SD_INIT(sd, SIBLING);
6944 set_domain_attribute(sd, attr);
6945 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
6946 sd->parent = parent;
6948 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
6953 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
6954 const struct cpumask *cpu_map, int cpu)
6957 #ifdef CONFIG_SCHED_SMT
6958 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
6959 cpumask_and(d->this_sibling_map, cpu_map,
6960 topology_thread_cpumask(cpu));
6961 if (cpu == cpumask_first(d->this_sibling_map))
6962 init_sched_build_groups(d->this_sibling_map, cpu_map,
6964 d->send_covered, d->tmpmask);
6967 #ifdef CONFIG_SCHED_MC
6968 case SD_LV_MC: /* set up multi-core groups */
6969 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
6970 if (cpu == cpumask_first(d->this_core_map))
6971 init_sched_build_groups(d->this_core_map, cpu_map,
6973 d->send_covered, d->tmpmask);
6976 case SD_LV_CPU: /* set up physical groups */
6977 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
6978 if (!cpumask_empty(d->nodemask))
6979 init_sched_build_groups(d->nodemask, cpu_map,
6981 d->send_covered, d->tmpmask);
6984 case SD_LV_ALLNODES:
6985 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
6986 d->send_covered, d->tmpmask);
6995 * Build sched domains for a given set of cpus and attach the sched domains
6996 * to the individual cpus
6998 static int __build_sched_domains(const struct cpumask *cpu_map,
6999 struct sched_domain_attr *attr)
7001 enum s_alloc alloc_state = sa_none;
7003 struct sched_domain *sd;
7009 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7010 if (alloc_state != sa_rootdomain)
7012 alloc_state = sa_sched_groups;
7015 * Set up domains for cpus specified by the cpu_map.
7017 for_each_cpu(i, cpu_map) {
7018 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7021 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7022 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7023 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7024 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7027 for_each_cpu(i, cpu_map) {
7028 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7029 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7032 /* Set up physical groups */
7033 for (i = 0; i < nr_node_ids; i++)
7034 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7037 /* Set up node groups */
7039 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7041 for (i = 0; i < nr_node_ids; i++)
7042 if (build_numa_sched_groups(&d, cpu_map, i))
7046 /* Calculate CPU power for physical packages and nodes */
7047 #ifdef CONFIG_SCHED_SMT
7048 for_each_cpu(i, cpu_map) {
7049 sd = &per_cpu(cpu_domains, i).sd;
7050 init_sched_groups_power(i, sd);
7053 #ifdef CONFIG_SCHED_MC
7054 for_each_cpu(i, cpu_map) {
7055 sd = &per_cpu(core_domains, i).sd;
7056 init_sched_groups_power(i, sd);
7060 for_each_cpu(i, cpu_map) {
7061 sd = &per_cpu(phys_domains, i).sd;
7062 init_sched_groups_power(i, sd);
7066 for (i = 0; i < nr_node_ids; i++)
7067 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7069 if (d.sd_allnodes) {
7070 struct sched_group *sg;
7072 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7074 init_numa_sched_groups_power(sg);
7078 /* Attach the domains */
7079 for_each_cpu(i, cpu_map) {
7080 #ifdef CONFIG_SCHED_SMT
7081 sd = &per_cpu(cpu_domains, i).sd;
7082 #elif defined(CONFIG_SCHED_MC)
7083 sd = &per_cpu(core_domains, i).sd;
7085 sd = &per_cpu(phys_domains, i).sd;
7087 cpu_attach_domain(sd, d.rd, i);
7090 d.sched_group_nodes = NULL; /* don't free this we still need it */
7091 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7095 __free_domain_allocs(&d, alloc_state, cpu_map);
7099 static int build_sched_domains(const struct cpumask *cpu_map)
7101 return __build_sched_domains(cpu_map, NULL);
7104 static cpumask_var_t *doms_cur; /* current sched domains */
7105 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7106 static struct sched_domain_attr *dattr_cur;
7107 /* attribues of custom domains in 'doms_cur' */
7110 * Special case: If a kmalloc of a doms_cur partition (array of
7111 * cpumask) fails, then fallback to a single sched domain,
7112 * as determined by the single cpumask fallback_doms.
7114 static cpumask_var_t fallback_doms;
7117 * arch_update_cpu_topology lets virtualized architectures update the
7118 * cpu core maps. It is supposed to return 1 if the topology changed
7119 * or 0 if it stayed the same.
7121 int __attribute__((weak)) arch_update_cpu_topology(void)
7126 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7129 cpumask_var_t *doms;
7131 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7134 for (i = 0; i < ndoms; i++) {
7135 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7136 free_sched_domains(doms, i);
7143 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7146 for (i = 0; i < ndoms; i++)
7147 free_cpumask_var(doms[i]);
7152 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7153 * For now this just excludes isolated cpus, but could be used to
7154 * exclude other special cases in the future.
7156 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7160 arch_update_cpu_topology();
7162 doms_cur = alloc_sched_domains(ndoms_cur);
7164 doms_cur = &fallback_doms;
7165 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7167 err = build_sched_domains(doms_cur[0]);
7168 register_sched_domain_sysctl();
7173 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7174 struct cpumask *tmpmask)
7176 free_sched_groups(cpu_map, tmpmask);
7180 * Detach sched domains from a group of cpus specified in cpu_map
7181 * These cpus will now be attached to the NULL domain
7183 static void detach_destroy_domains(const struct cpumask *cpu_map)
7185 /* Save because hotplug lock held. */
7186 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7189 for_each_cpu(i, cpu_map)
7190 cpu_attach_domain(NULL, &def_root_domain, i);
7191 synchronize_sched();
7192 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7195 /* handle null as "default" */
7196 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7197 struct sched_domain_attr *new, int idx_new)
7199 struct sched_domain_attr tmp;
7206 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7207 new ? (new + idx_new) : &tmp,
7208 sizeof(struct sched_domain_attr));
7212 * Partition sched domains as specified by the 'ndoms_new'
7213 * cpumasks in the array doms_new[] of cpumasks. This compares
7214 * doms_new[] to the current sched domain partitioning, doms_cur[].
7215 * It destroys each deleted domain and builds each new domain.
7217 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7218 * The masks don't intersect (don't overlap.) We should setup one
7219 * sched domain for each mask. CPUs not in any of the cpumasks will
7220 * not be load balanced. If the same cpumask appears both in the
7221 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7224 * The passed in 'doms_new' should be allocated using
7225 * alloc_sched_domains. This routine takes ownership of it and will
7226 * free_sched_domains it when done with it. If the caller failed the
7227 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7228 * and partition_sched_domains() will fallback to the single partition
7229 * 'fallback_doms', it also forces the domains to be rebuilt.
7231 * If doms_new == NULL it will be replaced with cpu_online_mask.
7232 * ndoms_new == 0 is a special case for destroying existing domains,
7233 * and it will not create the default domain.
7235 * Call with hotplug lock held
7237 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7238 struct sched_domain_attr *dattr_new)
7243 mutex_lock(&sched_domains_mutex);
7245 /* always unregister in case we don't destroy any domains */
7246 unregister_sched_domain_sysctl();
7248 /* Let architecture update cpu core mappings. */
7249 new_topology = arch_update_cpu_topology();
7251 n = doms_new ? ndoms_new : 0;
7253 /* Destroy deleted domains */
7254 for (i = 0; i < ndoms_cur; i++) {
7255 for (j = 0; j < n && !new_topology; j++) {
7256 if (cpumask_equal(doms_cur[i], doms_new[j])
7257 && dattrs_equal(dattr_cur, i, dattr_new, j))
7260 /* no match - a current sched domain not in new doms_new[] */
7261 detach_destroy_domains(doms_cur[i]);
7266 if (doms_new == NULL) {
7268 doms_new = &fallback_doms;
7269 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7270 WARN_ON_ONCE(dattr_new);
7273 /* Build new domains */
7274 for (i = 0; i < ndoms_new; i++) {
7275 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7276 if (cpumask_equal(doms_new[i], doms_cur[j])
7277 && dattrs_equal(dattr_new, i, dattr_cur, j))
7280 /* no match - add a new doms_new */
7281 __build_sched_domains(doms_new[i],
7282 dattr_new ? dattr_new + i : NULL);
7287 /* Remember the new sched domains */
7288 if (doms_cur != &fallback_doms)
7289 free_sched_domains(doms_cur, ndoms_cur);
7290 kfree(dattr_cur); /* kfree(NULL) is safe */
7291 doms_cur = doms_new;
7292 dattr_cur = dattr_new;
7293 ndoms_cur = ndoms_new;
7295 register_sched_domain_sysctl();
7297 mutex_unlock(&sched_domains_mutex);
7300 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7301 static void arch_reinit_sched_domains(void)
7305 /* Destroy domains first to force the rebuild */
7306 partition_sched_domains(0, NULL, NULL);
7308 rebuild_sched_domains();
7312 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7314 unsigned int level = 0;
7316 if (sscanf(buf, "%u", &level) != 1)
7320 * level is always be positive so don't check for
7321 * level < POWERSAVINGS_BALANCE_NONE which is 0
7322 * What happens on 0 or 1 byte write,
7323 * need to check for count as well?
7326 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7330 sched_smt_power_savings = level;
7332 sched_mc_power_savings = level;
7334 arch_reinit_sched_domains();
7339 #ifdef CONFIG_SCHED_MC
7340 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7343 return sprintf(page, "%u\n", sched_mc_power_savings);
7345 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7346 const char *buf, size_t count)
7348 return sched_power_savings_store(buf, count, 0);
7350 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7351 sched_mc_power_savings_show,
7352 sched_mc_power_savings_store);
7355 #ifdef CONFIG_SCHED_SMT
7356 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7359 return sprintf(page, "%u\n", sched_smt_power_savings);
7361 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7362 const char *buf, size_t count)
7364 return sched_power_savings_store(buf, count, 1);
7366 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7367 sched_smt_power_savings_show,
7368 sched_smt_power_savings_store);
7371 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7375 #ifdef CONFIG_SCHED_SMT
7377 err = sysfs_create_file(&cls->kset.kobj,
7378 &attr_sched_smt_power_savings.attr);
7380 #ifdef CONFIG_SCHED_MC
7381 if (!err && mc_capable())
7382 err = sysfs_create_file(&cls->kset.kobj,
7383 &attr_sched_mc_power_savings.attr);
7387 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7389 #ifndef CONFIG_CPUSETS
7391 * Add online and remove offline CPUs from the scheduler domains.
7392 * When cpusets are enabled they take over this function.
7394 static int update_sched_domains(struct notifier_block *nfb,
7395 unsigned long action, void *hcpu)
7399 case CPU_ONLINE_FROZEN:
7400 case CPU_DOWN_PREPARE:
7401 case CPU_DOWN_PREPARE_FROZEN:
7402 case CPU_DOWN_FAILED:
7403 case CPU_DOWN_FAILED_FROZEN:
7404 partition_sched_domains(1, NULL, NULL);
7413 static int update_runtime(struct notifier_block *nfb,
7414 unsigned long action, void *hcpu)
7416 int cpu = (int)(long)hcpu;
7419 case CPU_DOWN_PREPARE:
7420 case CPU_DOWN_PREPARE_FROZEN:
7421 disable_runtime(cpu_rq(cpu));
7424 case CPU_DOWN_FAILED:
7425 case CPU_DOWN_FAILED_FROZEN:
7427 case CPU_ONLINE_FROZEN:
7428 enable_runtime(cpu_rq(cpu));
7436 void __init sched_init_smp(void)
7438 cpumask_var_t non_isolated_cpus;
7440 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7441 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7443 #if defined(CONFIG_NUMA)
7444 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7446 BUG_ON(sched_group_nodes_bycpu == NULL);
7449 mutex_lock(&sched_domains_mutex);
7450 arch_init_sched_domains(cpu_active_mask);
7451 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7452 if (cpumask_empty(non_isolated_cpus))
7453 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7454 mutex_unlock(&sched_domains_mutex);
7457 #ifndef CONFIG_CPUSETS
7458 /* XXX: Theoretical race here - CPU may be hotplugged now */
7459 hotcpu_notifier(update_sched_domains, 0);
7462 /* RT runtime code needs to handle some hotplug events */
7463 hotcpu_notifier(update_runtime, 0);
7467 /* Move init over to a non-isolated CPU */
7468 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7470 sched_init_granularity();
7471 free_cpumask_var(non_isolated_cpus);
7473 init_sched_rt_class();
7476 void __init sched_init_smp(void)
7478 sched_init_granularity();
7480 #endif /* CONFIG_SMP */
7482 const_debug unsigned int sysctl_timer_migration = 1;
7484 int in_sched_functions(unsigned long addr)
7486 return in_lock_functions(addr) ||
7487 (addr >= (unsigned long)__sched_text_start
7488 && addr < (unsigned long)__sched_text_end);
7491 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7493 cfs_rq->tasks_timeline = RB_ROOT;
7494 INIT_LIST_HEAD(&cfs_rq->tasks);
7495 #ifdef CONFIG_FAIR_GROUP_SCHED
7498 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7501 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7503 struct rt_prio_array *array;
7506 array = &rt_rq->active;
7507 for (i = 0; i < MAX_RT_PRIO; i++) {
7508 INIT_LIST_HEAD(array->queue + i);
7509 __clear_bit(i, array->bitmap);
7511 /* delimiter for bitsearch: */
7512 __set_bit(MAX_RT_PRIO, array->bitmap);
7514 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7515 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7517 rt_rq->highest_prio.next = MAX_RT_PRIO;
7521 rt_rq->rt_nr_migratory = 0;
7522 rt_rq->overloaded = 0;
7523 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7527 rt_rq->rt_throttled = 0;
7528 rt_rq->rt_runtime = 0;
7529 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7531 #ifdef CONFIG_RT_GROUP_SCHED
7532 rt_rq->rt_nr_boosted = 0;
7537 #ifdef CONFIG_FAIR_GROUP_SCHED
7538 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7539 struct sched_entity *se, int cpu, int add,
7540 struct sched_entity *parent)
7542 struct rq *rq = cpu_rq(cpu);
7543 tg->cfs_rq[cpu] = cfs_rq;
7544 init_cfs_rq(cfs_rq, rq);
7547 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7550 /* se could be NULL for init_task_group */
7555 se->cfs_rq = &rq->cfs;
7557 se->cfs_rq = parent->my_q;
7560 se->load.weight = tg->shares;
7561 se->load.inv_weight = 0;
7562 se->parent = parent;
7566 #ifdef CONFIG_RT_GROUP_SCHED
7567 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7568 struct sched_rt_entity *rt_se, int cpu, int add,
7569 struct sched_rt_entity *parent)
7571 struct rq *rq = cpu_rq(cpu);
7573 tg->rt_rq[cpu] = rt_rq;
7574 init_rt_rq(rt_rq, rq);
7576 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7578 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7580 tg->rt_se[cpu] = rt_se;
7585 rt_se->rt_rq = &rq->rt;
7587 rt_se->rt_rq = parent->my_q;
7589 rt_se->my_q = rt_rq;
7590 rt_se->parent = parent;
7591 INIT_LIST_HEAD(&rt_se->run_list);
7595 void __init sched_init(void)
7598 unsigned long alloc_size = 0, ptr;
7600 #ifdef CONFIG_FAIR_GROUP_SCHED
7601 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7603 #ifdef CONFIG_RT_GROUP_SCHED
7604 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7606 #ifdef CONFIG_CPUMASK_OFFSTACK
7607 alloc_size += num_possible_cpus() * cpumask_size();
7610 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7612 #ifdef CONFIG_FAIR_GROUP_SCHED
7613 init_task_group.se = (struct sched_entity **)ptr;
7614 ptr += nr_cpu_ids * sizeof(void **);
7616 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7617 ptr += nr_cpu_ids * sizeof(void **);
7619 #endif /* CONFIG_FAIR_GROUP_SCHED */
7620 #ifdef CONFIG_RT_GROUP_SCHED
7621 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7622 ptr += nr_cpu_ids * sizeof(void **);
7624 init_task_group.rt_rq = (struct rt_rq **)ptr;
7625 ptr += nr_cpu_ids * sizeof(void **);
7627 #endif /* CONFIG_RT_GROUP_SCHED */
7628 #ifdef CONFIG_CPUMASK_OFFSTACK
7629 for_each_possible_cpu(i) {
7630 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7631 ptr += cpumask_size();
7633 #endif /* CONFIG_CPUMASK_OFFSTACK */
7637 init_defrootdomain();
7640 init_rt_bandwidth(&def_rt_bandwidth,
7641 global_rt_period(), global_rt_runtime());
7643 #ifdef CONFIG_RT_GROUP_SCHED
7644 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7645 global_rt_period(), global_rt_runtime());
7646 #endif /* CONFIG_RT_GROUP_SCHED */
7648 #ifdef CONFIG_CGROUP_SCHED
7649 list_add(&init_task_group.list, &task_groups);
7650 INIT_LIST_HEAD(&init_task_group.children);
7652 #endif /* CONFIG_CGROUP_SCHED */
7654 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7655 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7656 __alignof__(unsigned long));
7658 for_each_possible_cpu(i) {
7662 raw_spin_lock_init(&rq->lock);
7664 rq->calc_load_active = 0;
7665 rq->calc_load_update = jiffies + LOAD_FREQ;
7666 init_cfs_rq(&rq->cfs, rq);
7667 init_rt_rq(&rq->rt, rq);
7668 #ifdef CONFIG_FAIR_GROUP_SCHED
7669 init_task_group.shares = init_task_group_load;
7670 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7671 #ifdef CONFIG_CGROUP_SCHED
7673 * How much cpu bandwidth does init_task_group get?
7675 * In case of task-groups formed thr' the cgroup filesystem, it
7676 * gets 100% of the cpu resources in the system. This overall
7677 * system cpu resource is divided among the tasks of
7678 * init_task_group and its child task-groups in a fair manner,
7679 * based on each entity's (task or task-group's) weight
7680 * (se->load.weight).
7682 * In other words, if init_task_group has 10 tasks of weight
7683 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7684 * then A0's share of the cpu resource is:
7686 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7688 * We achieve this by letting init_task_group's tasks sit
7689 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7691 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7693 #endif /* CONFIG_FAIR_GROUP_SCHED */
7695 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7696 #ifdef CONFIG_RT_GROUP_SCHED
7697 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7698 #ifdef CONFIG_CGROUP_SCHED
7699 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7703 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7704 rq->cpu_load[j] = 0;
7708 rq->post_schedule = 0;
7709 rq->active_balance = 0;
7710 rq->next_balance = jiffies;
7714 rq->migration_thread = NULL;
7716 rq->avg_idle = 2*sysctl_sched_migration_cost;
7717 INIT_LIST_HEAD(&rq->migration_queue);
7718 rq_attach_root(rq, &def_root_domain);
7721 atomic_set(&rq->nr_iowait, 0);
7724 set_load_weight(&init_task);
7726 #ifdef CONFIG_PREEMPT_NOTIFIERS
7727 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7731 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7734 #ifdef CONFIG_RT_MUTEXES
7735 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7739 * The boot idle thread does lazy MMU switching as well:
7741 atomic_inc(&init_mm.mm_count);
7742 enter_lazy_tlb(&init_mm, current);
7745 * Make us the idle thread. Technically, schedule() should not be
7746 * called from this thread, however somewhere below it might be,
7747 * but because we are the idle thread, we just pick up running again
7748 * when this runqueue becomes "idle".
7750 init_idle(current, smp_processor_id());
7752 calc_load_update = jiffies + LOAD_FREQ;
7755 * During early bootup we pretend to be a normal task:
7757 current->sched_class = &fair_sched_class;
7759 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7760 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7763 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7764 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7766 /* May be allocated at isolcpus cmdline parse time */
7767 if (cpu_isolated_map == NULL)
7768 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7773 scheduler_running = 1;
7776 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7777 static inline int preempt_count_equals(int preempt_offset)
7779 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7781 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7784 void __might_sleep(const char *file, int line, int preempt_offset)
7787 static unsigned long prev_jiffy; /* ratelimiting */
7789 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7790 system_state != SYSTEM_RUNNING || oops_in_progress)
7792 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7794 prev_jiffy = jiffies;
7797 "BUG: sleeping function called from invalid context at %s:%d\n",
7800 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7801 in_atomic(), irqs_disabled(),
7802 current->pid, current->comm);
7804 debug_show_held_locks(current);
7805 if (irqs_disabled())
7806 print_irqtrace_events(current);
7810 EXPORT_SYMBOL(__might_sleep);
7813 #ifdef CONFIG_MAGIC_SYSRQ
7814 static void normalize_task(struct rq *rq, struct task_struct *p)
7818 update_rq_clock(rq);
7819 on_rq = p->se.on_rq;
7821 deactivate_task(rq, p, 0);
7822 __setscheduler(rq, p, SCHED_NORMAL, 0);
7824 activate_task(rq, p, 0);
7825 resched_task(rq->curr);
7829 void normalize_rt_tasks(void)
7831 struct task_struct *g, *p;
7832 unsigned long flags;
7835 read_lock_irqsave(&tasklist_lock, flags);
7836 do_each_thread(g, p) {
7838 * Only normalize user tasks:
7843 p->se.exec_start = 0;
7844 #ifdef CONFIG_SCHEDSTATS
7845 p->se.statistics.wait_start = 0;
7846 p->se.statistics.sleep_start = 0;
7847 p->se.statistics.block_start = 0;
7852 * Renice negative nice level userspace
7855 if (TASK_NICE(p) < 0 && p->mm)
7856 set_user_nice(p, 0);
7860 raw_spin_lock(&p->pi_lock);
7861 rq = __task_rq_lock(p);
7863 normalize_task(rq, p);
7865 __task_rq_unlock(rq);
7866 raw_spin_unlock(&p->pi_lock);
7867 } while_each_thread(g, p);
7869 read_unlock_irqrestore(&tasklist_lock, flags);
7872 #endif /* CONFIG_MAGIC_SYSRQ */
7876 * These functions are only useful for the IA64 MCA handling.
7878 * They can only be called when the whole system has been
7879 * stopped - every CPU needs to be quiescent, and no scheduling
7880 * activity can take place. Using them for anything else would
7881 * be a serious bug, and as a result, they aren't even visible
7882 * under any other configuration.
7886 * curr_task - return the current task for a given cpu.
7887 * @cpu: the processor in question.
7889 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7891 struct task_struct *curr_task(int cpu)
7893 return cpu_curr(cpu);
7897 * set_curr_task - set the current task for a given cpu.
7898 * @cpu: the processor in question.
7899 * @p: the task pointer to set.
7901 * Description: This function must only be used when non-maskable interrupts
7902 * are serviced on a separate stack. It allows the architecture to switch the
7903 * notion of the current task on a cpu in a non-blocking manner. This function
7904 * must be called with all CPU's synchronized, and interrupts disabled, the
7905 * and caller must save the original value of the current task (see
7906 * curr_task() above) and restore that value before reenabling interrupts and
7907 * re-starting the system.
7909 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7911 void set_curr_task(int cpu, struct task_struct *p)
7918 #ifdef CONFIG_FAIR_GROUP_SCHED
7919 static void free_fair_sched_group(struct task_group *tg)
7923 for_each_possible_cpu(i) {
7925 kfree(tg->cfs_rq[i]);
7935 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7937 struct cfs_rq *cfs_rq;
7938 struct sched_entity *se;
7942 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
7945 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
7949 tg->shares = NICE_0_LOAD;
7951 for_each_possible_cpu(i) {
7954 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
7955 GFP_KERNEL, cpu_to_node(i));
7959 se = kzalloc_node(sizeof(struct sched_entity),
7960 GFP_KERNEL, cpu_to_node(i));
7964 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
7975 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
7977 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
7978 &cpu_rq(cpu)->leaf_cfs_rq_list);
7981 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
7983 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
7985 #else /* !CONFG_FAIR_GROUP_SCHED */
7986 static inline void free_fair_sched_group(struct task_group *tg)
7991 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
7996 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8000 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8003 #endif /* CONFIG_FAIR_GROUP_SCHED */
8005 #ifdef CONFIG_RT_GROUP_SCHED
8006 static void free_rt_sched_group(struct task_group *tg)
8010 destroy_rt_bandwidth(&tg->rt_bandwidth);
8012 for_each_possible_cpu(i) {
8014 kfree(tg->rt_rq[i]);
8016 kfree(tg->rt_se[i]);
8024 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8026 struct rt_rq *rt_rq;
8027 struct sched_rt_entity *rt_se;
8031 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8034 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8038 init_rt_bandwidth(&tg->rt_bandwidth,
8039 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8041 for_each_possible_cpu(i) {
8044 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8045 GFP_KERNEL, cpu_to_node(i));
8049 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8050 GFP_KERNEL, cpu_to_node(i));
8054 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8065 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8067 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8068 &cpu_rq(cpu)->leaf_rt_rq_list);
8071 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8073 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8075 #else /* !CONFIG_RT_GROUP_SCHED */
8076 static inline void free_rt_sched_group(struct task_group *tg)
8081 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8086 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8090 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8093 #endif /* CONFIG_RT_GROUP_SCHED */
8095 #ifdef CONFIG_CGROUP_SCHED
8096 static void free_sched_group(struct task_group *tg)
8098 free_fair_sched_group(tg);
8099 free_rt_sched_group(tg);
8103 /* allocate runqueue etc for a new task group */
8104 struct task_group *sched_create_group(struct task_group *parent)
8106 struct task_group *tg;
8107 unsigned long flags;
8110 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8112 return ERR_PTR(-ENOMEM);
8114 if (!alloc_fair_sched_group(tg, parent))
8117 if (!alloc_rt_sched_group(tg, parent))
8120 spin_lock_irqsave(&task_group_lock, flags);
8121 for_each_possible_cpu(i) {
8122 register_fair_sched_group(tg, i);
8123 register_rt_sched_group(tg, i);
8125 list_add_rcu(&tg->list, &task_groups);
8127 WARN_ON(!parent); /* root should already exist */
8129 tg->parent = parent;
8130 INIT_LIST_HEAD(&tg->children);
8131 list_add_rcu(&tg->siblings, &parent->children);
8132 spin_unlock_irqrestore(&task_group_lock, flags);
8137 free_sched_group(tg);
8138 return ERR_PTR(-ENOMEM);
8141 /* rcu callback to free various structures associated with a task group */
8142 static void free_sched_group_rcu(struct rcu_head *rhp)
8144 /* now it should be safe to free those cfs_rqs */
8145 free_sched_group(container_of(rhp, struct task_group, rcu));
8148 /* Destroy runqueue etc associated with a task group */
8149 void sched_destroy_group(struct task_group *tg)
8151 unsigned long flags;
8154 spin_lock_irqsave(&task_group_lock, flags);
8155 for_each_possible_cpu(i) {
8156 unregister_fair_sched_group(tg, i);
8157 unregister_rt_sched_group(tg, i);
8159 list_del_rcu(&tg->list);
8160 list_del_rcu(&tg->siblings);
8161 spin_unlock_irqrestore(&task_group_lock, flags);
8163 /* wait for possible concurrent references to cfs_rqs complete */
8164 call_rcu(&tg->rcu, free_sched_group_rcu);
8167 /* change task's runqueue when it moves between groups.
8168 * The caller of this function should have put the task in its new group
8169 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8170 * reflect its new group.
8172 void sched_move_task(struct task_struct *tsk)
8175 unsigned long flags;
8178 rq = task_rq_lock(tsk, &flags);
8180 update_rq_clock(rq);
8182 running = task_current(rq, tsk);
8183 on_rq = tsk->se.on_rq;
8186 dequeue_task(rq, tsk, 0);
8187 if (unlikely(running))
8188 tsk->sched_class->put_prev_task(rq, tsk);
8190 set_task_rq(tsk, task_cpu(tsk));
8192 #ifdef CONFIG_FAIR_GROUP_SCHED
8193 if (tsk->sched_class->moved_group)
8194 tsk->sched_class->moved_group(tsk, on_rq);
8197 if (unlikely(running))
8198 tsk->sched_class->set_curr_task(rq);
8200 enqueue_task(rq, tsk, 0, false);
8202 task_rq_unlock(rq, &flags);
8204 #endif /* CONFIG_CGROUP_SCHED */
8206 #ifdef CONFIG_FAIR_GROUP_SCHED
8207 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8209 struct cfs_rq *cfs_rq = se->cfs_rq;
8214 dequeue_entity(cfs_rq, se, 0);
8216 se->load.weight = shares;
8217 se->load.inv_weight = 0;
8220 enqueue_entity(cfs_rq, se, 0);
8223 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8225 struct cfs_rq *cfs_rq = se->cfs_rq;
8226 struct rq *rq = cfs_rq->rq;
8227 unsigned long flags;
8229 raw_spin_lock_irqsave(&rq->lock, flags);
8230 __set_se_shares(se, shares);
8231 raw_spin_unlock_irqrestore(&rq->lock, flags);
8234 static DEFINE_MUTEX(shares_mutex);
8236 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8239 unsigned long flags;
8242 * We can't change the weight of the root cgroup.
8247 if (shares < MIN_SHARES)
8248 shares = MIN_SHARES;
8249 else if (shares > MAX_SHARES)
8250 shares = MAX_SHARES;
8252 mutex_lock(&shares_mutex);
8253 if (tg->shares == shares)
8256 spin_lock_irqsave(&task_group_lock, flags);
8257 for_each_possible_cpu(i)
8258 unregister_fair_sched_group(tg, i);
8259 list_del_rcu(&tg->siblings);
8260 spin_unlock_irqrestore(&task_group_lock, flags);
8262 /* wait for any ongoing reference to this group to finish */
8263 synchronize_sched();
8266 * Now we are free to modify the group's share on each cpu
8267 * w/o tripping rebalance_share or load_balance_fair.
8269 tg->shares = shares;
8270 for_each_possible_cpu(i) {
8274 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8275 set_se_shares(tg->se[i], shares);
8279 * Enable load balance activity on this group, by inserting it back on
8280 * each cpu's rq->leaf_cfs_rq_list.
8282 spin_lock_irqsave(&task_group_lock, flags);
8283 for_each_possible_cpu(i)
8284 register_fair_sched_group(tg, i);
8285 list_add_rcu(&tg->siblings, &tg->parent->children);
8286 spin_unlock_irqrestore(&task_group_lock, flags);
8288 mutex_unlock(&shares_mutex);
8292 unsigned long sched_group_shares(struct task_group *tg)
8298 #ifdef CONFIG_RT_GROUP_SCHED
8300 * Ensure that the real time constraints are schedulable.
8302 static DEFINE_MUTEX(rt_constraints_mutex);
8304 static unsigned long to_ratio(u64 period, u64 runtime)
8306 if (runtime == RUNTIME_INF)
8309 return div64_u64(runtime << 20, period);
8312 /* Must be called with tasklist_lock held */
8313 static inline int tg_has_rt_tasks(struct task_group *tg)
8315 struct task_struct *g, *p;
8317 do_each_thread(g, p) {
8318 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8320 } while_each_thread(g, p);
8325 struct rt_schedulable_data {
8326 struct task_group *tg;
8331 static int tg_schedulable(struct task_group *tg, void *data)
8333 struct rt_schedulable_data *d = data;
8334 struct task_group *child;
8335 unsigned long total, sum = 0;
8336 u64 period, runtime;
8338 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8339 runtime = tg->rt_bandwidth.rt_runtime;
8342 period = d->rt_period;
8343 runtime = d->rt_runtime;
8347 * Cannot have more runtime than the period.
8349 if (runtime > period && runtime != RUNTIME_INF)
8353 * Ensure we don't starve existing RT tasks.
8355 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8358 total = to_ratio(period, runtime);
8361 * Nobody can have more than the global setting allows.
8363 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8367 * The sum of our children's runtime should not exceed our own.
8369 list_for_each_entry_rcu(child, &tg->children, siblings) {
8370 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8371 runtime = child->rt_bandwidth.rt_runtime;
8373 if (child == d->tg) {
8374 period = d->rt_period;
8375 runtime = d->rt_runtime;
8378 sum += to_ratio(period, runtime);
8387 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8389 struct rt_schedulable_data data = {
8391 .rt_period = period,
8392 .rt_runtime = runtime,
8395 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8398 static int tg_set_bandwidth(struct task_group *tg,
8399 u64 rt_period, u64 rt_runtime)
8403 mutex_lock(&rt_constraints_mutex);
8404 read_lock(&tasklist_lock);
8405 err = __rt_schedulable(tg, rt_period, rt_runtime);
8409 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8410 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8411 tg->rt_bandwidth.rt_runtime = rt_runtime;
8413 for_each_possible_cpu(i) {
8414 struct rt_rq *rt_rq = tg->rt_rq[i];
8416 raw_spin_lock(&rt_rq->rt_runtime_lock);
8417 rt_rq->rt_runtime = rt_runtime;
8418 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8420 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8422 read_unlock(&tasklist_lock);
8423 mutex_unlock(&rt_constraints_mutex);
8428 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8430 u64 rt_runtime, rt_period;
8432 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8433 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8434 if (rt_runtime_us < 0)
8435 rt_runtime = RUNTIME_INF;
8437 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8440 long sched_group_rt_runtime(struct task_group *tg)
8444 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8447 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8448 do_div(rt_runtime_us, NSEC_PER_USEC);
8449 return rt_runtime_us;
8452 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8454 u64 rt_runtime, rt_period;
8456 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8457 rt_runtime = tg->rt_bandwidth.rt_runtime;
8462 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8465 long sched_group_rt_period(struct task_group *tg)
8469 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8470 do_div(rt_period_us, NSEC_PER_USEC);
8471 return rt_period_us;
8474 static int sched_rt_global_constraints(void)
8476 u64 runtime, period;
8479 if (sysctl_sched_rt_period <= 0)
8482 runtime = global_rt_runtime();
8483 period = global_rt_period();
8486 * Sanity check on the sysctl variables.
8488 if (runtime > period && runtime != RUNTIME_INF)
8491 mutex_lock(&rt_constraints_mutex);
8492 read_lock(&tasklist_lock);
8493 ret = __rt_schedulable(NULL, 0, 0);
8494 read_unlock(&tasklist_lock);
8495 mutex_unlock(&rt_constraints_mutex);
8500 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8502 /* Don't accept realtime tasks when there is no way for them to run */
8503 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8509 #else /* !CONFIG_RT_GROUP_SCHED */
8510 static int sched_rt_global_constraints(void)
8512 unsigned long flags;
8515 if (sysctl_sched_rt_period <= 0)
8519 * There's always some RT tasks in the root group
8520 * -- migration, kstopmachine etc..
8522 if (sysctl_sched_rt_runtime == 0)
8525 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8526 for_each_possible_cpu(i) {
8527 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8529 raw_spin_lock(&rt_rq->rt_runtime_lock);
8530 rt_rq->rt_runtime = global_rt_runtime();
8531 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8533 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8537 #endif /* CONFIG_RT_GROUP_SCHED */
8539 int sched_rt_handler(struct ctl_table *table, int write,
8540 void __user *buffer, size_t *lenp,
8544 int old_period, old_runtime;
8545 static DEFINE_MUTEX(mutex);
8548 old_period = sysctl_sched_rt_period;
8549 old_runtime = sysctl_sched_rt_runtime;
8551 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8553 if (!ret && write) {
8554 ret = sched_rt_global_constraints();
8556 sysctl_sched_rt_period = old_period;
8557 sysctl_sched_rt_runtime = old_runtime;
8559 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8560 def_rt_bandwidth.rt_period =
8561 ns_to_ktime(global_rt_period());
8564 mutex_unlock(&mutex);
8569 #ifdef CONFIG_CGROUP_SCHED
8571 /* return corresponding task_group object of a cgroup */
8572 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8574 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8575 struct task_group, css);
8578 static struct cgroup_subsys_state *
8579 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8581 struct task_group *tg, *parent;
8583 if (!cgrp->parent) {
8584 /* This is early initialization for the top cgroup */
8585 return &init_task_group.css;
8588 parent = cgroup_tg(cgrp->parent);
8589 tg = sched_create_group(parent);
8591 return ERR_PTR(-ENOMEM);
8597 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8599 struct task_group *tg = cgroup_tg(cgrp);
8601 sched_destroy_group(tg);
8605 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8607 #ifdef CONFIG_RT_GROUP_SCHED
8608 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8611 /* We don't support RT-tasks being in separate groups */
8612 if (tsk->sched_class != &fair_sched_class)
8619 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8620 struct task_struct *tsk, bool threadgroup)
8622 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8626 struct task_struct *c;
8628 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8629 retval = cpu_cgroup_can_attach_task(cgrp, c);
8641 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8642 struct cgroup *old_cont, struct task_struct *tsk,
8645 sched_move_task(tsk);
8647 struct task_struct *c;
8649 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8656 #ifdef CONFIG_FAIR_GROUP_SCHED
8657 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8660 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8663 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8665 struct task_group *tg = cgroup_tg(cgrp);
8667 return (u64) tg->shares;
8669 #endif /* CONFIG_FAIR_GROUP_SCHED */
8671 #ifdef CONFIG_RT_GROUP_SCHED
8672 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8675 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8678 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8680 return sched_group_rt_runtime(cgroup_tg(cgrp));
8683 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8686 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8689 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8691 return sched_group_rt_period(cgroup_tg(cgrp));
8693 #endif /* CONFIG_RT_GROUP_SCHED */
8695 static struct cftype cpu_files[] = {
8696 #ifdef CONFIG_FAIR_GROUP_SCHED
8699 .read_u64 = cpu_shares_read_u64,
8700 .write_u64 = cpu_shares_write_u64,
8703 #ifdef CONFIG_RT_GROUP_SCHED
8705 .name = "rt_runtime_us",
8706 .read_s64 = cpu_rt_runtime_read,
8707 .write_s64 = cpu_rt_runtime_write,
8710 .name = "rt_period_us",
8711 .read_u64 = cpu_rt_period_read_uint,
8712 .write_u64 = cpu_rt_period_write_uint,
8717 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8719 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8722 struct cgroup_subsys cpu_cgroup_subsys = {
8724 .create = cpu_cgroup_create,
8725 .destroy = cpu_cgroup_destroy,
8726 .can_attach = cpu_cgroup_can_attach,
8727 .attach = cpu_cgroup_attach,
8728 .populate = cpu_cgroup_populate,
8729 .subsys_id = cpu_cgroup_subsys_id,
8733 #endif /* CONFIG_CGROUP_SCHED */
8735 #ifdef CONFIG_CGROUP_CPUACCT
8738 * CPU accounting code for task groups.
8740 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8741 * (balbir@in.ibm.com).
8744 /* track cpu usage of a group of tasks and its child groups */
8746 struct cgroup_subsys_state css;
8747 /* cpuusage holds pointer to a u64-type object on every cpu */
8749 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8750 struct cpuacct *parent;
8753 struct cgroup_subsys cpuacct_subsys;
8755 /* return cpu accounting group corresponding to this container */
8756 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8758 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8759 struct cpuacct, css);
8762 /* return cpu accounting group to which this task belongs */
8763 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8765 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8766 struct cpuacct, css);
8769 /* create a new cpu accounting group */
8770 static struct cgroup_subsys_state *cpuacct_create(
8771 struct cgroup_subsys *ss, struct cgroup *cgrp)
8773 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8779 ca->cpuusage = alloc_percpu(u64);
8783 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8784 if (percpu_counter_init(&ca->cpustat[i], 0))
8785 goto out_free_counters;
8788 ca->parent = cgroup_ca(cgrp->parent);
8794 percpu_counter_destroy(&ca->cpustat[i]);
8795 free_percpu(ca->cpuusage);
8799 return ERR_PTR(-ENOMEM);
8802 /* destroy an existing cpu accounting group */
8804 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8806 struct cpuacct *ca = cgroup_ca(cgrp);
8809 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8810 percpu_counter_destroy(&ca->cpustat[i]);
8811 free_percpu(ca->cpuusage);
8815 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8817 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8820 #ifndef CONFIG_64BIT
8822 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8824 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8826 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8834 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8836 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8838 #ifndef CONFIG_64BIT
8840 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8842 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8844 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8850 /* return total cpu usage (in nanoseconds) of a group */
8851 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8853 struct cpuacct *ca = cgroup_ca(cgrp);
8854 u64 totalcpuusage = 0;
8857 for_each_present_cpu(i)
8858 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8860 return totalcpuusage;
8863 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8866 struct cpuacct *ca = cgroup_ca(cgrp);
8875 for_each_present_cpu(i)
8876 cpuacct_cpuusage_write(ca, i, 0);
8882 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8885 struct cpuacct *ca = cgroup_ca(cgroup);
8889 for_each_present_cpu(i) {
8890 percpu = cpuacct_cpuusage_read(ca, i);
8891 seq_printf(m, "%llu ", (unsigned long long) percpu);
8893 seq_printf(m, "\n");
8897 static const char *cpuacct_stat_desc[] = {
8898 [CPUACCT_STAT_USER] = "user",
8899 [CPUACCT_STAT_SYSTEM] = "system",
8902 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8903 struct cgroup_map_cb *cb)
8905 struct cpuacct *ca = cgroup_ca(cgrp);
8908 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8909 s64 val = percpu_counter_read(&ca->cpustat[i]);
8910 val = cputime64_to_clock_t(val);
8911 cb->fill(cb, cpuacct_stat_desc[i], val);
8916 static struct cftype files[] = {
8919 .read_u64 = cpuusage_read,
8920 .write_u64 = cpuusage_write,
8923 .name = "usage_percpu",
8924 .read_seq_string = cpuacct_percpu_seq_read,
8928 .read_map = cpuacct_stats_show,
8932 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8934 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8938 * charge this task's execution time to its accounting group.
8940 * called with rq->lock held.
8942 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8947 if (unlikely(!cpuacct_subsys.active))
8950 cpu = task_cpu(tsk);
8956 for (; ca; ca = ca->parent) {
8957 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8958 *cpuusage += cputime;
8965 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8966 * in cputime_t units. As a result, cpuacct_update_stats calls
8967 * percpu_counter_add with values large enough to always overflow the
8968 * per cpu batch limit causing bad SMP scalability.
8970 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8971 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8972 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8975 #define CPUACCT_BATCH \
8976 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8978 #define CPUACCT_BATCH 0
8982 * Charge the system/user time to the task's accounting group.
8984 static void cpuacct_update_stats(struct task_struct *tsk,
8985 enum cpuacct_stat_index idx, cputime_t val)
8988 int batch = CPUACCT_BATCH;
8990 if (unlikely(!cpuacct_subsys.active))
8997 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9003 struct cgroup_subsys cpuacct_subsys = {
9005 .create = cpuacct_create,
9006 .destroy = cpuacct_destroy,
9007 .populate = cpuacct_populate,
9008 .subsys_id = cpuacct_subsys_id,
9010 #endif /* CONFIG_CGROUP_CPUACCT */
9014 int rcu_expedited_torture_stats(char *page)
9018 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9020 void synchronize_sched_expedited(void)
9023 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9025 #else /* #ifndef CONFIG_SMP */
9027 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
9028 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
9030 #define RCU_EXPEDITED_STATE_POST -2
9031 #define RCU_EXPEDITED_STATE_IDLE -1
9033 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9035 int rcu_expedited_torture_stats(char *page)
9040 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
9041 for_each_online_cpu(cpu) {
9042 cnt += sprintf(&page[cnt], " %d:%d",
9043 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
9045 cnt += sprintf(&page[cnt], "\n");
9048 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
9050 static long synchronize_sched_expedited_count;
9053 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9054 * approach to force grace period to end quickly. This consumes
9055 * significant time on all CPUs, and is thus not recommended for
9056 * any sort of common-case code.
9058 * Note that it is illegal to call this function while holding any
9059 * lock that is acquired by a CPU-hotplug notifier. Failing to
9060 * observe this restriction will result in deadlock.
9062 void synchronize_sched_expedited(void)
9065 unsigned long flags;
9066 bool need_full_sync = 0;
9068 struct migration_req *req;
9072 smp_mb(); /* ensure prior mod happens before capturing snap. */
9073 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
9075 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
9077 if (trycount++ < 10)
9078 udelay(trycount * num_online_cpus());
9080 synchronize_sched();
9083 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
9084 smp_mb(); /* ensure test happens before caller kfree */
9089 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
9090 for_each_online_cpu(cpu) {
9092 req = &per_cpu(rcu_migration_req, cpu);
9093 init_completion(&req->done);
9095 req->dest_cpu = RCU_MIGRATION_NEED_QS;
9096 raw_spin_lock_irqsave(&rq->lock, flags);
9097 list_add(&req->list, &rq->migration_queue);
9098 raw_spin_unlock_irqrestore(&rq->lock, flags);
9099 wake_up_process(rq->migration_thread);
9101 for_each_online_cpu(cpu) {
9102 rcu_expedited_state = cpu;
9103 req = &per_cpu(rcu_migration_req, cpu);
9105 wait_for_completion(&req->done);
9106 raw_spin_lock_irqsave(&rq->lock, flags);
9107 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
9109 req->dest_cpu = RCU_MIGRATION_IDLE;
9110 raw_spin_unlock_irqrestore(&rq->lock, flags);
9112 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
9113 synchronize_sched_expedited_count++;
9114 mutex_unlock(&rcu_sched_expedited_mutex);
9117 synchronize_sched();
9119 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9121 #endif /* #else #ifndef CONFIG_SMP */