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
664 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
665 * @cpu: the processor in question.
667 * This interface allows printk to be called with the runqueue lock
668 * held and know whether or not it is OK to wake up the klogd.
670 int runqueue_is_locked(int cpu)
672 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
676 * Debugging: various feature bits
679 #define SCHED_FEAT(name, enabled) \
680 __SCHED_FEAT_##name ,
683 #include "sched_features.h"
688 #define SCHED_FEAT(name, enabled) \
689 (1UL << __SCHED_FEAT_##name) * enabled |
691 const_debug unsigned int sysctl_sched_features =
692 #include "sched_features.h"
697 #ifdef CONFIG_SCHED_DEBUG
698 #define SCHED_FEAT(name, enabled) \
701 static __read_mostly char *sched_feat_names[] = {
702 #include "sched_features.h"
708 static int sched_feat_show(struct seq_file *m, void *v)
712 for (i = 0; sched_feat_names[i]; i++) {
713 if (!(sysctl_sched_features & (1UL << i)))
715 seq_printf(m, "%s ", sched_feat_names[i]);
723 sched_feat_write(struct file *filp, const char __user *ubuf,
724 size_t cnt, loff_t *ppos)
734 if (copy_from_user(&buf, ubuf, cnt))
740 if (strncmp(cmp, "NO_", 3) == 0) {
745 for (i = 0; sched_feat_names[i]; i++) {
746 if (strcmp(cmp, sched_feat_names[i]) == 0) {
748 sysctl_sched_features &= ~(1UL << i);
750 sysctl_sched_features |= (1UL << i);
755 if (!sched_feat_names[i])
763 static int sched_feat_open(struct inode *inode, struct file *filp)
765 return single_open(filp, sched_feat_show, NULL);
768 static const struct file_operations sched_feat_fops = {
769 .open = sched_feat_open,
770 .write = sched_feat_write,
773 .release = single_release,
776 static __init int sched_init_debug(void)
778 debugfs_create_file("sched_features", 0644, NULL, NULL,
783 late_initcall(sched_init_debug);
787 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
790 * Number of tasks to iterate in a single balance run.
791 * Limited because this is done with IRQs disabled.
793 const_debug unsigned int sysctl_sched_nr_migrate = 32;
796 * period over which we average the RT time consumption, measured
801 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
804 * period over which we measure -rt task cpu usage in us.
807 unsigned int sysctl_sched_rt_period = 1000000;
809 static __read_mostly int scheduler_running;
812 * part of the period that we allow rt tasks to run in us.
815 int sysctl_sched_rt_runtime = 950000;
817 static inline u64 global_rt_period(void)
819 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
822 static inline u64 global_rt_runtime(void)
824 if (sysctl_sched_rt_runtime < 0)
827 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
830 #ifndef prepare_arch_switch
831 # define prepare_arch_switch(next) do { } while (0)
833 #ifndef finish_arch_switch
834 # define finish_arch_switch(prev) do { } while (0)
837 static inline int task_current(struct rq *rq, struct task_struct *p)
839 return rq->curr == p;
842 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
843 static inline int task_running(struct rq *rq, struct task_struct *p)
845 return task_current(rq, p);
848 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
852 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
854 #ifdef CONFIG_DEBUG_SPINLOCK
855 /* this is a valid case when another task releases the spinlock */
856 rq->lock.owner = current;
859 * If we are tracking spinlock dependencies then we have to
860 * fix up the runqueue lock - which gets 'carried over' from
863 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
865 raw_spin_unlock_irq(&rq->lock);
868 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
869 static inline int task_running(struct rq *rq, struct task_struct *p)
874 return task_current(rq, p);
878 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
882 * We can optimise this out completely for !SMP, because the
883 * SMP rebalancing from interrupt is the only thing that cares
888 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
889 raw_spin_unlock_irq(&rq->lock);
891 raw_spin_unlock(&rq->lock);
895 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
899 * After ->oncpu is cleared, the task can be moved to a different CPU.
900 * We must ensure this doesn't happen until the switch is completely
906 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
910 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
913 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
916 static inline int task_is_waking(struct task_struct *p)
918 return unlikely(p->state == TASK_WAKING);
922 * __task_rq_lock - lock the runqueue a given task resides on.
923 * Must be called interrupts disabled.
925 static inline struct rq *__task_rq_lock(struct task_struct *p)
932 raw_spin_lock(&rq->lock);
933 if (likely(rq == task_rq(p)))
935 raw_spin_unlock(&rq->lock);
940 * task_rq_lock - lock the runqueue a given task resides on and disable
941 * interrupts. Note the ordering: we can safely lookup the task_rq without
942 * explicitly disabling preemption.
944 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
950 local_irq_save(*flags);
952 raw_spin_lock(&rq->lock);
953 if (likely(rq == task_rq(p)))
955 raw_spin_unlock_irqrestore(&rq->lock, *flags);
959 static void __task_rq_unlock(struct rq *rq)
962 raw_spin_unlock(&rq->lock);
965 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
968 raw_spin_unlock_irqrestore(&rq->lock, *flags);
972 * this_rq_lock - lock this runqueue and disable interrupts.
974 static struct rq *this_rq_lock(void)
981 raw_spin_lock(&rq->lock);
986 #ifdef CONFIG_SCHED_HRTICK
988 * Use HR-timers to deliver accurate preemption points.
990 * Its all a bit involved since we cannot program an hrt while holding the
991 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
994 * When we get rescheduled we reprogram the hrtick_timer outside of the
1000 * - enabled by features
1001 * - hrtimer is actually high res
1003 static inline int hrtick_enabled(struct rq *rq)
1005 if (!sched_feat(HRTICK))
1007 if (!cpu_active(cpu_of(rq)))
1009 return hrtimer_is_hres_active(&rq->hrtick_timer);
1012 static void hrtick_clear(struct rq *rq)
1014 if (hrtimer_active(&rq->hrtick_timer))
1015 hrtimer_cancel(&rq->hrtick_timer);
1019 * High-resolution timer tick.
1020 * Runs from hardirq context with interrupts disabled.
1022 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1024 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1026 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1028 raw_spin_lock(&rq->lock);
1029 update_rq_clock(rq);
1030 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1031 raw_spin_unlock(&rq->lock);
1033 return HRTIMER_NORESTART;
1038 * called from hardirq (IPI) context
1040 static void __hrtick_start(void *arg)
1042 struct rq *rq = arg;
1044 raw_spin_lock(&rq->lock);
1045 hrtimer_restart(&rq->hrtick_timer);
1046 rq->hrtick_csd_pending = 0;
1047 raw_spin_unlock(&rq->lock);
1051 * Called to set the hrtick timer state.
1053 * called with rq->lock held and irqs disabled
1055 static void hrtick_start(struct rq *rq, u64 delay)
1057 struct hrtimer *timer = &rq->hrtick_timer;
1058 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1060 hrtimer_set_expires(timer, time);
1062 if (rq == this_rq()) {
1063 hrtimer_restart(timer);
1064 } else if (!rq->hrtick_csd_pending) {
1065 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1066 rq->hrtick_csd_pending = 1;
1071 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1073 int cpu = (int)(long)hcpu;
1076 case CPU_UP_CANCELED:
1077 case CPU_UP_CANCELED_FROZEN:
1078 case CPU_DOWN_PREPARE:
1079 case CPU_DOWN_PREPARE_FROZEN:
1081 case CPU_DEAD_FROZEN:
1082 hrtick_clear(cpu_rq(cpu));
1089 static __init void init_hrtick(void)
1091 hotcpu_notifier(hotplug_hrtick, 0);
1095 * Called to set the hrtick timer state.
1097 * called with rq->lock held and irqs disabled
1099 static void hrtick_start(struct rq *rq, u64 delay)
1101 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1102 HRTIMER_MODE_REL_PINNED, 0);
1105 static inline void init_hrtick(void)
1108 #endif /* CONFIG_SMP */
1110 static void init_rq_hrtick(struct rq *rq)
1113 rq->hrtick_csd_pending = 0;
1115 rq->hrtick_csd.flags = 0;
1116 rq->hrtick_csd.func = __hrtick_start;
1117 rq->hrtick_csd.info = rq;
1120 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1121 rq->hrtick_timer.function = hrtick;
1123 #else /* CONFIG_SCHED_HRTICK */
1124 static inline void hrtick_clear(struct rq *rq)
1128 static inline void init_rq_hrtick(struct rq *rq)
1132 static inline void init_hrtick(void)
1135 #endif /* CONFIG_SCHED_HRTICK */
1138 * resched_task - mark a task 'to be rescheduled now'.
1140 * On UP this means the setting of the need_resched flag, on SMP it
1141 * might also involve a cross-CPU call to trigger the scheduler on
1146 #ifndef tsk_is_polling
1147 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1150 static void resched_task(struct task_struct *p)
1154 assert_raw_spin_locked(&task_rq(p)->lock);
1156 if (test_tsk_need_resched(p))
1159 set_tsk_need_resched(p);
1162 if (cpu == smp_processor_id())
1165 /* NEED_RESCHED must be visible before we test polling */
1167 if (!tsk_is_polling(p))
1168 smp_send_reschedule(cpu);
1171 static void resched_cpu(int cpu)
1173 struct rq *rq = cpu_rq(cpu);
1174 unsigned long flags;
1176 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1178 resched_task(cpu_curr(cpu));
1179 raw_spin_unlock_irqrestore(&rq->lock, flags);
1184 * In the semi idle case, use the nearest busy cpu for migrating timers
1185 * from an idle cpu. This is good for power-savings.
1187 * We don't do similar optimization for completely idle system, as
1188 * selecting an idle cpu will add more delays to the timers than intended
1189 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1191 int get_nohz_timer_target(void)
1193 int cpu = smp_processor_id();
1195 struct sched_domain *sd;
1197 for_each_domain(cpu, sd) {
1198 for_each_cpu(i, sched_domain_span(sd))
1205 * When add_timer_on() enqueues a timer into the timer wheel of an
1206 * idle CPU then this timer might expire before the next timer event
1207 * which is scheduled to wake up that CPU. In case of a completely
1208 * idle system the next event might even be infinite time into the
1209 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1210 * leaves the inner idle loop so the newly added timer is taken into
1211 * account when the CPU goes back to idle and evaluates the timer
1212 * wheel for the next timer event.
1214 void wake_up_idle_cpu(int cpu)
1216 struct rq *rq = cpu_rq(cpu);
1218 if (cpu == smp_processor_id())
1222 * This is safe, as this function is called with the timer
1223 * wheel base lock of (cpu) held. When the CPU is on the way
1224 * to idle and has not yet set rq->curr to idle then it will
1225 * be serialized on the timer wheel base lock and take the new
1226 * timer into account automatically.
1228 if (rq->curr != rq->idle)
1232 * We can set TIF_RESCHED on the idle task of the other CPU
1233 * lockless. The worst case is that the other CPU runs the
1234 * idle task through an additional NOOP schedule()
1236 set_tsk_need_resched(rq->idle);
1238 /* NEED_RESCHED must be visible before we test polling */
1240 if (!tsk_is_polling(rq->idle))
1241 smp_send_reschedule(cpu);
1244 #endif /* CONFIG_NO_HZ */
1246 static u64 sched_avg_period(void)
1248 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1251 static void sched_avg_update(struct rq *rq)
1253 s64 period = sched_avg_period();
1255 while ((s64)(rq->clock - rq->age_stamp) > period) {
1257 * Inline assembly required to prevent the compiler
1258 * optimising this loop into a divmod call.
1259 * See __iter_div_u64_rem() for another example of this.
1261 asm("" : "+rm" (rq->age_stamp));
1262 rq->age_stamp += period;
1267 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1269 rq->rt_avg += rt_delta;
1270 sched_avg_update(rq);
1273 #else /* !CONFIG_SMP */
1274 static void resched_task(struct task_struct *p)
1276 assert_raw_spin_locked(&task_rq(p)->lock);
1277 set_tsk_need_resched(p);
1280 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1284 static void sched_avg_update(struct rq *rq)
1287 #endif /* CONFIG_SMP */
1289 #if BITS_PER_LONG == 32
1290 # define WMULT_CONST (~0UL)
1292 # define WMULT_CONST (1UL << 32)
1295 #define WMULT_SHIFT 32
1298 * Shift right and round:
1300 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1303 * delta *= weight / lw
1305 static unsigned long
1306 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1307 struct load_weight *lw)
1311 if (!lw->inv_weight) {
1312 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1315 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1319 tmp = (u64)delta_exec * weight;
1321 * Check whether we'd overflow the 64-bit multiplication:
1323 if (unlikely(tmp > WMULT_CONST))
1324 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1327 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1329 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1332 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1338 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1344 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1351 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1352 * of tasks with abnormal "nice" values across CPUs the contribution that
1353 * each task makes to its run queue's load is weighted according to its
1354 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1355 * scaled version of the new time slice allocation that they receive on time
1359 #define WEIGHT_IDLEPRIO 3
1360 #define WMULT_IDLEPRIO 1431655765
1363 * Nice levels are multiplicative, with a gentle 10% change for every
1364 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1365 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1366 * that remained on nice 0.
1368 * The "10% effect" is relative and cumulative: from _any_ nice level,
1369 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1370 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1371 * If a task goes up by ~10% and another task goes down by ~10% then
1372 * the relative distance between them is ~25%.)
1374 static const int prio_to_weight[40] = {
1375 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1376 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1377 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1378 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1379 /* 0 */ 1024, 820, 655, 526, 423,
1380 /* 5 */ 335, 272, 215, 172, 137,
1381 /* 10 */ 110, 87, 70, 56, 45,
1382 /* 15 */ 36, 29, 23, 18, 15,
1386 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1388 * In cases where the weight does not change often, we can use the
1389 * precalculated inverse to speed up arithmetics by turning divisions
1390 * into multiplications:
1392 static const u32 prio_to_wmult[40] = {
1393 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1394 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1395 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1396 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1397 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1398 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1399 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1400 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1403 /* Time spent by the tasks of the cpu accounting group executing in ... */
1404 enum cpuacct_stat_index {
1405 CPUACCT_STAT_USER, /* ... user mode */
1406 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1408 CPUACCT_STAT_NSTATS,
1411 #ifdef CONFIG_CGROUP_CPUACCT
1412 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1413 static void cpuacct_update_stats(struct task_struct *tsk,
1414 enum cpuacct_stat_index idx, cputime_t val);
1416 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1417 static inline void cpuacct_update_stats(struct task_struct *tsk,
1418 enum cpuacct_stat_index idx, cputime_t val) {}
1421 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1423 update_load_add(&rq->load, load);
1426 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1428 update_load_sub(&rq->load, load);
1431 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1432 typedef int (*tg_visitor)(struct task_group *, void *);
1435 * Iterate the full tree, calling @down when first entering a node and @up when
1436 * leaving it for the final time.
1438 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1440 struct task_group *parent, *child;
1444 parent = &root_task_group;
1446 ret = (*down)(parent, data);
1449 list_for_each_entry_rcu(child, &parent->children, siblings) {
1456 ret = (*up)(parent, data);
1461 parent = parent->parent;
1470 static int tg_nop(struct task_group *tg, void *data)
1477 /* Used instead of source_load when we know the type == 0 */
1478 static unsigned long weighted_cpuload(const int cpu)
1480 return cpu_rq(cpu)->load.weight;
1484 * Return a low guess at the load of a migration-source cpu weighted
1485 * according to the scheduling class and "nice" value.
1487 * We want to under-estimate the load of migration sources, to
1488 * balance conservatively.
1490 static unsigned long source_load(int cpu, int type)
1492 struct rq *rq = cpu_rq(cpu);
1493 unsigned long total = weighted_cpuload(cpu);
1495 if (type == 0 || !sched_feat(LB_BIAS))
1498 return min(rq->cpu_load[type-1], total);
1502 * Return a high guess at the load of a migration-target cpu weighted
1503 * according to the scheduling class and "nice" value.
1505 static unsigned long target_load(int cpu, int type)
1507 struct rq *rq = cpu_rq(cpu);
1508 unsigned long total = weighted_cpuload(cpu);
1510 if (type == 0 || !sched_feat(LB_BIAS))
1513 return max(rq->cpu_load[type-1], total);
1516 static unsigned long power_of(int cpu)
1518 return cpu_rq(cpu)->cpu_power;
1521 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1523 static unsigned long cpu_avg_load_per_task(int cpu)
1525 struct rq *rq = cpu_rq(cpu);
1526 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1529 rq->avg_load_per_task = rq->load.weight / nr_running;
1531 rq->avg_load_per_task = 0;
1533 return rq->avg_load_per_task;
1536 #ifdef CONFIG_FAIR_GROUP_SCHED
1539 * Compute the cpu's hierarchical load factor for each task group.
1540 * This needs to be done in a top-down fashion because the load of a child
1541 * group is a fraction of its parents load.
1543 static int tg_load_down(struct task_group *tg, void *data)
1546 long cpu = (long)data;
1549 load = cpu_rq(cpu)->load.weight;
1551 load = tg->parent->cfs_rq[cpu]->h_load;
1552 load *= tg->se[cpu]->load.weight;
1553 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1556 tg->cfs_rq[cpu]->h_load = load;
1561 static void update_h_load(long cpu)
1563 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1568 #ifdef CONFIG_PREEMPT
1570 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1573 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1574 * way at the expense of forcing extra atomic operations in all
1575 * invocations. This assures that the double_lock is acquired using the
1576 * same underlying policy as the spinlock_t on this architecture, which
1577 * reduces latency compared to the unfair variant below. However, it
1578 * also adds more overhead and therefore may reduce throughput.
1580 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1581 __releases(this_rq->lock)
1582 __acquires(busiest->lock)
1583 __acquires(this_rq->lock)
1585 raw_spin_unlock(&this_rq->lock);
1586 double_rq_lock(this_rq, busiest);
1593 * Unfair double_lock_balance: Optimizes throughput at the expense of
1594 * latency by eliminating extra atomic operations when the locks are
1595 * already in proper order on entry. This favors lower cpu-ids and will
1596 * grant the double lock to lower cpus over higher ids under contention,
1597 * regardless of entry order into the function.
1599 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1600 __releases(this_rq->lock)
1601 __acquires(busiest->lock)
1602 __acquires(this_rq->lock)
1606 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1607 if (busiest < this_rq) {
1608 raw_spin_unlock(&this_rq->lock);
1609 raw_spin_lock(&busiest->lock);
1610 raw_spin_lock_nested(&this_rq->lock,
1611 SINGLE_DEPTH_NESTING);
1614 raw_spin_lock_nested(&busiest->lock,
1615 SINGLE_DEPTH_NESTING);
1620 #endif /* CONFIG_PREEMPT */
1623 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1625 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1627 if (unlikely(!irqs_disabled())) {
1628 /* printk() doesn't work good under rq->lock */
1629 raw_spin_unlock(&this_rq->lock);
1633 return _double_lock_balance(this_rq, busiest);
1636 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1637 __releases(busiest->lock)
1639 raw_spin_unlock(&busiest->lock);
1640 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1644 * double_rq_lock - safely lock two runqueues
1646 * Note this does not disable interrupts like task_rq_lock,
1647 * you need to do so manually before calling.
1649 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1650 __acquires(rq1->lock)
1651 __acquires(rq2->lock)
1653 BUG_ON(!irqs_disabled());
1655 raw_spin_lock(&rq1->lock);
1656 __acquire(rq2->lock); /* Fake it out ;) */
1659 raw_spin_lock(&rq1->lock);
1660 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1662 raw_spin_lock(&rq2->lock);
1663 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1669 * double_rq_unlock - safely unlock two runqueues
1671 * Note this does not restore interrupts like task_rq_unlock,
1672 * you need to do so manually after calling.
1674 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1675 __releases(rq1->lock)
1676 __releases(rq2->lock)
1678 raw_spin_unlock(&rq1->lock);
1680 raw_spin_unlock(&rq2->lock);
1682 __release(rq2->lock);
1685 #else /* CONFIG_SMP */
1688 * double_rq_lock - safely lock two runqueues
1690 * Note this does not disable interrupts like task_rq_lock,
1691 * you need to do so manually before calling.
1693 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1694 __acquires(rq1->lock)
1695 __acquires(rq2->lock)
1697 BUG_ON(!irqs_disabled());
1699 raw_spin_lock(&rq1->lock);
1700 __acquire(rq2->lock); /* Fake it out ;) */
1704 * double_rq_unlock - safely unlock two runqueues
1706 * Note this does not restore interrupts like task_rq_unlock,
1707 * you need to do so manually after calling.
1709 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1710 __releases(rq1->lock)
1711 __releases(rq2->lock)
1714 raw_spin_unlock(&rq1->lock);
1715 __release(rq2->lock);
1720 static void calc_load_account_idle(struct rq *this_rq);
1721 static void update_sysctl(void);
1722 static int get_update_sysctl_factor(void);
1723 static void update_cpu_load(struct rq *this_rq);
1725 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1727 set_task_rq(p, cpu);
1730 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1731 * successfuly executed on another CPU. We must ensure that updates of
1732 * per-task data have been completed by this moment.
1735 task_thread_info(p)->cpu = cpu;
1739 static const struct sched_class rt_sched_class;
1741 #define sched_class_highest (&stop_sched_class)
1742 #define for_each_class(class) \
1743 for (class = sched_class_highest; class; class = class->next)
1745 #include "sched_stats.h"
1747 static void inc_nr_running(struct rq *rq)
1752 static void dec_nr_running(struct rq *rq)
1757 static void set_load_weight(struct task_struct *p)
1760 * SCHED_IDLE tasks get minimal weight:
1762 if (p->policy == SCHED_IDLE) {
1763 p->se.load.weight = WEIGHT_IDLEPRIO;
1764 p->se.load.inv_weight = WMULT_IDLEPRIO;
1768 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1769 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1772 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1774 update_rq_clock(rq);
1775 sched_info_queued(p);
1776 p->sched_class->enqueue_task(rq, p, flags);
1780 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1782 update_rq_clock(rq);
1783 sched_info_dequeued(p);
1784 p->sched_class->dequeue_task(rq, p, flags);
1789 * activate_task - move a task to the runqueue.
1791 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1793 if (task_contributes_to_load(p))
1794 rq->nr_uninterruptible--;
1796 enqueue_task(rq, p, flags);
1801 * deactivate_task - remove a task from the runqueue.
1803 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1805 if (task_contributes_to_load(p))
1806 rq->nr_uninterruptible++;
1808 dequeue_task(rq, p, flags);
1812 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1815 * There are no locks covering percpu hardirq/softirq time.
1816 * They are only modified in account_system_vtime, on corresponding CPU
1817 * with interrupts disabled. So, writes are safe.
1818 * They are read and saved off onto struct rq in update_rq_clock().
1819 * This may result in other CPU reading this CPU's irq time and can
1820 * race with irq/account_system_vtime on this CPU. We would either get old
1821 * or new value with a side effect of accounting a slice of irq time to wrong
1822 * task when irq is in progress while we read rq->clock. That is a worthy
1823 * compromise in place of having locks on each irq in account_system_time.
1825 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1826 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1828 static DEFINE_PER_CPU(u64, irq_start_time);
1829 static int sched_clock_irqtime;
1831 void enable_sched_clock_irqtime(void)
1833 sched_clock_irqtime = 1;
1836 void disable_sched_clock_irqtime(void)
1838 sched_clock_irqtime = 0;
1841 #ifndef CONFIG_64BIT
1842 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1844 static inline void irq_time_write_begin(void)
1846 __this_cpu_inc(irq_time_seq.sequence);
1850 static inline void irq_time_write_end(void)
1853 __this_cpu_inc(irq_time_seq.sequence);
1856 static inline u64 irq_time_read(int cpu)
1862 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1863 irq_time = per_cpu(cpu_softirq_time, cpu) +
1864 per_cpu(cpu_hardirq_time, cpu);
1865 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1869 #else /* CONFIG_64BIT */
1870 static inline void irq_time_write_begin(void)
1874 static inline void irq_time_write_end(void)
1878 static inline u64 irq_time_read(int cpu)
1880 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1882 #endif /* CONFIG_64BIT */
1885 * Called before incrementing preempt_count on {soft,}irq_enter
1886 * and before decrementing preempt_count on {soft,}irq_exit.
1888 void account_system_vtime(struct task_struct *curr)
1890 unsigned long flags;
1894 if (!sched_clock_irqtime)
1897 local_irq_save(flags);
1899 cpu = smp_processor_id();
1900 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1901 __this_cpu_add(irq_start_time, delta);
1903 irq_time_write_begin();
1905 * We do not account for softirq time from ksoftirqd here.
1906 * We want to continue accounting softirq time to ksoftirqd thread
1907 * in that case, so as not to confuse scheduler with a special task
1908 * that do not consume any time, but still wants to run.
1910 if (hardirq_count())
1911 __this_cpu_add(cpu_hardirq_time, delta);
1912 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1913 __this_cpu_add(cpu_softirq_time, delta);
1915 irq_time_write_end();
1916 local_irq_restore(flags);
1918 EXPORT_SYMBOL_GPL(account_system_vtime);
1920 static void update_rq_clock_task(struct rq *rq, s64 delta)
1924 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1927 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1928 * this case when a previous update_rq_clock() happened inside a
1929 * {soft,}irq region.
1931 * When this happens, we stop ->clock_task and only update the
1932 * prev_irq_time stamp to account for the part that fit, so that a next
1933 * update will consume the rest. This ensures ->clock_task is
1936 * It does however cause some slight miss-attribution of {soft,}irq
1937 * time, a more accurate solution would be to update the irq_time using
1938 * the current rq->clock timestamp, except that would require using
1941 if (irq_delta > delta)
1944 rq->prev_irq_time += irq_delta;
1946 rq->clock_task += delta;
1948 if (irq_delta && sched_feat(NONIRQ_POWER))
1949 sched_rt_avg_update(rq, irq_delta);
1952 static int irqtime_account_hi_update(void)
1954 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1955 unsigned long flags;
1959 local_irq_save(flags);
1960 latest_ns = this_cpu_read(cpu_hardirq_time);
1961 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
1963 local_irq_restore(flags);
1967 static int irqtime_account_si_update(void)
1969 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1970 unsigned long flags;
1974 local_irq_save(flags);
1975 latest_ns = this_cpu_read(cpu_softirq_time);
1976 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
1978 local_irq_restore(flags);
1982 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
1984 #define sched_clock_irqtime (0)
1986 static void update_rq_clock_task(struct rq *rq, s64 delta)
1988 rq->clock_task += delta;
1991 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1993 #include "sched_idletask.c"
1994 #include "sched_fair.c"
1995 #include "sched_rt.c"
1996 #include "sched_autogroup.c"
1997 #include "sched_stoptask.c"
1998 #ifdef CONFIG_SCHED_DEBUG
1999 # include "sched_debug.c"
2002 void sched_set_stop_task(int cpu, struct task_struct *stop)
2004 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2005 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2009 * Make it appear like a SCHED_FIFO task, its something
2010 * userspace knows about and won't get confused about.
2012 * Also, it will make PI more or less work without too
2013 * much confusion -- but then, stop work should not
2014 * rely on PI working anyway.
2016 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2018 stop->sched_class = &stop_sched_class;
2021 cpu_rq(cpu)->stop = stop;
2025 * Reset it back to a normal scheduling class so that
2026 * it can die in pieces.
2028 old_stop->sched_class = &rt_sched_class;
2033 * __normal_prio - return the priority that is based on the static prio
2035 static inline int __normal_prio(struct task_struct *p)
2037 return p->static_prio;
2041 * Calculate the expected normal priority: i.e. priority
2042 * without taking RT-inheritance into account. Might be
2043 * boosted by interactivity modifiers. Changes upon fork,
2044 * setprio syscalls, and whenever the interactivity
2045 * estimator recalculates.
2047 static inline int normal_prio(struct task_struct *p)
2051 if (task_has_rt_policy(p))
2052 prio = MAX_RT_PRIO-1 - p->rt_priority;
2054 prio = __normal_prio(p);
2059 * Calculate the current priority, i.e. the priority
2060 * taken into account by the scheduler. This value might
2061 * be boosted by RT tasks, or might be boosted by
2062 * interactivity modifiers. Will be RT if the task got
2063 * RT-boosted. If not then it returns p->normal_prio.
2065 static int effective_prio(struct task_struct *p)
2067 p->normal_prio = normal_prio(p);
2069 * If we are RT tasks or we were boosted to RT priority,
2070 * keep the priority unchanged. Otherwise, update priority
2071 * to the normal priority:
2073 if (!rt_prio(p->prio))
2074 return p->normal_prio;
2079 * task_curr - is this task currently executing on a CPU?
2080 * @p: the task in question.
2082 inline int task_curr(const struct task_struct *p)
2084 return cpu_curr(task_cpu(p)) == p;
2087 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2088 const struct sched_class *prev_class,
2091 if (prev_class != p->sched_class) {
2092 if (prev_class->switched_from)
2093 prev_class->switched_from(rq, p);
2094 p->sched_class->switched_to(rq, p);
2095 } else if (oldprio != p->prio)
2096 p->sched_class->prio_changed(rq, p, oldprio);
2099 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2101 const struct sched_class *class;
2103 if (p->sched_class == rq->curr->sched_class) {
2104 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2106 for_each_class(class) {
2107 if (class == rq->curr->sched_class)
2109 if (class == p->sched_class) {
2110 resched_task(rq->curr);
2117 * A queue event has occurred, and we're going to schedule. In
2118 * this case, we can save a useless back to back clock update.
2120 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
2121 rq->skip_clock_update = 1;
2126 * Is this task likely cache-hot:
2129 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2133 if (p->sched_class != &fair_sched_class)
2136 if (unlikely(p->policy == SCHED_IDLE))
2140 * Buddy candidates are cache hot:
2142 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2143 (&p->se == cfs_rq_of(&p->se)->next ||
2144 &p->se == cfs_rq_of(&p->se)->last))
2147 if (sysctl_sched_migration_cost == -1)
2149 if (sysctl_sched_migration_cost == 0)
2152 delta = now - p->se.exec_start;
2154 return delta < (s64)sysctl_sched_migration_cost;
2157 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2159 #ifdef CONFIG_SCHED_DEBUG
2161 * We should never call set_task_cpu() on a blocked task,
2162 * ttwu() will sort out the placement.
2164 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2165 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2168 trace_sched_migrate_task(p, new_cpu);
2170 if (task_cpu(p) != new_cpu) {
2171 p->se.nr_migrations++;
2172 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2175 __set_task_cpu(p, new_cpu);
2178 struct migration_arg {
2179 struct task_struct *task;
2183 static int migration_cpu_stop(void *data);
2186 * The task's runqueue lock must be held.
2187 * Returns true if you have to wait for migration thread.
2189 static bool migrate_task(struct task_struct *p, struct rq *rq)
2192 * If the task is not on a runqueue (and not running), then
2193 * the next wake-up will properly place the task.
2195 return p->se.on_rq || task_running(rq, p);
2199 * wait_task_inactive - wait for a thread to unschedule.
2201 * If @match_state is nonzero, it's the @p->state value just checked and
2202 * not expected to change. If it changes, i.e. @p might have woken up,
2203 * then return zero. When we succeed in waiting for @p to be off its CPU,
2204 * we return a positive number (its total switch count). If a second call
2205 * a short while later returns the same number, the caller can be sure that
2206 * @p has remained unscheduled the whole time.
2208 * The caller must ensure that the task *will* unschedule sometime soon,
2209 * else this function might spin for a *long* time. This function can't
2210 * be called with interrupts off, or it may introduce deadlock with
2211 * smp_call_function() if an IPI is sent by the same process we are
2212 * waiting to become inactive.
2214 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2216 unsigned long flags;
2223 * We do the initial early heuristics without holding
2224 * any task-queue locks at all. We'll only try to get
2225 * the runqueue lock when things look like they will
2231 * If the task is actively running on another CPU
2232 * still, just relax and busy-wait without holding
2235 * NOTE! Since we don't hold any locks, it's not
2236 * even sure that "rq" stays as the right runqueue!
2237 * But we don't care, since "task_running()" will
2238 * return false if the runqueue has changed and p
2239 * is actually now running somewhere else!
2241 while (task_running(rq, p)) {
2242 if (match_state && unlikely(p->state != match_state))
2248 * Ok, time to look more closely! We need the rq
2249 * lock now, to be *sure*. If we're wrong, we'll
2250 * just go back and repeat.
2252 rq = task_rq_lock(p, &flags);
2253 trace_sched_wait_task(p);
2254 running = task_running(rq, p);
2255 on_rq = p->se.on_rq;
2257 if (!match_state || p->state == match_state)
2258 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2259 task_rq_unlock(rq, &flags);
2262 * If it changed from the expected state, bail out now.
2264 if (unlikely(!ncsw))
2268 * Was it really running after all now that we
2269 * checked with the proper locks actually held?
2271 * Oops. Go back and try again..
2273 if (unlikely(running)) {
2279 * It's not enough that it's not actively running,
2280 * it must be off the runqueue _entirely_, and not
2283 * So if it was still runnable (but just not actively
2284 * running right now), it's preempted, and we should
2285 * yield - it could be a while.
2287 if (unlikely(on_rq)) {
2288 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2290 set_current_state(TASK_UNINTERRUPTIBLE);
2291 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2296 * Ahh, all good. It wasn't running, and it wasn't
2297 * runnable, which means that it will never become
2298 * running in the future either. We're all done!
2307 * kick_process - kick a running thread to enter/exit the kernel
2308 * @p: the to-be-kicked thread
2310 * Cause a process which is running on another CPU to enter
2311 * kernel-mode, without any delay. (to get signals handled.)
2313 * NOTE: this function doesnt have to take the runqueue lock,
2314 * because all it wants to ensure is that the remote task enters
2315 * the kernel. If the IPI races and the task has been migrated
2316 * to another CPU then no harm is done and the purpose has been
2319 void kick_process(struct task_struct *p)
2325 if ((cpu != smp_processor_id()) && task_curr(p))
2326 smp_send_reschedule(cpu);
2329 EXPORT_SYMBOL_GPL(kick_process);
2330 #endif /* CONFIG_SMP */
2334 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2336 static int select_fallback_rq(int cpu, struct task_struct *p)
2339 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2341 /* Look for allowed, online CPU in same node. */
2342 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2343 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2346 /* Any allowed, online CPU? */
2347 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2348 if (dest_cpu < nr_cpu_ids)
2351 /* No more Mr. Nice Guy. */
2352 dest_cpu = cpuset_cpus_allowed_fallback(p);
2354 * Don't tell them about moving exiting tasks or
2355 * kernel threads (both mm NULL), since they never
2358 if (p->mm && printk_ratelimit()) {
2359 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2360 task_pid_nr(p), p->comm, cpu);
2367 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2370 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2372 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2375 * In order not to call set_task_cpu() on a blocking task we need
2376 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2379 * Since this is common to all placement strategies, this lives here.
2381 * [ this allows ->select_task() to simply return task_cpu(p) and
2382 * not worry about this generic constraint ]
2384 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2386 cpu = select_fallback_rq(task_cpu(p), p);
2391 static void update_avg(u64 *avg, u64 sample)
2393 s64 diff = sample - *avg;
2398 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2399 bool is_sync, bool is_migrate, bool is_local,
2400 unsigned long en_flags)
2402 schedstat_inc(p, se.statistics.nr_wakeups);
2404 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2406 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2408 schedstat_inc(p, se.statistics.nr_wakeups_local);
2410 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2412 activate_task(rq, p, en_flags);
2415 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2416 int wake_flags, bool success)
2418 trace_sched_wakeup(p, success);
2419 check_preempt_curr(rq, p, wake_flags);
2421 p->state = TASK_RUNNING;
2423 if (p->sched_class->task_woken)
2424 p->sched_class->task_woken(rq, p);
2426 if (unlikely(rq->idle_stamp)) {
2427 u64 delta = rq->clock - rq->idle_stamp;
2428 u64 max = 2*sysctl_sched_migration_cost;
2433 update_avg(&rq->avg_idle, delta);
2437 /* if a worker is waking up, notify workqueue */
2438 if ((p->flags & PF_WQ_WORKER) && success)
2439 wq_worker_waking_up(p, cpu_of(rq));
2443 * try_to_wake_up - wake up a thread
2444 * @p: the thread to be awakened
2445 * @state: the mask of task states that can be woken
2446 * @wake_flags: wake modifier flags (WF_*)
2448 * Put it on the run-queue if it's not already there. The "current"
2449 * thread is always on the run-queue (except when the actual
2450 * re-schedule is in progress), and as such you're allowed to do
2451 * the simpler "current->state = TASK_RUNNING" to mark yourself
2452 * runnable without the overhead of this.
2454 * Returns %true if @p was woken up, %false if it was already running
2455 * or @state didn't match @p's state.
2457 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2460 int cpu, orig_cpu, this_cpu, success = 0;
2461 unsigned long flags;
2462 unsigned long en_flags = ENQUEUE_WAKEUP;
2465 this_cpu = get_cpu();
2468 rq = task_rq_lock(p, &flags);
2469 if (!(p->state & state))
2479 if (unlikely(task_running(rq, p)))
2483 * In order to handle concurrent wakeups and release the rq->lock
2484 * we put the task in TASK_WAKING state.
2486 * First fix up the nr_uninterruptible count:
2488 if (task_contributes_to_load(p)) {
2489 if (likely(cpu_online(orig_cpu)))
2490 rq->nr_uninterruptible--;
2492 this_rq()->nr_uninterruptible--;
2494 p->state = TASK_WAKING;
2496 if (p->sched_class->task_waking) {
2497 p->sched_class->task_waking(rq, p);
2498 en_flags |= ENQUEUE_WAKING;
2501 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2502 if (cpu != orig_cpu)
2503 set_task_cpu(p, cpu);
2504 __task_rq_unlock(rq);
2507 raw_spin_lock(&rq->lock);
2510 * We migrated the task without holding either rq->lock, however
2511 * since the task is not on the task list itself, nobody else
2512 * will try and migrate the task, hence the rq should match the
2513 * cpu we just moved it to.
2515 WARN_ON(task_cpu(p) != cpu);
2516 WARN_ON(p->state != TASK_WAKING);
2518 #ifdef CONFIG_SCHEDSTATS
2519 schedstat_inc(rq, ttwu_count);
2520 if (cpu == this_cpu)
2521 schedstat_inc(rq, ttwu_local);
2523 struct sched_domain *sd;
2524 for_each_domain(this_cpu, sd) {
2525 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2526 schedstat_inc(sd, ttwu_wake_remote);
2531 #endif /* CONFIG_SCHEDSTATS */
2534 #endif /* CONFIG_SMP */
2535 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2536 cpu == this_cpu, en_flags);
2539 ttwu_post_activation(p, rq, wake_flags, success);
2541 task_rq_unlock(rq, &flags);
2548 * try_to_wake_up_local - try to wake up a local task with rq lock held
2549 * @p: the thread to be awakened
2551 * Put @p on the run-queue if it's not already there. The caller must
2552 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2553 * the current task. this_rq() stays locked over invocation.
2555 static void try_to_wake_up_local(struct task_struct *p)
2557 struct rq *rq = task_rq(p);
2558 bool success = false;
2560 BUG_ON(rq != this_rq());
2561 BUG_ON(p == current);
2562 lockdep_assert_held(&rq->lock);
2564 if (!(p->state & TASK_NORMAL))
2568 if (likely(!task_running(rq, p))) {
2569 schedstat_inc(rq, ttwu_count);
2570 schedstat_inc(rq, ttwu_local);
2572 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2575 ttwu_post_activation(p, rq, 0, success);
2579 * wake_up_process - Wake up a specific process
2580 * @p: The process to be woken up.
2582 * Attempt to wake up the nominated process and move it to the set of runnable
2583 * processes. Returns 1 if the process was woken up, 0 if it was already
2586 * It may be assumed that this function implies a write memory barrier before
2587 * changing the task state if and only if any tasks are woken up.
2589 int wake_up_process(struct task_struct *p)
2591 return try_to_wake_up(p, TASK_ALL, 0);
2593 EXPORT_SYMBOL(wake_up_process);
2595 int wake_up_state(struct task_struct *p, unsigned int state)
2597 return try_to_wake_up(p, state, 0);
2601 * Perform scheduler related setup for a newly forked process p.
2602 * p is forked by current.
2604 * __sched_fork() is basic setup used by init_idle() too:
2606 static void __sched_fork(struct task_struct *p)
2608 p->se.exec_start = 0;
2609 p->se.sum_exec_runtime = 0;
2610 p->se.prev_sum_exec_runtime = 0;
2611 p->se.nr_migrations = 0;
2614 #ifdef CONFIG_SCHEDSTATS
2615 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2618 INIT_LIST_HEAD(&p->rt.run_list);
2620 INIT_LIST_HEAD(&p->se.group_node);
2622 #ifdef CONFIG_PREEMPT_NOTIFIERS
2623 INIT_HLIST_HEAD(&p->preempt_notifiers);
2628 * fork()/clone()-time setup:
2630 void sched_fork(struct task_struct *p, int clone_flags)
2632 int cpu = get_cpu();
2636 * We mark the process as running here. This guarantees that
2637 * nobody will actually run it, and a signal or other external
2638 * event cannot wake it up and insert it on the runqueue either.
2640 p->state = TASK_RUNNING;
2643 * Revert to default priority/policy on fork if requested.
2645 if (unlikely(p->sched_reset_on_fork)) {
2646 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2647 p->policy = SCHED_NORMAL;
2648 p->normal_prio = p->static_prio;
2651 if (PRIO_TO_NICE(p->static_prio) < 0) {
2652 p->static_prio = NICE_TO_PRIO(0);
2653 p->normal_prio = p->static_prio;
2658 * We don't need the reset flag anymore after the fork. It has
2659 * fulfilled its duty:
2661 p->sched_reset_on_fork = 0;
2665 * Make sure we do not leak PI boosting priority to the child.
2667 p->prio = current->normal_prio;
2669 if (!rt_prio(p->prio))
2670 p->sched_class = &fair_sched_class;
2672 if (p->sched_class->task_fork)
2673 p->sched_class->task_fork(p);
2676 * The child is not yet in the pid-hash so no cgroup attach races,
2677 * and the cgroup is pinned to this child due to cgroup_fork()
2678 * is ran before sched_fork().
2680 * Silence PROVE_RCU.
2683 set_task_cpu(p, cpu);
2686 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2687 if (likely(sched_info_on()))
2688 memset(&p->sched_info, 0, sizeof(p->sched_info));
2690 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2693 #ifdef CONFIG_PREEMPT
2694 /* Want to start with kernel preemption disabled. */
2695 task_thread_info(p)->preempt_count = 1;
2698 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2705 * wake_up_new_task - wake up a newly created task for the first time.
2707 * This function will do some initial scheduler statistics housekeeping
2708 * that must be done for every newly created context, then puts the task
2709 * on the runqueue and wakes it.
2711 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2713 unsigned long flags;
2715 int cpu __maybe_unused = get_cpu();
2718 rq = task_rq_lock(p, &flags);
2719 p->state = TASK_WAKING;
2722 * Fork balancing, do it here and not earlier because:
2723 * - cpus_allowed can change in the fork path
2724 * - any previously selected cpu might disappear through hotplug
2726 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2727 * without people poking at ->cpus_allowed.
2729 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2730 set_task_cpu(p, cpu);
2732 p->state = TASK_RUNNING;
2733 task_rq_unlock(rq, &flags);
2736 rq = task_rq_lock(p, &flags);
2737 activate_task(rq, p, 0);
2738 trace_sched_wakeup_new(p, 1);
2739 check_preempt_curr(rq, p, WF_FORK);
2741 if (p->sched_class->task_woken)
2742 p->sched_class->task_woken(rq, p);
2744 task_rq_unlock(rq, &flags);
2748 #ifdef CONFIG_PREEMPT_NOTIFIERS
2751 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2752 * @notifier: notifier struct to register
2754 void preempt_notifier_register(struct preempt_notifier *notifier)
2756 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2758 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2761 * preempt_notifier_unregister - no longer interested in preemption notifications
2762 * @notifier: notifier struct to unregister
2764 * This is safe to call from within a preemption notifier.
2766 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2768 hlist_del(¬ifier->link);
2770 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2772 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2774 struct preempt_notifier *notifier;
2775 struct hlist_node *node;
2777 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2778 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2782 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2783 struct task_struct *next)
2785 struct preempt_notifier *notifier;
2786 struct hlist_node *node;
2788 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2789 notifier->ops->sched_out(notifier, next);
2792 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2794 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2799 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2800 struct task_struct *next)
2804 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2807 * prepare_task_switch - prepare to switch tasks
2808 * @rq: the runqueue preparing to switch
2809 * @prev: the current task that is being switched out
2810 * @next: the task we are going to switch to.
2812 * This is called with the rq lock held and interrupts off. It must
2813 * be paired with a subsequent finish_task_switch after the context
2816 * prepare_task_switch sets up locking and calls architecture specific
2820 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2821 struct task_struct *next)
2823 sched_info_switch(prev, next);
2824 perf_event_task_sched_out(prev, next);
2825 fire_sched_out_preempt_notifiers(prev, next);
2826 prepare_lock_switch(rq, next);
2827 prepare_arch_switch(next);
2828 trace_sched_switch(prev, next);
2832 * finish_task_switch - clean up after a task-switch
2833 * @rq: runqueue associated with task-switch
2834 * @prev: the thread we just switched away from.
2836 * finish_task_switch must be called after the context switch, paired
2837 * with a prepare_task_switch call before the context switch.
2838 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2839 * and do any other architecture-specific cleanup actions.
2841 * Note that we may have delayed dropping an mm in context_switch(). If
2842 * so, we finish that here outside of the runqueue lock. (Doing it
2843 * with the lock held can cause deadlocks; see schedule() for
2846 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2847 __releases(rq->lock)
2849 struct mm_struct *mm = rq->prev_mm;
2855 * A task struct has one reference for the use as "current".
2856 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2857 * schedule one last time. The schedule call will never return, and
2858 * the scheduled task must drop that reference.
2859 * The test for TASK_DEAD must occur while the runqueue locks are
2860 * still held, otherwise prev could be scheduled on another cpu, die
2861 * there before we look at prev->state, and then the reference would
2863 * Manfred Spraul <manfred@colorfullife.com>
2865 prev_state = prev->state;
2866 finish_arch_switch(prev);
2867 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2868 local_irq_disable();
2869 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2870 perf_event_task_sched_in(current);
2871 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2873 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2874 finish_lock_switch(rq, prev);
2876 fire_sched_in_preempt_notifiers(current);
2879 if (unlikely(prev_state == TASK_DEAD)) {
2881 * Remove function-return probe instances associated with this
2882 * task and put them back on the free list.
2884 kprobe_flush_task(prev);
2885 put_task_struct(prev);
2891 /* assumes rq->lock is held */
2892 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2894 if (prev->sched_class->pre_schedule)
2895 prev->sched_class->pre_schedule(rq, prev);
2898 /* rq->lock is NOT held, but preemption is disabled */
2899 static inline void post_schedule(struct rq *rq)
2901 if (rq->post_schedule) {
2902 unsigned long flags;
2904 raw_spin_lock_irqsave(&rq->lock, flags);
2905 if (rq->curr->sched_class->post_schedule)
2906 rq->curr->sched_class->post_schedule(rq);
2907 raw_spin_unlock_irqrestore(&rq->lock, flags);
2909 rq->post_schedule = 0;
2915 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2919 static inline void post_schedule(struct rq *rq)
2926 * schedule_tail - first thing a freshly forked thread must call.
2927 * @prev: the thread we just switched away from.
2929 asmlinkage void schedule_tail(struct task_struct *prev)
2930 __releases(rq->lock)
2932 struct rq *rq = this_rq();
2934 finish_task_switch(rq, prev);
2937 * FIXME: do we need to worry about rq being invalidated by the
2942 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2943 /* In this case, finish_task_switch does not reenable preemption */
2946 if (current->set_child_tid)
2947 put_user(task_pid_vnr(current), current->set_child_tid);
2951 * context_switch - switch to the new MM and the new
2952 * thread's register state.
2955 context_switch(struct rq *rq, struct task_struct *prev,
2956 struct task_struct *next)
2958 struct mm_struct *mm, *oldmm;
2960 prepare_task_switch(rq, prev, next);
2963 oldmm = prev->active_mm;
2965 * For paravirt, this is coupled with an exit in switch_to to
2966 * combine the page table reload and the switch backend into
2969 arch_start_context_switch(prev);
2972 next->active_mm = oldmm;
2973 atomic_inc(&oldmm->mm_count);
2974 enter_lazy_tlb(oldmm, next);
2976 switch_mm(oldmm, mm, next);
2979 prev->active_mm = NULL;
2980 rq->prev_mm = oldmm;
2983 * Since the runqueue lock will be released by the next
2984 * task (which is an invalid locking op but in the case
2985 * of the scheduler it's an obvious special-case), so we
2986 * do an early lockdep release here:
2988 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2989 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2992 /* Here we just switch the register state and the stack. */
2993 switch_to(prev, next, prev);
2997 * this_rq must be evaluated again because prev may have moved
2998 * CPUs since it called schedule(), thus the 'rq' on its stack
2999 * frame will be invalid.
3001 finish_task_switch(this_rq(), prev);
3005 * nr_running, nr_uninterruptible and nr_context_switches:
3007 * externally visible scheduler statistics: current number of runnable
3008 * threads, current number of uninterruptible-sleeping threads, total
3009 * number of context switches performed since bootup.
3011 unsigned long nr_running(void)
3013 unsigned long i, sum = 0;
3015 for_each_online_cpu(i)
3016 sum += cpu_rq(i)->nr_running;
3021 unsigned long nr_uninterruptible(void)
3023 unsigned long i, sum = 0;
3025 for_each_possible_cpu(i)
3026 sum += cpu_rq(i)->nr_uninterruptible;
3029 * Since we read the counters lockless, it might be slightly
3030 * inaccurate. Do not allow it to go below zero though:
3032 if (unlikely((long)sum < 0))
3038 unsigned long long nr_context_switches(void)
3041 unsigned long long sum = 0;
3043 for_each_possible_cpu(i)
3044 sum += cpu_rq(i)->nr_switches;
3049 unsigned long nr_iowait(void)
3051 unsigned long i, sum = 0;
3053 for_each_possible_cpu(i)
3054 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3059 unsigned long nr_iowait_cpu(int cpu)
3061 struct rq *this = cpu_rq(cpu);
3062 return atomic_read(&this->nr_iowait);
3065 unsigned long this_cpu_load(void)
3067 struct rq *this = this_rq();
3068 return this->cpu_load[0];
3072 /* Variables and functions for calc_load */
3073 static atomic_long_t calc_load_tasks;
3074 static unsigned long calc_load_update;
3075 unsigned long avenrun[3];
3076 EXPORT_SYMBOL(avenrun);
3078 static long calc_load_fold_active(struct rq *this_rq)
3080 long nr_active, delta = 0;
3082 nr_active = this_rq->nr_running;
3083 nr_active += (long) this_rq->nr_uninterruptible;
3085 if (nr_active != this_rq->calc_load_active) {
3086 delta = nr_active - this_rq->calc_load_active;
3087 this_rq->calc_load_active = nr_active;
3093 static unsigned long
3094 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3097 load += active * (FIXED_1 - exp);
3098 load += 1UL << (FSHIFT - 1);
3099 return load >> FSHIFT;
3104 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3106 * When making the ILB scale, we should try to pull this in as well.
3108 static atomic_long_t calc_load_tasks_idle;
3110 static void calc_load_account_idle(struct rq *this_rq)
3114 delta = calc_load_fold_active(this_rq);
3116 atomic_long_add(delta, &calc_load_tasks_idle);
3119 static long calc_load_fold_idle(void)
3124 * Its got a race, we don't care...
3126 if (atomic_long_read(&calc_load_tasks_idle))
3127 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3133 * fixed_power_int - compute: x^n, in O(log n) time
3135 * @x: base of the power
3136 * @frac_bits: fractional bits of @x
3137 * @n: power to raise @x to.
3139 * By exploiting the relation between the definition of the natural power
3140 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3141 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3142 * (where: n_i \elem {0, 1}, the binary vector representing n),
3143 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3144 * of course trivially computable in O(log_2 n), the length of our binary
3147 static unsigned long
3148 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3150 unsigned long result = 1UL << frac_bits;
3155 result += 1UL << (frac_bits - 1);
3156 result >>= frac_bits;
3162 x += 1UL << (frac_bits - 1);
3170 * a1 = a0 * e + a * (1 - e)
3172 * a2 = a1 * e + a * (1 - e)
3173 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3174 * = a0 * e^2 + a * (1 - e) * (1 + e)
3176 * a3 = a2 * e + a * (1 - e)
3177 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3178 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3182 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3183 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3184 * = a0 * e^n + a * (1 - e^n)
3186 * [1] application of the geometric series:
3189 * S_n := \Sum x^i = -------------
3192 static unsigned long
3193 calc_load_n(unsigned long load, unsigned long exp,
3194 unsigned long active, unsigned int n)
3197 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3201 * NO_HZ can leave us missing all per-cpu ticks calling
3202 * calc_load_account_active(), but since an idle CPU folds its delta into
3203 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3204 * in the pending idle delta if our idle period crossed a load cycle boundary.
3206 * Once we've updated the global active value, we need to apply the exponential
3207 * weights adjusted to the number of cycles missed.
3209 static void calc_global_nohz(unsigned long ticks)
3211 long delta, active, n;
3213 if (time_before(jiffies, calc_load_update))
3217 * If we crossed a calc_load_update boundary, make sure to fold
3218 * any pending idle changes, the respective CPUs might have
3219 * missed the tick driven calc_load_account_active() update
3222 delta = calc_load_fold_idle();
3224 atomic_long_add(delta, &calc_load_tasks);
3227 * If we were idle for multiple load cycles, apply them.
3229 if (ticks >= LOAD_FREQ) {
3230 n = ticks / LOAD_FREQ;
3232 active = atomic_long_read(&calc_load_tasks);
3233 active = active > 0 ? active * FIXED_1 : 0;
3235 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3236 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3237 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3239 calc_load_update += n * LOAD_FREQ;
3243 * Its possible the remainder of the above division also crosses
3244 * a LOAD_FREQ period, the regular check in calc_global_load()
3245 * which comes after this will take care of that.
3247 * Consider us being 11 ticks before a cycle completion, and us
3248 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3249 * age us 4 cycles, and the test in calc_global_load() will
3250 * pick up the final one.
3254 static void calc_load_account_idle(struct rq *this_rq)
3258 static inline long calc_load_fold_idle(void)
3263 static void calc_global_nohz(unsigned long ticks)
3269 * get_avenrun - get the load average array
3270 * @loads: pointer to dest load array
3271 * @offset: offset to add
3272 * @shift: shift count to shift the result left
3274 * These values are estimates at best, so no need for locking.
3276 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3278 loads[0] = (avenrun[0] + offset) << shift;
3279 loads[1] = (avenrun[1] + offset) << shift;
3280 loads[2] = (avenrun[2] + offset) << shift;
3284 * calc_load - update the avenrun load estimates 10 ticks after the
3285 * CPUs have updated calc_load_tasks.
3287 void calc_global_load(unsigned long ticks)
3291 calc_global_nohz(ticks);
3293 if (time_before(jiffies, calc_load_update + 10))
3296 active = atomic_long_read(&calc_load_tasks);
3297 active = active > 0 ? active * FIXED_1 : 0;
3299 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3300 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3301 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3303 calc_load_update += LOAD_FREQ;
3307 * Called from update_cpu_load() to periodically update this CPU's
3310 static void calc_load_account_active(struct rq *this_rq)
3314 if (time_before(jiffies, this_rq->calc_load_update))
3317 delta = calc_load_fold_active(this_rq);
3318 delta += calc_load_fold_idle();
3320 atomic_long_add(delta, &calc_load_tasks);
3322 this_rq->calc_load_update += LOAD_FREQ;
3326 * The exact cpuload at various idx values, calculated at every tick would be
3327 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3329 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3330 * on nth tick when cpu may be busy, then we have:
3331 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3332 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3334 * decay_load_missed() below does efficient calculation of
3335 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3336 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3338 * The calculation is approximated on a 128 point scale.
3339 * degrade_zero_ticks is the number of ticks after which load at any
3340 * particular idx is approximated to be zero.
3341 * degrade_factor is a precomputed table, a row for each load idx.
3342 * Each column corresponds to degradation factor for a power of two ticks,
3343 * based on 128 point scale.
3345 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3346 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3348 * With this power of 2 load factors, we can degrade the load n times
3349 * by looking at 1 bits in n and doing as many mult/shift instead of
3350 * n mult/shifts needed by the exact degradation.
3352 #define DEGRADE_SHIFT 7
3353 static const unsigned char
3354 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3355 static const unsigned char
3356 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3357 {0, 0, 0, 0, 0, 0, 0, 0},
3358 {64, 32, 8, 0, 0, 0, 0, 0},
3359 {96, 72, 40, 12, 1, 0, 0},
3360 {112, 98, 75, 43, 15, 1, 0},
3361 {120, 112, 98, 76, 45, 16, 2} };
3364 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3365 * would be when CPU is idle and so we just decay the old load without
3366 * adding any new load.
3368 static unsigned long
3369 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3373 if (!missed_updates)
3376 if (missed_updates >= degrade_zero_ticks[idx])
3380 return load >> missed_updates;
3382 while (missed_updates) {
3383 if (missed_updates % 2)
3384 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3386 missed_updates >>= 1;
3393 * Update rq->cpu_load[] statistics. This function is usually called every
3394 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3395 * every tick. We fix it up based on jiffies.
3397 static void update_cpu_load(struct rq *this_rq)
3399 unsigned long this_load = this_rq->load.weight;
3400 unsigned long curr_jiffies = jiffies;
3401 unsigned long pending_updates;
3404 this_rq->nr_load_updates++;
3406 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3407 if (curr_jiffies == this_rq->last_load_update_tick)
3410 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3411 this_rq->last_load_update_tick = curr_jiffies;
3413 /* Update our load: */
3414 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3415 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3416 unsigned long old_load, new_load;
3418 /* scale is effectively 1 << i now, and >> i divides by scale */
3420 old_load = this_rq->cpu_load[i];
3421 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3422 new_load = this_load;
3424 * Round up the averaging division if load is increasing. This
3425 * prevents us from getting stuck on 9 if the load is 10, for
3428 if (new_load > old_load)
3429 new_load += scale - 1;
3431 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3434 sched_avg_update(this_rq);
3437 static void update_cpu_load_active(struct rq *this_rq)
3439 update_cpu_load(this_rq);
3441 calc_load_account_active(this_rq);
3447 * sched_exec - execve() is a valuable balancing opportunity, because at
3448 * this point the task has the smallest effective memory and cache footprint.
3450 void sched_exec(void)
3452 struct task_struct *p = current;
3453 unsigned long flags;
3457 rq = task_rq_lock(p, &flags);
3458 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3459 if (dest_cpu == smp_processor_id())
3463 * select_task_rq() can race against ->cpus_allowed
3465 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3466 likely(cpu_active(dest_cpu)) && migrate_task(p, rq)) {
3467 struct migration_arg arg = { p, dest_cpu };
3469 task_rq_unlock(rq, &flags);
3470 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3474 task_rq_unlock(rq, &flags);
3479 DEFINE_PER_CPU(struct kernel_stat, kstat);
3481 EXPORT_PER_CPU_SYMBOL(kstat);
3484 * Return any ns on the sched_clock that have not yet been accounted in
3485 * @p in case that task is currently running.
3487 * Called with task_rq_lock() held on @rq.
3489 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3493 if (task_current(rq, p)) {
3494 update_rq_clock(rq);
3495 ns = rq->clock_task - p->se.exec_start;
3503 unsigned long long task_delta_exec(struct task_struct *p)
3505 unsigned long flags;
3509 rq = task_rq_lock(p, &flags);
3510 ns = do_task_delta_exec(p, rq);
3511 task_rq_unlock(rq, &flags);
3517 * Return accounted runtime for the task.
3518 * In case the task is currently running, return the runtime plus current's
3519 * pending runtime that have not been accounted yet.
3521 unsigned long long task_sched_runtime(struct task_struct *p)
3523 unsigned long flags;
3527 rq = task_rq_lock(p, &flags);
3528 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3529 task_rq_unlock(rq, &flags);
3535 * Return sum_exec_runtime for the thread group.
3536 * In case the task is currently running, return the sum plus current's
3537 * pending runtime that have not been accounted yet.
3539 * Note that the thread group might have other running tasks as well,
3540 * so the return value not includes other pending runtime that other
3541 * running tasks might have.
3543 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3545 struct task_cputime totals;
3546 unsigned long flags;
3550 rq = task_rq_lock(p, &flags);
3551 thread_group_cputime(p, &totals);
3552 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3553 task_rq_unlock(rq, &flags);
3559 * Account user cpu time to a process.
3560 * @p: the process that the cpu time gets accounted to
3561 * @cputime: the cpu time spent in user space since the last update
3562 * @cputime_scaled: cputime scaled by cpu frequency
3564 void account_user_time(struct task_struct *p, cputime_t cputime,
3565 cputime_t cputime_scaled)
3567 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3570 /* Add user time to process. */
3571 p->utime = cputime_add(p->utime, cputime);
3572 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3573 account_group_user_time(p, cputime);
3575 /* Add user time to cpustat. */
3576 tmp = cputime_to_cputime64(cputime);
3577 if (TASK_NICE(p) > 0)
3578 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3580 cpustat->user = cputime64_add(cpustat->user, tmp);
3582 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3583 /* Account for user time used */
3584 acct_update_integrals(p);
3588 * Account guest cpu time to a process.
3589 * @p: the process that the cpu time gets accounted to
3590 * @cputime: the cpu time spent in virtual machine since the last update
3591 * @cputime_scaled: cputime scaled by cpu frequency
3593 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3594 cputime_t cputime_scaled)
3597 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3599 tmp = cputime_to_cputime64(cputime);
3601 /* Add guest time to process. */
3602 p->utime = cputime_add(p->utime, cputime);
3603 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3604 account_group_user_time(p, cputime);
3605 p->gtime = cputime_add(p->gtime, cputime);
3607 /* Add guest time to cpustat. */
3608 if (TASK_NICE(p) > 0) {
3609 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3610 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3612 cpustat->user = cputime64_add(cpustat->user, tmp);
3613 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3618 * Account system cpu time to a process and desired cpustat field
3619 * @p: the process that the cpu time gets accounted to
3620 * @cputime: the cpu time spent in kernel space since the last update
3621 * @cputime_scaled: cputime scaled by cpu frequency
3622 * @target_cputime64: pointer to cpustat field that has to be updated
3625 void __account_system_time(struct task_struct *p, cputime_t cputime,
3626 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3628 cputime64_t tmp = cputime_to_cputime64(cputime);
3630 /* Add system time to process. */
3631 p->stime = cputime_add(p->stime, cputime);
3632 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3633 account_group_system_time(p, cputime);
3635 /* Add system time to cpustat. */
3636 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3637 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3639 /* Account for system time used */
3640 acct_update_integrals(p);
3644 * Account system cpu time to a process.
3645 * @p: the process that the cpu time gets accounted to
3646 * @hardirq_offset: the offset to subtract from hardirq_count()
3647 * @cputime: the cpu time spent in kernel space since the last update
3648 * @cputime_scaled: cputime scaled by cpu frequency
3650 void account_system_time(struct task_struct *p, int hardirq_offset,
3651 cputime_t cputime, cputime_t cputime_scaled)
3653 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3654 cputime64_t *target_cputime64;
3656 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3657 account_guest_time(p, cputime, cputime_scaled);
3661 if (hardirq_count() - hardirq_offset)
3662 target_cputime64 = &cpustat->irq;
3663 else if (in_serving_softirq())
3664 target_cputime64 = &cpustat->softirq;
3666 target_cputime64 = &cpustat->system;
3668 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3672 * Account for involuntary wait time.
3673 * @cputime: the cpu time spent in involuntary wait
3675 void account_steal_time(cputime_t cputime)
3677 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3678 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3680 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3684 * Account for idle time.
3685 * @cputime: the cpu time spent in idle wait
3687 void account_idle_time(cputime_t cputime)
3689 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3690 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3691 struct rq *rq = this_rq();
3693 if (atomic_read(&rq->nr_iowait) > 0)
3694 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3696 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3699 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3701 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3703 * Account a tick to a process and cpustat
3704 * @p: the process that the cpu time gets accounted to
3705 * @user_tick: is the tick from userspace
3706 * @rq: the pointer to rq
3708 * Tick demultiplexing follows the order
3709 * - pending hardirq update
3710 * - pending softirq update
3714 * - check for guest_time
3715 * - else account as system_time
3717 * Check for hardirq is done both for system and user time as there is
3718 * no timer going off while we are on hardirq and hence we may never get an
3719 * opportunity to update it solely in system time.
3720 * p->stime and friends are only updated on system time and not on irq
3721 * softirq as those do not count in task exec_runtime any more.
3723 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3726 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3727 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3728 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3730 if (irqtime_account_hi_update()) {
3731 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3732 } else if (irqtime_account_si_update()) {
3733 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3734 } else if (this_cpu_ksoftirqd() == p) {
3736 * ksoftirqd time do not get accounted in cpu_softirq_time.
3737 * So, we have to handle it separately here.
3738 * Also, p->stime needs to be updated for ksoftirqd.
3740 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3742 } else if (user_tick) {
3743 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3744 } else if (p == rq->idle) {
3745 account_idle_time(cputime_one_jiffy);
3746 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3747 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3749 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3754 static void irqtime_account_idle_ticks(int ticks)
3757 struct rq *rq = this_rq();
3759 for (i = 0; i < ticks; i++)
3760 irqtime_account_process_tick(current, 0, rq);
3762 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3763 static void irqtime_account_idle_ticks(int ticks) {}
3764 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3766 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3769 * Account a single tick of cpu time.
3770 * @p: the process that the cpu time gets accounted to
3771 * @user_tick: indicates if the tick is a user or a system tick
3773 void account_process_tick(struct task_struct *p, int user_tick)
3775 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3776 struct rq *rq = this_rq();
3778 if (sched_clock_irqtime) {
3779 irqtime_account_process_tick(p, user_tick, rq);
3784 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3785 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3786 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3789 account_idle_time(cputime_one_jiffy);
3793 * Account multiple ticks of steal time.
3794 * @p: the process from which the cpu time has been stolen
3795 * @ticks: number of stolen ticks
3797 void account_steal_ticks(unsigned long ticks)
3799 account_steal_time(jiffies_to_cputime(ticks));
3803 * Account multiple ticks of idle time.
3804 * @ticks: number of stolen ticks
3806 void account_idle_ticks(unsigned long ticks)
3809 if (sched_clock_irqtime) {
3810 irqtime_account_idle_ticks(ticks);
3814 account_idle_time(jiffies_to_cputime(ticks));
3820 * Use precise platform statistics if available:
3822 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3823 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3829 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3831 struct task_cputime cputime;
3833 thread_group_cputime(p, &cputime);
3835 *ut = cputime.utime;
3836 *st = cputime.stime;
3840 #ifndef nsecs_to_cputime
3841 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3844 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3846 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3849 * Use CFS's precise accounting:
3851 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3857 do_div(temp, total);
3858 utime = (cputime_t)temp;
3863 * Compare with previous values, to keep monotonicity:
3865 p->prev_utime = max(p->prev_utime, utime);
3866 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3868 *ut = p->prev_utime;
3869 *st = p->prev_stime;
3873 * Must be called with siglock held.
3875 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3877 struct signal_struct *sig = p->signal;
3878 struct task_cputime cputime;
3879 cputime_t rtime, utime, total;
3881 thread_group_cputime(p, &cputime);
3883 total = cputime_add(cputime.utime, cputime.stime);
3884 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3889 temp *= cputime.utime;
3890 do_div(temp, total);
3891 utime = (cputime_t)temp;
3895 sig->prev_utime = max(sig->prev_utime, utime);
3896 sig->prev_stime = max(sig->prev_stime,
3897 cputime_sub(rtime, sig->prev_utime));
3899 *ut = sig->prev_utime;
3900 *st = sig->prev_stime;
3905 * This function gets called by the timer code, with HZ frequency.
3906 * We call it with interrupts disabled.
3908 * It also gets called by the fork code, when changing the parent's
3911 void scheduler_tick(void)
3913 int cpu = smp_processor_id();
3914 struct rq *rq = cpu_rq(cpu);
3915 struct task_struct *curr = rq->curr;
3919 raw_spin_lock(&rq->lock);
3920 update_rq_clock(rq);
3921 update_cpu_load_active(rq);
3922 curr->sched_class->task_tick(rq, curr, 0);
3923 raw_spin_unlock(&rq->lock);
3925 perf_event_task_tick();
3928 rq->idle_at_tick = idle_cpu(cpu);
3929 trigger_load_balance(rq, cpu);
3933 notrace unsigned long get_parent_ip(unsigned long addr)
3935 if (in_lock_functions(addr)) {
3936 addr = CALLER_ADDR2;
3937 if (in_lock_functions(addr))
3938 addr = CALLER_ADDR3;
3943 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3944 defined(CONFIG_PREEMPT_TRACER))
3946 void __kprobes add_preempt_count(int val)
3948 #ifdef CONFIG_DEBUG_PREEMPT
3952 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3955 preempt_count() += val;
3956 #ifdef CONFIG_DEBUG_PREEMPT
3958 * Spinlock count overflowing soon?
3960 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3963 if (preempt_count() == val)
3964 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3966 EXPORT_SYMBOL(add_preempt_count);
3968 void __kprobes sub_preempt_count(int val)
3970 #ifdef CONFIG_DEBUG_PREEMPT
3974 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3977 * Is the spinlock portion underflowing?
3979 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3980 !(preempt_count() & PREEMPT_MASK)))
3984 if (preempt_count() == val)
3985 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3986 preempt_count() -= val;
3988 EXPORT_SYMBOL(sub_preempt_count);
3993 * Print scheduling while atomic bug:
3995 static noinline void __schedule_bug(struct task_struct *prev)
3997 struct pt_regs *regs = get_irq_regs();
3999 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4000 prev->comm, prev->pid, preempt_count());
4002 debug_show_held_locks(prev);
4004 if (irqs_disabled())
4005 print_irqtrace_events(prev);
4014 * Various schedule()-time debugging checks and statistics:
4016 static inline void schedule_debug(struct task_struct *prev)
4019 * Test if we are atomic. Since do_exit() needs to call into
4020 * schedule() atomically, we ignore that path for now.
4021 * Otherwise, whine if we are scheduling when we should not be.
4023 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4024 __schedule_bug(prev);
4026 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4028 schedstat_inc(this_rq(), sched_count);
4029 #ifdef CONFIG_SCHEDSTATS
4030 if (unlikely(prev->lock_depth >= 0)) {
4031 schedstat_inc(this_rq(), rq_sched_info.bkl_count);
4032 schedstat_inc(prev, sched_info.bkl_count);
4037 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4040 update_rq_clock(rq);
4041 prev->sched_class->put_prev_task(rq, prev);
4045 * Pick up the highest-prio task:
4047 static inline struct task_struct *
4048 pick_next_task(struct rq *rq)
4050 const struct sched_class *class;
4051 struct task_struct *p;
4054 * Optimization: we know that if all tasks are in
4055 * the fair class we can call that function directly:
4057 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4058 p = fair_sched_class.pick_next_task(rq);
4063 for_each_class(class) {
4064 p = class->pick_next_task(rq);
4069 BUG(); /* the idle class will always have a runnable task */
4073 * schedule() is the main scheduler function.
4075 asmlinkage void __sched schedule(void)
4077 struct task_struct *prev, *next;
4078 unsigned long *switch_count;
4084 cpu = smp_processor_id();
4086 rcu_note_context_switch(cpu);
4089 release_kernel_lock(prev);
4090 need_resched_nonpreemptible:
4092 schedule_debug(prev);
4094 if (sched_feat(HRTICK))
4097 raw_spin_lock_irq(&rq->lock);
4099 switch_count = &prev->nivcsw;
4100 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4101 if (unlikely(signal_pending_state(prev->state, prev))) {
4102 prev->state = TASK_RUNNING;
4105 * If a worker is going to sleep, notify and
4106 * ask workqueue whether it wants to wake up a
4107 * task to maintain concurrency. If so, wake
4110 if (prev->flags & PF_WQ_WORKER) {
4111 struct task_struct *to_wakeup;
4113 to_wakeup = wq_worker_sleeping(prev, cpu);
4115 try_to_wake_up_local(to_wakeup);
4117 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4119 switch_count = &prev->nvcsw;
4122 pre_schedule(rq, prev);
4124 if (unlikely(!rq->nr_running))
4125 idle_balance(cpu, rq);
4127 put_prev_task(rq, prev);
4128 next = pick_next_task(rq);
4129 clear_tsk_need_resched(prev);
4130 rq->skip_clock_update = 0;
4132 if (likely(prev != next)) {
4137 context_switch(rq, prev, next); /* unlocks the rq */
4139 * The context switch have flipped the stack from under us
4140 * and restored the local variables which were saved when
4141 * this task called schedule() in the past. prev == current
4142 * is still correct, but it can be moved to another cpu/rq.
4144 cpu = smp_processor_id();
4147 raw_spin_unlock_irq(&rq->lock);
4151 if (unlikely(reacquire_kernel_lock(prev)))
4152 goto need_resched_nonpreemptible;
4154 preempt_enable_no_resched();
4158 EXPORT_SYMBOL(schedule);
4160 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4162 * Look out! "owner" is an entirely speculative pointer
4163 * access and not reliable.
4165 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
4170 if (!sched_feat(OWNER_SPIN))
4173 #ifdef CONFIG_DEBUG_PAGEALLOC
4175 * Need to access the cpu field knowing that
4176 * DEBUG_PAGEALLOC could have unmapped it if
4177 * the mutex owner just released it and exited.
4179 if (probe_kernel_address(&owner->cpu, cpu))
4186 * Even if the access succeeded (likely case),
4187 * the cpu field may no longer be valid.
4189 if (cpu >= nr_cpumask_bits)
4193 * We need to validate that we can do a
4194 * get_cpu() and that we have the percpu area.
4196 if (!cpu_online(cpu))
4203 * Owner changed, break to re-assess state.
4205 if (lock->owner != owner) {
4207 * If the lock has switched to a different owner,
4208 * we likely have heavy contention. Return 0 to quit
4209 * optimistic spinning and not contend further:
4217 * Is that owner really running on that cpu?
4219 if (task_thread_info(rq->curr) != owner || need_resched())
4222 arch_mutex_cpu_relax();
4229 #ifdef CONFIG_PREEMPT
4231 * this is the entry point to schedule() from in-kernel preemption
4232 * off of preempt_enable. Kernel preemptions off return from interrupt
4233 * occur there and call schedule directly.
4235 asmlinkage void __sched notrace preempt_schedule(void)
4237 struct thread_info *ti = current_thread_info();
4240 * If there is a non-zero preempt_count or interrupts are disabled,
4241 * we do not want to preempt the current task. Just return..
4243 if (likely(ti->preempt_count || irqs_disabled()))
4247 add_preempt_count_notrace(PREEMPT_ACTIVE);
4249 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4252 * Check again in case we missed a preemption opportunity
4253 * between schedule and now.
4256 } while (need_resched());
4258 EXPORT_SYMBOL(preempt_schedule);
4261 * this is the entry point to schedule() from kernel preemption
4262 * off of irq context.
4263 * Note, that this is called and return with irqs disabled. This will
4264 * protect us against recursive calling from irq.
4266 asmlinkage void __sched preempt_schedule_irq(void)
4268 struct thread_info *ti = current_thread_info();
4270 /* Catch callers which need to be fixed */
4271 BUG_ON(ti->preempt_count || !irqs_disabled());
4274 add_preempt_count(PREEMPT_ACTIVE);
4277 local_irq_disable();
4278 sub_preempt_count(PREEMPT_ACTIVE);
4281 * Check again in case we missed a preemption opportunity
4282 * between schedule and now.
4285 } while (need_resched());
4288 #endif /* CONFIG_PREEMPT */
4290 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4293 return try_to_wake_up(curr->private, mode, wake_flags);
4295 EXPORT_SYMBOL(default_wake_function);
4298 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4299 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4300 * number) then we wake all the non-exclusive tasks and one exclusive task.
4302 * There are circumstances in which we can try to wake a task which has already
4303 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4304 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4306 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4307 int nr_exclusive, int wake_flags, void *key)
4309 wait_queue_t *curr, *next;
4311 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4312 unsigned flags = curr->flags;
4314 if (curr->func(curr, mode, wake_flags, key) &&
4315 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4321 * __wake_up - wake up threads blocked on a waitqueue.
4323 * @mode: which threads
4324 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4325 * @key: is directly passed to the wakeup function
4327 * It may be assumed that this function implies a write memory barrier before
4328 * changing the task state if and only if any tasks are woken up.
4330 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4331 int nr_exclusive, void *key)
4333 unsigned long flags;
4335 spin_lock_irqsave(&q->lock, flags);
4336 __wake_up_common(q, mode, nr_exclusive, 0, key);
4337 spin_unlock_irqrestore(&q->lock, flags);
4339 EXPORT_SYMBOL(__wake_up);
4342 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4344 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4346 __wake_up_common(q, mode, 1, 0, NULL);
4348 EXPORT_SYMBOL_GPL(__wake_up_locked);
4350 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4352 __wake_up_common(q, mode, 1, 0, key);
4354 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4357 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4359 * @mode: which threads
4360 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4361 * @key: opaque value to be passed to wakeup targets
4363 * The sync wakeup differs that the waker knows that it will schedule
4364 * away soon, so while the target thread will be woken up, it will not
4365 * be migrated to another CPU - ie. the two threads are 'synchronized'
4366 * with each other. This can prevent needless bouncing between CPUs.
4368 * On UP it can prevent extra preemption.
4370 * It may be assumed that this function implies a write memory barrier before
4371 * changing the task state if and only if any tasks are woken up.
4373 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4374 int nr_exclusive, void *key)
4376 unsigned long flags;
4377 int wake_flags = WF_SYNC;
4382 if (unlikely(!nr_exclusive))
4385 spin_lock_irqsave(&q->lock, flags);
4386 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4387 spin_unlock_irqrestore(&q->lock, flags);
4389 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4392 * __wake_up_sync - see __wake_up_sync_key()
4394 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4396 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4398 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4401 * complete: - signals a single thread waiting on this completion
4402 * @x: holds the state of this particular completion
4404 * This will wake up a single thread waiting on this completion. Threads will be
4405 * awakened in the same order in which they were queued.
4407 * See also complete_all(), wait_for_completion() and related routines.
4409 * It may be assumed that this function implies a write memory barrier before
4410 * changing the task state if and only if any tasks are woken up.
4412 void complete(struct completion *x)
4414 unsigned long flags;
4416 spin_lock_irqsave(&x->wait.lock, flags);
4418 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4419 spin_unlock_irqrestore(&x->wait.lock, flags);
4421 EXPORT_SYMBOL(complete);
4424 * complete_all: - signals all threads waiting on this completion
4425 * @x: holds the state of this particular completion
4427 * This will wake up all threads waiting on this particular completion event.
4429 * It may be assumed that this function implies a write memory barrier before
4430 * changing the task state if and only if any tasks are woken up.
4432 void complete_all(struct completion *x)
4434 unsigned long flags;
4436 spin_lock_irqsave(&x->wait.lock, flags);
4437 x->done += UINT_MAX/2;
4438 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4439 spin_unlock_irqrestore(&x->wait.lock, flags);
4441 EXPORT_SYMBOL(complete_all);
4443 static inline long __sched
4444 do_wait_for_common(struct completion *x, long timeout, int state)
4447 DECLARE_WAITQUEUE(wait, current);
4449 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4451 if (signal_pending_state(state, current)) {
4452 timeout = -ERESTARTSYS;
4455 __set_current_state(state);
4456 spin_unlock_irq(&x->wait.lock);
4457 timeout = schedule_timeout(timeout);
4458 spin_lock_irq(&x->wait.lock);
4459 } while (!x->done && timeout);
4460 __remove_wait_queue(&x->wait, &wait);
4465 return timeout ?: 1;
4469 wait_for_common(struct completion *x, long timeout, int state)
4473 spin_lock_irq(&x->wait.lock);
4474 timeout = do_wait_for_common(x, timeout, state);
4475 spin_unlock_irq(&x->wait.lock);
4480 * wait_for_completion: - waits for completion of a task
4481 * @x: holds the state of this particular completion
4483 * This waits to be signaled for completion of a specific task. It is NOT
4484 * interruptible and there is no timeout.
4486 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4487 * and interrupt capability. Also see complete().
4489 void __sched wait_for_completion(struct completion *x)
4491 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4493 EXPORT_SYMBOL(wait_for_completion);
4496 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4497 * @x: holds the state of this particular completion
4498 * @timeout: timeout value in jiffies
4500 * This waits for either a completion of a specific task to be signaled or for a
4501 * specified timeout to expire. The timeout is in jiffies. It is not
4504 unsigned long __sched
4505 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4507 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4509 EXPORT_SYMBOL(wait_for_completion_timeout);
4512 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4513 * @x: holds the state of this particular completion
4515 * This waits for completion of a specific task to be signaled. It is
4518 int __sched wait_for_completion_interruptible(struct completion *x)
4520 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4521 if (t == -ERESTARTSYS)
4525 EXPORT_SYMBOL(wait_for_completion_interruptible);
4528 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4529 * @x: holds the state of this particular completion
4530 * @timeout: timeout value in jiffies
4532 * This waits for either a completion of a specific task to be signaled or for a
4533 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4536 wait_for_completion_interruptible_timeout(struct completion *x,
4537 unsigned long timeout)
4539 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4541 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4544 * wait_for_completion_killable: - waits for completion of a task (killable)
4545 * @x: holds the state of this particular completion
4547 * This waits to be signaled for completion of a specific task. It can be
4548 * interrupted by a kill signal.
4550 int __sched wait_for_completion_killable(struct completion *x)
4552 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4553 if (t == -ERESTARTSYS)
4557 EXPORT_SYMBOL(wait_for_completion_killable);
4560 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4561 * @x: holds the state of this particular completion
4562 * @timeout: timeout value in jiffies
4564 * This waits for either a completion of a specific task to be
4565 * signaled or for a specified timeout to expire. It can be
4566 * interrupted by a kill signal. The timeout is in jiffies.
4569 wait_for_completion_killable_timeout(struct completion *x,
4570 unsigned long timeout)
4572 return wait_for_common(x, timeout, TASK_KILLABLE);
4574 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4577 * try_wait_for_completion - try to decrement a completion without blocking
4578 * @x: completion structure
4580 * Returns: 0 if a decrement cannot be done without blocking
4581 * 1 if a decrement succeeded.
4583 * If a completion is being used as a counting completion,
4584 * attempt to decrement the counter without blocking. This
4585 * enables us to avoid waiting if the resource the completion
4586 * is protecting is not available.
4588 bool try_wait_for_completion(struct completion *x)
4590 unsigned long flags;
4593 spin_lock_irqsave(&x->wait.lock, flags);
4598 spin_unlock_irqrestore(&x->wait.lock, flags);
4601 EXPORT_SYMBOL(try_wait_for_completion);
4604 * completion_done - Test to see if a completion has any waiters
4605 * @x: completion structure
4607 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4608 * 1 if there are no waiters.
4611 bool completion_done(struct completion *x)
4613 unsigned long flags;
4616 spin_lock_irqsave(&x->wait.lock, flags);
4619 spin_unlock_irqrestore(&x->wait.lock, flags);
4622 EXPORT_SYMBOL(completion_done);
4625 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4627 unsigned long flags;
4630 init_waitqueue_entry(&wait, current);
4632 __set_current_state(state);
4634 spin_lock_irqsave(&q->lock, flags);
4635 __add_wait_queue(q, &wait);
4636 spin_unlock(&q->lock);
4637 timeout = schedule_timeout(timeout);
4638 spin_lock_irq(&q->lock);
4639 __remove_wait_queue(q, &wait);
4640 spin_unlock_irqrestore(&q->lock, flags);
4645 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4647 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4649 EXPORT_SYMBOL(interruptible_sleep_on);
4652 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4654 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4656 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4658 void __sched sleep_on(wait_queue_head_t *q)
4660 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4662 EXPORT_SYMBOL(sleep_on);
4664 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4666 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4668 EXPORT_SYMBOL(sleep_on_timeout);
4670 #ifdef CONFIG_RT_MUTEXES
4673 * rt_mutex_setprio - set the current priority of a task
4675 * @prio: prio value (kernel-internal form)
4677 * This function changes the 'effective' priority of a task. It does
4678 * not touch ->normal_prio like __setscheduler().
4680 * Used by the rt_mutex code to implement priority inheritance logic.
4682 void rt_mutex_setprio(struct task_struct *p, int prio)
4684 unsigned long flags;
4685 int oldprio, on_rq, running;
4687 const struct sched_class *prev_class;
4689 BUG_ON(prio < 0 || prio > MAX_PRIO);
4691 rq = task_rq_lock(p, &flags);
4693 trace_sched_pi_setprio(p, prio);
4695 prev_class = p->sched_class;
4696 on_rq = p->se.on_rq;
4697 running = task_current(rq, p);
4699 dequeue_task(rq, p, 0);
4701 p->sched_class->put_prev_task(rq, p);
4704 p->sched_class = &rt_sched_class;
4706 p->sched_class = &fair_sched_class;
4711 p->sched_class->set_curr_task(rq);
4713 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4715 check_class_changed(rq, p, prev_class, oldprio);
4716 task_rq_unlock(rq, &flags);
4721 void set_user_nice(struct task_struct *p, long nice)
4723 int old_prio, delta, on_rq;
4724 unsigned long flags;
4727 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4730 * We have to be careful, if called from sys_setpriority(),
4731 * the task might be in the middle of scheduling on another CPU.
4733 rq = task_rq_lock(p, &flags);
4735 * The RT priorities are set via sched_setscheduler(), but we still
4736 * allow the 'normal' nice value to be set - but as expected
4737 * it wont have any effect on scheduling until the task is
4738 * SCHED_FIFO/SCHED_RR:
4740 if (task_has_rt_policy(p)) {
4741 p->static_prio = NICE_TO_PRIO(nice);
4744 on_rq = p->se.on_rq;
4746 dequeue_task(rq, p, 0);
4748 p->static_prio = NICE_TO_PRIO(nice);
4751 p->prio = effective_prio(p);
4752 delta = p->prio - old_prio;
4755 enqueue_task(rq, p, 0);
4757 * If the task increased its priority or is running and
4758 * lowered its priority, then reschedule its CPU:
4760 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4761 resched_task(rq->curr);
4764 task_rq_unlock(rq, &flags);
4766 EXPORT_SYMBOL(set_user_nice);
4769 * can_nice - check if a task can reduce its nice value
4773 int can_nice(const struct task_struct *p, const int nice)
4775 /* convert nice value [19,-20] to rlimit style value [1,40] */
4776 int nice_rlim = 20 - nice;
4778 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4779 capable(CAP_SYS_NICE));
4782 #ifdef __ARCH_WANT_SYS_NICE
4785 * sys_nice - change the priority of the current process.
4786 * @increment: priority increment
4788 * sys_setpriority is a more generic, but much slower function that
4789 * does similar things.
4791 SYSCALL_DEFINE1(nice, int, increment)
4796 * Setpriority might change our priority at the same moment.
4797 * We don't have to worry. Conceptually one call occurs first
4798 * and we have a single winner.
4800 if (increment < -40)
4805 nice = TASK_NICE(current) + increment;
4811 if (increment < 0 && !can_nice(current, nice))
4814 retval = security_task_setnice(current, nice);
4818 set_user_nice(current, nice);
4825 * task_prio - return the priority value of a given task.
4826 * @p: the task in question.
4828 * This is the priority value as seen by users in /proc.
4829 * RT tasks are offset by -200. Normal tasks are centered
4830 * around 0, value goes from -16 to +15.
4832 int task_prio(const struct task_struct *p)
4834 return p->prio - MAX_RT_PRIO;
4838 * task_nice - return the nice value of a given task.
4839 * @p: the task in question.
4841 int task_nice(const struct task_struct *p)
4843 return TASK_NICE(p);
4845 EXPORT_SYMBOL(task_nice);
4848 * idle_cpu - is a given cpu idle currently?
4849 * @cpu: the processor in question.
4851 int idle_cpu(int cpu)
4853 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4857 * idle_task - return the idle task for a given cpu.
4858 * @cpu: the processor in question.
4860 struct task_struct *idle_task(int cpu)
4862 return cpu_rq(cpu)->idle;
4866 * find_process_by_pid - find a process with a matching PID value.
4867 * @pid: the pid in question.
4869 static struct task_struct *find_process_by_pid(pid_t pid)
4871 return pid ? find_task_by_vpid(pid) : current;
4874 /* Actually do priority change: must hold rq lock. */
4876 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4878 BUG_ON(p->se.on_rq);
4881 p->rt_priority = prio;
4882 p->normal_prio = normal_prio(p);
4883 /* we are holding p->pi_lock already */
4884 p->prio = rt_mutex_getprio(p);
4885 if (rt_prio(p->prio))
4886 p->sched_class = &rt_sched_class;
4888 p->sched_class = &fair_sched_class;
4893 * check the target process has a UID that matches the current process's
4895 static bool check_same_owner(struct task_struct *p)
4897 const struct cred *cred = current_cred(), *pcred;
4901 pcred = __task_cred(p);
4902 match = (cred->euid == pcred->euid ||
4903 cred->euid == pcred->uid);
4908 static int __sched_setscheduler(struct task_struct *p, int policy,
4909 const struct sched_param *param, bool user)
4911 int retval, oldprio, oldpolicy = -1, on_rq, running;
4912 unsigned long flags;
4913 const struct sched_class *prev_class;
4917 /* may grab non-irq protected spin_locks */
4918 BUG_ON(in_interrupt());
4920 /* double check policy once rq lock held */
4922 reset_on_fork = p->sched_reset_on_fork;
4923 policy = oldpolicy = p->policy;
4925 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4926 policy &= ~SCHED_RESET_ON_FORK;
4928 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4929 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4930 policy != SCHED_IDLE)
4935 * Valid priorities for SCHED_FIFO and SCHED_RR are
4936 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4937 * SCHED_BATCH and SCHED_IDLE is 0.
4939 if (param->sched_priority < 0 ||
4940 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4941 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4943 if (rt_policy(policy) != (param->sched_priority != 0))
4947 * Allow unprivileged RT tasks to decrease priority:
4949 if (user && !capable(CAP_SYS_NICE)) {
4950 if (rt_policy(policy)) {
4951 unsigned long rlim_rtprio =
4952 task_rlimit(p, RLIMIT_RTPRIO);
4954 /* can't set/change the rt policy */
4955 if (policy != p->policy && !rlim_rtprio)
4958 /* can't increase priority */
4959 if (param->sched_priority > p->rt_priority &&
4960 param->sched_priority > rlim_rtprio)
4965 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4966 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4968 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4969 if (!can_nice(p, TASK_NICE(p)))
4973 /* can't change other user's priorities */
4974 if (!check_same_owner(p))
4977 /* Normal users shall not reset the sched_reset_on_fork flag */
4978 if (p->sched_reset_on_fork && !reset_on_fork)
4983 retval = security_task_setscheduler(p);
4989 * make sure no PI-waiters arrive (or leave) while we are
4990 * changing the priority of the task:
4992 raw_spin_lock_irqsave(&p->pi_lock, flags);
4994 * To be able to change p->policy safely, the apropriate
4995 * runqueue lock must be held.
4997 rq = __task_rq_lock(p);
5000 * Changing the policy of the stop threads its a very bad idea
5002 if (p == rq->stop) {
5003 __task_rq_unlock(rq);
5004 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5008 #ifdef CONFIG_RT_GROUP_SCHED
5011 * Do not allow realtime tasks into groups that have no runtime
5014 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5015 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5016 !task_group_is_autogroup(task_group(p))) {
5017 __task_rq_unlock(rq);
5018 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5024 /* recheck policy now with rq lock held */
5025 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5026 policy = oldpolicy = -1;
5027 __task_rq_unlock(rq);
5028 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5031 on_rq = p->se.on_rq;
5032 running = task_current(rq, p);
5034 deactivate_task(rq, p, 0);
5036 p->sched_class->put_prev_task(rq, p);
5038 p->sched_reset_on_fork = reset_on_fork;
5041 prev_class = p->sched_class;
5042 __setscheduler(rq, p, policy, param->sched_priority);
5045 p->sched_class->set_curr_task(rq);
5047 activate_task(rq, p, 0);
5049 check_class_changed(rq, p, prev_class, oldprio);
5050 __task_rq_unlock(rq);
5051 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5053 rt_mutex_adjust_pi(p);
5059 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5060 * @p: the task in question.
5061 * @policy: new policy.
5062 * @param: structure containing the new RT priority.
5064 * NOTE that the task may be already dead.
5066 int sched_setscheduler(struct task_struct *p, int policy,
5067 const struct sched_param *param)
5069 return __sched_setscheduler(p, policy, param, true);
5071 EXPORT_SYMBOL_GPL(sched_setscheduler);
5074 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5075 * @p: the task in question.
5076 * @policy: new policy.
5077 * @param: structure containing the new RT priority.
5079 * Just like sched_setscheduler, only don't bother checking if the
5080 * current context has permission. For example, this is needed in
5081 * stop_machine(): we create temporary high priority worker threads,
5082 * but our caller might not have that capability.
5084 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5085 const struct sched_param *param)
5087 return __sched_setscheduler(p, policy, param, false);
5091 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5093 struct sched_param lparam;
5094 struct task_struct *p;
5097 if (!param || pid < 0)
5099 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5104 p = find_process_by_pid(pid);
5106 retval = sched_setscheduler(p, policy, &lparam);
5113 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5114 * @pid: the pid in question.
5115 * @policy: new policy.
5116 * @param: structure containing the new RT priority.
5118 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5119 struct sched_param __user *, param)
5121 /* negative values for policy are not valid */
5125 return do_sched_setscheduler(pid, policy, param);
5129 * sys_sched_setparam - set/change the RT priority of a thread
5130 * @pid: the pid in question.
5131 * @param: structure containing the new RT priority.
5133 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5135 return do_sched_setscheduler(pid, -1, param);
5139 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5140 * @pid: the pid in question.
5142 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5144 struct task_struct *p;
5152 p = find_process_by_pid(pid);
5154 retval = security_task_getscheduler(p);
5157 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5164 * sys_sched_getparam - get the RT priority of a thread
5165 * @pid: the pid in question.
5166 * @param: structure containing the RT priority.
5168 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5170 struct sched_param lp;
5171 struct task_struct *p;
5174 if (!param || pid < 0)
5178 p = find_process_by_pid(pid);
5183 retval = security_task_getscheduler(p);
5187 lp.sched_priority = p->rt_priority;
5191 * This one might sleep, we cannot do it with a spinlock held ...
5193 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5202 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5204 cpumask_var_t cpus_allowed, new_mask;
5205 struct task_struct *p;
5211 p = find_process_by_pid(pid);
5218 /* Prevent p going away */
5222 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5226 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5228 goto out_free_cpus_allowed;
5231 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5234 retval = security_task_setscheduler(p);
5238 cpuset_cpus_allowed(p, cpus_allowed);
5239 cpumask_and(new_mask, in_mask, cpus_allowed);
5241 retval = set_cpus_allowed_ptr(p, new_mask);
5244 cpuset_cpus_allowed(p, cpus_allowed);
5245 if (!cpumask_subset(new_mask, cpus_allowed)) {
5247 * We must have raced with a concurrent cpuset
5248 * update. Just reset the cpus_allowed to the
5249 * cpuset's cpus_allowed
5251 cpumask_copy(new_mask, cpus_allowed);
5256 free_cpumask_var(new_mask);
5257 out_free_cpus_allowed:
5258 free_cpumask_var(cpus_allowed);
5265 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5266 struct cpumask *new_mask)
5268 if (len < cpumask_size())
5269 cpumask_clear(new_mask);
5270 else if (len > cpumask_size())
5271 len = cpumask_size();
5273 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5277 * sys_sched_setaffinity - set the cpu affinity of a process
5278 * @pid: pid of the process
5279 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5280 * @user_mask_ptr: user-space pointer to the new cpu mask
5282 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5283 unsigned long __user *, user_mask_ptr)
5285 cpumask_var_t new_mask;
5288 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5291 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5293 retval = sched_setaffinity(pid, new_mask);
5294 free_cpumask_var(new_mask);
5298 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5300 struct task_struct *p;
5301 unsigned long flags;
5309 p = find_process_by_pid(pid);
5313 retval = security_task_getscheduler(p);
5317 rq = task_rq_lock(p, &flags);
5318 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5319 task_rq_unlock(rq, &flags);
5329 * sys_sched_getaffinity - get the cpu affinity of a process
5330 * @pid: pid of the process
5331 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5332 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5334 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5335 unsigned long __user *, user_mask_ptr)
5340 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5342 if (len & (sizeof(unsigned long)-1))
5345 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5348 ret = sched_getaffinity(pid, mask);
5350 size_t retlen = min_t(size_t, len, cpumask_size());
5352 if (copy_to_user(user_mask_ptr, mask, retlen))
5357 free_cpumask_var(mask);
5363 * sys_sched_yield - yield the current processor to other threads.
5365 * This function yields the current CPU to other tasks. If there are no
5366 * other threads running on this CPU then this function will return.
5368 SYSCALL_DEFINE0(sched_yield)
5370 struct rq *rq = this_rq_lock();
5372 schedstat_inc(rq, yld_count);
5373 current->sched_class->yield_task(rq);
5376 * Since we are going to call schedule() anyway, there's
5377 * no need to preempt or enable interrupts:
5379 __release(rq->lock);
5380 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5381 do_raw_spin_unlock(&rq->lock);
5382 preempt_enable_no_resched();
5389 static inline int should_resched(void)
5391 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5394 static void __cond_resched(void)
5396 add_preempt_count(PREEMPT_ACTIVE);
5398 sub_preempt_count(PREEMPT_ACTIVE);
5401 int __sched _cond_resched(void)
5403 if (should_resched()) {
5409 EXPORT_SYMBOL(_cond_resched);
5412 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5413 * call schedule, and on return reacquire the lock.
5415 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5416 * operations here to prevent schedule() from being called twice (once via
5417 * spin_unlock(), once by hand).
5419 int __cond_resched_lock(spinlock_t *lock)
5421 int resched = should_resched();
5424 lockdep_assert_held(lock);
5426 if (spin_needbreak(lock) || resched) {
5437 EXPORT_SYMBOL(__cond_resched_lock);
5439 int __sched __cond_resched_softirq(void)
5441 BUG_ON(!in_softirq());
5443 if (should_resched()) {
5451 EXPORT_SYMBOL(__cond_resched_softirq);
5454 * yield - yield the current processor to other threads.
5456 * This is a shortcut for kernel-space yielding - it marks the
5457 * thread runnable and calls sys_sched_yield().
5459 void __sched yield(void)
5461 set_current_state(TASK_RUNNING);
5464 EXPORT_SYMBOL(yield);
5467 * yield_to - yield the current processor to another thread in
5468 * your thread group, or accelerate that thread toward the
5469 * processor it's on.
5471 * It's the caller's job to ensure that the target task struct
5472 * can't go away on us before we can do any checks.
5474 * Returns true if we indeed boosted the target task.
5476 bool __sched yield_to(struct task_struct *p, bool preempt)
5478 struct task_struct *curr = current;
5479 struct rq *rq, *p_rq;
5480 unsigned long flags;
5483 local_irq_save(flags);
5488 double_rq_lock(rq, p_rq);
5489 while (task_rq(p) != p_rq) {
5490 double_rq_unlock(rq, p_rq);
5494 if (!curr->sched_class->yield_to_task)
5497 if (curr->sched_class != p->sched_class)
5500 if (task_running(p_rq, p) || p->state)
5503 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5505 schedstat_inc(rq, yld_count);
5507 * Make p's CPU reschedule; pick_next_entity takes care of
5510 if (preempt && rq != p_rq)
5511 resched_task(p_rq->curr);
5515 double_rq_unlock(rq, p_rq);
5516 local_irq_restore(flags);
5523 EXPORT_SYMBOL_GPL(yield_to);
5526 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5527 * that process accounting knows that this is a task in IO wait state.
5529 void __sched io_schedule(void)
5531 struct rq *rq = raw_rq();
5533 delayacct_blkio_start();
5534 atomic_inc(&rq->nr_iowait);
5535 current->in_iowait = 1;
5537 current->in_iowait = 0;
5538 atomic_dec(&rq->nr_iowait);
5539 delayacct_blkio_end();
5541 EXPORT_SYMBOL(io_schedule);
5543 long __sched io_schedule_timeout(long timeout)
5545 struct rq *rq = raw_rq();
5548 delayacct_blkio_start();
5549 atomic_inc(&rq->nr_iowait);
5550 current->in_iowait = 1;
5551 ret = schedule_timeout(timeout);
5552 current->in_iowait = 0;
5553 atomic_dec(&rq->nr_iowait);
5554 delayacct_blkio_end();
5559 * sys_sched_get_priority_max - return maximum RT priority.
5560 * @policy: scheduling class.
5562 * this syscall returns the maximum rt_priority that can be used
5563 * by a given scheduling class.
5565 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5572 ret = MAX_USER_RT_PRIO-1;
5584 * sys_sched_get_priority_min - return minimum RT priority.
5585 * @policy: scheduling class.
5587 * this syscall returns the minimum rt_priority that can be used
5588 * by a given scheduling class.
5590 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5608 * sys_sched_rr_get_interval - return the default timeslice of a process.
5609 * @pid: pid of the process.
5610 * @interval: userspace pointer to the timeslice value.
5612 * this syscall writes the default timeslice value of a given process
5613 * into the user-space timespec buffer. A value of '0' means infinity.
5615 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5616 struct timespec __user *, interval)
5618 struct task_struct *p;
5619 unsigned int time_slice;
5620 unsigned long flags;
5630 p = find_process_by_pid(pid);
5634 retval = security_task_getscheduler(p);
5638 rq = task_rq_lock(p, &flags);
5639 time_slice = p->sched_class->get_rr_interval(rq, p);
5640 task_rq_unlock(rq, &flags);
5643 jiffies_to_timespec(time_slice, &t);
5644 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5652 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5654 void sched_show_task(struct task_struct *p)
5656 unsigned long free = 0;
5659 state = p->state ? __ffs(p->state) + 1 : 0;
5660 printk(KERN_INFO "%-15.15s %c", p->comm,
5661 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5662 #if BITS_PER_LONG == 32
5663 if (state == TASK_RUNNING)
5664 printk(KERN_CONT " running ");
5666 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5668 if (state == TASK_RUNNING)
5669 printk(KERN_CONT " running task ");
5671 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5673 #ifdef CONFIG_DEBUG_STACK_USAGE
5674 free = stack_not_used(p);
5676 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5677 task_pid_nr(p), task_pid_nr(p->real_parent),
5678 (unsigned long)task_thread_info(p)->flags);
5680 show_stack(p, NULL);
5683 void show_state_filter(unsigned long state_filter)
5685 struct task_struct *g, *p;
5687 #if BITS_PER_LONG == 32
5689 " task PC stack pid father\n");
5692 " task PC stack pid father\n");
5694 read_lock(&tasklist_lock);
5695 do_each_thread(g, p) {
5697 * reset the NMI-timeout, listing all files on a slow
5698 * console might take alot of time:
5700 touch_nmi_watchdog();
5701 if (!state_filter || (p->state & state_filter))
5703 } while_each_thread(g, p);
5705 touch_all_softlockup_watchdogs();
5707 #ifdef CONFIG_SCHED_DEBUG
5708 sysrq_sched_debug_show();
5710 read_unlock(&tasklist_lock);
5712 * Only show locks if all tasks are dumped:
5715 debug_show_all_locks();
5718 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5720 idle->sched_class = &idle_sched_class;
5724 * init_idle - set up an idle thread for a given CPU
5725 * @idle: task in question
5726 * @cpu: cpu the idle task belongs to
5728 * NOTE: this function does not set the idle thread's NEED_RESCHED
5729 * flag, to make booting more robust.
5731 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5733 struct rq *rq = cpu_rq(cpu);
5734 unsigned long flags;
5736 raw_spin_lock_irqsave(&rq->lock, flags);
5739 idle->state = TASK_RUNNING;
5740 idle->se.exec_start = sched_clock();
5742 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5744 * We're having a chicken and egg problem, even though we are
5745 * holding rq->lock, the cpu isn't yet set to this cpu so the
5746 * lockdep check in task_group() will fail.
5748 * Similar case to sched_fork(). / Alternatively we could
5749 * use task_rq_lock() here and obtain the other rq->lock.
5754 __set_task_cpu(idle, cpu);
5757 rq->curr = rq->idle = idle;
5758 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5761 raw_spin_unlock_irqrestore(&rq->lock, flags);
5763 /* Set the preempt count _outside_ the spinlocks! */
5764 #if defined(CONFIG_PREEMPT)
5765 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5767 task_thread_info(idle)->preempt_count = 0;
5770 * The idle tasks have their own, simple scheduling class:
5772 idle->sched_class = &idle_sched_class;
5773 ftrace_graph_init_idle_task(idle, cpu);
5777 * In a system that switches off the HZ timer nohz_cpu_mask
5778 * indicates which cpus entered this state. This is used
5779 * in the rcu update to wait only for active cpus. For system
5780 * which do not switch off the HZ timer nohz_cpu_mask should
5781 * always be CPU_BITS_NONE.
5783 cpumask_var_t nohz_cpu_mask;
5786 * Increase the granularity value when there are more CPUs,
5787 * because with more CPUs the 'effective latency' as visible
5788 * to users decreases. But the relationship is not linear,
5789 * so pick a second-best guess by going with the log2 of the
5792 * This idea comes from the SD scheduler of Con Kolivas:
5794 static int get_update_sysctl_factor(void)
5796 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5797 unsigned int factor;
5799 switch (sysctl_sched_tunable_scaling) {
5800 case SCHED_TUNABLESCALING_NONE:
5803 case SCHED_TUNABLESCALING_LINEAR:
5806 case SCHED_TUNABLESCALING_LOG:
5808 factor = 1 + ilog2(cpus);
5815 static void update_sysctl(void)
5817 unsigned int factor = get_update_sysctl_factor();
5819 #define SET_SYSCTL(name) \
5820 (sysctl_##name = (factor) * normalized_sysctl_##name)
5821 SET_SYSCTL(sched_min_granularity);
5822 SET_SYSCTL(sched_latency);
5823 SET_SYSCTL(sched_wakeup_granularity);
5827 static inline void sched_init_granularity(void)
5834 * This is how migration works:
5836 * 1) we invoke migration_cpu_stop() on the target CPU using
5838 * 2) stopper starts to run (implicitly forcing the migrated thread
5840 * 3) it checks whether the migrated task is still in the wrong runqueue.
5841 * 4) if it's in the wrong runqueue then the migration thread removes
5842 * it and puts it into the right queue.
5843 * 5) stopper completes and stop_one_cpu() returns and the migration
5848 * Change a given task's CPU affinity. Migrate the thread to a
5849 * proper CPU and schedule it away if the CPU it's executing on
5850 * is removed from the allowed bitmask.
5852 * NOTE: the caller must have a valid reference to the task, the
5853 * task must not exit() & deallocate itself prematurely. The
5854 * call is not atomic; no spinlocks may be held.
5856 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5858 unsigned long flags;
5860 unsigned int dest_cpu;
5864 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5865 * drop the rq->lock and still rely on ->cpus_allowed.
5868 while (task_is_waking(p))
5870 rq = task_rq_lock(p, &flags);
5871 if (task_is_waking(p)) {
5872 task_rq_unlock(rq, &flags);
5876 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5881 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5882 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5887 if (p->sched_class->set_cpus_allowed)
5888 p->sched_class->set_cpus_allowed(p, new_mask);
5890 cpumask_copy(&p->cpus_allowed, new_mask);
5891 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5894 /* Can the task run on the task's current CPU? If so, we're done */
5895 if (cpumask_test_cpu(task_cpu(p), new_mask))
5898 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5899 if (migrate_task(p, rq)) {
5900 struct migration_arg arg = { p, dest_cpu };
5901 /* Need help from migration thread: drop lock and wait. */
5902 task_rq_unlock(rq, &flags);
5903 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5904 tlb_migrate_finish(p->mm);
5908 task_rq_unlock(rq, &flags);
5912 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5915 * Move (not current) task off this cpu, onto dest cpu. We're doing
5916 * this because either it can't run here any more (set_cpus_allowed()
5917 * away from this CPU, or CPU going down), or because we're
5918 * attempting to rebalance this task on exec (sched_exec).
5920 * So we race with normal scheduler movements, but that's OK, as long
5921 * as the task is no longer on this CPU.
5923 * Returns non-zero if task was successfully migrated.
5925 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5927 struct rq *rq_dest, *rq_src;
5930 if (unlikely(!cpu_active(dest_cpu)))
5933 rq_src = cpu_rq(src_cpu);
5934 rq_dest = cpu_rq(dest_cpu);
5936 double_rq_lock(rq_src, rq_dest);
5937 /* Already moved. */
5938 if (task_cpu(p) != src_cpu)
5940 /* Affinity changed (again). */
5941 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5945 * If we're not on a rq, the next wake-up will ensure we're
5949 deactivate_task(rq_src, p, 0);
5950 set_task_cpu(p, dest_cpu);
5951 activate_task(rq_dest, p, 0);
5952 check_preempt_curr(rq_dest, p, 0);
5957 double_rq_unlock(rq_src, rq_dest);
5962 * migration_cpu_stop - this will be executed by a highprio stopper thread
5963 * and performs thread migration by bumping thread off CPU then
5964 * 'pushing' onto another runqueue.
5966 static int migration_cpu_stop(void *data)
5968 struct migration_arg *arg = data;
5971 * The original target cpu might have gone down and we might
5972 * be on another cpu but it doesn't matter.
5974 local_irq_disable();
5975 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5980 #ifdef CONFIG_HOTPLUG_CPU
5983 * Ensures that the idle task is using init_mm right before its cpu goes
5986 void idle_task_exit(void)
5988 struct mm_struct *mm = current->active_mm;
5990 BUG_ON(cpu_online(smp_processor_id()));
5993 switch_mm(mm, &init_mm, current);
5998 * While a dead CPU has no uninterruptible tasks queued at this point,
5999 * it might still have a nonzero ->nr_uninterruptible counter, because
6000 * for performance reasons the counter is not stricly tracking tasks to
6001 * their home CPUs. So we just add the counter to another CPU's counter,
6002 * to keep the global sum constant after CPU-down:
6004 static void migrate_nr_uninterruptible(struct rq *rq_src)
6006 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6008 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6009 rq_src->nr_uninterruptible = 0;
6013 * remove the tasks which were accounted by rq from calc_load_tasks.
6015 static void calc_global_load_remove(struct rq *rq)
6017 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6018 rq->calc_load_active = 0;
6022 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6023 * try_to_wake_up()->select_task_rq().
6025 * Called with rq->lock held even though we'er in stop_machine() and
6026 * there's no concurrency possible, we hold the required locks anyway
6027 * because of lock validation efforts.
6029 static void migrate_tasks(unsigned int dead_cpu)
6031 struct rq *rq = cpu_rq(dead_cpu);
6032 struct task_struct *next, *stop = rq->stop;
6036 * Fudge the rq selection such that the below task selection loop
6037 * doesn't get stuck on the currently eligible stop task.
6039 * We're currently inside stop_machine() and the rq is either stuck
6040 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6041 * either way we should never end up calling schedule() until we're
6048 * There's this thread running, bail when that's the only
6051 if (rq->nr_running == 1)
6054 next = pick_next_task(rq);
6056 next->sched_class->put_prev_task(rq, next);
6058 /* Find suitable destination for @next, with force if needed. */
6059 dest_cpu = select_fallback_rq(dead_cpu, next);
6060 raw_spin_unlock(&rq->lock);
6062 __migrate_task(next, dead_cpu, dest_cpu);
6064 raw_spin_lock(&rq->lock);
6070 #endif /* CONFIG_HOTPLUG_CPU */
6072 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6074 static struct ctl_table sd_ctl_dir[] = {
6076 .procname = "sched_domain",
6082 static struct ctl_table sd_ctl_root[] = {
6084 .procname = "kernel",
6086 .child = sd_ctl_dir,
6091 static struct ctl_table *sd_alloc_ctl_entry(int n)
6093 struct ctl_table *entry =
6094 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6099 static void sd_free_ctl_entry(struct ctl_table **tablep)
6101 struct ctl_table *entry;
6104 * In the intermediate directories, both the child directory and
6105 * procname are dynamically allocated and could fail but the mode
6106 * will always be set. In the lowest directory the names are
6107 * static strings and all have proc handlers.
6109 for (entry = *tablep; entry->mode; entry++) {
6111 sd_free_ctl_entry(&entry->child);
6112 if (entry->proc_handler == NULL)
6113 kfree(entry->procname);
6121 set_table_entry(struct ctl_table *entry,
6122 const char *procname, void *data, int maxlen,
6123 mode_t mode, proc_handler *proc_handler)
6125 entry->procname = procname;
6127 entry->maxlen = maxlen;
6129 entry->proc_handler = proc_handler;
6132 static struct ctl_table *
6133 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6135 struct ctl_table *table = sd_alloc_ctl_entry(13);
6140 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6141 sizeof(long), 0644, proc_doulongvec_minmax);
6142 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6143 sizeof(long), 0644, proc_doulongvec_minmax);
6144 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6145 sizeof(int), 0644, proc_dointvec_minmax);
6146 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6147 sizeof(int), 0644, proc_dointvec_minmax);
6148 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6149 sizeof(int), 0644, proc_dointvec_minmax);
6150 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6151 sizeof(int), 0644, proc_dointvec_minmax);
6152 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6153 sizeof(int), 0644, proc_dointvec_minmax);
6154 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6155 sizeof(int), 0644, proc_dointvec_minmax);
6156 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6157 sizeof(int), 0644, proc_dointvec_minmax);
6158 set_table_entry(&table[9], "cache_nice_tries",
6159 &sd->cache_nice_tries,
6160 sizeof(int), 0644, proc_dointvec_minmax);
6161 set_table_entry(&table[10], "flags", &sd->flags,
6162 sizeof(int), 0644, proc_dointvec_minmax);
6163 set_table_entry(&table[11], "name", sd->name,
6164 CORENAME_MAX_SIZE, 0444, proc_dostring);
6165 /* &table[12] is terminator */
6170 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6172 struct ctl_table *entry, *table;
6173 struct sched_domain *sd;
6174 int domain_num = 0, i;
6177 for_each_domain(cpu, sd)
6179 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6184 for_each_domain(cpu, sd) {
6185 snprintf(buf, 32, "domain%d", i);
6186 entry->procname = kstrdup(buf, GFP_KERNEL);
6188 entry->child = sd_alloc_ctl_domain_table(sd);
6195 static struct ctl_table_header *sd_sysctl_header;
6196 static void register_sched_domain_sysctl(void)
6198 int i, cpu_num = num_possible_cpus();
6199 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6202 WARN_ON(sd_ctl_dir[0].child);
6203 sd_ctl_dir[0].child = entry;
6208 for_each_possible_cpu(i) {
6209 snprintf(buf, 32, "cpu%d", i);
6210 entry->procname = kstrdup(buf, GFP_KERNEL);
6212 entry->child = sd_alloc_ctl_cpu_table(i);
6216 WARN_ON(sd_sysctl_header);
6217 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6220 /* may be called multiple times per register */
6221 static void unregister_sched_domain_sysctl(void)
6223 if (sd_sysctl_header)
6224 unregister_sysctl_table(sd_sysctl_header);
6225 sd_sysctl_header = NULL;
6226 if (sd_ctl_dir[0].child)
6227 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6230 static void register_sched_domain_sysctl(void)
6233 static void unregister_sched_domain_sysctl(void)
6238 static void set_rq_online(struct rq *rq)
6241 const struct sched_class *class;
6243 cpumask_set_cpu(rq->cpu, rq->rd->online);
6246 for_each_class(class) {
6247 if (class->rq_online)
6248 class->rq_online(rq);
6253 static void set_rq_offline(struct rq *rq)
6256 const struct sched_class *class;
6258 for_each_class(class) {
6259 if (class->rq_offline)
6260 class->rq_offline(rq);
6263 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6269 * migration_call - callback that gets triggered when a CPU is added.
6270 * Here we can start up the necessary migration thread for the new CPU.
6272 static int __cpuinit
6273 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6275 int cpu = (long)hcpu;
6276 unsigned long flags;
6277 struct rq *rq = cpu_rq(cpu);
6279 switch (action & ~CPU_TASKS_FROZEN) {
6281 case CPU_UP_PREPARE:
6282 rq->calc_load_update = calc_load_update;
6286 /* Update our root-domain */
6287 raw_spin_lock_irqsave(&rq->lock, flags);
6289 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6293 raw_spin_unlock_irqrestore(&rq->lock, flags);
6296 #ifdef CONFIG_HOTPLUG_CPU
6298 /* Update our root-domain */
6299 raw_spin_lock_irqsave(&rq->lock, flags);
6301 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6305 BUG_ON(rq->nr_running != 1); /* the migration thread */
6306 raw_spin_unlock_irqrestore(&rq->lock, flags);
6308 migrate_nr_uninterruptible(rq);
6309 calc_global_load_remove(rq);
6317 * Register at high priority so that task migration (migrate_all_tasks)
6318 * happens before everything else. This has to be lower priority than
6319 * the notifier in the perf_event subsystem, though.
6321 static struct notifier_block __cpuinitdata migration_notifier = {
6322 .notifier_call = migration_call,
6323 .priority = CPU_PRI_MIGRATION,
6326 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6327 unsigned long action, void *hcpu)
6329 switch (action & ~CPU_TASKS_FROZEN) {
6331 case CPU_DOWN_FAILED:
6332 set_cpu_active((long)hcpu, true);
6339 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6340 unsigned long action, void *hcpu)
6342 switch (action & ~CPU_TASKS_FROZEN) {
6343 case CPU_DOWN_PREPARE:
6344 set_cpu_active((long)hcpu, false);
6351 static int __init migration_init(void)
6353 void *cpu = (void *)(long)smp_processor_id();
6356 /* Initialize migration for the boot CPU */
6357 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6358 BUG_ON(err == NOTIFY_BAD);
6359 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6360 register_cpu_notifier(&migration_notifier);
6362 /* Register cpu active notifiers */
6363 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6364 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6368 early_initcall(migration_init);
6373 #ifdef CONFIG_SCHED_DEBUG
6375 static __read_mostly int sched_domain_debug_enabled;
6377 static int __init sched_domain_debug_setup(char *str)
6379 sched_domain_debug_enabled = 1;
6383 early_param("sched_debug", sched_domain_debug_setup);
6385 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6386 struct cpumask *groupmask)
6388 struct sched_group *group = sd->groups;
6391 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6392 cpumask_clear(groupmask);
6394 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6396 if (!(sd->flags & SD_LOAD_BALANCE)) {
6397 printk("does not load-balance\n");
6399 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6404 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6406 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6407 printk(KERN_ERR "ERROR: domain->span does not contain "
6410 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6411 printk(KERN_ERR "ERROR: domain->groups does not contain"
6415 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6419 printk(KERN_ERR "ERROR: group is NULL\n");
6423 if (!group->cpu_power) {
6424 printk(KERN_CONT "\n");
6425 printk(KERN_ERR "ERROR: domain->cpu_power not "
6430 if (!cpumask_weight(sched_group_cpus(group))) {
6431 printk(KERN_CONT "\n");
6432 printk(KERN_ERR "ERROR: empty group\n");
6436 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6437 printk(KERN_CONT "\n");
6438 printk(KERN_ERR "ERROR: repeated CPUs\n");
6442 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6444 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6446 printk(KERN_CONT " %s", str);
6447 if (group->cpu_power != SCHED_LOAD_SCALE) {
6448 printk(KERN_CONT " (cpu_power = %d)",
6452 group = group->next;
6453 } while (group != sd->groups);
6454 printk(KERN_CONT "\n");
6456 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6457 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6460 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6461 printk(KERN_ERR "ERROR: parent span is not a superset "
6462 "of domain->span\n");
6466 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6468 cpumask_var_t groupmask;
6471 if (!sched_domain_debug_enabled)
6475 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6479 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6481 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6482 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6487 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6494 free_cpumask_var(groupmask);
6496 #else /* !CONFIG_SCHED_DEBUG */
6497 # define sched_domain_debug(sd, cpu) do { } while (0)
6498 #endif /* CONFIG_SCHED_DEBUG */
6500 static int sd_degenerate(struct sched_domain *sd)
6502 if (cpumask_weight(sched_domain_span(sd)) == 1)
6505 /* Following flags need at least 2 groups */
6506 if (sd->flags & (SD_LOAD_BALANCE |
6507 SD_BALANCE_NEWIDLE |
6511 SD_SHARE_PKG_RESOURCES)) {
6512 if (sd->groups != sd->groups->next)
6516 /* Following flags don't use groups */
6517 if (sd->flags & (SD_WAKE_AFFINE))
6524 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6526 unsigned long cflags = sd->flags, pflags = parent->flags;
6528 if (sd_degenerate(parent))
6531 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6534 /* Flags needing groups don't count if only 1 group in parent */
6535 if (parent->groups == parent->groups->next) {
6536 pflags &= ~(SD_LOAD_BALANCE |
6537 SD_BALANCE_NEWIDLE |
6541 SD_SHARE_PKG_RESOURCES);
6542 if (nr_node_ids == 1)
6543 pflags &= ~SD_SERIALIZE;
6545 if (~cflags & pflags)
6551 static void free_rootdomain(struct root_domain *rd)
6553 synchronize_sched();
6555 cpupri_cleanup(&rd->cpupri);
6557 free_cpumask_var(rd->rto_mask);
6558 free_cpumask_var(rd->online);
6559 free_cpumask_var(rd->span);
6563 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6565 struct root_domain *old_rd = NULL;
6566 unsigned long flags;
6568 raw_spin_lock_irqsave(&rq->lock, flags);
6573 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6576 cpumask_clear_cpu(rq->cpu, old_rd->span);
6579 * If we dont want to free the old_rt yet then
6580 * set old_rd to NULL to skip the freeing later
6583 if (!atomic_dec_and_test(&old_rd->refcount))
6587 atomic_inc(&rd->refcount);
6590 cpumask_set_cpu(rq->cpu, rd->span);
6591 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6594 raw_spin_unlock_irqrestore(&rq->lock, flags);
6597 free_rootdomain(old_rd);
6600 static int init_rootdomain(struct root_domain *rd)
6602 memset(rd, 0, sizeof(*rd));
6604 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6606 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6608 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6611 if (cpupri_init(&rd->cpupri) != 0)
6616 free_cpumask_var(rd->rto_mask);
6618 free_cpumask_var(rd->online);
6620 free_cpumask_var(rd->span);
6625 static void init_defrootdomain(void)
6627 init_rootdomain(&def_root_domain);
6629 atomic_set(&def_root_domain.refcount, 1);
6632 static struct root_domain *alloc_rootdomain(void)
6634 struct root_domain *rd;
6636 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6640 if (init_rootdomain(rd) != 0) {
6649 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6650 * hold the hotplug lock.
6653 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6655 struct rq *rq = cpu_rq(cpu);
6656 struct sched_domain *tmp;
6658 for (tmp = sd; tmp; tmp = tmp->parent)
6659 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6661 /* Remove the sched domains which do not contribute to scheduling. */
6662 for (tmp = sd; tmp; ) {
6663 struct sched_domain *parent = tmp->parent;
6667 if (sd_parent_degenerate(tmp, parent)) {
6668 tmp->parent = parent->parent;
6670 parent->parent->child = tmp;
6675 if (sd && sd_degenerate(sd)) {
6681 sched_domain_debug(sd, cpu);
6683 rq_attach_root(rq, rd);
6684 rcu_assign_pointer(rq->sd, sd);
6687 /* cpus with isolated domains */
6688 static cpumask_var_t cpu_isolated_map;
6690 /* Setup the mask of cpus configured for isolated domains */
6691 static int __init isolated_cpu_setup(char *str)
6693 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6694 cpulist_parse(str, cpu_isolated_map);
6698 __setup("isolcpus=", isolated_cpu_setup);
6701 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6702 * to a function which identifies what group(along with sched group) a CPU
6703 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6704 * (due to the fact that we keep track of groups covered with a struct cpumask).
6706 * init_sched_build_groups will build a circular linked list of the groups
6707 * covered by the given span, and will set each group's ->cpumask correctly,
6708 * and ->cpu_power to 0.
6711 init_sched_build_groups(const struct cpumask *span,
6712 const struct cpumask *cpu_map,
6713 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6714 struct sched_group **sg,
6715 struct cpumask *tmpmask),
6716 struct cpumask *covered, struct cpumask *tmpmask)
6718 struct sched_group *first = NULL, *last = NULL;
6721 cpumask_clear(covered);
6723 for_each_cpu(i, span) {
6724 struct sched_group *sg;
6725 int group = group_fn(i, cpu_map, &sg, tmpmask);
6728 if (cpumask_test_cpu(i, covered))
6731 cpumask_clear(sched_group_cpus(sg));
6734 for_each_cpu(j, span) {
6735 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6738 cpumask_set_cpu(j, covered);
6739 cpumask_set_cpu(j, sched_group_cpus(sg));
6750 #define SD_NODES_PER_DOMAIN 16
6755 * find_next_best_node - find the next node to include in a sched_domain
6756 * @node: node whose sched_domain we're building
6757 * @used_nodes: nodes already in the sched_domain
6759 * Find the next node to include in a given scheduling domain. Simply
6760 * finds the closest node not already in the @used_nodes map.
6762 * Should use nodemask_t.
6764 static int find_next_best_node(int node, nodemask_t *used_nodes)
6766 int i, n, val, min_val, best_node = 0;
6770 for (i = 0; i < nr_node_ids; i++) {
6771 /* Start at @node */
6772 n = (node + i) % nr_node_ids;
6774 if (!nr_cpus_node(n))
6777 /* Skip already used nodes */
6778 if (node_isset(n, *used_nodes))
6781 /* Simple min distance search */
6782 val = node_distance(node, n);
6784 if (val < min_val) {
6790 node_set(best_node, *used_nodes);
6795 * sched_domain_node_span - get a cpumask for a node's sched_domain
6796 * @node: node whose cpumask we're constructing
6797 * @span: resulting cpumask
6799 * Given a node, construct a good cpumask for its sched_domain to span. It
6800 * should be one that prevents unnecessary balancing, but also spreads tasks
6803 static void sched_domain_node_span(int node, struct cpumask *span)
6805 nodemask_t used_nodes;
6808 cpumask_clear(span);
6809 nodes_clear(used_nodes);
6811 cpumask_or(span, span, cpumask_of_node(node));
6812 node_set(node, used_nodes);
6814 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6815 int next_node = find_next_best_node(node, &used_nodes);
6817 cpumask_or(span, span, cpumask_of_node(next_node));
6820 #endif /* CONFIG_NUMA */
6822 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6825 * The cpus mask in sched_group and sched_domain hangs off the end.
6827 * ( See the the comments in include/linux/sched.h:struct sched_group
6828 * and struct sched_domain. )
6830 struct static_sched_group {
6831 struct sched_group sg;
6832 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6835 struct static_sched_domain {
6836 struct sched_domain sd;
6837 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6843 cpumask_var_t domainspan;
6844 cpumask_var_t covered;
6845 cpumask_var_t notcovered;
6847 cpumask_var_t nodemask;
6848 cpumask_var_t this_sibling_map;
6849 cpumask_var_t this_core_map;
6850 cpumask_var_t this_book_map;
6851 cpumask_var_t send_covered;
6852 cpumask_var_t tmpmask;
6853 struct sched_group **sched_group_nodes;
6854 struct root_domain *rd;
6858 sa_sched_groups = 0,
6864 sa_this_sibling_map,
6866 sa_sched_group_nodes,
6876 * SMT sched-domains:
6878 #ifdef CONFIG_SCHED_SMT
6879 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6880 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6883 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6884 struct sched_group **sg, struct cpumask *unused)
6887 *sg = &per_cpu(sched_groups, cpu).sg;
6890 #endif /* CONFIG_SCHED_SMT */
6893 * multi-core sched-domains:
6895 #ifdef CONFIG_SCHED_MC
6896 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6897 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6900 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6901 struct sched_group **sg, struct cpumask *mask)
6904 #ifdef CONFIG_SCHED_SMT
6905 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6906 group = cpumask_first(mask);
6911 *sg = &per_cpu(sched_group_core, group).sg;
6914 #endif /* CONFIG_SCHED_MC */
6917 * book sched-domains:
6919 #ifdef CONFIG_SCHED_BOOK
6920 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6921 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6924 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6925 struct sched_group **sg, struct cpumask *mask)
6928 #ifdef CONFIG_SCHED_MC
6929 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6930 group = cpumask_first(mask);
6931 #elif defined(CONFIG_SCHED_SMT)
6932 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6933 group = cpumask_first(mask);
6936 *sg = &per_cpu(sched_group_book, group).sg;
6939 #endif /* CONFIG_SCHED_BOOK */
6941 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6942 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6945 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6946 struct sched_group **sg, struct cpumask *mask)
6949 #ifdef CONFIG_SCHED_BOOK
6950 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6951 group = cpumask_first(mask);
6952 #elif defined(CONFIG_SCHED_MC)
6953 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6954 group = cpumask_first(mask);
6955 #elif defined(CONFIG_SCHED_SMT)
6956 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6957 group = cpumask_first(mask);
6962 *sg = &per_cpu(sched_group_phys, group).sg;
6968 * The init_sched_build_groups can't handle what we want to do with node
6969 * groups, so roll our own. Now each node has its own list of groups which
6970 * gets dynamically allocated.
6972 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6973 static struct sched_group ***sched_group_nodes_bycpu;
6975 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6976 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6978 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6979 struct sched_group **sg,
6980 struct cpumask *nodemask)
6984 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6985 group = cpumask_first(nodemask);
6988 *sg = &per_cpu(sched_group_allnodes, group).sg;
6992 static void init_numa_sched_groups_power(struct sched_group *group_head)
6994 struct sched_group *sg = group_head;
7000 for_each_cpu(j, sched_group_cpus(sg)) {
7001 struct sched_domain *sd;
7003 sd = &per_cpu(phys_domains, j).sd;
7004 if (j != group_first_cpu(sd->groups)) {
7006 * Only add "power" once for each
7012 sg->cpu_power += sd->groups->cpu_power;
7015 } while (sg != group_head);
7018 static int build_numa_sched_groups(struct s_data *d,
7019 const struct cpumask *cpu_map, int num)
7021 struct sched_domain *sd;
7022 struct sched_group *sg, *prev;
7025 cpumask_clear(d->covered);
7026 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
7027 if (cpumask_empty(d->nodemask)) {
7028 d->sched_group_nodes[num] = NULL;
7032 sched_domain_node_span(num, d->domainspan);
7033 cpumask_and(d->domainspan, d->domainspan, cpu_map);
7035 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7038 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
7042 d->sched_group_nodes[num] = sg;
7044 for_each_cpu(j, d->nodemask) {
7045 sd = &per_cpu(node_domains, j).sd;
7050 cpumask_copy(sched_group_cpus(sg), d->nodemask);
7052 cpumask_or(d->covered, d->covered, d->nodemask);
7055 for (j = 0; j < nr_node_ids; j++) {
7056 n = (num + j) % nr_node_ids;
7057 cpumask_complement(d->notcovered, d->covered);
7058 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
7059 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
7060 if (cpumask_empty(d->tmpmask))
7062 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
7063 if (cpumask_empty(d->tmpmask))
7065 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7069 "Can not alloc domain group for node %d\n", j);
7073 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
7074 sg->next = prev->next;
7075 cpumask_or(d->covered, d->covered, d->tmpmask);
7082 #endif /* CONFIG_NUMA */
7085 /* Free memory allocated for various sched_group structures */
7086 static void free_sched_groups(const struct cpumask *cpu_map,
7087 struct cpumask *nodemask)
7091 for_each_cpu(cpu, cpu_map) {
7092 struct sched_group **sched_group_nodes
7093 = sched_group_nodes_bycpu[cpu];
7095 if (!sched_group_nodes)
7098 for (i = 0; i < nr_node_ids; i++) {
7099 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7101 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7102 if (cpumask_empty(nodemask))
7112 if (oldsg != sched_group_nodes[i])
7115 kfree(sched_group_nodes);
7116 sched_group_nodes_bycpu[cpu] = NULL;
7119 #else /* !CONFIG_NUMA */
7120 static void free_sched_groups(const struct cpumask *cpu_map,
7121 struct cpumask *nodemask)
7124 #endif /* CONFIG_NUMA */
7127 * Initialize sched groups cpu_power.
7129 * cpu_power indicates the capacity of sched group, which is used while
7130 * distributing the load between different sched groups in a sched domain.
7131 * Typically cpu_power for all the groups in a sched domain will be same unless
7132 * there are asymmetries in the topology. If there are asymmetries, group
7133 * having more cpu_power will pickup more load compared to the group having
7136 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7138 struct sched_domain *child;
7139 struct sched_group *group;
7143 WARN_ON(!sd || !sd->groups);
7145 if (cpu != group_first_cpu(sd->groups))
7148 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7152 sd->groups->cpu_power = 0;
7155 power = SCHED_LOAD_SCALE;
7156 weight = cpumask_weight(sched_domain_span(sd));
7158 * SMT siblings share the power of a single core.
7159 * Usually multiple threads get a better yield out of
7160 * that one core than a single thread would have,
7161 * reflect that in sd->smt_gain.
7163 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
7164 power *= sd->smt_gain;
7166 power >>= SCHED_LOAD_SHIFT;
7168 sd->groups->cpu_power += power;
7173 * Add cpu_power of each child group to this groups cpu_power.
7175 group = child->groups;
7177 sd->groups->cpu_power += group->cpu_power;
7178 group = group->next;
7179 } while (group != child->groups);
7183 * Initializers for schedule domains
7184 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7187 #ifdef CONFIG_SCHED_DEBUG
7188 # define SD_INIT_NAME(sd, type) sd->name = #type
7190 # define SD_INIT_NAME(sd, type) do { } while (0)
7193 #define SD_INIT(sd, type) sd_init_##type(sd)
7195 #define SD_INIT_FUNC(type) \
7196 static noinline void sd_init_##type(struct sched_domain *sd) \
7198 memset(sd, 0, sizeof(*sd)); \
7199 *sd = SD_##type##_INIT; \
7200 sd->level = SD_LV_##type; \
7201 SD_INIT_NAME(sd, type); \
7206 SD_INIT_FUNC(ALLNODES)
7209 #ifdef CONFIG_SCHED_SMT
7210 SD_INIT_FUNC(SIBLING)
7212 #ifdef CONFIG_SCHED_MC
7215 #ifdef CONFIG_SCHED_BOOK
7219 static int default_relax_domain_level = -1;
7221 static int __init setup_relax_domain_level(char *str)
7225 val = simple_strtoul(str, NULL, 0);
7226 if (val < SD_LV_MAX)
7227 default_relax_domain_level = val;
7231 __setup("relax_domain_level=", setup_relax_domain_level);
7233 static void set_domain_attribute(struct sched_domain *sd,
7234 struct sched_domain_attr *attr)
7238 if (!attr || attr->relax_domain_level < 0) {
7239 if (default_relax_domain_level < 0)
7242 request = default_relax_domain_level;
7244 request = attr->relax_domain_level;
7245 if (request < sd->level) {
7246 /* turn off idle balance on this domain */
7247 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7249 /* turn on idle balance on this domain */
7250 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7254 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7255 const struct cpumask *cpu_map)
7258 case sa_sched_groups:
7259 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7260 d->sched_group_nodes = NULL;
7262 free_rootdomain(d->rd); /* fall through */
7264 free_cpumask_var(d->tmpmask); /* fall through */
7265 case sa_send_covered:
7266 free_cpumask_var(d->send_covered); /* fall through */
7267 case sa_this_book_map:
7268 free_cpumask_var(d->this_book_map); /* fall through */
7269 case sa_this_core_map:
7270 free_cpumask_var(d->this_core_map); /* fall through */
7271 case sa_this_sibling_map:
7272 free_cpumask_var(d->this_sibling_map); /* fall through */
7274 free_cpumask_var(d->nodemask); /* fall through */
7275 case sa_sched_group_nodes:
7277 kfree(d->sched_group_nodes); /* fall through */
7279 free_cpumask_var(d->notcovered); /* fall through */
7281 free_cpumask_var(d->covered); /* fall through */
7283 free_cpumask_var(d->domainspan); /* fall through */
7290 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7291 const struct cpumask *cpu_map)
7294 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7296 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7297 return sa_domainspan;
7298 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7300 /* Allocate the per-node list of sched groups */
7301 d->sched_group_nodes = kcalloc(nr_node_ids,
7302 sizeof(struct sched_group *), GFP_KERNEL);
7303 if (!d->sched_group_nodes) {
7304 printk(KERN_WARNING "Can not alloc sched group node list\n");
7305 return sa_notcovered;
7307 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7309 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7310 return sa_sched_group_nodes;
7311 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7313 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7314 return sa_this_sibling_map;
7315 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7316 return sa_this_core_map;
7317 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7318 return sa_this_book_map;
7319 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7320 return sa_send_covered;
7321 d->rd = alloc_rootdomain();
7323 printk(KERN_WARNING "Cannot alloc root domain\n");
7326 return sa_rootdomain;
7329 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7330 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7332 struct sched_domain *sd = NULL;
7334 struct sched_domain *parent;
7337 if (cpumask_weight(cpu_map) >
7338 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7339 sd = &per_cpu(allnodes_domains, i).sd;
7340 SD_INIT(sd, ALLNODES);
7341 set_domain_attribute(sd, attr);
7342 cpumask_copy(sched_domain_span(sd), cpu_map);
7343 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7348 sd = &per_cpu(node_domains, i).sd;
7350 set_domain_attribute(sd, attr);
7351 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7352 sd->parent = parent;
7355 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7360 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7361 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7362 struct sched_domain *parent, int i)
7364 struct sched_domain *sd;
7365 sd = &per_cpu(phys_domains, i).sd;
7367 set_domain_attribute(sd, attr);
7368 cpumask_copy(sched_domain_span(sd), d->nodemask);
7369 sd->parent = parent;
7372 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7376 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7377 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7378 struct sched_domain *parent, int i)
7380 struct sched_domain *sd = parent;
7381 #ifdef CONFIG_SCHED_BOOK
7382 sd = &per_cpu(book_domains, i).sd;
7384 set_domain_attribute(sd, attr);
7385 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7386 sd->parent = parent;
7388 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7393 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7394 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7395 struct sched_domain *parent, int i)
7397 struct sched_domain *sd = parent;
7398 #ifdef CONFIG_SCHED_MC
7399 sd = &per_cpu(core_domains, i).sd;
7401 set_domain_attribute(sd, attr);
7402 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7403 sd->parent = parent;
7405 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7410 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7411 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7412 struct sched_domain *parent, int i)
7414 struct sched_domain *sd = parent;
7415 #ifdef CONFIG_SCHED_SMT
7416 sd = &per_cpu(cpu_domains, i).sd;
7417 SD_INIT(sd, SIBLING);
7418 set_domain_attribute(sd, attr);
7419 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7420 sd->parent = parent;
7422 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7427 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7428 const struct cpumask *cpu_map, int cpu)
7431 #ifdef CONFIG_SCHED_SMT
7432 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7433 cpumask_and(d->this_sibling_map, cpu_map,
7434 topology_thread_cpumask(cpu));
7435 if (cpu == cpumask_first(d->this_sibling_map))
7436 init_sched_build_groups(d->this_sibling_map, cpu_map,
7438 d->send_covered, d->tmpmask);
7441 #ifdef CONFIG_SCHED_MC
7442 case SD_LV_MC: /* set up multi-core groups */
7443 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7444 if (cpu == cpumask_first(d->this_core_map))
7445 init_sched_build_groups(d->this_core_map, cpu_map,
7447 d->send_covered, d->tmpmask);
7450 #ifdef CONFIG_SCHED_BOOK
7451 case SD_LV_BOOK: /* set up book groups */
7452 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7453 if (cpu == cpumask_first(d->this_book_map))
7454 init_sched_build_groups(d->this_book_map, cpu_map,
7456 d->send_covered, d->tmpmask);
7459 case SD_LV_CPU: /* set up physical groups */
7460 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7461 if (!cpumask_empty(d->nodemask))
7462 init_sched_build_groups(d->nodemask, cpu_map,
7464 d->send_covered, d->tmpmask);
7467 case SD_LV_ALLNODES:
7468 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7469 d->send_covered, d->tmpmask);
7478 * Build sched domains for a given set of cpus and attach the sched domains
7479 * to the individual cpus
7481 static int __build_sched_domains(const struct cpumask *cpu_map,
7482 struct sched_domain_attr *attr)
7484 enum s_alloc alloc_state = sa_none;
7486 struct sched_domain *sd;
7492 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7493 if (alloc_state != sa_rootdomain)
7495 alloc_state = sa_sched_groups;
7498 * Set up domains for cpus specified by the cpu_map.
7500 for_each_cpu(i, cpu_map) {
7501 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7504 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7505 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7506 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7507 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7508 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7511 for_each_cpu(i, cpu_map) {
7512 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7513 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7514 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7517 /* Set up physical groups */
7518 for (i = 0; i < nr_node_ids; i++)
7519 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7522 /* Set up node groups */
7524 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7526 for (i = 0; i < nr_node_ids; i++)
7527 if (build_numa_sched_groups(&d, cpu_map, i))
7531 /* Calculate CPU power for physical packages and nodes */
7532 #ifdef CONFIG_SCHED_SMT
7533 for_each_cpu(i, cpu_map) {
7534 sd = &per_cpu(cpu_domains, i).sd;
7535 init_sched_groups_power(i, sd);
7538 #ifdef CONFIG_SCHED_MC
7539 for_each_cpu(i, cpu_map) {
7540 sd = &per_cpu(core_domains, i).sd;
7541 init_sched_groups_power(i, sd);
7544 #ifdef CONFIG_SCHED_BOOK
7545 for_each_cpu(i, cpu_map) {
7546 sd = &per_cpu(book_domains, i).sd;
7547 init_sched_groups_power(i, sd);
7551 for_each_cpu(i, cpu_map) {
7552 sd = &per_cpu(phys_domains, i).sd;
7553 init_sched_groups_power(i, sd);
7557 for (i = 0; i < nr_node_ids; i++)
7558 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7560 if (d.sd_allnodes) {
7561 struct sched_group *sg;
7563 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7565 init_numa_sched_groups_power(sg);
7569 /* Attach the domains */
7570 for_each_cpu(i, cpu_map) {
7571 #ifdef CONFIG_SCHED_SMT
7572 sd = &per_cpu(cpu_domains, i).sd;
7573 #elif defined(CONFIG_SCHED_MC)
7574 sd = &per_cpu(core_domains, i).sd;
7575 #elif defined(CONFIG_SCHED_BOOK)
7576 sd = &per_cpu(book_domains, i).sd;
7578 sd = &per_cpu(phys_domains, i).sd;
7580 cpu_attach_domain(sd, d.rd, i);
7583 d.sched_group_nodes = NULL; /* don't free this we still need it */
7584 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7588 __free_domain_allocs(&d, alloc_state, cpu_map);
7592 static int build_sched_domains(const struct cpumask *cpu_map)
7594 return __build_sched_domains(cpu_map, NULL);
7597 static cpumask_var_t *doms_cur; /* current sched domains */
7598 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7599 static struct sched_domain_attr *dattr_cur;
7600 /* attribues of custom domains in 'doms_cur' */
7603 * Special case: If a kmalloc of a doms_cur partition (array of
7604 * cpumask) fails, then fallback to a single sched domain,
7605 * as determined by the single cpumask fallback_doms.
7607 static cpumask_var_t fallback_doms;
7610 * arch_update_cpu_topology lets virtualized architectures update the
7611 * cpu core maps. It is supposed to return 1 if the topology changed
7612 * or 0 if it stayed the same.
7614 int __attribute__((weak)) arch_update_cpu_topology(void)
7619 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7622 cpumask_var_t *doms;
7624 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7627 for (i = 0; i < ndoms; i++) {
7628 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7629 free_sched_domains(doms, i);
7636 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7639 for (i = 0; i < ndoms; i++)
7640 free_cpumask_var(doms[i]);
7645 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7646 * For now this just excludes isolated cpus, but could be used to
7647 * exclude other special cases in the future.
7649 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7653 arch_update_cpu_topology();
7655 doms_cur = alloc_sched_domains(ndoms_cur);
7657 doms_cur = &fallback_doms;
7658 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7660 err = build_sched_domains(doms_cur[0]);
7661 register_sched_domain_sysctl();
7666 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7667 struct cpumask *tmpmask)
7669 free_sched_groups(cpu_map, tmpmask);
7673 * Detach sched domains from a group of cpus specified in cpu_map
7674 * These cpus will now be attached to the NULL domain
7676 static void detach_destroy_domains(const struct cpumask *cpu_map)
7678 /* Save because hotplug lock held. */
7679 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7682 for_each_cpu(i, cpu_map)
7683 cpu_attach_domain(NULL, &def_root_domain, i);
7684 synchronize_sched();
7685 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7688 /* handle null as "default" */
7689 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7690 struct sched_domain_attr *new, int idx_new)
7692 struct sched_domain_attr tmp;
7699 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7700 new ? (new + idx_new) : &tmp,
7701 sizeof(struct sched_domain_attr));
7705 * Partition sched domains as specified by the 'ndoms_new'
7706 * cpumasks in the array doms_new[] of cpumasks. This compares
7707 * doms_new[] to the current sched domain partitioning, doms_cur[].
7708 * It destroys each deleted domain and builds each new domain.
7710 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7711 * The masks don't intersect (don't overlap.) We should setup one
7712 * sched domain for each mask. CPUs not in any of the cpumasks will
7713 * not be load balanced. If the same cpumask appears both in the
7714 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7717 * The passed in 'doms_new' should be allocated using
7718 * alloc_sched_domains. This routine takes ownership of it and will
7719 * free_sched_domains it when done with it. If the caller failed the
7720 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7721 * and partition_sched_domains() will fallback to the single partition
7722 * 'fallback_doms', it also forces the domains to be rebuilt.
7724 * If doms_new == NULL it will be replaced with cpu_online_mask.
7725 * ndoms_new == 0 is a special case for destroying existing domains,
7726 * and it will not create the default domain.
7728 * Call with hotplug lock held
7730 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7731 struct sched_domain_attr *dattr_new)
7736 mutex_lock(&sched_domains_mutex);
7738 /* always unregister in case we don't destroy any domains */
7739 unregister_sched_domain_sysctl();
7741 /* Let architecture update cpu core mappings. */
7742 new_topology = arch_update_cpu_topology();
7744 n = doms_new ? ndoms_new : 0;
7746 /* Destroy deleted domains */
7747 for (i = 0; i < ndoms_cur; i++) {
7748 for (j = 0; j < n && !new_topology; j++) {
7749 if (cpumask_equal(doms_cur[i], doms_new[j])
7750 && dattrs_equal(dattr_cur, i, dattr_new, j))
7753 /* no match - a current sched domain not in new doms_new[] */
7754 detach_destroy_domains(doms_cur[i]);
7759 if (doms_new == NULL) {
7761 doms_new = &fallback_doms;
7762 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7763 WARN_ON_ONCE(dattr_new);
7766 /* Build new domains */
7767 for (i = 0; i < ndoms_new; i++) {
7768 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7769 if (cpumask_equal(doms_new[i], doms_cur[j])
7770 && dattrs_equal(dattr_new, i, dattr_cur, j))
7773 /* no match - add a new doms_new */
7774 __build_sched_domains(doms_new[i],
7775 dattr_new ? dattr_new + i : NULL);
7780 /* Remember the new sched domains */
7781 if (doms_cur != &fallback_doms)
7782 free_sched_domains(doms_cur, ndoms_cur);
7783 kfree(dattr_cur); /* kfree(NULL) is safe */
7784 doms_cur = doms_new;
7785 dattr_cur = dattr_new;
7786 ndoms_cur = ndoms_new;
7788 register_sched_domain_sysctl();
7790 mutex_unlock(&sched_domains_mutex);
7793 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7794 static void arch_reinit_sched_domains(void)
7798 /* Destroy domains first to force the rebuild */
7799 partition_sched_domains(0, NULL, NULL);
7801 rebuild_sched_domains();
7805 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7807 unsigned int level = 0;
7809 if (sscanf(buf, "%u", &level) != 1)
7813 * level is always be positive so don't check for
7814 * level < POWERSAVINGS_BALANCE_NONE which is 0
7815 * What happens on 0 or 1 byte write,
7816 * need to check for count as well?
7819 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7823 sched_smt_power_savings = level;
7825 sched_mc_power_savings = level;
7827 arch_reinit_sched_domains();
7832 #ifdef CONFIG_SCHED_MC
7833 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7834 struct sysdev_class_attribute *attr,
7837 return sprintf(page, "%u\n", sched_mc_power_savings);
7839 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7840 struct sysdev_class_attribute *attr,
7841 const char *buf, size_t count)
7843 return sched_power_savings_store(buf, count, 0);
7845 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7846 sched_mc_power_savings_show,
7847 sched_mc_power_savings_store);
7850 #ifdef CONFIG_SCHED_SMT
7851 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7852 struct sysdev_class_attribute *attr,
7855 return sprintf(page, "%u\n", sched_smt_power_savings);
7857 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7858 struct sysdev_class_attribute *attr,
7859 const char *buf, size_t count)
7861 return sched_power_savings_store(buf, count, 1);
7863 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7864 sched_smt_power_savings_show,
7865 sched_smt_power_savings_store);
7868 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7872 #ifdef CONFIG_SCHED_SMT
7874 err = sysfs_create_file(&cls->kset.kobj,
7875 &attr_sched_smt_power_savings.attr);
7877 #ifdef CONFIG_SCHED_MC
7878 if (!err && mc_capable())
7879 err = sysfs_create_file(&cls->kset.kobj,
7880 &attr_sched_mc_power_savings.attr);
7884 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7887 * Update cpusets according to cpu_active mask. If cpusets are
7888 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7889 * around partition_sched_domains().
7891 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7894 switch (action & ~CPU_TASKS_FROZEN) {
7896 case CPU_DOWN_FAILED:
7897 cpuset_update_active_cpus();
7904 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7907 switch (action & ~CPU_TASKS_FROZEN) {
7908 case CPU_DOWN_PREPARE:
7909 cpuset_update_active_cpus();
7916 static int update_runtime(struct notifier_block *nfb,
7917 unsigned long action, void *hcpu)
7919 int cpu = (int)(long)hcpu;
7922 case CPU_DOWN_PREPARE:
7923 case CPU_DOWN_PREPARE_FROZEN:
7924 disable_runtime(cpu_rq(cpu));
7927 case CPU_DOWN_FAILED:
7928 case CPU_DOWN_FAILED_FROZEN:
7930 case CPU_ONLINE_FROZEN:
7931 enable_runtime(cpu_rq(cpu));
7939 void __init sched_init_smp(void)
7941 cpumask_var_t non_isolated_cpus;
7943 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7944 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7946 #if defined(CONFIG_NUMA)
7947 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7949 BUG_ON(sched_group_nodes_bycpu == NULL);
7952 mutex_lock(&sched_domains_mutex);
7953 arch_init_sched_domains(cpu_active_mask);
7954 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7955 if (cpumask_empty(non_isolated_cpus))
7956 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7957 mutex_unlock(&sched_domains_mutex);
7960 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7961 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7963 /* RT runtime code needs to handle some hotplug events */
7964 hotcpu_notifier(update_runtime, 0);
7968 /* Move init over to a non-isolated CPU */
7969 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7971 sched_init_granularity();
7972 free_cpumask_var(non_isolated_cpus);
7974 init_sched_rt_class();
7977 void __init sched_init_smp(void)
7979 sched_init_granularity();
7981 #endif /* CONFIG_SMP */
7983 const_debug unsigned int sysctl_timer_migration = 1;
7985 int in_sched_functions(unsigned long addr)
7987 return in_lock_functions(addr) ||
7988 (addr >= (unsigned long)__sched_text_start
7989 && addr < (unsigned long)__sched_text_end);
7992 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7994 cfs_rq->tasks_timeline = RB_ROOT;
7995 INIT_LIST_HEAD(&cfs_rq->tasks);
7996 #ifdef CONFIG_FAIR_GROUP_SCHED
7998 /* allow initial update_cfs_load() to truncate */
8000 cfs_rq->load_stamp = 1;
8003 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8006 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8008 struct rt_prio_array *array;
8011 array = &rt_rq->active;
8012 for (i = 0; i < MAX_RT_PRIO; i++) {
8013 INIT_LIST_HEAD(array->queue + i);
8014 __clear_bit(i, array->bitmap);
8016 /* delimiter for bitsearch: */
8017 __set_bit(MAX_RT_PRIO, array->bitmap);
8019 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8020 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8022 rt_rq->highest_prio.next = MAX_RT_PRIO;
8026 rt_rq->rt_nr_migratory = 0;
8027 rt_rq->overloaded = 0;
8028 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
8032 rt_rq->rt_throttled = 0;
8033 rt_rq->rt_runtime = 0;
8034 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8036 #ifdef CONFIG_RT_GROUP_SCHED
8037 rt_rq->rt_nr_boosted = 0;
8042 #ifdef CONFIG_FAIR_GROUP_SCHED
8043 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8044 struct sched_entity *se, int cpu,
8045 struct sched_entity *parent)
8047 struct rq *rq = cpu_rq(cpu);
8048 tg->cfs_rq[cpu] = cfs_rq;
8049 init_cfs_rq(cfs_rq, rq);
8053 /* se could be NULL for root_task_group */
8058 se->cfs_rq = &rq->cfs;
8060 se->cfs_rq = parent->my_q;
8063 update_load_set(&se->load, 0);
8064 se->parent = parent;
8068 #ifdef CONFIG_RT_GROUP_SCHED
8069 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8070 struct sched_rt_entity *rt_se, int cpu,
8071 struct sched_rt_entity *parent)
8073 struct rq *rq = cpu_rq(cpu);
8075 tg->rt_rq[cpu] = rt_rq;
8076 init_rt_rq(rt_rq, rq);
8078 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8080 tg->rt_se[cpu] = rt_se;
8085 rt_se->rt_rq = &rq->rt;
8087 rt_se->rt_rq = parent->my_q;
8089 rt_se->my_q = rt_rq;
8090 rt_se->parent = parent;
8091 INIT_LIST_HEAD(&rt_se->run_list);
8095 void __init sched_init(void)
8098 unsigned long alloc_size = 0, ptr;
8100 #ifdef CONFIG_FAIR_GROUP_SCHED
8101 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8103 #ifdef CONFIG_RT_GROUP_SCHED
8104 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8106 #ifdef CONFIG_CPUMASK_OFFSTACK
8107 alloc_size += num_possible_cpus() * cpumask_size();
8110 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8112 #ifdef CONFIG_FAIR_GROUP_SCHED
8113 root_task_group.se = (struct sched_entity **)ptr;
8114 ptr += nr_cpu_ids * sizeof(void **);
8116 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8117 ptr += nr_cpu_ids * sizeof(void **);
8119 #endif /* CONFIG_FAIR_GROUP_SCHED */
8120 #ifdef CONFIG_RT_GROUP_SCHED
8121 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8122 ptr += nr_cpu_ids * sizeof(void **);
8124 root_task_group.rt_rq = (struct rt_rq **)ptr;
8125 ptr += nr_cpu_ids * sizeof(void **);
8127 #endif /* CONFIG_RT_GROUP_SCHED */
8128 #ifdef CONFIG_CPUMASK_OFFSTACK
8129 for_each_possible_cpu(i) {
8130 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8131 ptr += cpumask_size();
8133 #endif /* CONFIG_CPUMASK_OFFSTACK */
8137 init_defrootdomain();
8140 init_rt_bandwidth(&def_rt_bandwidth,
8141 global_rt_period(), global_rt_runtime());
8143 #ifdef CONFIG_RT_GROUP_SCHED
8144 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8145 global_rt_period(), global_rt_runtime());
8146 #endif /* CONFIG_RT_GROUP_SCHED */
8148 #ifdef CONFIG_CGROUP_SCHED
8149 list_add(&root_task_group.list, &task_groups);
8150 INIT_LIST_HEAD(&root_task_group.children);
8151 autogroup_init(&init_task);
8152 #endif /* CONFIG_CGROUP_SCHED */
8154 for_each_possible_cpu(i) {
8158 raw_spin_lock_init(&rq->lock);
8160 rq->calc_load_active = 0;
8161 rq->calc_load_update = jiffies + LOAD_FREQ;
8162 init_cfs_rq(&rq->cfs, rq);
8163 init_rt_rq(&rq->rt, rq);
8164 #ifdef CONFIG_FAIR_GROUP_SCHED
8165 root_task_group.shares = root_task_group_load;
8166 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8168 * How much cpu bandwidth does root_task_group get?
8170 * In case of task-groups formed thr' the cgroup filesystem, it
8171 * gets 100% of the cpu resources in the system. This overall
8172 * system cpu resource is divided among the tasks of
8173 * root_task_group and its child task-groups in a fair manner,
8174 * based on each entity's (task or task-group's) weight
8175 * (se->load.weight).
8177 * In other words, if root_task_group has 10 tasks of weight
8178 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8179 * then A0's share of the cpu resource is:
8181 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8183 * We achieve this by letting root_task_group's tasks sit
8184 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8186 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8187 #endif /* CONFIG_FAIR_GROUP_SCHED */
8189 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8190 #ifdef CONFIG_RT_GROUP_SCHED
8191 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8192 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8195 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8196 rq->cpu_load[j] = 0;
8198 rq->last_load_update_tick = jiffies;
8203 rq->cpu_power = SCHED_LOAD_SCALE;
8204 rq->post_schedule = 0;
8205 rq->active_balance = 0;
8206 rq->next_balance = jiffies;
8211 rq->avg_idle = 2*sysctl_sched_migration_cost;
8212 rq_attach_root(rq, &def_root_domain);
8214 rq->nohz_balance_kick = 0;
8215 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8219 atomic_set(&rq->nr_iowait, 0);
8222 set_load_weight(&init_task);
8224 #ifdef CONFIG_PREEMPT_NOTIFIERS
8225 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8229 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8232 #ifdef CONFIG_RT_MUTEXES
8233 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8237 * The boot idle thread does lazy MMU switching as well:
8239 atomic_inc(&init_mm.mm_count);
8240 enter_lazy_tlb(&init_mm, current);
8243 * Make us the idle thread. Technically, schedule() should not be
8244 * called from this thread, however somewhere below it might be,
8245 * but because we are the idle thread, we just pick up running again
8246 * when this runqueue becomes "idle".
8248 init_idle(current, smp_processor_id());
8250 calc_load_update = jiffies + LOAD_FREQ;
8253 * During early bootup we pretend to be a normal task:
8255 current->sched_class = &fair_sched_class;
8257 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8258 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8261 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8262 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8263 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8264 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8265 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8267 /* May be allocated at isolcpus cmdline parse time */
8268 if (cpu_isolated_map == NULL)
8269 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8272 scheduler_running = 1;
8275 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8276 static inline int preempt_count_equals(int preempt_offset)
8278 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8280 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
8283 void __might_sleep(const char *file, int line, int preempt_offset)
8286 static unsigned long prev_jiffy; /* ratelimiting */
8288 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8289 system_state != SYSTEM_RUNNING || oops_in_progress)
8291 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8293 prev_jiffy = jiffies;
8296 "BUG: sleeping function called from invalid context at %s:%d\n",
8299 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8300 in_atomic(), irqs_disabled(),
8301 current->pid, current->comm);
8303 debug_show_held_locks(current);
8304 if (irqs_disabled())
8305 print_irqtrace_events(current);
8309 EXPORT_SYMBOL(__might_sleep);
8312 #ifdef CONFIG_MAGIC_SYSRQ
8313 static void normalize_task(struct rq *rq, struct task_struct *p)
8315 const struct sched_class *prev_class = p->sched_class;
8316 int old_prio = p->prio;
8319 on_rq = p->se.on_rq;
8321 deactivate_task(rq, p, 0);
8322 __setscheduler(rq, p, SCHED_NORMAL, 0);
8324 activate_task(rq, p, 0);
8325 resched_task(rq->curr);
8328 check_class_changed(rq, p, prev_class, old_prio);
8331 void normalize_rt_tasks(void)
8333 struct task_struct *g, *p;
8334 unsigned long flags;
8337 read_lock_irqsave(&tasklist_lock, flags);
8338 do_each_thread(g, p) {
8340 * Only normalize user tasks:
8345 p->se.exec_start = 0;
8346 #ifdef CONFIG_SCHEDSTATS
8347 p->se.statistics.wait_start = 0;
8348 p->se.statistics.sleep_start = 0;
8349 p->se.statistics.block_start = 0;
8354 * Renice negative nice level userspace
8357 if (TASK_NICE(p) < 0 && p->mm)
8358 set_user_nice(p, 0);
8362 raw_spin_lock(&p->pi_lock);
8363 rq = __task_rq_lock(p);
8365 normalize_task(rq, p);
8367 __task_rq_unlock(rq);
8368 raw_spin_unlock(&p->pi_lock);
8369 } while_each_thread(g, p);
8371 read_unlock_irqrestore(&tasklist_lock, flags);
8374 #endif /* CONFIG_MAGIC_SYSRQ */
8376 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8378 * These functions are only useful for the IA64 MCA handling, or kdb.
8380 * They can only be called when the whole system has been
8381 * stopped - every CPU needs to be quiescent, and no scheduling
8382 * activity can take place. Using them for anything else would
8383 * be a serious bug, and as a result, they aren't even visible
8384 * under any other configuration.
8388 * curr_task - return the current task for a given cpu.
8389 * @cpu: the processor in question.
8391 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8393 struct task_struct *curr_task(int cpu)
8395 return cpu_curr(cpu);
8398 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8402 * set_curr_task - set the current task for a given cpu.
8403 * @cpu: the processor in question.
8404 * @p: the task pointer to set.
8406 * Description: This function must only be used when non-maskable interrupts
8407 * are serviced on a separate stack. It allows the architecture to switch the
8408 * notion of the current task on a cpu in a non-blocking manner. This function
8409 * must be called with all CPU's synchronized, and interrupts disabled, the
8410 * and caller must save the original value of the current task (see
8411 * curr_task() above) and restore that value before reenabling interrupts and
8412 * re-starting the system.
8414 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8416 void set_curr_task(int cpu, struct task_struct *p)
8423 #ifdef CONFIG_FAIR_GROUP_SCHED
8424 static void free_fair_sched_group(struct task_group *tg)
8428 for_each_possible_cpu(i) {
8430 kfree(tg->cfs_rq[i]);
8440 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8442 struct cfs_rq *cfs_rq;
8443 struct sched_entity *se;
8447 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8450 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8454 tg->shares = NICE_0_LOAD;
8456 for_each_possible_cpu(i) {
8459 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8460 GFP_KERNEL, cpu_to_node(i));
8464 se = kzalloc_node(sizeof(struct sched_entity),
8465 GFP_KERNEL, cpu_to_node(i));
8469 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8480 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8482 struct rq *rq = cpu_rq(cpu);
8483 unsigned long flags;
8486 * Only empty task groups can be destroyed; so we can speculatively
8487 * check on_list without danger of it being re-added.
8489 if (!tg->cfs_rq[cpu]->on_list)
8492 raw_spin_lock_irqsave(&rq->lock, flags);
8493 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8494 raw_spin_unlock_irqrestore(&rq->lock, flags);
8496 #else /* !CONFG_FAIR_GROUP_SCHED */
8497 static inline void free_fair_sched_group(struct task_group *tg)
8502 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8507 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8510 #endif /* CONFIG_FAIR_GROUP_SCHED */
8512 #ifdef CONFIG_RT_GROUP_SCHED
8513 static void free_rt_sched_group(struct task_group *tg)
8517 destroy_rt_bandwidth(&tg->rt_bandwidth);
8519 for_each_possible_cpu(i) {
8521 kfree(tg->rt_rq[i]);
8523 kfree(tg->rt_se[i]);
8531 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8533 struct rt_rq *rt_rq;
8534 struct sched_rt_entity *rt_se;
8538 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8541 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8545 init_rt_bandwidth(&tg->rt_bandwidth,
8546 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8548 for_each_possible_cpu(i) {
8551 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8552 GFP_KERNEL, cpu_to_node(i));
8556 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8557 GFP_KERNEL, cpu_to_node(i));
8561 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8571 #else /* !CONFIG_RT_GROUP_SCHED */
8572 static inline void free_rt_sched_group(struct task_group *tg)
8577 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8581 #endif /* CONFIG_RT_GROUP_SCHED */
8583 #ifdef CONFIG_CGROUP_SCHED
8584 static void free_sched_group(struct task_group *tg)
8586 free_fair_sched_group(tg);
8587 free_rt_sched_group(tg);
8592 /* allocate runqueue etc for a new task group */
8593 struct task_group *sched_create_group(struct task_group *parent)
8595 struct task_group *tg;
8596 unsigned long flags;
8598 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8600 return ERR_PTR(-ENOMEM);
8602 if (!alloc_fair_sched_group(tg, parent))
8605 if (!alloc_rt_sched_group(tg, parent))
8608 spin_lock_irqsave(&task_group_lock, flags);
8609 list_add_rcu(&tg->list, &task_groups);
8611 WARN_ON(!parent); /* root should already exist */
8613 tg->parent = parent;
8614 INIT_LIST_HEAD(&tg->children);
8615 list_add_rcu(&tg->siblings, &parent->children);
8616 spin_unlock_irqrestore(&task_group_lock, flags);
8621 free_sched_group(tg);
8622 return ERR_PTR(-ENOMEM);
8625 /* rcu callback to free various structures associated with a task group */
8626 static void free_sched_group_rcu(struct rcu_head *rhp)
8628 /* now it should be safe to free those cfs_rqs */
8629 free_sched_group(container_of(rhp, struct task_group, rcu));
8632 /* Destroy runqueue etc associated with a task group */
8633 void sched_destroy_group(struct task_group *tg)
8635 unsigned long flags;
8638 /* end participation in shares distribution */
8639 for_each_possible_cpu(i)
8640 unregister_fair_sched_group(tg, i);
8642 spin_lock_irqsave(&task_group_lock, flags);
8643 list_del_rcu(&tg->list);
8644 list_del_rcu(&tg->siblings);
8645 spin_unlock_irqrestore(&task_group_lock, flags);
8647 /* wait for possible concurrent references to cfs_rqs complete */
8648 call_rcu(&tg->rcu, free_sched_group_rcu);
8651 /* change task's runqueue when it moves between groups.
8652 * The caller of this function should have put the task in its new group
8653 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8654 * reflect its new group.
8656 void sched_move_task(struct task_struct *tsk)
8659 unsigned long flags;
8662 rq = task_rq_lock(tsk, &flags);
8664 running = task_current(rq, tsk);
8665 on_rq = tsk->se.on_rq;
8668 dequeue_task(rq, tsk, 0);
8669 if (unlikely(running))
8670 tsk->sched_class->put_prev_task(rq, tsk);
8672 #ifdef CONFIG_FAIR_GROUP_SCHED
8673 if (tsk->sched_class->task_move_group)
8674 tsk->sched_class->task_move_group(tsk, on_rq);
8677 set_task_rq(tsk, task_cpu(tsk));
8679 if (unlikely(running))
8680 tsk->sched_class->set_curr_task(rq);
8682 enqueue_task(rq, tsk, 0);
8684 task_rq_unlock(rq, &flags);
8686 #endif /* CONFIG_CGROUP_SCHED */
8688 #ifdef CONFIG_FAIR_GROUP_SCHED
8689 static DEFINE_MUTEX(shares_mutex);
8691 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8694 unsigned long flags;
8697 * We can't change the weight of the root cgroup.
8702 if (shares < MIN_SHARES)
8703 shares = MIN_SHARES;
8704 else if (shares > MAX_SHARES)
8705 shares = MAX_SHARES;
8707 mutex_lock(&shares_mutex);
8708 if (tg->shares == shares)
8711 tg->shares = shares;
8712 for_each_possible_cpu(i) {
8713 struct rq *rq = cpu_rq(i);
8714 struct sched_entity *se;
8717 /* Propagate contribution to hierarchy */
8718 raw_spin_lock_irqsave(&rq->lock, flags);
8719 for_each_sched_entity(se)
8720 update_cfs_shares(group_cfs_rq(se));
8721 raw_spin_unlock_irqrestore(&rq->lock, flags);
8725 mutex_unlock(&shares_mutex);
8729 unsigned long sched_group_shares(struct task_group *tg)
8735 #ifdef CONFIG_RT_GROUP_SCHED
8737 * Ensure that the real time constraints are schedulable.
8739 static DEFINE_MUTEX(rt_constraints_mutex);
8741 static unsigned long to_ratio(u64 period, u64 runtime)
8743 if (runtime == RUNTIME_INF)
8746 return div64_u64(runtime << 20, period);
8749 /* Must be called with tasklist_lock held */
8750 static inline int tg_has_rt_tasks(struct task_group *tg)
8752 struct task_struct *g, *p;
8754 do_each_thread(g, p) {
8755 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8757 } while_each_thread(g, p);
8762 struct rt_schedulable_data {
8763 struct task_group *tg;
8768 static int tg_schedulable(struct task_group *tg, void *data)
8770 struct rt_schedulable_data *d = data;
8771 struct task_group *child;
8772 unsigned long total, sum = 0;
8773 u64 period, runtime;
8775 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8776 runtime = tg->rt_bandwidth.rt_runtime;
8779 period = d->rt_period;
8780 runtime = d->rt_runtime;
8784 * Cannot have more runtime than the period.
8786 if (runtime > period && runtime != RUNTIME_INF)
8790 * Ensure we don't starve existing RT tasks.
8792 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8795 total = to_ratio(period, runtime);
8798 * Nobody can have more than the global setting allows.
8800 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8804 * The sum of our children's runtime should not exceed our own.
8806 list_for_each_entry_rcu(child, &tg->children, siblings) {
8807 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8808 runtime = child->rt_bandwidth.rt_runtime;
8810 if (child == d->tg) {
8811 period = d->rt_period;
8812 runtime = d->rt_runtime;
8815 sum += to_ratio(period, runtime);
8824 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8826 struct rt_schedulable_data data = {
8828 .rt_period = period,
8829 .rt_runtime = runtime,
8832 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8835 static int tg_set_bandwidth(struct task_group *tg,
8836 u64 rt_period, u64 rt_runtime)
8840 mutex_lock(&rt_constraints_mutex);
8841 read_lock(&tasklist_lock);
8842 err = __rt_schedulable(tg, rt_period, rt_runtime);
8846 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8847 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8848 tg->rt_bandwidth.rt_runtime = rt_runtime;
8850 for_each_possible_cpu(i) {
8851 struct rt_rq *rt_rq = tg->rt_rq[i];
8853 raw_spin_lock(&rt_rq->rt_runtime_lock);
8854 rt_rq->rt_runtime = rt_runtime;
8855 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8857 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8859 read_unlock(&tasklist_lock);
8860 mutex_unlock(&rt_constraints_mutex);
8865 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8867 u64 rt_runtime, rt_period;
8869 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8870 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8871 if (rt_runtime_us < 0)
8872 rt_runtime = RUNTIME_INF;
8874 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8877 long sched_group_rt_runtime(struct task_group *tg)
8881 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8884 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8885 do_div(rt_runtime_us, NSEC_PER_USEC);
8886 return rt_runtime_us;
8889 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8891 u64 rt_runtime, rt_period;
8893 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8894 rt_runtime = tg->rt_bandwidth.rt_runtime;
8899 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8902 long sched_group_rt_period(struct task_group *tg)
8906 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8907 do_div(rt_period_us, NSEC_PER_USEC);
8908 return rt_period_us;
8911 static int sched_rt_global_constraints(void)
8913 u64 runtime, period;
8916 if (sysctl_sched_rt_period <= 0)
8919 runtime = global_rt_runtime();
8920 period = global_rt_period();
8923 * Sanity check on the sysctl variables.
8925 if (runtime > period && runtime != RUNTIME_INF)
8928 mutex_lock(&rt_constraints_mutex);
8929 read_lock(&tasklist_lock);
8930 ret = __rt_schedulable(NULL, 0, 0);
8931 read_unlock(&tasklist_lock);
8932 mutex_unlock(&rt_constraints_mutex);
8937 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8939 /* Don't accept realtime tasks when there is no way for them to run */
8940 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8946 #else /* !CONFIG_RT_GROUP_SCHED */
8947 static int sched_rt_global_constraints(void)
8949 unsigned long flags;
8952 if (sysctl_sched_rt_period <= 0)
8956 * There's always some RT tasks in the root group
8957 * -- migration, kstopmachine etc..
8959 if (sysctl_sched_rt_runtime == 0)
8962 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8963 for_each_possible_cpu(i) {
8964 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8966 raw_spin_lock(&rt_rq->rt_runtime_lock);
8967 rt_rq->rt_runtime = global_rt_runtime();
8968 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8970 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8974 #endif /* CONFIG_RT_GROUP_SCHED */
8976 int sched_rt_handler(struct ctl_table *table, int write,
8977 void __user *buffer, size_t *lenp,
8981 int old_period, old_runtime;
8982 static DEFINE_MUTEX(mutex);
8985 old_period = sysctl_sched_rt_period;
8986 old_runtime = sysctl_sched_rt_runtime;
8988 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8990 if (!ret && write) {
8991 ret = sched_rt_global_constraints();
8993 sysctl_sched_rt_period = old_period;
8994 sysctl_sched_rt_runtime = old_runtime;
8996 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8997 def_rt_bandwidth.rt_period =
8998 ns_to_ktime(global_rt_period());
9001 mutex_unlock(&mutex);
9006 #ifdef CONFIG_CGROUP_SCHED
9008 /* return corresponding task_group object of a cgroup */
9009 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9011 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9012 struct task_group, css);
9015 static struct cgroup_subsys_state *
9016 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9018 struct task_group *tg, *parent;
9020 if (!cgrp->parent) {
9021 /* This is early initialization for the top cgroup */
9022 return &root_task_group.css;
9025 parent = cgroup_tg(cgrp->parent);
9026 tg = sched_create_group(parent);
9028 return ERR_PTR(-ENOMEM);
9034 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9036 struct task_group *tg = cgroup_tg(cgrp);
9038 sched_destroy_group(tg);
9042 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9044 #ifdef CONFIG_RT_GROUP_SCHED
9045 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9048 /* We don't support RT-tasks being in separate groups */
9049 if (tsk->sched_class != &fair_sched_class)
9056 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9057 struct task_struct *tsk, bool threadgroup)
9059 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
9063 struct task_struct *c;
9065 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9066 retval = cpu_cgroup_can_attach_task(cgrp, c);
9078 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9079 struct cgroup *old_cont, struct task_struct *tsk,
9082 sched_move_task(tsk);
9084 struct task_struct *c;
9086 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9094 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9095 struct cgroup *old_cgrp, struct task_struct *task)
9098 * cgroup_exit() is called in the copy_process() failure path.
9099 * Ignore this case since the task hasn't ran yet, this avoids
9100 * trying to poke a half freed task state from generic code.
9102 if (!(task->flags & PF_EXITING))
9105 sched_move_task(task);
9108 #ifdef CONFIG_FAIR_GROUP_SCHED
9109 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9112 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9115 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9117 struct task_group *tg = cgroup_tg(cgrp);
9119 return (u64) tg->shares;
9121 #endif /* CONFIG_FAIR_GROUP_SCHED */
9123 #ifdef CONFIG_RT_GROUP_SCHED
9124 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9127 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9130 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9132 return sched_group_rt_runtime(cgroup_tg(cgrp));
9135 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9138 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9141 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9143 return sched_group_rt_period(cgroup_tg(cgrp));
9145 #endif /* CONFIG_RT_GROUP_SCHED */
9147 static struct cftype cpu_files[] = {
9148 #ifdef CONFIG_FAIR_GROUP_SCHED
9151 .read_u64 = cpu_shares_read_u64,
9152 .write_u64 = cpu_shares_write_u64,
9155 #ifdef CONFIG_RT_GROUP_SCHED
9157 .name = "rt_runtime_us",
9158 .read_s64 = cpu_rt_runtime_read,
9159 .write_s64 = cpu_rt_runtime_write,
9162 .name = "rt_period_us",
9163 .read_u64 = cpu_rt_period_read_uint,
9164 .write_u64 = cpu_rt_period_write_uint,
9169 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9171 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9174 struct cgroup_subsys cpu_cgroup_subsys = {
9176 .create = cpu_cgroup_create,
9177 .destroy = cpu_cgroup_destroy,
9178 .can_attach = cpu_cgroup_can_attach,
9179 .attach = cpu_cgroup_attach,
9180 .exit = cpu_cgroup_exit,
9181 .populate = cpu_cgroup_populate,
9182 .subsys_id = cpu_cgroup_subsys_id,
9186 #endif /* CONFIG_CGROUP_SCHED */
9188 #ifdef CONFIG_CGROUP_CPUACCT
9191 * CPU accounting code for task groups.
9193 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9194 * (balbir@in.ibm.com).
9197 /* track cpu usage of a group of tasks and its child groups */
9199 struct cgroup_subsys_state css;
9200 /* cpuusage holds pointer to a u64-type object on every cpu */
9201 u64 __percpu *cpuusage;
9202 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9203 struct cpuacct *parent;
9206 struct cgroup_subsys cpuacct_subsys;
9208 /* return cpu accounting group corresponding to this container */
9209 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9211 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9212 struct cpuacct, css);
9215 /* return cpu accounting group to which this task belongs */
9216 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9218 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9219 struct cpuacct, css);
9222 /* create a new cpu accounting group */
9223 static struct cgroup_subsys_state *cpuacct_create(
9224 struct cgroup_subsys *ss, struct cgroup *cgrp)
9226 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9232 ca->cpuusage = alloc_percpu(u64);
9236 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9237 if (percpu_counter_init(&ca->cpustat[i], 0))
9238 goto out_free_counters;
9241 ca->parent = cgroup_ca(cgrp->parent);
9247 percpu_counter_destroy(&ca->cpustat[i]);
9248 free_percpu(ca->cpuusage);
9252 return ERR_PTR(-ENOMEM);
9255 /* destroy an existing cpu accounting group */
9257 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9259 struct cpuacct *ca = cgroup_ca(cgrp);
9262 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9263 percpu_counter_destroy(&ca->cpustat[i]);
9264 free_percpu(ca->cpuusage);
9268 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9270 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9273 #ifndef CONFIG_64BIT
9275 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9277 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9279 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9287 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9289 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9291 #ifndef CONFIG_64BIT
9293 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9295 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9297 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9303 /* return total cpu usage (in nanoseconds) of a group */
9304 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9306 struct cpuacct *ca = cgroup_ca(cgrp);
9307 u64 totalcpuusage = 0;
9310 for_each_present_cpu(i)
9311 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9313 return totalcpuusage;
9316 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9319 struct cpuacct *ca = cgroup_ca(cgrp);
9328 for_each_present_cpu(i)
9329 cpuacct_cpuusage_write(ca, i, 0);
9335 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9338 struct cpuacct *ca = cgroup_ca(cgroup);
9342 for_each_present_cpu(i) {
9343 percpu = cpuacct_cpuusage_read(ca, i);
9344 seq_printf(m, "%llu ", (unsigned long long) percpu);
9346 seq_printf(m, "\n");
9350 static const char *cpuacct_stat_desc[] = {
9351 [CPUACCT_STAT_USER] = "user",
9352 [CPUACCT_STAT_SYSTEM] = "system",
9355 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9356 struct cgroup_map_cb *cb)
9358 struct cpuacct *ca = cgroup_ca(cgrp);
9361 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9362 s64 val = percpu_counter_read(&ca->cpustat[i]);
9363 val = cputime64_to_clock_t(val);
9364 cb->fill(cb, cpuacct_stat_desc[i], val);
9369 static struct cftype files[] = {
9372 .read_u64 = cpuusage_read,
9373 .write_u64 = cpuusage_write,
9376 .name = "usage_percpu",
9377 .read_seq_string = cpuacct_percpu_seq_read,
9381 .read_map = cpuacct_stats_show,
9385 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9387 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9391 * charge this task's execution time to its accounting group.
9393 * called with rq->lock held.
9395 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9400 if (unlikely(!cpuacct_subsys.active))
9403 cpu = task_cpu(tsk);
9409 for (; ca; ca = ca->parent) {
9410 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9411 *cpuusage += cputime;
9418 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9419 * in cputime_t units. As a result, cpuacct_update_stats calls
9420 * percpu_counter_add with values large enough to always overflow the
9421 * per cpu batch limit causing bad SMP scalability.
9423 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9424 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9425 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9428 #define CPUACCT_BATCH \
9429 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9431 #define CPUACCT_BATCH 0
9435 * Charge the system/user time to the task's accounting group.
9437 static void cpuacct_update_stats(struct task_struct *tsk,
9438 enum cpuacct_stat_index idx, cputime_t val)
9441 int batch = CPUACCT_BATCH;
9443 if (unlikely(!cpuacct_subsys.active))
9450 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9456 struct cgroup_subsys cpuacct_subsys = {
9458 .create = cpuacct_create,
9459 .destroy = cpuacct_destroy,
9460 .populate = cpuacct_populate,
9461 .subsys_id = cpuacct_subsys_id,
9463 #endif /* CONFIG_CGROUP_CPUACCT */