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>
78 #include <asm/mutex.h>
80 #include "sched_cpupri.h"
81 #include "workqueue_sched.h"
82 #include "sched_autogroup.h"
84 #define CREATE_TRACE_POINTS
85 #include <trace/events/sched.h>
88 * Convert user-nice values [ -20 ... 0 ... 19 ]
89 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
92 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
93 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
94 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
97 * 'User priority' is the nice value converted to something we
98 * can work with better when scaling various scheduler parameters,
99 * it's a [ 0 ... 39 ] range.
101 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
102 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
103 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
106 * Helpers for converting nanosecond timing to jiffy resolution
108 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
110 #define NICE_0_LOAD SCHED_LOAD_SCALE
111 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
114 * These are the 'tuning knobs' of the scheduler:
116 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
117 * Timeslices get refilled after they expire.
119 #define DEF_TIMESLICE (100 * HZ / 1000)
122 * single value that denotes runtime == period, ie unlimited time.
124 #define RUNTIME_INF ((u64)~0ULL)
126 static inline int rt_policy(int policy)
128 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
133 static inline int task_has_rt_policy(struct task_struct *p)
135 return rt_policy(p->policy);
139 * This is the priority-queue data structure of the RT scheduling class:
141 struct rt_prio_array {
142 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
143 struct list_head queue[MAX_RT_PRIO];
146 struct rt_bandwidth {
147 /* nests inside the rq lock: */
148 raw_spinlock_t rt_runtime_lock;
151 struct hrtimer rt_period_timer;
154 static struct rt_bandwidth def_rt_bandwidth;
156 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
158 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
160 struct rt_bandwidth *rt_b =
161 container_of(timer, struct rt_bandwidth, rt_period_timer);
167 now = hrtimer_cb_get_time(timer);
168 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
173 idle = do_sched_rt_period_timer(rt_b, overrun);
176 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
180 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
182 rt_b->rt_period = ns_to_ktime(period);
183 rt_b->rt_runtime = runtime;
185 raw_spin_lock_init(&rt_b->rt_runtime_lock);
187 hrtimer_init(&rt_b->rt_period_timer,
188 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
189 rt_b->rt_period_timer.function = sched_rt_period_timer;
192 static inline int rt_bandwidth_enabled(void)
194 return sysctl_sched_rt_runtime >= 0;
197 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
201 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
204 if (hrtimer_active(&rt_b->rt_period_timer))
207 raw_spin_lock(&rt_b->rt_runtime_lock);
212 if (hrtimer_active(&rt_b->rt_period_timer))
215 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
216 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
218 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
219 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
220 delta = ktime_to_ns(ktime_sub(hard, soft));
221 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
222 HRTIMER_MODE_ABS_PINNED, 0);
224 raw_spin_unlock(&rt_b->rt_runtime_lock);
227 #ifdef CONFIG_RT_GROUP_SCHED
228 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
230 hrtimer_cancel(&rt_b->rt_period_timer);
235 * sched_domains_mutex serializes calls to arch_init_sched_domains,
236 * detach_destroy_domains and partition_sched_domains.
238 static DEFINE_MUTEX(sched_domains_mutex);
240 #ifdef CONFIG_CGROUP_SCHED
242 #include <linux/cgroup.h>
246 static LIST_HEAD(task_groups);
248 /* task group related information */
250 struct cgroup_subsys_state css;
252 #ifdef CONFIG_FAIR_GROUP_SCHED
253 /* schedulable entities of this group on each cpu */
254 struct sched_entity **se;
255 /* runqueue "owned" by this group on each cpu */
256 struct cfs_rq **cfs_rq;
257 unsigned long shares;
259 atomic_t load_weight;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
276 #ifdef CONFIG_SCHED_AUTOGROUP
277 struct autogroup *autogroup;
281 /* task_group_lock serializes the addition/removal of task groups */
282 static DEFINE_SPINLOCK(task_group_lock);
284 #ifdef CONFIG_FAIR_GROUP_SCHED
286 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
289 * A weight of 0 or 1 can cause arithmetics problems.
290 * A weight of a cfs_rq is the sum of weights of which entities
291 * are queued on this cfs_rq, so a weight of a entity should not be
292 * too large, so as the shares value of a task group.
293 * (The default weight is 1024 - so there's no practical
294 * limitation from this.)
297 #define MAX_SHARES (1UL << 18)
299 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
302 /* Default task group.
303 * Every task in system belong to this group at bootup.
305 struct task_group root_task_group;
307 #endif /* CONFIG_CGROUP_SCHED */
309 /* CFS-related fields in a runqueue */
311 struct load_weight load;
312 unsigned long nr_running;
317 struct rb_root tasks_timeline;
318 struct rb_node *rb_leftmost;
320 struct list_head tasks;
321 struct list_head *balance_iterator;
324 * 'curr' points to currently running entity on this cfs_rq.
325 * It is set to NULL otherwise (i.e when none are currently running).
327 struct sched_entity *curr, *next, *last, *skip;
329 unsigned int nr_spread_over;
331 #ifdef CONFIG_FAIR_GROUP_SCHED
332 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
335 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
336 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
337 * (like users, containers etc.)
339 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
340 * list is used during load balance.
343 struct list_head leaf_cfs_rq_list;
344 struct task_group *tg; /* group that "owns" this runqueue */
348 * the part of load.weight contributed by tasks
350 unsigned long task_weight;
353 * h_load = weight * f(tg)
355 * Where f(tg) is the recursive weight fraction assigned to
358 unsigned long h_load;
361 * Maintaining per-cpu shares distribution for group scheduling
363 * load_stamp is the last time we updated the load average
364 * load_last is the last time we updated the load average and saw load
365 * load_unacc_exec_time is currently unaccounted execution time
369 u64 load_stamp, load_last, load_unacc_exec_time;
371 unsigned long load_contribution;
376 /* Real-Time classes' related field in a runqueue: */
378 struct rt_prio_array active;
379 unsigned long rt_nr_running;
380 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
382 int curr; /* highest queued rt task prio */
384 int next; /* next highest */
389 unsigned long rt_nr_migratory;
390 unsigned long rt_nr_total;
392 struct plist_head pushable_tasks;
397 /* Nests inside the rq lock: */
398 raw_spinlock_t rt_runtime_lock;
400 #ifdef CONFIG_RT_GROUP_SCHED
401 unsigned long rt_nr_boosted;
404 struct list_head leaf_rt_rq_list;
405 struct task_group *tg;
412 * We add the notion of a root-domain which will be used to define per-domain
413 * variables. Each exclusive cpuset essentially defines an island domain by
414 * fully partitioning the member cpus from any other cpuset. Whenever a new
415 * exclusive cpuset is created, we also create and attach a new root-domain
422 cpumask_var_t online;
425 * The "RT overload" flag: it gets set if a CPU has more than
426 * one runnable RT task.
428 cpumask_var_t rto_mask;
430 struct cpupri cpupri;
434 * By default the system creates a single root-domain with all cpus as
435 * members (mimicking the global state we have today).
437 static struct root_domain def_root_domain;
439 #endif /* CONFIG_SMP */
442 * This is the main, per-CPU runqueue data structure.
444 * Locking rule: those places that want to lock multiple runqueues
445 * (such as the load balancing or the thread migration code), lock
446 * acquire operations must be ordered by ascending &runqueue.
453 * nr_running and cpu_load should be in the same cacheline because
454 * remote CPUs use both these fields when doing load calculation.
456 unsigned long nr_running;
457 #define CPU_LOAD_IDX_MAX 5
458 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
459 unsigned long last_load_update_tick;
462 unsigned char nohz_balance_kick;
464 unsigned int skip_clock_update;
466 /* capture load from *all* tasks on this cpu: */
467 struct load_weight load;
468 unsigned long nr_load_updates;
474 #ifdef CONFIG_FAIR_GROUP_SCHED
475 /* list of leaf cfs_rq on this cpu: */
476 struct list_head leaf_cfs_rq_list;
478 #ifdef CONFIG_RT_GROUP_SCHED
479 struct list_head leaf_rt_rq_list;
483 * This is part of a global counter where only the total sum
484 * over all CPUs matters. A task can increase this counter on
485 * one CPU and if it got migrated afterwards it may decrease
486 * it on another CPU. Always updated under the runqueue lock:
488 unsigned long nr_uninterruptible;
490 struct task_struct *curr, *idle, *stop;
491 unsigned long next_balance;
492 struct mm_struct *prev_mm;
500 struct root_domain *rd;
501 struct sched_domain *sd;
503 unsigned long cpu_power;
505 unsigned char idle_at_tick;
506 /* For active balancing */
510 struct cpu_stop_work active_balance_work;
511 /* cpu of this runqueue: */
515 unsigned long avg_load_per_task;
523 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
527 /* calc_load related fields */
528 unsigned long calc_load_update;
529 long calc_load_active;
531 #ifdef CONFIG_SCHED_HRTICK
533 int hrtick_csd_pending;
534 struct call_single_data hrtick_csd;
536 struct hrtimer hrtick_timer;
539 #ifdef CONFIG_SCHEDSTATS
541 struct sched_info rq_sched_info;
542 unsigned long long rq_cpu_time;
543 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
545 /* sys_sched_yield() stats */
546 unsigned int yld_count;
548 /* schedule() stats */
549 unsigned int sched_switch;
550 unsigned int sched_count;
551 unsigned int sched_goidle;
553 /* try_to_wake_up() stats */
554 unsigned int ttwu_count;
555 unsigned int ttwu_local;
559 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
562 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
564 static inline int cpu_of(struct rq *rq)
573 #define rcu_dereference_check_sched_domain(p) \
574 rcu_dereference_check((p), \
575 rcu_read_lock_sched_held() || \
576 lockdep_is_held(&sched_domains_mutex))
579 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
580 * See detach_destroy_domains: synchronize_sched for details.
582 * The domain tree of any CPU may only be accessed from within
583 * preempt-disabled sections.
585 #define for_each_domain(cpu, __sd) \
586 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
588 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
589 #define this_rq() (&__get_cpu_var(runqueues))
590 #define task_rq(p) cpu_rq(task_cpu(p))
591 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
592 #define raw_rq() (&__raw_get_cpu_var(runqueues))
594 #ifdef CONFIG_CGROUP_SCHED
597 * Return the group to which this tasks belongs.
599 * We use task_subsys_state_check() and extend the RCU verification
600 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
601 * holds that lock for each task it moves into the cgroup. Therefore
602 * by holding that lock, we pin the task to the current cgroup.
604 static inline struct task_group *task_group(struct task_struct *p)
606 struct task_group *tg;
607 struct cgroup_subsys_state *css;
609 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
610 lockdep_is_held(&task_rq(p)->lock));
611 tg = container_of(css, struct task_group, css);
613 return autogroup_task_group(p, tg);
616 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
617 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
619 #ifdef CONFIG_FAIR_GROUP_SCHED
620 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
621 p->se.parent = task_group(p)->se[cpu];
624 #ifdef CONFIG_RT_GROUP_SCHED
625 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
626 p->rt.parent = task_group(p)->rt_se[cpu];
630 #else /* CONFIG_CGROUP_SCHED */
632 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
633 static inline struct task_group *task_group(struct task_struct *p)
638 #endif /* CONFIG_CGROUP_SCHED */
640 static void update_rq_clock_task(struct rq *rq, s64 delta);
642 static void update_rq_clock(struct rq *rq)
646 if (rq->skip_clock_update)
649 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
651 update_rq_clock_task(rq, delta);
655 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
657 #ifdef CONFIG_SCHED_DEBUG
658 # define const_debug __read_mostly
660 # define const_debug static const
665 * @cpu: the processor in question.
667 * Returns true if the current cpu runqueue is locked.
668 * This interface allows printk to be called with the runqueue lock
669 * held and know whether or not it is OK to wake up the klogd.
671 int runqueue_is_locked(int cpu)
673 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
677 * Debugging: various feature bits
680 #define SCHED_FEAT(name, enabled) \
681 __SCHED_FEAT_##name ,
684 #include "sched_features.h"
689 #define SCHED_FEAT(name, enabled) \
690 (1UL << __SCHED_FEAT_##name) * enabled |
692 const_debug unsigned int sysctl_sched_features =
693 #include "sched_features.h"
698 #ifdef CONFIG_SCHED_DEBUG
699 #define SCHED_FEAT(name, enabled) \
702 static __read_mostly char *sched_feat_names[] = {
703 #include "sched_features.h"
709 static int sched_feat_show(struct seq_file *m, void *v)
713 for (i = 0; sched_feat_names[i]; i++) {
714 if (!(sysctl_sched_features & (1UL << i)))
716 seq_printf(m, "%s ", sched_feat_names[i]);
724 sched_feat_write(struct file *filp, const char __user *ubuf,
725 size_t cnt, loff_t *ppos)
735 if (copy_from_user(&buf, ubuf, cnt))
741 if (strncmp(cmp, "NO_", 3) == 0) {
746 for (i = 0; sched_feat_names[i]; i++) {
747 if (strcmp(cmp, sched_feat_names[i]) == 0) {
749 sysctl_sched_features &= ~(1UL << i);
751 sysctl_sched_features |= (1UL << i);
756 if (!sched_feat_names[i])
764 static int sched_feat_open(struct inode *inode, struct file *filp)
766 return single_open(filp, sched_feat_show, NULL);
769 static const struct file_operations sched_feat_fops = {
770 .open = sched_feat_open,
771 .write = sched_feat_write,
774 .release = single_release,
777 static __init int sched_init_debug(void)
779 debugfs_create_file("sched_features", 0644, NULL, NULL,
784 late_initcall(sched_init_debug);
788 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
791 * Number of tasks to iterate in a single balance run.
792 * Limited because this is done with IRQs disabled.
794 const_debug unsigned int sysctl_sched_nr_migrate = 32;
797 * period over which we average the RT time consumption, measured
802 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
805 * period over which we measure -rt task cpu usage in us.
808 unsigned int sysctl_sched_rt_period = 1000000;
810 static __read_mostly int scheduler_running;
813 * part of the period that we allow rt tasks to run in us.
816 int sysctl_sched_rt_runtime = 950000;
818 static inline u64 global_rt_period(void)
820 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
823 static inline u64 global_rt_runtime(void)
825 if (sysctl_sched_rt_runtime < 0)
828 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
831 #ifndef prepare_arch_switch
832 # define prepare_arch_switch(next) do { } while (0)
834 #ifndef finish_arch_switch
835 # define finish_arch_switch(prev) do { } while (0)
838 static inline int task_current(struct rq *rq, struct task_struct *p)
840 return rq->curr == p;
843 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
844 static inline int task_running(struct rq *rq, struct task_struct *p)
846 return task_current(rq, p);
849 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
853 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
855 #ifdef CONFIG_DEBUG_SPINLOCK
856 /* this is a valid case when another task releases the spinlock */
857 rq->lock.owner = current;
860 * If we are tracking spinlock dependencies then we have to
861 * fix up the runqueue lock - which gets 'carried over' from
864 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
866 raw_spin_unlock_irq(&rq->lock);
869 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
870 static inline int task_running(struct rq *rq, struct task_struct *p)
875 return task_current(rq, p);
879 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
883 * We can optimise this out completely for !SMP, because the
884 * SMP rebalancing from interrupt is the only thing that cares
889 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
890 raw_spin_unlock_irq(&rq->lock);
892 raw_spin_unlock(&rq->lock);
896 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
900 * After ->oncpu is cleared, the task can be moved to a different CPU.
901 * We must ensure this doesn't happen until the switch is completely
907 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
911 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
914 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
917 static inline int task_is_waking(struct task_struct *p)
919 return unlikely(p->state == TASK_WAKING);
923 * __task_rq_lock - lock the runqueue a given task resides on.
924 * Must be called interrupts disabled.
926 static inline struct rq *__task_rq_lock(struct task_struct *p)
933 raw_spin_lock(&rq->lock);
934 if (likely(rq == task_rq(p)))
936 raw_spin_unlock(&rq->lock);
941 * task_rq_lock - lock the runqueue a given task resides on and disable
942 * interrupts. Note the ordering: we can safely lookup the task_rq without
943 * explicitly disabling preemption.
945 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
951 local_irq_save(*flags);
953 raw_spin_lock(&rq->lock);
954 if (likely(rq == task_rq(p)))
956 raw_spin_unlock_irqrestore(&rq->lock, *flags);
960 static void __task_rq_unlock(struct rq *rq)
963 raw_spin_unlock(&rq->lock);
966 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
969 raw_spin_unlock_irqrestore(&rq->lock, *flags);
973 * this_rq_lock - lock this runqueue and disable interrupts.
975 static struct rq *this_rq_lock(void)
982 raw_spin_lock(&rq->lock);
987 #ifdef CONFIG_SCHED_HRTICK
989 * Use HR-timers to deliver accurate preemption points.
991 * Its all a bit involved since we cannot program an hrt while holding the
992 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
995 * When we get rescheduled we reprogram the hrtick_timer outside of the
1001 * - enabled by features
1002 * - hrtimer is actually high res
1004 static inline int hrtick_enabled(struct rq *rq)
1006 if (!sched_feat(HRTICK))
1008 if (!cpu_active(cpu_of(rq)))
1010 return hrtimer_is_hres_active(&rq->hrtick_timer);
1013 static void hrtick_clear(struct rq *rq)
1015 if (hrtimer_active(&rq->hrtick_timer))
1016 hrtimer_cancel(&rq->hrtick_timer);
1020 * High-resolution timer tick.
1021 * Runs from hardirq context with interrupts disabled.
1023 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1025 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1027 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1029 raw_spin_lock(&rq->lock);
1030 update_rq_clock(rq);
1031 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1032 raw_spin_unlock(&rq->lock);
1034 return HRTIMER_NORESTART;
1039 * called from hardirq (IPI) context
1041 static void __hrtick_start(void *arg)
1043 struct rq *rq = arg;
1045 raw_spin_lock(&rq->lock);
1046 hrtimer_restart(&rq->hrtick_timer);
1047 rq->hrtick_csd_pending = 0;
1048 raw_spin_unlock(&rq->lock);
1052 * Called to set the hrtick timer state.
1054 * called with rq->lock held and irqs disabled
1056 static void hrtick_start(struct rq *rq, u64 delay)
1058 struct hrtimer *timer = &rq->hrtick_timer;
1059 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1061 hrtimer_set_expires(timer, time);
1063 if (rq == this_rq()) {
1064 hrtimer_restart(timer);
1065 } else if (!rq->hrtick_csd_pending) {
1066 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1067 rq->hrtick_csd_pending = 1;
1072 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1074 int cpu = (int)(long)hcpu;
1077 case CPU_UP_CANCELED:
1078 case CPU_UP_CANCELED_FROZEN:
1079 case CPU_DOWN_PREPARE:
1080 case CPU_DOWN_PREPARE_FROZEN:
1082 case CPU_DEAD_FROZEN:
1083 hrtick_clear(cpu_rq(cpu));
1090 static __init void init_hrtick(void)
1092 hotcpu_notifier(hotplug_hrtick, 0);
1096 * Called to set the hrtick timer state.
1098 * called with rq->lock held and irqs disabled
1100 static void hrtick_start(struct rq *rq, u64 delay)
1102 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1103 HRTIMER_MODE_REL_PINNED, 0);
1106 static inline void init_hrtick(void)
1109 #endif /* CONFIG_SMP */
1111 static void init_rq_hrtick(struct rq *rq)
1114 rq->hrtick_csd_pending = 0;
1116 rq->hrtick_csd.flags = 0;
1117 rq->hrtick_csd.func = __hrtick_start;
1118 rq->hrtick_csd.info = rq;
1121 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1122 rq->hrtick_timer.function = hrtick;
1124 #else /* CONFIG_SCHED_HRTICK */
1125 static inline void hrtick_clear(struct rq *rq)
1129 static inline void init_rq_hrtick(struct rq *rq)
1133 static inline void init_hrtick(void)
1136 #endif /* CONFIG_SCHED_HRTICK */
1139 * resched_task - mark a task 'to be rescheduled now'.
1141 * On UP this means the setting of the need_resched flag, on SMP it
1142 * might also involve a cross-CPU call to trigger the scheduler on
1147 #ifndef tsk_is_polling
1148 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1151 static void resched_task(struct task_struct *p)
1155 assert_raw_spin_locked(&task_rq(p)->lock);
1157 if (test_tsk_need_resched(p))
1160 set_tsk_need_resched(p);
1163 if (cpu == smp_processor_id())
1166 /* NEED_RESCHED must be visible before we test polling */
1168 if (!tsk_is_polling(p))
1169 smp_send_reschedule(cpu);
1172 static void resched_cpu(int cpu)
1174 struct rq *rq = cpu_rq(cpu);
1175 unsigned long flags;
1177 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1179 resched_task(cpu_curr(cpu));
1180 raw_spin_unlock_irqrestore(&rq->lock, flags);
1185 * In the semi idle case, use the nearest busy cpu for migrating timers
1186 * from an idle cpu. This is good for power-savings.
1188 * We don't do similar optimization for completely idle system, as
1189 * selecting an idle cpu will add more delays to the timers than intended
1190 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1192 int get_nohz_timer_target(void)
1194 int cpu = smp_processor_id();
1196 struct sched_domain *sd;
1198 for_each_domain(cpu, sd) {
1199 for_each_cpu(i, sched_domain_span(sd))
1206 * When add_timer_on() enqueues a timer into the timer wheel of an
1207 * idle CPU then this timer might expire before the next timer event
1208 * which is scheduled to wake up that CPU. In case of a completely
1209 * idle system the next event might even be infinite time into the
1210 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1211 * leaves the inner idle loop so the newly added timer is taken into
1212 * account when the CPU goes back to idle and evaluates the timer
1213 * wheel for the next timer event.
1215 void wake_up_idle_cpu(int cpu)
1217 struct rq *rq = cpu_rq(cpu);
1219 if (cpu == smp_processor_id())
1223 * This is safe, as this function is called with the timer
1224 * wheel base lock of (cpu) held. When the CPU is on the way
1225 * to idle and has not yet set rq->curr to idle then it will
1226 * be serialized on the timer wheel base lock and take the new
1227 * timer into account automatically.
1229 if (rq->curr != rq->idle)
1233 * We can set TIF_RESCHED on the idle task of the other CPU
1234 * lockless. The worst case is that the other CPU runs the
1235 * idle task through an additional NOOP schedule()
1237 set_tsk_need_resched(rq->idle);
1239 /* NEED_RESCHED must be visible before we test polling */
1241 if (!tsk_is_polling(rq->idle))
1242 smp_send_reschedule(cpu);
1245 #endif /* CONFIG_NO_HZ */
1247 static u64 sched_avg_period(void)
1249 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1252 static void sched_avg_update(struct rq *rq)
1254 s64 period = sched_avg_period();
1256 while ((s64)(rq->clock - rq->age_stamp) > period) {
1258 * Inline assembly required to prevent the compiler
1259 * optimising this loop into a divmod call.
1260 * See __iter_div_u64_rem() for another example of this.
1262 asm("" : "+rm" (rq->age_stamp));
1263 rq->age_stamp += period;
1268 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1270 rq->rt_avg += rt_delta;
1271 sched_avg_update(rq);
1274 #else /* !CONFIG_SMP */
1275 static void resched_task(struct task_struct *p)
1277 assert_raw_spin_locked(&task_rq(p)->lock);
1278 set_tsk_need_resched(p);
1281 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1285 static void sched_avg_update(struct rq *rq)
1288 #endif /* CONFIG_SMP */
1290 #if BITS_PER_LONG == 32
1291 # define WMULT_CONST (~0UL)
1293 # define WMULT_CONST (1UL << 32)
1296 #define WMULT_SHIFT 32
1299 * Shift right and round:
1301 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1304 * delta *= weight / lw
1306 static unsigned long
1307 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1308 struct load_weight *lw)
1312 if (!lw->inv_weight) {
1313 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1316 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1320 tmp = (u64)delta_exec * weight;
1322 * Check whether we'd overflow the 64-bit multiplication:
1324 if (unlikely(tmp > WMULT_CONST))
1325 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1328 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1330 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1333 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1339 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1345 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1352 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1353 * of tasks with abnormal "nice" values across CPUs the contribution that
1354 * each task makes to its run queue's load is weighted according to its
1355 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1356 * scaled version of the new time slice allocation that they receive on time
1360 #define WEIGHT_IDLEPRIO 3
1361 #define WMULT_IDLEPRIO 1431655765
1364 * Nice levels are multiplicative, with a gentle 10% change for every
1365 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1366 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1367 * that remained on nice 0.
1369 * The "10% effect" is relative and cumulative: from _any_ nice level,
1370 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1371 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1372 * If a task goes up by ~10% and another task goes down by ~10% then
1373 * the relative distance between them is ~25%.)
1375 static const int prio_to_weight[40] = {
1376 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1377 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1378 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1379 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1380 /* 0 */ 1024, 820, 655, 526, 423,
1381 /* 5 */ 335, 272, 215, 172, 137,
1382 /* 10 */ 110, 87, 70, 56, 45,
1383 /* 15 */ 36, 29, 23, 18, 15,
1387 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1389 * In cases where the weight does not change often, we can use the
1390 * precalculated inverse to speed up arithmetics by turning divisions
1391 * into multiplications:
1393 static const u32 prio_to_wmult[40] = {
1394 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1395 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1396 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1397 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1398 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1399 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1400 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1401 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1404 /* Time spent by the tasks of the cpu accounting group executing in ... */
1405 enum cpuacct_stat_index {
1406 CPUACCT_STAT_USER, /* ... user mode */
1407 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1409 CPUACCT_STAT_NSTATS,
1412 #ifdef CONFIG_CGROUP_CPUACCT
1413 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1414 static void cpuacct_update_stats(struct task_struct *tsk,
1415 enum cpuacct_stat_index idx, cputime_t val);
1417 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1418 static inline void cpuacct_update_stats(struct task_struct *tsk,
1419 enum cpuacct_stat_index idx, cputime_t val) {}
1422 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1424 update_load_add(&rq->load, load);
1427 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1429 update_load_sub(&rq->load, load);
1432 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1433 typedef int (*tg_visitor)(struct task_group *, void *);
1436 * Iterate the full tree, calling @down when first entering a node and @up when
1437 * leaving it for the final time.
1439 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1441 struct task_group *parent, *child;
1445 parent = &root_task_group;
1447 ret = (*down)(parent, data);
1450 list_for_each_entry_rcu(child, &parent->children, siblings) {
1457 ret = (*up)(parent, data);
1462 parent = parent->parent;
1471 static int tg_nop(struct task_group *tg, void *data)
1478 /* Used instead of source_load when we know the type == 0 */
1479 static unsigned long weighted_cpuload(const int cpu)
1481 return cpu_rq(cpu)->load.weight;
1485 * Return a low guess at the load of a migration-source cpu weighted
1486 * according to the scheduling class and "nice" value.
1488 * We want to under-estimate the load of migration sources, to
1489 * balance conservatively.
1491 static unsigned long source_load(int cpu, int type)
1493 struct rq *rq = cpu_rq(cpu);
1494 unsigned long total = weighted_cpuload(cpu);
1496 if (type == 0 || !sched_feat(LB_BIAS))
1499 return min(rq->cpu_load[type-1], total);
1503 * Return a high guess at the load of a migration-target cpu weighted
1504 * according to the scheduling class and "nice" value.
1506 static unsigned long target_load(int cpu, int type)
1508 struct rq *rq = cpu_rq(cpu);
1509 unsigned long total = weighted_cpuload(cpu);
1511 if (type == 0 || !sched_feat(LB_BIAS))
1514 return max(rq->cpu_load[type-1], total);
1517 static unsigned long power_of(int cpu)
1519 return cpu_rq(cpu)->cpu_power;
1522 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1524 static unsigned long cpu_avg_load_per_task(int cpu)
1526 struct rq *rq = cpu_rq(cpu);
1527 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1530 rq->avg_load_per_task = rq->load.weight / nr_running;
1532 rq->avg_load_per_task = 0;
1534 return rq->avg_load_per_task;
1537 #ifdef CONFIG_FAIR_GROUP_SCHED
1540 * Compute the cpu's hierarchical load factor for each task group.
1541 * This needs to be done in a top-down fashion because the load of a child
1542 * group is a fraction of its parents load.
1544 static int tg_load_down(struct task_group *tg, void *data)
1547 long cpu = (long)data;
1550 load = cpu_rq(cpu)->load.weight;
1552 load = tg->parent->cfs_rq[cpu]->h_load;
1553 load *= tg->se[cpu]->load.weight;
1554 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1557 tg->cfs_rq[cpu]->h_load = load;
1562 static void update_h_load(long cpu)
1564 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1569 #ifdef CONFIG_PREEMPT
1571 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1574 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1575 * way at the expense of forcing extra atomic operations in all
1576 * invocations. This assures that the double_lock is acquired using the
1577 * same underlying policy as the spinlock_t on this architecture, which
1578 * reduces latency compared to the unfair variant below. However, it
1579 * also adds more overhead and therefore may reduce throughput.
1581 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1582 __releases(this_rq->lock)
1583 __acquires(busiest->lock)
1584 __acquires(this_rq->lock)
1586 raw_spin_unlock(&this_rq->lock);
1587 double_rq_lock(this_rq, busiest);
1594 * Unfair double_lock_balance: Optimizes throughput at the expense of
1595 * latency by eliminating extra atomic operations when the locks are
1596 * already in proper order on entry. This favors lower cpu-ids and will
1597 * grant the double lock to lower cpus over higher ids under contention,
1598 * regardless of entry order into the function.
1600 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1601 __releases(this_rq->lock)
1602 __acquires(busiest->lock)
1603 __acquires(this_rq->lock)
1607 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1608 if (busiest < this_rq) {
1609 raw_spin_unlock(&this_rq->lock);
1610 raw_spin_lock(&busiest->lock);
1611 raw_spin_lock_nested(&this_rq->lock,
1612 SINGLE_DEPTH_NESTING);
1615 raw_spin_lock_nested(&busiest->lock,
1616 SINGLE_DEPTH_NESTING);
1621 #endif /* CONFIG_PREEMPT */
1624 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1626 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1628 if (unlikely(!irqs_disabled())) {
1629 /* printk() doesn't work good under rq->lock */
1630 raw_spin_unlock(&this_rq->lock);
1634 return _double_lock_balance(this_rq, busiest);
1637 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1638 __releases(busiest->lock)
1640 raw_spin_unlock(&busiest->lock);
1641 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1645 * double_rq_lock - safely lock two runqueues
1647 * Note this does not disable interrupts like task_rq_lock,
1648 * you need to do so manually before calling.
1650 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1651 __acquires(rq1->lock)
1652 __acquires(rq2->lock)
1654 BUG_ON(!irqs_disabled());
1656 raw_spin_lock(&rq1->lock);
1657 __acquire(rq2->lock); /* Fake it out ;) */
1660 raw_spin_lock(&rq1->lock);
1661 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1663 raw_spin_lock(&rq2->lock);
1664 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1670 * double_rq_unlock - safely unlock two runqueues
1672 * Note this does not restore interrupts like task_rq_unlock,
1673 * you need to do so manually after calling.
1675 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1676 __releases(rq1->lock)
1677 __releases(rq2->lock)
1679 raw_spin_unlock(&rq1->lock);
1681 raw_spin_unlock(&rq2->lock);
1683 __release(rq2->lock);
1686 #else /* CONFIG_SMP */
1689 * double_rq_lock - safely lock two runqueues
1691 * Note this does not disable interrupts like task_rq_lock,
1692 * you need to do so manually before calling.
1694 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1695 __acquires(rq1->lock)
1696 __acquires(rq2->lock)
1698 BUG_ON(!irqs_disabled());
1700 raw_spin_lock(&rq1->lock);
1701 __acquire(rq2->lock); /* Fake it out ;) */
1705 * double_rq_unlock - safely unlock two runqueues
1707 * Note this does not restore interrupts like task_rq_unlock,
1708 * you need to do so manually after calling.
1710 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1711 __releases(rq1->lock)
1712 __releases(rq2->lock)
1715 raw_spin_unlock(&rq1->lock);
1716 __release(rq2->lock);
1721 static void calc_load_account_idle(struct rq *this_rq);
1722 static void update_sysctl(void);
1723 static int get_update_sysctl_factor(void);
1724 static void update_cpu_load(struct rq *this_rq);
1726 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1728 set_task_rq(p, cpu);
1731 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1732 * successfuly executed on another CPU. We must ensure that updates of
1733 * per-task data have been completed by this moment.
1736 task_thread_info(p)->cpu = cpu;
1740 static const struct sched_class rt_sched_class;
1742 #define sched_class_highest (&stop_sched_class)
1743 #define for_each_class(class) \
1744 for (class = sched_class_highest; class; class = class->next)
1746 #include "sched_stats.h"
1748 static void inc_nr_running(struct rq *rq)
1753 static void dec_nr_running(struct rq *rq)
1758 static void set_load_weight(struct task_struct *p)
1761 * SCHED_IDLE tasks get minimal weight:
1763 if (p->policy == SCHED_IDLE) {
1764 p->se.load.weight = WEIGHT_IDLEPRIO;
1765 p->se.load.inv_weight = WMULT_IDLEPRIO;
1769 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1770 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1773 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1775 update_rq_clock(rq);
1776 sched_info_queued(p);
1777 p->sched_class->enqueue_task(rq, p, flags);
1781 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1783 update_rq_clock(rq);
1784 sched_info_dequeued(p);
1785 p->sched_class->dequeue_task(rq, p, flags);
1790 * activate_task - move a task to the runqueue.
1792 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1794 if (task_contributes_to_load(p))
1795 rq->nr_uninterruptible--;
1797 enqueue_task(rq, p, flags);
1802 * deactivate_task - remove a task from the runqueue.
1804 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1806 if (task_contributes_to_load(p))
1807 rq->nr_uninterruptible++;
1809 dequeue_task(rq, p, flags);
1813 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1816 * There are no locks covering percpu hardirq/softirq time.
1817 * They are only modified in account_system_vtime, on corresponding CPU
1818 * with interrupts disabled. So, writes are safe.
1819 * They are read and saved off onto struct rq in update_rq_clock().
1820 * This may result in other CPU reading this CPU's irq time and can
1821 * race with irq/account_system_vtime on this CPU. We would either get old
1822 * or new value with a side effect of accounting a slice of irq time to wrong
1823 * task when irq is in progress while we read rq->clock. That is a worthy
1824 * compromise in place of having locks on each irq in account_system_time.
1826 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1827 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1829 static DEFINE_PER_CPU(u64, irq_start_time);
1830 static int sched_clock_irqtime;
1832 void enable_sched_clock_irqtime(void)
1834 sched_clock_irqtime = 1;
1837 void disable_sched_clock_irqtime(void)
1839 sched_clock_irqtime = 0;
1842 #ifndef CONFIG_64BIT
1843 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1845 static inline void irq_time_write_begin(void)
1847 __this_cpu_inc(irq_time_seq.sequence);
1851 static inline void irq_time_write_end(void)
1854 __this_cpu_inc(irq_time_seq.sequence);
1857 static inline u64 irq_time_read(int cpu)
1863 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1864 irq_time = per_cpu(cpu_softirq_time, cpu) +
1865 per_cpu(cpu_hardirq_time, cpu);
1866 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1870 #else /* CONFIG_64BIT */
1871 static inline void irq_time_write_begin(void)
1875 static inline void irq_time_write_end(void)
1879 static inline u64 irq_time_read(int cpu)
1881 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1883 #endif /* CONFIG_64BIT */
1886 * Called before incrementing preempt_count on {soft,}irq_enter
1887 * and before decrementing preempt_count on {soft,}irq_exit.
1889 void account_system_vtime(struct task_struct *curr)
1891 unsigned long flags;
1895 if (!sched_clock_irqtime)
1898 local_irq_save(flags);
1900 cpu = smp_processor_id();
1901 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1902 __this_cpu_add(irq_start_time, delta);
1904 irq_time_write_begin();
1906 * We do not account for softirq time from ksoftirqd here.
1907 * We want to continue accounting softirq time to ksoftirqd thread
1908 * in that case, so as not to confuse scheduler with a special task
1909 * that do not consume any time, but still wants to run.
1911 if (hardirq_count())
1912 __this_cpu_add(cpu_hardirq_time, delta);
1913 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1914 __this_cpu_add(cpu_softirq_time, delta);
1916 irq_time_write_end();
1917 local_irq_restore(flags);
1919 EXPORT_SYMBOL_GPL(account_system_vtime);
1921 static void update_rq_clock_task(struct rq *rq, s64 delta)
1925 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1928 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1929 * this case when a previous update_rq_clock() happened inside a
1930 * {soft,}irq region.
1932 * When this happens, we stop ->clock_task and only update the
1933 * prev_irq_time stamp to account for the part that fit, so that a next
1934 * update will consume the rest. This ensures ->clock_task is
1937 * It does however cause some slight miss-attribution of {soft,}irq
1938 * time, a more accurate solution would be to update the irq_time using
1939 * the current rq->clock timestamp, except that would require using
1942 if (irq_delta > delta)
1945 rq->prev_irq_time += irq_delta;
1947 rq->clock_task += delta;
1949 if (irq_delta && sched_feat(NONIRQ_POWER))
1950 sched_rt_avg_update(rq, irq_delta);
1953 static int irqtime_account_hi_update(void)
1955 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1956 unsigned long flags;
1960 local_irq_save(flags);
1961 latest_ns = this_cpu_read(cpu_hardirq_time);
1962 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
1964 local_irq_restore(flags);
1968 static int irqtime_account_si_update(void)
1970 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1971 unsigned long flags;
1975 local_irq_save(flags);
1976 latest_ns = this_cpu_read(cpu_softirq_time);
1977 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
1979 local_irq_restore(flags);
1983 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
1985 #define sched_clock_irqtime (0)
1987 static void update_rq_clock_task(struct rq *rq, s64 delta)
1989 rq->clock_task += delta;
1992 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1994 #include "sched_idletask.c"
1995 #include "sched_fair.c"
1996 #include "sched_rt.c"
1997 #include "sched_autogroup.c"
1998 #include "sched_stoptask.c"
1999 #ifdef CONFIG_SCHED_DEBUG
2000 # include "sched_debug.c"
2003 void sched_set_stop_task(int cpu, struct task_struct *stop)
2005 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2006 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2010 * Make it appear like a SCHED_FIFO task, its something
2011 * userspace knows about and won't get confused about.
2013 * Also, it will make PI more or less work without too
2014 * much confusion -- but then, stop work should not
2015 * rely on PI working anyway.
2017 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2019 stop->sched_class = &stop_sched_class;
2022 cpu_rq(cpu)->stop = stop;
2026 * Reset it back to a normal scheduling class so that
2027 * it can die in pieces.
2029 old_stop->sched_class = &rt_sched_class;
2034 * __normal_prio - return the priority that is based on the static prio
2036 static inline int __normal_prio(struct task_struct *p)
2038 return p->static_prio;
2042 * Calculate the expected normal priority: i.e. priority
2043 * without taking RT-inheritance into account. Might be
2044 * boosted by interactivity modifiers. Changes upon fork,
2045 * setprio syscalls, and whenever the interactivity
2046 * estimator recalculates.
2048 static inline int normal_prio(struct task_struct *p)
2052 if (task_has_rt_policy(p))
2053 prio = MAX_RT_PRIO-1 - p->rt_priority;
2055 prio = __normal_prio(p);
2060 * Calculate the current priority, i.e. the priority
2061 * taken into account by the scheduler. This value might
2062 * be boosted by RT tasks, or might be boosted by
2063 * interactivity modifiers. Will be RT if the task got
2064 * RT-boosted. If not then it returns p->normal_prio.
2066 static int effective_prio(struct task_struct *p)
2068 p->normal_prio = normal_prio(p);
2070 * If we are RT tasks or we were boosted to RT priority,
2071 * keep the priority unchanged. Otherwise, update priority
2072 * to the normal priority:
2074 if (!rt_prio(p->prio))
2075 return p->normal_prio;
2080 * task_curr - is this task currently executing on a CPU?
2081 * @p: the task in question.
2083 inline int task_curr(const struct task_struct *p)
2085 return cpu_curr(task_cpu(p)) == p;
2088 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2089 const struct sched_class *prev_class,
2092 if (prev_class != p->sched_class) {
2093 if (prev_class->switched_from)
2094 prev_class->switched_from(rq, p);
2095 p->sched_class->switched_to(rq, p);
2096 } else if (oldprio != p->prio)
2097 p->sched_class->prio_changed(rq, p, oldprio);
2100 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2102 const struct sched_class *class;
2104 if (p->sched_class == rq->curr->sched_class) {
2105 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2107 for_each_class(class) {
2108 if (class == rq->curr->sched_class)
2110 if (class == p->sched_class) {
2111 resched_task(rq->curr);
2118 * A queue event has occurred, and we're going to schedule. In
2119 * this case, we can save a useless back to back clock update.
2121 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
2122 rq->skip_clock_update = 1;
2127 * Is this task likely cache-hot:
2130 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2134 if (p->sched_class != &fair_sched_class)
2137 if (unlikely(p->policy == SCHED_IDLE))
2141 * Buddy candidates are cache hot:
2143 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2144 (&p->se == cfs_rq_of(&p->se)->next ||
2145 &p->se == cfs_rq_of(&p->se)->last))
2148 if (sysctl_sched_migration_cost == -1)
2150 if (sysctl_sched_migration_cost == 0)
2153 delta = now - p->se.exec_start;
2155 return delta < (s64)sysctl_sched_migration_cost;
2158 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2160 #ifdef CONFIG_SCHED_DEBUG
2162 * We should never call set_task_cpu() on a blocked task,
2163 * ttwu() will sort out the placement.
2165 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2166 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2169 trace_sched_migrate_task(p, new_cpu);
2171 if (task_cpu(p) != new_cpu) {
2172 p->se.nr_migrations++;
2173 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2176 __set_task_cpu(p, new_cpu);
2179 struct migration_arg {
2180 struct task_struct *task;
2184 static int migration_cpu_stop(void *data);
2187 * The task's runqueue lock must be held.
2188 * Returns true if you have to wait for migration thread.
2190 static bool migrate_task(struct task_struct *p, struct rq *rq)
2193 * If the task is not on a runqueue (and not running), then
2194 * the next wake-up will properly place the task.
2196 return p->se.on_rq || task_running(rq, p);
2200 * wait_task_inactive - wait for a thread to unschedule.
2202 * If @match_state is nonzero, it's the @p->state value just checked and
2203 * not expected to change. If it changes, i.e. @p might have woken up,
2204 * then return zero. When we succeed in waiting for @p to be off its CPU,
2205 * we return a positive number (its total switch count). If a second call
2206 * a short while later returns the same number, the caller can be sure that
2207 * @p has remained unscheduled the whole time.
2209 * The caller must ensure that the task *will* unschedule sometime soon,
2210 * else this function might spin for a *long* time. This function can't
2211 * be called with interrupts off, or it may introduce deadlock with
2212 * smp_call_function() if an IPI is sent by the same process we are
2213 * waiting to become inactive.
2215 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2217 unsigned long flags;
2224 * We do the initial early heuristics without holding
2225 * any task-queue locks at all. We'll only try to get
2226 * the runqueue lock when things look like they will
2232 * If the task is actively running on another CPU
2233 * still, just relax and busy-wait without holding
2236 * NOTE! Since we don't hold any locks, it's not
2237 * even sure that "rq" stays as the right runqueue!
2238 * But we don't care, since "task_running()" will
2239 * return false if the runqueue has changed and p
2240 * is actually now running somewhere else!
2242 while (task_running(rq, p)) {
2243 if (match_state && unlikely(p->state != match_state))
2249 * Ok, time to look more closely! We need the rq
2250 * lock now, to be *sure*. If we're wrong, we'll
2251 * just go back and repeat.
2253 rq = task_rq_lock(p, &flags);
2254 trace_sched_wait_task(p);
2255 running = task_running(rq, p);
2256 on_rq = p->se.on_rq;
2258 if (!match_state || p->state == match_state)
2259 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2260 task_rq_unlock(rq, &flags);
2263 * If it changed from the expected state, bail out now.
2265 if (unlikely(!ncsw))
2269 * Was it really running after all now that we
2270 * checked with the proper locks actually held?
2272 * Oops. Go back and try again..
2274 if (unlikely(running)) {
2280 * It's not enough that it's not actively running,
2281 * it must be off the runqueue _entirely_, and not
2284 * So if it was still runnable (but just not actively
2285 * running right now), it's preempted, and we should
2286 * yield - it could be a while.
2288 if (unlikely(on_rq)) {
2289 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2291 set_current_state(TASK_UNINTERRUPTIBLE);
2292 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2297 * Ahh, all good. It wasn't running, and it wasn't
2298 * runnable, which means that it will never become
2299 * running in the future either. We're all done!
2308 * kick_process - kick a running thread to enter/exit the kernel
2309 * @p: the to-be-kicked thread
2311 * Cause a process which is running on another CPU to enter
2312 * kernel-mode, without any delay. (to get signals handled.)
2314 * NOTE: this function doesnt have to take the runqueue lock,
2315 * because all it wants to ensure is that the remote task enters
2316 * the kernel. If the IPI races and the task has been migrated
2317 * to another CPU then no harm is done and the purpose has been
2320 void kick_process(struct task_struct *p)
2326 if ((cpu != smp_processor_id()) && task_curr(p))
2327 smp_send_reschedule(cpu);
2330 EXPORT_SYMBOL_GPL(kick_process);
2331 #endif /* CONFIG_SMP */
2335 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2337 static int select_fallback_rq(int cpu, struct task_struct *p)
2340 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2342 /* Look for allowed, online CPU in same node. */
2343 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2344 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2347 /* Any allowed, online CPU? */
2348 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2349 if (dest_cpu < nr_cpu_ids)
2352 /* No more Mr. Nice Guy. */
2353 dest_cpu = cpuset_cpus_allowed_fallback(p);
2355 * Don't tell them about moving exiting tasks or
2356 * kernel threads (both mm NULL), since they never
2359 if (p->mm && printk_ratelimit()) {
2360 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2361 task_pid_nr(p), p->comm, cpu);
2368 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2371 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2373 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2376 * In order not to call set_task_cpu() on a blocking task we need
2377 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2380 * Since this is common to all placement strategies, this lives here.
2382 * [ this allows ->select_task() to simply return task_cpu(p) and
2383 * not worry about this generic constraint ]
2385 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2387 cpu = select_fallback_rq(task_cpu(p), p);
2392 static void update_avg(u64 *avg, u64 sample)
2394 s64 diff = sample - *avg;
2399 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2400 bool is_sync, bool is_migrate, bool is_local,
2401 unsigned long en_flags)
2403 schedstat_inc(p, se.statistics.nr_wakeups);
2405 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2407 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2409 schedstat_inc(p, se.statistics.nr_wakeups_local);
2411 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2413 activate_task(rq, p, en_flags);
2416 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2417 int wake_flags, bool success)
2419 trace_sched_wakeup(p, success);
2420 check_preempt_curr(rq, p, wake_flags);
2422 p->state = TASK_RUNNING;
2424 if (p->sched_class->task_woken)
2425 p->sched_class->task_woken(rq, p);
2427 if (unlikely(rq->idle_stamp)) {
2428 u64 delta = rq->clock - rq->idle_stamp;
2429 u64 max = 2*sysctl_sched_migration_cost;
2434 update_avg(&rq->avg_idle, delta);
2438 /* if a worker is waking up, notify workqueue */
2439 if ((p->flags & PF_WQ_WORKER) && success)
2440 wq_worker_waking_up(p, cpu_of(rq));
2444 * try_to_wake_up - wake up a thread
2445 * @p: the thread to be awakened
2446 * @state: the mask of task states that can be woken
2447 * @wake_flags: wake modifier flags (WF_*)
2449 * Put it on the run-queue if it's not already there. The "current"
2450 * thread is always on the run-queue (except when the actual
2451 * re-schedule is in progress), and as such you're allowed to do
2452 * the simpler "current->state = TASK_RUNNING" to mark yourself
2453 * runnable without the overhead of this.
2455 * Returns %true if @p was woken up, %false if it was already running
2456 * or @state didn't match @p's state.
2458 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2461 int cpu, orig_cpu, this_cpu, success = 0;
2462 unsigned long flags;
2463 unsigned long en_flags = ENQUEUE_WAKEUP;
2466 this_cpu = get_cpu();
2469 rq = task_rq_lock(p, &flags);
2470 if (!(p->state & state))
2480 if (unlikely(task_running(rq, p)))
2484 * In order to handle concurrent wakeups and release the rq->lock
2485 * we put the task in TASK_WAKING state.
2487 * First fix up the nr_uninterruptible count:
2489 if (task_contributes_to_load(p)) {
2490 if (likely(cpu_online(orig_cpu)))
2491 rq->nr_uninterruptible--;
2493 this_rq()->nr_uninterruptible--;
2495 p->state = TASK_WAKING;
2497 if (p->sched_class->task_waking) {
2498 p->sched_class->task_waking(rq, p);
2499 en_flags |= ENQUEUE_WAKING;
2502 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2503 if (cpu != orig_cpu)
2504 set_task_cpu(p, cpu);
2505 __task_rq_unlock(rq);
2508 raw_spin_lock(&rq->lock);
2511 * We migrated the task without holding either rq->lock, however
2512 * since the task is not on the task list itself, nobody else
2513 * will try and migrate the task, hence the rq should match the
2514 * cpu we just moved it to.
2516 WARN_ON(task_cpu(p) != cpu);
2517 WARN_ON(p->state != TASK_WAKING);
2519 #ifdef CONFIG_SCHEDSTATS
2520 schedstat_inc(rq, ttwu_count);
2521 if (cpu == this_cpu)
2522 schedstat_inc(rq, ttwu_local);
2524 struct sched_domain *sd;
2525 for_each_domain(this_cpu, sd) {
2526 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2527 schedstat_inc(sd, ttwu_wake_remote);
2532 #endif /* CONFIG_SCHEDSTATS */
2535 #endif /* CONFIG_SMP */
2536 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2537 cpu == this_cpu, en_flags);
2540 ttwu_post_activation(p, rq, wake_flags, success);
2542 task_rq_unlock(rq, &flags);
2549 * try_to_wake_up_local - try to wake up a local task with rq lock held
2550 * @p: the thread to be awakened
2552 * Put @p on the run-queue if it's not already there. The caller must
2553 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2554 * the current task. this_rq() stays locked over invocation.
2556 static void try_to_wake_up_local(struct task_struct *p)
2558 struct rq *rq = task_rq(p);
2559 bool success = false;
2561 BUG_ON(rq != this_rq());
2562 BUG_ON(p == current);
2563 lockdep_assert_held(&rq->lock);
2565 if (!(p->state & TASK_NORMAL))
2569 if (likely(!task_running(rq, p))) {
2570 schedstat_inc(rq, ttwu_count);
2571 schedstat_inc(rq, ttwu_local);
2573 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2576 ttwu_post_activation(p, rq, 0, success);
2580 * wake_up_process - Wake up a specific process
2581 * @p: The process to be woken up.
2583 * Attempt to wake up the nominated process and move it to the set of runnable
2584 * processes. Returns 1 if the process was woken up, 0 if it was already
2587 * It may be assumed that this function implies a write memory barrier before
2588 * changing the task state if and only if any tasks are woken up.
2590 int wake_up_process(struct task_struct *p)
2592 return try_to_wake_up(p, TASK_ALL, 0);
2594 EXPORT_SYMBOL(wake_up_process);
2596 int wake_up_state(struct task_struct *p, unsigned int state)
2598 return try_to_wake_up(p, state, 0);
2602 * Perform scheduler related setup for a newly forked process p.
2603 * p is forked by current.
2605 * __sched_fork() is basic setup used by init_idle() too:
2607 static void __sched_fork(struct task_struct *p)
2609 p->se.exec_start = 0;
2610 p->se.sum_exec_runtime = 0;
2611 p->se.prev_sum_exec_runtime = 0;
2612 p->se.nr_migrations = 0;
2615 #ifdef CONFIG_SCHEDSTATS
2616 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2619 INIT_LIST_HEAD(&p->rt.run_list);
2621 INIT_LIST_HEAD(&p->se.group_node);
2623 #ifdef CONFIG_PREEMPT_NOTIFIERS
2624 INIT_HLIST_HEAD(&p->preempt_notifiers);
2629 * fork()/clone()-time setup:
2631 void sched_fork(struct task_struct *p, int clone_flags)
2633 int cpu = get_cpu();
2637 * We mark the process as running here. This guarantees that
2638 * nobody will actually run it, and a signal or other external
2639 * event cannot wake it up and insert it on the runqueue either.
2641 p->state = TASK_RUNNING;
2644 * Revert to default priority/policy on fork if requested.
2646 if (unlikely(p->sched_reset_on_fork)) {
2647 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2648 p->policy = SCHED_NORMAL;
2649 p->normal_prio = p->static_prio;
2652 if (PRIO_TO_NICE(p->static_prio) < 0) {
2653 p->static_prio = NICE_TO_PRIO(0);
2654 p->normal_prio = p->static_prio;
2659 * We don't need the reset flag anymore after the fork. It has
2660 * fulfilled its duty:
2662 p->sched_reset_on_fork = 0;
2666 * Make sure we do not leak PI boosting priority to the child.
2668 p->prio = current->normal_prio;
2670 if (!rt_prio(p->prio))
2671 p->sched_class = &fair_sched_class;
2673 if (p->sched_class->task_fork)
2674 p->sched_class->task_fork(p);
2677 * The child is not yet in the pid-hash so no cgroup attach races,
2678 * and the cgroup is pinned to this child due to cgroup_fork()
2679 * is ran before sched_fork().
2681 * Silence PROVE_RCU.
2684 set_task_cpu(p, cpu);
2687 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2688 if (likely(sched_info_on()))
2689 memset(&p->sched_info, 0, sizeof(p->sched_info));
2691 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2694 #ifdef CONFIG_PREEMPT
2695 /* Want to start with kernel preemption disabled. */
2696 task_thread_info(p)->preempt_count = 1;
2699 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2706 * wake_up_new_task - wake up a newly created task for the first time.
2708 * This function will do some initial scheduler statistics housekeeping
2709 * that must be done for every newly created context, then puts the task
2710 * on the runqueue and wakes it.
2712 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2714 unsigned long flags;
2716 int cpu __maybe_unused = get_cpu();
2719 rq = task_rq_lock(p, &flags);
2720 p->state = TASK_WAKING;
2723 * Fork balancing, do it here and not earlier because:
2724 * - cpus_allowed can change in the fork path
2725 * - any previously selected cpu might disappear through hotplug
2727 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2728 * without people poking at ->cpus_allowed.
2730 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2731 set_task_cpu(p, cpu);
2733 p->state = TASK_RUNNING;
2734 task_rq_unlock(rq, &flags);
2737 rq = task_rq_lock(p, &flags);
2738 activate_task(rq, p, 0);
2739 trace_sched_wakeup_new(p, 1);
2740 check_preempt_curr(rq, p, WF_FORK);
2742 if (p->sched_class->task_woken)
2743 p->sched_class->task_woken(rq, p);
2745 task_rq_unlock(rq, &flags);
2749 #ifdef CONFIG_PREEMPT_NOTIFIERS
2752 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2753 * @notifier: notifier struct to register
2755 void preempt_notifier_register(struct preempt_notifier *notifier)
2757 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2759 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2762 * preempt_notifier_unregister - no longer interested in preemption notifications
2763 * @notifier: notifier struct to unregister
2765 * This is safe to call from within a preemption notifier.
2767 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2769 hlist_del(¬ifier->link);
2771 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2773 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2775 struct preempt_notifier *notifier;
2776 struct hlist_node *node;
2778 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2779 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2783 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2784 struct task_struct *next)
2786 struct preempt_notifier *notifier;
2787 struct hlist_node *node;
2789 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2790 notifier->ops->sched_out(notifier, next);
2793 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2795 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2800 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2801 struct task_struct *next)
2805 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2808 * prepare_task_switch - prepare to switch tasks
2809 * @rq: the runqueue preparing to switch
2810 * @prev: the current task that is being switched out
2811 * @next: the task we are going to switch to.
2813 * This is called with the rq lock held and interrupts off. It must
2814 * be paired with a subsequent finish_task_switch after the context
2817 * prepare_task_switch sets up locking and calls architecture specific
2821 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2822 struct task_struct *next)
2824 sched_info_switch(prev, next);
2825 perf_event_task_sched_out(prev, next);
2826 fire_sched_out_preempt_notifiers(prev, next);
2827 prepare_lock_switch(rq, next);
2828 prepare_arch_switch(next);
2829 trace_sched_switch(prev, next);
2833 * finish_task_switch - clean up after a task-switch
2834 * @rq: runqueue associated with task-switch
2835 * @prev: the thread we just switched away from.
2837 * finish_task_switch must be called after the context switch, paired
2838 * with a prepare_task_switch call before the context switch.
2839 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2840 * and do any other architecture-specific cleanup actions.
2842 * Note that we may have delayed dropping an mm in context_switch(). If
2843 * so, we finish that here outside of the runqueue lock. (Doing it
2844 * with the lock held can cause deadlocks; see schedule() for
2847 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2848 __releases(rq->lock)
2850 struct mm_struct *mm = rq->prev_mm;
2856 * A task struct has one reference for the use as "current".
2857 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2858 * schedule one last time. The schedule call will never return, and
2859 * the scheduled task must drop that reference.
2860 * The test for TASK_DEAD must occur while the runqueue locks are
2861 * still held, otherwise prev could be scheduled on another cpu, die
2862 * there before we look at prev->state, and then the reference would
2864 * Manfred Spraul <manfred@colorfullife.com>
2866 prev_state = prev->state;
2867 finish_arch_switch(prev);
2868 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2869 local_irq_disable();
2870 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2871 perf_event_task_sched_in(current);
2872 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2874 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2875 finish_lock_switch(rq, prev);
2877 fire_sched_in_preempt_notifiers(current);
2880 if (unlikely(prev_state == TASK_DEAD)) {
2882 * Remove function-return probe instances associated with this
2883 * task and put them back on the free list.
2885 kprobe_flush_task(prev);
2886 put_task_struct(prev);
2892 /* assumes rq->lock is held */
2893 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2895 if (prev->sched_class->pre_schedule)
2896 prev->sched_class->pre_schedule(rq, prev);
2899 /* rq->lock is NOT held, but preemption is disabled */
2900 static inline void post_schedule(struct rq *rq)
2902 if (rq->post_schedule) {
2903 unsigned long flags;
2905 raw_spin_lock_irqsave(&rq->lock, flags);
2906 if (rq->curr->sched_class->post_schedule)
2907 rq->curr->sched_class->post_schedule(rq);
2908 raw_spin_unlock_irqrestore(&rq->lock, flags);
2910 rq->post_schedule = 0;
2916 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2920 static inline void post_schedule(struct rq *rq)
2927 * schedule_tail - first thing a freshly forked thread must call.
2928 * @prev: the thread we just switched away from.
2930 asmlinkage void schedule_tail(struct task_struct *prev)
2931 __releases(rq->lock)
2933 struct rq *rq = this_rq();
2935 finish_task_switch(rq, prev);
2938 * FIXME: do we need to worry about rq being invalidated by the
2943 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2944 /* In this case, finish_task_switch does not reenable preemption */
2947 if (current->set_child_tid)
2948 put_user(task_pid_vnr(current), current->set_child_tid);
2952 * context_switch - switch to the new MM and the new
2953 * thread's register state.
2956 context_switch(struct rq *rq, struct task_struct *prev,
2957 struct task_struct *next)
2959 struct mm_struct *mm, *oldmm;
2961 prepare_task_switch(rq, prev, next);
2964 oldmm = prev->active_mm;
2966 * For paravirt, this is coupled with an exit in switch_to to
2967 * combine the page table reload and the switch backend into
2970 arch_start_context_switch(prev);
2973 next->active_mm = oldmm;
2974 atomic_inc(&oldmm->mm_count);
2975 enter_lazy_tlb(oldmm, next);
2977 switch_mm(oldmm, mm, next);
2980 prev->active_mm = NULL;
2981 rq->prev_mm = oldmm;
2984 * Since the runqueue lock will be released by the next
2985 * task (which is an invalid locking op but in the case
2986 * of the scheduler it's an obvious special-case), so we
2987 * do an early lockdep release here:
2989 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2990 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2993 /* Here we just switch the register state and the stack. */
2994 switch_to(prev, next, prev);
2998 * this_rq must be evaluated again because prev may have moved
2999 * CPUs since it called schedule(), thus the 'rq' on its stack
3000 * frame will be invalid.
3002 finish_task_switch(this_rq(), prev);
3006 * nr_running, nr_uninterruptible and nr_context_switches:
3008 * externally visible scheduler statistics: current number of runnable
3009 * threads, current number of uninterruptible-sleeping threads, total
3010 * number of context switches performed since bootup.
3012 unsigned long nr_running(void)
3014 unsigned long i, sum = 0;
3016 for_each_online_cpu(i)
3017 sum += cpu_rq(i)->nr_running;
3022 unsigned long nr_uninterruptible(void)
3024 unsigned long i, sum = 0;
3026 for_each_possible_cpu(i)
3027 sum += cpu_rq(i)->nr_uninterruptible;
3030 * Since we read the counters lockless, it might be slightly
3031 * inaccurate. Do not allow it to go below zero though:
3033 if (unlikely((long)sum < 0))
3039 unsigned long long nr_context_switches(void)
3042 unsigned long long sum = 0;
3044 for_each_possible_cpu(i)
3045 sum += cpu_rq(i)->nr_switches;
3050 unsigned long nr_iowait(void)
3052 unsigned long i, sum = 0;
3054 for_each_possible_cpu(i)
3055 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3060 unsigned long nr_iowait_cpu(int cpu)
3062 struct rq *this = cpu_rq(cpu);
3063 return atomic_read(&this->nr_iowait);
3066 unsigned long this_cpu_load(void)
3068 struct rq *this = this_rq();
3069 return this->cpu_load[0];
3073 /* Variables and functions for calc_load */
3074 static atomic_long_t calc_load_tasks;
3075 static unsigned long calc_load_update;
3076 unsigned long avenrun[3];
3077 EXPORT_SYMBOL(avenrun);
3079 static long calc_load_fold_active(struct rq *this_rq)
3081 long nr_active, delta = 0;
3083 nr_active = this_rq->nr_running;
3084 nr_active += (long) this_rq->nr_uninterruptible;
3086 if (nr_active != this_rq->calc_load_active) {
3087 delta = nr_active - this_rq->calc_load_active;
3088 this_rq->calc_load_active = nr_active;
3094 static unsigned long
3095 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3098 load += active * (FIXED_1 - exp);
3099 load += 1UL << (FSHIFT - 1);
3100 return load >> FSHIFT;
3105 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3107 * When making the ILB scale, we should try to pull this in as well.
3109 static atomic_long_t calc_load_tasks_idle;
3111 static void calc_load_account_idle(struct rq *this_rq)
3115 delta = calc_load_fold_active(this_rq);
3117 atomic_long_add(delta, &calc_load_tasks_idle);
3120 static long calc_load_fold_idle(void)
3125 * Its got a race, we don't care...
3127 if (atomic_long_read(&calc_load_tasks_idle))
3128 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3134 * fixed_power_int - compute: x^n, in O(log n) time
3136 * @x: base of the power
3137 * @frac_bits: fractional bits of @x
3138 * @n: power to raise @x to.
3140 * By exploiting the relation between the definition of the natural power
3141 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3142 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3143 * (where: n_i \elem {0, 1}, the binary vector representing n),
3144 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3145 * of course trivially computable in O(log_2 n), the length of our binary
3148 static unsigned long
3149 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3151 unsigned long result = 1UL << frac_bits;
3156 result += 1UL << (frac_bits - 1);
3157 result >>= frac_bits;
3163 x += 1UL << (frac_bits - 1);
3171 * a1 = a0 * e + a * (1 - e)
3173 * a2 = a1 * e + a * (1 - e)
3174 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3175 * = a0 * e^2 + a * (1 - e) * (1 + e)
3177 * a3 = a2 * e + a * (1 - e)
3178 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3179 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3183 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3184 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3185 * = a0 * e^n + a * (1 - e^n)
3187 * [1] application of the geometric series:
3190 * S_n := \Sum x^i = -------------
3193 static unsigned long
3194 calc_load_n(unsigned long load, unsigned long exp,
3195 unsigned long active, unsigned int n)
3198 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3202 * NO_HZ can leave us missing all per-cpu ticks calling
3203 * calc_load_account_active(), but since an idle CPU folds its delta into
3204 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3205 * in the pending idle delta if our idle period crossed a load cycle boundary.
3207 * Once we've updated the global active value, we need to apply the exponential
3208 * weights adjusted to the number of cycles missed.
3210 static void calc_global_nohz(unsigned long ticks)
3212 long delta, active, n;
3214 if (time_before(jiffies, calc_load_update))
3218 * If we crossed a calc_load_update boundary, make sure to fold
3219 * any pending idle changes, the respective CPUs might have
3220 * missed the tick driven calc_load_account_active() update
3223 delta = calc_load_fold_idle();
3225 atomic_long_add(delta, &calc_load_tasks);
3228 * If we were idle for multiple load cycles, apply them.
3230 if (ticks >= LOAD_FREQ) {
3231 n = ticks / LOAD_FREQ;
3233 active = atomic_long_read(&calc_load_tasks);
3234 active = active > 0 ? active * FIXED_1 : 0;
3236 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3237 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3238 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3240 calc_load_update += n * LOAD_FREQ;
3244 * Its possible the remainder of the above division also crosses
3245 * a LOAD_FREQ period, the regular check in calc_global_load()
3246 * which comes after this will take care of that.
3248 * Consider us being 11 ticks before a cycle completion, and us
3249 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3250 * age us 4 cycles, and the test in calc_global_load() will
3251 * pick up the final one.
3255 static void calc_load_account_idle(struct rq *this_rq)
3259 static inline long calc_load_fold_idle(void)
3264 static void calc_global_nohz(unsigned long ticks)
3270 * get_avenrun - get the load average array
3271 * @loads: pointer to dest load array
3272 * @offset: offset to add
3273 * @shift: shift count to shift the result left
3275 * These values are estimates at best, so no need for locking.
3277 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3279 loads[0] = (avenrun[0] + offset) << shift;
3280 loads[1] = (avenrun[1] + offset) << shift;
3281 loads[2] = (avenrun[2] + offset) << shift;
3285 * calc_load - update the avenrun load estimates 10 ticks after the
3286 * CPUs have updated calc_load_tasks.
3288 void calc_global_load(unsigned long ticks)
3292 calc_global_nohz(ticks);
3294 if (time_before(jiffies, calc_load_update + 10))
3297 active = atomic_long_read(&calc_load_tasks);
3298 active = active > 0 ? active * FIXED_1 : 0;
3300 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3301 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3302 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3304 calc_load_update += LOAD_FREQ;
3308 * Called from update_cpu_load() to periodically update this CPU's
3311 static void calc_load_account_active(struct rq *this_rq)
3315 if (time_before(jiffies, this_rq->calc_load_update))
3318 delta = calc_load_fold_active(this_rq);
3319 delta += calc_load_fold_idle();
3321 atomic_long_add(delta, &calc_load_tasks);
3323 this_rq->calc_load_update += LOAD_FREQ;
3327 * The exact cpuload at various idx values, calculated at every tick would be
3328 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3330 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3331 * on nth tick when cpu may be busy, then we have:
3332 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3333 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3335 * decay_load_missed() below does efficient calculation of
3336 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3337 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3339 * The calculation is approximated on a 128 point scale.
3340 * degrade_zero_ticks is the number of ticks after which load at any
3341 * particular idx is approximated to be zero.
3342 * degrade_factor is a precomputed table, a row for each load idx.
3343 * Each column corresponds to degradation factor for a power of two ticks,
3344 * based on 128 point scale.
3346 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3347 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3349 * With this power of 2 load factors, we can degrade the load n times
3350 * by looking at 1 bits in n and doing as many mult/shift instead of
3351 * n mult/shifts needed by the exact degradation.
3353 #define DEGRADE_SHIFT 7
3354 static const unsigned char
3355 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3356 static const unsigned char
3357 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3358 {0, 0, 0, 0, 0, 0, 0, 0},
3359 {64, 32, 8, 0, 0, 0, 0, 0},
3360 {96, 72, 40, 12, 1, 0, 0},
3361 {112, 98, 75, 43, 15, 1, 0},
3362 {120, 112, 98, 76, 45, 16, 2} };
3365 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3366 * would be when CPU is idle and so we just decay the old load without
3367 * adding any new load.
3369 static unsigned long
3370 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3374 if (!missed_updates)
3377 if (missed_updates >= degrade_zero_ticks[idx])
3381 return load >> missed_updates;
3383 while (missed_updates) {
3384 if (missed_updates % 2)
3385 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3387 missed_updates >>= 1;
3394 * Update rq->cpu_load[] statistics. This function is usually called every
3395 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3396 * every tick. We fix it up based on jiffies.
3398 static void update_cpu_load(struct rq *this_rq)
3400 unsigned long this_load = this_rq->load.weight;
3401 unsigned long curr_jiffies = jiffies;
3402 unsigned long pending_updates;
3405 this_rq->nr_load_updates++;
3407 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3408 if (curr_jiffies == this_rq->last_load_update_tick)
3411 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3412 this_rq->last_load_update_tick = curr_jiffies;
3414 /* Update our load: */
3415 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3416 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3417 unsigned long old_load, new_load;
3419 /* scale is effectively 1 << i now, and >> i divides by scale */
3421 old_load = this_rq->cpu_load[i];
3422 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3423 new_load = this_load;
3425 * Round up the averaging division if load is increasing. This
3426 * prevents us from getting stuck on 9 if the load is 10, for
3429 if (new_load > old_load)
3430 new_load += scale - 1;
3432 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3435 sched_avg_update(this_rq);
3438 static void update_cpu_load_active(struct rq *this_rq)
3440 update_cpu_load(this_rq);
3442 calc_load_account_active(this_rq);
3448 * sched_exec - execve() is a valuable balancing opportunity, because at
3449 * this point the task has the smallest effective memory and cache footprint.
3451 void sched_exec(void)
3453 struct task_struct *p = current;
3454 unsigned long flags;
3458 rq = task_rq_lock(p, &flags);
3459 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3460 if (dest_cpu == smp_processor_id())
3464 * select_task_rq() can race against ->cpus_allowed
3466 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3467 likely(cpu_active(dest_cpu)) && migrate_task(p, rq)) {
3468 struct migration_arg arg = { p, dest_cpu };
3470 task_rq_unlock(rq, &flags);
3471 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3475 task_rq_unlock(rq, &flags);
3480 DEFINE_PER_CPU(struct kernel_stat, kstat);
3482 EXPORT_PER_CPU_SYMBOL(kstat);
3485 * Return any ns on the sched_clock that have not yet been accounted in
3486 * @p in case that task is currently running.
3488 * Called with task_rq_lock() held on @rq.
3490 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3494 if (task_current(rq, p)) {
3495 update_rq_clock(rq);
3496 ns = rq->clock_task - p->se.exec_start;
3504 unsigned long long task_delta_exec(struct task_struct *p)
3506 unsigned long flags;
3510 rq = task_rq_lock(p, &flags);
3511 ns = do_task_delta_exec(p, rq);
3512 task_rq_unlock(rq, &flags);
3518 * Return accounted runtime for the task.
3519 * In case the task is currently running, return the runtime plus current's
3520 * pending runtime that have not been accounted yet.
3522 unsigned long long task_sched_runtime(struct task_struct *p)
3524 unsigned long flags;
3528 rq = task_rq_lock(p, &flags);
3529 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3530 task_rq_unlock(rq, &flags);
3536 * Return sum_exec_runtime for the thread group.
3537 * In case the task is currently running, return the sum plus current's
3538 * pending runtime that have not been accounted yet.
3540 * Note that the thread group might have other running tasks as well,
3541 * so the return value not includes other pending runtime that other
3542 * running tasks might have.
3544 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3546 struct task_cputime totals;
3547 unsigned long flags;
3551 rq = task_rq_lock(p, &flags);
3552 thread_group_cputime(p, &totals);
3553 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3554 task_rq_unlock(rq, &flags);
3560 * Account user cpu time to a process.
3561 * @p: the process that the cpu time gets accounted to
3562 * @cputime: the cpu time spent in user space since the last update
3563 * @cputime_scaled: cputime scaled by cpu frequency
3565 void account_user_time(struct task_struct *p, cputime_t cputime,
3566 cputime_t cputime_scaled)
3568 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3571 /* Add user time to process. */
3572 p->utime = cputime_add(p->utime, cputime);
3573 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3574 account_group_user_time(p, cputime);
3576 /* Add user time to cpustat. */
3577 tmp = cputime_to_cputime64(cputime);
3578 if (TASK_NICE(p) > 0)
3579 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3581 cpustat->user = cputime64_add(cpustat->user, tmp);
3583 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3584 /* Account for user time used */
3585 acct_update_integrals(p);
3589 * Account guest cpu time to a process.
3590 * @p: the process that the cpu time gets accounted to
3591 * @cputime: the cpu time spent in virtual machine since the last update
3592 * @cputime_scaled: cputime scaled by cpu frequency
3594 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3595 cputime_t cputime_scaled)
3598 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3600 tmp = cputime_to_cputime64(cputime);
3602 /* Add guest time to process. */
3603 p->utime = cputime_add(p->utime, cputime);
3604 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3605 account_group_user_time(p, cputime);
3606 p->gtime = cputime_add(p->gtime, cputime);
3608 /* Add guest time to cpustat. */
3609 if (TASK_NICE(p) > 0) {
3610 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3611 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3613 cpustat->user = cputime64_add(cpustat->user, tmp);
3614 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3619 * Account system cpu time to a process and desired cpustat field
3620 * @p: the process that the cpu time gets accounted to
3621 * @cputime: the cpu time spent in kernel space since the last update
3622 * @cputime_scaled: cputime scaled by cpu frequency
3623 * @target_cputime64: pointer to cpustat field that has to be updated
3626 void __account_system_time(struct task_struct *p, cputime_t cputime,
3627 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3629 cputime64_t tmp = cputime_to_cputime64(cputime);
3631 /* Add system time to process. */
3632 p->stime = cputime_add(p->stime, cputime);
3633 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3634 account_group_system_time(p, cputime);
3636 /* Add system time to cpustat. */
3637 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3638 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3640 /* Account for system time used */
3641 acct_update_integrals(p);
3645 * Account system cpu time to a process.
3646 * @p: the process that the cpu time gets accounted to
3647 * @hardirq_offset: the offset to subtract from hardirq_count()
3648 * @cputime: the cpu time spent in kernel space since the last update
3649 * @cputime_scaled: cputime scaled by cpu frequency
3651 void account_system_time(struct task_struct *p, int hardirq_offset,
3652 cputime_t cputime, cputime_t cputime_scaled)
3654 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3655 cputime64_t *target_cputime64;
3657 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3658 account_guest_time(p, cputime, cputime_scaled);
3662 if (hardirq_count() - hardirq_offset)
3663 target_cputime64 = &cpustat->irq;
3664 else if (in_serving_softirq())
3665 target_cputime64 = &cpustat->softirq;
3667 target_cputime64 = &cpustat->system;
3669 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3673 * Account for involuntary wait time.
3674 * @cputime: the cpu time spent in involuntary wait
3676 void account_steal_time(cputime_t cputime)
3678 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3679 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3681 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3685 * Account for idle time.
3686 * @cputime: the cpu time spent in idle wait
3688 void account_idle_time(cputime_t cputime)
3690 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3691 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3692 struct rq *rq = this_rq();
3694 if (atomic_read(&rq->nr_iowait) > 0)
3695 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3697 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3700 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3702 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3704 * Account a tick to a process and cpustat
3705 * @p: the process that the cpu time gets accounted to
3706 * @user_tick: is the tick from userspace
3707 * @rq: the pointer to rq
3709 * Tick demultiplexing follows the order
3710 * - pending hardirq update
3711 * - pending softirq update
3715 * - check for guest_time
3716 * - else account as system_time
3718 * Check for hardirq is done both for system and user time as there is
3719 * no timer going off while we are on hardirq and hence we may never get an
3720 * opportunity to update it solely in system time.
3721 * p->stime and friends are only updated on system time and not on irq
3722 * softirq as those do not count in task exec_runtime any more.
3724 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3727 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3728 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3729 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3731 if (irqtime_account_hi_update()) {
3732 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3733 } else if (irqtime_account_si_update()) {
3734 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3735 } else if (this_cpu_ksoftirqd() == p) {
3737 * ksoftirqd time do not get accounted in cpu_softirq_time.
3738 * So, we have to handle it separately here.
3739 * Also, p->stime needs to be updated for ksoftirqd.
3741 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3743 } else if (user_tick) {
3744 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3745 } else if (p == rq->idle) {
3746 account_idle_time(cputime_one_jiffy);
3747 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3748 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3750 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3755 static void irqtime_account_idle_ticks(int ticks)
3758 struct rq *rq = this_rq();
3760 for (i = 0; i < ticks; i++)
3761 irqtime_account_process_tick(current, 0, rq);
3763 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3764 static void irqtime_account_idle_ticks(int ticks) {}
3765 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3767 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3770 * Account a single tick of cpu time.
3771 * @p: the process that the cpu time gets accounted to
3772 * @user_tick: indicates if the tick is a user or a system tick
3774 void account_process_tick(struct task_struct *p, int user_tick)
3776 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3777 struct rq *rq = this_rq();
3779 if (sched_clock_irqtime) {
3780 irqtime_account_process_tick(p, user_tick, rq);
3785 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3786 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3787 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3790 account_idle_time(cputime_one_jiffy);
3794 * Account multiple ticks of steal time.
3795 * @p: the process from which the cpu time has been stolen
3796 * @ticks: number of stolen ticks
3798 void account_steal_ticks(unsigned long ticks)
3800 account_steal_time(jiffies_to_cputime(ticks));
3804 * Account multiple ticks of idle time.
3805 * @ticks: number of stolen ticks
3807 void account_idle_ticks(unsigned long ticks)
3810 if (sched_clock_irqtime) {
3811 irqtime_account_idle_ticks(ticks);
3815 account_idle_time(jiffies_to_cputime(ticks));
3821 * Use precise platform statistics if available:
3823 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3824 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3830 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3832 struct task_cputime cputime;
3834 thread_group_cputime(p, &cputime);
3836 *ut = cputime.utime;
3837 *st = cputime.stime;
3841 #ifndef nsecs_to_cputime
3842 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3845 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3847 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3850 * Use CFS's precise accounting:
3852 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3858 do_div(temp, total);
3859 utime = (cputime_t)temp;
3864 * Compare with previous values, to keep monotonicity:
3866 p->prev_utime = max(p->prev_utime, utime);
3867 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3869 *ut = p->prev_utime;
3870 *st = p->prev_stime;
3874 * Must be called with siglock held.
3876 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3878 struct signal_struct *sig = p->signal;
3879 struct task_cputime cputime;
3880 cputime_t rtime, utime, total;
3882 thread_group_cputime(p, &cputime);
3884 total = cputime_add(cputime.utime, cputime.stime);
3885 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3890 temp *= cputime.utime;
3891 do_div(temp, total);
3892 utime = (cputime_t)temp;
3896 sig->prev_utime = max(sig->prev_utime, utime);
3897 sig->prev_stime = max(sig->prev_stime,
3898 cputime_sub(rtime, sig->prev_utime));
3900 *ut = sig->prev_utime;
3901 *st = sig->prev_stime;
3906 * This function gets called by the timer code, with HZ frequency.
3907 * We call it with interrupts disabled.
3909 * It also gets called by the fork code, when changing the parent's
3912 void scheduler_tick(void)
3914 int cpu = smp_processor_id();
3915 struct rq *rq = cpu_rq(cpu);
3916 struct task_struct *curr = rq->curr;
3920 raw_spin_lock(&rq->lock);
3921 update_rq_clock(rq);
3922 update_cpu_load_active(rq);
3923 curr->sched_class->task_tick(rq, curr, 0);
3924 raw_spin_unlock(&rq->lock);
3926 perf_event_task_tick();
3929 rq->idle_at_tick = idle_cpu(cpu);
3930 trigger_load_balance(rq, cpu);
3934 notrace unsigned long get_parent_ip(unsigned long addr)
3936 if (in_lock_functions(addr)) {
3937 addr = CALLER_ADDR2;
3938 if (in_lock_functions(addr))
3939 addr = CALLER_ADDR3;
3944 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3945 defined(CONFIG_PREEMPT_TRACER))
3947 void __kprobes add_preempt_count(int val)
3949 #ifdef CONFIG_DEBUG_PREEMPT
3953 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3956 preempt_count() += val;
3957 #ifdef CONFIG_DEBUG_PREEMPT
3959 * Spinlock count overflowing soon?
3961 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3964 if (preempt_count() == val)
3965 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3967 EXPORT_SYMBOL(add_preempt_count);
3969 void __kprobes sub_preempt_count(int val)
3971 #ifdef CONFIG_DEBUG_PREEMPT
3975 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3978 * Is the spinlock portion underflowing?
3980 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3981 !(preempt_count() & PREEMPT_MASK)))
3985 if (preempt_count() == val)
3986 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3987 preempt_count() -= val;
3989 EXPORT_SYMBOL(sub_preempt_count);
3994 * Print scheduling while atomic bug:
3996 static noinline void __schedule_bug(struct task_struct *prev)
3998 struct pt_regs *regs = get_irq_regs();
4000 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4001 prev->comm, prev->pid, preempt_count());
4003 debug_show_held_locks(prev);
4005 if (irqs_disabled())
4006 print_irqtrace_events(prev);
4015 * Various schedule()-time debugging checks and statistics:
4017 static inline void schedule_debug(struct task_struct *prev)
4020 * Test if we are atomic. Since do_exit() needs to call into
4021 * schedule() atomically, we ignore that path for now.
4022 * Otherwise, whine if we are scheduling when we should not be.
4024 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4025 __schedule_bug(prev);
4027 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4029 schedstat_inc(this_rq(), sched_count);
4030 #ifdef CONFIG_SCHEDSTATS
4031 if (unlikely(prev->lock_depth >= 0)) {
4032 schedstat_inc(this_rq(), rq_sched_info.bkl_count);
4033 schedstat_inc(prev, sched_info.bkl_count);
4038 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4041 update_rq_clock(rq);
4042 prev->sched_class->put_prev_task(rq, prev);
4046 * Pick up the highest-prio task:
4048 static inline struct task_struct *
4049 pick_next_task(struct rq *rq)
4051 const struct sched_class *class;
4052 struct task_struct *p;
4055 * Optimization: we know that if all tasks are in
4056 * the fair class we can call that function directly:
4058 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4059 p = fair_sched_class.pick_next_task(rq);
4064 for_each_class(class) {
4065 p = class->pick_next_task(rq);
4070 BUG(); /* the idle class will always have a runnable task */
4074 * schedule() is the main scheduler function.
4076 asmlinkage void __sched schedule(void)
4078 struct task_struct *prev, *next;
4079 unsigned long *switch_count;
4085 cpu = smp_processor_id();
4087 rcu_note_context_switch(cpu);
4090 release_kernel_lock(prev);
4091 need_resched_nonpreemptible:
4093 schedule_debug(prev);
4095 if (sched_feat(HRTICK))
4098 raw_spin_lock_irq(&rq->lock);
4100 switch_count = &prev->nivcsw;
4101 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4102 if (unlikely(signal_pending_state(prev->state, prev))) {
4103 prev->state = TASK_RUNNING;
4106 * If a worker is going to sleep, notify and
4107 * ask workqueue whether it wants to wake up a
4108 * task to maintain concurrency. If so, wake
4111 if (prev->flags & PF_WQ_WORKER) {
4112 struct task_struct *to_wakeup;
4114 to_wakeup = wq_worker_sleeping(prev, cpu);
4116 try_to_wake_up_local(to_wakeup);
4118 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4120 switch_count = &prev->nvcsw;
4123 pre_schedule(rq, prev);
4125 if (unlikely(!rq->nr_running))
4126 idle_balance(cpu, rq);
4128 put_prev_task(rq, prev);
4129 next = pick_next_task(rq);
4130 clear_tsk_need_resched(prev);
4131 rq->skip_clock_update = 0;
4133 if (likely(prev != next)) {
4138 context_switch(rq, prev, next); /* unlocks the rq */
4140 * The context switch have flipped the stack from under us
4141 * and restored the local variables which were saved when
4142 * this task called schedule() in the past. prev == current
4143 * is still correct, but it can be moved to another cpu/rq.
4145 cpu = smp_processor_id();
4148 raw_spin_unlock_irq(&rq->lock);
4152 if (unlikely(reacquire_kernel_lock(prev)))
4153 goto need_resched_nonpreemptible;
4155 preempt_enable_no_resched();
4159 EXPORT_SYMBOL(schedule);
4161 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4163 * Look out! "owner" is an entirely speculative pointer
4164 * access and not reliable.
4166 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
4171 if (!sched_feat(OWNER_SPIN))
4174 #ifdef CONFIG_DEBUG_PAGEALLOC
4176 * Need to access the cpu field knowing that
4177 * DEBUG_PAGEALLOC could have unmapped it if
4178 * the mutex owner just released it and exited.
4180 if (probe_kernel_address(&owner->cpu, cpu))
4187 * Even if the access succeeded (likely case),
4188 * the cpu field may no longer be valid.
4190 if (cpu >= nr_cpumask_bits)
4194 * We need to validate that we can do a
4195 * get_cpu() and that we have the percpu area.
4197 if (!cpu_online(cpu))
4204 * Owner changed, break to re-assess state.
4206 if (lock->owner != owner) {
4208 * If the lock has switched to a different owner,
4209 * we likely have heavy contention. Return 0 to quit
4210 * optimistic spinning and not contend further:
4218 * Is that owner really running on that cpu?
4220 if (task_thread_info(rq->curr) != owner || need_resched())
4223 arch_mutex_cpu_relax();
4230 #ifdef CONFIG_PREEMPT
4232 * this is the entry point to schedule() from in-kernel preemption
4233 * off of preempt_enable. Kernel preemptions off return from interrupt
4234 * occur there and call schedule directly.
4236 asmlinkage void __sched notrace preempt_schedule(void)
4238 struct thread_info *ti = current_thread_info();
4241 * If there is a non-zero preempt_count or interrupts are disabled,
4242 * we do not want to preempt the current task. Just return..
4244 if (likely(ti->preempt_count || irqs_disabled()))
4248 add_preempt_count_notrace(PREEMPT_ACTIVE);
4250 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4253 * Check again in case we missed a preemption opportunity
4254 * between schedule and now.
4257 } while (need_resched());
4259 EXPORT_SYMBOL(preempt_schedule);
4262 * this is the entry point to schedule() from kernel preemption
4263 * off of irq context.
4264 * Note, that this is called and return with irqs disabled. This will
4265 * protect us against recursive calling from irq.
4267 asmlinkage void __sched preempt_schedule_irq(void)
4269 struct thread_info *ti = current_thread_info();
4271 /* Catch callers which need to be fixed */
4272 BUG_ON(ti->preempt_count || !irqs_disabled());
4275 add_preempt_count(PREEMPT_ACTIVE);
4278 local_irq_disable();
4279 sub_preempt_count(PREEMPT_ACTIVE);
4282 * Check again in case we missed a preemption opportunity
4283 * between schedule and now.
4286 } while (need_resched());
4289 #endif /* CONFIG_PREEMPT */
4291 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4294 return try_to_wake_up(curr->private, mode, wake_flags);
4296 EXPORT_SYMBOL(default_wake_function);
4299 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4300 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4301 * number) then we wake all the non-exclusive tasks and one exclusive task.
4303 * There are circumstances in which we can try to wake a task which has already
4304 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4305 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4307 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4308 int nr_exclusive, int wake_flags, void *key)
4310 wait_queue_t *curr, *next;
4312 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4313 unsigned flags = curr->flags;
4315 if (curr->func(curr, mode, wake_flags, key) &&
4316 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4322 * __wake_up - wake up threads blocked on a waitqueue.
4324 * @mode: which threads
4325 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4326 * @key: is directly passed to the wakeup function
4328 * It may be assumed that this function implies a write memory barrier before
4329 * changing the task state if and only if any tasks are woken up.
4331 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4332 int nr_exclusive, void *key)
4334 unsigned long flags;
4336 spin_lock_irqsave(&q->lock, flags);
4337 __wake_up_common(q, mode, nr_exclusive, 0, key);
4338 spin_unlock_irqrestore(&q->lock, flags);
4340 EXPORT_SYMBOL(__wake_up);
4343 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4345 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4347 __wake_up_common(q, mode, 1, 0, NULL);
4349 EXPORT_SYMBOL_GPL(__wake_up_locked);
4351 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4353 __wake_up_common(q, mode, 1, 0, key);
4355 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4358 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4360 * @mode: which threads
4361 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4362 * @key: opaque value to be passed to wakeup targets
4364 * The sync wakeup differs that the waker knows that it will schedule
4365 * away soon, so while the target thread will be woken up, it will not
4366 * be migrated to another CPU - ie. the two threads are 'synchronized'
4367 * with each other. This can prevent needless bouncing between CPUs.
4369 * On UP it can prevent extra preemption.
4371 * It may be assumed that this function implies a write memory barrier before
4372 * changing the task state if and only if any tasks are woken up.
4374 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4375 int nr_exclusive, void *key)
4377 unsigned long flags;
4378 int wake_flags = WF_SYNC;
4383 if (unlikely(!nr_exclusive))
4386 spin_lock_irqsave(&q->lock, flags);
4387 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4388 spin_unlock_irqrestore(&q->lock, flags);
4390 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4393 * __wake_up_sync - see __wake_up_sync_key()
4395 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4397 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4399 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4402 * complete: - signals a single thread waiting on this completion
4403 * @x: holds the state of this particular completion
4405 * This will wake up a single thread waiting on this completion. Threads will be
4406 * awakened in the same order in which they were queued.
4408 * See also complete_all(), wait_for_completion() and related routines.
4410 * It may be assumed that this function implies a write memory barrier before
4411 * changing the task state if and only if any tasks are woken up.
4413 void complete(struct completion *x)
4415 unsigned long flags;
4417 spin_lock_irqsave(&x->wait.lock, flags);
4419 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4420 spin_unlock_irqrestore(&x->wait.lock, flags);
4422 EXPORT_SYMBOL(complete);
4425 * complete_all: - signals all threads waiting on this completion
4426 * @x: holds the state of this particular completion
4428 * This will wake up all threads waiting on this particular completion event.
4430 * It may be assumed that this function implies a write memory barrier before
4431 * changing the task state if and only if any tasks are woken up.
4433 void complete_all(struct completion *x)
4435 unsigned long flags;
4437 spin_lock_irqsave(&x->wait.lock, flags);
4438 x->done += UINT_MAX/2;
4439 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4440 spin_unlock_irqrestore(&x->wait.lock, flags);
4442 EXPORT_SYMBOL(complete_all);
4444 static inline long __sched
4445 do_wait_for_common(struct completion *x, long timeout, int state)
4448 DECLARE_WAITQUEUE(wait, current);
4450 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4452 if (signal_pending_state(state, current)) {
4453 timeout = -ERESTARTSYS;
4456 __set_current_state(state);
4457 spin_unlock_irq(&x->wait.lock);
4458 timeout = schedule_timeout(timeout);
4459 spin_lock_irq(&x->wait.lock);
4460 } while (!x->done && timeout);
4461 __remove_wait_queue(&x->wait, &wait);
4466 return timeout ?: 1;
4470 wait_for_common(struct completion *x, long timeout, int state)
4474 spin_lock_irq(&x->wait.lock);
4475 timeout = do_wait_for_common(x, timeout, state);
4476 spin_unlock_irq(&x->wait.lock);
4481 * wait_for_completion: - waits for completion of a task
4482 * @x: holds the state of this particular completion
4484 * This waits to be signaled for completion of a specific task. It is NOT
4485 * interruptible and there is no timeout.
4487 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4488 * and interrupt capability. Also see complete().
4490 void __sched wait_for_completion(struct completion *x)
4492 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4494 EXPORT_SYMBOL(wait_for_completion);
4497 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4498 * @x: holds the state of this particular completion
4499 * @timeout: timeout value in jiffies
4501 * This waits for either a completion of a specific task to be signaled or for a
4502 * specified timeout to expire. The timeout is in jiffies. It is not
4505 unsigned long __sched
4506 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4508 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4510 EXPORT_SYMBOL(wait_for_completion_timeout);
4513 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4514 * @x: holds the state of this particular completion
4516 * This waits for completion of a specific task to be signaled. It is
4519 int __sched wait_for_completion_interruptible(struct completion *x)
4521 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4522 if (t == -ERESTARTSYS)
4526 EXPORT_SYMBOL(wait_for_completion_interruptible);
4529 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4530 * @x: holds the state of this particular completion
4531 * @timeout: timeout value in jiffies
4533 * This waits for either a completion of a specific task to be signaled or for a
4534 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4537 wait_for_completion_interruptible_timeout(struct completion *x,
4538 unsigned long timeout)
4540 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4542 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4545 * wait_for_completion_killable: - waits for completion of a task (killable)
4546 * @x: holds the state of this particular completion
4548 * This waits to be signaled for completion of a specific task. It can be
4549 * interrupted by a kill signal.
4551 int __sched wait_for_completion_killable(struct completion *x)
4553 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4554 if (t == -ERESTARTSYS)
4558 EXPORT_SYMBOL(wait_for_completion_killable);
4561 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4562 * @x: holds the state of this particular completion
4563 * @timeout: timeout value in jiffies
4565 * This waits for either a completion of a specific task to be
4566 * signaled or for a specified timeout to expire. It can be
4567 * interrupted by a kill signal. The timeout is in jiffies.
4570 wait_for_completion_killable_timeout(struct completion *x,
4571 unsigned long timeout)
4573 return wait_for_common(x, timeout, TASK_KILLABLE);
4575 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4578 * try_wait_for_completion - try to decrement a completion without blocking
4579 * @x: completion structure
4581 * Returns: 0 if a decrement cannot be done without blocking
4582 * 1 if a decrement succeeded.
4584 * If a completion is being used as a counting completion,
4585 * attempt to decrement the counter without blocking. This
4586 * enables us to avoid waiting if the resource the completion
4587 * is protecting is not available.
4589 bool try_wait_for_completion(struct completion *x)
4591 unsigned long flags;
4594 spin_lock_irqsave(&x->wait.lock, flags);
4599 spin_unlock_irqrestore(&x->wait.lock, flags);
4602 EXPORT_SYMBOL(try_wait_for_completion);
4605 * completion_done - Test to see if a completion has any waiters
4606 * @x: completion structure
4608 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4609 * 1 if there are no waiters.
4612 bool completion_done(struct completion *x)
4614 unsigned long flags;
4617 spin_lock_irqsave(&x->wait.lock, flags);
4620 spin_unlock_irqrestore(&x->wait.lock, flags);
4623 EXPORT_SYMBOL(completion_done);
4626 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4628 unsigned long flags;
4631 init_waitqueue_entry(&wait, current);
4633 __set_current_state(state);
4635 spin_lock_irqsave(&q->lock, flags);
4636 __add_wait_queue(q, &wait);
4637 spin_unlock(&q->lock);
4638 timeout = schedule_timeout(timeout);
4639 spin_lock_irq(&q->lock);
4640 __remove_wait_queue(q, &wait);
4641 spin_unlock_irqrestore(&q->lock, flags);
4646 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4648 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4650 EXPORT_SYMBOL(interruptible_sleep_on);
4653 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4655 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4657 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4659 void __sched sleep_on(wait_queue_head_t *q)
4661 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4663 EXPORT_SYMBOL(sleep_on);
4665 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4667 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4669 EXPORT_SYMBOL(sleep_on_timeout);
4671 #ifdef CONFIG_RT_MUTEXES
4674 * rt_mutex_setprio - set the current priority of a task
4676 * @prio: prio value (kernel-internal form)
4678 * This function changes the 'effective' priority of a task. It does
4679 * not touch ->normal_prio like __setscheduler().
4681 * Used by the rt_mutex code to implement priority inheritance logic.
4683 void rt_mutex_setprio(struct task_struct *p, int prio)
4685 unsigned long flags;
4686 int oldprio, on_rq, running;
4688 const struct sched_class *prev_class;
4690 BUG_ON(prio < 0 || prio > MAX_PRIO);
4692 rq = task_rq_lock(p, &flags);
4694 trace_sched_pi_setprio(p, prio);
4696 prev_class = p->sched_class;
4697 on_rq = p->se.on_rq;
4698 running = task_current(rq, p);
4700 dequeue_task(rq, p, 0);
4702 p->sched_class->put_prev_task(rq, p);
4705 p->sched_class = &rt_sched_class;
4707 p->sched_class = &fair_sched_class;
4712 p->sched_class->set_curr_task(rq);
4714 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4716 check_class_changed(rq, p, prev_class, oldprio);
4717 task_rq_unlock(rq, &flags);
4722 void set_user_nice(struct task_struct *p, long nice)
4724 int old_prio, delta, on_rq;
4725 unsigned long flags;
4728 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4731 * We have to be careful, if called from sys_setpriority(),
4732 * the task might be in the middle of scheduling on another CPU.
4734 rq = task_rq_lock(p, &flags);
4736 * The RT priorities are set via sched_setscheduler(), but we still
4737 * allow the 'normal' nice value to be set - but as expected
4738 * it wont have any effect on scheduling until the task is
4739 * SCHED_FIFO/SCHED_RR:
4741 if (task_has_rt_policy(p)) {
4742 p->static_prio = NICE_TO_PRIO(nice);
4745 on_rq = p->se.on_rq;
4747 dequeue_task(rq, p, 0);
4749 p->static_prio = NICE_TO_PRIO(nice);
4752 p->prio = effective_prio(p);
4753 delta = p->prio - old_prio;
4756 enqueue_task(rq, p, 0);
4758 * If the task increased its priority or is running and
4759 * lowered its priority, then reschedule its CPU:
4761 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4762 resched_task(rq->curr);
4765 task_rq_unlock(rq, &flags);
4767 EXPORT_SYMBOL(set_user_nice);
4770 * can_nice - check if a task can reduce its nice value
4774 int can_nice(const struct task_struct *p, const int nice)
4776 /* convert nice value [19,-20] to rlimit style value [1,40] */
4777 int nice_rlim = 20 - nice;
4779 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4780 capable(CAP_SYS_NICE));
4783 #ifdef __ARCH_WANT_SYS_NICE
4786 * sys_nice - change the priority of the current process.
4787 * @increment: priority increment
4789 * sys_setpriority is a more generic, but much slower function that
4790 * does similar things.
4792 SYSCALL_DEFINE1(nice, int, increment)
4797 * Setpriority might change our priority at the same moment.
4798 * We don't have to worry. Conceptually one call occurs first
4799 * and we have a single winner.
4801 if (increment < -40)
4806 nice = TASK_NICE(current) + increment;
4812 if (increment < 0 && !can_nice(current, nice))
4815 retval = security_task_setnice(current, nice);
4819 set_user_nice(current, nice);
4826 * task_prio - return the priority value of a given task.
4827 * @p: the task in question.
4829 * This is the priority value as seen by users in /proc.
4830 * RT tasks are offset by -200. Normal tasks are centered
4831 * around 0, value goes from -16 to +15.
4833 int task_prio(const struct task_struct *p)
4835 return p->prio - MAX_RT_PRIO;
4839 * task_nice - return the nice value of a given task.
4840 * @p: the task in question.
4842 int task_nice(const struct task_struct *p)
4844 return TASK_NICE(p);
4846 EXPORT_SYMBOL(task_nice);
4849 * idle_cpu - is a given cpu idle currently?
4850 * @cpu: the processor in question.
4852 int idle_cpu(int cpu)
4854 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4858 * idle_task - return the idle task for a given cpu.
4859 * @cpu: the processor in question.
4861 struct task_struct *idle_task(int cpu)
4863 return cpu_rq(cpu)->idle;
4867 * find_process_by_pid - find a process with a matching PID value.
4868 * @pid: the pid in question.
4870 static struct task_struct *find_process_by_pid(pid_t pid)
4872 return pid ? find_task_by_vpid(pid) : current;
4875 /* Actually do priority change: must hold rq lock. */
4877 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4879 BUG_ON(p->se.on_rq);
4882 p->rt_priority = prio;
4883 p->normal_prio = normal_prio(p);
4884 /* we are holding p->pi_lock already */
4885 p->prio = rt_mutex_getprio(p);
4886 if (rt_prio(p->prio))
4887 p->sched_class = &rt_sched_class;
4889 p->sched_class = &fair_sched_class;
4894 * check the target process has a UID that matches the current process's
4896 static bool check_same_owner(struct task_struct *p)
4898 const struct cred *cred = current_cred(), *pcred;
4902 pcred = __task_cred(p);
4903 match = (cred->euid == pcred->euid ||
4904 cred->euid == pcred->uid);
4909 static int __sched_setscheduler(struct task_struct *p, int policy,
4910 const struct sched_param *param, bool user)
4912 int retval, oldprio, oldpolicy = -1, on_rq, running;
4913 unsigned long flags;
4914 const struct sched_class *prev_class;
4918 /* may grab non-irq protected spin_locks */
4919 BUG_ON(in_interrupt());
4921 /* double check policy once rq lock held */
4923 reset_on_fork = p->sched_reset_on_fork;
4924 policy = oldpolicy = p->policy;
4926 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4927 policy &= ~SCHED_RESET_ON_FORK;
4929 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4930 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4931 policy != SCHED_IDLE)
4936 * Valid priorities for SCHED_FIFO and SCHED_RR are
4937 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4938 * SCHED_BATCH and SCHED_IDLE is 0.
4940 if (param->sched_priority < 0 ||
4941 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4942 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4944 if (rt_policy(policy) != (param->sched_priority != 0))
4948 * Allow unprivileged RT tasks to decrease priority:
4950 if (user && !capable(CAP_SYS_NICE)) {
4951 if (rt_policy(policy)) {
4952 unsigned long rlim_rtprio =
4953 task_rlimit(p, RLIMIT_RTPRIO);
4955 /* can't set/change the rt policy */
4956 if (policy != p->policy && !rlim_rtprio)
4959 /* can't increase priority */
4960 if (param->sched_priority > p->rt_priority &&
4961 param->sched_priority > rlim_rtprio)
4966 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4967 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4969 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4970 if (!can_nice(p, TASK_NICE(p)))
4974 /* can't change other user's priorities */
4975 if (!check_same_owner(p))
4978 /* Normal users shall not reset the sched_reset_on_fork flag */
4979 if (p->sched_reset_on_fork && !reset_on_fork)
4984 retval = security_task_setscheduler(p);
4990 * make sure no PI-waiters arrive (or leave) while we are
4991 * changing the priority of the task:
4993 raw_spin_lock_irqsave(&p->pi_lock, flags);
4995 * To be able to change p->policy safely, the apropriate
4996 * runqueue lock must be held.
4998 rq = __task_rq_lock(p);
5001 * Changing the policy of the stop threads its a very bad idea
5003 if (p == rq->stop) {
5004 __task_rq_unlock(rq);
5005 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5009 #ifdef CONFIG_RT_GROUP_SCHED
5012 * Do not allow realtime tasks into groups that have no runtime
5015 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5016 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5017 !task_group_is_autogroup(task_group(p))) {
5018 __task_rq_unlock(rq);
5019 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5025 /* recheck policy now with rq lock held */
5026 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5027 policy = oldpolicy = -1;
5028 __task_rq_unlock(rq);
5029 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5032 on_rq = p->se.on_rq;
5033 running = task_current(rq, p);
5035 deactivate_task(rq, p, 0);
5037 p->sched_class->put_prev_task(rq, p);
5039 p->sched_reset_on_fork = reset_on_fork;
5042 prev_class = p->sched_class;
5043 __setscheduler(rq, p, policy, param->sched_priority);
5046 p->sched_class->set_curr_task(rq);
5048 activate_task(rq, p, 0);
5050 check_class_changed(rq, p, prev_class, oldprio);
5051 __task_rq_unlock(rq);
5052 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5054 rt_mutex_adjust_pi(p);
5060 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5061 * @p: the task in question.
5062 * @policy: new policy.
5063 * @param: structure containing the new RT priority.
5065 * NOTE that the task may be already dead.
5067 int sched_setscheduler(struct task_struct *p, int policy,
5068 const struct sched_param *param)
5070 return __sched_setscheduler(p, policy, param, true);
5072 EXPORT_SYMBOL_GPL(sched_setscheduler);
5075 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5076 * @p: the task in question.
5077 * @policy: new policy.
5078 * @param: structure containing the new RT priority.
5080 * Just like sched_setscheduler, only don't bother checking if the
5081 * current context has permission. For example, this is needed in
5082 * stop_machine(): we create temporary high priority worker threads,
5083 * but our caller might not have that capability.
5085 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5086 const struct sched_param *param)
5088 return __sched_setscheduler(p, policy, param, false);
5092 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5094 struct sched_param lparam;
5095 struct task_struct *p;
5098 if (!param || pid < 0)
5100 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5105 p = find_process_by_pid(pid);
5107 retval = sched_setscheduler(p, policy, &lparam);
5114 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5115 * @pid: the pid in question.
5116 * @policy: new policy.
5117 * @param: structure containing the new RT priority.
5119 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5120 struct sched_param __user *, param)
5122 /* negative values for policy are not valid */
5126 return do_sched_setscheduler(pid, policy, param);
5130 * sys_sched_setparam - set/change the RT priority of a thread
5131 * @pid: the pid in question.
5132 * @param: structure containing the new RT priority.
5134 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5136 return do_sched_setscheduler(pid, -1, param);
5140 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5141 * @pid: the pid in question.
5143 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5145 struct task_struct *p;
5153 p = find_process_by_pid(pid);
5155 retval = security_task_getscheduler(p);
5158 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5165 * sys_sched_getparam - get the RT priority of a thread
5166 * @pid: the pid in question.
5167 * @param: structure containing the RT priority.
5169 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5171 struct sched_param lp;
5172 struct task_struct *p;
5175 if (!param || pid < 0)
5179 p = find_process_by_pid(pid);
5184 retval = security_task_getscheduler(p);
5188 lp.sched_priority = p->rt_priority;
5192 * This one might sleep, we cannot do it with a spinlock held ...
5194 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5203 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5205 cpumask_var_t cpus_allowed, new_mask;
5206 struct task_struct *p;
5212 p = find_process_by_pid(pid);
5219 /* Prevent p going away */
5223 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5227 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5229 goto out_free_cpus_allowed;
5232 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5235 retval = security_task_setscheduler(p);
5239 cpuset_cpus_allowed(p, cpus_allowed);
5240 cpumask_and(new_mask, in_mask, cpus_allowed);
5242 retval = set_cpus_allowed_ptr(p, new_mask);
5245 cpuset_cpus_allowed(p, cpus_allowed);
5246 if (!cpumask_subset(new_mask, cpus_allowed)) {
5248 * We must have raced with a concurrent cpuset
5249 * update. Just reset the cpus_allowed to the
5250 * cpuset's cpus_allowed
5252 cpumask_copy(new_mask, cpus_allowed);
5257 free_cpumask_var(new_mask);
5258 out_free_cpus_allowed:
5259 free_cpumask_var(cpus_allowed);
5266 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5267 struct cpumask *new_mask)
5269 if (len < cpumask_size())
5270 cpumask_clear(new_mask);
5271 else if (len > cpumask_size())
5272 len = cpumask_size();
5274 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5278 * sys_sched_setaffinity - set the cpu affinity of a process
5279 * @pid: pid of the process
5280 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5281 * @user_mask_ptr: user-space pointer to the new cpu mask
5283 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5284 unsigned long __user *, user_mask_ptr)
5286 cpumask_var_t new_mask;
5289 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5292 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5294 retval = sched_setaffinity(pid, new_mask);
5295 free_cpumask_var(new_mask);
5299 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5301 struct task_struct *p;
5302 unsigned long flags;
5310 p = find_process_by_pid(pid);
5314 retval = security_task_getscheduler(p);
5318 rq = task_rq_lock(p, &flags);
5319 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5320 task_rq_unlock(rq, &flags);
5330 * sys_sched_getaffinity - get the cpu affinity of a process
5331 * @pid: pid of the process
5332 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5333 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5335 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5336 unsigned long __user *, user_mask_ptr)
5341 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5343 if (len & (sizeof(unsigned long)-1))
5346 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5349 ret = sched_getaffinity(pid, mask);
5351 size_t retlen = min_t(size_t, len, cpumask_size());
5353 if (copy_to_user(user_mask_ptr, mask, retlen))
5358 free_cpumask_var(mask);
5364 * sys_sched_yield - yield the current processor to other threads.
5366 * This function yields the current CPU to other tasks. If there are no
5367 * other threads running on this CPU then this function will return.
5369 SYSCALL_DEFINE0(sched_yield)
5371 struct rq *rq = this_rq_lock();
5373 schedstat_inc(rq, yld_count);
5374 current->sched_class->yield_task(rq);
5377 * Since we are going to call schedule() anyway, there's
5378 * no need to preempt or enable interrupts:
5380 __release(rq->lock);
5381 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5382 do_raw_spin_unlock(&rq->lock);
5383 preempt_enable_no_resched();
5390 static inline int should_resched(void)
5392 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5395 static void __cond_resched(void)
5397 add_preempt_count(PREEMPT_ACTIVE);
5399 sub_preempt_count(PREEMPT_ACTIVE);
5402 int __sched _cond_resched(void)
5404 if (should_resched()) {
5410 EXPORT_SYMBOL(_cond_resched);
5413 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5414 * call schedule, and on return reacquire the lock.
5416 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5417 * operations here to prevent schedule() from being called twice (once via
5418 * spin_unlock(), once by hand).
5420 int __cond_resched_lock(spinlock_t *lock)
5422 int resched = should_resched();
5425 lockdep_assert_held(lock);
5427 if (spin_needbreak(lock) || resched) {
5438 EXPORT_SYMBOL(__cond_resched_lock);
5440 int __sched __cond_resched_softirq(void)
5442 BUG_ON(!in_softirq());
5444 if (should_resched()) {
5452 EXPORT_SYMBOL(__cond_resched_softirq);
5455 * yield - yield the current processor to other threads.
5457 * This is a shortcut for kernel-space yielding - it marks the
5458 * thread runnable and calls sys_sched_yield().
5460 void __sched yield(void)
5462 set_current_state(TASK_RUNNING);
5465 EXPORT_SYMBOL(yield);
5468 * yield_to - yield the current processor to another thread in
5469 * your thread group, or accelerate that thread toward the
5470 * processor it's on.
5472 * It's the caller's job to ensure that the target task struct
5473 * can't go away on us before we can do any checks.
5475 * Returns true if we indeed boosted the target task.
5477 bool __sched yield_to(struct task_struct *p, bool preempt)
5479 struct task_struct *curr = current;
5480 struct rq *rq, *p_rq;
5481 unsigned long flags;
5484 local_irq_save(flags);
5489 double_rq_lock(rq, p_rq);
5490 while (task_rq(p) != p_rq) {
5491 double_rq_unlock(rq, p_rq);
5495 if (!curr->sched_class->yield_to_task)
5498 if (curr->sched_class != p->sched_class)
5501 if (task_running(p_rq, p) || p->state)
5504 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5506 schedstat_inc(rq, yld_count);
5508 * Make p's CPU reschedule; pick_next_entity takes care of
5511 if (preempt && rq != p_rq)
5512 resched_task(p_rq->curr);
5516 double_rq_unlock(rq, p_rq);
5517 local_irq_restore(flags);
5524 EXPORT_SYMBOL_GPL(yield_to);
5527 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5528 * that process accounting knows that this is a task in IO wait state.
5530 void __sched io_schedule(void)
5532 struct rq *rq = raw_rq();
5534 delayacct_blkio_start();
5535 atomic_inc(&rq->nr_iowait);
5536 current->in_iowait = 1;
5538 current->in_iowait = 0;
5539 atomic_dec(&rq->nr_iowait);
5540 delayacct_blkio_end();
5542 EXPORT_SYMBOL(io_schedule);
5544 long __sched io_schedule_timeout(long timeout)
5546 struct rq *rq = raw_rq();
5549 delayacct_blkio_start();
5550 atomic_inc(&rq->nr_iowait);
5551 current->in_iowait = 1;
5552 ret = schedule_timeout(timeout);
5553 current->in_iowait = 0;
5554 atomic_dec(&rq->nr_iowait);
5555 delayacct_blkio_end();
5560 * sys_sched_get_priority_max - return maximum RT priority.
5561 * @policy: scheduling class.
5563 * this syscall returns the maximum rt_priority that can be used
5564 * by a given scheduling class.
5566 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5573 ret = MAX_USER_RT_PRIO-1;
5585 * sys_sched_get_priority_min - return minimum RT priority.
5586 * @policy: scheduling class.
5588 * this syscall returns the minimum rt_priority that can be used
5589 * by a given scheduling class.
5591 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5609 * sys_sched_rr_get_interval - return the default timeslice of a process.
5610 * @pid: pid of the process.
5611 * @interval: userspace pointer to the timeslice value.
5613 * this syscall writes the default timeslice value of a given process
5614 * into the user-space timespec buffer. A value of '0' means infinity.
5616 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5617 struct timespec __user *, interval)
5619 struct task_struct *p;
5620 unsigned int time_slice;
5621 unsigned long flags;
5631 p = find_process_by_pid(pid);
5635 retval = security_task_getscheduler(p);
5639 rq = task_rq_lock(p, &flags);
5640 time_slice = p->sched_class->get_rr_interval(rq, p);
5641 task_rq_unlock(rq, &flags);
5644 jiffies_to_timespec(time_slice, &t);
5645 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5653 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5655 void sched_show_task(struct task_struct *p)
5657 unsigned long free = 0;
5660 state = p->state ? __ffs(p->state) + 1 : 0;
5661 printk(KERN_INFO "%-15.15s %c", p->comm,
5662 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5663 #if BITS_PER_LONG == 32
5664 if (state == TASK_RUNNING)
5665 printk(KERN_CONT " running ");
5667 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5669 if (state == TASK_RUNNING)
5670 printk(KERN_CONT " running task ");
5672 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5674 #ifdef CONFIG_DEBUG_STACK_USAGE
5675 free = stack_not_used(p);
5677 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5678 task_pid_nr(p), task_pid_nr(p->real_parent),
5679 (unsigned long)task_thread_info(p)->flags);
5681 show_stack(p, NULL);
5684 void show_state_filter(unsigned long state_filter)
5686 struct task_struct *g, *p;
5688 #if BITS_PER_LONG == 32
5690 " task PC stack pid father\n");
5693 " task PC stack pid father\n");
5695 read_lock(&tasklist_lock);
5696 do_each_thread(g, p) {
5698 * reset the NMI-timeout, listing all files on a slow
5699 * console might take alot of time:
5701 touch_nmi_watchdog();
5702 if (!state_filter || (p->state & state_filter))
5704 } while_each_thread(g, p);
5706 touch_all_softlockup_watchdogs();
5708 #ifdef CONFIG_SCHED_DEBUG
5709 sysrq_sched_debug_show();
5711 read_unlock(&tasklist_lock);
5713 * Only show locks if all tasks are dumped:
5716 debug_show_all_locks();
5719 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5721 idle->sched_class = &idle_sched_class;
5725 * init_idle - set up an idle thread for a given CPU
5726 * @idle: task in question
5727 * @cpu: cpu the idle task belongs to
5729 * NOTE: this function does not set the idle thread's NEED_RESCHED
5730 * flag, to make booting more robust.
5732 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5734 struct rq *rq = cpu_rq(cpu);
5735 unsigned long flags;
5737 raw_spin_lock_irqsave(&rq->lock, flags);
5740 idle->state = TASK_RUNNING;
5741 idle->se.exec_start = sched_clock();
5743 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5745 * We're having a chicken and egg problem, even though we are
5746 * holding rq->lock, the cpu isn't yet set to this cpu so the
5747 * lockdep check in task_group() will fail.
5749 * Similar case to sched_fork(). / Alternatively we could
5750 * use task_rq_lock() here and obtain the other rq->lock.
5755 __set_task_cpu(idle, cpu);
5758 rq->curr = rq->idle = idle;
5759 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5762 raw_spin_unlock_irqrestore(&rq->lock, flags);
5764 /* Set the preempt count _outside_ the spinlocks! */
5765 #if defined(CONFIG_PREEMPT)
5766 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5768 task_thread_info(idle)->preempt_count = 0;
5771 * The idle tasks have their own, simple scheduling class:
5773 idle->sched_class = &idle_sched_class;
5774 ftrace_graph_init_idle_task(idle, cpu);
5778 * In a system that switches off the HZ timer nohz_cpu_mask
5779 * indicates which cpus entered this state. This is used
5780 * in the rcu update to wait only for active cpus. For system
5781 * which do not switch off the HZ timer nohz_cpu_mask should
5782 * always be CPU_BITS_NONE.
5784 cpumask_var_t nohz_cpu_mask;
5787 * Increase the granularity value when there are more CPUs,
5788 * because with more CPUs the 'effective latency' as visible
5789 * to users decreases. But the relationship is not linear,
5790 * so pick a second-best guess by going with the log2 of the
5793 * This idea comes from the SD scheduler of Con Kolivas:
5795 static int get_update_sysctl_factor(void)
5797 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5798 unsigned int factor;
5800 switch (sysctl_sched_tunable_scaling) {
5801 case SCHED_TUNABLESCALING_NONE:
5804 case SCHED_TUNABLESCALING_LINEAR:
5807 case SCHED_TUNABLESCALING_LOG:
5809 factor = 1 + ilog2(cpus);
5816 static void update_sysctl(void)
5818 unsigned int factor = get_update_sysctl_factor();
5820 #define SET_SYSCTL(name) \
5821 (sysctl_##name = (factor) * normalized_sysctl_##name)
5822 SET_SYSCTL(sched_min_granularity);
5823 SET_SYSCTL(sched_latency);
5824 SET_SYSCTL(sched_wakeup_granularity);
5828 static inline void sched_init_granularity(void)
5835 * This is how migration works:
5837 * 1) we invoke migration_cpu_stop() on the target CPU using
5839 * 2) stopper starts to run (implicitly forcing the migrated thread
5841 * 3) it checks whether the migrated task is still in the wrong runqueue.
5842 * 4) if it's in the wrong runqueue then the migration thread removes
5843 * it and puts it into the right queue.
5844 * 5) stopper completes and stop_one_cpu() returns and the migration
5849 * Change a given task's CPU affinity. Migrate the thread to a
5850 * proper CPU and schedule it away if the CPU it's executing on
5851 * is removed from the allowed bitmask.
5853 * NOTE: the caller must have a valid reference to the task, the
5854 * task must not exit() & deallocate itself prematurely. The
5855 * call is not atomic; no spinlocks may be held.
5857 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5859 unsigned long flags;
5861 unsigned int dest_cpu;
5865 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5866 * drop the rq->lock and still rely on ->cpus_allowed.
5869 while (task_is_waking(p))
5871 rq = task_rq_lock(p, &flags);
5872 if (task_is_waking(p)) {
5873 task_rq_unlock(rq, &flags);
5877 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5882 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5883 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5888 if (p->sched_class->set_cpus_allowed)
5889 p->sched_class->set_cpus_allowed(p, new_mask);
5891 cpumask_copy(&p->cpus_allowed, new_mask);
5892 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5895 /* Can the task run on the task's current CPU? If so, we're done */
5896 if (cpumask_test_cpu(task_cpu(p), new_mask))
5899 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5900 if (migrate_task(p, rq)) {
5901 struct migration_arg arg = { p, dest_cpu };
5902 /* Need help from migration thread: drop lock and wait. */
5903 task_rq_unlock(rq, &flags);
5904 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5905 tlb_migrate_finish(p->mm);
5909 task_rq_unlock(rq, &flags);
5913 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5916 * Move (not current) task off this cpu, onto dest cpu. We're doing
5917 * this because either it can't run here any more (set_cpus_allowed()
5918 * away from this CPU, or CPU going down), or because we're
5919 * attempting to rebalance this task on exec (sched_exec).
5921 * So we race with normal scheduler movements, but that's OK, as long
5922 * as the task is no longer on this CPU.
5924 * Returns non-zero if task was successfully migrated.
5926 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5928 struct rq *rq_dest, *rq_src;
5931 if (unlikely(!cpu_active(dest_cpu)))
5934 rq_src = cpu_rq(src_cpu);
5935 rq_dest = cpu_rq(dest_cpu);
5937 double_rq_lock(rq_src, rq_dest);
5938 /* Already moved. */
5939 if (task_cpu(p) != src_cpu)
5941 /* Affinity changed (again). */
5942 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5946 * If we're not on a rq, the next wake-up will ensure we're
5950 deactivate_task(rq_src, p, 0);
5951 set_task_cpu(p, dest_cpu);
5952 activate_task(rq_dest, p, 0);
5953 check_preempt_curr(rq_dest, p, 0);
5958 double_rq_unlock(rq_src, rq_dest);
5963 * migration_cpu_stop - this will be executed by a highprio stopper thread
5964 * and performs thread migration by bumping thread off CPU then
5965 * 'pushing' onto another runqueue.
5967 static int migration_cpu_stop(void *data)
5969 struct migration_arg *arg = data;
5972 * The original target cpu might have gone down and we might
5973 * be on another cpu but it doesn't matter.
5975 local_irq_disable();
5976 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5981 #ifdef CONFIG_HOTPLUG_CPU
5984 * Ensures that the idle task is using init_mm right before its cpu goes
5987 void idle_task_exit(void)
5989 struct mm_struct *mm = current->active_mm;
5991 BUG_ON(cpu_online(smp_processor_id()));
5994 switch_mm(mm, &init_mm, current);
5999 * While a dead CPU has no uninterruptible tasks queued at this point,
6000 * it might still have a nonzero ->nr_uninterruptible counter, because
6001 * for performance reasons the counter is not stricly tracking tasks to
6002 * their home CPUs. So we just add the counter to another CPU's counter,
6003 * to keep the global sum constant after CPU-down:
6005 static void migrate_nr_uninterruptible(struct rq *rq_src)
6007 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6009 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6010 rq_src->nr_uninterruptible = 0;
6014 * remove the tasks which were accounted by rq from calc_load_tasks.
6016 static void calc_global_load_remove(struct rq *rq)
6018 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6019 rq->calc_load_active = 0;
6023 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6024 * try_to_wake_up()->select_task_rq().
6026 * Called with rq->lock held even though we'er in stop_machine() and
6027 * there's no concurrency possible, we hold the required locks anyway
6028 * because of lock validation efforts.
6030 static void migrate_tasks(unsigned int dead_cpu)
6032 struct rq *rq = cpu_rq(dead_cpu);
6033 struct task_struct *next, *stop = rq->stop;
6037 * Fudge the rq selection such that the below task selection loop
6038 * doesn't get stuck on the currently eligible stop task.
6040 * We're currently inside stop_machine() and the rq is either stuck
6041 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6042 * either way we should never end up calling schedule() until we're
6049 * There's this thread running, bail when that's the only
6052 if (rq->nr_running == 1)
6055 next = pick_next_task(rq);
6057 next->sched_class->put_prev_task(rq, next);
6059 /* Find suitable destination for @next, with force if needed. */
6060 dest_cpu = select_fallback_rq(dead_cpu, next);
6061 raw_spin_unlock(&rq->lock);
6063 __migrate_task(next, dead_cpu, dest_cpu);
6065 raw_spin_lock(&rq->lock);
6071 #endif /* CONFIG_HOTPLUG_CPU */
6073 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6075 static struct ctl_table sd_ctl_dir[] = {
6077 .procname = "sched_domain",
6083 static struct ctl_table sd_ctl_root[] = {
6085 .procname = "kernel",
6087 .child = sd_ctl_dir,
6092 static struct ctl_table *sd_alloc_ctl_entry(int n)
6094 struct ctl_table *entry =
6095 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6100 static void sd_free_ctl_entry(struct ctl_table **tablep)
6102 struct ctl_table *entry;
6105 * In the intermediate directories, both the child directory and
6106 * procname are dynamically allocated and could fail but the mode
6107 * will always be set. In the lowest directory the names are
6108 * static strings and all have proc handlers.
6110 for (entry = *tablep; entry->mode; entry++) {
6112 sd_free_ctl_entry(&entry->child);
6113 if (entry->proc_handler == NULL)
6114 kfree(entry->procname);
6122 set_table_entry(struct ctl_table *entry,
6123 const char *procname, void *data, int maxlen,
6124 mode_t mode, proc_handler *proc_handler)
6126 entry->procname = procname;
6128 entry->maxlen = maxlen;
6130 entry->proc_handler = proc_handler;
6133 static struct ctl_table *
6134 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6136 struct ctl_table *table = sd_alloc_ctl_entry(13);
6141 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6142 sizeof(long), 0644, proc_doulongvec_minmax);
6143 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6144 sizeof(long), 0644, proc_doulongvec_minmax);
6145 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6146 sizeof(int), 0644, proc_dointvec_minmax);
6147 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6148 sizeof(int), 0644, proc_dointvec_minmax);
6149 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6150 sizeof(int), 0644, proc_dointvec_minmax);
6151 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6152 sizeof(int), 0644, proc_dointvec_minmax);
6153 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6154 sizeof(int), 0644, proc_dointvec_minmax);
6155 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6156 sizeof(int), 0644, proc_dointvec_minmax);
6157 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6158 sizeof(int), 0644, proc_dointvec_minmax);
6159 set_table_entry(&table[9], "cache_nice_tries",
6160 &sd->cache_nice_tries,
6161 sizeof(int), 0644, proc_dointvec_minmax);
6162 set_table_entry(&table[10], "flags", &sd->flags,
6163 sizeof(int), 0644, proc_dointvec_minmax);
6164 set_table_entry(&table[11], "name", sd->name,
6165 CORENAME_MAX_SIZE, 0444, proc_dostring);
6166 /* &table[12] is terminator */
6171 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6173 struct ctl_table *entry, *table;
6174 struct sched_domain *sd;
6175 int domain_num = 0, i;
6178 for_each_domain(cpu, sd)
6180 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6185 for_each_domain(cpu, sd) {
6186 snprintf(buf, 32, "domain%d", i);
6187 entry->procname = kstrdup(buf, GFP_KERNEL);
6189 entry->child = sd_alloc_ctl_domain_table(sd);
6196 static struct ctl_table_header *sd_sysctl_header;
6197 static void register_sched_domain_sysctl(void)
6199 int i, cpu_num = num_possible_cpus();
6200 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6203 WARN_ON(sd_ctl_dir[0].child);
6204 sd_ctl_dir[0].child = entry;
6209 for_each_possible_cpu(i) {
6210 snprintf(buf, 32, "cpu%d", i);
6211 entry->procname = kstrdup(buf, GFP_KERNEL);
6213 entry->child = sd_alloc_ctl_cpu_table(i);
6217 WARN_ON(sd_sysctl_header);
6218 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6221 /* may be called multiple times per register */
6222 static void unregister_sched_domain_sysctl(void)
6224 if (sd_sysctl_header)
6225 unregister_sysctl_table(sd_sysctl_header);
6226 sd_sysctl_header = NULL;
6227 if (sd_ctl_dir[0].child)
6228 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6231 static void register_sched_domain_sysctl(void)
6234 static void unregister_sched_domain_sysctl(void)
6239 static void set_rq_online(struct rq *rq)
6242 const struct sched_class *class;
6244 cpumask_set_cpu(rq->cpu, rq->rd->online);
6247 for_each_class(class) {
6248 if (class->rq_online)
6249 class->rq_online(rq);
6254 static void set_rq_offline(struct rq *rq)
6257 const struct sched_class *class;
6259 for_each_class(class) {
6260 if (class->rq_offline)
6261 class->rq_offline(rq);
6264 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6270 * migration_call - callback that gets triggered when a CPU is added.
6271 * Here we can start up the necessary migration thread for the new CPU.
6273 static int __cpuinit
6274 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6276 int cpu = (long)hcpu;
6277 unsigned long flags;
6278 struct rq *rq = cpu_rq(cpu);
6280 switch (action & ~CPU_TASKS_FROZEN) {
6282 case CPU_UP_PREPARE:
6283 rq->calc_load_update = calc_load_update;
6287 /* Update our root-domain */
6288 raw_spin_lock_irqsave(&rq->lock, flags);
6290 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6294 raw_spin_unlock_irqrestore(&rq->lock, flags);
6297 #ifdef CONFIG_HOTPLUG_CPU
6299 /* Update our root-domain */
6300 raw_spin_lock_irqsave(&rq->lock, flags);
6302 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6306 BUG_ON(rq->nr_running != 1); /* the migration thread */
6307 raw_spin_unlock_irqrestore(&rq->lock, flags);
6309 migrate_nr_uninterruptible(rq);
6310 calc_global_load_remove(rq);
6318 * Register at high priority so that task migration (migrate_all_tasks)
6319 * happens before everything else. This has to be lower priority than
6320 * the notifier in the perf_event subsystem, though.
6322 static struct notifier_block __cpuinitdata migration_notifier = {
6323 .notifier_call = migration_call,
6324 .priority = CPU_PRI_MIGRATION,
6327 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6328 unsigned long action, void *hcpu)
6330 switch (action & ~CPU_TASKS_FROZEN) {
6332 case CPU_DOWN_FAILED:
6333 set_cpu_active((long)hcpu, true);
6340 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6341 unsigned long action, void *hcpu)
6343 switch (action & ~CPU_TASKS_FROZEN) {
6344 case CPU_DOWN_PREPARE:
6345 set_cpu_active((long)hcpu, false);
6352 static int __init migration_init(void)
6354 void *cpu = (void *)(long)smp_processor_id();
6357 /* Initialize migration for the boot CPU */
6358 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6359 BUG_ON(err == NOTIFY_BAD);
6360 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6361 register_cpu_notifier(&migration_notifier);
6363 /* Register cpu active notifiers */
6364 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6365 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6369 early_initcall(migration_init);
6374 #ifdef CONFIG_SCHED_DEBUG
6376 static __read_mostly int sched_domain_debug_enabled;
6378 static int __init sched_domain_debug_setup(char *str)
6380 sched_domain_debug_enabled = 1;
6384 early_param("sched_debug", sched_domain_debug_setup);
6386 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6387 struct cpumask *groupmask)
6389 struct sched_group *group = sd->groups;
6392 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6393 cpumask_clear(groupmask);
6395 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6397 if (!(sd->flags & SD_LOAD_BALANCE)) {
6398 printk("does not load-balance\n");
6400 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6405 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6407 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6408 printk(KERN_ERR "ERROR: domain->span does not contain "
6411 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6412 printk(KERN_ERR "ERROR: domain->groups does not contain"
6416 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6420 printk(KERN_ERR "ERROR: group is NULL\n");
6424 if (!group->cpu_power) {
6425 printk(KERN_CONT "\n");
6426 printk(KERN_ERR "ERROR: domain->cpu_power not "
6431 if (!cpumask_weight(sched_group_cpus(group))) {
6432 printk(KERN_CONT "\n");
6433 printk(KERN_ERR "ERROR: empty group\n");
6437 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6438 printk(KERN_CONT "\n");
6439 printk(KERN_ERR "ERROR: repeated CPUs\n");
6443 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6445 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6447 printk(KERN_CONT " %s", str);
6448 if (group->cpu_power != SCHED_LOAD_SCALE) {
6449 printk(KERN_CONT " (cpu_power = %d)",
6453 group = group->next;
6454 } while (group != sd->groups);
6455 printk(KERN_CONT "\n");
6457 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6458 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6461 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6462 printk(KERN_ERR "ERROR: parent span is not a superset "
6463 "of domain->span\n");
6467 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6469 cpumask_var_t groupmask;
6472 if (!sched_domain_debug_enabled)
6476 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6480 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6482 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6483 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6488 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6495 free_cpumask_var(groupmask);
6497 #else /* !CONFIG_SCHED_DEBUG */
6498 # define sched_domain_debug(sd, cpu) do { } while (0)
6499 #endif /* CONFIG_SCHED_DEBUG */
6501 static int sd_degenerate(struct sched_domain *sd)
6503 if (cpumask_weight(sched_domain_span(sd)) == 1)
6506 /* Following flags need at least 2 groups */
6507 if (sd->flags & (SD_LOAD_BALANCE |
6508 SD_BALANCE_NEWIDLE |
6512 SD_SHARE_PKG_RESOURCES)) {
6513 if (sd->groups != sd->groups->next)
6517 /* Following flags don't use groups */
6518 if (sd->flags & (SD_WAKE_AFFINE))
6525 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6527 unsigned long cflags = sd->flags, pflags = parent->flags;
6529 if (sd_degenerate(parent))
6532 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6535 /* Flags needing groups don't count if only 1 group in parent */
6536 if (parent->groups == parent->groups->next) {
6537 pflags &= ~(SD_LOAD_BALANCE |
6538 SD_BALANCE_NEWIDLE |
6542 SD_SHARE_PKG_RESOURCES);
6543 if (nr_node_ids == 1)
6544 pflags &= ~SD_SERIALIZE;
6546 if (~cflags & pflags)
6552 static void free_rootdomain(struct root_domain *rd)
6554 synchronize_sched();
6556 cpupri_cleanup(&rd->cpupri);
6558 free_cpumask_var(rd->rto_mask);
6559 free_cpumask_var(rd->online);
6560 free_cpumask_var(rd->span);
6564 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6566 struct root_domain *old_rd = NULL;
6567 unsigned long flags;
6569 raw_spin_lock_irqsave(&rq->lock, flags);
6574 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6577 cpumask_clear_cpu(rq->cpu, old_rd->span);
6580 * If we dont want to free the old_rt yet then
6581 * set old_rd to NULL to skip the freeing later
6584 if (!atomic_dec_and_test(&old_rd->refcount))
6588 atomic_inc(&rd->refcount);
6591 cpumask_set_cpu(rq->cpu, rd->span);
6592 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6595 raw_spin_unlock_irqrestore(&rq->lock, flags);
6598 free_rootdomain(old_rd);
6601 static int init_rootdomain(struct root_domain *rd)
6603 memset(rd, 0, sizeof(*rd));
6605 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6607 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6609 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6612 if (cpupri_init(&rd->cpupri) != 0)
6617 free_cpumask_var(rd->rto_mask);
6619 free_cpumask_var(rd->online);
6621 free_cpumask_var(rd->span);
6626 static void init_defrootdomain(void)
6628 init_rootdomain(&def_root_domain);
6630 atomic_set(&def_root_domain.refcount, 1);
6633 static struct root_domain *alloc_rootdomain(void)
6635 struct root_domain *rd;
6637 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6641 if (init_rootdomain(rd) != 0) {
6650 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6651 * hold the hotplug lock.
6654 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6656 struct rq *rq = cpu_rq(cpu);
6657 struct sched_domain *tmp;
6659 for (tmp = sd; tmp; tmp = tmp->parent)
6660 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6662 /* Remove the sched domains which do not contribute to scheduling. */
6663 for (tmp = sd; tmp; ) {
6664 struct sched_domain *parent = tmp->parent;
6668 if (sd_parent_degenerate(tmp, parent)) {
6669 tmp->parent = parent->parent;
6671 parent->parent->child = tmp;
6676 if (sd && sd_degenerate(sd)) {
6682 sched_domain_debug(sd, cpu);
6684 rq_attach_root(rq, rd);
6685 rcu_assign_pointer(rq->sd, sd);
6688 /* cpus with isolated domains */
6689 static cpumask_var_t cpu_isolated_map;
6691 /* Setup the mask of cpus configured for isolated domains */
6692 static int __init isolated_cpu_setup(char *str)
6694 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6695 cpulist_parse(str, cpu_isolated_map);
6699 __setup("isolcpus=", isolated_cpu_setup);
6702 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6703 * to a function which identifies what group(along with sched group) a CPU
6704 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6705 * (due to the fact that we keep track of groups covered with a struct cpumask).
6707 * init_sched_build_groups will build a circular linked list of the groups
6708 * covered by the given span, and will set each group's ->cpumask correctly,
6709 * and ->cpu_power to 0.
6712 init_sched_build_groups(const struct cpumask *span,
6713 const struct cpumask *cpu_map,
6714 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6715 struct sched_group **sg,
6716 struct cpumask *tmpmask),
6717 struct cpumask *covered, struct cpumask *tmpmask)
6719 struct sched_group *first = NULL, *last = NULL;
6722 cpumask_clear(covered);
6724 for_each_cpu(i, span) {
6725 struct sched_group *sg;
6726 int group = group_fn(i, cpu_map, &sg, tmpmask);
6729 if (cpumask_test_cpu(i, covered))
6732 cpumask_clear(sched_group_cpus(sg));
6735 for_each_cpu(j, span) {
6736 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6739 cpumask_set_cpu(j, covered);
6740 cpumask_set_cpu(j, sched_group_cpus(sg));
6751 #define SD_NODES_PER_DOMAIN 16
6756 * find_next_best_node - find the next node to include in a sched_domain
6757 * @node: node whose sched_domain we're building
6758 * @used_nodes: nodes already in the sched_domain
6760 * Find the next node to include in a given scheduling domain. Simply
6761 * finds the closest node not already in the @used_nodes map.
6763 * Should use nodemask_t.
6765 static int find_next_best_node(int node, nodemask_t *used_nodes)
6767 int i, n, val, min_val, best_node = 0;
6771 for (i = 0; i < nr_node_ids; i++) {
6772 /* Start at @node */
6773 n = (node + i) % nr_node_ids;
6775 if (!nr_cpus_node(n))
6778 /* Skip already used nodes */
6779 if (node_isset(n, *used_nodes))
6782 /* Simple min distance search */
6783 val = node_distance(node, n);
6785 if (val < min_val) {
6791 node_set(best_node, *used_nodes);
6796 * sched_domain_node_span - get a cpumask for a node's sched_domain
6797 * @node: node whose cpumask we're constructing
6798 * @span: resulting cpumask
6800 * Given a node, construct a good cpumask for its sched_domain to span. It
6801 * should be one that prevents unnecessary balancing, but also spreads tasks
6804 static void sched_domain_node_span(int node, struct cpumask *span)
6806 nodemask_t used_nodes;
6809 cpumask_clear(span);
6810 nodes_clear(used_nodes);
6812 cpumask_or(span, span, cpumask_of_node(node));
6813 node_set(node, used_nodes);
6815 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6816 int next_node = find_next_best_node(node, &used_nodes);
6818 cpumask_or(span, span, cpumask_of_node(next_node));
6821 #endif /* CONFIG_NUMA */
6823 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6826 * The cpus mask in sched_group and sched_domain hangs off the end.
6828 * ( See the the comments in include/linux/sched.h:struct sched_group
6829 * and struct sched_domain. )
6831 struct static_sched_group {
6832 struct sched_group sg;
6833 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6836 struct static_sched_domain {
6837 struct sched_domain sd;
6838 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6844 cpumask_var_t domainspan;
6845 cpumask_var_t covered;
6846 cpumask_var_t notcovered;
6848 cpumask_var_t nodemask;
6849 cpumask_var_t this_sibling_map;
6850 cpumask_var_t this_core_map;
6851 cpumask_var_t this_book_map;
6852 cpumask_var_t send_covered;
6853 cpumask_var_t tmpmask;
6854 struct sched_group **sched_group_nodes;
6855 struct root_domain *rd;
6859 sa_sched_groups = 0,
6865 sa_this_sibling_map,
6867 sa_sched_group_nodes,
6877 * SMT sched-domains:
6879 #ifdef CONFIG_SCHED_SMT
6880 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6881 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6884 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6885 struct sched_group **sg, struct cpumask *unused)
6888 *sg = &per_cpu(sched_groups, cpu).sg;
6891 #endif /* CONFIG_SCHED_SMT */
6894 * multi-core sched-domains:
6896 #ifdef CONFIG_SCHED_MC
6897 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6898 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6901 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6902 struct sched_group **sg, struct cpumask *mask)
6905 #ifdef CONFIG_SCHED_SMT
6906 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6907 group = cpumask_first(mask);
6912 *sg = &per_cpu(sched_group_core, group).sg;
6915 #endif /* CONFIG_SCHED_MC */
6918 * book sched-domains:
6920 #ifdef CONFIG_SCHED_BOOK
6921 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6922 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6925 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6926 struct sched_group **sg, struct cpumask *mask)
6929 #ifdef CONFIG_SCHED_MC
6930 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6931 group = cpumask_first(mask);
6932 #elif defined(CONFIG_SCHED_SMT)
6933 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6934 group = cpumask_first(mask);
6937 *sg = &per_cpu(sched_group_book, group).sg;
6940 #endif /* CONFIG_SCHED_BOOK */
6942 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6943 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6946 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6947 struct sched_group **sg, struct cpumask *mask)
6950 #ifdef CONFIG_SCHED_BOOK
6951 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6952 group = cpumask_first(mask);
6953 #elif defined(CONFIG_SCHED_MC)
6954 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6955 group = cpumask_first(mask);
6956 #elif defined(CONFIG_SCHED_SMT)
6957 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6958 group = cpumask_first(mask);
6963 *sg = &per_cpu(sched_group_phys, group).sg;
6969 * The init_sched_build_groups can't handle what we want to do with node
6970 * groups, so roll our own. Now each node has its own list of groups which
6971 * gets dynamically allocated.
6973 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6974 static struct sched_group ***sched_group_nodes_bycpu;
6976 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6977 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6979 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6980 struct sched_group **sg,
6981 struct cpumask *nodemask)
6985 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6986 group = cpumask_first(nodemask);
6989 *sg = &per_cpu(sched_group_allnodes, group).sg;
6993 static void init_numa_sched_groups_power(struct sched_group *group_head)
6995 struct sched_group *sg = group_head;
7001 for_each_cpu(j, sched_group_cpus(sg)) {
7002 struct sched_domain *sd;
7004 sd = &per_cpu(phys_domains, j).sd;
7005 if (j != group_first_cpu(sd->groups)) {
7007 * Only add "power" once for each
7013 sg->cpu_power += sd->groups->cpu_power;
7016 } while (sg != group_head);
7019 static int build_numa_sched_groups(struct s_data *d,
7020 const struct cpumask *cpu_map, int num)
7022 struct sched_domain *sd;
7023 struct sched_group *sg, *prev;
7026 cpumask_clear(d->covered);
7027 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
7028 if (cpumask_empty(d->nodemask)) {
7029 d->sched_group_nodes[num] = NULL;
7033 sched_domain_node_span(num, d->domainspan);
7034 cpumask_and(d->domainspan, d->domainspan, cpu_map);
7036 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7039 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
7043 d->sched_group_nodes[num] = sg;
7045 for_each_cpu(j, d->nodemask) {
7046 sd = &per_cpu(node_domains, j).sd;
7051 cpumask_copy(sched_group_cpus(sg), d->nodemask);
7053 cpumask_or(d->covered, d->covered, d->nodemask);
7056 for (j = 0; j < nr_node_ids; j++) {
7057 n = (num + j) % nr_node_ids;
7058 cpumask_complement(d->notcovered, d->covered);
7059 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
7060 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
7061 if (cpumask_empty(d->tmpmask))
7063 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
7064 if (cpumask_empty(d->tmpmask))
7066 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7070 "Can not alloc domain group for node %d\n", j);
7074 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
7075 sg->next = prev->next;
7076 cpumask_or(d->covered, d->covered, d->tmpmask);
7083 #endif /* CONFIG_NUMA */
7086 /* Free memory allocated for various sched_group structures */
7087 static void free_sched_groups(const struct cpumask *cpu_map,
7088 struct cpumask *nodemask)
7092 for_each_cpu(cpu, cpu_map) {
7093 struct sched_group **sched_group_nodes
7094 = sched_group_nodes_bycpu[cpu];
7096 if (!sched_group_nodes)
7099 for (i = 0; i < nr_node_ids; i++) {
7100 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7102 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7103 if (cpumask_empty(nodemask))
7113 if (oldsg != sched_group_nodes[i])
7116 kfree(sched_group_nodes);
7117 sched_group_nodes_bycpu[cpu] = NULL;
7120 #else /* !CONFIG_NUMA */
7121 static void free_sched_groups(const struct cpumask *cpu_map,
7122 struct cpumask *nodemask)
7125 #endif /* CONFIG_NUMA */
7128 * Initialize sched groups cpu_power.
7130 * cpu_power indicates the capacity of sched group, which is used while
7131 * distributing the load between different sched groups in a sched domain.
7132 * Typically cpu_power for all the groups in a sched domain will be same unless
7133 * there are asymmetries in the topology. If there are asymmetries, group
7134 * having more cpu_power will pickup more load compared to the group having
7137 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7139 struct sched_domain *child;
7140 struct sched_group *group;
7144 WARN_ON(!sd || !sd->groups);
7146 if (cpu != group_first_cpu(sd->groups))
7149 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7153 sd->groups->cpu_power = 0;
7156 power = SCHED_LOAD_SCALE;
7157 weight = cpumask_weight(sched_domain_span(sd));
7159 * SMT siblings share the power of a single core.
7160 * Usually multiple threads get a better yield out of
7161 * that one core than a single thread would have,
7162 * reflect that in sd->smt_gain.
7164 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
7165 power *= sd->smt_gain;
7167 power >>= SCHED_LOAD_SHIFT;
7169 sd->groups->cpu_power += power;
7174 * Add cpu_power of each child group to this groups cpu_power.
7176 group = child->groups;
7178 sd->groups->cpu_power += group->cpu_power;
7179 group = group->next;
7180 } while (group != child->groups);
7184 * Initializers for schedule domains
7185 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7188 #ifdef CONFIG_SCHED_DEBUG
7189 # define SD_INIT_NAME(sd, type) sd->name = #type
7191 # define SD_INIT_NAME(sd, type) do { } while (0)
7194 #define SD_INIT(sd, type) sd_init_##type(sd)
7196 #define SD_INIT_FUNC(type) \
7197 static noinline void sd_init_##type(struct sched_domain *sd) \
7199 memset(sd, 0, sizeof(*sd)); \
7200 *sd = SD_##type##_INIT; \
7201 sd->level = SD_LV_##type; \
7202 SD_INIT_NAME(sd, type); \
7207 SD_INIT_FUNC(ALLNODES)
7210 #ifdef CONFIG_SCHED_SMT
7211 SD_INIT_FUNC(SIBLING)
7213 #ifdef CONFIG_SCHED_MC
7216 #ifdef CONFIG_SCHED_BOOK
7220 static int default_relax_domain_level = -1;
7222 static int __init setup_relax_domain_level(char *str)
7226 val = simple_strtoul(str, NULL, 0);
7227 if (val < SD_LV_MAX)
7228 default_relax_domain_level = val;
7232 __setup("relax_domain_level=", setup_relax_domain_level);
7234 static void set_domain_attribute(struct sched_domain *sd,
7235 struct sched_domain_attr *attr)
7239 if (!attr || attr->relax_domain_level < 0) {
7240 if (default_relax_domain_level < 0)
7243 request = default_relax_domain_level;
7245 request = attr->relax_domain_level;
7246 if (request < sd->level) {
7247 /* turn off idle balance on this domain */
7248 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7250 /* turn on idle balance on this domain */
7251 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7255 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7256 const struct cpumask *cpu_map)
7259 case sa_sched_groups:
7260 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7261 d->sched_group_nodes = NULL;
7263 free_rootdomain(d->rd); /* fall through */
7265 free_cpumask_var(d->tmpmask); /* fall through */
7266 case sa_send_covered:
7267 free_cpumask_var(d->send_covered); /* fall through */
7268 case sa_this_book_map:
7269 free_cpumask_var(d->this_book_map); /* fall through */
7270 case sa_this_core_map:
7271 free_cpumask_var(d->this_core_map); /* fall through */
7272 case sa_this_sibling_map:
7273 free_cpumask_var(d->this_sibling_map); /* fall through */
7275 free_cpumask_var(d->nodemask); /* fall through */
7276 case sa_sched_group_nodes:
7278 kfree(d->sched_group_nodes); /* fall through */
7280 free_cpumask_var(d->notcovered); /* fall through */
7282 free_cpumask_var(d->covered); /* fall through */
7284 free_cpumask_var(d->domainspan); /* fall through */
7291 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7292 const struct cpumask *cpu_map)
7295 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7297 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7298 return sa_domainspan;
7299 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7301 /* Allocate the per-node list of sched groups */
7302 d->sched_group_nodes = kcalloc(nr_node_ids,
7303 sizeof(struct sched_group *), GFP_KERNEL);
7304 if (!d->sched_group_nodes) {
7305 printk(KERN_WARNING "Can not alloc sched group node list\n");
7306 return sa_notcovered;
7308 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7310 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7311 return sa_sched_group_nodes;
7312 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7314 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7315 return sa_this_sibling_map;
7316 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7317 return sa_this_core_map;
7318 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7319 return sa_this_book_map;
7320 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7321 return sa_send_covered;
7322 d->rd = alloc_rootdomain();
7324 printk(KERN_WARNING "Cannot alloc root domain\n");
7327 return sa_rootdomain;
7330 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7331 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7333 struct sched_domain *sd = NULL;
7335 struct sched_domain *parent;
7338 if (cpumask_weight(cpu_map) >
7339 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7340 sd = &per_cpu(allnodes_domains, i).sd;
7341 SD_INIT(sd, ALLNODES);
7342 set_domain_attribute(sd, attr);
7343 cpumask_copy(sched_domain_span(sd), cpu_map);
7344 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7349 sd = &per_cpu(node_domains, i).sd;
7351 set_domain_attribute(sd, attr);
7352 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7353 sd->parent = parent;
7356 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7361 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7362 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7363 struct sched_domain *parent, int i)
7365 struct sched_domain *sd;
7366 sd = &per_cpu(phys_domains, i).sd;
7368 set_domain_attribute(sd, attr);
7369 cpumask_copy(sched_domain_span(sd), d->nodemask);
7370 sd->parent = parent;
7373 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7377 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7378 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7379 struct sched_domain *parent, int i)
7381 struct sched_domain *sd = parent;
7382 #ifdef CONFIG_SCHED_BOOK
7383 sd = &per_cpu(book_domains, i).sd;
7385 set_domain_attribute(sd, attr);
7386 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7387 sd->parent = parent;
7389 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7394 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7395 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7396 struct sched_domain *parent, int i)
7398 struct sched_domain *sd = parent;
7399 #ifdef CONFIG_SCHED_MC
7400 sd = &per_cpu(core_domains, i).sd;
7402 set_domain_attribute(sd, attr);
7403 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7404 sd->parent = parent;
7406 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7411 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7412 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7413 struct sched_domain *parent, int i)
7415 struct sched_domain *sd = parent;
7416 #ifdef CONFIG_SCHED_SMT
7417 sd = &per_cpu(cpu_domains, i).sd;
7418 SD_INIT(sd, SIBLING);
7419 set_domain_attribute(sd, attr);
7420 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7421 sd->parent = parent;
7423 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7428 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7429 const struct cpumask *cpu_map, int cpu)
7432 #ifdef CONFIG_SCHED_SMT
7433 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7434 cpumask_and(d->this_sibling_map, cpu_map,
7435 topology_thread_cpumask(cpu));
7436 if (cpu == cpumask_first(d->this_sibling_map))
7437 init_sched_build_groups(d->this_sibling_map, cpu_map,
7439 d->send_covered, d->tmpmask);
7442 #ifdef CONFIG_SCHED_MC
7443 case SD_LV_MC: /* set up multi-core groups */
7444 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7445 if (cpu == cpumask_first(d->this_core_map))
7446 init_sched_build_groups(d->this_core_map, cpu_map,
7448 d->send_covered, d->tmpmask);
7451 #ifdef CONFIG_SCHED_BOOK
7452 case SD_LV_BOOK: /* set up book groups */
7453 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7454 if (cpu == cpumask_first(d->this_book_map))
7455 init_sched_build_groups(d->this_book_map, cpu_map,
7457 d->send_covered, d->tmpmask);
7460 case SD_LV_CPU: /* set up physical groups */
7461 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7462 if (!cpumask_empty(d->nodemask))
7463 init_sched_build_groups(d->nodemask, cpu_map,
7465 d->send_covered, d->tmpmask);
7468 case SD_LV_ALLNODES:
7469 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7470 d->send_covered, d->tmpmask);
7479 * Build sched domains for a given set of cpus and attach the sched domains
7480 * to the individual cpus
7482 static int __build_sched_domains(const struct cpumask *cpu_map,
7483 struct sched_domain_attr *attr)
7485 enum s_alloc alloc_state = sa_none;
7487 struct sched_domain *sd;
7493 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7494 if (alloc_state != sa_rootdomain)
7496 alloc_state = sa_sched_groups;
7499 * Set up domains for cpus specified by the cpu_map.
7501 for_each_cpu(i, cpu_map) {
7502 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7505 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7506 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7507 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7508 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7509 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7512 for_each_cpu(i, cpu_map) {
7513 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7514 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7515 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7518 /* Set up physical groups */
7519 for (i = 0; i < nr_node_ids; i++)
7520 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7523 /* Set up node groups */
7525 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7527 for (i = 0; i < nr_node_ids; i++)
7528 if (build_numa_sched_groups(&d, cpu_map, i))
7532 /* Calculate CPU power for physical packages and nodes */
7533 #ifdef CONFIG_SCHED_SMT
7534 for_each_cpu(i, cpu_map) {
7535 sd = &per_cpu(cpu_domains, i).sd;
7536 init_sched_groups_power(i, sd);
7539 #ifdef CONFIG_SCHED_MC
7540 for_each_cpu(i, cpu_map) {
7541 sd = &per_cpu(core_domains, i).sd;
7542 init_sched_groups_power(i, sd);
7545 #ifdef CONFIG_SCHED_BOOK
7546 for_each_cpu(i, cpu_map) {
7547 sd = &per_cpu(book_domains, i).sd;
7548 init_sched_groups_power(i, sd);
7552 for_each_cpu(i, cpu_map) {
7553 sd = &per_cpu(phys_domains, i).sd;
7554 init_sched_groups_power(i, sd);
7558 for (i = 0; i < nr_node_ids; i++)
7559 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7561 if (d.sd_allnodes) {
7562 struct sched_group *sg;
7564 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7566 init_numa_sched_groups_power(sg);
7570 /* Attach the domains */
7571 for_each_cpu(i, cpu_map) {
7572 #ifdef CONFIG_SCHED_SMT
7573 sd = &per_cpu(cpu_domains, i).sd;
7574 #elif defined(CONFIG_SCHED_MC)
7575 sd = &per_cpu(core_domains, i).sd;
7576 #elif defined(CONFIG_SCHED_BOOK)
7577 sd = &per_cpu(book_domains, i).sd;
7579 sd = &per_cpu(phys_domains, i).sd;
7581 cpu_attach_domain(sd, d.rd, i);
7584 d.sched_group_nodes = NULL; /* don't free this we still need it */
7585 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7589 __free_domain_allocs(&d, alloc_state, cpu_map);
7593 static int build_sched_domains(const struct cpumask *cpu_map)
7595 return __build_sched_domains(cpu_map, NULL);
7598 static cpumask_var_t *doms_cur; /* current sched domains */
7599 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7600 static struct sched_domain_attr *dattr_cur;
7601 /* attribues of custom domains in 'doms_cur' */
7604 * Special case: If a kmalloc of a doms_cur partition (array of
7605 * cpumask) fails, then fallback to a single sched domain,
7606 * as determined by the single cpumask fallback_doms.
7608 static cpumask_var_t fallback_doms;
7611 * arch_update_cpu_topology lets virtualized architectures update the
7612 * cpu core maps. It is supposed to return 1 if the topology changed
7613 * or 0 if it stayed the same.
7615 int __attribute__((weak)) arch_update_cpu_topology(void)
7620 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7623 cpumask_var_t *doms;
7625 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7628 for (i = 0; i < ndoms; i++) {
7629 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7630 free_sched_domains(doms, i);
7637 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7640 for (i = 0; i < ndoms; i++)
7641 free_cpumask_var(doms[i]);
7646 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7647 * For now this just excludes isolated cpus, but could be used to
7648 * exclude other special cases in the future.
7650 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7654 arch_update_cpu_topology();
7656 doms_cur = alloc_sched_domains(ndoms_cur);
7658 doms_cur = &fallback_doms;
7659 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7661 err = build_sched_domains(doms_cur[0]);
7662 register_sched_domain_sysctl();
7667 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7668 struct cpumask *tmpmask)
7670 free_sched_groups(cpu_map, tmpmask);
7674 * Detach sched domains from a group of cpus specified in cpu_map
7675 * These cpus will now be attached to the NULL domain
7677 static void detach_destroy_domains(const struct cpumask *cpu_map)
7679 /* Save because hotplug lock held. */
7680 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7683 for_each_cpu(i, cpu_map)
7684 cpu_attach_domain(NULL, &def_root_domain, i);
7685 synchronize_sched();
7686 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7689 /* handle null as "default" */
7690 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7691 struct sched_domain_attr *new, int idx_new)
7693 struct sched_domain_attr tmp;
7700 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7701 new ? (new + idx_new) : &tmp,
7702 sizeof(struct sched_domain_attr));
7706 * Partition sched domains as specified by the 'ndoms_new'
7707 * cpumasks in the array doms_new[] of cpumasks. This compares
7708 * doms_new[] to the current sched domain partitioning, doms_cur[].
7709 * It destroys each deleted domain and builds each new domain.
7711 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7712 * The masks don't intersect (don't overlap.) We should setup one
7713 * sched domain for each mask. CPUs not in any of the cpumasks will
7714 * not be load balanced. If the same cpumask appears both in the
7715 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7718 * The passed in 'doms_new' should be allocated using
7719 * alloc_sched_domains. This routine takes ownership of it and will
7720 * free_sched_domains it when done with it. If the caller failed the
7721 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7722 * and partition_sched_domains() will fallback to the single partition
7723 * 'fallback_doms', it also forces the domains to be rebuilt.
7725 * If doms_new == NULL it will be replaced with cpu_online_mask.
7726 * ndoms_new == 0 is a special case for destroying existing domains,
7727 * and it will not create the default domain.
7729 * Call with hotplug lock held
7731 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7732 struct sched_domain_attr *dattr_new)
7737 mutex_lock(&sched_domains_mutex);
7739 /* always unregister in case we don't destroy any domains */
7740 unregister_sched_domain_sysctl();
7742 /* Let architecture update cpu core mappings. */
7743 new_topology = arch_update_cpu_topology();
7745 n = doms_new ? ndoms_new : 0;
7747 /* Destroy deleted domains */
7748 for (i = 0; i < ndoms_cur; i++) {
7749 for (j = 0; j < n && !new_topology; j++) {
7750 if (cpumask_equal(doms_cur[i], doms_new[j])
7751 && dattrs_equal(dattr_cur, i, dattr_new, j))
7754 /* no match - a current sched domain not in new doms_new[] */
7755 detach_destroy_domains(doms_cur[i]);
7760 if (doms_new == NULL) {
7762 doms_new = &fallback_doms;
7763 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7764 WARN_ON_ONCE(dattr_new);
7767 /* Build new domains */
7768 for (i = 0; i < ndoms_new; i++) {
7769 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7770 if (cpumask_equal(doms_new[i], doms_cur[j])
7771 && dattrs_equal(dattr_new, i, dattr_cur, j))
7774 /* no match - add a new doms_new */
7775 __build_sched_domains(doms_new[i],
7776 dattr_new ? dattr_new + i : NULL);
7781 /* Remember the new sched domains */
7782 if (doms_cur != &fallback_doms)
7783 free_sched_domains(doms_cur, ndoms_cur);
7784 kfree(dattr_cur); /* kfree(NULL) is safe */
7785 doms_cur = doms_new;
7786 dattr_cur = dattr_new;
7787 ndoms_cur = ndoms_new;
7789 register_sched_domain_sysctl();
7791 mutex_unlock(&sched_domains_mutex);
7794 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7795 static void arch_reinit_sched_domains(void)
7799 /* Destroy domains first to force the rebuild */
7800 partition_sched_domains(0, NULL, NULL);
7802 rebuild_sched_domains();
7806 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7808 unsigned int level = 0;
7810 if (sscanf(buf, "%u", &level) != 1)
7814 * level is always be positive so don't check for
7815 * level < POWERSAVINGS_BALANCE_NONE which is 0
7816 * What happens on 0 or 1 byte write,
7817 * need to check for count as well?
7820 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7824 sched_smt_power_savings = level;
7826 sched_mc_power_savings = level;
7828 arch_reinit_sched_domains();
7833 #ifdef CONFIG_SCHED_MC
7834 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7835 struct sysdev_class_attribute *attr,
7838 return sprintf(page, "%u\n", sched_mc_power_savings);
7840 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7841 struct sysdev_class_attribute *attr,
7842 const char *buf, size_t count)
7844 return sched_power_savings_store(buf, count, 0);
7846 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7847 sched_mc_power_savings_show,
7848 sched_mc_power_savings_store);
7851 #ifdef CONFIG_SCHED_SMT
7852 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7853 struct sysdev_class_attribute *attr,
7856 return sprintf(page, "%u\n", sched_smt_power_savings);
7858 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7859 struct sysdev_class_attribute *attr,
7860 const char *buf, size_t count)
7862 return sched_power_savings_store(buf, count, 1);
7864 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7865 sched_smt_power_savings_show,
7866 sched_smt_power_savings_store);
7869 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7873 #ifdef CONFIG_SCHED_SMT
7875 err = sysfs_create_file(&cls->kset.kobj,
7876 &attr_sched_smt_power_savings.attr);
7878 #ifdef CONFIG_SCHED_MC
7879 if (!err && mc_capable())
7880 err = sysfs_create_file(&cls->kset.kobj,
7881 &attr_sched_mc_power_savings.attr);
7885 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7888 * Update cpusets according to cpu_active mask. If cpusets are
7889 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7890 * around partition_sched_domains().
7892 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7895 switch (action & ~CPU_TASKS_FROZEN) {
7897 case CPU_DOWN_FAILED:
7898 cpuset_update_active_cpus();
7905 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7908 switch (action & ~CPU_TASKS_FROZEN) {
7909 case CPU_DOWN_PREPARE:
7910 cpuset_update_active_cpus();
7917 static int update_runtime(struct notifier_block *nfb,
7918 unsigned long action, void *hcpu)
7920 int cpu = (int)(long)hcpu;
7923 case CPU_DOWN_PREPARE:
7924 case CPU_DOWN_PREPARE_FROZEN:
7925 disable_runtime(cpu_rq(cpu));
7928 case CPU_DOWN_FAILED:
7929 case CPU_DOWN_FAILED_FROZEN:
7931 case CPU_ONLINE_FROZEN:
7932 enable_runtime(cpu_rq(cpu));
7940 void __init sched_init_smp(void)
7942 cpumask_var_t non_isolated_cpus;
7944 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7945 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7947 #if defined(CONFIG_NUMA)
7948 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7950 BUG_ON(sched_group_nodes_bycpu == NULL);
7953 mutex_lock(&sched_domains_mutex);
7954 arch_init_sched_domains(cpu_active_mask);
7955 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7956 if (cpumask_empty(non_isolated_cpus))
7957 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7958 mutex_unlock(&sched_domains_mutex);
7961 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7962 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7964 /* RT runtime code needs to handle some hotplug events */
7965 hotcpu_notifier(update_runtime, 0);
7969 /* Move init over to a non-isolated CPU */
7970 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7972 sched_init_granularity();
7973 free_cpumask_var(non_isolated_cpus);
7975 init_sched_rt_class();
7978 void __init sched_init_smp(void)
7980 sched_init_granularity();
7982 #endif /* CONFIG_SMP */
7984 const_debug unsigned int sysctl_timer_migration = 1;
7986 int in_sched_functions(unsigned long addr)
7988 return in_lock_functions(addr) ||
7989 (addr >= (unsigned long)__sched_text_start
7990 && addr < (unsigned long)__sched_text_end);
7993 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7995 cfs_rq->tasks_timeline = RB_ROOT;
7996 INIT_LIST_HEAD(&cfs_rq->tasks);
7997 #ifdef CONFIG_FAIR_GROUP_SCHED
7999 /* allow initial update_cfs_load() to truncate */
8001 cfs_rq->load_stamp = 1;
8004 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8007 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8009 struct rt_prio_array *array;
8012 array = &rt_rq->active;
8013 for (i = 0; i < MAX_RT_PRIO; i++) {
8014 INIT_LIST_HEAD(array->queue + i);
8015 __clear_bit(i, array->bitmap);
8017 /* delimiter for bitsearch: */
8018 __set_bit(MAX_RT_PRIO, array->bitmap);
8020 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8021 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8023 rt_rq->highest_prio.next = MAX_RT_PRIO;
8027 rt_rq->rt_nr_migratory = 0;
8028 rt_rq->overloaded = 0;
8029 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
8033 rt_rq->rt_throttled = 0;
8034 rt_rq->rt_runtime = 0;
8035 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8037 #ifdef CONFIG_RT_GROUP_SCHED
8038 rt_rq->rt_nr_boosted = 0;
8043 #ifdef CONFIG_FAIR_GROUP_SCHED
8044 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8045 struct sched_entity *se, int cpu,
8046 struct sched_entity *parent)
8048 struct rq *rq = cpu_rq(cpu);
8049 tg->cfs_rq[cpu] = cfs_rq;
8050 init_cfs_rq(cfs_rq, rq);
8054 /* se could be NULL for root_task_group */
8059 se->cfs_rq = &rq->cfs;
8061 se->cfs_rq = parent->my_q;
8064 update_load_set(&se->load, 0);
8065 se->parent = parent;
8069 #ifdef CONFIG_RT_GROUP_SCHED
8070 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8071 struct sched_rt_entity *rt_se, int cpu,
8072 struct sched_rt_entity *parent)
8074 struct rq *rq = cpu_rq(cpu);
8076 tg->rt_rq[cpu] = rt_rq;
8077 init_rt_rq(rt_rq, rq);
8079 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8081 tg->rt_se[cpu] = rt_se;
8086 rt_se->rt_rq = &rq->rt;
8088 rt_se->rt_rq = parent->my_q;
8090 rt_se->my_q = rt_rq;
8091 rt_se->parent = parent;
8092 INIT_LIST_HEAD(&rt_se->run_list);
8096 void __init sched_init(void)
8099 unsigned long alloc_size = 0, ptr;
8101 #ifdef CONFIG_FAIR_GROUP_SCHED
8102 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8104 #ifdef CONFIG_RT_GROUP_SCHED
8105 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8107 #ifdef CONFIG_CPUMASK_OFFSTACK
8108 alloc_size += num_possible_cpus() * cpumask_size();
8111 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8113 #ifdef CONFIG_FAIR_GROUP_SCHED
8114 root_task_group.se = (struct sched_entity **)ptr;
8115 ptr += nr_cpu_ids * sizeof(void **);
8117 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8118 ptr += nr_cpu_ids * sizeof(void **);
8120 #endif /* CONFIG_FAIR_GROUP_SCHED */
8121 #ifdef CONFIG_RT_GROUP_SCHED
8122 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8123 ptr += nr_cpu_ids * sizeof(void **);
8125 root_task_group.rt_rq = (struct rt_rq **)ptr;
8126 ptr += nr_cpu_ids * sizeof(void **);
8128 #endif /* CONFIG_RT_GROUP_SCHED */
8129 #ifdef CONFIG_CPUMASK_OFFSTACK
8130 for_each_possible_cpu(i) {
8131 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8132 ptr += cpumask_size();
8134 #endif /* CONFIG_CPUMASK_OFFSTACK */
8138 init_defrootdomain();
8141 init_rt_bandwidth(&def_rt_bandwidth,
8142 global_rt_period(), global_rt_runtime());
8144 #ifdef CONFIG_RT_GROUP_SCHED
8145 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8146 global_rt_period(), global_rt_runtime());
8147 #endif /* CONFIG_RT_GROUP_SCHED */
8149 #ifdef CONFIG_CGROUP_SCHED
8150 list_add(&root_task_group.list, &task_groups);
8151 INIT_LIST_HEAD(&root_task_group.children);
8152 autogroup_init(&init_task);
8153 #endif /* CONFIG_CGROUP_SCHED */
8155 for_each_possible_cpu(i) {
8159 raw_spin_lock_init(&rq->lock);
8161 rq->calc_load_active = 0;
8162 rq->calc_load_update = jiffies + LOAD_FREQ;
8163 init_cfs_rq(&rq->cfs, rq);
8164 init_rt_rq(&rq->rt, rq);
8165 #ifdef CONFIG_FAIR_GROUP_SCHED
8166 root_task_group.shares = root_task_group_load;
8167 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8169 * How much cpu bandwidth does root_task_group get?
8171 * In case of task-groups formed thr' the cgroup filesystem, it
8172 * gets 100% of the cpu resources in the system. This overall
8173 * system cpu resource is divided among the tasks of
8174 * root_task_group and its child task-groups in a fair manner,
8175 * based on each entity's (task or task-group's) weight
8176 * (se->load.weight).
8178 * In other words, if root_task_group has 10 tasks of weight
8179 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8180 * then A0's share of the cpu resource is:
8182 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8184 * We achieve this by letting root_task_group's tasks sit
8185 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8187 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8188 #endif /* CONFIG_FAIR_GROUP_SCHED */
8190 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8191 #ifdef CONFIG_RT_GROUP_SCHED
8192 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8193 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8196 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8197 rq->cpu_load[j] = 0;
8199 rq->last_load_update_tick = jiffies;
8204 rq->cpu_power = SCHED_LOAD_SCALE;
8205 rq->post_schedule = 0;
8206 rq->active_balance = 0;
8207 rq->next_balance = jiffies;
8212 rq->avg_idle = 2*sysctl_sched_migration_cost;
8213 rq_attach_root(rq, &def_root_domain);
8215 rq->nohz_balance_kick = 0;
8216 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8220 atomic_set(&rq->nr_iowait, 0);
8223 set_load_weight(&init_task);
8225 #ifdef CONFIG_PREEMPT_NOTIFIERS
8226 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8230 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8233 #ifdef CONFIG_RT_MUTEXES
8234 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8238 * The boot idle thread does lazy MMU switching as well:
8240 atomic_inc(&init_mm.mm_count);
8241 enter_lazy_tlb(&init_mm, current);
8244 * Make us the idle thread. Technically, schedule() should not be
8245 * called from this thread, however somewhere below it might be,
8246 * but because we are the idle thread, we just pick up running again
8247 * when this runqueue becomes "idle".
8249 init_idle(current, smp_processor_id());
8251 calc_load_update = jiffies + LOAD_FREQ;
8254 * During early bootup we pretend to be a normal task:
8256 current->sched_class = &fair_sched_class;
8258 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8259 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8262 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8263 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8264 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8265 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8266 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8268 /* May be allocated at isolcpus cmdline parse time */
8269 if (cpu_isolated_map == NULL)
8270 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8273 scheduler_running = 1;
8276 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8277 static inline int preempt_count_equals(int preempt_offset)
8279 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8281 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
8284 void __might_sleep(const char *file, int line, int preempt_offset)
8287 static unsigned long prev_jiffy; /* ratelimiting */
8289 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8290 system_state != SYSTEM_RUNNING || oops_in_progress)
8292 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8294 prev_jiffy = jiffies;
8297 "BUG: sleeping function called from invalid context at %s:%d\n",
8300 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8301 in_atomic(), irqs_disabled(),
8302 current->pid, current->comm);
8304 debug_show_held_locks(current);
8305 if (irqs_disabled())
8306 print_irqtrace_events(current);
8310 EXPORT_SYMBOL(__might_sleep);
8313 #ifdef CONFIG_MAGIC_SYSRQ
8314 static void normalize_task(struct rq *rq, struct task_struct *p)
8316 const struct sched_class *prev_class = p->sched_class;
8317 int old_prio = p->prio;
8320 on_rq = p->se.on_rq;
8322 deactivate_task(rq, p, 0);
8323 __setscheduler(rq, p, SCHED_NORMAL, 0);
8325 activate_task(rq, p, 0);
8326 resched_task(rq->curr);
8329 check_class_changed(rq, p, prev_class, old_prio);
8332 void normalize_rt_tasks(void)
8334 struct task_struct *g, *p;
8335 unsigned long flags;
8338 read_lock_irqsave(&tasklist_lock, flags);
8339 do_each_thread(g, p) {
8341 * Only normalize user tasks:
8346 p->se.exec_start = 0;
8347 #ifdef CONFIG_SCHEDSTATS
8348 p->se.statistics.wait_start = 0;
8349 p->se.statistics.sleep_start = 0;
8350 p->se.statistics.block_start = 0;
8355 * Renice negative nice level userspace
8358 if (TASK_NICE(p) < 0 && p->mm)
8359 set_user_nice(p, 0);
8363 raw_spin_lock(&p->pi_lock);
8364 rq = __task_rq_lock(p);
8366 normalize_task(rq, p);
8368 __task_rq_unlock(rq);
8369 raw_spin_unlock(&p->pi_lock);
8370 } while_each_thread(g, p);
8372 read_unlock_irqrestore(&tasklist_lock, flags);
8375 #endif /* CONFIG_MAGIC_SYSRQ */
8377 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8379 * These functions are only useful for the IA64 MCA handling, or kdb.
8381 * They can only be called when the whole system has been
8382 * stopped - every CPU needs to be quiescent, and no scheduling
8383 * activity can take place. Using them for anything else would
8384 * be a serious bug, and as a result, they aren't even visible
8385 * under any other configuration.
8389 * curr_task - return the current task for a given cpu.
8390 * @cpu: the processor in question.
8392 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8394 struct task_struct *curr_task(int cpu)
8396 return cpu_curr(cpu);
8399 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8403 * set_curr_task - set the current task for a given cpu.
8404 * @cpu: the processor in question.
8405 * @p: the task pointer to set.
8407 * Description: This function must only be used when non-maskable interrupts
8408 * are serviced on a separate stack. It allows the architecture to switch the
8409 * notion of the current task on a cpu in a non-blocking manner. This function
8410 * must be called with all CPU's synchronized, and interrupts disabled, the
8411 * and caller must save the original value of the current task (see
8412 * curr_task() above) and restore that value before reenabling interrupts and
8413 * re-starting the system.
8415 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8417 void set_curr_task(int cpu, struct task_struct *p)
8424 #ifdef CONFIG_FAIR_GROUP_SCHED
8425 static void free_fair_sched_group(struct task_group *tg)
8429 for_each_possible_cpu(i) {
8431 kfree(tg->cfs_rq[i]);
8441 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8443 struct cfs_rq *cfs_rq;
8444 struct sched_entity *se;
8448 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8451 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8455 tg->shares = NICE_0_LOAD;
8457 for_each_possible_cpu(i) {
8460 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8461 GFP_KERNEL, cpu_to_node(i));
8465 se = kzalloc_node(sizeof(struct sched_entity),
8466 GFP_KERNEL, cpu_to_node(i));
8470 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8481 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8483 struct rq *rq = cpu_rq(cpu);
8484 unsigned long flags;
8487 * Only empty task groups can be destroyed; so we can speculatively
8488 * check on_list without danger of it being re-added.
8490 if (!tg->cfs_rq[cpu]->on_list)
8493 raw_spin_lock_irqsave(&rq->lock, flags);
8494 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8495 raw_spin_unlock_irqrestore(&rq->lock, flags);
8497 #else /* !CONFG_FAIR_GROUP_SCHED */
8498 static inline void free_fair_sched_group(struct task_group *tg)
8503 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8508 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8511 #endif /* CONFIG_FAIR_GROUP_SCHED */
8513 #ifdef CONFIG_RT_GROUP_SCHED
8514 static void free_rt_sched_group(struct task_group *tg)
8518 destroy_rt_bandwidth(&tg->rt_bandwidth);
8520 for_each_possible_cpu(i) {
8522 kfree(tg->rt_rq[i]);
8524 kfree(tg->rt_se[i]);
8532 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8534 struct rt_rq *rt_rq;
8535 struct sched_rt_entity *rt_se;
8539 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8542 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8546 init_rt_bandwidth(&tg->rt_bandwidth,
8547 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8549 for_each_possible_cpu(i) {
8552 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8553 GFP_KERNEL, cpu_to_node(i));
8557 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8558 GFP_KERNEL, cpu_to_node(i));
8562 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8572 #else /* !CONFIG_RT_GROUP_SCHED */
8573 static inline void free_rt_sched_group(struct task_group *tg)
8578 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8582 #endif /* CONFIG_RT_GROUP_SCHED */
8584 #ifdef CONFIG_CGROUP_SCHED
8585 static void free_sched_group(struct task_group *tg)
8587 free_fair_sched_group(tg);
8588 free_rt_sched_group(tg);
8593 /* allocate runqueue etc for a new task group */
8594 struct task_group *sched_create_group(struct task_group *parent)
8596 struct task_group *tg;
8597 unsigned long flags;
8599 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8601 return ERR_PTR(-ENOMEM);
8603 if (!alloc_fair_sched_group(tg, parent))
8606 if (!alloc_rt_sched_group(tg, parent))
8609 spin_lock_irqsave(&task_group_lock, flags);
8610 list_add_rcu(&tg->list, &task_groups);
8612 WARN_ON(!parent); /* root should already exist */
8614 tg->parent = parent;
8615 INIT_LIST_HEAD(&tg->children);
8616 list_add_rcu(&tg->siblings, &parent->children);
8617 spin_unlock_irqrestore(&task_group_lock, flags);
8622 free_sched_group(tg);
8623 return ERR_PTR(-ENOMEM);
8626 /* rcu callback to free various structures associated with a task group */
8627 static void free_sched_group_rcu(struct rcu_head *rhp)
8629 /* now it should be safe to free those cfs_rqs */
8630 free_sched_group(container_of(rhp, struct task_group, rcu));
8633 /* Destroy runqueue etc associated with a task group */
8634 void sched_destroy_group(struct task_group *tg)
8636 unsigned long flags;
8639 /* end participation in shares distribution */
8640 for_each_possible_cpu(i)
8641 unregister_fair_sched_group(tg, i);
8643 spin_lock_irqsave(&task_group_lock, flags);
8644 list_del_rcu(&tg->list);
8645 list_del_rcu(&tg->siblings);
8646 spin_unlock_irqrestore(&task_group_lock, flags);
8648 /* wait for possible concurrent references to cfs_rqs complete */
8649 call_rcu(&tg->rcu, free_sched_group_rcu);
8652 /* change task's runqueue when it moves between groups.
8653 * The caller of this function should have put the task in its new group
8654 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8655 * reflect its new group.
8657 void sched_move_task(struct task_struct *tsk)
8660 unsigned long flags;
8663 rq = task_rq_lock(tsk, &flags);
8665 running = task_current(rq, tsk);
8666 on_rq = tsk->se.on_rq;
8669 dequeue_task(rq, tsk, 0);
8670 if (unlikely(running))
8671 tsk->sched_class->put_prev_task(rq, tsk);
8673 #ifdef CONFIG_FAIR_GROUP_SCHED
8674 if (tsk->sched_class->task_move_group)
8675 tsk->sched_class->task_move_group(tsk, on_rq);
8678 set_task_rq(tsk, task_cpu(tsk));
8680 if (unlikely(running))
8681 tsk->sched_class->set_curr_task(rq);
8683 enqueue_task(rq, tsk, 0);
8685 task_rq_unlock(rq, &flags);
8687 #endif /* CONFIG_CGROUP_SCHED */
8689 #ifdef CONFIG_FAIR_GROUP_SCHED
8690 static DEFINE_MUTEX(shares_mutex);
8692 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8695 unsigned long flags;
8698 * We can't change the weight of the root cgroup.
8703 if (shares < MIN_SHARES)
8704 shares = MIN_SHARES;
8705 else if (shares > MAX_SHARES)
8706 shares = MAX_SHARES;
8708 mutex_lock(&shares_mutex);
8709 if (tg->shares == shares)
8712 tg->shares = shares;
8713 for_each_possible_cpu(i) {
8714 struct rq *rq = cpu_rq(i);
8715 struct sched_entity *se;
8718 /* Propagate contribution to hierarchy */
8719 raw_spin_lock_irqsave(&rq->lock, flags);
8720 for_each_sched_entity(se)
8721 update_cfs_shares(group_cfs_rq(se));
8722 raw_spin_unlock_irqrestore(&rq->lock, flags);
8726 mutex_unlock(&shares_mutex);
8730 unsigned long sched_group_shares(struct task_group *tg)
8736 #ifdef CONFIG_RT_GROUP_SCHED
8738 * Ensure that the real time constraints are schedulable.
8740 static DEFINE_MUTEX(rt_constraints_mutex);
8742 static unsigned long to_ratio(u64 period, u64 runtime)
8744 if (runtime == RUNTIME_INF)
8747 return div64_u64(runtime << 20, period);
8750 /* Must be called with tasklist_lock held */
8751 static inline int tg_has_rt_tasks(struct task_group *tg)
8753 struct task_struct *g, *p;
8755 do_each_thread(g, p) {
8756 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8758 } while_each_thread(g, p);
8763 struct rt_schedulable_data {
8764 struct task_group *tg;
8769 static int tg_schedulable(struct task_group *tg, void *data)
8771 struct rt_schedulable_data *d = data;
8772 struct task_group *child;
8773 unsigned long total, sum = 0;
8774 u64 period, runtime;
8776 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8777 runtime = tg->rt_bandwidth.rt_runtime;
8780 period = d->rt_period;
8781 runtime = d->rt_runtime;
8785 * Cannot have more runtime than the period.
8787 if (runtime > period && runtime != RUNTIME_INF)
8791 * Ensure we don't starve existing RT tasks.
8793 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8796 total = to_ratio(period, runtime);
8799 * Nobody can have more than the global setting allows.
8801 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8805 * The sum of our children's runtime should not exceed our own.
8807 list_for_each_entry_rcu(child, &tg->children, siblings) {
8808 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8809 runtime = child->rt_bandwidth.rt_runtime;
8811 if (child == d->tg) {
8812 period = d->rt_period;
8813 runtime = d->rt_runtime;
8816 sum += to_ratio(period, runtime);
8825 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8827 struct rt_schedulable_data data = {
8829 .rt_period = period,
8830 .rt_runtime = runtime,
8833 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8836 static int tg_set_bandwidth(struct task_group *tg,
8837 u64 rt_period, u64 rt_runtime)
8841 mutex_lock(&rt_constraints_mutex);
8842 read_lock(&tasklist_lock);
8843 err = __rt_schedulable(tg, rt_period, rt_runtime);
8847 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8848 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8849 tg->rt_bandwidth.rt_runtime = rt_runtime;
8851 for_each_possible_cpu(i) {
8852 struct rt_rq *rt_rq = tg->rt_rq[i];
8854 raw_spin_lock(&rt_rq->rt_runtime_lock);
8855 rt_rq->rt_runtime = rt_runtime;
8856 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8858 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8860 read_unlock(&tasklist_lock);
8861 mutex_unlock(&rt_constraints_mutex);
8866 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8868 u64 rt_runtime, rt_period;
8870 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8871 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8872 if (rt_runtime_us < 0)
8873 rt_runtime = RUNTIME_INF;
8875 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8878 long sched_group_rt_runtime(struct task_group *tg)
8882 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8885 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8886 do_div(rt_runtime_us, NSEC_PER_USEC);
8887 return rt_runtime_us;
8890 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8892 u64 rt_runtime, rt_period;
8894 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8895 rt_runtime = tg->rt_bandwidth.rt_runtime;
8900 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8903 long sched_group_rt_period(struct task_group *tg)
8907 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8908 do_div(rt_period_us, NSEC_PER_USEC);
8909 return rt_period_us;
8912 static int sched_rt_global_constraints(void)
8914 u64 runtime, period;
8917 if (sysctl_sched_rt_period <= 0)
8920 runtime = global_rt_runtime();
8921 period = global_rt_period();
8924 * Sanity check on the sysctl variables.
8926 if (runtime > period && runtime != RUNTIME_INF)
8929 mutex_lock(&rt_constraints_mutex);
8930 read_lock(&tasklist_lock);
8931 ret = __rt_schedulable(NULL, 0, 0);
8932 read_unlock(&tasklist_lock);
8933 mutex_unlock(&rt_constraints_mutex);
8938 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8940 /* Don't accept realtime tasks when there is no way for them to run */
8941 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8947 #else /* !CONFIG_RT_GROUP_SCHED */
8948 static int sched_rt_global_constraints(void)
8950 unsigned long flags;
8953 if (sysctl_sched_rt_period <= 0)
8957 * There's always some RT tasks in the root group
8958 * -- migration, kstopmachine etc..
8960 if (sysctl_sched_rt_runtime == 0)
8963 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8964 for_each_possible_cpu(i) {
8965 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8967 raw_spin_lock(&rt_rq->rt_runtime_lock);
8968 rt_rq->rt_runtime = global_rt_runtime();
8969 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8971 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8975 #endif /* CONFIG_RT_GROUP_SCHED */
8977 int sched_rt_handler(struct ctl_table *table, int write,
8978 void __user *buffer, size_t *lenp,
8982 int old_period, old_runtime;
8983 static DEFINE_MUTEX(mutex);
8986 old_period = sysctl_sched_rt_period;
8987 old_runtime = sysctl_sched_rt_runtime;
8989 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8991 if (!ret && write) {
8992 ret = sched_rt_global_constraints();
8994 sysctl_sched_rt_period = old_period;
8995 sysctl_sched_rt_runtime = old_runtime;
8997 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8998 def_rt_bandwidth.rt_period =
8999 ns_to_ktime(global_rt_period());
9002 mutex_unlock(&mutex);
9007 #ifdef CONFIG_CGROUP_SCHED
9009 /* return corresponding task_group object of a cgroup */
9010 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9012 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9013 struct task_group, css);
9016 static struct cgroup_subsys_state *
9017 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9019 struct task_group *tg, *parent;
9021 if (!cgrp->parent) {
9022 /* This is early initialization for the top cgroup */
9023 return &root_task_group.css;
9026 parent = cgroup_tg(cgrp->parent);
9027 tg = sched_create_group(parent);
9029 return ERR_PTR(-ENOMEM);
9035 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9037 struct task_group *tg = cgroup_tg(cgrp);
9039 sched_destroy_group(tg);
9043 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9045 #ifdef CONFIG_RT_GROUP_SCHED
9046 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9049 /* We don't support RT-tasks being in separate groups */
9050 if (tsk->sched_class != &fair_sched_class)
9057 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9058 struct task_struct *tsk, bool threadgroup)
9060 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
9064 struct task_struct *c;
9066 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9067 retval = cpu_cgroup_can_attach_task(cgrp, c);
9079 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9080 struct cgroup *old_cont, struct task_struct *tsk,
9083 sched_move_task(tsk);
9085 struct task_struct *c;
9087 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9095 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9096 struct cgroup *old_cgrp, struct task_struct *task)
9099 * cgroup_exit() is called in the copy_process() failure path.
9100 * Ignore this case since the task hasn't ran yet, this avoids
9101 * trying to poke a half freed task state from generic code.
9103 if (!(task->flags & PF_EXITING))
9106 sched_move_task(task);
9109 #ifdef CONFIG_FAIR_GROUP_SCHED
9110 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9113 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9116 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9118 struct task_group *tg = cgroup_tg(cgrp);
9120 return (u64) tg->shares;
9122 #endif /* CONFIG_FAIR_GROUP_SCHED */
9124 #ifdef CONFIG_RT_GROUP_SCHED
9125 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9128 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9131 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9133 return sched_group_rt_runtime(cgroup_tg(cgrp));
9136 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9139 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9142 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9144 return sched_group_rt_period(cgroup_tg(cgrp));
9146 #endif /* CONFIG_RT_GROUP_SCHED */
9148 static struct cftype cpu_files[] = {
9149 #ifdef CONFIG_FAIR_GROUP_SCHED
9152 .read_u64 = cpu_shares_read_u64,
9153 .write_u64 = cpu_shares_write_u64,
9156 #ifdef CONFIG_RT_GROUP_SCHED
9158 .name = "rt_runtime_us",
9159 .read_s64 = cpu_rt_runtime_read,
9160 .write_s64 = cpu_rt_runtime_write,
9163 .name = "rt_period_us",
9164 .read_u64 = cpu_rt_period_read_uint,
9165 .write_u64 = cpu_rt_period_write_uint,
9170 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9172 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9175 struct cgroup_subsys cpu_cgroup_subsys = {
9177 .create = cpu_cgroup_create,
9178 .destroy = cpu_cgroup_destroy,
9179 .can_attach = cpu_cgroup_can_attach,
9180 .attach = cpu_cgroup_attach,
9181 .exit = cpu_cgroup_exit,
9182 .populate = cpu_cgroup_populate,
9183 .subsys_id = cpu_cgroup_subsys_id,
9187 #endif /* CONFIG_CGROUP_SCHED */
9189 #ifdef CONFIG_CGROUP_CPUACCT
9192 * CPU accounting code for task groups.
9194 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9195 * (balbir@in.ibm.com).
9198 /* track cpu usage of a group of tasks and its child groups */
9200 struct cgroup_subsys_state css;
9201 /* cpuusage holds pointer to a u64-type object on every cpu */
9202 u64 __percpu *cpuusage;
9203 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9204 struct cpuacct *parent;
9207 struct cgroup_subsys cpuacct_subsys;
9209 /* return cpu accounting group corresponding to this container */
9210 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9212 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9213 struct cpuacct, css);
9216 /* return cpu accounting group to which this task belongs */
9217 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9219 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9220 struct cpuacct, css);
9223 /* create a new cpu accounting group */
9224 static struct cgroup_subsys_state *cpuacct_create(
9225 struct cgroup_subsys *ss, struct cgroup *cgrp)
9227 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9233 ca->cpuusage = alloc_percpu(u64);
9237 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9238 if (percpu_counter_init(&ca->cpustat[i], 0))
9239 goto out_free_counters;
9242 ca->parent = cgroup_ca(cgrp->parent);
9248 percpu_counter_destroy(&ca->cpustat[i]);
9249 free_percpu(ca->cpuusage);
9253 return ERR_PTR(-ENOMEM);
9256 /* destroy an existing cpu accounting group */
9258 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9260 struct cpuacct *ca = cgroup_ca(cgrp);
9263 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9264 percpu_counter_destroy(&ca->cpustat[i]);
9265 free_percpu(ca->cpuusage);
9269 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9271 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9274 #ifndef CONFIG_64BIT
9276 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9278 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9280 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9288 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9290 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9292 #ifndef CONFIG_64BIT
9294 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9296 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9298 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9304 /* return total cpu usage (in nanoseconds) of a group */
9305 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9307 struct cpuacct *ca = cgroup_ca(cgrp);
9308 u64 totalcpuusage = 0;
9311 for_each_present_cpu(i)
9312 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9314 return totalcpuusage;
9317 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9320 struct cpuacct *ca = cgroup_ca(cgrp);
9329 for_each_present_cpu(i)
9330 cpuacct_cpuusage_write(ca, i, 0);
9336 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9339 struct cpuacct *ca = cgroup_ca(cgroup);
9343 for_each_present_cpu(i) {
9344 percpu = cpuacct_cpuusage_read(ca, i);
9345 seq_printf(m, "%llu ", (unsigned long long) percpu);
9347 seq_printf(m, "\n");
9351 static const char *cpuacct_stat_desc[] = {
9352 [CPUACCT_STAT_USER] = "user",
9353 [CPUACCT_STAT_SYSTEM] = "system",
9356 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9357 struct cgroup_map_cb *cb)
9359 struct cpuacct *ca = cgroup_ca(cgrp);
9362 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9363 s64 val = percpu_counter_read(&ca->cpustat[i]);
9364 val = cputime64_to_clock_t(val);
9365 cb->fill(cb, cpuacct_stat_desc[i], val);
9370 static struct cftype files[] = {
9373 .read_u64 = cpuusage_read,
9374 .write_u64 = cpuusage_write,
9377 .name = "usage_percpu",
9378 .read_seq_string = cpuacct_percpu_seq_read,
9382 .read_map = cpuacct_stats_show,
9386 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9388 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9392 * charge this task's execution time to its accounting group.
9394 * called with rq->lock held.
9396 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9401 if (unlikely(!cpuacct_subsys.active))
9404 cpu = task_cpu(tsk);
9410 for (; ca; ca = ca->parent) {
9411 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9412 *cpuusage += cputime;
9419 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9420 * in cputime_t units. As a result, cpuacct_update_stats calls
9421 * percpu_counter_add with values large enough to always overflow the
9422 * per cpu batch limit causing bad SMP scalability.
9424 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9425 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9426 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9429 #define CPUACCT_BATCH \
9430 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9432 #define CPUACCT_BATCH 0
9436 * Charge the system/user time to the task's accounting group.
9438 static void cpuacct_update_stats(struct task_struct *tsk,
9439 enum cpuacct_stat_index idx, cputime_t val)
9442 int batch = CPUACCT_BATCH;
9444 if (unlikely(!cpuacct_subsys.active))
9451 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9457 struct cgroup_subsys cpuacct_subsys = {
9459 .create = cpuacct_create,
9460 .destroy = cpuacct_destroy,
9461 .populate = cpuacct_populate,
9462 .subsys_id = cpuacct_subsys_id,
9464 #endif /* CONFIG_CGROUP_CPUACCT */