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/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/sched.h>
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
124 static inline int rt_policy(int policy)
126 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
131 static inline int task_has_rt_policy(struct task_struct *p)
133 return rt_policy(p->policy);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array {
140 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
141 struct list_head queue[MAX_RT_PRIO];
144 struct rt_bandwidth {
145 /* nests inside the rq lock: */
146 raw_spinlock_t rt_runtime_lock;
149 struct hrtimer rt_period_timer;
152 static struct rt_bandwidth def_rt_bandwidth;
154 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
156 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
158 struct rt_bandwidth *rt_b =
159 container_of(timer, struct rt_bandwidth, rt_period_timer);
165 now = hrtimer_cb_get_time(timer);
166 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
171 idle = do_sched_rt_period_timer(rt_b, overrun);
174 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
180 rt_b->rt_period = ns_to_ktime(period);
181 rt_b->rt_runtime = runtime;
183 raw_spin_lock_init(&rt_b->rt_runtime_lock);
185 hrtimer_init(&rt_b->rt_period_timer,
186 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
187 rt_b->rt_period_timer.function = sched_rt_period_timer;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime >= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
199 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
202 if (hrtimer_active(&rt_b->rt_period_timer))
205 raw_spin_lock(&rt_b->rt_runtime_lock);
210 if (hrtimer_active(&rt_b->rt_period_timer))
213 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
214 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
216 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
217 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
218 delta = ktime_to_ns(ktime_sub(hard, soft));
219 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
220 HRTIMER_MODE_ABS_PINNED, 0);
222 raw_spin_unlock(&rt_b->rt_runtime_lock);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
228 hrtimer_cancel(&rt_b->rt_period_timer);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex);
238 #ifdef CONFIG_CGROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups);
246 /* task group related information */
248 struct cgroup_subsys_state css;
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 /* schedulable entities of this group on each cpu */
252 struct sched_entity **se;
253 /* runqueue "owned" by this group on each cpu */
254 struct cfs_rq **cfs_rq;
255 unsigned long shares;
257 atomic_t load_weight;
260 #ifdef CONFIG_RT_GROUP_SCHED
261 struct sched_rt_entity **rt_se;
262 struct rt_rq **rt_rq;
264 struct rt_bandwidth rt_bandwidth;
268 struct list_head list;
270 struct task_group *parent;
271 struct list_head siblings;
272 struct list_head children;
275 #define root_task_group init_task_group
277 /* task_group_lock serializes the addition/removal of task groups */
278 static DEFINE_SPINLOCK(task_group_lock);
280 #ifdef CONFIG_FAIR_GROUP_SCHED
282 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
285 * A weight of 0 or 1 can cause arithmetics problems.
286 * A weight of a cfs_rq is the sum of weights of which entities
287 * are queued on this cfs_rq, so a weight of a entity should not be
288 * too large, so as the shares value of a task group.
289 * (The default weight is 1024 - so there's no practical
290 * limitation from this.)
293 #define MAX_SHARES (1UL << 18)
295 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
298 /* Default task group.
299 * Every task in system belong to this group at bootup.
301 struct task_group init_task_group;
303 #endif /* CONFIG_CGROUP_SCHED */
305 /* CFS-related fields in a runqueue */
307 struct load_weight load;
308 unsigned long nr_running;
313 struct rb_root tasks_timeline;
314 struct rb_node *rb_leftmost;
316 struct list_head tasks;
317 struct list_head *balance_iterator;
320 * 'curr' points to currently running entity on this cfs_rq.
321 * It is set to NULL otherwise (i.e when none are currently running).
323 struct sched_entity *curr, *next, *last;
325 unsigned int nr_spread_over;
327 #ifdef CONFIG_FAIR_GROUP_SCHED
328 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
331 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
332 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
333 * (like users, containers etc.)
335 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
336 * list is used during load balance.
339 struct list_head leaf_cfs_rq_list;
340 struct task_group *tg; /* group that "owns" this runqueue */
344 * the part of load.weight contributed by tasks
346 unsigned long task_weight;
349 * h_load = weight * f(tg)
351 * Where f(tg) is the recursive weight fraction assigned to
354 unsigned long h_load;
357 * Maintaining per-cpu shares distribution for group scheduling
359 * load_stamp is the last time we updated the load average
360 * load_last is the last time we updated the load average and saw load
361 * load_unacc_exec_time is currently unaccounted execution time
365 u64 load_stamp, load_last, load_unacc_exec_time;
367 unsigned long load_contribution;
372 /* Real-Time classes' related field in a runqueue: */
374 struct rt_prio_array active;
375 unsigned long rt_nr_running;
376 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
378 int curr; /* highest queued rt task prio */
380 int next; /* next highest */
385 unsigned long rt_nr_migratory;
386 unsigned long rt_nr_total;
388 struct plist_head pushable_tasks;
393 /* Nests inside the rq lock: */
394 raw_spinlock_t rt_runtime_lock;
396 #ifdef CONFIG_RT_GROUP_SCHED
397 unsigned long rt_nr_boosted;
400 struct list_head leaf_rt_rq_list;
401 struct task_group *tg;
408 * We add the notion of a root-domain which will be used to define per-domain
409 * variables. Each exclusive cpuset essentially defines an island domain by
410 * fully partitioning the member cpus from any other cpuset. Whenever a new
411 * exclusive cpuset is created, we also create and attach a new root-domain
418 cpumask_var_t online;
421 * The "RT overload" flag: it gets set if a CPU has more than
422 * one runnable RT task.
424 cpumask_var_t rto_mask;
426 struct cpupri cpupri;
430 * By default the system creates a single root-domain with all cpus as
431 * members (mimicking the global state we have today).
433 static struct root_domain def_root_domain;
435 #endif /* CONFIG_SMP */
438 * This is the main, per-CPU runqueue data structure.
440 * Locking rule: those places that want to lock multiple runqueues
441 * (such as the load balancing or the thread migration code), lock
442 * acquire operations must be ordered by ascending &runqueue.
449 * nr_running and cpu_load should be in the same cacheline because
450 * remote CPUs use both these fields when doing load calculation.
452 unsigned long nr_running;
453 #define CPU_LOAD_IDX_MAX 5
454 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
455 unsigned long last_load_update_tick;
458 unsigned char nohz_balance_kick;
460 unsigned int skip_clock_update;
462 /* capture load from *all* tasks on this cpu: */
463 struct load_weight load;
464 unsigned long nr_load_updates;
470 #ifdef CONFIG_FAIR_GROUP_SCHED
471 /* list of leaf cfs_rq on this cpu: */
472 struct list_head leaf_cfs_rq_list;
474 #ifdef CONFIG_RT_GROUP_SCHED
475 struct list_head leaf_rt_rq_list;
479 * This is part of a global counter where only the total sum
480 * over all CPUs matters. A task can increase this counter on
481 * one CPU and if it got migrated afterwards it may decrease
482 * it on another CPU. Always updated under the runqueue lock:
484 unsigned long nr_uninterruptible;
486 struct task_struct *curr, *idle, *stop;
487 unsigned long next_balance;
488 struct mm_struct *prev_mm;
496 struct root_domain *rd;
497 struct sched_domain *sd;
499 unsigned long cpu_power;
501 unsigned char idle_at_tick;
502 /* For active balancing */
506 struct cpu_stop_work active_balance_work;
507 /* cpu of this runqueue: */
511 unsigned long avg_load_per_task;
519 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
523 /* calc_load related fields */
524 unsigned long calc_load_update;
525 long calc_load_active;
527 #ifdef CONFIG_SCHED_HRTICK
529 int hrtick_csd_pending;
530 struct call_single_data hrtick_csd;
532 struct hrtimer hrtick_timer;
535 #ifdef CONFIG_SCHEDSTATS
537 struct sched_info rq_sched_info;
538 unsigned long long rq_cpu_time;
539 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
541 /* sys_sched_yield() stats */
542 unsigned int yld_count;
544 /* schedule() stats */
545 unsigned int sched_switch;
546 unsigned int sched_count;
547 unsigned int sched_goidle;
549 /* try_to_wake_up() stats */
550 unsigned int ttwu_count;
551 unsigned int ttwu_local;
554 unsigned int bkl_count;
558 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
561 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
563 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
566 * A queue event has occurred, and we're going to schedule. In
567 * this case, we can save a useless back to back clock update.
569 if (test_tsk_need_resched(p))
570 rq->skip_clock_update = 1;
573 static inline int cpu_of(struct rq *rq)
582 #define rcu_dereference_check_sched_domain(p) \
583 rcu_dereference_check((p), \
584 rcu_read_lock_sched_held() || \
585 lockdep_is_held(&sched_domains_mutex))
588 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
589 * See detach_destroy_domains: synchronize_sched for details.
591 * The domain tree of any CPU may only be accessed from within
592 * preempt-disabled sections.
594 #define for_each_domain(cpu, __sd) \
595 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
597 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
598 #define this_rq() (&__get_cpu_var(runqueues))
599 #define task_rq(p) cpu_rq(task_cpu(p))
600 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
601 #define raw_rq() (&__raw_get_cpu_var(runqueues))
603 #ifdef CONFIG_CGROUP_SCHED
606 * Return the group to which this tasks belongs.
608 * We use task_subsys_state_check() and extend the RCU verification
609 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
610 * holds that lock for each task it moves into the cgroup. Therefore
611 * by holding that lock, we pin the task to the current cgroup.
613 static inline struct task_group *task_group(struct task_struct *p)
615 struct cgroup_subsys_state *css;
617 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
618 lockdep_is_held(&task_rq(p)->lock));
619 return container_of(css, struct task_group, css);
622 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
623 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
625 #ifdef CONFIG_FAIR_GROUP_SCHED
626 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
627 p->se.parent = task_group(p)->se[cpu];
630 #ifdef CONFIG_RT_GROUP_SCHED
631 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
632 p->rt.parent = task_group(p)->rt_se[cpu];
636 #else /* CONFIG_CGROUP_SCHED */
638 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
639 static inline struct task_group *task_group(struct task_struct *p)
644 #endif /* CONFIG_CGROUP_SCHED */
646 static u64 irq_time_cpu(int cpu);
647 static void sched_irq_time_avg_update(struct rq *rq, u64 irq_time);
649 inline void update_rq_clock(struct rq *rq)
651 if (!rq->skip_clock_update) {
652 int cpu = cpu_of(rq);
655 rq->clock = sched_clock_cpu(cpu);
656 irq_time = irq_time_cpu(cpu);
657 if (rq->clock - irq_time > rq->clock_task)
658 rq->clock_task = rq->clock - irq_time;
660 sched_irq_time_avg_update(rq, irq_time);
665 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
667 #ifdef CONFIG_SCHED_DEBUG
668 # define const_debug __read_mostly
670 # define const_debug static const
675 * @cpu: the processor in question.
677 * Returns true if the current cpu runqueue is locked.
678 * This interface allows printk to be called with the runqueue lock
679 * held and know whether or not it is OK to wake up the klogd.
681 int runqueue_is_locked(int cpu)
683 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
687 * Debugging: various feature bits
690 #define SCHED_FEAT(name, enabled) \
691 __SCHED_FEAT_##name ,
694 #include "sched_features.h"
699 #define SCHED_FEAT(name, enabled) \
700 (1UL << __SCHED_FEAT_##name) * enabled |
702 const_debug unsigned int sysctl_sched_features =
703 #include "sched_features.h"
708 #ifdef CONFIG_SCHED_DEBUG
709 #define SCHED_FEAT(name, enabled) \
712 static __read_mostly char *sched_feat_names[] = {
713 #include "sched_features.h"
719 static int sched_feat_show(struct seq_file *m, void *v)
723 for (i = 0; sched_feat_names[i]; i++) {
724 if (!(sysctl_sched_features & (1UL << i)))
726 seq_printf(m, "%s ", sched_feat_names[i]);
734 sched_feat_write(struct file *filp, const char __user *ubuf,
735 size_t cnt, loff_t *ppos)
745 if (copy_from_user(&buf, ubuf, cnt))
751 if (strncmp(buf, "NO_", 3) == 0) {
756 for (i = 0; sched_feat_names[i]; i++) {
757 if (strcmp(cmp, sched_feat_names[i]) == 0) {
759 sysctl_sched_features &= ~(1UL << i);
761 sysctl_sched_features |= (1UL << i);
766 if (!sched_feat_names[i])
774 static int sched_feat_open(struct inode *inode, struct file *filp)
776 return single_open(filp, sched_feat_show, NULL);
779 static const struct file_operations sched_feat_fops = {
780 .open = sched_feat_open,
781 .write = sched_feat_write,
784 .release = single_release,
787 static __init int sched_init_debug(void)
789 debugfs_create_file("sched_features", 0644, NULL, NULL,
794 late_initcall(sched_init_debug);
798 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
801 * Number of tasks to iterate in a single balance run.
802 * Limited because this is done with IRQs disabled.
804 const_debug unsigned int sysctl_sched_nr_migrate = 32;
807 * period over which we average the RT time consumption, measured
812 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
815 * period over which we measure -rt task cpu usage in us.
818 unsigned int sysctl_sched_rt_period = 1000000;
820 static __read_mostly int scheduler_running;
823 * part of the period that we allow rt tasks to run in us.
826 int sysctl_sched_rt_runtime = 950000;
828 static inline u64 global_rt_period(void)
830 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
833 static inline u64 global_rt_runtime(void)
835 if (sysctl_sched_rt_runtime < 0)
838 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
841 #ifndef prepare_arch_switch
842 # define prepare_arch_switch(next) do { } while (0)
844 #ifndef finish_arch_switch
845 # define finish_arch_switch(prev) do { } while (0)
848 static inline int task_current(struct rq *rq, struct task_struct *p)
850 return rq->curr == p;
853 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
854 static inline int task_running(struct rq *rq, struct task_struct *p)
856 return task_current(rq, p);
859 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
863 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
865 #ifdef CONFIG_DEBUG_SPINLOCK
866 /* this is a valid case when another task releases the spinlock */
867 rq->lock.owner = current;
870 * If we are tracking spinlock dependencies then we have to
871 * fix up the runqueue lock - which gets 'carried over' from
874 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
876 raw_spin_unlock_irq(&rq->lock);
879 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
880 static inline int task_running(struct rq *rq, struct task_struct *p)
885 return task_current(rq, p);
889 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
893 * We can optimise this out completely for !SMP, because the
894 * SMP rebalancing from interrupt is the only thing that cares
899 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
900 raw_spin_unlock_irq(&rq->lock);
902 raw_spin_unlock(&rq->lock);
906 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
910 * After ->oncpu is cleared, the task can be moved to a different CPU.
911 * We must ensure this doesn't happen until the switch is completely
917 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
924 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
927 static inline int task_is_waking(struct task_struct *p)
929 return unlikely(p->state == TASK_WAKING);
933 * __task_rq_lock - lock the runqueue a given task resides on.
934 * Must be called interrupts disabled.
936 static inline struct rq *__task_rq_lock(struct task_struct *p)
943 raw_spin_lock(&rq->lock);
944 if (likely(rq == task_rq(p)))
946 raw_spin_unlock(&rq->lock);
951 * task_rq_lock - lock the runqueue a given task resides on and disable
952 * interrupts. Note the ordering: we can safely lookup the task_rq without
953 * explicitly disabling preemption.
955 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
961 local_irq_save(*flags);
963 raw_spin_lock(&rq->lock);
964 if (likely(rq == task_rq(p)))
966 raw_spin_unlock_irqrestore(&rq->lock, *flags);
970 static void __task_rq_unlock(struct rq *rq)
973 raw_spin_unlock(&rq->lock);
976 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
979 raw_spin_unlock_irqrestore(&rq->lock, *flags);
983 * this_rq_lock - lock this runqueue and disable interrupts.
985 static struct rq *this_rq_lock(void)
992 raw_spin_lock(&rq->lock);
997 #ifdef CONFIG_SCHED_HRTICK
999 * Use HR-timers to deliver accurate preemption points.
1001 * Its all a bit involved since we cannot program an hrt while holding the
1002 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1005 * When we get rescheduled we reprogram the hrtick_timer outside of the
1011 * - enabled by features
1012 * - hrtimer is actually high res
1014 static inline int hrtick_enabled(struct rq *rq)
1016 if (!sched_feat(HRTICK))
1018 if (!cpu_active(cpu_of(rq)))
1020 return hrtimer_is_hres_active(&rq->hrtick_timer);
1023 static void hrtick_clear(struct rq *rq)
1025 if (hrtimer_active(&rq->hrtick_timer))
1026 hrtimer_cancel(&rq->hrtick_timer);
1030 * High-resolution timer tick.
1031 * Runs from hardirq context with interrupts disabled.
1033 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1035 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1037 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1039 raw_spin_lock(&rq->lock);
1040 update_rq_clock(rq);
1041 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1042 raw_spin_unlock(&rq->lock);
1044 return HRTIMER_NORESTART;
1049 * called from hardirq (IPI) context
1051 static void __hrtick_start(void *arg)
1053 struct rq *rq = arg;
1055 raw_spin_lock(&rq->lock);
1056 hrtimer_restart(&rq->hrtick_timer);
1057 rq->hrtick_csd_pending = 0;
1058 raw_spin_unlock(&rq->lock);
1062 * Called to set the hrtick timer state.
1064 * called with rq->lock held and irqs disabled
1066 static void hrtick_start(struct rq *rq, u64 delay)
1068 struct hrtimer *timer = &rq->hrtick_timer;
1069 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1071 hrtimer_set_expires(timer, time);
1073 if (rq == this_rq()) {
1074 hrtimer_restart(timer);
1075 } else if (!rq->hrtick_csd_pending) {
1076 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1077 rq->hrtick_csd_pending = 1;
1082 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1084 int cpu = (int)(long)hcpu;
1087 case CPU_UP_CANCELED:
1088 case CPU_UP_CANCELED_FROZEN:
1089 case CPU_DOWN_PREPARE:
1090 case CPU_DOWN_PREPARE_FROZEN:
1092 case CPU_DEAD_FROZEN:
1093 hrtick_clear(cpu_rq(cpu));
1100 static __init void init_hrtick(void)
1102 hotcpu_notifier(hotplug_hrtick, 0);
1106 * Called to set the hrtick timer state.
1108 * called with rq->lock held and irqs disabled
1110 static void hrtick_start(struct rq *rq, u64 delay)
1112 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1113 HRTIMER_MODE_REL_PINNED, 0);
1116 static inline void init_hrtick(void)
1119 #endif /* CONFIG_SMP */
1121 static void init_rq_hrtick(struct rq *rq)
1124 rq->hrtick_csd_pending = 0;
1126 rq->hrtick_csd.flags = 0;
1127 rq->hrtick_csd.func = __hrtick_start;
1128 rq->hrtick_csd.info = rq;
1131 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1132 rq->hrtick_timer.function = hrtick;
1134 #else /* CONFIG_SCHED_HRTICK */
1135 static inline void hrtick_clear(struct rq *rq)
1139 static inline void init_rq_hrtick(struct rq *rq)
1143 static inline void init_hrtick(void)
1146 #endif /* CONFIG_SCHED_HRTICK */
1149 * resched_task - mark a task 'to be rescheduled now'.
1151 * On UP this means the setting of the need_resched flag, on SMP it
1152 * might also involve a cross-CPU call to trigger the scheduler on
1157 #ifndef tsk_is_polling
1158 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1161 static void resched_task(struct task_struct *p)
1165 assert_raw_spin_locked(&task_rq(p)->lock);
1167 if (test_tsk_need_resched(p))
1170 set_tsk_need_resched(p);
1173 if (cpu == smp_processor_id())
1176 /* NEED_RESCHED must be visible before we test polling */
1178 if (!tsk_is_polling(p))
1179 smp_send_reschedule(cpu);
1182 static void resched_cpu(int cpu)
1184 struct rq *rq = cpu_rq(cpu);
1185 unsigned long flags;
1187 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1189 resched_task(cpu_curr(cpu));
1190 raw_spin_unlock_irqrestore(&rq->lock, flags);
1195 * In the semi idle case, use the nearest busy cpu for migrating timers
1196 * from an idle cpu. This is good for power-savings.
1198 * We don't do similar optimization for completely idle system, as
1199 * selecting an idle cpu will add more delays to the timers than intended
1200 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1202 int get_nohz_timer_target(void)
1204 int cpu = smp_processor_id();
1206 struct sched_domain *sd;
1208 for_each_domain(cpu, sd) {
1209 for_each_cpu(i, sched_domain_span(sd))
1216 * When add_timer_on() enqueues a timer into the timer wheel of an
1217 * idle CPU then this timer might expire before the next timer event
1218 * which is scheduled to wake up that CPU. In case of a completely
1219 * idle system the next event might even be infinite time into the
1220 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1221 * leaves the inner idle loop so the newly added timer is taken into
1222 * account when the CPU goes back to idle and evaluates the timer
1223 * wheel for the next timer event.
1225 void wake_up_idle_cpu(int cpu)
1227 struct rq *rq = cpu_rq(cpu);
1229 if (cpu == smp_processor_id())
1233 * This is safe, as this function is called with the timer
1234 * wheel base lock of (cpu) held. When the CPU is on the way
1235 * to idle and has not yet set rq->curr to idle then it will
1236 * be serialized on the timer wheel base lock and take the new
1237 * timer into account automatically.
1239 if (rq->curr != rq->idle)
1243 * We can set TIF_RESCHED on the idle task of the other CPU
1244 * lockless. The worst case is that the other CPU runs the
1245 * idle task through an additional NOOP schedule()
1247 set_tsk_need_resched(rq->idle);
1249 /* NEED_RESCHED must be visible before we test polling */
1251 if (!tsk_is_polling(rq->idle))
1252 smp_send_reschedule(cpu);
1255 #endif /* CONFIG_NO_HZ */
1257 static u64 sched_avg_period(void)
1259 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1262 static void sched_avg_update(struct rq *rq)
1264 s64 period = sched_avg_period();
1266 while ((s64)(rq->clock - rq->age_stamp) > period) {
1268 * Inline assembly required to prevent the compiler
1269 * optimising this loop into a divmod call.
1270 * See __iter_div_u64_rem() for another example of this.
1272 asm("" : "+rm" (rq->age_stamp));
1273 rq->age_stamp += period;
1278 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1280 rq->rt_avg += rt_delta;
1281 sched_avg_update(rq);
1284 #else /* !CONFIG_SMP */
1285 static void resched_task(struct task_struct *p)
1287 assert_raw_spin_locked(&task_rq(p)->lock);
1288 set_tsk_need_resched(p);
1291 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1295 static void sched_avg_update(struct rq *rq)
1298 #endif /* CONFIG_SMP */
1300 #if BITS_PER_LONG == 32
1301 # define WMULT_CONST (~0UL)
1303 # define WMULT_CONST (1UL << 32)
1306 #define WMULT_SHIFT 32
1309 * Shift right and round:
1311 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1314 * delta *= weight / lw
1316 static unsigned long
1317 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1318 struct load_weight *lw)
1322 if (!lw->inv_weight) {
1323 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1326 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1330 tmp = (u64)delta_exec * weight;
1332 * Check whether we'd overflow the 64-bit multiplication:
1334 if (unlikely(tmp > WMULT_CONST))
1335 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1338 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1340 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1343 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1349 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1355 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1362 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1363 * of tasks with abnormal "nice" values across CPUs the contribution that
1364 * each task makes to its run queue's load is weighted according to its
1365 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1366 * scaled version of the new time slice allocation that they receive on time
1370 #define WEIGHT_IDLEPRIO 3
1371 #define WMULT_IDLEPRIO 1431655765
1374 * Nice levels are multiplicative, with a gentle 10% change for every
1375 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1376 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1377 * that remained on nice 0.
1379 * The "10% effect" is relative and cumulative: from _any_ nice level,
1380 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1381 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1382 * If a task goes up by ~10% and another task goes down by ~10% then
1383 * the relative distance between them is ~25%.)
1385 static const int prio_to_weight[40] = {
1386 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1387 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1388 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1389 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1390 /* 0 */ 1024, 820, 655, 526, 423,
1391 /* 5 */ 335, 272, 215, 172, 137,
1392 /* 10 */ 110, 87, 70, 56, 45,
1393 /* 15 */ 36, 29, 23, 18, 15,
1397 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1399 * In cases where the weight does not change often, we can use the
1400 * precalculated inverse to speed up arithmetics by turning divisions
1401 * into multiplications:
1403 static const u32 prio_to_wmult[40] = {
1404 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1405 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1406 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1407 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1408 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1409 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1410 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1411 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1414 /* Time spent by the tasks of the cpu accounting group executing in ... */
1415 enum cpuacct_stat_index {
1416 CPUACCT_STAT_USER, /* ... user mode */
1417 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1419 CPUACCT_STAT_NSTATS,
1422 #ifdef CONFIG_CGROUP_CPUACCT
1423 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1424 static void cpuacct_update_stats(struct task_struct *tsk,
1425 enum cpuacct_stat_index idx, cputime_t val);
1427 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1428 static inline void cpuacct_update_stats(struct task_struct *tsk,
1429 enum cpuacct_stat_index idx, cputime_t val) {}
1432 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1434 update_load_add(&rq->load, load);
1437 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1439 update_load_sub(&rq->load, load);
1442 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1443 typedef int (*tg_visitor)(struct task_group *, void *);
1446 * Iterate the full tree, calling @down when first entering a node and @up when
1447 * leaving it for the final time.
1449 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1451 struct task_group *parent, *child;
1455 parent = &root_task_group;
1457 ret = (*down)(parent, data);
1460 list_for_each_entry_rcu(child, &parent->children, siblings) {
1467 ret = (*up)(parent, data);
1472 parent = parent->parent;
1481 static int tg_nop(struct task_group *tg, void *data)
1488 /* Used instead of source_load when we know the type == 0 */
1489 static unsigned long weighted_cpuload(const int cpu)
1491 return cpu_rq(cpu)->load.weight;
1495 * Return a low guess at the load of a migration-source cpu weighted
1496 * according to the scheduling class and "nice" value.
1498 * We want to under-estimate the load of migration sources, to
1499 * balance conservatively.
1501 static unsigned long source_load(int cpu, int type)
1503 struct rq *rq = cpu_rq(cpu);
1504 unsigned long total = weighted_cpuload(cpu);
1506 if (type == 0 || !sched_feat(LB_BIAS))
1509 return min(rq->cpu_load[type-1], total);
1513 * Return a high guess at the load of a migration-target cpu weighted
1514 * according to the scheduling class and "nice" value.
1516 static unsigned long target_load(int cpu, int type)
1518 struct rq *rq = cpu_rq(cpu);
1519 unsigned long total = weighted_cpuload(cpu);
1521 if (type == 0 || !sched_feat(LB_BIAS))
1524 return max(rq->cpu_load[type-1], total);
1527 static unsigned long power_of(int cpu)
1529 return cpu_rq(cpu)->cpu_power;
1532 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1534 static unsigned long cpu_avg_load_per_task(int cpu)
1536 struct rq *rq = cpu_rq(cpu);
1537 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1540 rq->avg_load_per_task = rq->load.weight / nr_running;
1542 rq->avg_load_per_task = 0;
1544 return rq->avg_load_per_task;
1547 #ifdef CONFIG_FAIR_GROUP_SCHED
1550 * Compute the cpu's hierarchical load factor for each task group.
1551 * This needs to be done in a top-down fashion because the load of a child
1552 * group is a fraction of its parents load.
1554 static int tg_load_down(struct task_group *tg, void *data)
1557 long cpu = (long)data;
1560 load = cpu_rq(cpu)->load.weight;
1562 load = tg->parent->cfs_rq[cpu]->h_load;
1563 load *= tg->se[cpu]->load.weight;
1564 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1567 tg->cfs_rq[cpu]->h_load = load;
1572 static void update_h_load(long cpu)
1574 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1579 #ifdef CONFIG_PREEMPT
1581 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1584 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1585 * way at the expense of forcing extra atomic operations in all
1586 * invocations. This assures that the double_lock is acquired using the
1587 * same underlying policy as the spinlock_t on this architecture, which
1588 * reduces latency compared to the unfair variant below. However, it
1589 * also adds more overhead and therefore may reduce throughput.
1591 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1592 __releases(this_rq->lock)
1593 __acquires(busiest->lock)
1594 __acquires(this_rq->lock)
1596 raw_spin_unlock(&this_rq->lock);
1597 double_rq_lock(this_rq, busiest);
1604 * Unfair double_lock_balance: Optimizes throughput at the expense of
1605 * latency by eliminating extra atomic operations when the locks are
1606 * already in proper order on entry. This favors lower cpu-ids and will
1607 * grant the double lock to lower cpus over higher ids under contention,
1608 * regardless of entry order into the function.
1610 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1611 __releases(this_rq->lock)
1612 __acquires(busiest->lock)
1613 __acquires(this_rq->lock)
1617 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1618 if (busiest < this_rq) {
1619 raw_spin_unlock(&this_rq->lock);
1620 raw_spin_lock(&busiest->lock);
1621 raw_spin_lock_nested(&this_rq->lock,
1622 SINGLE_DEPTH_NESTING);
1625 raw_spin_lock_nested(&busiest->lock,
1626 SINGLE_DEPTH_NESTING);
1631 #endif /* CONFIG_PREEMPT */
1634 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1636 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1638 if (unlikely(!irqs_disabled())) {
1639 /* printk() doesn't work good under rq->lock */
1640 raw_spin_unlock(&this_rq->lock);
1644 return _double_lock_balance(this_rq, busiest);
1647 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1648 __releases(busiest->lock)
1650 raw_spin_unlock(&busiest->lock);
1651 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1655 * double_rq_lock - safely lock two runqueues
1657 * Note this does not disable interrupts like task_rq_lock,
1658 * you need to do so manually before calling.
1660 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1661 __acquires(rq1->lock)
1662 __acquires(rq2->lock)
1664 BUG_ON(!irqs_disabled());
1666 raw_spin_lock(&rq1->lock);
1667 __acquire(rq2->lock); /* Fake it out ;) */
1670 raw_spin_lock(&rq1->lock);
1671 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1673 raw_spin_lock(&rq2->lock);
1674 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1680 * double_rq_unlock - safely unlock two runqueues
1682 * Note this does not restore interrupts like task_rq_unlock,
1683 * you need to do so manually after calling.
1685 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1686 __releases(rq1->lock)
1687 __releases(rq2->lock)
1689 raw_spin_unlock(&rq1->lock);
1691 raw_spin_unlock(&rq2->lock);
1693 __release(rq2->lock);
1698 static void calc_load_account_idle(struct rq *this_rq);
1699 static void update_sysctl(void);
1700 static int get_update_sysctl_factor(void);
1701 static void update_cpu_load(struct rq *this_rq);
1703 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1705 set_task_rq(p, cpu);
1708 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1709 * successfuly executed on another CPU. We must ensure that updates of
1710 * per-task data have been completed by this moment.
1713 task_thread_info(p)->cpu = cpu;
1717 static const struct sched_class rt_sched_class;
1719 #define sched_class_highest (&stop_sched_class)
1720 #define for_each_class(class) \
1721 for (class = sched_class_highest; class; class = class->next)
1723 #include "sched_stats.h"
1725 static void inc_nr_running(struct rq *rq)
1730 static void dec_nr_running(struct rq *rq)
1735 static void set_load_weight(struct task_struct *p)
1738 * SCHED_IDLE tasks get minimal weight:
1740 if (p->policy == SCHED_IDLE) {
1741 p->se.load.weight = WEIGHT_IDLEPRIO;
1742 p->se.load.inv_weight = WMULT_IDLEPRIO;
1746 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1747 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1750 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1752 update_rq_clock(rq);
1753 sched_info_queued(p);
1754 p->sched_class->enqueue_task(rq, p, flags);
1758 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1760 update_rq_clock(rq);
1761 sched_info_dequeued(p);
1762 p->sched_class->dequeue_task(rq, p, flags);
1767 * activate_task - move a task to the runqueue.
1769 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1771 if (task_contributes_to_load(p))
1772 rq->nr_uninterruptible--;
1774 enqueue_task(rq, p, flags);
1779 * deactivate_task - remove a task from the runqueue.
1781 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1783 if (task_contributes_to_load(p))
1784 rq->nr_uninterruptible++;
1786 dequeue_task(rq, p, flags);
1790 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1793 * There are no locks covering percpu hardirq/softirq time.
1794 * They are only modified in account_system_vtime, on corresponding CPU
1795 * with interrupts disabled. So, writes are safe.
1796 * They are read and saved off onto struct rq in update_rq_clock().
1797 * This may result in other CPU reading this CPU's irq time and can
1798 * race with irq/account_system_vtime on this CPU. We would either get old
1799 * or new value (or semi updated value on 32 bit) with a side effect of
1800 * accounting a slice of irq time to wrong task when irq is in progress
1801 * while we read rq->clock. That is a worthy compromise in place of having
1802 * locks on each irq in account_system_time.
1804 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1805 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1807 static DEFINE_PER_CPU(u64, irq_start_time);
1808 static int sched_clock_irqtime;
1810 void enable_sched_clock_irqtime(void)
1812 sched_clock_irqtime = 1;
1815 void disable_sched_clock_irqtime(void)
1817 sched_clock_irqtime = 0;
1820 static u64 irq_time_cpu(int cpu)
1822 if (!sched_clock_irqtime)
1825 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1828 void account_system_vtime(struct task_struct *curr)
1830 unsigned long flags;
1834 if (!sched_clock_irqtime)
1837 local_irq_save(flags);
1839 cpu = smp_processor_id();
1840 now = sched_clock_cpu(cpu);
1841 delta = now - per_cpu(irq_start_time, cpu);
1842 per_cpu(irq_start_time, cpu) = now;
1844 * We do not account for softirq time from ksoftirqd here.
1845 * We want to continue accounting softirq time to ksoftirqd thread
1846 * in that case, so as not to confuse scheduler with a special task
1847 * that do not consume any time, but still wants to run.
1849 if (hardirq_count())
1850 per_cpu(cpu_hardirq_time, cpu) += delta;
1851 else if (in_serving_softirq() && !(curr->flags & PF_KSOFTIRQD))
1852 per_cpu(cpu_softirq_time, cpu) += delta;
1854 local_irq_restore(flags);
1856 EXPORT_SYMBOL_GPL(account_system_vtime);
1858 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time)
1860 if (sched_clock_irqtime && sched_feat(NONIRQ_POWER)) {
1861 u64 delta_irq = curr_irq_time - rq->prev_irq_time;
1862 rq->prev_irq_time = curr_irq_time;
1863 sched_rt_avg_update(rq, delta_irq);
1869 static u64 irq_time_cpu(int cpu)
1874 static void sched_irq_time_avg_update(struct rq *rq, u64 curr_irq_time) { }
1878 #include "sched_idletask.c"
1879 #include "sched_fair.c"
1880 #include "sched_rt.c"
1881 #include "sched_stoptask.c"
1882 #ifdef CONFIG_SCHED_DEBUG
1883 # include "sched_debug.c"
1886 void sched_set_stop_task(int cpu, struct task_struct *stop)
1888 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
1889 struct task_struct *old_stop = cpu_rq(cpu)->stop;
1893 * Make it appear like a SCHED_FIFO task, its something
1894 * userspace knows about and won't get confused about.
1896 * Also, it will make PI more or less work without too
1897 * much confusion -- but then, stop work should not
1898 * rely on PI working anyway.
1900 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
1902 stop->sched_class = &stop_sched_class;
1905 cpu_rq(cpu)->stop = stop;
1909 * Reset it back to a normal scheduling class so that
1910 * it can die in pieces.
1912 old_stop->sched_class = &rt_sched_class;
1917 * __normal_prio - return the priority that is based on the static prio
1919 static inline int __normal_prio(struct task_struct *p)
1921 return p->static_prio;
1925 * Calculate the expected normal priority: i.e. priority
1926 * without taking RT-inheritance into account. Might be
1927 * boosted by interactivity modifiers. Changes upon fork,
1928 * setprio syscalls, and whenever the interactivity
1929 * estimator recalculates.
1931 static inline int normal_prio(struct task_struct *p)
1935 if (task_has_rt_policy(p))
1936 prio = MAX_RT_PRIO-1 - p->rt_priority;
1938 prio = __normal_prio(p);
1943 * Calculate the current priority, i.e. the priority
1944 * taken into account by the scheduler. This value might
1945 * be boosted by RT tasks, or might be boosted by
1946 * interactivity modifiers. Will be RT if the task got
1947 * RT-boosted. If not then it returns p->normal_prio.
1949 static int effective_prio(struct task_struct *p)
1951 p->normal_prio = normal_prio(p);
1953 * If we are RT tasks or we were boosted to RT priority,
1954 * keep the priority unchanged. Otherwise, update priority
1955 * to the normal priority:
1957 if (!rt_prio(p->prio))
1958 return p->normal_prio;
1963 * task_curr - is this task currently executing on a CPU?
1964 * @p: the task in question.
1966 inline int task_curr(const struct task_struct *p)
1968 return cpu_curr(task_cpu(p)) == p;
1971 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1972 const struct sched_class *prev_class,
1973 int oldprio, int running)
1975 if (prev_class != p->sched_class) {
1976 if (prev_class->switched_from)
1977 prev_class->switched_from(rq, p, running);
1978 p->sched_class->switched_to(rq, p, running);
1980 p->sched_class->prio_changed(rq, p, oldprio, running);
1985 * Is this task likely cache-hot:
1988 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1992 if (p->sched_class != &fair_sched_class)
1995 if (unlikely(p->policy == SCHED_IDLE))
1999 * Buddy candidates are cache hot:
2001 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2002 (&p->se == cfs_rq_of(&p->se)->next ||
2003 &p->se == cfs_rq_of(&p->se)->last))
2006 if (sysctl_sched_migration_cost == -1)
2008 if (sysctl_sched_migration_cost == 0)
2011 delta = now - p->se.exec_start;
2013 return delta < (s64)sysctl_sched_migration_cost;
2016 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2018 #ifdef CONFIG_SCHED_DEBUG
2020 * We should never call set_task_cpu() on a blocked task,
2021 * ttwu() will sort out the placement.
2023 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2024 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2027 trace_sched_migrate_task(p, new_cpu);
2029 if (task_cpu(p) != new_cpu) {
2030 p->se.nr_migrations++;
2031 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2034 __set_task_cpu(p, new_cpu);
2037 struct migration_arg {
2038 struct task_struct *task;
2042 static int migration_cpu_stop(void *data);
2045 * The task's runqueue lock must be held.
2046 * Returns true if you have to wait for migration thread.
2048 static bool migrate_task(struct task_struct *p, int dest_cpu)
2050 struct rq *rq = task_rq(p);
2053 * If the task is not on a runqueue (and not running), then
2054 * the next wake-up will properly place the task.
2056 return p->se.on_rq || task_running(rq, p);
2060 * wait_task_inactive - wait for a thread to unschedule.
2062 * If @match_state is nonzero, it's the @p->state value just checked and
2063 * not expected to change. If it changes, i.e. @p might have woken up,
2064 * then return zero. When we succeed in waiting for @p to be off its CPU,
2065 * we return a positive number (its total switch count). If a second call
2066 * a short while later returns the same number, the caller can be sure that
2067 * @p has remained unscheduled the whole time.
2069 * The caller must ensure that the task *will* unschedule sometime soon,
2070 * else this function might spin for a *long* time. This function can't
2071 * be called with interrupts off, or it may introduce deadlock with
2072 * smp_call_function() if an IPI is sent by the same process we are
2073 * waiting to become inactive.
2075 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2077 unsigned long flags;
2084 * We do the initial early heuristics without holding
2085 * any task-queue locks at all. We'll only try to get
2086 * the runqueue lock when things look like they will
2092 * If the task is actively running on another CPU
2093 * still, just relax and busy-wait without holding
2096 * NOTE! Since we don't hold any locks, it's not
2097 * even sure that "rq" stays as the right runqueue!
2098 * But we don't care, since "task_running()" will
2099 * return false if the runqueue has changed and p
2100 * is actually now running somewhere else!
2102 while (task_running(rq, p)) {
2103 if (match_state && unlikely(p->state != match_state))
2109 * Ok, time to look more closely! We need the rq
2110 * lock now, to be *sure*. If we're wrong, we'll
2111 * just go back and repeat.
2113 rq = task_rq_lock(p, &flags);
2114 trace_sched_wait_task(p);
2115 running = task_running(rq, p);
2116 on_rq = p->se.on_rq;
2118 if (!match_state || p->state == match_state)
2119 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2120 task_rq_unlock(rq, &flags);
2123 * If it changed from the expected state, bail out now.
2125 if (unlikely(!ncsw))
2129 * Was it really running after all now that we
2130 * checked with the proper locks actually held?
2132 * Oops. Go back and try again..
2134 if (unlikely(running)) {
2140 * It's not enough that it's not actively running,
2141 * it must be off the runqueue _entirely_, and not
2144 * So if it was still runnable (but just not actively
2145 * running right now), it's preempted, and we should
2146 * yield - it could be a while.
2148 if (unlikely(on_rq)) {
2149 schedule_timeout_uninterruptible(1);
2154 * Ahh, all good. It wasn't running, and it wasn't
2155 * runnable, which means that it will never become
2156 * running in the future either. We're all done!
2165 * kick_process - kick a running thread to enter/exit the kernel
2166 * @p: the to-be-kicked thread
2168 * Cause a process which is running on another CPU to enter
2169 * kernel-mode, without any delay. (to get signals handled.)
2171 * NOTE: this function doesnt have to take the runqueue lock,
2172 * because all it wants to ensure is that the remote task enters
2173 * the kernel. If the IPI races and the task has been migrated
2174 * to another CPU then no harm is done and the purpose has been
2177 void kick_process(struct task_struct *p)
2183 if ((cpu != smp_processor_id()) && task_curr(p))
2184 smp_send_reschedule(cpu);
2187 EXPORT_SYMBOL_GPL(kick_process);
2188 #endif /* CONFIG_SMP */
2191 * task_oncpu_function_call - call a function on the cpu on which a task runs
2192 * @p: the task to evaluate
2193 * @func: the function to be called
2194 * @info: the function call argument
2196 * Calls the function @func when the task is currently running. This might
2197 * be on the current CPU, which just calls the function directly
2199 void task_oncpu_function_call(struct task_struct *p,
2200 void (*func) (void *info), void *info)
2207 smp_call_function_single(cpu, func, info, 1);
2213 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2215 static int select_fallback_rq(int cpu, struct task_struct *p)
2218 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2220 /* Look for allowed, online CPU in same node. */
2221 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2222 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2225 /* Any allowed, online CPU? */
2226 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2227 if (dest_cpu < nr_cpu_ids)
2230 /* No more Mr. Nice Guy. */
2231 dest_cpu = cpuset_cpus_allowed_fallback(p);
2233 * Don't tell them about moving exiting tasks or
2234 * kernel threads (both mm NULL), since they never
2237 if (p->mm && printk_ratelimit()) {
2238 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2239 task_pid_nr(p), p->comm, cpu);
2246 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2249 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2251 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2254 * In order not to call set_task_cpu() on a blocking task we need
2255 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2258 * Since this is common to all placement strategies, this lives here.
2260 * [ this allows ->select_task() to simply return task_cpu(p) and
2261 * not worry about this generic constraint ]
2263 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2265 cpu = select_fallback_rq(task_cpu(p), p);
2270 static void update_avg(u64 *avg, u64 sample)
2272 s64 diff = sample - *avg;
2277 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2278 bool is_sync, bool is_migrate, bool is_local,
2279 unsigned long en_flags)
2281 schedstat_inc(p, se.statistics.nr_wakeups);
2283 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2285 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2287 schedstat_inc(p, se.statistics.nr_wakeups_local);
2289 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2291 activate_task(rq, p, en_flags);
2294 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2295 int wake_flags, bool success)
2297 trace_sched_wakeup(p, success);
2298 check_preempt_curr(rq, p, wake_flags);
2300 p->state = TASK_RUNNING;
2302 if (p->sched_class->task_woken)
2303 p->sched_class->task_woken(rq, p);
2305 if (unlikely(rq->idle_stamp)) {
2306 u64 delta = rq->clock - rq->idle_stamp;
2307 u64 max = 2*sysctl_sched_migration_cost;
2312 update_avg(&rq->avg_idle, delta);
2316 /* if a worker is waking up, notify workqueue */
2317 if ((p->flags & PF_WQ_WORKER) && success)
2318 wq_worker_waking_up(p, cpu_of(rq));
2322 * try_to_wake_up - wake up a thread
2323 * @p: the thread to be awakened
2324 * @state: the mask of task states that can be woken
2325 * @wake_flags: wake modifier flags (WF_*)
2327 * Put it on the run-queue if it's not already there. The "current"
2328 * thread is always on the run-queue (except when the actual
2329 * re-schedule is in progress), and as such you're allowed to do
2330 * the simpler "current->state = TASK_RUNNING" to mark yourself
2331 * runnable without the overhead of this.
2333 * Returns %true if @p was woken up, %false if it was already running
2334 * or @state didn't match @p's state.
2336 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2339 int cpu, orig_cpu, this_cpu, success = 0;
2340 unsigned long flags;
2341 unsigned long en_flags = ENQUEUE_WAKEUP;
2344 this_cpu = get_cpu();
2347 rq = task_rq_lock(p, &flags);
2348 if (!(p->state & state))
2358 if (unlikely(task_running(rq, p)))
2362 * In order to handle concurrent wakeups and release the rq->lock
2363 * we put the task in TASK_WAKING state.
2365 * First fix up the nr_uninterruptible count:
2367 if (task_contributes_to_load(p)) {
2368 if (likely(cpu_online(orig_cpu)))
2369 rq->nr_uninterruptible--;
2371 this_rq()->nr_uninterruptible--;
2373 p->state = TASK_WAKING;
2375 if (p->sched_class->task_waking) {
2376 p->sched_class->task_waking(rq, p);
2377 en_flags |= ENQUEUE_WAKING;
2380 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2381 if (cpu != orig_cpu)
2382 set_task_cpu(p, cpu);
2383 __task_rq_unlock(rq);
2386 raw_spin_lock(&rq->lock);
2389 * We migrated the task without holding either rq->lock, however
2390 * since the task is not on the task list itself, nobody else
2391 * will try and migrate the task, hence the rq should match the
2392 * cpu we just moved it to.
2394 WARN_ON(task_cpu(p) != cpu);
2395 WARN_ON(p->state != TASK_WAKING);
2397 #ifdef CONFIG_SCHEDSTATS
2398 schedstat_inc(rq, ttwu_count);
2399 if (cpu == this_cpu)
2400 schedstat_inc(rq, ttwu_local);
2402 struct sched_domain *sd;
2403 for_each_domain(this_cpu, sd) {
2404 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2405 schedstat_inc(sd, ttwu_wake_remote);
2410 #endif /* CONFIG_SCHEDSTATS */
2413 #endif /* CONFIG_SMP */
2414 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2415 cpu == this_cpu, en_flags);
2418 ttwu_post_activation(p, rq, wake_flags, success);
2420 task_rq_unlock(rq, &flags);
2427 * try_to_wake_up_local - try to wake up a local task with rq lock held
2428 * @p: the thread to be awakened
2430 * Put @p on the run-queue if it's not alredy there. The caller must
2431 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2432 * the current task. this_rq() stays locked over invocation.
2434 static void try_to_wake_up_local(struct task_struct *p)
2436 struct rq *rq = task_rq(p);
2437 bool success = false;
2439 BUG_ON(rq != this_rq());
2440 BUG_ON(p == current);
2441 lockdep_assert_held(&rq->lock);
2443 if (!(p->state & TASK_NORMAL))
2447 if (likely(!task_running(rq, p))) {
2448 schedstat_inc(rq, ttwu_count);
2449 schedstat_inc(rq, ttwu_local);
2451 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2454 ttwu_post_activation(p, rq, 0, success);
2458 * wake_up_process - Wake up a specific process
2459 * @p: The process to be woken up.
2461 * Attempt to wake up the nominated process and move it to the set of runnable
2462 * processes. Returns 1 if the process was woken up, 0 if it was already
2465 * It may be assumed that this function implies a write memory barrier before
2466 * changing the task state if and only if any tasks are woken up.
2468 int wake_up_process(struct task_struct *p)
2470 return try_to_wake_up(p, TASK_ALL, 0);
2472 EXPORT_SYMBOL(wake_up_process);
2474 int wake_up_state(struct task_struct *p, unsigned int state)
2476 return try_to_wake_up(p, state, 0);
2480 * Perform scheduler related setup for a newly forked process p.
2481 * p is forked by current.
2483 * __sched_fork() is basic setup used by init_idle() too:
2485 static void __sched_fork(struct task_struct *p)
2487 p->se.exec_start = 0;
2488 p->se.sum_exec_runtime = 0;
2489 p->se.prev_sum_exec_runtime = 0;
2490 p->se.nr_migrations = 0;
2492 #ifdef CONFIG_SCHEDSTATS
2493 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2496 INIT_LIST_HEAD(&p->rt.run_list);
2498 INIT_LIST_HEAD(&p->se.group_node);
2500 #ifdef CONFIG_PREEMPT_NOTIFIERS
2501 INIT_HLIST_HEAD(&p->preempt_notifiers);
2506 * fork()/clone()-time setup:
2508 void sched_fork(struct task_struct *p, int clone_flags)
2510 int cpu = get_cpu();
2514 * We mark the process as running here. This guarantees that
2515 * nobody will actually run it, and a signal or other external
2516 * event cannot wake it up and insert it on the runqueue either.
2518 p->state = TASK_RUNNING;
2521 * Revert to default priority/policy on fork if requested.
2523 if (unlikely(p->sched_reset_on_fork)) {
2524 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2525 p->policy = SCHED_NORMAL;
2526 p->normal_prio = p->static_prio;
2529 if (PRIO_TO_NICE(p->static_prio) < 0) {
2530 p->static_prio = NICE_TO_PRIO(0);
2531 p->normal_prio = p->static_prio;
2536 * We don't need the reset flag anymore after the fork. It has
2537 * fulfilled its duty:
2539 p->sched_reset_on_fork = 0;
2543 * Make sure we do not leak PI boosting priority to the child.
2545 p->prio = current->normal_prio;
2547 if (!rt_prio(p->prio))
2548 p->sched_class = &fair_sched_class;
2550 if (p->sched_class->task_fork)
2551 p->sched_class->task_fork(p);
2554 * The child is not yet in the pid-hash so no cgroup attach races,
2555 * and the cgroup is pinned to this child due to cgroup_fork()
2556 * is ran before sched_fork().
2558 * Silence PROVE_RCU.
2561 set_task_cpu(p, cpu);
2564 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2565 if (likely(sched_info_on()))
2566 memset(&p->sched_info, 0, sizeof(p->sched_info));
2568 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2571 #ifdef CONFIG_PREEMPT
2572 /* Want to start with kernel preemption disabled. */
2573 task_thread_info(p)->preempt_count = 1;
2575 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2581 * wake_up_new_task - wake up a newly created task for the first time.
2583 * This function will do some initial scheduler statistics housekeeping
2584 * that must be done for every newly created context, then puts the task
2585 * on the runqueue and wakes it.
2587 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2589 unsigned long flags;
2591 int cpu __maybe_unused = get_cpu();
2594 rq = task_rq_lock(p, &flags);
2595 p->state = TASK_WAKING;
2598 * Fork balancing, do it here and not earlier because:
2599 * - cpus_allowed can change in the fork path
2600 * - any previously selected cpu might disappear through hotplug
2602 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2603 * without people poking at ->cpus_allowed.
2605 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2606 set_task_cpu(p, cpu);
2608 p->state = TASK_RUNNING;
2609 task_rq_unlock(rq, &flags);
2612 rq = task_rq_lock(p, &flags);
2613 activate_task(rq, p, 0);
2614 trace_sched_wakeup_new(p, 1);
2615 check_preempt_curr(rq, p, WF_FORK);
2617 if (p->sched_class->task_woken)
2618 p->sched_class->task_woken(rq, p);
2620 task_rq_unlock(rq, &flags);
2624 #ifdef CONFIG_PREEMPT_NOTIFIERS
2627 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2628 * @notifier: notifier struct to register
2630 void preempt_notifier_register(struct preempt_notifier *notifier)
2632 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2634 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2637 * preempt_notifier_unregister - no longer interested in preemption notifications
2638 * @notifier: notifier struct to unregister
2640 * This is safe to call from within a preemption notifier.
2642 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2644 hlist_del(¬ifier->link);
2646 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2648 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2650 struct preempt_notifier *notifier;
2651 struct hlist_node *node;
2653 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2654 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2658 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2659 struct task_struct *next)
2661 struct preempt_notifier *notifier;
2662 struct hlist_node *node;
2664 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2665 notifier->ops->sched_out(notifier, next);
2668 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2670 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2675 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2676 struct task_struct *next)
2680 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2683 * prepare_task_switch - prepare to switch tasks
2684 * @rq: the runqueue preparing to switch
2685 * @prev: the current task that is being switched out
2686 * @next: the task we are going to switch to.
2688 * This is called with the rq lock held and interrupts off. It must
2689 * be paired with a subsequent finish_task_switch after the context
2692 * prepare_task_switch sets up locking and calls architecture specific
2696 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2697 struct task_struct *next)
2699 fire_sched_out_preempt_notifiers(prev, next);
2700 prepare_lock_switch(rq, next);
2701 prepare_arch_switch(next);
2705 * finish_task_switch - clean up after a task-switch
2706 * @rq: runqueue associated with task-switch
2707 * @prev: the thread we just switched away from.
2709 * finish_task_switch must be called after the context switch, paired
2710 * with a prepare_task_switch call before the context switch.
2711 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2712 * and do any other architecture-specific cleanup actions.
2714 * Note that we may have delayed dropping an mm in context_switch(). If
2715 * so, we finish that here outside of the runqueue lock. (Doing it
2716 * with the lock held can cause deadlocks; see schedule() for
2719 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2720 __releases(rq->lock)
2722 struct mm_struct *mm = rq->prev_mm;
2728 * A task struct has one reference for the use as "current".
2729 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2730 * schedule one last time. The schedule call will never return, and
2731 * the scheduled task must drop that reference.
2732 * The test for TASK_DEAD must occur while the runqueue locks are
2733 * still held, otherwise prev could be scheduled on another cpu, die
2734 * there before we look at prev->state, and then the reference would
2736 * Manfred Spraul <manfred@colorfullife.com>
2738 prev_state = prev->state;
2739 finish_arch_switch(prev);
2740 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2741 local_irq_disable();
2742 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2743 perf_event_task_sched_in(current);
2744 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2746 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2747 finish_lock_switch(rq, prev);
2749 fire_sched_in_preempt_notifiers(current);
2752 if (unlikely(prev_state == TASK_DEAD)) {
2754 * Remove function-return probe instances associated with this
2755 * task and put them back on the free list.
2757 kprobe_flush_task(prev);
2758 put_task_struct(prev);
2764 /* assumes rq->lock is held */
2765 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2767 if (prev->sched_class->pre_schedule)
2768 prev->sched_class->pre_schedule(rq, prev);
2771 /* rq->lock is NOT held, but preemption is disabled */
2772 static inline void post_schedule(struct rq *rq)
2774 if (rq->post_schedule) {
2775 unsigned long flags;
2777 raw_spin_lock_irqsave(&rq->lock, flags);
2778 if (rq->curr->sched_class->post_schedule)
2779 rq->curr->sched_class->post_schedule(rq);
2780 raw_spin_unlock_irqrestore(&rq->lock, flags);
2782 rq->post_schedule = 0;
2788 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2792 static inline void post_schedule(struct rq *rq)
2799 * schedule_tail - first thing a freshly forked thread must call.
2800 * @prev: the thread we just switched away from.
2802 asmlinkage void schedule_tail(struct task_struct *prev)
2803 __releases(rq->lock)
2805 struct rq *rq = this_rq();
2807 finish_task_switch(rq, prev);
2810 * FIXME: do we need to worry about rq being invalidated by the
2815 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2816 /* In this case, finish_task_switch does not reenable preemption */
2819 if (current->set_child_tid)
2820 put_user(task_pid_vnr(current), current->set_child_tid);
2824 * context_switch - switch to the new MM and the new
2825 * thread's register state.
2828 context_switch(struct rq *rq, struct task_struct *prev,
2829 struct task_struct *next)
2831 struct mm_struct *mm, *oldmm;
2833 prepare_task_switch(rq, prev, next);
2834 trace_sched_switch(prev, next);
2836 oldmm = prev->active_mm;
2838 * For paravirt, this is coupled with an exit in switch_to to
2839 * combine the page table reload and the switch backend into
2842 arch_start_context_switch(prev);
2845 next->active_mm = oldmm;
2846 atomic_inc(&oldmm->mm_count);
2847 enter_lazy_tlb(oldmm, next);
2849 switch_mm(oldmm, mm, next);
2852 prev->active_mm = NULL;
2853 rq->prev_mm = oldmm;
2856 * Since the runqueue lock will be released by the next
2857 * task (which is an invalid locking op but in the case
2858 * of the scheduler it's an obvious special-case), so we
2859 * do an early lockdep release here:
2861 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2862 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2865 /* Here we just switch the register state and the stack. */
2866 switch_to(prev, next, prev);
2870 * this_rq must be evaluated again because prev may have moved
2871 * CPUs since it called schedule(), thus the 'rq' on its stack
2872 * frame will be invalid.
2874 finish_task_switch(this_rq(), prev);
2878 * nr_running, nr_uninterruptible and nr_context_switches:
2880 * externally visible scheduler statistics: current number of runnable
2881 * threads, current number of uninterruptible-sleeping threads, total
2882 * number of context switches performed since bootup.
2884 unsigned long nr_running(void)
2886 unsigned long i, sum = 0;
2888 for_each_online_cpu(i)
2889 sum += cpu_rq(i)->nr_running;
2894 unsigned long nr_uninterruptible(void)
2896 unsigned long i, sum = 0;
2898 for_each_possible_cpu(i)
2899 sum += cpu_rq(i)->nr_uninterruptible;
2902 * Since we read the counters lockless, it might be slightly
2903 * inaccurate. Do not allow it to go below zero though:
2905 if (unlikely((long)sum < 0))
2911 unsigned long long nr_context_switches(void)
2914 unsigned long long sum = 0;
2916 for_each_possible_cpu(i)
2917 sum += cpu_rq(i)->nr_switches;
2922 unsigned long nr_iowait(void)
2924 unsigned long i, sum = 0;
2926 for_each_possible_cpu(i)
2927 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2932 unsigned long nr_iowait_cpu(int cpu)
2934 struct rq *this = cpu_rq(cpu);
2935 return atomic_read(&this->nr_iowait);
2938 unsigned long this_cpu_load(void)
2940 struct rq *this = this_rq();
2941 return this->cpu_load[0];
2945 /* Variables and functions for calc_load */
2946 static atomic_long_t calc_load_tasks;
2947 static unsigned long calc_load_update;
2948 unsigned long avenrun[3];
2949 EXPORT_SYMBOL(avenrun);
2951 static long calc_load_fold_active(struct rq *this_rq)
2953 long nr_active, delta = 0;
2955 nr_active = this_rq->nr_running;
2956 nr_active += (long) this_rq->nr_uninterruptible;
2958 if (nr_active != this_rq->calc_load_active) {
2959 delta = nr_active - this_rq->calc_load_active;
2960 this_rq->calc_load_active = nr_active;
2968 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2970 * When making the ILB scale, we should try to pull this in as well.
2972 static atomic_long_t calc_load_tasks_idle;
2974 static void calc_load_account_idle(struct rq *this_rq)
2978 delta = calc_load_fold_active(this_rq);
2980 atomic_long_add(delta, &calc_load_tasks_idle);
2983 static long calc_load_fold_idle(void)
2988 * Its got a race, we don't care...
2990 if (atomic_long_read(&calc_load_tasks_idle))
2991 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2996 static void calc_load_account_idle(struct rq *this_rq)
3000 static inline long calc_load_fold_idle(void)
3007 * get_avenrun - get the load average array
3008 * @loads: pointer to dest load array
3009 * @offset: offset to add
3010 * @shift: shift count to shift the result left
3012 * These values are estimates at best, so no need for locking.
3014 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3016 loads[0] = (avenrun[0] + offset) << shift;
3017 loads[1] = (avenrun[1] + offset) << shift;
3018 loads[2] = (avenrun[2] + offset) << shift;
3021 static unsigned long
3022 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3025 load += active * (FIXED_1 - exp);
3026 return load >> FSHIFT;
3030 * calc_load - update the avenrun load estimates 10 ticks after the
3031 * CPUs have updated calc_load_tasks.
3033 void calc_global_load(void)
3035 unsigned long upd = calc_load_update + 10;
3038 if (time_before(jiffies, upd))
3041 active = atomic_long_read(&calc_load_tasks);
3042 active = active > 0 ? active * FIXED_1 : 0;
3044 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3045 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3046 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3048 calc_load_update += LOAD_FREQ;
3052 * Called from update_cpu_load() to periodically update this CPU's
3055 static void calc_load_account_active(struct rq *this_rq)
3059 if (time_before(jiffies, this_rq->calc_load_update))
3062 delta = calc_load_fold_active(this_rq);
3063 delta += calc_load_fold_idle();
3065 atomic_long_add(delta, &calc_load_tasks);
3067 this_rq->calc_load_update += LOAD_FREQ;
3071 * The exact cpuload at various idx values, calculated at every tick would be
3072 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3074 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3075 * on nth tick when cpu may be busy, then we have:
3076 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3077 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3079 * decay_load_missed() below does efficient calculation of
3080 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3081 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3083 * The calculation is approximated on a 128 point scale.
3084 * degrade_zero_ticks is the number of ticks after which load at any
3085 * particular idx is approximated to be zero.
3086 * degrade_factor is a precomputed table, a row for each load idx.
3087 * Each column corresponds to degradation factor for a power of two ticks,
3088 * based on 128 point scale.
3090 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3091 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3093 * With this power of 2 load factors, we can degrade the load n times
3094 * by looking at 1 bits in n and doing as many mult/shift instead of
3095 * n mult/shifts needed by the exact degradation.
3097 #define DEGRADE_SHIFT 7
3098 static const unsigned char
3099 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3100 static const unsigned char
3101 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3102 {0, 0, 0, 0, 0, 0, 0, 0},
3103 {64, 32, 8, 0, 0, 0, 0, 0},
3104 {96, 72, 40, 12, 1, 0, 0},
3105 {112, 98, 75, 43, 15, 1, 0},
3106 {120, 112, 98, 76, 45, 16, 2} };
3109 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3110 * would be when CPU is idle and so we just decay the old load without
3111 * adding any new load.
3113 static unsigned long
3114 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3118 if (!missed_updates)
3121 if (missed_updates >= degrade_zero_ticks[idx])
3125 return load >> missed_updates;
3127 while (missed_updates) {
3128 if (missed_updates % 2)
3129 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3131 missed_updates >>= 1;
3138 * Update rq->cpu_load[] statistics. This function is usually called every
3139 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3140 * every tick. We fix it up based on jiffies.
3142 static void update_cpu_load(struct rq *this_rq)
3144 unsigned long this_load = this_rq->load.weight;
3145 unsigned long curr_jiffies = jiffies;
3146 unsigned long pending_updates;
3149 this_rq->nr_load_updates++;
3151 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3152 if (curr_jiffies == this_rq->last_load_update_tick)
3155 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3156 this_rq->last_load_update_tick = curr_jiffies;
3158 /* Update our load: */
3159 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3160 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3161 unsigned long old_load, new_load;
3163 /* scale is effectively 1 << i now, and >> i divides by scale */
3165 old_load = this_rq->cpu_load[i];
3166 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3167 new_load = this_load;
3169 * Round up the averaging division if load is increasing. This
3170 * prevents us from getting stuck on 9 if the load is 10, for
3173 if (new_load > old_load)
3174 new_load += scale - 1;
3176 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3179 sched_avg_update(this_rq);
3182 static void update_cpu_load_active(struct rq *this_rq)
3184 update_cpu_load(this_rq);
3186 calc_load_account_active(this_rq);
3192 * sched_exec - execve() is a valuable balancing opportunity, because at
3193 * this point the task has the smallest effective memory and cache footprint.
3195 void sched_exec(void)
3197 struct task_struct *p = current;
3198 unsigned long flags;
3202 rq = task_rq_lock(p, &flags);
3203 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3204 if (dest_cpu == smp_processor_id())
3208 * select_task_rq() can race against ->cpus_allowed
3210 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3211 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3212 struct migration_arg arg = { p, dest_cpu };
3214 task_rq_unlock(rq, &flags);
3215 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3219 task_rq_unlock(rq, &flags);
3224 DEFINE_PER_CPU(struct kernel_stat, kstat);
3226 EXPORT_PER_CPU_SYMBOL(kstat);
3229 * Return any ns on the sched_clock that have not yet been accounted in
3230 * @p in case that task is currently running.
3232 * Called with task_rq_lock() held on @rq.
3234 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3238 if (task_current(rq, p)) {
3239 update_rq_clock(rq);
3240 ns = rq->clock_task - p->se.exec_start;
3248 unsigned long long task_delta_exec(struct task_struct *p)
3250 unsigned long flags;
3254 rq = task_rq_lock(p, &flags);
3255 ns = do_task_delta_exec(p, rq);
3256 task_rq_unlock(rq, &flags);
3262 * Return accounted runtime for the task.
3263 * In case the task is currently running, return the runtime plus current's
3264 * pending runtime that have not been accounted yet.
3266 unsigned long long task_sched_runtime(struct task_struct *p)
3268 unsigned long flags;
3272 rq = task_rq_lock(p, &flags);
3273 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3274 task_rq_unlock(rq, &flags);
3280 * Return sum_exec_runtime for the thread group.
3281 * In case the task is currently running, return the sum plus current's
3282 * pending runtime that have not been accounted yet.
3284 * Note that the thread group might have other running tasks as well,
3285 * so the return value not includes other pending runtime that other
3286 * running tasks might have.
3288 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3290 struct task_cputime totals;
3291 unsigned long flags;
3295 rq = task_rq_lock(p, &flags);
3296 thread_group_cputime(p, &totals);
3297 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3298 task_rq_unlock(rq, &flags);
3304 * Account user cpu time to a process.
3305 * @p: the process that the cpu time gets accounted to
3306 * @cputime: the cpu time spent in user space since the last update
3307 * @cputime_scaled: cputime scaled by cpu frequency
3309 void account_user_time(struct task_struct *p, cputime_t cputime,
3310 cputime_t cputime_scaled)
3312 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3315 /* Add user time to process. */
3316 p->utime = cputime_add(p->utime, cputime);
3317 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3318 account_group_user_time(p, cputime);
3320 /* Add user time to cpustat. */
3321 tmp = cputime_to_cputime64(cputime);
3322 if (TASK_NICE(p) > 0)
3323 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3325 cpustat->user = cputime64_add(cpustat->user, tmp);
3327 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3328 /* Account for user time used */
3329 acct_update_integrals(p);
3333 * Account guest cpu time to a process.
3334 * @p: the process that the cpu time gets accounted to
3335 * @cputime: the cpu time spent in virtual machine since the last update
3336 * @cputime_scaled: cputime scaled by cpu frequency
3338 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3339 cputime_t cputime_scaled)
3342 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3344 tmp = cputime_to_cputime64(cputime);
3346 /* Add guest time to process. */
3347 p->utime = cputime_add(p->utime, cputime);
3348 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3349 account_group_user_time(p, cputime);
3350 p->gtime = cputime_add(p->gtime, cputime);
3352 /* Add guest time to cpustat. */
3353 if (TASK_NICE(p) > 0) {
3354 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3355 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3357 cpustat->user = cputime64_add(cpustat->user, tmp);
3358 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3363 * Account system cpu time to a process.
3364 * @p: the process that the cpu time gets accounted to
3365 * @hardirq_offset: the offset to subtract from hardirq_count()
3366 * @cputime: the cpu time spent in kernel space since the last update
3367 * @cputime_scaled: cputime scaled by cpu frequency
3369 void account_system_time(struct task_struct *p, int hardirq_offset,
3370 cputime_t cputime, cputime_t cputime_scaled)
3372 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3375 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3376 account_guest_time(p, cputime, cputime_scaled);
3380 /* Add system time to process. */
3381 p->stime = cputime_add(p->stime, cputime);
3382 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3383 account_group_system_time(p, cputime);
3385 /* Add system time to cpustat. */
3386 tmp = cputime_to_cputime64(cputime);
3387 if (hardirq_count() - hardirq_offset)
3388 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3389 else if (in_serving_softirq())
3390 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3392 cpustat->system = cputime64_add(cpustat->system, tmp);
3394 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3396 /* Account for system time used */
3397 acct_update_integrals(p);
3401 * Account for involuntary wait time.
3402 * @steal: the cpu time spent in involuntary wait
3404 void account_steal_time(cputime_t cputime)
3406 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3407 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3409 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3413 * Account for idle time.
3414 * @cputime: the cpu time spent in idle wait
3416 void account_idle_time(cputime_t cputime)
3418 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3419 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3420 struct rq *rq = this_rq();
3422 if (atomic_read(&rq->nr_iowait) > 0)
3423 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3425 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3428 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3431 * Account a single tick of cpu time.
3432 * @p: the process that the cpu time gets accounted to
3433 * @user_tick: indicates if the tick is a user or a system tick
3435 void account_process_tick(struct task_struct *p, int user_tick)
3437 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3438 struct rq *rq = this_rq();
3441 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3442 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3443 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3446 account_idle_time(cputime_one_jiffy);
3450 * Account multiple ticks of steal time.
3451 * @p: the process from which the cpu time has been stolen
3452 * @ticks: number of stolen ticks
3454 void account_steal_ticks(unsigned long ticks)
3456 account_steal_time(jiffies_to_cputime(ticks));
3460 * Account multiple ticks of idle time.
3461 * @ticks: number of stolen ticks
3463 void account_idle_ticks(unsigned long ticks)
3465 account_idle_time(jiffies_to_cputime(ticks));
3471 * Use precise platform statistics if available:
3473 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3474 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3480 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3482 struct task_cputime cputime;
3484 thread_group_cputime(p, &cputime);
3486 *ut = cputime.utime;
3487 *st = cputime.stime;
3491 #ifndef nsecs_to_cputime
3492 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3495 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3497 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3500 * Use CFS's precise accounting:
3502 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3508 do_div(temp, total);
3509 utime = (cputime_t)temp;
3514 * Compare with previous values, to keep monotonicity:
3516 p->prev_utime = max(p->prev_utime, utime);
3517 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3519 *ut = p->prev_utime;
3520 *st = p->prev_stime;
3524 * Must be called with siglock held.
3526 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3528 struct signal_struct *sig = p->signal;
3529 struct task_cputime cputime;
3530 cputime_t rtime, utime, total;
3532 thread_group_cputime(p, &cputime);
3534 total = cputime_add(cputime.utime, cputime.stime);
3535 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3540 temp *= cputime.utime;
3541 do_div(temp, total);
3542 utime = (cputime_t)temp;
3546 sig->prev_utime = max(sig->prev_utime, utime);
3547 sig->prev_stime = max(sig->prev_stime,
3548 cputime_sub(rtime, sig->prev_utime));
3550 *ut = sig->prev_utime;
3551 *st = sig->prev_stime;
3556 * This function gets called by the timer code, with HZ frequency.
3557 * We call it with interrupts disabled.
3559 * It also gets called by the fork code, when changing the parent's
3562 void scheduler_tick(void)
3564 int cpu = smp_processor_id();
3565 struct rq *rq = cpu_rq(cpu);
3566 struct task_struct *curr = rq->curr;
3570 raw_spin_lock(&rq->lock);
3571 update_rq_clock(rq);
3572 update_cpu_load_active(rq);
3573 curr->sched_class->task_tick(rq, curr, 0);
3574 raw_spin_unlock(&rq->lock);
3576 perf_event_task_tick();
3579 rq->idle_at_tick = idle_cpu(cpu);
3580 trigger_load_balance(rq, cpu);
3584 notrace unsigned long get_parent_ip(unsigned long addr)
3586 if (in_lock_functions(addr)) {
3587 addr = CALLER_ADDR2;
3588 if (in_lock_functions(addr))
3589 addr = CALLER_ADDR3;
3594 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3595 defined(CONFIG_PREEMPT_TRACER))
3597 void __kprobes add_preempt_count(int val)
3599 #ifdef CONFIG_DEBUG_PREEMPT
3603 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3606 preempt_count() += val;
3607 #ifdef CONFIG_DEBUG_PREEMPT
3609 * Spinlock count overflowing soon?
3611 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3614 if (preempt_count() == val)
3615 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3617 EXPORT_SYMBOL(add_preempt_count);
3619 void __kprobes sub_preempt_count(int val)
3621 #ifdef CONFIG_DEBUG_PREEMPT
3625 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3628 * Is the spinlock portion underflowing?
3630 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3631 !(preempt_count() & PREEMPT_MASK)))
3635 if (preempt_count() == val)
3636 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3637 preempt_count() -= val;
3639 EXPORT_SYMBOL(sub_preempt_count);
3644 * Print scheduling while atomic bug:
3646 static noinline void __schedule_bug(struct task_struct *prev)
3648 struct pt_regs *regs = get_irq_regs();
3650 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3651 prev->comm, prev->pid, preempt_count());
3653 debug_show_held_locks(prev);
3655 if (irqs_disabled())
3656 print_irqtrace_events(prev);
3665 * Various schedule()-time debugging checks and statistics:
3667 static inline void schedule_debug(struct task_struct *prev)
3670 * Test if we are atomic. Since do_exit() needs to call into
3671 * schedule() atomically, we ignore that path for now.
3672 * Otherwise, whine if we are scheduling when we should not be.
3674 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3675 __schedule_bug(prev);
3677 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3679 schedstat_inc(this_rq(), sched_count);
3680 #ifdef CONFIG_SCHEDSTATS
3681 if (unlikely(prev->lock_depth >= 0)) {
3682 schedstat_inc(this_rq(), bkl_count);
3683 schedstat_inc(prev, sched_info.bkl_count);
3688 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3691 update_rq_clock(rq);
3692 rq->skip_clock_update = 0;
3693 prev->sched_class->put_prev_task(rq, prev);
3697 * Pick up the highest-prio task:
3699 static inline struct task_struct *
3700 pick_next_task(struct rq *rq)
3702 const struct sched_class *class;
3703 struct task_struct *p;
3706 * Optimization: we know that if all tasks are in
3707 * the fair class we can call that function directly:
3709 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3710 p = fair_sched_class.pick_next_task(rq);
3715 for_each_class(class) {
3716 p = class->pick_next_task(rq);
3721 BUG(); /* the idle class will always have a runnable task */
3725 * schedule() is the main scheduler function.
3727 asmlinkage void __sched schedule(void)
3729 struct task_struct *prev, *next;
3730 unsigned long *switch_count;
3736 cpu = smp_processor_id();
3738 rcu_note_context_switch(cpu);
3741 release_kernel_lock(prev);
3742 need_resched_nonpreemptible:
3744 schedule_debug(prev);
3746 if (sched_feat(HRTICK))
3749 raw_spin_lock_irq(&rq->lock);
3750 clear_tsk_need_resched(prev);
3752 switch_count = &prev->nivcsw;
3753 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3754 if (unlikely(signal_pending_state(prev->state, prev))) {
3755 prev->state = TASK_RUNNING;
3758 * If a worker is going to sleep, notify and
3759 * ask workqueue whether it wants to wake up a
3760 * task to maintain concurrency. If so, wake
3763 if (prev->flags & PF_WQ_WORKER) {
3764 struct task_struct *to_wakeup;
3766 to_wakeup = wq_worker_sleeping(prev, cpu);
3768 try_to_wake_up_local(to_wakeup);
3770 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3772 switch_count = &prev->nvcsw;
3775 pre_schedule(rq, prev);
3777 if (unlikely(!rq->nr_running))
3778 idle_balance(cpu, rq);
3780 put_prev_task(rq, prev);
3781 next = pick_next_task(rq);
3783 if (likely(prev != next)) {
3784 sched_info_switch(prev, next);
3785 perf_event_task_sched_out(prev, next);
3791 context_switch(rq, prev, next); /* unlocks the rq */
3793 * The context switch have flipped the stack from under us
3794 * and restored the local variables which were saved when
3795 * this task called schedule() in the past. prev == current
3796 * is still correct, but it can be moved to another cpu/rq.
3798 cpu = smp_processor_id();
3801 raw_spin_unlock_irq(&rq->lock);
3805 if (unlikely(reacquire_kernel_lock(prev)))
3806 goto need_resched_nonpreemptible;
3808 preempt_enable_no_resched();
3812 EXPORT_SYMBOL(schedule);
3814 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3816 * Look out! "owner" is an entirely speculative pointer
3817 * access and not reliable.
3819 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3824 if (!sched_feat(OWNER_SPIN))
3827 #ifdef CONFIG_DEBUG_PAGEALLOC
3829 * Need to access the cpu field knowing that
3830 * DEBUG_PAGEALLOC could have unmapped it if
3831 * the mutex owner just released it and exited.
3833 if (probe_kernel_address(&owner->cpu, cpu))
3840 * Even if the access succeeded (likely case),
3841 * the cpu field may no longer be valid.
3843 if (cpu >= nr_cpumask_bits)
3847 * We need to validate that we can do a
3848 * get_cpu() and that we have the percpu area.
3850 if (!cpu_online(cpu))
3857 * Owner changed, break to re-assess state.
3859 if (lock->owner != owner) {
3861 * If the lock has switched to a different owner,
3862 * we likely have heavy contention. Return 0 to quit
3863 * optimistic spinning and not contend further:
3871 * Is that owner really running on that cpu?
3873 if (task_thread_info(rq->curr) != owner || need_resched())
3883 #ifdef CONFIG_PREEMPT
3885 * this is the entry point to schedule() from in-kernel preemption
3886 * off of preempt_enable. Kernel preemptions off return from interrupt
3887 * occur there and call schedule directly.
3889 asmlinkage void __sched notrace preempt_schedule(void)
3891 struct thread_info *ti = current_thread_info();
3894 * If there is a non-zero preempt_count or interrupts are disabled,
3895 * we do not want to preempt the current task. Just return..
3897 if (likely(ti->preempt_count || irqs_disabled()))
3901 add_preempt_count_notrace(PREEMPT_ACTIVE);
3903 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3906 * Check again in case we missed a preemption opportunity
3907 * between schedule and now.
3910 } while (need_resched());
3912 EXPORT_SYMBOL(preempt_schedule);
3915 * this is the entry point to schedule() from kernel preemption
3916 * off of irq context.
3917 * Note, that this is called and return with irqs disabled. This will
3918 * protect us against recursive calling from irq.
3920 asmlinkage void __sched preempt_schedule_irq(void)
3922 struct thread_info *ti = current_thread_info();
3924 /* Catch callers which need to be fixed */
3925 BUG_ON(ti->preempt_count || !irqs_disabled());
3928 add_preempt_count(PREEMPT_ACTIVE);
3931 local_irq_disable();
3932 sub_preempt_count(PREEMPT_ACTIVE);
3935 * Check again in case we missed a preemption opportunity
3936 * between schedule and now.
3939 } while (need_resched());
3942 #endif /* CONFIG_PREEMPT */
3944 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3947 return try_to_wake_up(curr->private, mode, wake_flags);
3949 EXPORT_SYMBOL(default_wake_function);
3952 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3953 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3954 * number) then we wake all the non-exclusive tasks and one exclusive task.
3956 * There are circumstances in which we can try to wake a task which has already
3957 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3958 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3960 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3961 int nr_exclusive, int wake_flags, void *key)
3963 wait_queue_t *curr, *next;
3965 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3966 unsigned flags = curr->flags;
3968 if (curr->func(curr, mode, wake_flags, key) &&
3969 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3975 * __wake_up - wake up threads blocked on a waitqueue.
3977 * @mode: which threads
3978 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3979 * @key: is directly passed to the wakeup function
3981 * It may be assumed that this function implies a write memory barrier before
3982 * changing the task state if and only if any tasks are woken up.
3984 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3985 int nr_exclusive, void *key)
3987 unsigned long flags;
3989 spin_lock_irqsave(&q->lock, flags);
3990 __wake_up_common(q, mode, nr_exclusive, 0, key);
3991 spin_unlock_irqrestore(&q->lock, flags);
3993 EXPORT_SYMBOL(__wake_up);
3996 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3998 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4000 __wake_up_common(q, mode, 1, 0, NULL);
4002 EXPORT_SYMBOL_GPL(__wake_up_locked);
4004 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4006 __wake_up_common(q, mode, 1, 0, key);
4010 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4012 * @mode: which threads
4013 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4014 * @key: opaque value to be passed to wakeup targets
4016 * The sync wakeup differs that the waker knows that it will schedule
4017 * away soon, so while the target thread will be woken up, it will not
4018 * be migrated to another CPU - ie. the two threads are 'synchronized'
4019 * with each other. This can prevent needless bouncing between CPUs.
4021 * On UP it can prevent extra preemption.
4023 * It may be assumed that this function implies a write memory barrier before
4024 * changing the task state if and only if any tasks are woken up.
4026 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4027 int nr_exclusive, void *key)
4029 unsigned long flags;
4030 int wake_flags = WF_SYNC;
4035 if (unlikely(!nr_exclusive))
4038 spin_lock_irqsave(&q->lock, flags);
4039 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4040 spin_unlock_irqrestore(&q->lock, flags);
4042 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4045 * __wake_up_sync - see __wake_up_sync_key()
4047 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4049 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4051 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4054 * complete: - signals a single thread waiting on this completion
4055 * @x: holds the state of this particular completion
4057 * This will wake up a single thread waiting on this completion. Threads will be
4058 * awakened in the same order in which they were queued.
4060 * See also complete_all(), wait_for_completion() and related routines.
4062 * It may be assumed that this function implies a write memory barrier before
4063 * changing the task state if and only if any tasks are woken up.
4065 void complete(struct completion *x)
4067 unsigned long flags;
4069 spin_lock_irqsave(&x->wait.lock, flags);
4071 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4072 spin_unlock_irqrestore(&x->wait.lock, flags);
4074 EXPORT_SYMBOL(complete);
4077 * complete_all: - signals all threads waiting on this completion
4078 * @x: holds the state of this particular completion
4080 * This will wake up all threads waiting on this particular completion event.
4082 * It may be assumed that this function implies a write memory barrier before
4083 * changing the task state if and only if any tasks are woken up.
4085 void complete_all(struct completion *x)
4087 unsigned long flags;
4089 spin_lock_irqsave(&x->wait.lock, flags);
4090 x->done += UINT_MAX/2;
4091 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4092 spin_unlock_irqrestore(&x->wait.lock, flags);
4094 EXPORT_SYMBOL(complete_all);
4096 static inline long __sched
4097 do_wait_for_common(struct completion *x, long timeout, int state)
4100 DECLARE_WAITQUEUE(wait, current);
4102 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4104 if (signal_pending_state(state, current)) {
4105 timeout = -ERESTARTSYS;
4108 __set_current_state(state);
4109 spin_unlock_irq(&x->wait.lock);
4110 timeout = schedule_timeout(timeout);
4111 spin_lock_irq(&x->wait.lock);
4112 } while (!x->done && timeout);
4113 __remove_wait_queue(&x->wait, &wait);
4118 return timeout ?: 1;
4122 wait_for_common(struct completion *x, long timeout, int state)
4126 spin_lock_irq(&x->wait.lock);
4127 timeout = do_wait_for_common(x, timeout, state);
4128 spin_unlock_irq(&x->wait.lock);
4133 * wait_for_completion: - waits for completion of a task
4134 * @x: holds the state of this particular completion
4136 * This waits to be signaled for completion of a specific task. It is NOT
4137 * interruptible and there is no timeout.
4139 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4140 * and interrupt capability. Also see complete().
4142 void __sched wait_for_completion(struct completion *x)
4144 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4146 EXPORT_SYMBOL(wait_for_completion);
4149 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4150 * @x: holds the state of this particular completion
4151 * @timeout: timeout value in jiffies
4153 * This waits for either a completion of a specific task to be signaled or for a
4154 * specified timeout to expire. The timeout is in jiffies. It is not
4157 unsigned long __sched
4158 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4160 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4162 EXPORT_SYMBOL(wait_for_completion_timeout);
4165 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4166 * @x: holds the state of this particular completion
4168 * This waits for completion of a specific task to be signaled. It is
4171 int __sched wait_for_completion_interruptible(struct completion *x)
4173 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4174 if (t == -ERESTARTSYS)
4178 EXPORT_SYMBOL(wait_for_completion_interruptible);
4181 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4182 * @x: holds the state of this particular completion
4183 * @timeout: timeout value in jiffies
4185 * This waits for either a completion of a specific task to be signaled or for a
4186 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4188 unsigned long __sched
4189 wait_for_completion_interruptible_timeout(struct completion *x,
4190 unsigned long timeout)
4192 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4194 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4197 * wait_for_completion_killable: - waits for completion of a task (killable)
4198 * @x: holds the state of this particular completion
4200 * This waits to be signaled for completion of a specific task. It can be
4201 * interrupted by a kill signal.
4203 int __sched wait_for_completion_killable(struct completion *x)
4205 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4206 if (t == -ERESTARTSYS)
4210 EXPORT_SYMBOL(wait_for_completion_killable);
4213 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4214 * @x: holds the state of this particular completion
4215 * @timeout: timeout value in jiffies
4217 * This waits for either a completion of a specific task to be
4218 * signaled or for a specified timeout to expire. It can be
4219 * interrupted by a kill signal. The timeout is in jiffies.
4221 unsigned long __sched
4222 wait_for_completion_killable_timeout(struct completion *x,
4223 unsigned long timeout)
4225 return wait_for_common(x, timeout, TASK_KILLABLE);
4227 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4230 * try_wait_for_completion - try to decrement a completion without blocking
4231 * @x: completion structure
4233 * Returns: 0 if a decrement cannot be done without blocking
4234 * 1 if a decrement succeeded.
4236 * If a completion is being used as a counting completion,
4237 * attempt to decrement the counter without blocking. This
4238 * enables us to avoid waiting if the resource the completion
4239 * is protecting is not available.
4241 bool try_wait_for_completion(struct completion *x)
4243 unsigned long flags;
4246 spin_lock_irqsave(&x->wait.lock, flags);
4251 spin_unlock_irqrestore(&x->wait.lock, flags);
4254 EXPORT_SYMBOL(try_wait_for_completion);
4257 * completion_done - Test to see if a completion has any waiters
4258 * @x: completion structure
4260 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4261 * 1 if there are no waiters.
4264 bool completion_done(struct completion *x)
4266 unsigned long flags;
4269 spin_lock_irqsave(&x->wait.lock, flags);
4272 spin_unlock_irqrestore(&x->wait.lock, flags);
4275 EXPORT_SYMBOL(completion_done);
4278 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4280 unsigned long flags;
4283 init_waitqueue_entry(&wait, current);
4285 __set_current_state(state);
4287 spin_lock_irqsave(&q->lock, flags);
4288 __add_wait_queue(q, &wait);
4289 spin_unlock(&q->lock);
4290 timeout = schedule_timeout(timeout);
4291 spin_lock_irq(&q->lock);
4292 __remove_wait_queue(q, &wait);
4293 spin_unlock_irqrestore(&q->lock, flags);
4298 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4300 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4302 EXPORT_SYMBOL(interruptible_sleep_on);
4305 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4307 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4309 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4311 void __sched sleep_on(wait_queue_head_t *q)
4313 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4315 EXPORT_SYMBOL(sleep_on);
4317 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4319 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4321 EXPORT_SYMBOL(sleep_on_timeout);
4323 #ifdef CONFIG_RT_MUTEXES
4326 * rt_mutex_setprio - set the current priority of a task
4328 * @prio: prio value (kernel-internal form)
4330 * This function changes the 'effective' priority of a task. It does
4331 * not touch ->normal_prio like __setscheduler().
4333 * Used by the rt_mutex code to implement priority inheritance logic.
4335 void rt_mutex_setprio(struct task_struct *p, int prio)
4337 unsigned long flags;
4338 int oldprio, on_rq, running;
4340 const struct sched_class *prev_class;
4342 BUG_ON(prio < 0 || prio > MAX_PRIO);
4344 rq = task_rq_lock(p, &flags);
4346 trace_sched_pi_setprio(p, prio);
4348 prev_class = p->sched_class;
4349 on_rq = p->se.on_rq;
4350 running = task_current(rq, p);
4352 dequeue_task(rq, p, 0);
4354 p->sched_class->put_prev_task(rq, p);
4357 p->sched_class = &rt_sched_class;
4359 p->sched_class = &fair_sched_class;
4364 p->sched_class->set_curr_task(rq);
4366 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4368 check_class_changed(rq, p, prev_class, oldprio, running);
4370 task_rq_unlock(rq, &flags);
4375 void set_user_nice(struct task_struct *p, long nice)
4377 int old_prio, delta, on_rq;
4378 unsigned long flags;
4381 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4384 * We have to be careful, if called from sys_setpriority(),
4385 * the task might be in the middle of scheduling on another CPU.
4387 rq = task_rq_lock(p, &flags);
4389 * The RT priorities are set via sched_setscheduler(), but we still
4390 * allow the 'normal' nice value to be set - but as expected
4391 * it wont have any effect on scheduling until the task is
4392 * SCHED_FIFO/SCHED_RR:
4394 if (task_has_rt_policy(p)) {
4395 p->static_prio = NICE_TO_PRIO(nice);
4398 on_rq = p->se.on_rq;
4400 dequeue_task(rq, p, 0);
4402 p->static_prio = NICE_TO_PRIO(nice);
4405 p->prio = effective_prio(p);
4406 delta = p->prio - old_prio;
4409 enqueue_task(rq, p, 0);
4411 * If the task increased its priority or is running and
4412 * lowered its priority, then reschedule its CPU:
4414 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4415 resched_task(rq->curr);
4418 task_rq_unlock(rq, &flags);
4420 EXPORT_SYMBOL(set_user_nice);
4423 * can_nice - check if a task can reduce its nice value
4427 int can_nice(const struct task_struct *p, const int nice)
4429 /* convert nice value [19,-20] to rlimit style value [1,40] */
4430 int nice_rlim = 20 - nice;
4432 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4433 capable(CAP_SYS_NICE));
4436 #ifdef __ARCH_WANT_SYS_NICE
4439 * sys_nice - change the priority of the current process.
4440 * @increment: priority increment
4442 * sys_setpriority is a more generic, but much slower function that
4443 * does similar things.
4445 SYSCALL_DEFINE1(nice, int, increment)
4450 * Setpriority might change our priority at the same moment.
4451 * We don't have to worry. Conceptually one call occurs first
4452 * and we have a single winner.
4454 if (increment < -40)
4459 nice = TASK_NICE(current) + increment;
4465 if (increment < 0 && !can_nice(current, nice))
4468 retval = security_task_setnice(current, nice);
4472 set_user_nice(current, nice);
4479 * task_prio - return the priority value of a given task.
4480 * @p: the task in question.
4482 * This is the priority value as seen by users in /proc.
4483 * RT tasks are offset by -200. Normal tasks are centered
4484 * around 0, value goes from -16 to +15.
4486 int task_prio(const struct task_struct *p)
4488 return p->prio - MAX_RT_PRIO;
4492 * task_nice - return the nice value of a given task.
4493 * @p: the task in question.
4495 int task_nice(const struct task_struct *p)
4497 return TASK_NICE(p);
4499 EXPORT_SYMBOL(task_nice);
4502 * idle_cpu - is a given cpu idle currently?
4503 * @cpu: the processor in question.
4505 int idle_cpu(int cpu)
4507 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4511 * idle_task - return the idle task for a given cpu.
4512 * @cpu: the processor in question.
4514 struct task_struct *idle_task(int cpu)
4516 return cpu_rq(cpu)->idle;
4520 * find_process_by_pid - find a process with a matching PID value.
4521 * @pid: the pid in question.
4523 static struct task_struct *find_process_by_pid(pid_t pid)
4525 return pid ? find_task_by_vpid(pid) : current;
4528 /* Actually do priority change: must hold rq lock. */
4530 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4532 BUG_ON(p->se.on_rq);
4535 p->rt_priority = prio;
4536 p->normal_prio = normal_prio(p);
4537 /* we are holding p->pi_lock already */
4538 p->prio = rt_mutex_getprio(p);
4539 if (rt_prio(p->prio))
4540 p->sched_class = &rt_sched_class;
4542 p->sched_class = &fair_sched_class;
4547 * check the target process has a UID that matches the current process's
4549 static bool check_same_owner(struct task_struct *p)
4551 const struct cred *cred = current_cred(), *pcred;
4555 pcred = __task_cred(p);
4556 match = (cred->euid == pcred->euid ||
4557 cred->euid == pcred->uid);
4562 static int __sched_setscheduler(struct task_struct *p, int policy,
4563 const struct sched_param *param, bool user)
4565 int retval, oldprio, oldpolicy = -1, on_rq, running;
4566 unsigned long flags;
4567 const struct sched_class *prev_class;
4571 /* may grab non-irq protected spin_locks */
4572 BUG_ON(in_interrupt());
4574 /* double check policy once rq lock held */
4576 reset_on_fork = p->sched_reset_on_fork;
4577 policy = oldpolicy = p->policy;
4579 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4580 policy &= ~SCHED_RESET_ON_FORK;
4582 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4583 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4584 policy != SCHED_IDLE)
4589 * Valid priorities for SCHED_FIFO and SCHED_RR are
4590 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4591 * SCHED_BATCH and SCHED_IDLE is 0.
4593 if (param->sched_priority < 0 ||
4594 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4595 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4597 if (rt_policy(policy) != (param->sched_priority != 0))
4601 * Allow unprivileged RT tasks to decrease priority:
4603 if (user && !capable(CAP_SYS_NICE)) {
4604 if (rt_policy(policy)) {
4605 unsigned long rlim_rtprio =
4606 task_rlimit(p, RLIMIT_RTPRIO);
4608 /* can't set/change the rt policy */
4609 if (policy != p->policy && !rlim_rtprio)
4612 /* can't increase priority */
4613 if (param->sched_priority > p->rt_priority &&
4614 param->sched_priority > rlim_rtprio)
4618 * Like positive nice levels, dont allow tasks to
4619 * move out of SCHED_IDLE either:
4621 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4624 /* can't change other user's priorities */
4625 if (!check_same_owner(p))
4628 /* Normal users shall not reset the sched_reset_on_fork flag */
4629 if (p->sched_reset_on_fork && !reset_on_fork)
4634 retval = security_task_setscheduler(p);
4640 * make sure no PI-waiters arrive (or leave) while we are
4641 * changing the priority of the task:
4643 raw_spin_lock_irqsave(&p->pi_lock, flags);
4645 * To be able to change p->policy safely, the apropriate
4646 * runqueue lock must be held.
4648 rq = __task_rq_lock(p);
4651 * Changing the policy of the stop threads its a very bad idea
4653 if (p == rq->stop) {
4654 __task_rq_unlock(rq);
4655 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4659 #ifdef CONFIG_RT_GROUP_SCHED
4662 * Do not allow realtime tasks into groups that have no runtime
4665 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4666 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4667 __task_rq_unlock(rq);
4668 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4674 /* recheck policy now with rq lock held */
4675 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4676 policy = oldpolicy = -1;
4677 __task_rq_unlock(rq);
4678 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4681 on_rq = p->se.on_rq;
4682 running = task_current(rq, p);
4684 deactivate_task(rq, p, 0);
4686 p->sched_class->put_prev_task(rq, p);
4688 p->sched_reset_on_fork = reset_on_fork;
4691 prev_class = p->sched_class;
4692 __setscheduler(rq, p, policy, param->sched_priority);
4695 p->sched_class->set_curr_task(rq);
4697 activate_task(rq, p, 0);
4699 check_class_changed(rq, p, prev_class, oldprio, running);
4701 __task_rq_unlock(rq);
4702 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4704 rt_mutex_adjust_pi(p);
4710 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4711 * @p: the task in question.
4712 * @policy: new policy.
4713 * @param: structure containing the new RT priority.
4715 * NOTE that the task may be already dead.
4717 int sched_setscheduler(struct task_struct *p, int policy,
4718 const struct sched_param *param)
4720 return __sched_setscheduler(p, policy, param, true);
4722 EXPORT_SYMBOL_GPL(sched_setscheduler);
4725 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4726 * @p: the task in question.
4727 * @policy: new policy.
4728 * @param: structure containing the new RT priority.
4730 * Just like sched_setscheduler, only don't bother checking if the
4731 * current context has permission. For example, this is needed in
4732 * stop_machine(): we create temporary high priority worker threads,
4733 * but our caller might not have that capability.
4735 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4736 const struct sched_param *param)
4738 return __sched_setscheduler(p, policy, param, false);
4742 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4744 struct sched_param lparam;
4745 struct task_struct *p;
4748 if (!param || pid < 0)
4750 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4755 p = find_process_by_pid(pid);
4757 retval = sched_setscheduler(p, policy, &lparam);
4764 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4765 * @pid: the pid in question.
4766 * @policy: new policy.
4767 * @param: structure containing the new RT priority.
4769 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4770 struct sched_param __user *, param)
4772 /* negative values for policy are not valid */
4776 return do_sched_setscheduler(pid, policy, param);
4780 * sys_sched_setparam - set/change the RT priority of a thread
4781 * @pid: the pid in question.
4782 * @param: structure containing the new RT priority.
4784 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4786 return do_sched_setscheduler(pid, -1, param);
4790 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4791 * @pid: the pid in question.
4793 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4795 struct task_struct *p;
4803 p = find_process_by_pid(pid);
4805 retval = security_task_getscheduler(p);
4808 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4815 * sys_sched_getparam - get the RT priority of a thread
4816 * @pid: the pid in question.
4817 * @param: structure containing the RT priority.
4819 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4821 struct sched_param lp;
4822 struct task_struct *p;
4825 if (!param || pid < 0)
4829 p = find_process_by_pid(pid);
4834 retval = security_task_getscheduler(p);
4838 lp.sched_priority = p->rt_priority;
4842 * This one might sleep, we cannot do it with a spinlock held ...
4844 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4853 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4855 cpumask_var_t cpus_allowed, new_mask;
4856 struct task_struct *p;
4862 p = find_process_by_pid(pid);
4869 /* Prevent p going away */
4873 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4877 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4879 goto out_free_cpus_allowed;
4882 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4885 retval = security_task_setscheduler(p);
4889 cpuset_cpus_allowed(p, cpus_allowed);
4890 cpumask_and(new_mask, in_mask, cpus_allowed);
4892 retval = set_cpus_allowed_ptr(p, new_mask);
4895 cpuset_cpus_allowed(p, cpus_allowed);
4896 if (!cpumask_subset(new_mask, cpus_allowed)) {
4898 * We must have raced with a concurrent cpuset
4899 * update. Just reset the cpus_allowed to the
4900 * cpuset's cpus_allowed
4902 cpumask_copy(new_mask, cpus_allowed);
4907 free_cpumask_var(new_mask);
4908 out_free_cpus_allowed:
4909 free_cpumask_var(cpus_allowed);
4916 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4917 struct cpumask *new_mask)
4919 if (len < cpumask_size())
4920 cpumask_clear(new_mask);
4921 else if (len > cpumask_size())
4922 len = cpumask_size();
4924 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4928 * sys_sched_setaffinity - set the cpu affinity of a process
4929 * @pid: pid of the process
4930 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4931 * @user_mask_ptr: user-space pointer to the new cpu mask
4933 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4934 unsigned long __user *, user_mask_ptr)
4936 cpumask_var_t new_mask;
4939 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4942 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4944 retval = sched_setaffinity(pid, new_mask);
4945 free_cpumask_var(new_mask);
4949 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4951 struct task_struct *p;
4952 unsigned long flags;
4960 p = find_process_by_pid(pid);
4964 retval = security_task_getscheduler(p);
4968 rq = task_rq_lock(p, &flags);
4969 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4970 task_rq_unlock(rq, &flags);
4980 * sys_sched_getaffinity - get the cpu affinity of a process
4981 * @pid: pid of the process
4982 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4983 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4985 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4986 unsigned long __user *, user_mask_ptr)
4991 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4993 if (len & (sizeof(unsigned long)-1))
4996 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4999 ret = sched_getaffinity(pid, mask);
5001 size_t retlen = min_t(size_t, len, cpumask_size());
5003 if (copy_to_user(user_mask_ptr, mask, retlen))
5008 free_cpumask_var(mask);
5014 * sys_sched_yield - yield the current processor to other threads.
5016 * This function yields the current CPU to other tasks. If there are no
5017 * other threads running on this CPU then this function will return.
5019 SYSCALL_DEFINE0(sched_yield)
5021 struct rq *rq = this_rq_lock();
5023 schedstat_inc(rq, yld_count);
5024 current->sched_class->yield_task(rq);
5027 * Since we are going to call schedule() anyway, there's
5028 * no need to preempt or enable interrupts:
5030 __release(rq->lock);
5031 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5032 do_raw_spin_unlock(&rq->lock);
5033 preempt_enable_no_resched();
5040 static inline int should_resched(void)
5042 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5045 static void __cond_resched(void)
5047 add_preempt_count(PREEMPT_ACTIVE);
5049 sub_preempt_count(PREEMPT_ACTIVE);
5052 int __sched _cond_resched(void)
5054 if (should_resched()) {
5060 EXPORT_SYMBOL(_cond_resched);
5063 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5064 * call schedule, and on return reacquire the lock.
5066 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5067 * operations here to prevent schedule() from being called twice (once via
5068 * spin_unlock(), once by hand).
5070 int __cond_resched_lock(spinlock_t *lock)
5072 int resched = should_resched();
5075 lockdep_assert_held(lock);
5077 if (spin_needbreak(lock) || resched) {
5088 EXPORT_SYMBOL(__cond_resched_lock);
5090 int __sched __cond_resched_softirq(void)
5092 BUG_ON(!in_softirq());
5094 if (should_resched()) {
5102 EXPORT_SYMBOL(__cond_resched_softirq);
5105 * yield - yield the current processor to other threads.
5107 * This is a shortcut for kernel-space yielding - it marks the
5108 * thread runnable and calls sys_sched_yield().
5110 void __sched yield(void)
5112 set_current_state(TASK_RUNNING);
5115 EXPORT_SYMBOL(yield);
5118 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5119 * that process accounting knows that this is a task in IO wait state.
5121 void __sched io_schedule(void)
5123 struct rq *rq = raw_rq();
5125 delayacct_blkio_start();
5126 atomic_inc(&rq->nr_iowait);
5127 current->in_iowait = 1;
5129 current->in_iowait = 0;
5130 atomic_dec(&rq->nr_iowait);
5131 delayacct_blkio_end();
5133 EXPORT_SYMBOL(io_schedule);
5135 long __sched io_schedule_timeout(long timeout)
5137 struct rq *rq = raw_rq();
5140 delayacct_blkio_start();
5141 atomic_inc(&rq->nr_iowait);
5142 current->in_iowait = 1;
5143 ret = schedule_timeout(timeout);
5144 current->in_iowait = 0;
5145 atomic_dec(&rq->nr_iowait);
5146 delayacct_blkio_end();
5151 * sys_sched_get_priority_max - return maximum RT priority.
5152 * @policy: scheduling class.
5154 * this syscall returns the maximum rt_priority that can be used
5155 * by a given scheduling class.
5157 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5164 ret = MAX_USER_RT_PRIO-1;
5176 * sys_sched_get_priority_min - return minimum RT priority.
5177 * @policy: scheduling class.
5179 * this syscall returns the minimum rt_priority that can be used
5180 * by a given scheduling class.
5182 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5200 * sys_sched_rr_get_interval - return the default timeslice of a process.
5201 * @pid: pid of the process.
5202 * @interval: userspace pointer to the timeslice value.
5204 * this syscall writes the default timeslice value of a given process
5205 * into the user-space timespec buffer. A value of '0' means infinity.
5207 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5208 struct timespec __user *, interval)
5210 struct task_struct *p;
5211 unsigned int time_slice;
5212 unsigned long flags;
5222 p = find_process_by_pid(pid);
5226 retval = security_task_getscheduler(p);
5230 rq = task_rq_lock(p, &flags);
5231 time_slice = p->sched_class->get_rr_interval(rq, p);
5232 task_rq_unlock(rq, &flags);
5235 jiffies_to_timespec(time_slice, &t);
5236 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5244 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5246 void sched_show_task(struct task_struct *p)
5248 unsigned long free = 0;
5251 state = p->state ? __ffs(p->state) + 1 : 0;
5252 printk(KERN_INFO "%-13.13s %c", p->comm,
5253 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5254 #if BITS_PER_LONG == 32
5255 if (state == TASK_RUNNING)
5256 printk(KERN_CONT " running ");
5258 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5260 if (state == TASK_RUNNING)
5261 printk(KERN_CONT " running task ");
5263 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5265 #ifdef CONFIG_DEBUG_STACK_USAGE
5266 free = stack_not_used(p);
5268 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5269 task_pid_nr(p), task_pid_nr(p->real_parent),
5270 (unsigned long)task_thread_info(p)->flags);
5272 show_stack(p, NULL);
5275 void show_state_filter(unsigned long state_filter)
5277 struct task_struct *g, *p;
5279 #if BITS_PER_LONG == 32
5281 " task PC stack pid father\n");
5284 " task PC stack pid father\n");
5286 read_lock(&tasklist_lock);
5287 do_each_thread(g, p) {
5289 * reset the NMI-timeout, listing all files on a slow
5290 * console might take alot of time:
5292 touch_nmi_watchdog();
5293 if (!state_filter || (p->state & state_filter))
5295 } while_each_thread(g, p);
5297 touch_all_softlockup_watchdogs();
5299 #ifdef CONFIG_SCHED_DEBUG
5300 sysrq_sched_debug_show();
5302 read_unlock(&tasklist_lock);
5304 * Only show locks if all tasks are dumped:
5307 debug_show_all_locks();
5310 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5312 idle->sched_class = &idle_sched_class;
5316 * init_idle - set up an idle thread for a given CPU
5317 * @idle: task in question
5318 * @cpu: cpu the idle task belongs to
5320 * NOTE: this function does not set the idle thread's NEED_RESCHED
5321 * flag, to make booting more robust.
5323 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5325 struct rq *rq = cpu_rq(cpu);
5326 unsigned long flags;
5328 raw_spin_lock_irqsave(&rq->lock, flags);
5331 idle->state = TASK_RUNNING;
5332 idle->se.exec_start = sched_clock();
5334 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5336 * We're having a chicken and egg problem, even though we are
5337 * holding rq->lock, the cpu isn't yet set to this cpu so the
5338 * lockdep check in task_group() will fail.
5340 * Similar case to sched_fork(). / Alternatively we could
5341 * use task_rq_lock() here and obtain the other rq->lock.
5346 __set_task_cpu(idle, cpu);
5349 rq->curr = rq->idle = idle;
5350 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5353 raw_spin_unlock_irqrestore(&rq->lock, flags);
5355 /* Set the preempt count _outside_ the spinlocks! */
5356 #if defined(CONFIG_PREEMPT)
5357 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5359 task_thread_info(idle)->preempt_count = 0;
5362 * The idle tasks have their own, simple scheduling class:
5364 idle->sched_class = &idle_sched_class;
5365 ftrace_graph_init_task(idle);
5369 * In a system that switches off the HZ timer nohz_cpu_mask
5370 * indicates which cpus entered this state. This is used
5371 * in the rcu update to wait only for active cpus. For system
5372 * which do not switch off the HZ timer nohz_cpu_mask should
5373 * always be CPU_BITS_NONE.
5375 cpumask_var_t nohz_cpu_mask;
5378 * Increase the granularity value when there are more CPUs,
5379 * because with more CPUs the 'effective latency' as visible
5380 * to users decreases. But the relationship is not linear,
5381 * so pick a second-best guess by going with the log2 of the
5384 * This idea comes from the SD scheduler of Con Kolivas:
5386 static int get_update_sysctl_factor(void)
5388 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5389 unsigned int factor;
5391 switch (sysctl_sched_tunable_scaling) {
5392 case SCHED_TUNABLESCALING_NONE:
5395 case SCHED_TUNABLESCALING_LINEAR:
5398 case SCHED_TUNABLESCALING_LOG:
5400 factor = 1 + ilog2(cpus);
5407 static void update_sysctl(void)
5409 unsigned int factor = get_update_sysctl_factor();
5411 #define SET_SYSCTL(name) \
5412 (sysctl_##name = (factor) * normalized_sysctl_##name)
5413 SET_SYSCTL(sched_min_granularity);
5414 SET_SYSCTL(sched_latency);
5415 SET_SYSCTL(sched_wakeup_granularity);
5419 static inline void sched_init_granularity(void)
5426 * This is how migration works:
5428 * 1) we invoke migration_cpu_stop() on the target CPU using
5430 * 2) stopper starts to run (implicitly forcing the migrated thread
5432 * 3) it checks whether the migrated task is still in the wrong runqueue.
5433 * 4) if it's in the wrong runqueue then the migration thread removes
5434 * it and puts it into the right queue.
5435 * 5) stopper completes and stop_one_cpu() returns and the migration
5440 * Change a given task's CPU affinity. Migrate the thread to a
5441 * proper CPU and schedule it away if the CPU it's executing on
5442 * is removed from the allowed bitmask.
5444 * NOTE: the caller must have a valid reference to the task, the
5445 * task must not exit() & deallocate itself prematurely. The
5446 * call is not atomic; no spinlocks may be held.
5448 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5450 unsigned long flags;
5452 unsigned int dest_cpu;
5456 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5457 * drop the rq->lock and still rely on ->cpus_allowed.
5460 while (task_is_waking(p))
5462 rq = task_rq_lock(p, &flags);
5463 if (task_is_waking(p)) {
5464 task_rq_unlock(rq, &flags);
5468 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5473 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5474 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5479 if (p->sched_class->set_cpus_allowed)
5480 p->sched_class->set_cpus_allowed(p, new_mask);
5482 cpumask_copy(&p->cpus_allowed, new_mask);
5483 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5486 /* Can the task run on the task's current CPU? If so, we're done */
5487 if (cpumask_test_cpu(task_cpu(p), new_mask))
5490 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5491 if (migrate_task(p, dest_cpu)) {
5492 struct migration_arg arg = { p, dest_cpu };
5493 /* Need help from migration thread: drop lock and wait. */
5494 task_rq_unlock(rq, &flags);
5495 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5496 tlb_migrate_finish(p->mm);
5500 task_rq_unlock(rq, &flags);
5504 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5507 * Move (not current) task off this cpu, onto dest cpu. We're doing
5508 * this because either it can't run here any more (set_cpus_allowed()
5509 * away from this CPU, or CPU going down), or because we're
5510 * attempting to rebalance this task on exec (sched_exec).
5512 * So we race with normal scheduler movements, but that's OK, as long
5513 * as the task is no longer on this CPU.
5515 * Returns non-zero if task was successfully migrated.
5517 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5519 struct rq *rq_dest, *rq_src;
5522 if (unlikely(!cpu_active(dest_cpu)))
5525 rq_src = cpu_rq(src_cpu);
5526 rq_dest = cpu_rq(dest_cpu);
5528 double_rq_lock(rq_src, rq_dest);
5529 /* Already moved. */
5530 if (task_cpu(p) != src_cpu)
5532 /* Affinity changed (again). */
5533 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5537 * If we're not on a rq, the next wake-up will ensure we're
5541 deactivate_task(rq_src, p, 0);
5542 set_task_cpu(p, dest_cpu);
5543 activate_task(rq_dest, p, 0);
5544 check_preempt_curr(rq_dest, p, 0);
5549 double_rq_unlock(rq_src, rq_dest);
5554 * migration_cpu_stop - this will be executed by a highprio stopper thread
5555 * and performs thread migration by bumping thread off CPU then
5556 * 'pushing' onto another runqueue.
5558 static int migration_cpu_stop(void *data)
5560 struct migration_arg *arg = data;
5563 * The original target cpu might have gone down and we might
5564 * be on another cpu but it doesn't matter.
5566 local_irq_disable();
5567 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5572 #ifdef CONFIG_HOTPLUG_CPU
5575 * Ensures that the idle task is using init_mm right before its cpu goes
5578 void idle_task_exit(void)
5580 struct mm_struct *mm = current->active_mm;
5582 BUG_ON(cpu_online(smp_processor_id()));
5585 switch_mm(mm, &init_mm, current);
5590 * While a dead CPU has no uninterruptible tasks queued at this point,
5591 * it might still have a nonzero ->nr_uninterruptible counter, because
5592 * for performance reasons the counter is not stricly tracking tasks to
5593 * their home CPUs. So we just add the counter to another CPU's counter,
5594 * to keep the global sum constant after CPU-down:
5596 static void migrate_nr_uninterruptible(struct rq *rq_src)
5598 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5600 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5601 rq_src->nr_uninterruptible = 0;
5605 * remove the tasks which were accounted by rq from calc_load_tasks.
5607 static void calc_global_load_remove(struct rq *rq)
5609 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5610 rq->calc_load_active = 0;
5614 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5615 * try_to_wake_up()->select_task_rq().
5617 * Called with rq->lock held even though we'er in stop_machine() and
5618 * there's no concurrency possible, we hold the required locks anyway
5619 * because of lock validation efforts.
5621 static void migrate_tasks(unsigned int dead_cpu)
5623 struct rq *rq = cpu_rq(dead_cpu);
5624 struct task_struct *next, *stop = rq->stop;
5628 * Fudge the rq selection such that the below task selection loop
5629 * doesn't get stuck on the currently eligible stop task.
5631 * We're currently inside stop_machine() and the rq is either stuck
5632 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5633 * either way we should never end up calling schedule() until we're
5640 * There's this thread running, bail when that's the only
5643 if (rq->nr_running == 1)
5646 next = pick_next_task(rq);
5648 next->sched_class->put_prev_task(rq, next);
5650 /* Find suitable destination for @next, with force if needed. */
5651 dest_cpu = select_fallback_rq(dead_cpu, next);
5652 raw_spin_unlock(&rq->lock);
5654 __migrate_task(next, dead_cpu, dest_cpu);
5656 raw_spin_lock(&rq->lock);
5662 #endif /* CONFIG_HOTPLUG_CPU */
5664 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5666 static struct ctl_table sd_ctl_dir[] = {
5668 .procname = "sched_domain",
5674 static struct ctl_table sd_ctl_root[] = {
5676 .procname = "kernel",
5678 .child = sd_ctl_dir,
5683 static struct ctl_table *sd_alloc_ctl_entry(int n)
5685 struct ctl_table *entry =
5686 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5691 static void sd_free_ctl_entry(struct ctl_table **tablep)
5693 struct ctl_table *entry;
5696 * In the intermediate directories, both the child directory and
5697 * procname are dynamically allocated and could fail but the mode
5698 * will always be set. In the lowest directory the names are
5699 * static strings and all have proc handlers.
5701 for (entry = *tablep; entry->mode; entry++) {
5703 sd_free_ctl_entry(&entry->child);
5704 if (entry->proc_handler == NULL)
5705 kfree(entry->procname);
5713 set_table_entry(struct ctl_table *entry,
5714 const char *procname, void *data, int maxlen,
5715 mode_t mode, proc_handler *proc_handler)
5717 entry->procname = procname;
5719 entry->maxlen = maxlen;
5721 entry->proc_handler = proc_handler;
5724 static struct ctl_table *
5725 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5727 struct ctl_table *table = sd_alloc_ctl_entry(13);
5732 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5733 sizeof(long), 0644, proc_doulongvec_minmax);
5734 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5735 sizeof(long), 0644, proc_doulongvec_minmax);
5736 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5737 sizeof(int), 0644, proc_dointvec_minmax);
5738 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5739 sizeof(int), 0644, proc_dointvec_minmax);
5740 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5741 sizeof(int), 0644, proc_dointvec_minmax);
5742 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5743 sizeof(int), 0644, proc_dointvec_minmax);
5744 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5745 sizeof(int), 0644, proc_dointvec_minmax);
5746 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5747 sizeof(int), 0644, proc_dointvec_minmax);
5748 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5749 sizeof(int), 0644, proc_dointvec_minmax);
5750 set_table_entry(&table[9], "cache_nice_tries",
5751 &sd->cache_nice_tries,
5752 sizeof(int), 0644, proc_dointvec_minmax);
5753 set_table_entry(&table[10], "flags", &sd->flags,
5754 sizeof(int), 0644, proc_dointvec_minmax);
5755 set_table_entry(&table[11], "name", sd->name,
5756 CORENAME_MAX_SIZE, 0444, proc_dostring);
5757 /* &table[12] is terminator */
5762 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5764 struct ctl_table *entry, *table;
5765 struct sched_domain *sd;
5766 int domain_num = 0, i;
5769 for_each_domain(cpu, sd)
5771 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5776 for_each_domain(cpu, sd) {
5777 snprintf(buf, 32, "domain%d", i);
5778 entry->procname = kstrdup(buf, GFP_KERNEL);
5780 entry->child = sd_alloc_ctl_domain_table(sd);
5787 static struct ctl_table_header *sd_sysctl_header;
5788 static void register_sched_domain_sysctl(void)
5790 int i, cpu_num = num_possible_cpus();
5791 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5794 WARN_ON(sd_ctl_dir[0].child);
5795 sd_ctl_dir[0].child = entry;
5800 for_each_possible_cpu(i) {
5801 snprintf(buf, 32, "cpu%d", i);
5802 entry->procname = kstrdup(buf, GFP_KERNEL);
5804 entry->child = sd_alloc_ctl_cpu_table(i);
5808 WARN_ON(sd_sysctl_header);
5809 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5812 /* may be called multiple times per register */
5813 static void unregister_sched_domain_sysctl(void)
5815 if (sd_sysctl_header)
5816 unregister_sysctl_table(sd_sysctl_header);
5817 sd_sysctl_header = NULL;
5818 if (sd_ctl_dir[0].child)
5819 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5822 static void register_sched_domain_sysctl(void)
5825 static void unregister_sched_domain_sysctl(void)
5830 static void set_rq_online(struct rq *rq)
5833 const struct sched_class *class;
5835 cpumask_set_cpu(rq->cpu, rq->rd->online);
5838 for_each_class(class) {
5839 if (class->rq_online)
5840 class->rq_online(rq);
5845 static void set_rq_offline(struct rq *rq)
5848 const struct sched_class *class;
5850 for_each_class(class) {
5851 if (class->rq_offline)
5852 class->rq_offline(rq);
5855 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5861 * migration_call - callback that gets triggered when a CPU is added.
5862 * Here we can start up the necessary migration thread for the new CPU.
5864 static int __cpuinit
5865 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5867 int cpu = (long)hcpu;
5868 unsigned long flags;
5869 struct rq *rq = cpu_rq(cpu);
5871 switch (action & ~CPU_TASKS_FROZEN) {
5873 case CPU_UP_PREPARE:
5874 rq->calc_load_update = calc_load_update;
5878 /* Update our root-domain */
5879 raw_spin_lock_irqsave(&rq->lock, flags);
5881 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5885 raw_spin_unlock_irqrestore(&rq->lock, flags);
5888 #ifdef CONFIG_HOTPLUG_CPU
5890 /* Update our root-domain */
5891 raw_spin_lock_irqsave(&rq->lock, flags);
5893 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5897 BUG_ON(rq->nr_running != 1); /* the migration thread */
5898 raw_spin_unlock_irqrestore(&rq->lock, flags);
5900 migrate_nr_uninterruptible(rq);
5901 calc_global_load_remove(rq);
5909 * Register at high priority so that task migration (migrate_all_tasks)
5910 * happens before everything else. This has to be lower priority than
5911 * the notifier in the perf_event subsystem, though.
5913 static struct notifier_block __cpuinitdata migration_notifier = {
5914 .notifier_call = migration_call,
5915 .priority = CPU_PRI_MIGRATION,
5918 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5919 unsigned long action, void *hcpu)
5921 switch (action & ~CPU_TASKS_FROZEN) {
5923 case CPU_DOWN_FAILED:
5924 set_cpu_active((long)hcpu, true);
5931 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5932 unsigned long action, void *hcpu)
5934 switch (action & ~CPU_TASKS_FROZEN) {
5935 case CPU_DOWN_PREPARE:
5936 set_cpu_active((long)hcpu, false);
5943 static int __init migration_init(void)
5945 void *cpu = (void *)(long)smp_processor_id();
5948 /* Initialize migration for the boot CPU */
5949 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5950 BUG_ON(err == NOTIFY_BAD);
5951 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5952 register_cpu_notifier(&migration_notifier);
5954 /* Register cpu active notifiers */
5955 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5956 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
5960 early_initcall(migration_init);
5965 #ifdef CONFIG_SCHED_DEBUG
5967 static __read_mostly int sched_domain_debug_enabled;
5969 static int __init sched_domain_debug_setup(char *str)
5971 sched_domain_debug_enabled = 1;
5975 early_param("sched_debug", sched_domain_debug_setup);
5977 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
5978 struct cpumask *groupmask)
5980 struct sched_group *group = sd->groups;
5983 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
5984 cpumask_clear(groupmask);
5986 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
5988 if (!(sd->flags & SD_LOAD_BALANCE)) {
5989 printk("does not load-balance\n");
5991 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
5996 printk(KERN_CONT "span %s level %s\n", str, sd->name);
5998 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
5999 printk(KERN_ERR "ERROR: domain->span does not contain "
6002 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6003 printk(KERN_ERR "ERROR: domain->groups does not contain"
6007 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6011 printk(KERN_ERR "ERROR: group is NULL\n");
6015 if (!group->cpu_power) {
6016 printk(KERN_CONT "\n");
6017 printk(KERN_ERR "ERROR: domain->cpu_power not "
6022 if (!cpumask_weight(sched_group_cpus(group))) {
6023 printk(KERN_CONT "\n");
6024 printk(KERN_ERR "ERROR: empty group\n");
6028 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6029 printk(KERN_CONT "\n");
6030 printk(KERN_ERR "ERROR: repeated CPUs\n");
6034 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6036 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6038 printk(KERN_CONT " %s", str);
6039 if (group->cpu_power != SCHED_LOAD_SCALE) {
6040 printk(KERN_CONT " (cpu_power = %d)",
6044 group = group->next;
6045 } while (group != sd->groups);
6046 printk(KERN_CONT "\n");
6048 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6049 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6052 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6053 printk(KERN_ERR "ERROR: parent span is not a superset "
6054 "of domain->span\n");
6058 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6060 cpumask_var_t groupmask;
6063 if (!sched_domain_debug_enabled)
6067 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6071 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6073 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6074 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6079 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6086 free_cpumask_var(groupmask);
6088 #else /* !CONFIG_SCHED_DEBUG */
6089 # define sched_domain_debug(sd, cpu) do { } while (0)
6090 #endif /* CONFIG_SCHED_DEBUG */
6092 static int sd_degenerate(struct sched_domain *sd)
6094 if (cpumask_weight(sched_domain_span(sd)) == 1)
6097 /* Following flags need at least 2 groups */
6098 if (sd->flags & (SD_LOAD_BALANCE |
6099 SD_BALANCE_NEWIDLE |
6103 SD_SHARE_PKG_RESOURCES)) {
6104 if (sd->groups != sd->groups->next)
6108 /* Following flags don't use groups */
6109 if (sd->flags & (SD_WAKE_AFFINE))
6116 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6118 unsigned long cflags = sd->flags, pflags = parent->flags;
6120 if (sd_degenerate(parent))
6123 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6126 /* Flags needing groups don't count if only 1 group in parent */
6127 if (parent->groups == parent->groups->next) {
6128 pflags &= ~(SD_LOAD_BALANCE |
6129 SD_BALANCE_NEWIDLE |
6133 SD_SHARE_PKG_RESOURCES);
6134 if (nr_node_ids == 1)
6135 pflags &= ~SD_SERIALIZE;
6137 if (~cflags & pflags)
6143 static void free_rootdomain(struct root_domain *rd)
6145 synchronize_sched();
6147 cpupri_cleanup(&rd->cpupri);
6149 free_cpumask_var(rd->rto_mask);
6150 free_cpumask_var(rd->online);
6151 free_cpumask_var(rd->span);
6155 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6157 struct root_domain *old_rd = NULL;
6158 unsigned long flags;
6160 raw_spin_lock_irqsave(&rq->lock, flags);
6165 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6168 cpumask_clear_cpu(rq->cpu, old_rd->span);
6171 * If we dont want to free the old_rt yet then
6172 * set old_rd to NULL to skip the freeing later
6175 if (!atomic_dec_and_test(&old_rd->refcount))
6179 atomic_inc(&rd->refcount);
6182 cpumask_set_cpu(rq->cpu, rd->span);
6183 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6186 raw_spin_unlock_irqrestore(&rq->lock, flags);
6189 free_rootdomain(old_rd);
6192 static int init_rootdomain(struct root_domain *rd)
6194 memset(rd, 0, sizeof(*rd));
6196 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6198 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6200 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6203 if (cpupri_init(&rd->cpupri) != 0)
6208 free_cpumask_var(rd->rto_mask);
6210 free_cpumask_var(rd->online);
6212 free_cpumask_var(rd->span);
6217 static void init_defrootdomain(void)
6219 init_rootdomain(&def_root_domain);
6221 atomic_set(&def_root_domain.refcount, 1);
6224 static struct root_domain *alloc_rootdomain(void)
6226 struct root_domain *rd;
6228 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6232 if (init_rootdomain(rd) != 0) {
6241 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6242 * hold the hotplug lock.
6245 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6247 struct rq *rq = cpu_rq(cpu);
6248 struct sched_domain *tmp;
6250 for (tmp = sd; tmp; tmp = tmp->parent)
6251 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6253 /* Remove the sched domains which do not contribute to scheduling. */
6254 for (tmp = sd; tmp; ) {
6255 struct sched_domain *parent = tmp->parent;
6259 if (sd_parent_degenerate(tmp, parent)) {
6260 tmp->parent = parent->parent;
6262 parent->parent->child = tmp;
6267 if (sd && sd_degenerate(sd)) {
6273 sched_domain_debug(sd, cpu);
6275 rq_attach_root(rq, rd);
6276 rcu_assign_pointer(rq->sd, sd);
6279 /* cpus with isolated domains */
6280 static cpumask_var_t cpu_isolated_map;
6282 /* Setup the mask of cpus configured for isolated domains */
6283 static int __init isolated_cpu_setup(char *str)
6285 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6286 cpulist_parse(str, cpu_isolated_map);
6290 __setup("isolcpus=", isolated_cpu_setup);
6293 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6294 * to a function which identifies what group(along with sched group) a CPU
6295 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6296 * (due to the fact that we keep track of groups covered with a struct cpumask).
6298 * init_sched_build_groups will build a circular linked list of the groups
6299 * covered by the given span, and will set each group's ->cpumask correctly,
6300 * and ->cpu_power to 0.
6303 init_sched_build_groups(const struct cpumask *span,
6304 const struct cpumask *cpu_map,
6305 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6306 struct sched_group **sg,
6307 struct cpumask *tmpmask),
6308 struct cpumask *covered, struct cpumask *tmpmask)
6310 struct sched_group *first = NULL, *last = NULL;
6313 cpumask_clear(covered);
6315 for_each_cpu(i, span) {
6316 struct sched_group *sg;
6317 int group = group_fn(i, cpu_map, &sg, tmpmask);
6320 if (cpumask_test_cpu(i, covered))
6323 cpumask_clear(sched_group_cpus(sg));
6326 for_each_cpu(j, span) {
6327 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6330 cpumask_set_cpu(j, covered);
6331 cpumask_set_cpu(j, sched_group_cpus(sg));
6342 #define SD_NODES_PER_DOMAIN 16
6347 * find_next_best_node - find the next node to include in a sched_domain
6348 * @node: node whose sched_domain we're building
6349 * @used_nodes: nodes already in the sched_domain
6351 * Find the next node to include in a given scheduling domain. Simply
6352 * finds the closest node not already in the @used_nodes map.
6354 * Should use nodemask_t.
6356 static int find_next_best_node(int node, nodemask_t *used_nodes)
6358 int i, n, val, min_val, best_node = 0;
6362 for (i = 0; i < nr_node_ids; i++) {
6363 /* Start at @node */
6364 n = (node + i) % nr_node_ids;
6366 if (!nr_cpus_node(n))
6369 /* Skip already used nodes */
6370 if (node_isset(n, *used_nodes))
6373 /* Simple min distance search */
6374 val = node_distance(node, n);
6376 if (val < min_val) {
6382 node_set(best_node, *used_nodes);
6387 * sched_domain_node_span - get a cpumask for a node's sched_domain
6388 * @node: node whose cpumask we're constructing
6389 * @span: resulting cpumask
6391 * Given a node, construct a good cpumask for its sched_domain to span. It
6392 * should be one that prevents unnecessary balancing, but also spreads tasks
6395 static void sched_domain_node_span(int node, struct cpumask *span)
6397 nodemask_t used_nodes;
6400 cpumask_clear(span);
6401 nodes_clear(used_nodes);
6403 cpumask_or(span, span, cpumask_of_node(node));
6404 node_set(node, used_nodes);
6406 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6407 int next_node = find_next_best_node(node, &used_nodes);
6409 cpumask_or(span, span, cpumask_of_node(next_node));
6412 #endif /* CONFIG_NUMA */
6414 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6417 * The cpus mask in sched_group and sched_domain hangs off the end.
6419 * ( See the the comments in include/linux/sched.h:struct sched_group
6420 * and struct sched_domain. )
6422 struct static_sched_group {
6423 struct sched_group sg;
6424 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6427 struct static_sched_domain {
6428 struct sched_domain sd;
6429 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6435 cpumask_var_t domainspan;
6436 cpumask_var_t covered;
6437 cpumask_var_t notcovered;
6439 cpumask_var_t nodemask;
6440 cpumask_var_t this_sibling_map;
6441 cpumask_var_t this_core_map;
6442 cpumask_var_t this_book_map;
6443 cpumask_var_t send_covered;
6444 cpumask_var_t tmpmask;
6445 struct sched_group **sched_group_nodes;
6446 struct root_domain *rd;
6450 sa_sched_groups = 0,
6456 sa_this_sibling_map,
6458 sa_sched_group_nodes,
6468 * SMT sched-domains:
6470 #ifdef CONFIG_SCHED_SMT
6471 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6472 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6475 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6476 struct sched_group **sg, struct cpumask *unused)
6479 *sg = &per_cpu(sched_groups, cpu).sg;
6482 #endif /* CONFIG_SCHED_SMT */
6485 * multi-core sched-domains:
6487 #ifdef CONFIG_SCHED_MC
6488 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6489 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6492 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6493 struct sched_group **sg, struct cpumask *mask)
6496 #ifdef CONFIG_SCHED_SMT
6497 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6498 group = cpumask_first(mask);
6503 *sg = &per_cpu(sched_group_core, group).sg;
6506 #endif /* CONFIG_SCHED_MC */
6509 * book sched-domains:
6511 #ifdef CONFIG_SCHED_BOOK
6512 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6513 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6516 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6517 struct sched_group **sg, struct cpumask *mask)
6520 #ifdef CONFIG_SCHED_MC
6521 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6522 group = cpumask_first(mask);
6523 #elif defined(CONFIG_SCHED_SMT)
6524 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6525 group = cpumask_first(mask);
6528 *sg = &per_cpu(sched_group_book, group).sg;
6531 #endif /* CONFIG_SCHED_BOOK */
6533 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6534 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6537 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6538 struct sched_group **sg, struct cpumask *mask)
6541 #ifdef CONFIG_SCHED_BOOK
6542 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6543 group = cpumask_first(mask);
6544 #elif defined(CONFIG_SCHED_MC)
6545 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6546 group = cpumask_first(mask);
6547 #elif defined(CONFIG_SCHED_SMT)
6548 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6549 group = cpumask_first(mask);
6554 *sg = &per_cpu(sched_group_phys, group).sg;
6560 * The init_sched_build_groups can't handle what we want to do with node
6561 * groups, so roll our own. Now each node has its own list of groups which
6562 * gets dynamically allocated.
6564 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6565 static struct sched_group ***sched_group_nodes_bycpu;
6567 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6568 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6570 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6571 struct sched_group **sg,
6572 struct cpumask *nodemask)
6576 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6577 group = cpumask_first(nodemask);
6580 *sg = &per_cpu(sched_group_allnodes, group).sg;
6584 static void init_numa_sched_groups_power(struct sched_group *group_head)
6586 struct sched_group *sg = group_head;
6592 for_each_cpu(j, sched_group_cpus(sg)) {
6593 struct sched_domain *sd;
6595 sd = &per_cpu(phys_domains, j).sd;
6596 if (j != group_first_cpu(sd->groups)) {
6598 * Only add "power" once for each
6604 sg->cpu_power += sd->groups->cpu_power;
6607 } while (sg != group_head);
6610 static int build_numa_sched_groups(struct s_data *d,
6611 const struct cpumask *cpu_map, int num)
6613 struct sched_domain *sd;
6614 struct sched_group *sg, *prev;
6617 cpumask_clear(d->covered);
6618 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6619 if (cpumask_empty(d->nodemask)) {
6620 d->sched_group_nodes[num] = NULL;
6624 sched_domain_node_span(num, d->domainspan);
6625 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6627 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6630 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6634 d->sched_group_nodes[num] = sg;
6636 for_each_cpu(j, d->nodemask) {
6637 sd = &per_cpu(node_domains, j).sd;
6642 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6644 cpumask_or(d->covered, d->covered, d->nodemask);
6647 for (j = 0; j < nr_node_ids; j++) {
6648 n = (num + j) % nr_node_ids;
6649 cpumask_complement(d->notcovered, d->covered);
6650 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6651 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6652 if (cpumask_empty(d->tmpmask))
6654 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6655 if (cpumask_empty(d->tmpmask))
6657 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6661 "Can not alloc domain group for node %d\n", j);
6665 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6666 sg->next = prev->next;
6667 cpumask_or(d->covered, d->covered, d->tmpmask);
6674 #endif /* CONFIG_NUMA */
6677 /* Free memory allocated for various sched_group structures */
6678 static void free_sched_groups(const struct cpumask *cpu_map,
6679 struct cpumask *nodemask)
6683 for_each_cpu(cpu, cpu_map) {
6684 struct sched_group **sched_group_nodes
6685 = sched_group_nodes_bycpu[cpu];
6687 if (!sched_group_nodes)
6690 for (i = 0; i < nr_node_ids; i++) {
6691 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6693 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6694 if (cpumask_empty(nodemask))
6704 if (oldsg != sched_group_nodes[i])
6707 kfree(sched_group_nodes);
6708 sched_group_nodes_bycpu[cpu] = NULL;
6711 #else /* !CONFIG_NUMA */
6712 static void free_sched_groups(const struct cpumask *cpu_map,
6713 struct cpumask *nodemask)
6716 #endif /* CONFIG_NUMA */
6719 * Initialize sched groups cpu_power.
6721 * cpu_power indicates the capacity of sched group, which is used while
6722 * distributing the load between different sched groups in a sched domain.
6723 * Typically cpu_power for all the groups in a sched domain will be same unless
6724 * there are asymmetries in the topology. If there are asymmetries, group
6725 * having more cpu_power will pickup more load compared to the group having
6728 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6730 struct sched_domain *child;
6731 struct sched_group *group;
6735 WARN_ON(!sd || !sd->groups);
6737 if (cpu != group_first_cpu(sd->groups))
6742 sd->groups->cpu_power = 0;
6745 power = SCHED_LOAD_SCALE;
6746 weight = cpumask_weight(sched_domain_span(sd));
6748 * SMT siblings share the power of a single core.
6749 * Usually multiple threads get a better yield out of
6750 * that one core than a single thread would have,
6751 * reflect that in sd->smt_gain.
6753 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6754 power *= sd->smt_gain;
6756 power >>= SCHED_LOAD_SHIFT;
6758 sd->groups->cpu_power += power;
6763 * Add cpu_power of each child group to this groups cpu_power.
6765 group = child->groups;
6767 sd->groups->cpu_power += group->cpu_power;
6768 group = group->next;
6769 } while (group != child->groups);
6773 * Initializers for schedule domains
6774 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6777 #ifdef CONFIG_SCHED_DEBUG
6778 # define SD_INIT_NAME(sd, type) sd->name = #type
6780 # define SD_INIT_NAME(sd, type) do { } while (0)
6783 #define SD_INIT(sd, type) sd_init_##type(sd)
6785 #define SD_INIT_FUNC(type) \
6786 static noinline void sd_init_##type(struct sched_domain *sd) \
6788 memset(sd, 0, sizeof(*sd)); \
6789 *sd = SD_##type##_INIT; \
6790 sd->level = SD_LV_##type; \
6791 SD_INIT_NAME(sd, type); \
6796 SD_INIT_FUNC(ALLNODES)
6799 #ifdef CONFIG_SCHED_SMT
6800 SD_INIT_FUNC(SIBLING)
6802 #ifdef CONFIG_SCHED_MC
6805 #ifdef CONFIG_SCHED_BOOK
6809 static int default_relax_domain_level = -1;
6811 static int __init setup_relax_domain_level(char *str)
6815 val = simple_strtoul(str, NULL, 0);
6816 if (val < SD_LV_MAX)
6817 default_relax_domain_level = val;
6821 __setup("relax_domain_level=", setup_relax_domain_level);
6823 static void set_domain_attribute(struct sched_domain *sd,
6824 struct sched_domain_attr *attr)
6828 if (!attr || attr->relax_domain_level < 0) {
6829 if (default_relax_domain_level < 0)
6832 request = default_relax_domain_level;
6834 request = attr->relax_domain_level;
6835 if (request < sd->level) {
6836 /* turn off idle balance on this domain */
6837 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6839 /* turn on idle balance on this domain */
6840 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6844 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6845 const struct cpumask *cpu_map)
6848 case sa_sched_groups:
6849 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6850 d->sched_group_nodes = NULL;
6852 free_rootdomain(d->rd); /* fall through */
6854 free_cpumask_var(d->tmpmask); /* fall through */
6855 case sa_send_covered:
6856 free_cpumask_var(d->send_covered); /* fall through */
6857 case sa_this_book_map:
6858 free_cpumask_var(d->this_book_map); /* fall through */
6859 case sa_this_core_map:
6860 free_cpumask_var(d->this_core_map); /* fall through */
6861 case sa_this_sibling_map:
6862 free_cpumask_var(d->this_sibling_map); /* fall through */
6864 free_cpumask_var(d->nodemask); /* fall through */
6865 case sa_sched_group_nodes:
6867 kfree(d->sched_group_nodes); /* fall through */
6869 free_cpumask_var(d->notcovered); /* fall through */
6871 free_cpumask_var(d->covered); /* fall through */
6873 free_cpumask_var(d->domainspan); /* fall through */
6880 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6881 const struct cpumask *cpu_map)
6884 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6886 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6887 return sa_domainspan;
6888 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6890 /* Allocate the per-node list of sched groups */
6891 d->sched_group_nodes = kcalloc(nr_node_ids,
6892 sizeof(struct sched_group *), GFP_KERNEL);
6893 if (!d->sched_group_nodes) {
6894 printk(KERN_WARNING "Can not alloc sched group node list\n");
6895 return sa_notcovered;
6897 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6899 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6900 return sa_sched_group_nodes;
6901 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6903 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6904 return sa_this_sibling_map;
6905 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
6906 return sa_this_core_map;
6907 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6908 return sa_this_book_map;
6909 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6910 return sa_send_covered;
6911 d->rd = alloc_rootdomain();
6913 printk(KERN_WARNING "Cannot alloc root domain\n");
6916 return sa_rootdomain;
6919 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6920 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6922 struct sched_domain *sd = NULL;
6924 struct sched_domain *parent;
6927 if (cpumask_weight(cpu_map) >
6928 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6929 sd = &per_cpu(allnodes_domains, i).sd;
6930 SD_INIT(sd, ALLNODES);
6931 set_domain_attribute(sd, attr);
6932 cpumask_copy(sched_domain_span(sd), cpu_map);
6933 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6938 sd = &per_cpu(node_domains, i).sd;
6940 set_domain_attribute(sd, attr);
6941 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6942 sd->parent = parent;
6945 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6950 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6951 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6952 struct sched_domain *parent, int i)
6954 struct sched_domain *sd;
6955 sd = &per_cpu(phys_domains, i).sd;
6957 set_domain_attribute(sd, attr);
6958 cpumask_copy(sched_domain_span(sd), d->nodemask);
6959 sd->parent = parent;
6962 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6966 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
6967 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6968 struct sched_domain *parent, int i)
6970 struct sched_domain *sd = parent;
6971 #ifdef CONFIG_SCHED_BOOK
6972 sd = &per_cpu(book_domains, i).sd;
6974 set_domain_attribute(sd, attr);
6975 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
6976 sd->parent = parent;
6978 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
6983 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6984 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6985 struct sched_domain *parent, int i)
6987 struct sched_domain *sd = parent;
6988 #ifdef CONFIG_SCHED_MC
6989 sd = &per_cpu(core_domains, i).sd;
6991 set_domain_attribute(sd, attr);
6992 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6993 sd->parent = parent;
6995 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7000 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7001 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7002 struct sched_domain *parent, int i)
7004 struct sched_domain *sd = parent;
7005 #ifdef CONFIG_SCHED_SMT
7006 sd = &per_cpu(cpu_domains, i).sd;
7007 SD_INIT(sd, SIBLING);
7008 set_domain_attribute(sd, attr);
7009 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7010 sd->parent = parent;
7012 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7017 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7018 const struct cpumask *cpu_map, int cpu)
7021 #ifdef CONFIG_SCHED_SMT
7022 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7023 cpumask_and(d->this_sibling_map, cpu_map,
7024 topology_thread_cpumask(cpu));
7025 if (cpu == cpumask_first(d->this_sibling_map))
7026 init_sched_build_groups(d->this_sibling_map, cpu_map,
7028 d->send_covered, d->tmpmask);
7031 #ifdef CONFIG_SCHED_MC
7032 case SD_LV_MC: /* set up multi-core groups */
7033 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7034 if (cpu == cpumask_first(d->this_core_map))
7035 init_sched_build_groups(d->this_core_map, cpu_map,
7037 d->send_covered, d->tmpmask);
7040 #ifdef CONFIG_SCHED_BOOK
7041 case SD_LV_BOOK: /* set up book groups */
7042 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7043 if (cpu == cpumask_first(d->this_book_map))
7044 init_sched_build_groups(d->this_book_map, cpu_map,
7046 d->send_covered, d->tmpmask);
7049 case SD_LV_CPU: /* set up physical groups */
7050 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7051 if (!cpumask_empty(d->nodemask))
7052 init_sched_build_groups(d->nodemask, cpu_map,
7054 d->send_covered, d->tmpmask);
7057 case SD_LV_ALLNODES:
7058 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7059 d->send_covered, d->tmpmask);
7068 * Build sched domains for a given set of cpus and attach the sched domains
7069 * to the individual cpus
7071 static int __build_sched_domains(const struct cpumask *cpu_map,
7072 struct sched_domain_attr *attr)
7074 enum s_alloc alloc_state = sa_none;
7076 struct sched_domain *sd;
7082 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7083 if (alloc_state != sa_rootdomain)
7085 alloc_state = sa_sched_groups;
7088 * Set up domains for cpus specified by the cpu_map.
7090 for_each_cpu(i, cpu_map) {
7091 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7094 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7095 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7096 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7097 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7098 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7101 for_each_cpu(i, cpu_map) {
7102 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7103 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7104 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7107 /* Set up physical groups */
7108 for (i = 0; i < nr_node_ids; i++)
7109 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7112 /* Set up node groups */
7114 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7116 for (i = 0; i < nr_node_ids; i++)
7117 if (build_numa_sched_groups(&d, cpu_map, i))
7121 /* Calculate CPU power for physical packages and nodes */
7122 #ifdef CONFIG_SCHED_SMT
7123 for_each_cpu(i, cpu_map) {
7124 sd = &per_cpu(cpu_domains, i).sd;
7125 init_sched_groups_power(i, sd);
7128 #ifdef CONFIG_SCHED_MC
7129 for_each_cpu(i, cpu_map) {
7130 sd = &per_cpu(core_domains, i).sd;
7131 init_sched_groups_power(i, sd);
7134 #ifdef CONFIG_SCHED_BOOK
7135 for_each_cpu(i, cpu_map) {
7136 sd = &per_cpu(book_domains, i).sd;
7137 init_sched_groups_power(i, sd);
7141 for_each_cpu(i, cpu_map) {
7142 sd = &per_cpu(phys_domains, i).sd;
7143 init_sched_groups_power(i, sd);
7147 for (i = 0; i < nr_node_ids; i++)
7148 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7150 if (d.sd_allnodes) {
7151 struct sched_group *sg;
7153 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7155 init_numa_sched_groups_power(sg);
7159 /* Attach the domains */
7160 for_each_cpu(i, cpu_map) {
7161 #ifdef CONFIG_SCHED_SMT
7162 sd = &per_cpu(cpu_domains, i).sd;
7163 #elif defined(CONFIG_SCHED_MC)
7164 sd = &per_cpu(core_domains, i).sd;
7165 #elif defined(CONFIG_SCHED_BOOK)
7166 sd = &per_cpu(book_domains, i).sd;
7168 sd = &per_cpu(phys_domains, i).sd;
7170 cpu_attach_domain(sd, d.rd, i);
7173 d.sched_group_nodes = NULL; /* don't free this we still need it */
7174 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7178 __free_domain_allocs(&d, alloc_state, cpu_map);
7182 static int build_sched_domains(const struct cpumask *cpu_map)
7184 return __build_sched_domains(cpu_map, NULL);
7187 static cpumask_var_t *doms_cur; /* current sched domains */
7188 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7189 static struct sched_domain_attr *dattr_cur;
7190 /* attribues of custom domains in 'doms_cur' */
7193 * Special case: If a kmalloc of a doms_cur partition (array of
7194 * cpumask) fails, then fallback to a single sched domain,
7195 * as determined by the single cpumask fallback_doms.
7197 static cpumask_var_t fallback_doms;
7200 * arch_update_cpu_topology lets virtualized architectures update the
7201 * cpu core maps. It is supposed to return 1 if the topology changed
7202 * or 0 if it stayed the same.
7204 int __attribute__((weak)) arch_update_cpu_topology(void)
7209 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7212 cpumask_var_t *doms;
7214 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7217 for (i = 0; i < ndoms; i++) {
7218 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7219 free_sched_domains(doms, i);
7226 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7229 for (i = 0; i < ndoms; i++)
7230 free_cpumask_var(doms[i]);
7235 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7236 * For now this just excludes isolated cpus, but could be used to
7237 * exclude other special cases in the future.
7239 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7243 arch_update_cpu_topology();
7245 doms_cur = alloc_sched_domains(ndoms_cur);
7247 doms_cur = &fallback_doms;
7248 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7250 err = build_sched_domains(doms_cur[0]);
7251 register_sched_domain_sysctl();
7256 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7257 struct cpumask *tmpmask)
7259 free_sched_groups(cpu_map, tmpmask);
7263 * Detach sched domains from a group of cpus specified in cpu_map
7264 * These cpus will now be attached to the NULL domain
7266 static void detach_destroy_domains(const struct cpumask *cpu_map)
7268 /* Save because hotplug lock held. */
7269 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7272 for_each_cpu(i, cpu_map)
7273 cpu_attach_domain(NULL, &def_root_domain, i);
7274 synchronize_sched();
7275 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7278 /* handle null as "default" */
7279 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7280 struct sched_domain_attr *new, int idx_new)
7282 struct sched_domain_attr tmp;
7289 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7290 new ? (new + idx_new) : &tmp,
7291 sizeof(struct sched_domain_attr));
7295 * Partition sched domains as specified by the 'ndoms_new'
7296 * cpumasks in the array doms_new[] of cpumasks. This compares
7297 * doms_new[] to the current sched domain partitioning, doms_cur[].
7298 * It destroys each deleted domain and builds each new domain.
7300 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7301 * The masks don't intersect (don't overlap.) We should setup one
7302 * sched domain for each mask. CPUs not in any of the cpumasks will
7303 * not be load balanced. If the same cpumask appears both in the
7304 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7307 * The passed in 'doms_new' should be allocated using
7308 * alloc_sched_domains. This routine takes ownership of it and will
7309 * free_sched_domains it when done with it. If the caller failed the
7310 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7311 * and partition_sched_domains() will fallback to the single partition
7312 * 'fallback_doms', it also forces the domains to be rebuilt.
7314 * If doms_new == NULL it will be replaced with cpu_online_mask.
7315 * ndoms_new == 0 is a special case for destroying existing domains,
7316 * and it will not create the default domain.
7318 * Call with hotplug lock held
7320 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7321 struct sched_domain_attr *dattr_new)
7326 mutex_lock(&sched_domains_mutex);
7328 /* always unregister in case we don't destroy any domains */
7329 unregister_sched_domain_sysctl();
7331 /* Let architecture update cpu core mappings. */
7332 new_topology = arch_update_cpu_topology();
7334 n = doms_new ? ndoms_new : 0;
7336 /* Destroy deleted domains */
7337 for (i = 0; i < ndoms_cur; i++) {
7338 for (j = 0; j < n && !new_topology; j++) {
7339 if (cpumask_equal(doms_cur[i], doms_new[j])
7340 && dattrs_equal(dattr_cur, i, dattr_new, j))
7343 /* no match - a current sched domain not in new doms_new[] */
7344 detach_destroy_domains(doms_cur[i]);
7349 if (doms_new == NULL) {
7351 doms_new = &fallback_doms;
7352 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7353 WARN_ON_ONCE(dattr_new);
7356 /* Build new domains */
7357 for (i = 0; i < ndoms_new; i++) {
7358 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7359 if (cpumask_equal(doms_new[i], doms_cur[j])
7360 && dattrs_equal(dattr_new, i, dattr_cur, j))
7363 /* no match - add a new doms_new */
7364 __build_sched_domains(doms_new[i],
7365 dattr_new ? dattr_new + i : NULL);
7370 /* Remember the new sched domains */
7371 if (doms_cur != &fallback_doms)
7372 free_sched_domains(doms_cur, ndoms_cur);
7373 kfree(dattr_cur); /* kfree(NULL) is safe */
7374 doms_cur = doms_new;
7375 dattr_cur = dattr_new;
7376 ndoms_cur = ndoms_new;
7378 register_sched_domain_sysctl();
7380 mutex_unlock(&sched_domains_mutex);
7383 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7384 static void arch_reinit_sched_domains(void)
7388 /* Destroy domains first to force the rebuild */
7389 partition_sched_domains(0, NULL, NULL);
7391 rebuild_sched_domains();
7395 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7397 unsigned int level = 0;
7399 if (sscanf(buf, "%u", &level) != 1)
7403 * level is always be positive so don't check for
7404 * level < POWERSAVINGS_BALANCE_NONE which is 0
7405 * What happens on 0 or 1 byte write,
7406 * need to check for count as well?
7409 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7413 sched_smt_power_savings = level;
7415 sched_mc_power_savings = level;
7417 arch_reinit_sched_domains();
7422 #ifdef CONFIG_SCHED_MC
7423 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7424 struct sysdev_class_attribute *attr,
7427 return sprintf(page, "%u\n", sched_mc_power_savings);
7429 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7430 struct sysdev_class_attribute *attr,
7431 const char *buf, size_t count)
7433 return sched_power_savings_store(buf, count, 0);
7435 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7436 sched_mc_power_savings_show,
7437 sched_mc_power_savings_store);
7440 #ifdef CONFIG_SCHED_SMT
7441 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7442 struct sysdev_class_attribute *attr,
7445 return sprintf(page, "%u\n", sched_smt_power_savings);
7447 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7448 struct sysdev_class_attribute *attr,
7449 const char *buf, size_t count)
7451 return sched_power_savings_store(buf, count, 1);
7453 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7454 sched_smt_power_savings_show,
7455 sched_smt_power_savings_store);
7458 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7462 #ifdef CONFIG_SCHED_SMT
7464 err = sysfs_create_file(&cls->kset.kobj,
7465 &attr_sched_smt_power_savings.attr);
7467 #ifdef CONFIG_SCHED_MC
7468 if (!err && mc_capable())
7469 err = sysfs_create_file(&cls->kset.kobj,
7470 &attr_sched_mc_power_savings.attr);
7474 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7477 * Update cpusets according to cpu_active mask. If cpusets are
7478 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7479 * around partition_sched_domains().
7481 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7484 switch (action & ~CPU_TASKS_FROZEN) {
7486 case CPU_DOWN_FAILED:
7487 cpuset_update_active_cpus();
7494 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7497 switch (action & ~CPU_TASKS_FROZEN) {
7498 case CPU_DOWN_PREPARE:
7499 cpuset_update_active_cpus();
7506 static int update_runtime(struct notifier_block *nfb,
7507 unsigned long action, void *hcpu)
7509 int cpu = (int)(long)hcpu;
7512 case CPU_DOWN_PREPARE:
7513 case CPU_DOWN_PREPARE_FROZEN:
7514 disable_runtime(cpu_rq(cpu));
7517 case CPU_DOWN_FAILED:
7518 case CPU_DOWN_FAILED_FROZEN:
7520 case CPU_ONLINE_FROZEN:
7521 enable_runtime(cpu_rq(cpu));
7529 void __init sched_init_smp(void)
7531 cpumask_var_t non_isolated_cpus;
7533 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7534 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7536 #if defined(CONFIG_NUMA)
7537 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7539 BUG_ON(sched_group_nodes_bycpu == NULL);
7542 mutex_lock(&sched_domains_mutex);
7543 arch_init_sched_domains(cpu_active_mask);
7544 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7545 if (cpumask_empty(non_isolated_cpus))
7546 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7547 mutex_unlock(&sched_domains_mutex);
7550 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7551 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7553 /* RT runtime code needs to handle some hotplug events */
7554 hotcpu_notifier(update_runtime, 0);
7558 /* Move init over to a non-isolated CPU */
7559 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7561 sched_init_granularity();
7562 free_cpumask_var(non_isolated_cpus);
7564 init_sched_rt_class();
7567 void __init sched_init_smp(void)
7569 sched_init_granularity();
7571 #endif /* CONFIG_SMP */
7573 const_debug unsigned int sysctl_timer_migration = 1;
7575 int in_sched_functions(unsigned long addr)
7577 return in_lock_functions(addr) ||
7578 (addr >= (unsigned long)__sched_text_start
7579 && addr < (unsigned long)__sched_text_end);
7582 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7584 cfs_rq->tasks_timeline = RB_ROOT;
7585 INIT_LIST_HEAD(&cfs_rq->tasks);
7586 #ifdef CONFIG_FAIR_GROUP_SCHED
7589 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7592 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7594 struct rt_prio_array *array;
7597 array = &rt_rq->active;
7598 for (i = 0; i < MAX_RT_PRIO; i++) {
7599 INIT_LIST_HEAD(array->queue + i);
7600 __clear_bit(i, array->bitmap);
7602 /* delimiter for bitsearch: */
7603 __set_bit(MAX_RT_PRIO, array->bitmap);
7605 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7606 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7608 rt_rq->highest_prio.next = MAX_RT_PRIO;
7612 rt_rq->rt_nr_migratory = 0;
7613 rt_rq->overloaded = 0;
7614 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7618 rt_rq->rt_throttled = 0;
7619 rt_rq->rt_runtime = 0;
7620 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7622 #ifdef CONFIG_RT_GROUP_SCHED
7623 rt_rq->rt_nr_boosted = 0;
7628 #ifdef CONFIG_FAIR_GROUP_SCHED
7629 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7630 struct sched_entity *se, int cpu,
7631 struct sched_entity *parent)
7633 struct rq *rq = cpu_rq(cpu);
7634 tg->cfs_rq[cpu] = cfs_rq;
7635 init_cfs_rq(cfs_rq, rq);
7639 /* se could be NULL for init_task_group */
7644 se->cfs_rq = &rq->cfs;
7646 se->cfs_rq = parent->my_q;
7649 update_load_set(&se->load, tg->shares);
7650 se->parent = parent;
7654 #ifdef CONFIG_RT_GROUP_SCHED
7655 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7656 struct sched_rt_entity *rt_se, int cpu,
7657 struct sched_rt_entity *parent)
7659 struct rq *rq = cpu_rq(cpu);
7661 tg->rt_rq[cpu] = rt_rq;
7662 init_rt_rq(rt_rq, rq);
7664 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7666 tg->rt_se[cpu] = rt_se;
7671 rt_se->rt_rq = &rq->rt;
7673 rt_se->rt_rq = parent->my_q;
7675 rt_se->my_q = rt_rq;
7676 rt_se->parent = parent;
7677 INIT_LIST_HEAD(&rt_se->run_list);
7681 void __init sched_init(void)
7684 unsigned long alloc_size = 0, ptr;
7686 #ifdef CONFIG_FAIR_GROUP_SCHED
7687 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7689 #ifdef CONFIG_RT_GROUP_SCHED
7690 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7692 #ifdef CONFIG_CPUMASK_OFFSTACK
7693 alloc_size += num_possible_cpus() * cpumask_size();
7696 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7698 #ifdef CONFIG_FAIR_GROUP_SCHED
7699 init_task_group.se = (struct sched_entity **)ptr;
7700 ptr += nr_cpu_ids * sizeof(void **);
7702 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7703 ptr += nr_cpu_ids * sizeof(void **);
7705 #endif /* CONFIG_FAIR_GROUP_SCHED */
7706 #ifdef CONFIG_RT_GROUP_SCHED
7707 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7708 ptr += nr_cpu_ids * sizeof(void **);
7710 init_task_group.rt_rq = (struct rt_rq **)ptr;
7711 ptr += nr_cpu_ids * sizeof(void **);
7713 #endif /* CONFIG_RT_GROUP_SCHED */
7714 #ifdef CONFIG_CPUMASK_OFFSTACK
7715 for_each_possible_cpu(i) {
7716 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7717 ptr += cpumask_size();
7719 #endif /* CONFIG_CPUMASK_OFFSTACK */
7723 init_defrootdomain();
7726 init_rt_bandwidth(&def_rt_bandwidth,
7727 global_rt_period(), global_rt_runtime());
7729 #ifdef CONFIG_RT_GROUP_SCHED
7730 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7731 global_rt_period(), global_rt_runtime());
7732 #endif /* CONFIG_RT_GROUP_SCHED */
7734 #ifdef CONFIG_CGROUP_SCHED
7735 list_add(&init_task_group.list, &task_groups);
7736 INIT_LIST_HEAD(&init_task_group.children);
7738 #endif /* CONFIG_CGROUP_SCHED */
7740 for_each_possible_cpu(i) {
7744 raw_spin_lock_init(&rq->lock);
7746 rq->calc_load_active = 0;
7747 rq->calc_load_update = jiffies + LOAD_FREQ;
7748 init_cfs_rq(&rq->cfs, rq);
7749 init_rt_rq(&rq->rt, rq);
7750 #ifdef CONFIG_FAIR_GROUP_SCHED
7751 init_task_group.shares = init_task_group_load;
7752 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7753 #ifdef CONFIG_CGROUP_SCHED
7755 * How much cpu bandwidth does init_task_group get?
7757 * In case of task-groups formed thr' the cgroup filesystem, it
7758 * gets 100% of the cpu resources in the system. This overall
7759 * system cpu resource is divided among the tasks of
7760 * init_task_group and its child task-groups in a fair manner,
7761 * based on each entity's (task or task-group's) weight
7762 * (se->load.weight).
7764 * In other words, if init_task_group has 10 tasks of weight
7765 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7766 * then A0's share of the cpu resource is:
7768 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7770 * We achieve this by letting init_task_group's tasks sit
7771 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7773 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, NULL);
7775 #endif /* CONFIG_FAIR_GROUP_SCHED */
7777 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7778 #ifdef CONFIG_RT_GROUP_SCHED
7779 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7780 #ifdef CONFIG_CGROUP_SCHED
7781 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, NULL);
7785 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7786 rq->cpu_load[j] = 0;
7788 rq->last_load_update_tick = jiffies;
7793 rq->cpu_power = SCHED_LOAD_SCALE;
7794 rq->post_schedule = 0;
7795 rq->active_balance = 0;
7796 rq->next_balance = jiffies;
7801 rq->avg_idle = 2*sysctl_sched_migration_cost;
7802 rq_attach_root(rq, &def_root_domain);
7804 rq->nohz_balance_kick = 0;
7805 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7809 atomic_set(&rq->nr_iowait, 0);
7812 set_load_weight(&init_task);
7814 #ifdef CONFIG_PREEMPT_NOTIFIERS
7815 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7819 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7822 #ifdef CONFIG_RT_MUTEXES
7823 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7827 * The boot idle thread does lazy MMU switching as well:
7829 atomic_inc(&init_mm.mm_count);
7830 enter_lazy_tlb(&init_mm, current);
7833 * Make us the idle thread. Technically, schedule() should not be
7834 * called from this thread, however somewhere below it might be,
7835 * but because we are the idle thread, we just pick up running again
7836 * when this runqueue becomes "idle".
7838 init_idle(current, smp_processor_id());
7840 calc_load_update = jiffies + LOAD_FREQ;
7843 * During early bootup we pretend to be a normal task:
7845 current->sched_class = &fair_sched_class;
7847 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7848 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7851 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7852 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
7853 atomic_set(&nohz.load_balancer, nr_cpu_ids);
7854 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
7855 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
7857 /* May be allocated at isolcpus cmdline parse time */
7858 if (cpu_isolated_map == NULL)
7859 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7864 scheduler_running = 1;
7867 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7868 static inline int preempt_count_equals(int preempt_offset)
7870 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7872 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7875 void __might_sleep(const char *file, int line, int preempt_offset)
7878 static unsigned long prev_jiffy; /* ratelimiting */
7880 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7881 system_state != SYSTEM_RUNNING || oops_in_progress)
7883 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7885 prev_jiffy = jiffies;
7888 "BUG: sleeping function called from invalid context at %s:%d\n",
7891 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7892 in_atomic(), irqs_disabled(),
7893 current->pid, current->comm);
7895 debug_show_held_locks(current);
7896 if (irqs_disabled())
7897 print_irqtrace_events(current);
7901 EXPORT_SYMBOL(__might_sleep);
7904 #ifdef CONFIG_MAGIC_SYSRQ
7905 static void normalize_task(struct rq *rq, struct task_struct *p)
7909 on_rq = p->se.on_rq;
7911 deactivate_task(rq, p, 0);
7912 __setscheduler(rq, p, SCHED_NORMAL, 0);
7914 activate_task(rq, p, 0);
7915 resched_task(rq->curr);
7919 void normalize_rt_tasks(void)
7921 struct task_struct *g, *p;
7922 unsigned long flags;
7925 read_lock_irqsave(&tasklist_lock, flags);
7926 do_each_thread(g, p) {
7928 * Only normalize user tasks:
7933 p->se.exec_start = 0;
7934 #ifdef CONFIG_SCHEDSTATS
7935 p->se.statistics.wait_start = 0;
7936 p->se.statistics.sleep_start = 0;
7937 p->se.statistics.block_start = 0;
7942 * Renice negative nice level userspace
7945 if (TASK_NICE(p) < 0 && p->mm)
7946 set_user_nice(p, 0);
7950 raw_spin_lock(&p->pi_lock);
7951 rq = __task_rq_lock(p);
7953 normalize_task(rq, p);
7955 __task_rq_unlock(rq);
7956 raw_spin_unlock(&p->pi_lock);
7957 } while_each_thread(g, p);
7959 read_unlock_irqrestore(&tasklist_lock, flags);
7962 #endif /* CONFIG_MAGIC_SYSRQ */
7964 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7966 * These functions are only useful for the IA64 MCA handling, or kdb.
7968 * They can only be called when the whole system has been
7969 * stopped - every CPU needs to be quiescent, and no scheduling
7970 * activity can take place. Using them for anything else would
7971 * be a serious bug, and as a result, they aren't even visible
7972 * under any other configuration.
7976 * curr_task - return the current task for a given cpu.
7977 * @cpu: the processor in question.
7979 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7981 struct task_struct *curr_task(int cpu)
7983 return cpu_curr(cpu);
7986 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7990 * set_curr_task - set the current task for a given cpu.
7991 * @cpu: the processor in question.
7992 * @p: the task pointer to set.
7994 * Description: This function must only be used when non-maskable interrupts
7995 * are serviced on a separate stack. It allows the architecture to switch the
7996 * notion of the current task on a cpu in a non-blocking manner. This function
7997 * must be called with all CPU's synchronized, and interrupts disabled, the
7998 * and caller must save the original value of the current task (see
7999 * curr_task() above) and restore that value before reenabling interrupts and
8000 * re-starting the system.
8002 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8004 void set_curr_task(int cpu, struct task_struct *p)
8011 #ifdef CONFIG_FAIR_GROUP_SCHED
8012 static void free_fair_sched_group(struct task_group *tg)
8016 for_each_possible_cpu(i) {
8018 kfree(tg->cfs_rq[i]);
8028 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8030 struct cfs_rq *cfs_rq;
8031 struct sched_entity *se;
8035 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8038 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8042 tg->shares = NICE_0_LOAD;
8044 for_each_possible_cpu(i) {
8047 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8048 GFP_KERNEL, cpu_to_node(i));
8052 se = kzalloc_node(sizeof(struct sched_entity),
8053 GFP_KERNEL, cpu_to_node(i));
8057 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8068 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8070 struct rq *rq = cpu_rq(cpu);
8071 unsigned long flags;
8075 * Only empty task groups can be destroyed; so we can speculatively
8076 * check on_list without danger of it being re-added.
8078 if (!tg->cfs_rq[cpu]->on_list)
8081 raw_spin_lock_irqsave(&rq->lock, flags);
8082 list_del_leaf_cfs_rq(tg->cfs_rq[i]);
8083 raw_spin_unlock_irqrestore(&rq->lock, flags);
8085 #else /* !CONFG_FAIR_GROUP_SCHED */
8086 static inline void free_fair_sched_group(struct task_group *tg)
8091 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8096 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8099 #endif /* CONFIG_FAIR_GROUP_SCHED */
8101 #ifdef CONFIG_RT_GROUP_SCHED
8102 static void free_rt_sched_group(struct task_group *tg)
8106 destroy_rt_bandwidth(&tg->rt_bandwidth);
8108 for_each_possible_cpu(i) {
8110 kfree(tg->rt_rq[i]);
8112 kfree(tg->rt_se[i]);
8120 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8122 struct rt_rq *rt_rq;
8123 struct sched_rt_entity *rt_se;
8127 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8130 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8134 init_rt_bandwidth(&tg->rt_bandwidth,
8135 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8137 for_each_possible_cpu(i) {
8140 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8141 GFP_KERNEL, cpu_to_node(i));
8145 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8146 GFP_KERNEL, cpu_to_node(i));
8150 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8160 #else /* !CONFIG_RT_GROUP_SCHED */
8161 static inline void free_rt_sched_group(struct task_group *tg)
8166 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8170 #endif /* CONFIG_RT_GROUP_SCHED */
8172 #ifdef CONFIG_CGROUP_SCHED
8173 static void free_sched_group(struct task_group *tg)
8175 free_fair_sched_group(tg);
8176 free_rt_sched_group(tg);
8180 /* allocate runqueue etc for a new task group */
8181 struct task_group *sched_create_group(struct task_group *parent)
8183 struct task_group *tg;
8184 unsigned long flags;
8186 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8188 return ERR_PTR(-ENOMEM);
8190 if (!alloc_fair_sched_group(tg, parent))
8193 if (!alloc_rt_sched_group(tg, parent))
8196 spin_lock_irqsave(&task_group_lock, flags);
8197 list_add_rcu(&tg->list, &task_groups);
8199 WARN_ON(!parent); /* root should already exist */
8201 tg->parent = parent;
8202 INIT_LIST_HEAD(&tg->children);
8203 list_add_rcu(&tg->siblings, &parent->children);
8204 spin_unlock_irqrestore(&task_group_lock, flags);
8209 free_sched_group(tg);
8210 return ERR_PTR(-ENOMEM);
8213 /* rcu callback to free various structures associated with a task group */
8214 static void free_sched_group_rcu(struct rcu_head *rhp)
8216 /* now it should be safe to free those cfs_rqs */
8217 free_sched_group(container_of(rhp, struct task_group, rcu));
8220 /* Destroy runqueue etc associated with a task group */
8221 void sched_destroy_group(struct task_group *tg)
8223 unsigned long flags;
8226 /* end participation in shares distribution */
8227 for_each_possible_cpu(i)
8228 unregister_fair_sched_group(tg, i);
8230 spin_lock_irqsave(&task_group_lock, flags);
8231 list_del_rcu(&tg->list);
8232 list_del_rcu(&tg->siblings);
8233 spin_unlock_irqrestore(&task_group_lock, flags);
8235 /* wait for possible concurrent references to cfs_rqs complete */
8236 call_rcu(&tg->rcu, free_sched_group_rcu);
8239 /* change task's runqueue when it moves between groups.
8240 * The caller of this function should have put the task in its new group
8241 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8242 * reflect its new group.
8244 void sched_move_task(struct task_struct *tsk)
8247 unsigned long flags;
8250 rq = task_rq_lock(tsk, &flags);
8252 running = task_current(rq, tsk);
8253 on_rq = tsk->se.on_rq;
8256 dequeue_task(rq, tsk, 0);
8257 if (unlikely(running))
8258 tsk->sched_class->put_prev_task(rq, tsk);
8260 #ifdef CONFIG_FAIR_GROUP_SCHED
8261 if (tsk->sched_class->task_move_group)
8262 tsk->sched_class->task_move_group(tsk, on_rq);
8265 set_task_rq(tsk, task_cpu(tsk));
8267 if (unlikely(running))
8268 tsk->sched_class->set_curr_task(rq);
8270 enqueue_task(rq, tsk, 0);
8272 task_rq_unlock(rq, &flags);
8274 #endif /* CONFIG_CGROUP_SCHED */
8276 #ifdef CONFIG_FAIR_GROUP_SCHED
8277 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8279 struct cfs_rq *cfs_rq = se->cfs_rq;
8284 dequeue_entity(cfs_rq, se, 0);
8286 update_load_set(&se->load, shares);
8289 enqueue_entity(cfs_rq, se, 0);
8292 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8294 struct cfs_rq *cfs_rq = se->cfs_rq;
8295 struct rq *rq = cfs_rq->rq;
8296 unsigned long flags;
8298 raw_spin_lock_irqsave(&rq->lock, flags);
8299 __set_se_shares(se, shares);
8300 raw_spin_unlock_irqrestore(&rq->lock, flags);
8303 static DEFINE_MUTEX(shares_mutex);
8305 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8310 * We can't change the weight of the root cgroup.
8315 if (shares < MIN_SHARES)
8316 shares = MIN_SHARES;
8317 else if (shares > MAX_SHARES)
8318 shares = MAX_SHARES;
8320 mutex_lock(&shares_mutex);
8321 if (tg->shares == shares)
8324 tg->shares = shares;
8325 for_each_possible_cpu(i) {
8329 set_se_shares(tg->se[i], shares);
8333 mutex_unlock(&shares_mutex);
8337 unsigned long sched_group_shares(struct task_group *tg)
8343 #ifdef CONFIG_RT_GROUP_SCHED
8345 * Ensure that the real time constraints are schedulable.
8347 static DEFINE_MUTEX(rt_constraints_mutex);
8349 static unsigned long to_ratio(u64 period, u64 runtime)
8351 if (runtime == RUNTIME_INF)
8354 return div64_u64(runtime << 20, period);
8357 /* Must be called with tasklist_lock held */
8358 static inline int tg_has_rt_tasks(struct task_group *tg)
8360 struct task_struct *g, *p;
8362 do_each_thread(g, p) {
8363 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8365 } while_each_thread(g, p);
8370 struct rt_schedulable_data {
8371 struct task_group *tg;
8376 static int tg_schedulable(struct task_group *tg, void *data)
8378 struct rt_schedulable_data *d = data;
8379 struct task_group *child;
8380 unsigned long total, sum = 0;
8381 u64 period, runtime;
8383 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8384 runtime = tg->rt_bandwidth.rt_runtime;
8387 period = d->rt_period;
8388 runtime = d->rt_runtime;
8392 * Cannot have more runtime than the period.
8394 if (runtime > period && runtime != RUNTIME_INF)
8398 * Ensure we don't starve existing RT tasks.
8400 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8403 total = to_ratio(period, runtime);
8406 * Nobody can have more than the global setting allows.
8408 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8412 * The sum of our children's runtime should not exceed our own.
8414 list_for_each_entry_rcu(child, &tg->children, siblings) {
8415 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8416 runtime = child->rt_bandwidth.rt_runtime;
8418 if (child == d->tg) {
8419 period = d->rt_period;
8420 runtime = d->rt_runtime;
8423 sum += to_ratio(period, runtime);
8432 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8434 struct rt_schedulable_data data = {
8436 .rt_period = period,
8437 .rt_runtime = runtime,
8440 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8443 static int tg_set_bandwidth(struct task_group *tg,
8444 u64 rt_period, u64 rt_runtime)
8448 mutex_lock(&rt_constraints_mutex);
8449 read_lock(&tasklist_lock);
8450 err = __rt_schedulable(tg, rt_period, rt_runtime);
8454 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8455 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8456 tg->rt_bandwidth.rt_runtime = rt_runtime;
8458 for_each_possible_cpu(i) {
8459 struct rt_rq *rt_rq = tg->rt_rq[i];
8461 raw_spin_lock(&rt_rq->rt_runtime_lock);
8462 rt_rq->rt_runtime = rt_runtime;
8463 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8465 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8467 read_unlock(&tasklist_lock);
8468 mutex_unlock(&rt_constraints_mutex);
8473 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8475 u64 rt_runtime, rt_period;
8477 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8478 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8479 if (rt_runtime_us < 0)
8480 rt_runtime = RUNTIME_INF;
8482 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8485 long sched_group_rt_runtime(struct task_group *tg)
8489 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8492 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8493 do_div(rt_runtime_us, NSEC_PER_USEC);
8494 return rt_runtime_us;
8497 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8499 u64 rt_runtime, rt_period;
8501 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8502 rt_runtime = tg->rt_bandwidth.rt_runtime;
8507 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8510 long sched_group_rt_period(struct task_group *tg)
8514 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8515 do_div(rt_period_us, NSEC_PER_USEC);
8516 return rt_period_us;
8519 static int sched_rt_global_constraints(void)
8521 u64 runtime, period;
8524 if (sysctl_sched_rt_period <= 0)
8527 runtime = global_rt_runtime();
8528 period = global_rt_period();
8531 * Sanity check on the sysctl variables.
8533 if (runtime > period && runtime != RUNTIME_INF)
8536 mutex_lock(&rt_constraints_mutex);
8537 read_lock(&tasklist_lock);
8538 ret = __rt_schedulable(NULL, 0, 0);
8539 read_unlock(&tasklist_lock);
8540 mutex_unlock(&rt_constraints_mutex);
8545 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8547 /* Don't accept realtime tasks when there is no way for them to run */
8548 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8554 #else /* !CONFIG_RT_GROUP_SCHED */
8555 static int sched_rt_global_constraints(void)
8557 unsigned long flags;
8560 if (sysctl_sched_rt_period <= 0)
8564 * There's always some RT tasks in the root group
8565 * -- migration, kstopmachine etc..
8567 if (sysctl_sched_rt_runtime == 0)
8570 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8571 for_each_possible_cpu(i) {
8572 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8574 raw_spin_lock(&rt_rq->rt_runtime_lock);
8575 rt_rq->rt_runtime = global_rt_runtime();
8576 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8578 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8582 #endif /* CONFIG_RT_GROUP_SCHED */
8584 int sched_rt_handler(struct ctl_table *table, int write,
8585 void __user *buffer, size_t *lenp,
8589 int old_period, old_runtime;
8590 static DEFINE_MUTEX(mutex);
8593 old_period = sysctl_sched_rt_period;
8594 old_runtime = sysctl_sched_rt_runtime;
8596 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8598 if (!ret && write) {
8599 ret = sched_rt_global_constraints();
8601 sysctl_sched_rt_period = old_period;
8602 sysctl_sched_rt_runtime = old_runtime;
8604 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8605 def_rt_bandwidth.rt_period =
8606 ns_to_ktime(global_rt_period());
8609 mutex_unlock(&mutex);
8614 #ifdef CONFIG_CGROUP_SCHED
8616 /* return corresponding task_group object of a cgroup */
8617 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8619 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8620 struct task_group, css);
8623 static struct cgroup_subsys_state *
8624 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8626 struct task_group *tg, *parent;
8628 if (!cgrp->parent) {
8629 /* This is early initialization for the top cgroup */
8630 return &init_task_group.css;
8633 parent = cgroup_tg(cgrp->parent);
8634 tg = sched_create_group(parent);
8636 return ERR_PTR(-ENOMEM);
8642 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8644 struct task_group *tg = cgroup_tg(cgrp);
8646 sched_destroy_group(tg);
8650 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8652 #ifdef CONFIG_RT_GROUP_SCHED
8653 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8656 /* We don't support RT-tasks being in separate groups */
8657 if (tsk->sched_class != &fair_sched_class)
8664 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8665 struct task_struct *tsk, bool threadgroup)
8667 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8671 struct task_struct *c;
8673 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8674 retval = cpu_cgroup_can_attach_task(cgrp, c);
8686 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8687 struct cgroup *old_cont, struct task_struct *tsk,
8690 sched_move_task(tsk);
8692 struct task_struct *c;
8694 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8701 #ifdef CONFIG_FAIR_GROUP_SCHED
8702 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8705 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8708 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8710 struct task_group *tg = cgroup_tg(cgrp);
8712 return (u64) tg->shares;
8714 #endif /* CONFIG_FAIR_GROUP_SCHED */
8716 #ifdef CONFIG_RT_GROUP_SCHED
8717 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8720 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8723 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8725 return sched_group_rt_runtime(cgroup_tg(cgrp));
8728 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8731 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8734 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8736 return sched_group_rt_period(cgroup_tg(cgrp));
8738 #endif /* CONFIG_RT_GROUP_SCHED */
8740 static struct cftype cpu_files[] = {
8741 #ifdef CONFIG_FAIR_GROUP_SCHED
8744 .read_u64 = cpu_shares_read_u64,
8745 .write_u64 = cpu_shares_write_u64,
8748 #ifdef CONFIG_RT_GROUP_SCHED
8750 .name = "rt_runtime_us",
8751 .read_s64 = cpu_rt_runtime_read,
8752 .write_s64 = cpu_rt_runtime_write,
8755 .name = "rt_period_us",
8756 .read_u64 = cpu_rt_period_read_uint,
8757 .write_u64 = cpu_rt_period_write_uint,
8762 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8764 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8767 struct cgroup_subsys cpu_cgroup_subsys = {
8769 .create = cpu_cgroup_create,
8770 .destroy = cpu_cgroup_destroy,
8771 .can_attach = cpu_cgroup_can_attach,
8772 .attach = cpu_cgroup_attach,
8773 .populate = cpu_cgroup_populate,
8774 .subsys_id = cpu_cgroup_subsys_id,
8778 #endif /* CONFIG_CGROUP_SCHED */
8780 #ifdef CONFIG_CGROUP_CPUACCT
8783 * CPU accounting code for task groups.
8785 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8786 * (balbir@in.ibm.com).
8789 /* track cpu usage of a group of tasks and its child groups */
8791 struct cgroup_subsys_state css;
8792 /* cpuusage holds pointer to a u64-type object on every cpu */
8793 u64 __percpu *cpuusage;
8794 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8795 struct cpuacct *parent;
8798 struct cgroup_subsys cpuacct_subsys;
8800 /* return cpu accounting group corresponding to this container */
8801 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8803 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8804 struct cpuacct, css);
8807 /* return cpu accounting group to which this task belongs */
8808 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8810 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8811 struct cpuacct, css);
8814 /* create a new cpu accounting group */
8815 static struct cgroup_subsys_state *cpuacct_create(
8816 struct cgroup_subsys *ss, struct cgroup *cgrp)
8818 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8824 ca->cpuusage = alloc_percpu(u64);
8828 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8829 if (percpu_counter_init(&ca->cpustat[i], 0))
8830 goto out_free_counters;
8833 ca->parent = cgroup_ca(cgrp->parent);
8839 percpu_counter_destroy(&ca->cpustat[i]);
8840 free_percpu(ca->cpuusage);
8844 return ERR_PTR(-ENOMEM);
8847 /* destroy an existing cpu accounting group */
8849 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8851 struct cpuacct *ca = cgroup_ca(cgrp);
8854 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8855 percpu_counter_destroy(&ca->cpustat[i]);
8856 free_percpu(ca->cpuusage);
8860 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8862 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8865 #ifndef CONFIG_64BIT
8867 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8869 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8871 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8879 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8881 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8883 #ifndef CONFIG_64BIT
8885 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8887 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8889 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8895 /* return total cpu usage (in nanoseconds) of a group */
8896 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8898 struct cpuacct *ca = cgroup_ca(cgrp);
8899 u64 totalcpuusage = 0;
8902 for_each_present_cpu(i)
8903 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8905 return totalcpuusage;
8908 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8911 struct cpuacct *ca = cgroup_ca(cgrp);
8920 for_each_present_cpu(i)
8921 cpuacct_cpuusage_write(ca, i, 0);
8927 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8930 struct cpuacct *ca = cgroup_ca(cgroup);
8934 for_each_present_cpu(i) {
8935 percpu = cpuacct_cpuusage_read(ca, i);
8936 seq_printf(m, "%llu ", (unsigned long long) percpu);
8938 seq_printf(m, "\n");
8942 static const char *cpuacct_stat_desc[] = {
8943 [CPUACCT_STAT_USER] = "user",
8944 [CPUACCT_STAT_SYSTEM] = "system",
8947 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8948 struct cgroup_map_cb *cb)
8950 struct cpuacct *ca = cgroup_ca(cgrp);
8953 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8954 s64 val = percpu_counter_read(&ca->cpustat[i]);
8955 val = cputime64_to_clock_t(val);
8956 cb->fill(cb, cpuacct_stat_desc[i], val);
8961 static struct cftype files[] = {
8964 .read_u64 = cpuusage_read,
8965 .write_u64 = cpuusage_write,
8968 .name = "usage_percpu",
8969 .read_seq_string = cpuacct_percpu_seq_read,
8973 .read_map = cpuacct_stats_show,
8977 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
8979 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
8983 * charge this task's execution time to its accounting group.
8985 * called with rq->lock held.
8987 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
8992 if (unlikely(!cpuacct_subsys.active))
8995 cpu = task_cpu(tsk);
9001 for (; ca; ca = ca->parent) {
9002 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9003 *cpuusage += cputime;
9010 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9011 * in cputime_t units. As a result, cpuacct_update_stats calls
9012 * percpu_counter_add with values large enough to always overflow the
9013 * per cpu batch limit causing bad SMP scalability.
9015 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9016 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9017 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9020 #define CPUACCT_BATCH \
9021 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9023 #define CPUACCT_BATCH 0
9027 * Charge the system/user time to the task's accounting group.
9029 static void cpuacct_update_stats(struct task_struct *tsk,
9030 enum cpuacct_stat_index idx, cputime_t val)
9033 int batch = CPUACCT_BATCH;
9035 if (unlikely(!cpuacct_subsys.active))
9042 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9048 struct cgroup_subsys cpuacct_subsys = {
9050 .create = cpuacct_create,
9051 .destroy = cpuacct_destroy,
9052 .populate = cpuacct_populate,
9053 .subsys_id = cpuacct_subsys_id,
9055 #endif /* CONFIG_CGROUP_CPUACCT */
9059 void synchronize_sched_expedited(void)
9063 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9065 #else /* #ifndef CONFIG_SMP */
9067 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9069 static int synchronize_sched_expedited_cpu_stop(void *data)
9072 * There must be a full memory barrier on each affected CPU
9073 * between the time that try_stop_cpus() is called and the
9074 * time that it returns.
9076 * In the current initial implementation of cpu_stop, the
9077 * above condition is already met when the control reaches
9078 * this point and the following smp_mb() is not strictly
9079 * necessary. Do smp_mb() anyway for documentation and
9080 * robustness against future implementation changes.
9082 smp_mb(); /* See above comment block. */
9087 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9088 * approach to force grace period to end quickly. This consumes
9089 * significant time on all CPUs, and is thus not recommended for
9090 * any sort of common-case code.
9092 * Note that it is illegal to call this function while holding any
9093 * lock that is acquired by a CPU-hotplug notifier. Failing to
9094 * observe this restriction will result in deadlock.
9096 void synchronize_sched_expedited(void)
9098 int snap, trycount = 0;
9100 smp_mb(); /* ensure prior mod happens before capturing snap. */
9101 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9103 while (try_stop_cpus(cpu_online_mask,
9104 synchronize_sched_expedited_cpu_stop,
9107 if (trycount++ < 10)
9108 udelay(trycount * num_online_cpus());
9110 synchronize_sched();
9113 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9114 smp_mb(); /* ensure test happens before caller kfree */
9119 atomic_inc(&synchronize_sched_expedited_count);
9120 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9123 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9125 #endif /* #else #ifndef CONFIG_SMP */