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 <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
78 #ifdef CONFIG_PARAVIRT
79 #include <asm/paravirt.h>
82 #include "sched_cpupri.h"
83 #include "workqueue_sched.h"
84 #include "sched_autogroup.h"
86 #define CREATE_TRACE_POINTS
87 #include <trace/events/sched.h>
90 * Convert user-nice values [ -20 ... 0 ... 19 ]
91 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
94 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
95 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
96 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
99 * 'User priority' is the nice value converted to something we
100 * can work with better when scaling various scheduler parameters,
101 * it's a [ 0 ... 39 ] range.
103 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
104 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
105 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
108 * Helpers for converting nanosecond timing to jiffy resolution
110 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
112 #define NICE_0_LOAD SCHED_LOAD_SCALE
113 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
116 * These are the 'tuning knobs' of the scheduler:
118 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
119 * Timeslices get refilled after they expire.
121 #define DEF_TIMESLICE (100 * HZ / 1000)
124 * single value that denotes runtime == period, ie unlimited time.
126 #define RUNTIME_INF ((u64)~0ULL)
128 static inline int rt_policy(int policy)
130 if (policy == SCHED_FIFO || policy == SCHED_RR)
135 static inline int task_has_rt_policy(struct task_struct *p)
137 return rt_policy(p->policy);
141 * This is the priority-queue data structure of the RT scheduling class:
143 struct rt_prio_array {
144 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
145 struct list_head queue[MAX_RT_PRIO];
148 struct rt_bandwidth {
149 /* nests inside the rq lock: */
150 raw_spinlock_t rt_runtime_lock;
153 struct hrtimer rt_period_timer;
156 static struct rt_bandwidth def_rt_bandwidth;
158 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
160 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
162 struct rt_bandwidth *rt_b =
163 container_of(timer, struct rt_bandwidth, rt_period_timer);
169 now = hrtimer_cb_get_time(timer);
170 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
175 idle = do_sched_rt_period_timer(rt_b, overrun);
178 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
182 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
184 rt_b->rt_period = ns_to_ktime(period);
185 rt_b->rt_runtime = runtime;
187 raw_spin_lock_init(&rt_b->rt_runtime_lock);
189 hrtimer_init(&rt_b->rt_period_timer,
190 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
191 rt_b->rt_period_timer.function = sched_rt_period_timer;
194 static inline int rt_bandwidth_enabled(void)
196 return sysctl_sched_rt_runtime >= 0;
199 static void start_bandwidth_timer(struct hrtimer *period_timer, ktime_t period)
202 ktime_t soft, hard, now;
205 if (hrtimer_active(period_timer))
208 now = hrtimer_cb_get_time(period_timer);
209 hrtimer_forward(period_timer, now, period);
211 soft = hrtimer_get_softexpires(period_timer);
212 hard = hrtimer_get_expires(period_timer);
213 delta = ktime_to_ns(ktime_sub(hard, soft));
214 __hrtimer_start_range_ns(period_timer, soft, delta,
215 HRTIMER_MODE_ABS_PINNED, 0);
219 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
221 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
224 if (hrtimer_active(&rt_b->rt_period_timer))
227 raw_spin_lock(&rt_b->rt_runtime_lock);
228 start_bandwidth_timer(&rt_b->rt_period_timer, rt_b->rt_period);
229 raw_spin_unlock(&rt_b->rt_runtime_lock);
232 #ifdef CONFIG_RT_GROUP_SCHED
233 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
235 hrtimer_cancel(&rt_b->rt_period_timer);
240 * sched_domains_mutex serializes calls to init_sched_domains,
241 * detach_destroy_domains and partition_sched_domains.
243 static DEFINE_MUTEX(sched_domains_mutex);
245 #ifdef CONFIG_CGROUP_SCHED
247 #include <linux/cgroup.h>
251 static LIST_HEAD(task_groups);
253 struct cfs_bandwidth {
254 #ifdef CONFIG_CFS_BANDWIDTH
258 s64 hierarchal_quota;
261 int idle, timer_active;
262 struct hrtimer period_timer;
263 struct list_head throttled_cfs_rq;
268 /* task group related information */
270 struct cgroup_subsys_state css;
272 #ifdef CONFIG_FAIR_GROUP_SCHED
273 /* schedulable entities of this group on each cpu */
274 struct sched_entity **se;
275 /* runqueue "owned" by this group on each cpu */
276 struct cfs_rq **cfs_rq;
277 unsigned long shares;
279 atomic_t load_weight;
282 #ifdef CONFIG_RT_GROUP_SCHED
283 struct sched_rt_entity **rt_se;
284 struct rt_rq **rt_rq;
286 struct rt_bandwidth rt_bandwidth;
290 struct list_head list;
292 struct task_group *parent;
293 struct list_head siblings;
294 struct list_head children;
296 #ifdef CONFIG_SCHED_AUTOGROUP
297 struct autogroup *autogroup;
300 struct cfs_bandwidth cfs_bandwidth;
303 /* task_group_lock serializes the addition/removal of task groups */
304 static DEFINE_SPINLOCK(task_group_lock);
306 #ifdef CONFIG_FAIR_GROUP_SCHED
308 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
311 * A weight of 0 or 1 can cause arithmetics problems.
312 * A weight of a cfs_rq is the sum of weights of which entities
313 * are queued on this cfs_rq, so a weight of a entity should not be
314 * too large, so as the shares value of a task group.
315 * (The default weight is 1024 - so there's no practical
316 * limitation from this.)
318 #define MIN_SHARES (1UL << 1)
319 #define MAX_SHARES (1UL << 18)
321 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
324 /* Default task group.
325 * Every task in system belong to this group at bootup.
327 struct task_group root_task_group;
329 #endif /* CONFIG_CGROUP_SCHED */
331 /* CFS-related fields in a runqueue */
333 struct load_weight load;
334 unsigned long nr_running, h_nr_running;
339 u64 min_vruntime_copy;
342 struct rb_root tasks_timeline;
343 struct rb_node *rb_leftmost;
345 struct list_head tasks;
346 struct list_head *balance_iterator;
349 * 'curr' points to currently running entity on this cfs_rq.
350 * It is set to NULL otherwise (i.e when none are currently running).
352 struct sched_entity *curr, *next, *last, *skip;
354 #ifdef CONFIG_SCHED_DEBUG
355 unsigned int nr_spread_over;
358 #ifdef CONFIG_FAIR_GROUP_SCHED
359 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
362 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
363 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
364 * (like users, containers etc.)
366 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
367 * list is used during load balance.
370 struct list_head leaf_cfs_rq_list;
371 struct task_group *tg; /* group that "owns" this runqueue */
375 * the part of load.weight contributed by tasks
377 unsigned long task_weight;
380 * h_load = weight * f(tg)
382 * Where f(tg) is the recursive weight fraction assigned to
385 unsigned long h_load;
388 * Maintaining per-cpu shares distribution for group scheduling
390 * load_stamp is the last time we updated the load average
391 * load_last is the last time we updated the load average and saw load
392 * load_unacc_exec_time is currently unaccounted execution time
396 u64 load_stamp, load_last, load_unacc_exec_time;
398 unsigned long load_contribution;
400 #ifdef CONFIG_CFS_BANDWIDTH
403 s64 runtime_remaining;
406 struct list_head throttled_list;
411 #ifdef CONFIG_FAIR_GROUP_SCHED
412 #ifdef CONFIG_CFS_BANDWIDTH
413 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
415 return &tg->cfs_bandwidth;
418 static inline u64 default_cfs_period(void);
419 static int do_sched_cfs_period_timer(struct cfs_bandwidth *cfs_b, int overrun);
421 static enum hrtimer_restart sched_cfs_period_timer(struct hrtimer *timer)
423 struct cfs_bandwidth *cfs_b =
424 container_of(timer, struct cfs_bandwidth, period_timer);
430 now = hrtimer_cb_get_time(timer);
431 overrun = hrtimer_forward(timer, now, cfs_b->period);
436 idle = do_sched_cfs_period_timer(cfs_b, overrun);
439 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
442 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
444 raw_spin_lock_init(&cfs_b->lock);
446 cfs_b->quota = RUNTIME_INF;
447 cfs_b->period = ns_to_ktime(default_cfs_period());
449 INIT_LIST_HEAD(&cfs_b->throttled_cfs_rq);
450 hrtimer_init(&cfs_b->period_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
451 cfs_b->period_timer.function = sched_cfs_period_timer;
454 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq)
456 cfs_rq->runtime_enabled = 0;
457 INIT_LIST_HEAD(&cfs_rq->throttled_list);
460 /* requires cfs_b->lock, may release to reprogram timer */
461 static void __start_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
464 * The timer may be active because we're trying to set a new bandwidth
465 * period or because we're racing with the tear-down path
466 * (timer_active==0 becomes visible before the hrtimer call-back
467 * terminates). In either case we ensure that it's re-programmed
469 while (unlikely(hrtimer_active(&cfs_b->period_timer))) {
470 raw_spin_unlock(&cfs_b->lock);
471 /* ensure cfs_b->lock is available while we wait */
472 hrtimer_cancel(&cfs_b->period_timer);
474 raw_spin_lock(&cfs_b->lock);
475 /* if someone else restarted the timer then we're done */
476 if (cfs_b->timer_active)
480 cfs_b->timer_active = 1;
481 start_bandwidth_timer(&cfs_b->period_timer, cfs_b->period);
484 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b)
486 hrtimer_cancel(&cfs_b->period_timer);
489 static void init_cfs_rq_runtime(struct cfs_rq *cfs_rq) {}
490 static void init_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
491 static void destroy_cfs_bandwidth(struct cfs_bandwidth *cfs_b) {}
493 static inline struct cfs_bandwidth *tg_cfs_bandwidth(struct task_group *tg)
497 #endif /* CONFIG_CFS_BANDWIDTH */
498 #endif /* CONFIG_FAIR_GROUP_SCHED */
500 /* Real-Time classes' related field in a runqueue: */
502 struct rt_prio_array active;
503 unsigned long rt_nr_running;
504 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
506 int curr; /* highest queued rt task prio */
508 int next; /* next highest */
513 unsigned long rt_nr_migratory;
514 unsigned long rt_nr_total;
516 struct plist_head pushable_tasks;
521 /* Nests inside the rq lock: */
522 raw_spinlock_t rt_runtime_lock;
524 #ifdef CONFIG_RT_GROUP_SCHED
525 unsigned long rt_nr_boosted;
528 struct list_head leaf_rt_rq_list;
529 struct task_group *tg;
536 * We add the notion of a root-domain which will be used to define per-domain
537 * variables. Each exclusive cpuset essentially defines an island domain by
538 * fully partitioning the member cpus from any other cpuset. Whenever a new
539 * exclusive cpuset is created, we also create and attach a new root-domain
548 cpumask_var_t online;
551 * The "RT overload" flag: it gets set if a CPU has more than
552 * one runnable RT task.
554 cpumask_var_t rto_mask;
555 struct cpupri cpupri;
559 * By default the system creates a single root-domain with all cpus as
560 * members (mimicking the global state we have today).
562 static struct root_domain def_root_domain;
564 #endif /* CONFIG_SMP */
567 * This is the main, per-CPU runqueue data structure.
569 * Locking rule: those places that want to lock multiple runqueues
570 * (such as the load balancing or the thread migration code), lock
571 * acquire operations must be ordered by ascending &runqueue.
578 * nr_running and cpu_load should be in the same cacheline because
579 * remote CPUs use both these fields when doing load calculation.
581 unsigned long nr_running;
582 #define CPU_LOAD_IDX_MAX 5
583 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
584 unsigned long last_load_update_tick;
587 unsigned char nohz_balance_kick;
589 int skip_clock_update;
591 /* capture load from *all* tasks on this cpu: */
592 struct load_weight load;
593 unsigned long nr_load_updates;
599 #ifdef CONFIG_FAIR_GROUP_SCHED
600 /* list of leaf cfs_rq on this cpu: */
601 struct list_head leaf_cfs_rq_list;
603 #ifdef CONFIG_RT_GROUP_SCHED
604 struct list_head leaf_rt_rq_list;
608 * This is part of a global counter where only the total sum
609 * over all CPUs matters. A task can increase this counter on
610 * one CPU and if it got migrated afterwards it may decrease
611 * it on another CPU. Always updated under the runqueue lock:
613 unsigned long nr_uninterruptible;
615 struct task_struct *curr, *idle, *stop;
616 unsigned long next_balance;
617 struct mm_struct *prev_mm;
625 struct root_domain *rd;
626 struct sched_domain *sd;
628 unsigned long cpu_power;
630 unsigned char idle_at_tick;
631 /* For active balancing */
635 struct cpu_stop_work active_balance_work;
636 /* cpu of this runqueue: */
646 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
649 #ifdef CONFIG_PARAVIRT
652 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
653 u64 prev_steal_time_rq;
656 /* calc_load related fields */
657 unsigned long calc_load_update;
658 long calc_load_active;
660 #ifdef CONFIG_SCHED_HRTICK
662 int hrtick_csd_pending;
663 struct call_single_data hrtick_csd;
665 struct hrtimer hrtick_timer;
668 #ifdef CONFIG_SCHEDSTATS
670 struct sched_info rq_sched_info;
671 unsigned long long rq_cpu_time;
672 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
674 /* sys_sched_yield() stats */
675 unsigned int yld_count;
677 /* schedule() stats */
678 unsigned int sched_switch;
679 unsigned int sched_count;
680 unsigned int sched_goidle;
682 /* try_to_wake_up() stats */
683 unsigned int ttwu_count;
684 unsigned int ttwu_local;
688 struct task_struct *wake_list;
692 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
695 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
697 static inline int cpu_of(struct rq *rq)
706 #define rcu_dereference_check_sched_domain(p) \
707 rcu_dereference_check((p), \
708 lockdep_is_held(&sched_domains_mutex))
711 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
712 * See detach_destroy_domains: synchronize_sched for details.
714 * The domain tree of any CPU may only be accessed from within
715 * preempt-disabled sections.
717 #define for_each_domain(cpu, __sd) \
718 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
720 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
721 #define this_rq() (&__get_cpu_var(runqueues))
722 #define task_rq(p) cpu_rq(task_cpu(p))
723 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
724 #define raw_rq() (&__raw_get_cpu_var(runqueues))
726 #ifdef CONFIG_CGROUP_SCHED
729 * Return the group to which this tasks belongs.
731 * We use task_subsys_state_check() and extend the RCU verification with
732 * pi->lock and rq->lock because cpu_cgroup_attach() holds those locks for each
733 * task it moves into the cgroup. Therefore by holding either of those locks,
734 * we pin the task to the current cgroup.
736 static inline struct task_group *task_group(struct task_struct *p)
738 struct task_group *tg;
739 struct cgroup_subsys_state *css;
741 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
742 lockdep_is_held(&p->pi_lock) ||
743 lockdep_is_held(&task_rq(p)->lock));
744 tg = container_of(css, struct task_group, css);
746 return autogroup_task_group(p, tg);
749 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
750 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
752 #ifdef CONFIG_FAIR_GROUP_SCHED
753 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
754 p->se.parent = task_group(p)->se[cpu];
757 #ifdef CONFIG_RT_GROUP_SCHED
758 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
759 p->rt.parent = task_group(p)->rt_se[cpu];
763 #else /* CONFIG_CGROUP_SCHED */
765 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
766 static inline struct task_group *task_group(struct task_struct *p)
771 #endif /* CONFIG_CGROUP_SCHED */
773 static void update_rq_clock_task(struct rq *rq, s64 delta);
775 static void update_rq_clock(struct rq *rq)
779 if (rq->skip_clock_update > 0)
782 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
784 update_rq_clock_task(rq, delta);
788 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
790 #ifdef CONFIG_SCHED_DEBUG
791 # define const_debug __read_mostly
793 # define const_debug static const
797 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
798 * @cpu: the processor in question.
800 * This interface allows printk to be called with the runqueue lock
801 * held and know whether or not it is OK to wake up the klogd.
803 int runqueue_is_locked(int cpu)
805 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
809 * Debugging: various feature bits
812 #define SCHED_FEAT(name, enabled) \
813 __SCHED_FEAT_##name ,
816 #include "sched_features.h"
821 #define SCHED_FEAT(name, enabled) \
822 (1UL << __SCHED_FEAT_##name) * enabled |
824 const_debug unsigned int sysctl_sched_features =
825 #include "sched_features.h"
830 #ifdef CONFIG_SCHED_DEBUG
831 #define SCHED_FEAT(name, enabled) \
834 static __read_mostly char *sched_feat_names[] = {
835 #include "sched_features.h"
841 static int sched_feat_show(struct seq_file *m, void *v)
845 for (i = 0; sched_feat_names[i]; i++) {
846 if (!(sysctl_sched_features & (1UL << i)))
848 seq_printf(m, "%s ", sched_feat_names[i]);
856 sched_feat_write(struct file *filp, const char __user *ubuf,
857 size_t cnt, loff_t *ppos)
867 if (copy_from_user(&buf, ubuf, cnt))
873 if (strncmp(cmp, "NO_", 3) == 0) {
878 for (i = 0; sched_feat_names[i]; i++) {
879 if (strcmp(cmp, sched_feat_names[i]) == 0) {
881 sysctl_sched_features &= ~(1UL << i);
883 sysctl_sched_features |= (1UL << i);
888 if (!sched_feat_names[i])
896 static int sched_feat_open(struct inode *inode, struct file *filp)
898 return single_open(filp, sched_feat_show, NULL);
901 static const struct file_operations sched_feat_fops = {
902 .open = sched_feat_open,
903 .write = sched_feat_write,
906 .release = single_release,
909 static __init int sched_init_debug(void)
911 debugfs_create_file("sched_features", 0644, NULL, NULL,
916 late_initcall(sched_init_debug);
920 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
923 * Number of tasks to iterate in a single balance run.
924 * Limited because this is done with IRQs disabled.
926 const_debug unsigned int sysctl_sched_nr_migrate = 32;
929 * period over which we average the RT time consumption, measured
934 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
937 * period over which we measure -rt task cpu usage in us.
940 unsigned int sysctl_sched_rt_period = 1000000;
942 static __read_mostly int scheduler_running;
945 * part of the period that we allow rt tasks to run in us.
948 int sysctl_sched_rt_runtime = 950000;
950 static inline u64 global_rt_period(void)
952 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
955 static inline u64 global_rt_runtime(void)
957 if (sysctl_sched_rt_runtime < 0)
960 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
963 #ifndef prepare_arch_switch
964 # define prepare_arch_switch(next) do { } while (0)
966 #ifndef finish_arch_switch
967 # define finish_arch_switch(prev) do { } while (0)
970 static inline int task_current(struct rq *rq, struct task_struct *p)
972 return rq->curr == p;
975 static inline int task_running(struct rq *rq, struct task_struct *p)
980 return task_current(rq, p);
984 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
985 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
989 * We can optimise this out completely for !SMP, because the
990 * SMP rebalancing from interrupt is the only thing that cares
997 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1001 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1002 * We must ensure this doesn't happen until the switch is completely
1008 #ifdef CONFIG_DEBUG_SPINLOCK
1009 /* this is a valid case when another task releases the spinlock */
1010 rq->lock.owner = current;
1013 * If we are tracking spinlock dependencies then we have to
1014 * fix up the runqueue lock - which gets 'carried over' from
1015 * prev into current:
1017 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1019 raw_spin_unlock_irq(&rq->lock);
1022 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1023 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1027 * We can optimise this out completely for !SMP, because the
1028 * SMP rebalancing from interrupt is the only thing that cares
1033 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1034 raw_spin_unlock_irq(&rq->lock);
1036 raw_spin_unlock(&rq->lock);
1040 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1044 * After ->on_cpu is cleared, the task can be moved to a different CPU.
1045 * We must ensure this doesn't happen until the switch is completely
1051 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1055 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1058 * __task_rq_lock - lock the rq @p resides on.
1060 static inline struct rq *__task_rq_lock(struct task_struct *p)
1061 __acquires(rq->lock)
1065 lockdep_assert_held(&p->pi_lock);
1069 raw_spin_lock(&rq->lock);
1070 if (likely(rq == task_rq(p)))
1072 raw_spin_unlock(&rq->lock);
1077 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
1079 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1080 __acquires(p->pi_lock)
1081 __acquires(rq->lock)
1086 raw_spin_lock_irqsave(&p->pi_lock, *flags);
1088 raw_spin_lock(&rq->lock);
1089 if (likely(rq == task_rq(p)))
1091 raw_spin_unlock(&rq->lock);
1092 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1096 static void __task_rq_unlock(struct rq *rq)
1097 __releases(rq->lock)
1099 raw_spin_unlock(&rq->lock);
1103 task_rq_unlock(struct rq *rq, struct task_struct *p, unsigned long *flags)
1104 __releases(rq->lock)
1105 __releases(p->pi_lock)
1107 raw_spin_unlock(&rq->lock);
1108 raw_spin_unlock_irqrestore(&p->pi_lock, *flags);
1112 * this_rq_lock - lock this runqueue and disable interrupts.
1114 static struct rq *this_rq_lock(void)
1115 __acquires(rq->lock)
1119 local_irq_disable();
1121 raw_spin_lock(&rq->lock);
1126 #ifdef CONFIG_SCHED_HRTICK
1128 * Use HR-timers to deliver accurate preemption points.
1130 * Its all a bit involved since we cannot program an hrt while holding the
1131 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1134 * When we get rescheduled we reprogram the hrtick_timer outside of the
1140 * - enabled by features
1141 * - hrtimer is actually high res
1143 static inline int hrtick_enabled(struct rq *rq)
1145 if (!sched_feat(HRTICK))
1147 if (!cpu_active(cpu_of(rq)))
1149 return hrtimer_is_hres_active(&rq->hrtick_timer);
1152 static void hrtick_clear(struct rq *rq)
1154 if (hrtimer_active(&rq->hrtick_timer))
1155 hrtimer_cancel(&rq->hrtick_timer);
1159 * High-resolution timer tick.
1160 * Runs from hardirq context with interrupts disabled.
1162 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1164 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1166 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1168 raw_spin_lock(&rq->lock);
1169 update_rq_clock(rq);
1170 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1171 raw_spin_unlock(&rq->lock);
1173 return HRTIMER_NORESTART;
1178 * called from hardirq (IPI) context
1180 static void __hrtick_start(void *arg)
1182 struct rq *rq = arg;
1184 raw_spin_lock(&rq->lock);
1185 hrtimer_restart(&rq->hrtick_timer);
1186 rq->hrtick_csd_pending = 0;
1187 raw_spin_unlock(&rq->lock);
1191 * Called to set the hrtick timer state.
1193 * called with rq->lock held and irqs disabled
1195 static void hrtick_start(struct rq *rq, u64 delay)
1197 struct hrtimer *timer = &rq->hrtick_timer;
1198 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1200 hrtimer_set_expires(timer, time);
1202 if (rq == this_rq()) {
1203 hrtimer_restart(timer);
1204 } else if (!rq->hrtick_csd_pending) {
1205 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1206 rq->hrtick_csd_pending = 1;
1211 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1213 int cpu = (int)(long)hcpu;
1216 case CPU_UP_CANCELED:
1217 case CPU_UP_CANCELED_FROZEN:
1218 case CPU_DOWN_PREPARE:
1219 case CPU_DOWN_PREPARE_FROZEN:
1221 case CPU_DEAD_FROZEN:
1222 hrtick_clear(cpu_rq(cpu));
1229 static __init void init_hrtick(void)
1231 hotcpu_notifier(hotplug_hrtick, 0);
1235 * Called to set the hrtick timer state.
1237 * called with rq->lock held and irqs disabled
1239 static void hrtick_start(struct rq *rq, u64 delay)
1241 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1242 HRTIMER_MODE_REL_PINNED, 0);
1245 static inline void init_hrtick(void)
1248 #endif /* CONFIG_SMP */
1250 static void init_rq_hrtick(struct rq *rq)
1253 rq->hrtick_csd_pending = 0;
1255 rq->hrtick_csd.flags = 0;
1256 rq->hrtick_csd.func = __hrtick_start;
1257 rq->hrtick_csd.info = rq;
1260 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1261 rq->hrtick_timer.function = hrtick;
1263 #else /* CONFIG_SCHED_HRTICK */
1264 static inline void hrtick_clear(struct rq *rq)
1268 static inline void init_rq_hrtick(struct rq *rq)
1272 static inline void init_hrtick(void)
1275 #endif /* CONFIG_SCHED_HRTICK */
1278 * resched_task - mark a task 'to be rescheduled now'.
1280 * On UP this means the setting of the need_resched flag, on SMP it
1281 * might also involve a cross-CPU call to trigger the scheduler on
1286 #ifndef tsk_is_polling
1287 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1290 static void resched_task(struct task_struct *p)
1294 assert_raw_spin_locked(&task_rq(p)->lock);
1296 if (test_tsk_need_resched(p))
1299 set_tsk_need_resched(p);
1302 if (cpu == smp_processor_id())
1305 /* NEED_RESCHED must be visible before we test polling */
1307 if (!tsk_is_polling(p))
1308 smp_send_reschedule(cpu);
1311 static void resched_cpu(int cpu)
1313 struct rq *rq = cpu_rq(cpu);
1314 unsigned long flags;
1316 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1318 resched_task(cpu_curr(cpu));
1319 raw_spin_unlock_irqrestore(&rq->lock, flags);
1324 * In the semi idle case, use the nearest busy cpu for migrating timers
1325 * from an idle cpu. This is good for power-savings.
1327 * We don't do similar optimization for completely idle system, as
1328 * selecting an idle cpu will add more delays to the timers than intended
1329 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1331 int get_nohz_timer_target(void)
1333 int cpu = smp_processor_id();
1335 struct sched_domain *sd;
1338 for_each_domain(cpu, sd) {
1339 for_each_cpu(i, sched_domain_span(sd)) {
1351 * When add_timer_on() enqueues a timer into the timer wheel of an
1352 * idle CPU then this timer might expire before the next timer event
1353 * which is scheduled to wake up that CPU. In case of a completely
1354 * idle system the next event might even be infinite time into the
1355 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1356 * leaves the inner idle loop so the newly added timer is taken into
1357 * account when the CPU goes back to idle and evaluates the timer
1358 * wheel for the next timer event.
1360 void wake_up_idle_cpu(int cpu)
1362 struct rq *rq = cpu_rq(cpu);
1364 if (cpu == smp_processor_id())
1368 * This is safe, as this function is called with the timer
1369 * wheel base lock of (cpu) held. When the CPU is on the way
1370 * to idle and has not yet set rq->curr to idle then it will
1371 * be serialized on the timer wheel base lock and take the new
1372 * timer into account automatically.
1374 if (rq->curr != rq->idle)
1378 * We can set TIF_RESCHED on the idle task of the other CPU
1379 * lockless. The worst case is that the other CPU runs the
1380 * idle task through an additional NOOP schedule()
1382 set_tsk_need_resched(rq->idle);
1384 /* NEED_RESCHED must be visible before we test polling */
1386 if (!tsk_is_polling(rq->idle))
1387 smp_send_reschedule(cpu);
1390 #endif /* CONFIG_NO_HZ */
1392 static u64 sched_avg_period(void)
1394 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1397 static void sched_avg_update(struct rq *rq)
1399 s64 period = sched_avg_period();
1401 while ((s64)(rq->clock - rq->age_stamp) > period) {
1403 * Inline assembly required to prevent the compiler
1404 * optimising this loop into a divmod call.
1405 * See __iter_div_u64_rem() for another example of this.
1407 asm("" : "+rm" (rq->age_stamp));
1408 rq->age_stamp += period;
1413 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1415 rq->rt_avg += rt_delta;
1416 sched_avg_update(rq);
1419 #else /* !CONFIG_SMP */
1420 static void resched_task(struct task_struct *p)
1422 assert_raw_spin_locked(&task_rq(p)->lock);
1423 set_tsk_need_resched(p);
1426 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1430 static void sched_avg_update(struct rq *rq)
1433 #endif /* CONFIG_SMP */
1435 #if BITS_PER_LONG == 32
1436 # define WMULT_CONST (~0UL)
1438 # define WMULT_CONST (1UL << 32)
1441 #define WMULT_SHIFT 32
1444 * Shift right and round:
1446 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1449 * delta *= weight / lw
1451 static unsigned long
1452 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1453 struct load_weight *lw)
1458 * weight can be less than 2^SCHED_LOAD_RESOLUTION for task group sched
1459 * entities since MIN_SHARES = 2. Treat weight as 1 if less than
1460 * 2^SCHED_LOAD_RESOLUTION.
1462 if (likely(weight > (1UL << SCHED_LOAD_RESOLUTION)))
1463 tmp = (u64)delta_exec * scale_load_down(weight);
1465 tmp = (u64)delta_exec;
1467 if (!lw->inv_weight) {
1468 unsigned long w = scale_load_down(lw->weight);
1470 if (BITS_PER_LONG > 32 && unlikely(w >= WMULT_CONST))
1472 else if (unlikely(!w))
1473 lw->inv_weight = WMULT_CONST;
1475 lw->inv_weight = WMULT_CONST / w;
1479 * Check whether we'd overflow the 64-bit multiplication:
1481 if (unlikely(tmp > WMULT_CONST))
1482 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1485 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1487 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1490 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1496 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1502 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1509 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1510 * of tasks with abnormal "nice" values across CPUs the contribution that
1511 * each task makes to its run queue's load is weighted according to its
1512 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1513 * scaled version of the new time slice allocation that they receive on time
1517 #define WEIGHT_IDLEPRIO 3
1518 #define WMULT_IDLEPRIO 1431655765
1521 * Nice levels are multiplicative, with a gentle 10% change for every
1522 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1523 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1524 * that remained on nice 0.
1526 * The "10% effect" is relative and cumulative: from _any_ nice level,
1527 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1528 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1529 * If a task goes up by ~10% and another task goes down by ~10% then
1530 * the relative distance between them is ~25%.)
1532 static const int prio_to_weight[40] = {
1533 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1534 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1535 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1536 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1537 /* 0 */ 1024, 820, 655, 526, 423,
1538 /* 5 */ 335, 272, 215, 172, 137,
1539 /* 10 */ 110, 87, 70, 56, 45,
1540 /* 15 */ 36, 29, 23, 18, 15,
1544 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1546 * In cases where the weight does not change often, we can use the
1547 * precalculated inverse to speed up arithmetics by turning divisions
1548 * into multiplications:
1550 static const u32 prio_to_wmult[40] = {
1551 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1552 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1553 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1554 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1555 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1556 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1557 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1558 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1561 /* Time spent by the tasks of the cpu accounting group executing in ... */
1562 enum cpuacct_stat_index {
1563 CPUACCT_STAT_USER, /* ... user mode */
1564 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1566 CPUACCT_STAT_NSTATS,
1569 #ifdef CONFIG_CGROUP_CPUACCT
1570 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1571 static void cpuacct_update_stats(struct task_struct *tsk,
1572 enum cpuacct_stat_index idx, cputime_t val);
1574 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1575 static inline void cpuacct_update_stats(struct task_struct *tsk,
1576 enum cpuacct_stat_index idx, cputime_t val) {}
1579 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1581 update_load_add(&rq->load, load);
1584 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1586 update_load_sub(&rq->load, load);
1589 #if defined(CONFIG_RT_GROUP_SCHED) || (defined(CONFIG_FAIR_GROUP_SCHED) && \
1590 (defined(CONFIG_SMP) || defined(CONFIG_CFS_BANDWIDTH)))
1591 typedef int (*tg_visitor)(struct task_group *, void *);
1594 * Iterate the full tree, calling @down when first entering a node and @up when
1595 * leaving it for the final time.
1597 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1599 struct task_group *parent, *child;
1603 parent = &root_task_group;
1605 ret = (*down)(parent, data);
1608 list_for_each_entry_rcu(child, &parent->children, siblings) {
1615 ret = (*up)(parent, data);
1620 parent = parent->parent;
1629 static int tg_nop(struct task_group *tg, void *data)
1636 /* Used instead of source_load when we know the type == 0 */
1637 static unsigned long weighted_cpuload(const int cpu)
1639 return cpu_rq(cpu)->load.weight;
1643 * Return a low guess at the load of a migration-source cpu weighted
1644 * according to the scheduling class and "nice" value.
1646 * We want to under-estimate the load of migration sources, to
1647 * balance conservatively.
1649 static unsigned long source_load(int cpu, int type)
1651 struct rq *rq = cpu_rq(cpu);
1652 unsigned long total = weighted_cpuload(cpu);
1654 if (type == 0 || !sched_feat(LB_BIAS))
1657 return min(rq->cpu_load[type-1], total);
1661 * Return a high guess at the load of a migration-target cpu weighted
1662 * according to the scheduling class and "nice" value.
1664 static unsigned long target_load(int cpu, int type)
1666 struct rq *rq = cpu_rq(cpu);
1667 unsigned long total = weighted_cpuload(cpu);
1669 if (type == 0 || !sched_feat(LB_BIAS))
1672 return max(rq->cpu_load[type-1], total);
1675 static unsigned long power_of(int cpu)
1677 return cpu_rq(cpu)->cpu_power;
1680 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1682 static unsigned long cpu_avg_load_per_task(int cpu)
1684 struct rq *rq = cpu_rq(cpu);
1685 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1688 return rq->load.weight / nr_running;
1693 #ifdef CONFIG_PREEMPT
1695 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1698 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1699 * way at the expense of forcing extra atomic operations in all
1700 * invocations. This assures that the double_lock is acquired using the
1701 * same underlying policy as the spinlock_t on this architecture, which
1702 * reduces latency compared to the unfair variant below. However, it
1703 * also adds more overhead and therefore may reduce throughput.
1705 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1706 __releases(this_rq->lock)
1707 __acquires(busiest->lock)
1708 __acquires(this_rq->lock)
1710 raw_spin_unlock(&this_rq->lock);
1711 double_rq_lock(this_rq, busiest);
1718 * Unfair double_lock_balance: Optimizes throughput at the expense of
1719 * latency by eliminating extra atomic operations when the locks are
1720 * already in proper order on entry. This favors lower cpu-ids and will
1721 * grant the double lock to lower cpus over higher ids under contention,
1722 * regardless of entry order into the function.
1724 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1725 __releases(this_rq->lock)
1726 __acquires(busiest->lock)
1727 __acquires(this_rq->lock)
1731 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1732 if (busiest < this_rq) {
1733 raw_spin_unlock(&this_rq->lock);
1734 raw_spin_lock(&busiest->lock);
1735 raw_spin_lock_nested(&this_rq->lock,
1736 SINGLE_DEPTH_NESTING);
1739 raw_spin_lock_nested(&busiest->lock,
1740 SINGLE_DEPTH_NESTING);
1745 #endif /* CONFIG_PREEMPT */
1748 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1750 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1752 if (unlikely(!irqs_disabled())) {
1753 /* printk() doesn't work good under rq->lock */
1754 raw_spin_unlock(&this_rq->lock);
1758 return _double_lock_balance(this_rq, busiest);
1761 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1762 __releases(busiest->lock)
1764 raw_spin_unlock(&busiest->lock);
1765 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1769 * double_rq_lock - safely lock two runqueues
1771 * Note this does not disable interrupts like task_rq_lock,
1772 * you need to do so manually before calling.
1774 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1775 __acquires(rq1->lock)
1776 __acquires(rq2->lock)
1778 BUG_ON(!irqs_disabled());
1780 raw_spin_lock(&rq1->lock);
1781 __acquire(rq2->lock); /* Fake it out ;) */
1784 raw_spin_lock(&rq1->lock);
1785 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1787 raw_spin_lock(&rq2->lock);
1788 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1794 * double_rq_unlock - safely unlock two runqueues
1796 * Note this does not restore interrupts like task_rq_unlock,
1797 * you need to do so manually after calling.
1799 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1800 __releases(rq1->lock)
1801 __releases(rq2->lock)
1803 raw_spin_unlock(&rq1->lock);
1805 raw_spin_unlock(&rq2->lock);
1807 __release(rq2->lock);
1810 #else /* CONFIG_SMP */
1813 * double_rq_lock - safely lock two runqueues
1815 * Note this does not disable interrupts like task_rq_lock,
1816 * you need to do so manually before calling.
1818 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1819 __acquires(rq1->lock)
1820 __acquires(rq2->lock)
1822 BUG_ON(!irqs_disabled());
1824 raw_spin_lock(&rq1->lock);
1825 __acquire(rq2->lock); /* Fake it out ;) */
1829 * double_rq_unlock - safely unlock two runqueues
1831 * Note this does not restore interrupts like task_rq_unlock,
1832 * you need to do so manually after calling.
1834 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1835 __releases(rq1->lock)
1836 __releases(rq2->lock)
1839 raw_spin_unlock(&rq1->lock);
1840 __release(rq2->lock);
1845 static void calc_load_account_idle(struct rq *this_rq);
1846 static void update_sysctl(void);
1847 static int get_update_sysctl_factor(void);
1848 static void update_cpu_load(struct rq *this_rq);
1850 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1852 set_task_rq(p, cpu);
1855 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1856 * successfuly executed on another CPU. We must ensure that updates of
1857 * per-task data have been completed by this moment.
1860 task_thread_info(p)->cpu = cpu;
1864 static const struct sched_class rt_sched_class;
1866 #define sched_class_highest (&stop_sched_class)
1867 #define for_each_class(class) \
1868 for (class = sched_class_highest; class; class = class->next)
1870 #include "sched_stats.h"
1872 static void inc_nr_running(struct rq *rq)
1877 static void dec_nr_running(struct rq *rq)
1882 static void set_load_weight(struct task_struct *p)
1884 int prio = p->static_prio - MAX_RT_PRIO;
1885 struct load_weight *load = &p->se.load;
1888 * SCHED_IDLE tasks get minimal weight:
1890 if (p->policy == SCHED_IDLE) {
1891 load->weight = scale_load(WEIGHT_IDLEPRIO);
1892 load->inv_weight = WMULT_IDLEPRIO;
1896 load->weight = scale_load(prio_to_weight[prio]);
1897 load->inv_weight = prio_to_wmult[prio];
1900 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1902 update_rq_clock(rq);
1903 sched_info_queued(p);
1904 p->sched_class->enqueue_task(rq, p, flags);
1907 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1909 update_rq_clock(rq);
1910 sched_info_dequeued(p);
1911 p->sched_class->dequeue_task(rq, p, flags);
1915 * activate_task - move a task to the runqueue.
1917 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1919 if (task_contributes_to_load(p))
1920 rq->nr_uninterruptible--;
1922 enqueue_task(rq, p, flags);
1926 * deactivate_task - remove a task from the runqueue.
1928 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1930 if (task_contributes_to_load(p))
1931 rq->nr_uninterruptible++;
1933 dequeue_task(rq, p, flags);
1936 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1939 * There are no locks covering percpu hardirq/softirq time.
1940 * They are only modified in account_system_vtime, on corresponding CPU
1941 * with interrupts disabled. So, writes are safe.
1942 * They are read and saved off onto struct rq in update_rq_clock().
1943 * This may result in other CPU reading this CPU's irq time and can
1944 * race with irq/account_system_vtime on this CPU. We would either get old
1945 * or new value with a side effect of accounting a slice of irq time to wrong
1946 * task when irq is in progress while we read rq->clock. That is a worthy
1947 * compromise in place of having locks on each irq in account_system_time.
1949 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1950 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1952 static DEFINE_PER_CPU(u64, irq_start_time);
1953 static int sched_clock_irqtime;
1955 void enable_sched_clock_irqtime(void)
1957 sched_clock_irqtime = 1;
1960 void disable_sched_clock_irqtime(void)
1962 sched_clock_irqtime = 0;
1965 #ifndef CONFIG_64BIT
1966 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1968 static inline void irq_time_write_begin(void)
1970 __this_cpu_inc(irq_time_seq.sequence);
1974 static inline void irq_time_write_end(void)
1977 __this_cpu_inc(irq_time_seq.sequence);
1980 static inline u64 irq_time_read(int cpu)
1986 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1987 irq_time = per_cpu(cpu_softirq_time, cpu) +
1988 per_cpu(cpu_hardirq_time, cpu);
1989 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1993 #else /* CONFIG_64BIT */
1994 static inline void irq_time_write_begin(void)
1998 static inline void irq_time_write_end(void)
2002 static inline u64 irq_time_read(int cpu)
2004 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
2006 #endif /* CONFIG_64BIT */
2009 * Called before incrementing preempt_count on {soft,}irq_enter
2010 * and before decrementing preempt_count on {soft,}irq_exit.
2012 void account_system_vtime(struct task_struct *curr)
2014 unsigned long flags;
2018 if (!sched_clock_irqtime)
2021 local_irq_save(flags);
2023 cpu = smp_processor_id();
2024 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
2025 __this_cpu_add(irq_start_time, delta);
2027 irq_time_write_begin();
2029 * We do not account for softirq time from ksoftirqd here.
2030 * We want to continue accounting softirq time to ksoftirqd thread
2031 * in that case, so as not to confuse scheduler with a special task
2032 * that do not consume any time, but still wants to run.
2034 if (hardirq_count())
2035 __this_cpu_add(cpu_hardirq_time, delta);
2036 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
2037 __this_cpu_add(cpu_softirq_time, delta);
2039 irq_time_write_end();
2040 local_irq_restore(flags);
2042 EXPORT_SYMBOL_GPL(account_system_vtime);
2044 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2046 #ifdef CONFIG_PARAVIRT
2047 static inline u64 steal_ticks(u64 steal)
2049 if (unlikely(steal > NSEC_PER_SEC))
2050 return div_u64(steal, TICK_NSEC);
2052 return __iter_div_u64_rem(steal, TICK_NSEC, &steal);
2056 static void update_rq_clock_task(struct rq *rq, s64 delta)
2059 * In theory, the compile should just see 0 here, and optimize out the call
2060 * to sched_rt_avg_update. But I don't trust it...
2062 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2063 s64 steal = 0, irq_delta = 0;
2065 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2066 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
2069 * Since irq_time is only updated on {soft,}irq_exit, we might run into
2070 * this case when a previous update_rq_clock() happened inside a
2071 * {soft,}irq region.
2073 * When this happens, we stop ->clock_task and only update the
2074 * prev_irq_time stamp to account for the part that fit, so that a next
2075 * update will consume the rest. This ensures ->clock_task is
2078 * It does however cause some slight miss-attribution of {soft,}irq
2079 * time, a more accurate solution would be to update the irq_time using
2080 * the current rq->clock timestamp, except that would require using
2083 if (irq_delta > delta)
2086 rq->prev_irq_time += irq_delta;
2089 #ifdef CONFIG_PARAVIRT_TIME_ACCOUNTING
2090 if (static_branch((¶virt_steal_rq_enabled))) {
2093 steal = paravirt_steal_clock(cpu_of(rq));
2094 steal -= rq->prev_steal_time_rq;
2096 if (unlikely(steal > delta))
2099 st = steal_ticks(steal);
2100 steal = st * TICK_NSEC;
2102 rq->prev_steal_time_rq += steal;
2108 rq->clock_task += delta;
2110 #if defined(CONFIG_IRQ_TIME_ACCOUNTING) || defined(CONFIG_PARAVIRT_TIME_ACCOUNTING)
2111 if ((irq_delta + steal) && sched_feat(NONTASK_POWER))
2112 sched_rt_avg_update(rq, irq_delta + steal);
2116 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
2117 static int irqtime_account_hi_update(void)
2119 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2120 unsigned long flags;
2124 local_irq_save(flags);
2125 latest_ns = this_cpu_read(cpu_hardirq_time);
2126 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
2128 local_irq_restore(flags);
2132 static int irqtime_account_si_update(void)
2134 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
2135 unsigned long flags;
2139 local_irq_save(flags);
2140 latest_ns = this_cpu_read(cpu_softirq_time);
2141 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
2143 local_irq_restore(flags);
2147 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
2149 #define sched_clock_irqtime (0)
2153 #include "sched_idletask.c"
2154 #include "sched_fair.c"
2155 #include "sched_rt.c"
2156 #include "sched_autogroup.c"
2157 #include "sched_stoptask.c"
2158 #ifdef CONFIG_SCHED_DEBUG
2159 # include "sched_debug.c"
2162 void sched_set_stop_task(int cpu, struct task_struct *stop)
2164 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2165 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2169 * Make it appear like a SCHED_FIFO task, its something
2170 * userspace knows about and won't get confused about.
2172 * Also, it will make PI more or less work without too
2173 * much confusion -- but then, stop work should not
2174 * rely on PI working anyway.
2176 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2178 stop->sched_class = &stop_sched_class;
2181 cpu_rq(cpu)->stop = stop;
2185 * Reset it back to a normal scheduling class so that
2186 * it can die in pieces.
2188 old_stop->sched_class = &rt_sched_class;
2193 * __normal_prio - return the priority that is based on the static prio
2195 static inline int __normal_prio(struct task_struct *p)
2197 return p->static_prio;
2201 * Calculate the expected normal priority: i.e. priority
2202 * without taking RT-inheritance into account. Might be
2203 * boosted by interactivity modifiers. Changes upon fork,
2204 * setprio syscalls, and whenever the interactivity
2205 * estimator recalculates.
2207 static inline int normal_prio(struct task_struct *p)
2211 if (task_has_rt_policy(p))
2212 prio = MAX_RT_PRIO-1 - p->rt_priority;
2214 prio = __normal_prio(p);
2219 * Calculate the current priority, i.e. the priority
2220 * taken into account by the scheduler. This value might
2221 * be boosted by RT tasks, or might be boosted by
2222 * interactivity modifiers. Will be RT if the task got
2223 * RT-boosted. If not then it returns p->normal_prio.
2225 static int effective_prio(struct task_struct *p)
2227 p->normal_prio = normal_prio(p);
2229 * If we are RT tasks or we were boosted to RT priority,
2230 * keep the priority unchanged. Otherwise, update priority
2231 * to the normal priority:
2233 if (!rt_prio(p->prio))
2234 return p->normal_prio;
2239 * task_curr - is this task currently executing on a CPU?
2240 * @p: the task in question.
2242 inline int task_curr(const struct task_struct *p)
2244 return cpu_curr(task_cpu(p)) == p;
2247 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2248 const struct sched_class *prev_class,
2251 if (prev_class != p->sched_class) {
2252 if (prev_class->switched_from)
2253 prev_class->switched_from(rq, p);
2254 p->sched_class->switched_to(rq, p);
2255 } else if (oldprio != p->prio)
2256 p->sched_class->prio_changed(rq, p, oldprio);
2259 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2261 const struct sched_class *class;
2263 if (p->sched_class == rq->curr->sched_class) {
2264 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2266 for_each_class(class) {
2267 if (class == rq->curr->sched_class)
2269 if (class == p->sched_class) {
2270 resched_task(rq->curr);
2277 * A queue event has occurred, and we're going to schedule. In
2278 * this case, we can save a useless back to back clock update.
2280 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2281 rq->skip_clock_update = 1;
2286 * Is this task likely cache-hot:
2289 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2293 if (p->sched_class != &fair_sched_class)
2296 if (unlikely(p->policy == SCHED_IDLE))
2300 * Buddy candidates are cache hot:
2302 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2303 (&p->se == cfs_rq_of(&p->se)->next ||
2304 &p->se == cfs_rq_of(&p->se)->last))
2307 if (sysctl_sched_migration_cost == -1)
2309 if (sysctl_sched_migration_cost == 0)
2312 delta = now - p->se.exec_start;
2314 return delta < (s64)sysctl_sched_migration_cost;
2317 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2319 #ifdef CONFIG_SCHED_DEBUG
2321 * We should never call set_task_cpu() on a blocked task,
2322 * ttwu() will sort out the placement.
2324 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2325 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2327 #ifdef CONFIG_LOCKDEP
2329 * The caller should hold either p->pi_lock or rq->lock, when changing
2330 * a task's CPU. ->pi_lock for waking tasks, rq->lock for runnable tasks.
2332 * sched_move_task() holds both and thus holding either pins the cgroup,
2333 * see set_task_rq().
2335 * Furthermore, all task_rq users should acquire both locks, see
2338 WARN_ON_ONCE(debug_locks && !(lockdep_is_held(&p->pi_lock) ||
2339 lockdep_is_held(&task_rq(p)->lock)));
2343 trace_sched_migrate_task(p, new_cpu);
2345 if (task_cpu(p) != new_cpu) {
2346 p->se.nr_migrations++;
2347 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, NULL, 0);
2350 __set_task_cpu(p, new_cpu);
2353 struct migration_arg {
2354 struct task_struct *task;
2358 static int migration_cpu_stop(void *data);
2361 * wait_task_inactive - wait for a thread to unschedule.
2363 * If @match_state is nonzero, it's the @p->state value just checked and
2364 * not expected to change. If it changes, i.e. @p might have woken up,
2365 * then return zero. When we succeed in waiting for @p to be off its CPU,
2366 * we return a positive number (its total switch count). If a second call
2367 * a short while later returns the same number, the caller can be sure that
2368 * @p has remained unscheduled the whole time.
2370 * The caller must ensure that the task *will* unschedule sometime soon,
2371 * else this function might spin for a *long* time. This function can't
2372 * be called with interrupts off, or it may introduce deadlock with
2373 * smp_call_function() if an IPI is sent by the same process we are
2374 * waiting to become inactive.
2376 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2378 unsigned long flags;
2385 * We do the initial early heuristics without holding
2386 * any task-queue locks at all. We'll only try to get
2387 * the runqueue lock when things look like they will
2393 * If the task is actively running on another CPU
2394 * still, just relax and busy-wait without holding
2397 * NOTE! Since we don't hold any locks, it's not
2398 * even sure that "rq" stays as the right runqueue!
2399 * But we don't care, since "task_running()" will
2400 * return false if the runqueue has changed and p
2401 * is actually now running somewhere else!
2403 while (task_running(rq, p)) {
2404 if (match_state && unlikely(p->state != match_state))
2410 * Ok, time to look more closely! We need the rq
2411 * lock now, to be *sure*. If we're wrong, we'll
2412 * just go back and repeat.
2414 rq = task_rq_lock(p, &flags);
2415 trace_sched_wait_task(p);
2416 running = task_running(rq, p);
2419 if (!match_state || p->state == match_state)
2420 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2421 task_rq_unlock(rq, p, &flags);
2424 * If it changed from the expected state, bail out now.
2426 if (unlikely(!ncsw))
2430 * Was it really running after all now that we
2431 * checked with the proper locks actually held?
2433 * Oops. Go back and try again..
2435 if (unlikely(running)) {
2441 * It's not enough that it's not actively running,
2442 * it must be off the runqueue _entirely_, and not
2445 * So if it was still runnable (but just not actively
2446 * running right now), it's preempted, and we should
2447 * yield - it could be a while.
2449 if (unlikely(on_rq)) {
2450 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2452 set_current_state(TASK_UNINTERRUPTIBLE);
2453 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2458 * Ahh, all good. It wasn't running, and it wasn't
2459 * runnable, which means that it will never become
2460 * running in the future either. We're all done!
2469 * kick_process - kick a running thread to enter/exit the kernel
2470 * @p: the to-be-kicked thread
2472 * Cause a process which is running on another CPU to enter
2473 * kernel-mode, without any delay. (to get signals handled.)
2475 * NOTE: this function doesn't have to take the runqueue lock,
2476 * because all it wants to ensure is that the remote task enters
2477 * the kernel. If the IPI races and the task has been migrated
2478 * to another CPU then no harm is done and the purpose has been
2481 void kick_process(struct task_struct *p)
2487 if ((cpu != smp_processor_id()) && task_curr(p))
2488 smp_send_reschedule(cpu);
2491 EXPORT_SYMBOL_GPL(kick_process);
2492 #endif /* CONFIG_SMP */
2496 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2498 static int select_fallback_rq(int cpu, struct task_struct *p)
2501 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2503 /* Look for allowed, online CPU in same node. */
2504 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2505 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2508 /* Any allowed, online CPU? */
2509 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2510 if (dest_cpu < nr_cpu_ids)
2513 /* No more Mr. Nice Guy. */
2514 dest_cpu = cpuset_cpus_allowed_fallback(p);
2516 * Don't tell them about moving exiting tasks or
2517 * kernel threads (both mm NULL), since they never
2520 if (p->mm && printk_ratelimit()) {
2521 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2522 task_pid_nr(p), p->comm, cpu);
2529 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2532 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2534 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2537 * In order not to call set_task_cpu() on a blocking task we need
2538 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2541 * Since this is common to all placement strategies, this lives here.
2543 * [ this allows ->select_task() to simply return task_cpu(p) and
2544 * not worry about this generic constraint ]
2546 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2548 cpu = select_fallback_rq(task_cpu(p), p);
2553 static void update_avg(u64 *avg, u64 sample)
2555 s64 diff = sample - *avg;
2561 ttwu_stat(struct task_struct *p, int cpu, int wake_flags)
2563 #ifdef CONFIG_SCHEDSTATS
2564 struct rq *rq = this_rq();
2567 int this_cpu = smp_processor_id();
2569 if (cpu == this_cpu) {
2570 schedstat_inc(rq, ttwu_local);
2571 schedstat_inc(p, se.statistics.nr_wakeups_local);
2573 struct sched_domain *sd;
2575 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2577 for_each_domain(this_cpu, sd) {
2578 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2579 schedstat_inc(sd, ttwu_wake_remote);
2586 if (wake_flags & WF_MIGRATED)
2587 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2589 #endif /* CONFIG_SMP */
2591 schedstat_inc(rq, ttwu_count);
2592 schedstat_inc(p, se.statistics.nr_wakeups);
2594 if (wake_flags & WF_SYNC)
2595 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2597 #endif /* CONFIG_SCHEDSTATS */
2600 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2602 activate_task(rq, p, en_flags);
2605 /* if a worker is waking up, notify workqueue */
2606 if (p->flags & PF_WQ_WORKER)
2607 wq_worker_waking_up(p, cpu_of(rq));
2611 * Mark the task runnable and perform wakeup-preemption.
2614 ttwu_do_wakeup(struct rq *rq, struct task_struct *p, int wake_flags)
2616 trace_sched_wakeup(p, true);
2617 check_preempt_curr(rq, p, wake_flags);
2619 p->state = TASK_RUNNING;
2621 if (p->sched_class->task_woken)
2622 p->sched_class->task_woken(rq, p);
2624 if (rq->idle_stamp) {
2625 u64 delta = rq->clock - rq->idle_stamp;
2626 u64 max = 2*sysctl_sched_migration_cost;
2631 update_avg(&rq->avg_idle, delta);
2638 ttwu_do_activate(struct rq *rq, struct task_struct *p, int wake_flags)
2641 if (p->sched_contributes_to_load)
2642 rq->nr_uninterruptible--;
2645 ttwu_activate(rq, p, ENQUEUE_WAKEUP | ENQUEUE_WAKING);
2646 ttwu_do_wakeup(rq, p, wake_flags);
2650 * Called in case the task @p isn't fully descheduled from its runqueue,
2651 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2652 * since all we need to do is flip p->state to TASK_RUNNING, since
2653 * the task is still ->on_rq.
2655 static int ttwu_remote(struct task_struct *p, int wake_flags)
2660 rq = __task_rq_lock(p);
2662 ttwu_do_wakeup(rq, p, wake_flags);
2665 __task_rq_unlock(rq);
2671 static void sched_ttwu_do_pending(struct task_struct *list)
2673 struct rq *rq = this_rq();
2675 raw_spin_lock(&rq->lock);
2678 struct task_struct *p = list;
2679 list = list->wake_entry;
2680 ttwu_do_activate(rq, p, 0);
2683 raw_spin_unlock(&rq->lock);
2686 #ifdef CONFIG_HOTPLUG_CPU
2688 static void sched_ttwu_pending(void)
2690 struct rq *rq = this_rq();
2691 struct task_struct *list = xchg(&rq->wake_list, NULL);
2696 sched_ttwu_do_pending(list);
2699 #endif /* CONFIG_HOTPLUG_CPU */
2701 void scheduler_ipi(void)
2703 struct rq *rq = this_rq();
2704 struct task_struct *list = xchg(&rq->wake_list, NULL);
2710 * Not all reschedule IPI handlers call irq_enter/irq_exit, since
2711 * traditionally all their work was done from the interrupt return
2712 * path. Now that we actually do some work, we need to make sure
2715 * Some archs already do call them, luckily irq_enter/exit nest
2718 * Arguably we should visit all archs and update all handlers,
2719 * however a fair share of IPIs are still resched only so this would
2720 * somewhat pessimize the simple resched case.
2723 sched_ttwu_do_pending(list);
2727 static void ttwu_queue_remote(struct task_struct *p, int cpu)
2729 struct rq *rq = cpu_rq(cpu);
2730 struct task_struct *next = rq->wake_list;
2733 struct task_struct *old = next;
2735 p->wake_entry = next;
2736 next = cmpxchg(&rq->wake_list, old, p);
2742 smp_send_reschedule(cpu);
2745 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2746 static int ttwu_activate_remote(struct task_struct *p, int wake_flags)
2751 rq = __task_rq_lock(p);
2753 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2754 ttwu_do_wakeup(rq, p, wake_flags);
2757 __task_rq_unlock(rq);
2762 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2763 #endif /* CONFIG_SMP */
2765 static void ttwu_queue(struct task_struct *p, int cpu)
2767 struct rq *rq = cpu_rq(cpu);
2769 #if defined(CONFIG_SMP)
2770 if (sched_feat(TTWU_QUEUE) && cpu != smp_processor_id()) {
2771 sched_clock_cpu(cpu); /* sync clocks x-cpu */
2772 ttwu_queue_remote(p, cpu);
2777 raw_spin_lock(&rq->lock);
2778 ttwu_do_activate(rq, p, 0);
2779 raw_spin_unlock(&rq->lock);
2783 * try_to_wake_up - wake up a thread
2784 * @p: the thread to be awakened
2785 * @state: the mask of task states that can be woken
2786 * @wake_flags: wake modifier flags (WF_*)
2788 * Put it on the run-queue if it's not already there. The "current"
2789 * thread is always on the run-queue (except when the actual
2790 * re-schedule is in progress), and as such you're allowed to do
2791 * the simpler "current->state = TASK_RUNNING" to mark yourself
2792 * runnable without the overhead of this.
2794 * Returns %true if @p was woken up, %false if it was already running
2795 * or @state didn't match @p's state.
2798 try_to_wake_up(struct task_struct *p, unsigned int state, int wake_flags)
2800 unsigned long flags;
2801 int cpu, success = 0;
2804 raw_spin_lock_irqsave(&p->pi_lock, flags);
2805 if (!(p->state & state))
2808 success = 1; /* we're going to change ->state */
2811 if (p->on_rq && ttwu_remote(p, wake_flags))
2816 * If the owning (remote) cpu is still in the middle of schedule() with
2817 * this task as prev, wait until its done referencing the task.
2820 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2822 * In case the architecture enables interrupts in
2823 * context_switch(), we cannot busy wait, since that
2824 * would lead to deadlocks when an interrupt hits and
2825 * tries to wake up @prev. So bail and do a complete
2828 if (ttwu_activate_remote(p, wake_flags))
2835 * Pairs with the smp_wmb() in finish_lock_switch().
2839 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2840 p->state = TASK_WAKING;
2842 if (p->sched_class->task_waking)
2843 p->sched_class->task_waking(p);
2845 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2846 if (task_cpu(p) != cpu) {
2847 wake_flags |= WF_MIGRATED;
2848 set_task_cpu(p, cpu);
2850 #endif /* CONFIG_SMP */
2854 ttwu_stat(p, cpu, wake_flags);
2856 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2862 * try_to_wake_up_local - try to wake up a local task with rq lock held
2863 * @p: the thread to be awakened
2865 * Put @p on the run-queue if it's not already there. The caller must
2866 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2869 static void try_to_wake_up_local(struct task_struct *p)
2871 struct rq *rq = task_rq(p);
2873 BUG_ON(rq != this_rq());
2874 BUG_ON(p == current);
2875 lockdep_assert_held(&rq->lock);
2877 if (!raw_spin_trylock(&p->pi_lock)) {
2878 raw_spin_unlock(&rq->lock);
2879 raw_spin_lock(&p->pi_lock);
2880 raw_spin_lock(&rq->lock);
2883 if (!(p->state & TASK_NORMAL))
2887 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2889 ttwu_do_wakeup(rq, p, 0);
2890 ttwu_stat(p, smp_processor_id(), 0);
2892 raw_spin_unlock(&p->pi_lock);
2896 * wake_up_process - Wake up a specific process
2897 * @p: The process to be woken up.
2899 * Attempt to wake up the nominated process and move it to the set of runnable
2900 * processes. Returns 1 if the process was woken up, 0 if it was already
2903 * It may be assumed that this function implies a write memory barrier before
2904 * changing the task state if and only if any tasks are woken up.
2906 int wake_up_process(struct task_struct *p)
2908 return try_to_wake_up(p, TASK_ALL, 0);
2910 EXPORT_SYMBOL(wake_up_process);
2912 int wake_up_state(struct task_struct *p, unsigned int state)
2914 return try_to_wake_up(p, state, 0);
2918 * Perform scheduler related setup for a newly forked process p.
2919 * p is forked by current.
2921 * __sched_fork() is basic setup used by init_idle() too:
2923 static void __sched_fork(struct task_struct *p)
2928 p->se.exec_start = 0;
2929 p->se.sum_exec_runtime = 0;
2930 p->se.prev_sum_exec_runtime = 0;
2931 p->se.nr_migrations = 0;
2933 INIT_LIST_HEAD(&p->se.group_node);
2935 #ifdef CONFIG_SCHEDSTATS
2936 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2939 INIT_LIST_HEAD(&p->rt.run_list);
2941 #ifdef CONFIG_PREEMPT_NOTIFIERS
2942 INIT_HLIST_HEAD(&p->preempt_notifiers);
2947 * fork()/clone()-time setup:
2949 void sched_fork(struct task_struct *p)
2951 unsigned long flags;
2952 int cpu = get_cpu();
2956 * We mark the process as running here. This guarantees that
2957 * nobody will actually run it, and a signal or other external
2958 * event cannot wake it up and insert it on the runqueue either.
2960 p->state = TASK_RUNNING;
2963 * Make sure we do not leak PI boosting priority to the child.
2965 p->prio = current->normal_prio;
2968 * Revert to default priority/policy on fork if requested.
2970 if (unlikely(p->sched_reset_on_fork)) {
2971 if (task_has_rt_policy(p)) {
2972 p->policy = SCHED_NORMAL;
2973 p->static_prio = NICE_TO_PRIO(0);
2975 } else if (PRIO_TO_NICE(p->static_prio) < 0)
2976 p->static_prio = NICE_TO_PRIO(0);
2978 p->prio = p->normal_prio = __normal_prio(p);
2982 * We don't need the reset flag anymore after the fork. It has
2983 * fulfilled its duty:
2985 p->sched_reset_on_fork = 0;
2988 if (!rt_prio(p->prio))
2989 p->sched_class = &fair_sched_class;
2991 if (p->sched_class->task_fork)
2992 p->sched_class->task_fork(p);
2995 * The child is not yet in the pid-hash so no cgroup attach races,
2996 * and the cgroup is pinned to this child due to cgroup_fork()
2997 * is ran before sched_fork().
2999 * Silence PROVE_RCU.
3001 raw_spin_lock_irqsave(&p->pi_lock, flags);
3002 set_task_cpu(p, cpu);
3003 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3005 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
3006 if (likely(sched_info_on()))
3007 memset(&p->sched_info, 0, sizeof(p->sched_info));
3009 #if defined(CONFIG_SMP)
3012 #ifdef CONFIG_PREEMPT_COUNT
3013 /* Want to start with kernel preemption disabled. */
3014 task_thread_info(p)->preempt_count = 1;
3017 plist_node_init(&p->pushable_tasks, MAX_PRIO);
3024 * wake_up_new_task - wake up a newly created task for the first time.
3026 * This function will do some initial scheduler statistics housekeeping
3027 * that must be done for every newly created context, then puts the task
3028 * on the runqueue and wakes it.
3030 void wake_up_new_task(struct task_struct *p)
3032 unsigned long flags;
3035 raw_spin_lock_irqsave(&p->pi_lock, flags);
3038 * Fork balancing, do it here and not earlier because:
3039 * - cpus_allowed can change in the fork path
3040 * - any previously selected cpu might disappear through hotplug
3042 set_task_cpu(p, select_task_rq(p, SD_BALANCE_FORK, 0));
3045 rq = __task_rq_lock(p);
3046 activate_task(rq, p, 0);
3048 trace_sched_wakeup_new(p, true);
3049 check_preempt_curr(rq, p, WF_FORK);
3051 if (p->sched_class->task_woken)
3052 p->sched_class->task_woken(rq, p);
3054 task_rq_unlock(rq, p, &flags);
3057 #ifdef CONFIG_PREEMPT_NOTIFIERS
3060 * preempt_notifier_register - tell me when current is being preempted & rescheduled
3061 * @notifier: notifier struct to register
3063 void preempt_notifier_register(struct preempt_notifier *notifier)
3065 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
3067 EXPORT_SYMBOL_GPL(preempt_notifier_register);
3070 * preempt_notifier_unregister - no longer interested in preemption notifications
3071 * @notifier: notifier struct to unregister
3073 * This is safe to call from within a preemption notifier.
3075 void preempt_notifier_unregister(struct preempt_notifier *notifier)
3077 hlist_del(¬ifier->link);
3079 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
3081 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3083 struct preempt_notifier *notifier;
3084 struct hlist_node *node;
3086 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3087 notifier->ops->sched_in(notifier, raw_smp_processor_id());
3091 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3092 struct task_struct *next)
3094 struct preempt_notifier *notifier;
3095 struct hlist_node *node;
3097 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
3098 notifier->ops->sched_out(notifier, next);
3101 #else /* !CONFIG_PREEMPT_NOTIFIERS */
3103 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
3108 fire_sched_out_preempt_notifiers(struct task_struct *curr,
3109 struct task_struct *next)
3113 #endif /* CONFIG_PREEMPT_NOTIFIERS */
3116 * prepare_task_switch - prepare to switch tasks
3117 * @rq: the runqueue preparing to switch
3118 * @prev: the current task that is being switched out
3119 * @next: the task we are going to switch to.
3121 * This is called with the rq lock held and interrupts off. It must
3122 * be paired with a subsequent finish_task_switch after the context
3125 * prepare_task_switch sets up locking and calls architecture specific
3129 prepare_task_switch(struct rq *rq, struct task_struct *prev,
3130 struct task_struct *next)
3132 sched_info_switch(prev, next);
3133 perf_event_task_sched_out(prev, next);
3134 fire_sched_out_preempt_notifiers(prev, next);
3135 prepare_lock_switch(rq, next);
3136 prepare_arch_switch(next);
3137 trace_sched_switch(prev, next);
3141 * finish_task_switch - clean up after a task-switch
3142 * @rq: runqueue associated with task-switch
3143 * @prev: the thread we just switched away from.
3145 * finish_task_switch must be called after the context switch, paired
3146 * with a prepare_task_switch call before the context switch.
3147 * finish_task_switch will reconcile locking set up by prepare_task_switch,
3148 * and do any other architecture-specific cleanup actions.
3150 * Note that we may have delayed dropping an mm in context_switch(). If
3151 * so, we finish that here outside of the runqueue lock. (Doing it
3152 * with the lock held can cause deadlocks; see schedule() for
3155 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
3156 __releases(rq->lock)
3158 struct mm_struct *mm = rq->prev_mm;
3164 * A task struct has one reference for the use as "current".
3165 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
3166 * schedule one last time. The schedule call will never return, and
3167 * the scheduled task must drop that reference.
3168 * The test for TASK_DEAD must occur while the runqueue locks are
3169 * still held, otherwise prev could be scheduled on another cpu, die
3170 * there before we look at prev->state, and then the reference would
3172 * Manfred Spraul <manfred@colorfullife.com>
3174 prev_state = prev->state;
3175 finish_arch_switch(prev);
3176 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3177 local_irq_disable();
3178 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3179 perf_event_task_sched_in(current);
3180 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
3182 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
3183 finish_lock_switch(rq, prev);
3185 fire_sched_in_preempt_notifiers(current);
3188 if (unlikely(prev_state == TASK_DEAD)) {
3190 * Remove function-return probe instances associated with this
3191 * task and put them back on the free list.
3193 kprobe_flush_task(prev);
3194 put_task_struct(prev);
3200 /* assumes rq->lock is held */
3201 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
3203 if (prev->sched_class->pre_schedule)
3204 prev->sched_class->pre_schedule(rq, prev);
3207 /* rq->lock is NOT held, but preemption is disabled */
3208 static inline void post_schedule(struct rq *rq)
3210 if (rq->post_schedule) {
3211 unsigned long flags;
3213 raw_spin_lock_irqsave(&rq->lock, flags);
3214 if (rq->curr->sched_class->post_schedule)
3215 rq->curr->sched_class->post_schedule(rq);
3216 raw_spin_unlock_irqrestore(&rq->lock, flags);
3218 rq->post_schedule = 0;
3224 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
3228 static inline void post_schedule(struct rq *rq)
3235 * schedule_tail - first thing a freshly forked thread must call.
3236 * @prev: the thread we just switched away from.
3238 asmlinkage void schedule_tail(struct task_struct *prev)
3239 __releases(rq->lock)
3241 struct rq *rq = this_rq();
3243 finish_task_switch(rq, prev);
3246 * FIXME: do we need to worry about rq being invalidated by the
3251 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3252 /* In this case, finish_task_switch does not reenable preemption */
3255 if (current->set_child_tid)
3256 put_user(task_pid_vnr(current), current->set_child_tid);
3260 * context_switch - switch to the new MM and the new
3261 * thread's register state.
3264 context_switch(struct rq *rq, struct task_struct *prev,
3265 struct task_struct *next)
3267 struct mm_struct *mm, *oldmm;
3269 prepare_task_switch(rq, prev, next);
3272 oldmm = prev->active_mm;
3274 * For paravirt, this is coupled with an exit in switch_to to
3275 * combine the page table reload and the switch backend into
3278 arch_start_context_switch(prev);
3281 next->active_mm = oldmm;
3282 atomic_inc(&oldmm->mm_count);
3283 enter_lazy_tlb(oldmm, next);
3285 switch_mm(oldmm, mm, next);
3288 prev->active_mm = NULL;
3289 rq->prev_mm = oldmm;
3292 * Since the runqueue lock will be released by the next
3293 * task (which is an invalid locking op but in the case
3294 * of the scheduler it's an obvious special-case), so we
3295 * do an early lockdep release here:
3297 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3298 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3301 /* Here we just switch the register state and the stack. */
3302 switch_to(prev, next, prev);
3306 * this_rq must be evaluated again because prev may have moved
3307 * CPUs since it called schedule(), thus the 'rq' on its stack
3308 * frame will be invalid.
3310 finish_task_switch(this_rq(), prev);
3314 * nr_running, nr_uninterruptible and nr_context_switches:
3316 * externally visible scheduler statistics: current number of runnable
3317 * threads, current number of uninterruptible-sleeping threads, total
3318 * number of context switches performed since bootup.
3320 unsigned long nr_running(void)
3322 unsigned long i, sum = 0;
3324 for_each_online_cpu(i)
3325 sum += cpu_rq(i)->nr_running;
3330 unsigned long nr_uninterruptible(void)
3332 unsigned long i, sum = 0;
3334 for_each_possible_cpu(i)
3335 sum += cpu_rq(i)->nr_uninterruptible;
3338 * Since we read the counters lockless, it might be slightly
3339 * inaccurate. Do not allow it to go below zero though:
3341 if (unlikely((long)sum < 0))
3347 unsigned long long nr_context_switches(void)
3350 unsigned long long sum = 0;
3352 for_each_possible_cpu(i)
3353 sum += cpu_rq(i)->nr_switches;
3358 unsigned long nr_iowait(void)
3360 unsigned long i, sum = 0;
3362 for_each_possible_cpu(i)
3363 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3368 unsigned long nr_iowait_cpu(int cpu)
3370 struct rq *this = cpu_rq(cpu);
3371 return atomic_read(&this->nr_iowait);
3374 unsigned long this_cpu_load(void)
3376 struct rq *this = this_rq();
3377 return this->cpu_load[0];
3381 /* Variables and functions for calc_load */
3382 static atomic_long_t calc_load_tasks;
3383 static unsigned long calc_load_update;
3384 unsigned long avenrun[3];
3385 EXPORT_SYMBOL(avenrun);
3387 static long calc_load_fold_active(struct rq *this_rq)
3389 long nr_active, delta = 0;
3391 nr_active = this_rq->nr_running;
3392 nr_active += (long) this_rq->nr_uninterruptible;
3394 if (nr_active != this_rq->calc_load_active) {
3395 delta = nr_active - this_rq->calc_load_active;
3396 this_rq->calc_load_active = nr_active;
3402 static unsigned long
3403 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3406 load += active * (FIXED_1 - exp);
3407 load += 1UL << (FSHIFT - 1);
3408 return load >> FSHIFT;
3413 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3415 * When making the ILB scale, we should try to pull this in as well.
3417 static atomic_long_t calc_load_tasks_idle;
3419 static void calc_load_account_idle(struct rq *this_rq)
3423 delta = calc_load_fold_active(this_rq);
3425 atomic_long_add(delta, &calc_load_tasks_idle);
3428 static long calc_load_fold_idle(void)
3433 * Its got a race, we don't care...
3435 if (atomic_long_read(&calc_load_tasks_idle))
3436 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3442 * fixed_power_int - compute: x^n, in O(log n) time
3444 * @x: base of the power
3445 * @frac_bits: fractional bits of @x
3446 * @n: power to raise @x to.
3448 * By exploiting the relation between the definition of the natural power
3449 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3450 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3451 * (where: n_i \elem {0, 1}, the binary vector representing n),
3452 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3453 * of course trivially computable in O(log_2 n), the length of our binary
3456 static unsigned long
3457 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3459 unsigned long result = 1UL << frac_bits;
3464 result += 1UL << (frac_bits - 1);
3465 result >>= frac_bits;
3471 x += 1UL << (frac_bits - 1);
3479 * a1 = a0 * e + a * (1 - e)
3481 * a2 = a1 * e + a * (1 - e)
3482 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3483 * = a0 * e^2 + a * (1 - e) * (1 + e)
3485 * a3 = a2 * e + a * (1 - e)
3486 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3487 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3491 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3492 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3493 * = a0 * e^n + a * (1 - e^n)
3495 * [1] application of the geometric series:
3498 * S_n := \Sum x^i = -------------
3501 static unsigned long
3502 calc_load_n(unsigned long load, unsigned long exp,
3503 unsigned long active, unsigned int n)
3506 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3510 * NO_HZ can leave us missing all per-cpu ticks calling
3511 * calc_load_account_active(), but since an idle CPU folds its delta into
3512 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3513 * in the pending idle delta if our idle period crossed a load cycle boundary.
3515 * Once we've updated the global active value, we need to apply the exponential
3516 * weights adjusted to the number of cycles missed.
3518 static void calc_global_nohz(unsigned long ticks)
3520 long delta, active, n;
3522 if (time_before(jiffies, calc_load_update))
3526 * If we crossed a calc_load_update boundary, make sure to fold
3527 * any pending idle changes, the respective CPUs might have
3528 * missed the tick driven calc_load_account_active() update
3531 delta = calc_load_fold_idle();
3533 atomic_long_add(delta, &calc_load_tasks);
3536 * If we were idle for multiple load cycles, apply them.
3538 if (ticks >= LOAD_FREQ) {
3539 n = ticks / LOAD_FREQ;
3541 active = atomic_long_read(&calc_load_tasks);
3542 active = active > 0 ? active * FIXED_1 : 0;
3544 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3545 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3546 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3548 calc_load_update += n * LOAD_FREQ;
3552 * Its possible the remainder of the above division also crosses
3553 * a LOAD_FREQ period, the regular check in calc_global_load()
3554 * which comes after this will take care of that.
3556 * Consider us being 11 ticks before a cycle completion, and us
3557 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3558 * age us 4 cycles, and the test in calc_global_load() will
3559 * pick up the final one.
3563 static void calc_load_account_idle(struct rq *this_rq)
3567 static inline long calc_load_fold_idle(void)
3572 static void calc_global_nohz(unsigned long ticks)
3578 * get_avenrun - get the load average array
3579 * @loads: pointer to dest load array
3580 * @offset: offset to add
3581 * @shift: shift count to shift the result left
3583 * These values are estimates at best, so no need for locking.
3585 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3587 loads[0] = (avenrun[0] + offset) << shift;
3588 loads[1] = (avenrun[1] + offset) << shift;
3589 loads[2] = (avenrun[2] + offset) << shift;
3593 * calc_load - update the avenrun load estimates 10 ticks after the
3594 * CPUs have updated calc_load_tasks.
3596 void calc_global_load(unsigned long ticks)
3600 calc_global_nohz(ticks);
3602 if (time_before(jiffies, calc_load_update + 10))
3605 active = atomic_long_read(&calc_load_tasks);
3606 active = active > 0 ? active * FIXED_1 : 0;
3608 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3609 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3610 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3612 calc_load_update += LOAD_FREQ;
3616 * Called from update_cpu_load() to periodically update this CPU's
3619 static void calc_load_account_active(struct rq *this_rq)
3623 if (time_before(jiffies, this_rq->calc_load_update))
3626 delta = calc_load_fold_active(this_rq);
3627 delta += calc_load_fold_idle();
3629 atomic_long_add(delta, &calc_load_tasks);
3631 this_rq->calc_load_update += LOAD_FREQ;
3635 * The exact cpuload at various idx values, calculated at every tick would be
3636 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3638 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3639 * on nth tick when cpu may be busy, then we have:
3640 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3641 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3643 * decay_load_missed() below does efficient calculation of
3644 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3645 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3647 * The calculation is approximated on a 128 point scale.
3648 * degrade_zero_ticks is the number of ticks after which load at any
3649 * particular idx is approximated to be zero.
3650 * degrade_factor is a precomputed table, a row for each load idx.
3651 * Each column corresponds to degradation factor for a power of two ticks,
3652 * based on 128 point scale.
3654 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3655 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3657 * With this power of 2 load factors, we can degrade the load n times
3658 * by looking at 1 bits in n and doing as many mult/shift instead of
3659 * n mult/shifts needed by the exact degradation.
3661 #define DEGRADE_SHIFT 7
3662 static const unsigned char
3663 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3664 static const unsigned char
3665 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3666 {0, 0, 0, 0, 0, 0, 0, 0},
3667 {64, 32, 8, 0, 0, 0, 0, 0},
3668 {96, 72, 40, 12, 1, 0, 0},
3669 {112, 98, 75, 43, 15, 1, 0},
3670 {120, 112, 98, 76, 45, 16, 2} };
3673 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3674 * would be when CPU is idle and so we just decay the old load without
3675 * adding any new load.
3677 static unsigned long
3678 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3682 if (!missed_updates)
3685 if (missed_updates >= degrade_zero_ticks[idx])
3689 return load >> missed_updates;
3691 while (missed_updates) {
3692 if (missed_updates % 2)
3693 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3695 missed_updates >>= 1;
3702 * Update rq->cpu_load[] statistics. This function is usually called every
3703 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3704 * every tick. We fix it up based on jiffies.
3706 static void update_cpu_load(struct rq *this_rq)
3708 unsigned long this_load = this_rq->load.weight;
3709 unsigned long curr_jiffies = jiffies;
3710 unsigned long pending_updates;
3713 this_rq->nr_load_updates++;
3715 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3716 if (curr_jiffies == this_rq->last_load_update_tick)
3719 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3720 this_rq->last_load_update_tick = curr_jiffies;
3722 /* Update our load: */
3723 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3724 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3725 unsigned long old_load, new_load;
3727 /* scale is effectively 1 << i now, and >> i divides by scale */
3729 old_load = this_rq->cpu_load[i];
3730 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3731 new_load = this_load;
3733 * Round up the averaging division if load is increasing. This
3734 * prevents us from getting stuck on 9 if the load is 10, for
3737 if (new_load > old_load)
3738 new_load += scale - 1;
3740 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3743 sched_avg_update(this_rq);
3746 static void update_cpu_load_active(struct rq *this_rq)
3748 update_cpu_load(this_rq);
3750 calc_load_account_active(this_rq);
3756 * sched_exec - execve() is a valuable balancing opportunity, because at
3757 * this point the task has the smallest effective memory and cache footprint.
3759 void sched_exec(void)
3761 struct task_struct *p = current;
3762 unsigned long flags;
3765 raw_spin_lock_irqsave(&p->pi_lock, flags);
3766 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3767 if (dest_cpu == smp_processor_id())
3770 if (likely(cpu_active(dest_cpu))) {
3771 struct migration_arg arg = { p, dest_cpu };
3773 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3774 stop_one_cpu(task_cpu(p), migration_cpu_stop, &arg);
3778 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
3783 DEFINE_PER_CPU(struct kernel_stat, kstat);
3785 EXPORT_PER_CPU_SYMBOL(kstat);
3788 * Return any ns on the sched_clock that have not yet been accounted in
3789 * @p in case that task is currently running.
3791 * Called with task_rq_lock() held on @rq.
3793 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3797 if (task_current(rq, p)) {
3798 update_rq_clock(rq);
3799 ns = rq->clock_task - p->se.exec_start;
3807 unsigned long long task_delta_exec(struct task_struct *p)
3809 unsigned long flags;
3813 rq = task_rq_lock(p, &flags);
3814 ns = do_task_delta_exec(p, rq);
3815 task_rq_unlock(rq, p, &flags);
3821 * Return accounted runtime for the task.
3822 * In case the task is currently running, return the runtime plus current's
3823 * pending runtime that have not been accounted yet.
3825 unsigned long long task_sched_runtime(struct task_struct *p)
3827 unsigned long flags;
3831 rq = task_rq_lock(p, &flags);
3832 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3833 task_rq_unlock(rq, p, &flags);
3839 * Return sum_exec_runtime for the thread group.
3840 * In case the task is currently running, return the sum plus current's
3841 * pending runtime that have not been accounted yet.
3843 * Note that the thread group might have other running tasks as well,
3844 * so the return value not includes other pending runtime that other
3845 * running tasks might have.
3847 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3849 struct task_cputime totals;
3850 unsigned long flags;
3854 rq = task_rq_lock(p, &flags);
3855 thread_group_cputime(p, &totals);
3856 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3857 task_rq_unlock(rq, p, &flags);
3863 * Account user cpu time to a process.
3864 * @p: the process that the cpu time gets accounted to
3865 * @cputime: the cpu time spent in user space since the last update
3866 * @cputime_scaled: cputime scaled by cpu frequency
3868 void account_user_time(struct task_struct *p, cputime_t cputime,
3869 cputime_t cputime_scaled)
3871 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3874 /* Add user time to process. */
3875 p->utime = cputime_add(p->utime, cputime);
3876 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3877 account_group_user_time(p, cputime);
3879 /* Add user time to cpustat. */
3880 tmp = cputime_to_cputime64(cputime);
3881 if (TASK_NICE(p) > 0)
3882 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3884 cpustat->user = cputime64_add(cpustat->user, tmp);
3886 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3887 /* Account for user time used */
3888 acct_update_integrals(p);
3892 * Account guest cpu time to a process.
3893 * @p: the process that the cpu time gets accounted to
3894 * @cputime: the cpu time spent in virtual machine since the last update
3895 * @cputime_scaled: cputime scaled by cpu frequency
3897 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3898 cputime_t cputime_scaled)
3901 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3903 tmp = cputime_to_cputime64(cputime);
3905 /* Add guest time to process. */
3906 p->utime = cputime_add(p->utime, cputime);
3907 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3908 account_group_user_time(p, cputime);
3909 p->gtime = cputime_add(p->gtime, cputime);
3911 /* Add guest time to cpustat. */
3912 if (TASK_NICE(p) > 0) {
3913 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3914 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3916 cpustat->user = cputime64_add(cpustat->user, tmp);
3917 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3922 * Account system cpu time to a process and desired cpustat field
3923 * @p: the process that the cpu time gets accounted to
3924 * @cputime: the cpu time spent in kernel space since the last update
3925 * @cputime_scaled: cputime scaled by cpu frequency
3926 * @target_cputime64: pointer to cpustat field that has to be updated
3929 void __account_system_time(struct task_struct *p, cputime_t cputime,
3930 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3932 cputime64_t tmp = cputime_to_cputime64(cputime);
3934 /* Add system time to process. */
3935 p->stime = cputime_add(p->stime, cputime);
3936 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3937 account_group_system_time(p, cputime);
3939 /* Add system time to cpustat. */
3940 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3941 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3943 /* Account for system time used */
3944 acct_update_integrals(p);
3948 * Account system cpu time to a process.
3949 * @p: the process that the cpu time gets accounted to
3950 * @hardirq_offset: the offset to subtract from hardirq_count()
3951 * @cputime: the cpu time spent in kernel space since the last update
3952 * @cputime_scaled: cputime scaled by cpu frequency
3954 void account_system_time(struct task_struct *p, int hardirq_offset,
3955 cputime_t cputime, cputime_t cputime_scaled)
3957 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3958 cputime64_t *target_cputime64;
3960 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3961 account_guest_time(p, cputime, cputime_scaled);
3965 if (hardirq_count() - hardirq_offset)
3966 target_cputime64 = &cpustat->irq;
3967 else if (in_serving_softirq())
3968 target_cputime64 = &cpustat->softirq;
3970 target_cputime64 = &cpustat->system;
3972 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3976 * Account for involuntary wait time.
3977 * @cputime: the cpu time spent in involuntary wait
3979 void account_steal_time(cputime_t cputime)
3981 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3982 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3984 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3988 * Account for idle time.
3989 * @cputime: the cpu time spent in idle wait
3991 void account_idle_time(cputime_t cputime)
3993 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3994 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3995 struct rq *rq = this_rq();
3997 if (atomic_read(&rq->nr_iowait) > 0)
3998 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4000 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4003 static __always_inline bool steal_account_process_tick(void)
4005 #ifdef CONFIG_PARAVIRT
4006 if (static_branch(¶virt_steal_enabled)) {
4009 steal = paravirt_steal_clock(smp_processor_id());
4010 steal -= this_rq()->prev_steal_time;
4012 st = steal_ticks(steal);
4013 this_rq()->prev_steal_time += st * TICK_NSEC;
4015 account_steal_time(st);
4022 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4024 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
4026 * Account a tick to a process and cpustat
4027 * @p: the process that the cpu time gets accounted to
4028 * @user_tick: is the tick from userspace
4029 * @rq: the pointer to rq
4031 * Tick demultiplexing follows the order
4032 * - pending hardirq update
4033 * - pending softirq update
4037 * - check for guest_time
4038 * - else account as system_time
4040 * Check for hardirq is done both for system and user time as there is
4041 * no timer going off while we are on hardirq and hence we may never get an
4042 * opportunity to update it solely in system time.
4043 * p->stime and friends are only updated on system time and not on irq
4044 * softirq as those do not count in task exec_runtime any more.
4046 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4049 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4050 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
4051 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4053 if (steal_account_process_tick())
4056 if (irqtime_account_hi_update()) {
4057 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4058 } else if (irqtime_account_si_update()) {
4059 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4060 } else if (this_cpu_ksoftirqd() == p) {
4062 * ksoftirqd time do not get accounted in cpu_softirq_time.
4063 * So, we have to handle it separately here.
4064 * Also, p->stime needs to be updated for ksoftirqd.
4066 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4068 } else if (user_tick) {
4069 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4070 } else if (p == rq->idle) {
4071 account_idle_time(cputime_one_jiffy);
4072 } else if (p->flags & PF_VCPU) { /* System time or guest time */
4073 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
4075 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
4080 static void irqtime_account_idle_ticks(int ticks)
4083 struct rq *rq = this_rq();
4085 for (i = 0; i < ticks; i++)
4086 irqtime_account_process_tick(current, 0, rq);
4088 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
4089 static void irqtime_account_idle_ticks(int ticks) {}
4090 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
4092 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
4095 * Account a single tick of cpu time.
4096 * @p: the process that the cpu time gets accounted to
4097 * @user_tick: indicates if the tick is a user or a system tick
4099 void account_process_tick(struct task_struct *p, int user_tick)
4101 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
4102 struct rq *rq = this_rq();
4104 if (sched_clock_irqtime) {
4105 irqtime_account_process_tick(p, user_tick, rq);
4109 if (steal_account_process_tick())
4113 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
4114 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4115 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
4118 account_idle_time(cputime_one_jiffy);
4122 * Account multiple ticks of steal time.
4123 * @p: the process from which the cpu time has been stolen
4124 * @ticks: number of stolen ticks
4126 void account_steal_ticks(unsigned long ticks)
4128 account_steal_time(jiffies_to_cputime(ticks));
4132 * Account multiple ticks of idle time.
4133 * @ticks: number of stolen ticks
4135 void account_idle_ticks(unsigned long ticks)
4138 if (sched_clock_irqtime) {
4139 irqtime_account_idle_ticks(ticks);
4143 account_idle_time(jiffies_to_cputime(ticks));
4149 * Use precise platform statistics if available:
4151 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4152 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4158 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4160 struct task_cputime cputime;
4162 thread_group_cputime(p, &cputime);
4164 *ut = cputime.utime;
4165 *st = cputime.stime;
4169 #ifndef nsecs_to_cputime
4170 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
4173 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4175 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
4178 * Use CFS's precise accounting:
4180 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
4186 do_div(temp, total);
4187 utime = (cputime_t)temp;
4192 * Compare with previous values, to keep monotonicity:
4194 p->prev_utime = max(p->prev_utime, utime);
4195 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
4197 *ut = p->prev_utime;
4198 *st = p->prev_stime;
4202 * Must be called with siglock held.
4204 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
4206 struct signal_struct *sig = p->signal;
4207 struct task_cputime cputime;
4208 cputime_t rtime, utime, total;
4210 thread_group_cputime(p, &cputime);
4212 total = cputime_add(cputime.utime, cputime.stime);
4213 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
4218 temp *= cputime.utime;
4219 do_div(temp, total);
4220 utime = (cputime_t)temp;
4224 sig->prev_utime = max(sig->prev_utime, utime);
4225 sig->prev_stime = max(sig->prev_stime,
4226 cputime_sub(rtime, sig->prev_utime));
4228 *ut = sig->prev_utime;
4229 *st = sig->prev_stime;
4234 * This function gets called by the timer code, with HZ frequency.
4235 * We call it with interrupts disabled.
4237 void scheduler_tick(void)
4239 int cpu = smp_processor_id();
4240 struct rq *rq = cpu_rq(cpu);
4241 struct task_struct *curr = rq->curr;
4245 raw_spin_lock(&rq->lock);
4246 update_rq_clock(rq);
4247 update_cpu_load_active(rq);
4248 curr->sched_class->task_tick(rq, curr, 0);
4249 raw_spin_unlock(&rq->lock);
4251 perf_event_task_tick();
4254 rq->idle_at_tick = idle_cpu(cpu);
4255 trigger_load_balance(rq, cpu);
4259 notrace unsigned long get_parent_ip(unsigned long addr)
4261 if (in_lock_functions(addr)) {
4262 addr = CALLER_ADDR2;
4263 if (in_lock_functions(addr))
4264 addr = CALLER_ADDR3;
4269 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4270 defined(CONFIG_PREEMPT_TRACER))
4272 void __kprobes add_preempt_count(int val)
4274 #ifdef CONFIG_DEBUG_PREEMPT
4278 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4281 preempt_count() += val;
4282 #ifdef CONFIG_DEBUG_PREEMPT
4284 * Spinlock count overflowing soon?
4286 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4289 if (preempt_count() == val)
4290 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4292 EXPORT_SYMBOL(add_preempt_count);
4294 void __kprobes sub_preempt_count(int val)
4296 #ifdef CONFIG_DEBUG_PREEMPT
4300 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4303 * Is the spinlock portion underflowing?
4305 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4306 !(preempt_count() & PREEMPT_MASK)))
4310 if (preempt_count() == val)
4311 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4312 preempt_count() -= val;
4314 EXPORT_SYMBOL(sub_preempt_count);
4319 * Print scheduling while atomic bug:
4321 static noinline void __schedule_bug(struct task_struct *prev)
4323 struct pt_regs *regs = get_irq_regs();
4325 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4326 prev->comm, prev->pid, preempt_count());
4328 debug_show_held_locks(prev);
4330 if (irqs_disabled())
4331 print_irqtrace_events(prev);
4340 * Various schedule()-time debugging checks and statistics:
4342 static inline void schedule_debug(struct task_struct *prev)
4345 * Test if we are atomic. Since do_exit() needs to call into
4346 * schedule() atomically, we ignore that path for now.
4347 * Otherwise, whine if we are scheduling when we should not be.
4349 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4350 __schedule_bug(prev);
4352 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4354 schedstat_inc(this_rq(), sched_count);
4357 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4359 if (prev->on_rq || rq->skip_clock_update < 0)
4360 update_rq_clock(rq);
4361 prev->sched_class->put_prev_task(rq, prev);
4365 * Pick up the highest-prio task:
4367 static inline struct task_struct *
4368 pick_next_task(struct rq *rq)
4370 const struct sched_class *class;
4371 struct task_struct *p;
4374 * Optimization: we know that if all tasks are in
4375 * the fair class we can call that function directly:
4377 if (likely(rq->nr_running == rq->cfs.h_nr_running)) {
4378 p = fair_sched_class.pick_next_task(rq);
4383 for_each_class(class) {
4384 p = class->pick_next_task(rq);
4389 BUG(); /* the idle class will always have a runnable task */
4393 * schedule() is the main scheduler function.
4395 asmlinkage void __sched schedule(void)
4397 struct task_struct *prev, *next;
4398 unsigned long *switch_count;
4404 cpu = smp_processor_id();
4406 rcu_note_context_switch(cpu);
4409 schedule_debug(prev);
4411 if (sched_feat(HRTICK))
4414 raw_spin_lock_irq(&rq->lock);
4416 switch_count = &prev->nivcsw;
4417 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4418 if (unlikely(signal_pending_state(prev->state, prev))) {
4419 prev->state = TASK_RUNNING;
4421 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4425 * If a worker went to sleep, notify and ask workqueue
4426 * whether it wants to wake up a task to maintain
4429 if (prev->flags & PF_WQ_WORKER) {
4430 struct task_struct *to_wakeup;
4432 to_wakeup = wq_worker_sleeping(prev, cpu);
4434 try_to_wake_up_local(to_wakeup);
4438 * If we are going to sleep and we have plugged IO
4439 * queued, make sure to submit it to avoid deadlocks.
4441 if (blk_needs_flush_plug(prev)) {
4442 raw_spin_unlock(&rq->lock);
4443 blk_schedule_flush_plug(prev);
4444 raw_spin_lock(&rq->lock);
4447 switch_count = &prev->nvcsw;
4450 pre_schedule(rq, prev);
4452 if (unlikely(!rq->nr_running))
4453 idle_balance(cpu, rq);
4455 put_prev_task(rq, prev);
4456 next = pick_next_task(rq);
4457 clear_tsk_need_resched(prev);
4458 rq->skip_clock_update = 0;
4460 if (likely(prev != next)) {
4465 context_switch(rq, prev, next); /* unlocks the rq */
4467 * The context switch have flipped the stack from under us
4468 * and restored the local variables which were saved when
4469 * this task called schedule() in the past. prev == current
4470 * is still correct, but it can be moved to another cpu/rq.
4472 cpu = smp_processor_id();
4475 raw_spin_unlock_irq(&rq->lock);
4479 preempt_enable_no_resched();
4483 EXPORT_SYMBOL(schedule);
4485 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4487 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4489 if (lock->owner != owner)
4493 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4494 * lock->owner still matches owner, if that fails, owner might
4495 * point to free()d memory, if it still matches, the rcu_read_lock()
4496 * ensures the memory stays valid.
4500 return owner->on_cpu;
4504 * Look out! "owner" is an entirely speculative pointer
4505 * access and not reliable.
4507 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4509 if (!sched_feat(OWNER_SPIN))
4513 while (owner_running(lock, owner)) {
4517 arch_mutex_cpu_relax();
4522 * We break out the loop above on need_resched() and when the
4523 * owner changed, which is a sign for heavy contention. Return
4524 * success only when lock->owner is NULL.
4526 return lock->owner == NULL;
4530 #ifdef CONFIG_PREEMPT
4532 * this is the entry point to schedule() from in-kernel preemption
4533 * off of preempt_enable. Kernel preemptions off return from interrupt
4534 * occur there and call schedule directly.
4536 asmlinkage void __sched notrace preempt_schedule(void)
4538 struct thread_info *ti = current_thread_info();
4541 * If there is a non-zero preempt_count or interrupts are disabled,
4542 * we do not want to preempt the current task. Just return..
4544 if (likely(ti->preempt_count || irqs_disabled()))
4548 add_preempt_count_notrace(PREEMPT_ACTIVE);
4550 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4553 * Check again in case we missed a preemption opportunity
4554 * between schedule and now.
4557 } while (need_resched());
4559 EXPORT_SYMBOL(preempt_schedule);
4562 * this is the entry point to schedule() from kernel preemption
4563 * off of irq context.
4564 * Note, that this is called and return with irqs disabled. This will
4565 * protect us against recursive calling from irq.
4567 asmlinkage void __sched preempt_schedule_irq(void)
4569 struct thread_info *ti = current_thread_info();
4571 /* Catch callers which need to be fixed */
4572 BUG_ON(ti->preempt_count || !irqs_disabled());
4575 add_preempt_count(PREEMPT_ACTIVE);
4578 local_irq_disable();
4579 sub_preempt_count(PREEMPT_ACTIVE);
4582 * Check again in case we missed a preemption opportunity
4583 * between schedule and now.
4586 } while (need_resched());
4589 #endif /* CONFIG_PREEMPT */
4591 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4594 return try_to_wake_up(curr->private, mode, wake_flags);
4596 EXPORT_SYMBOL(default_wake_function);
4599 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4600 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4601 * number) then we wake all the non-exclusive tasks and one exclusive task.
4603 * There are circumstances in which we can try to wake a task which has already
4604 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4605 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4607 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4608 int nr_exclusive, int wake_flags, void *key)
4610 wait_queue_t *curr, *next;
4612 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4613 unsigned flags = curr->flags;
4615 if (curr->func(curr, mode, wake_flags, key) &&
4616 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4622 * __wake_up - wake up threads blocked on a waitqueue.
4624 * @mode: which threads
4625 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4626 * @key: is directly passed to the wakeup function
4628 * It may be assumed that this function implies a write memory barrier before
4629 * changing the task state if and only if any tasks are woken up.
4631 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4632 int nr_exclusive, void *key)
4634 unsigned long flags;
4636 spin_lock_irqsave(&q->lock, flags);
4637 __wake_up_common(q, mode, nr_exclusive, 0, key);
4638 spin_unlock_irqrestore(&q->lock, flags);
4640 EXPORT_SYMBOL(__wake_up);
4643 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4645 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4647 __wake_up_common(q, mode, 1, 0, NULL);
4649 EXPORT_SYMBOL_GPL(__wake_up_locked);
4651 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4653 __wake_up_common(q, mode, 1, 0, key);
4655 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4658 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4660 * @mode: which threads
4661 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4662 * @key: opaque value to be passed to wakeup targets
4664 * The sync wakeup differs that the waker knows that it will schedule
4665 * away soon, so while the target thread will be woken up, it will not
4666 * be migrated to another CPU - ie. the two threads are 'synchronized'
4667 * with each other. This can prevent needless bouncing between CPUs.
4669 * On UP it can prevent extra preemption.
4671 * It may be assumed that this function implies a write memory barrier before
4672 * changing the task state if and only if any tasks are woken up.
4674 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4675 int nr_exclusive, void *key)
4677 unsigned long flags;
4678 int wake_flags = WF_SYNC;
4683 if (unlikely(!nr_exclusive))
4686 spin_lock_irqsave(&q->lock, flags);
4687 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4688 spin_unlock_irqrestore(&q->lock, flags);
4690 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4693 * __wake_up_sync - see __wake_up_sync_key()
4695 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4697 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4699 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4702 * complete: - signals a single thread waiting on this completion
4703 * @x: holds the state of this particular completion
4705 * This will wake up a single thread waiting on this completion. Threads will be
4706 * awakened in the same order in which they were queued.
4708 * See also complete_all(), wait_for_completion() and related routines.
4710 * It may be assumed that this function implies a write memory barrier before
4711 * changing the task state if and only if any tasks are woken up.
4713 void complete(struct completion *x)
4715 unsigned long flags;
4717 spin_lock_irqsave(&x->wait.lock, flags);
4719 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4720 spin_unlock_irqrestore(&x->wait.lock, flags);
4722 EXPORT_SYMBOL(complete);
4725 * complete_all: - signals all threads waiting on this completion
4726 * @x: holds the state of this particular completion
4728 * This will wake up all threads waiting on this particular completion event.
4730 * It may be assumed that this function implies a write memory barrier before
4731 * changing the task state if and only if any tasks are woken up.
4733 void complete_all(struct completion *x)
4735 unsigned long flags;
4737 spin_lock_irqsave(&x->wait.lock, flags);
4738 x->done += UINT_MAX/2;
4739 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4740 spin_unlock_irqrestore(&x->wait.lock, flags);
4742 EXPORT_SYMBOL(complete_all);
4744 static inline long __sched
4745 do_wait_for_common(struct completion *x, long timeout, int state)
4748 DECLARE_WAITQUEUE(wait, current);
4750 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4752 if (signal_pending_state(state, current)) {
4753 timeout = -ERESTARTSYS;
4756 __set_current_state(state);
4757 spin_unlock_irq(&x->wait.lock);
4758 timeout = schedule_timeout(timeout);
4759 spin_lock_irq(&x->wait.lock);
4760 } while (!x->done && timeout);
4761 __remove_wait_queue(&x->wait, &wait);
4766 return timeout ?: 1;
4770 wait_for_common(struct completion *x, long timeout, int state)
4774 spin_lock_irq(&x->wait.lock);
4775 timeout = do_wait_for_common(x, timeout, state);
4776 spin_unlock_irq(&x->wait.lock);
4781 * wait_for_completion: - waits for completion of a task
4782 * @x: holds the state of this particular completion
4784 * This waits to be signaled for completion of a specific task. It is NOT
4785 * interruptible and there is no timeout.
4787 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4788 * and interrupt capability. Also see complete().
4790 void __sched wait_for_completion(struct completion *x)
4792 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4794 EXPORT_SYMBOL(wait_for_completion);
4797 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4798 * @x: holds the state of this particular completion
4799 * @timeout: timeout value in jiffies
4801 * This waits for either a completion of a specific task to be signaled or for a
4802 * specified timeout to expire. The timeout is in jiffies. It is not
4805 unsigned long __sched
4806 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4808 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4810 EXPORT_SYMBOL(wait_for_completion_timeout);
4813 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4814 * @x: holds the state of this particular completion
4816 * This waits for completion of a specific task to be signaled. It is
4819 int __sched wait_for_completion_interruptible(struct completion *x)
4821 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4822 if (t == -ERESTARTSYS)
4826 EXPORT_SYMBOL(wait_for_completion_interruptible);
4829 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4830 * @x: holds the state of this particular completion
4831 * @timeout: timeout value in jiffies
4833 * This waits for either a completion of a specific task to be signaled or for a
4834 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4837 wait_for_completion_interruptible_timeout(struct completion *x,
4838 unsigned long timeout)
4840 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4842 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4845 * wait_for_completion_killable: - waits for completion of a task (killable)
4846 * @x: holds the state of this particular completion
4848 * This waits to be signaled for completion of a specific task. It can be
4849 * interrupted by a kill signal.
4851 int __sched wait_for_completion_killable(struct completion *x)
4853 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4854 if (t == -ERESTARTSYS)
4858 EXPORT_SYMBOL(wait_for_completion_killable);
4861 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4862 * @x: holds the state of this particular completion
4863 * @timeout: timeout value in jiffies
4865 * This waits for either a completion of a specific task to be
4866 * signaled or for a specified timeout to expire. It can be
4867 * interrupted by a kill signal. The timeout is in jiffies.
4870 wait_for_completion_killable_timeout(struct completion *x,
4871 unsigned long timeout)
4873 return wait_for_common(x, timeout, TASK_KILLABLE);
4875 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4878 * try_wait_for_completion - try to decrement a completion without blocking
4879 * @x: completion structure
4881 * Returns: 0 if a decrement cannot be done without blocking
4882 * 1 if a decrement succeeded.
4884 * If a completion is being used as a counting completion,
4885 * attempt to decrement the counter without blocking. This
4886 * enables us to avoid waiting if the resource the completion
4887 * is protecting is not available.
4889 bool try_wait_for_completion(struct completion *x)
4891 unsigned long flags;
4894 spin_lock_irqsave(&x->wait.lock, flags);
4899 spin_unlock_irqrestore(&x->wait.lock, flags);
4902 EXPORT_SYMBOL(try_wait_for_completion);
4905 * completion_done - Test to see if a completion has any waiters
4906 * @x: completion structure
4908 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4909 * 1 if there are no waiters.
4912 bool completion_done(struct completion *x)
4914 unsigned long flags;
4917 spin_lock_irqsave(&x->wait.lock, flags);
4920 spin_unlock_irqrestore(&x->wait.lock, flags);
4923 EXPORT_SYMBOL(completion_done);
4926 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4928 unsigned long flags;
4931 init_waitqueue_entry(&wait, current);
4933 __set_current_state(state);
4935 spin_lock_irqsave(&q->lock, flags);
4936 __add_wait_queue(q, &wait);
4937 spin_unlock(&q->lock);
4938 timeout = schedule_timeout(timeout);
4939 spin_lock_irq(&q->lock);
4940 __remove_wait_queue(q, &wait);
4941 spin_unlock_irqrestore(&q->lock, flags);
4946 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4948 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4950 EXPORT_SYMBOL(interruptible_sleep_on);
4953 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4955 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4957 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4959 void __sched sleep_on(wait_queue_head_t *q)
4961 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4963 EXPORT_SYMBOL(sleep_on);
4965 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4967 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4969 EXPORT_SYMBOL(sleep_on_timeout);
4971 #ifdef CONFIG_RT_MUTEXES
4974 * rt_mutex_setprio - set the current priority of a task
4976 * @prio: prio value (kernel-internal form)
4978 * This function changes the 'effective' priority of a task. It does
4979 * not touch ->normal_prio like __setscheduler().
4981 * Used by the rt_mutex code to implement priority inheritance logic.
4983 void rt_mutex_setprio(struct task_struct *p, int prio)
4985 int oldprio, on_rq, running;
4987 const struct sched_class *prev_class;
4989 BUG_ON(prio < 0 || prio > MAX_PRIO);
4991 rq = __task_rq_lock(p);
4993 trace_sched_pi_setprio(p, prio);
4995 prev_class = p->sched_class;
4997 running = task_current(rq, p);
4999 dequeue_task(rq, p, 0);
5001 p->sched_class->put_prev_task(rq, p);
5004 p->sched_class = &rt_sched_class;
5006 p->sched_class = &fair_sched_class;
5011 p->sched_class->set_curr_task(rq);
5013 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
5015 check_class_changed(rq, p, prev_class, oldprio);
5016 __task_rq_unlock(rq);
5021 void set_user_nice(struct task_struct *p, long nice)
5023 int old_prio, delta, on_rq;
5024 unsigned long flags;
5027 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5030 * We have to be careful, if called from sys_setpriority(),
5031 * the task might be in the middle of scheduling on another CPU.
5033 rq = task_rq_lock(p, &flags);
5035 * The RT priorities are set via sched_setscheduler(), but we still
5036 * allow the 'normal' nice value to be set - but as expected
5037 * it wont have any effect on scheduling until the task is
5038 * SCHED_FIFO/SCHED_RR:
5040 if (task_has_rt_policy(p)) {
5041 p->static_prio = NICE_TO_PRIO(nice);
5046 dequeue_task(rq, p, 0);
5048 p->static_prio = NICE_TO_PRIO(nice);
5051 p->prio = effective_prio(p);
5052 delta = p->prio - old_prio;
5055 enqueue_task(rq, p, 0);
5057 * If the task increased its priority or is running and
5058 * lowered its priority, then reschedule its CPU:
5060 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5061 resched_task(rq->curr);
5064 task_rq_unlock(rq, p, &flags);
5066 EXPORT_SYMBOL(set_user_nice);
5069 * can_nice - check if a task can reduce its nice value
5073 int can_nice(const struct task_struct *p, const int nice)
5075 /* convert nice value [19,-20] to rlimit style value [1,40] */
5076 int nice_rlim = 20 - nice;
5078 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
5079 capable(CAP_SYS_NICE));
5082 #ifdef __ARCH_WANT_SYS_NICE
5085 * sys_nice - change the priority of the current process.
5086 * @increment: priority increment
5088 * sys_setpriority is a more generic, but much slower function that
5089 * does similar things.
5091 SYSCALL_DEFINE1(nice, int, increment)
5096 * Setpriority might change our priority at the same moment.
5097 * We don't have to worry. Conceptually one call occurs first
5098 * and we have a single winner.
5100 if (increment < -40)
5105 nice = TASK_NICE(current) + increment;
5111 if (increment < 0 && !can_nice(current, nice))
5114 retval = security_task_setnice(current, nice);
5118 set_user_nice(current, nice);
5125 * task_prio - return the priority value of a given task.
5126 * @p: the task in question.
5128 * This is the priority value as seen by users in /proc.
5129 * RT tasks are offset by -200. Normal tasks are centered
5130 * around 0, value goes from -16 to +15.
5132 int task_prio(const struct task_struct *p)
5134 return p->prio - MAX_RT_PRIO;
5138 * task_nice - return the nice value of a given task.
5139 * @p: the task in question.
5141 int task_nice(const struct task_struct *p)
5143 return TASK_NICE(p);
5145 EXPORT_SYMBOL(task_nice);
5148 * idle_cpu - is a given cpu idle currently?
5149 * @cpu: the processor in question.
5151 int idle_cpu(int cpu)
5153 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5157 * idle_task - return the idle task for a given cpu.
5158 * @cpu: the processor in question.
5160 struct task_struct *idle_task(int cpu)
5162 return cpu_rq(cpu)->idle;
5166 * find_process_by_pid - find a process with a matching PID value.
5167 * @pid: the pid in question.
5169 static struct task_struct *find_process_by_pid(pid_t pid)
5171 return pid ? find_task_by_vpid(pid) : current;
5174 /* Actually do priority change: must hold rq lock. */
5176 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5179 p->rt_priority = prio;
5180 p->normal_prio = normal_prio(p);
5181 /* we are holding p->pi_lock already */
5182 p->prio = rt_mutex_getprio(p);
5183 if (rt_prio(p->prio))
5184 p->sched_class = &rt_sched_class;
5186 p->sched_class = &fair_sched_class;
5191 * check the target process has a UID that matches the current process's
5193 static bool check_same_owner(struct task_struct *p)
5195 const struct cred *cred = current_cred(), *pcred;
5199 pcred = __task_cred(p);
5200 if (cred->user->user_ns == pcred->user->user_ns)
5201 match = (cred->euid == pcred->euid ||
5202 cred->euid == pcred->uid);
5209 static int __sched_setscheduler(struct task_struct *p, int policy,
5210 const struct sched_param *param, bool user)
5212 int retval, oldprio, oldpolicy = -1, on_rq, running;
5213 unsigned long flags;
5214 const struct sched_class *prev_class;
5218 /* may grab non-irq protected spin_locks */
5219 BUG_ON(in_interrupt());
5221 /* double check policy once rq lock held */
5223 reset_on_fork = p->sched_reset_on_fork;
5224 policy = oldpolicy = p->policy;
5226 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
5227 policy &= ~SCHED_RESET_ON_FORK;
5229 if (policy != SCHED_FIFO && policy != SCHED_RR &&
5230 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5231 policy != SCHED_IDLE)
5236 * Valid priorities for SCHED_FIFO and SCHED_RR are
5237 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5238 * SCHED_BATCH and SCHED_IDLE is 0.
5240 if (param->sched_priority < 0 ||
5241 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5242 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5244 if (rt_policy(policy) != (param->sched_priority != 0))
5248 * Allow unprivileged RT tasks to decrease priority:
5250 if (user && !capable(CAP_SYS_NICE)) {
5251 if (rt_policy(policy)) {
5252 unsigned long rlim_rtprio =
5253 task_rlimit(p, RLIMIT_RTPRIO);
5255 /* can't set/change the rt policy */
5256 if (policy != p->policy && !rlim_rtprio)
5259 /* can't increase priority */
5260 if (param->sched_priority > p->rt_priority &&
5261 param->sched_priority > rlim_rtprio)
5266 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5267 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5269 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
5270 if (!can_nice(p, TASK_NICE(p)))
5274 /* can't change other user's priorities */
5275 if (!check_same_owner(p))
5278 /* Normal users shall not reset the sched_reset_on_fork flag */
5279 if (p->sched_reset_on_fork && !reset_on_fork)
5284 retval = security_task_setscheduler(p);
5290 * make sure no PI-waiters arrive (or leave) while we are
5291 * changing the priority of the task:
5293 * To be able to change p->policy safely, the appropriate
5294 * runqueue lock must be held.
5296 rq = task_rq_lock(p, &flags);
5299 * Changing the policy of the stop threads its a very bad idea
5301 if (p == rq->stop) {
5302 task_rq_unlock(rq, p, &flags);
5307 * If not changing anything there's no need to proceed further:
5309 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5310 param->sched_priority == p->rt_priority))) {
5312 __task_rq_unlock(rq);
5313 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5317 #ifdef CONFIG_RT_GROUP_SCHED
5320 * Do not allow realtime tasks into groups that have no runtime
5323 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5324 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5325 !task_group_is_autogroup(task_group(p))) {
5326 task_rq_unlock(rq, p, &flags);
5332 /* recheck policy now with rq lock held */
5333 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5334 policy = oldpolicy = -1;
5335 task_rq_unlock(rq, p, &flags);
5339 running = task_current(rq, p);
5341 deactivate_task(rq, p, 0);
5343 p->sched_class->put_prev_task(rq, p);
5345 p->sched_reset_on_fork = reset_on_fork;
5348 prev_class = p->sched_class;
5349 __setscheduler(rq, p, policy, param->sched_priority);
5352 p->sched_class->set_curr_task(rq);
5354 activate_task(rq, p, 0);
5356 check_class_changed(rq, p, prev_class, oldprio);
5357 task_rq_unlock(rq, p, &flags);
5359 rt_mutex_adjust_pi(p);
5365 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5366 * @p: the task in question.
5367 * @policy: new policy.
5368 * @param: structure containing the new RT priority.
5370 * NOTE that the task may be already dead.
5372 int sched_setscheduler(struct task_struct *p, int policy,
5373 const struct sched_param *param)
5375 return __sched_setscheduler(p, policy, param, true);
5377 EXPORT_SYMBOL_GPL(sched_setscheduler);
5380 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5381 * @p: the task in question.
5382 * @policy: new policy.
5383 * @param: structure containing the new RT priority.
5385 * Just like sched_setscheduler, only don't bother checking if the
5386 * current context has permission. For example, this is needed in
5387 * stop_machine(): we create temporary high priority worker threads,
5388 * but our caller might not have that capability.
5390 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5391 const struct sched_param *param)
5393 return __sched_setscheduler(p, policy, param, false);
5397 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5399 struct sched_param lparam;
5400 struct task_struct *p;
5403 if (!param || pid < 0)
5405 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5410 p = find_process_by_pid(pid);
5412 retval = sched_setscheduler(p, policy, &lparam);
5419 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5420 * @pid: the pid in question.
5421 * @policy: new policy.
5422 * @param: structure containing the new RT priority.
5424 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5425 struct sched_param __user *, param)
5427 /* negative values for policy are not valid */
5431 return do_sched_setscheduler(pid, policy, param);
5435 * sys_sched_setparam - set/change the RT priority of a thread
5436 * @pid: the pid in question.
5437 * @param: structure containing the new RT priority.
5439 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5441 return do_sched_setscheduler(pid, -1, param);
5445 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5446 * @pid: the pid in question.
5448 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5450 struct task_struct *p;
5458 p = find_process_by_pid(pid);
5460 retval = security_task_getscheduler(p);
5463 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5470 * sys_sched_getparam - get the RT priority of a thread
5471 * @pid: the pid in question.
5472 * @param: structure containing the RT priority.
5474 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5476 struct sched_param lp;
5477 struct task_struct *p;
5480 if (!param || pid < 0)
5484 p = find_process_by_pid(pid);
5489 retval = security_task_getscheduler(p);
5493 lp.sched_priority = p->rt_priority;
5497 * This one might sleep, we cannot do it with a spinlock held ...
5499 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5508 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5510 cpumask_var_t cpus_allowed, new_mask;
5511 struct task_struct *p;
5517 p = find_process_by_pid(pid);
5524 /* Prevent p going away */
5528 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5532 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5534 goto out_free_cpus_allowed;
5537 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5540 retval = security_task_setscheduler(p);
5544 cpuset_cpus_allowed(p, cpus_allowed);
5545 cpumask_and(new_mask, in_mask, cpus_allowed);
5547 retval = set_cpus_allowed_ptr(p, new_mask);
5550 cpuset_cpus_allowed(p, cpus_allowed);
5551 if (!cpumask_subset(new_mask, cpus_allowed)) {
5553 * We must have raced with a concurrent cpuset
5554 * update. Just reset the cpus_allowed to the
5555 * cpuset's cpus_allowed
5557 cpumask_copy(new_mask, cpus_allowed);
5562 free_cpumask_var(new_mask);
5563 out_free_cpus_allowed:
5564 free_cpumask_var(cpus_allowed);
5571 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5572 struct cpumask *new_mask)
5574 if (len < cpumask_size())
5575 cpumask_clear(new_mask);
5576 else if (len > cpumask_size())
5577 len = cpumask_size();
5579 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5583 * sys_sched_setaffinity - set the cpu affinity of a process
5584 * @pid: pid of the process
5585 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5586 * @user_mask_ptr: user-space pointer to the new cpu mask
5588 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5589 unsigned long __user *, user_mask_ptr)
5591 cpumask_var_t new_mask;
5594 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5597 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5599 retval = sched_setaffinity(pid, new_mask);
5600 free_cpumask_var(new_mask);
5604 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5606 struct task_struct *p;
5607 unsigned long flags;
5614 p = find_process_by_pid(pid);
5618 retval = security_task_getscheduler(p);
5622 raw_spin_lock_irqsave(&p->pi_lock, flags);
5623 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5624 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5634 * sys_sched_getaffinity - get the cpu affinity of a process
5635 * @pid: pid of the process
5636 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5637 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5639 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5640 unsigned long __user *, user_mask_ptr)
5645 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5647 if (len & (sizeof(unsigned long)-1))
5650 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5653 ret = sched_getaffinity(pid, mask);
5655 size_t retlen = min_t(size_t, len, cpumask_size());
5657 if (copy_to_user(user_mask_ptr, mask, retlen))
5662 free_cpumask_var(mask);
5668 * sys_sched_yield - yield the current processor to other threads.
5670 * This function yields the current CPU to other tasks. If there are no
5671 * other threads running on this CPU then this function will return.
5673 SYSCALL_DEFINE0(sched_yield)
5675 struct rq *rq = this_rq_lock();
5677 schedstat_inc(rq, yld_count);
5678 current->sched_class->yield_task(rq);
5681 * Since we are going to call schedule() anyway, there's
5682 * no need to preempt or enable interrupts:
5684 __release(rq->lock);
5685 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5686 do_raw_spin_unlock(&rq->lock);
5687 preempt_enable_no_resched();
5694 static inline int should_resched(void)
5696 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5699 static void __cond_resched(void)
5701 add_preempt_count(PREEMPT_ACTIVE);
5703 sub_preempt_count(PREEMPT_ACTIVE);
5706 int __sched _cond_resched(void)
5708 if (should_resched()) {
5714 EXPORT_SYMBOL(_cond_resched);
5717 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5718 * call schedule, and on return reacquire the lock.
5720 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5721 * operations here to prevent schedule() from being called twice (once via
5722 * spin_unlock(), once by hand).
5724 int __cond_resched_lock(spinlock_t *lock)
5726 int resched = should_resched();
5729 lockdep_assert_held(lock);
5731 if (spin_needbreak(lock) || resched) {
5742 EXPORT_SYMBOL(__cond_resched_lock);
5744 int __sched __cond_resched_softirq(void)
5746 BUG_ON(!in_softirq());
5748 if (should_resched()) {
5756 EXPORT_SYMBOL(__cond_resched_softirq);
5759 * yield - yield the current processor to other threads.
5761 * This is a shortcut for kernel-space yielding - it marks the
5762 * thread runnable and calls sys_sched_yield().
5764 void __sched yield(void)
5766 set_current_state(TASK_RUNNING);
5769 EXPORT_SYMBOL(yield);
5772 * yield_to - yield the current processor to another thread in
5773 * your thread group, or accelerate that thread toward the
5774 * processor it's on.
5776 * @preempt: whether task preemption is allowed or not
5778 * It's the caller's job to ensure that the target task struct
5779 * can't go away on us before we can do any checks.
5781 * Returns true if we indeed boosted the target task.
5783 bool __sched yield_to(struct task_struct *p, bool preempt)
5785 struct task_struct *curr = current;
5786 struct rq *rq, *p_rq;
5787 unsigned long flags;
5790 local_irq_save(flags);
5795 double_rq_lock(rq, p_rq);
5796 while (task_rq(p) != p_rq) {
5797 double_rq_unlock(rq, p_rq);
5801 if (!curr->sched_class->yield_to_task)
5804 if (curr->sched_class != p->sched_class)
5807 if (task_running(p_rq, p) || p->state)
5810 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5812 schedstat_inc(rq, yld_count);
5814 * Make p's CPU reschedule; pick_next_entity takes care of
5817 if (preempt && rq != p_rq)
5818 resched_task(p_rq->curr);
5822 double_rq_unlock(rq, p_rq);
5823 local_irq_restore(flags);
5830 EXPORT_SYMBOL_GPL(yield_to);
5833 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5834 * that process accounting knows that this is a task in IO wait state.
5836 void __sched io_schedule(void)
5838 struct rq *rq = raw_rq();
5840 delayacct_blkio_start();
5841 atomic_inc(&rq->nr_iowait);
5842 blk_flush_plug(current);
5843 current->in_iowait = 1;
5845 current->in_iowait = 0;
5846 atomic_dec(&rq->nr_iowait);
5847 delayacct_blkio_end();
5849 EXPORT_SYMBOL(io_schedule);
5851 long __sched io_schedule_timeout(long timeout)
5853 struct rq *rq = raw_rq();
5856 delayacct_blkio_start();
5857 atomic_inc(&rq->nr_iowait);
5858 blk_flush_plug(current);
5859 current->in_iowait = 1;
5860 ret = schedule_timeout(timeout);
5861 current->in_iowait = 0;
5862 atomic_dec(&rq->nr_iowait);
5863 delayacct_blkio_end();
5868 * sys_sched_get_priority_max - return maximum RT priority.
5869 * @policy: scheduling class.
5871 * this syscall returns the maximum rt_priority that can be used
5872 * by a given scheduling class.
5874 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5881 ret = MAX_USER_RT_PRIO-1;
5893 * sys_sched_get_priority_min - return minimum RT priority.
5894 * @policy: scheduling class.
5896 * this syscall returns the minimum rt_priority that can be used
5897 * by a given scheduling class.
5899 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5917 * sys_sched_rr_get_interval - return the default timeslice of a process.
5918 * @pid: pid of the process.
5919 * @interval: userspace pointer to the timeslice value.
5921 * this syscall writes the default timeslice value of a given process
5922 * into the user-space timespec buffer. A value of '0' means infinity.
5924 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5925 struct timespec __user *, interval)
5927 struct task_struct *p;
5928 unsigned int time_slice;
5929 unsigned long flags;
5939 p = find_process_by_pid(pid);
5943 retval = security_task_getscheduler(p);
5947 rq = task_rq_lock(p, &flags);
5948 time_slice = p->sched_class->get_rr_interval(rq, p);
5949 task_rq_unlock(rq, p, &flags);
5952 jiffies_to_timespec(time_slice, &t);
5953 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5961 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5963 void sched_show_task(struct task_struct *p)
5965 unsigned long free = 0;
5968 state = p->state ? __ffs(p->state) + 1 : 0;
5969 printk(KERN_INFO "%-15.15s %c", p->comm,
5970 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5971 #if BITS_PER_LONG == 32
5972 if (state == TASK_RUNNING)
5973 printk(KERN_CONT " running ");
5975 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5977 if (state == TASK_RUNNING)
5978 printk(KERN_CONT " running task ");
5980 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5982 #ifdef CONFIG_DEBUG_STACK_USAGE
5983 free = stack_not_used(p);
5985 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5986 task_pid_nr(p), task_pid_nr(p->real_parent),
5987 (unsigned long)task_thread_info(p)->flags);
5989 show_stack(p, NULL);
5992 void show_state_filter(unsigned long state_filter)
5994 struct task_struct *g, *p;
5996 #if BITS_PER_LONG == 32
5998 " task PC stack pid father\n");
6001 " task PC stack pid father\n");
6003 read_lock(&tasklist_lock);
6004 do_each_thread(g, p) {
6006 * reset the NMI-timeout, listing all files on a slow
6007 * console might take a lot of time:
6009 touch_nmi_watchdog();
6010 if (!state_filter || (p->state & state_filter))
6012 } while_each_thread(g, p);
6014 touch_all_softlockup_watchdogs();
6016 #ifdef CONFIG_SCHED_DEBUG
6017 sysrq_sched_debug_show();
6019 read_unlock(&tasklist_lock);
6021 * Only show locks if all tasks are dumped:
6024 debug_show_all_locks();
6027 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6029 idle->sched_class = &idle_sched_class;
6033 * init_idle - set up an idle thread for a given CPU
6034 * @idle: task in question
6035 * @cpu: cpu the idle task belongs to
6037 * NOTE: this function does not set the idle thread's NEED_RESCHED
6038 * flag, to make booting more robust.
6040 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6042 struct rq *rq = cpu_rq(cpu);
6043 unsigned long flags;
6045 raw_spin_lock_irqsave(&rq->lock, flags);
6048 idle->state = TASK_RUNNING;
6049 idle->se.exec_start = sched_clock();
6051 do_set_cpus_allowed(idle, cpumask_of(cpu));
6053 * We're having a chicken and egg problem, even though we are
6054 * holding rq->lock, the cpu isn't yet set to this cpu so the
6055 * lockdep check in task_group() will fail.
6057 * Similar case to sched_fork(). / Alternatively we could
6058 * use task_rq_lock() here and obtain the other rq->lock.
6063 __set_task_cpu(idle, cpu);
6066 rq->curr = rq->idle = idle;
6067 #if defined(CONFIG_SMP)
6070 raw_spin_unlock_irqrestore(&rq->lock, flags);
6072 /* Set the preempt count _outside_ the spinlocks! */
6073 task_thread_info(idle)->preempt_count = 0;
6076 * The idle tasks have their own, simple scheduling class:
6078 idle->sched_class = &idle_sched_class;
6079 ftrace_graph_init_idle_task(idle, cpu);
6083 * In a system that switches off the HZ timer nohz_cpu_mask
6084 * indicates which cpus entered this state. This is used
6085 * in the rcu update to wait only for active cpus. For system
6086 * which do not switch off the HZ timer nohz_cpu_mask should
6087 * always be CPU_BITS_NONE.
6089 cpumask_var_t nohz_cpu_mask;
6092 * Increase the granularity value when there are more CPUs,
6093 * because with more CPUs the 'effective latency' as visible
6094 * to users decreases. But the relationship is not linear,
6095 * so pick a second-best guess by going with the log2 of the
6098 * This idea comes from the SD scheduler of Con Kolivas:
6100 static int get_update_sysctl_factor(void)
6102 unsigned int cpus = min_t(int, num_online_cpus(), 8);
6103 unsigned int factor;
6105 switch (sysctl_sched_tunable_scaling) {
6106 case SCHED_TUNABLESCALING_NONE:
6109 case SCHED_TUNABLESCALING_LINEAR:
6112 case SCHED_TUNABLESCALING_LOG:
6114 factor = 1 + ilog2(cpus);
6121 static void update_sysctl(void)
6123 unsigned int factor = get_update_sysctl_factor();
6125 #define SET_SYSCTL(name) \
6126 (sysctl_##name = (factor) * normalized_sysctl_##name)
6127 SET_SYSCTL(sched_min_granularity);
6128 SET_SYSCTL(sched_latency);
6129 SET_SYSCTL(sched_wakeup_granularity);
6133 static inline void sched_init_granularity(void)
6139 void do_set_cpus_allowed(struct task_struct *p, const struct cpumask *new_mask)
6141 if (p->sched_class && p->sched_class->set_cpus_allowed)
6142 p->sched_class->set_cpus_allowed(p, new_mask);
6144 cpumask_copy(&p->cpus_allowed, new_mask);
6145 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6150 * This is how migration works:
6152 * 1) we invoke migration_cpu_stop() on the target CPU using
6154 * 2) stopper starts to run (implicitly forcing the migrated thread
6156 * 3) it checks whether the migrated task is still in the wrong runqueue.
6157 * 4) if it's in the wrong runqueue then the migration thread removes
6158 * it and puts it into the right queue.
6159 * 5) stopper completes and stop_one_cpu() returns and the migration
6164 * Change a given task's CPU affinity. Migrate the thread to a
6165 * proper CPU and schedule it away if the CPU it's executing on
6166 * is removed from the allowed bitmask.
6168 * NOTE: the caller must have a valid reference to the task, the
6169 * task must not exit() & deallocate itself prematurely. The
6170 * call is not atomic; no spinlocks may be held.
6172 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6174 unsigned long flags;
6176 unsigned int dest_cpu;
6179 rq = task_rq_lock(p, &flags);
6181 if (cpumask_equal(&p->cpus_allowed, new_mask))
6184 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
6189 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current)) {
6194 do_set_cpus_allowed(p, new_mask);
6196 /* Can the task run on the task's current CPU? If so, we're done */
6197 if (cpumask_test_cpu(task_cpu(p), new_mask))
6200 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
6202 struct migration_arg arg = { p, dest_cpu };
6203 /* Need help from migration thread: drop lock and wait. */
6204 task_rq_unlock(rq, p, &flags);
6205 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
6206 tlb_migrate_finish(p->mm);
6210 task_rq_unlock(rq, p, &flags);
6214 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6217 * Move (not current) task off this cpu, onto dest cpu. We're doing
6218 * this because either it can't run here any more (set_cpus_allowed()
6219 * away from this CPU, or CPU going down), or because we're
6220 * attempting to rebalance this task on exec (sched_exec).
6222 * So we race with normal scheduler movements, but that's OK, as long
6223 * as the task is no longer on this CPU.
6225 * Returns non-zero if task was successfully migrated.
6227 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6229 struct rq *rq_dest, *rq_src;
6232 if (unlikely(!cpu_active(dest_cpu)))
6235 rq_src = cpu_rq(src_cpu);
6236 rq_dest = cpu_rq(dest_cpu);
6238 raw_spin_lock(&p->pi_lock);
6239 double_rq_lock(rq_src, rq_dest);
6240 /* Already moved. */
6241 if (task_cpu(p) != src_cpu)
6243 /* Affinity changed (again). */
6244 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6248 * If we're not on a rq, the next wake-up will ensure we're
6252 deactivate_task(rq_src, p, 0);
6253 set_task_cpu(p, dest_cpu);
6254 activate_task(rq_dest, p, 0);
6255 check_preempt_curr(rq_dest, p, 0);
6260 double_rq_unlock(rq_src, rq_dest);
6261 raw_spin_unlock(&p->pi_lock);
6266 * migration_cpu_stop - this will be executed by a highprio stopper thread
6267 * and performs thread migration by bumping thread off CPU then
6268 * 'pushing' onto another runqueue.
6270 static int migration_cpu_stop(void *data)
6272 struct migration_arg *arg = data;
6275 * The original target cpu might have gone down and we might
6276 * be on another cpu but it doesn't matter.
6278 local_irq_disable();
6279 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
6284 #ifdef CONFIG_HOTPLUG_CPU
6287 * Ensures that the idle task is using init_mm right before its cpu goes
6290 void idle_task_exit(void)
6292 struct mm_struct *mm = current->active_mm;
6294 BUG_ON(cpu_online(smp_processor_id()));
6297 switch_mm(mm, &init_mm, current);
6302 * While a dead CPU has no uninterruptible tasks queued at this point,
6303 * it might still have a nonzero ->nr_uninterruptible counter, because
6304 * for performance reasons the counter is not stricly tracking tasks to
6305 * their home CPUs. So we just add the counter to another CPU's counter,
6306 * to keep the global sum constant after CPU-down:
6308 static void migrate_nr_uninterruptible(struct rq *rq_src)
6310 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6312 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6313 rq_src->nr_uninterruptible = 0;
6317 * remove the tasks which were accounted by rq from calc_load_tasks.
6319 static void calc_global_load_remove(struct rq *rq)
6321 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6322 rq->calc_load_active = 0;
6326 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6327 * try_to_wake_up()->select_task_rq().
6329 * Called with rq->lock held even though we'er in stop_machine() and
6330 * there's no concurrency possible, we hold the required locks anyway
6331 * because of lock validation efforts.
6333 static void migrate_tasks(unsigned int dead_cpu)
6335 struct rq *rq = cpu_rq(dead_cpu);
6336 struct task_struct *next, *stop = rq->stop;
6340 * Fudge the rq selection such that the below task selection loop
6341 * doesn't get stuck on the currently eligible stop task.
6343 * We're currently inside stop_machine() and the rq is either stuck
6344 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6345 * either way we should never end up calling schedule() until we're
6352 * There's this thread running, bail when that's the only
6355 if (rq->nr_running == 1)
6358 next = pick_next_task(rq);
6360 next->sched_class->put_prev_task(rq, next);
6362 /* Find suitable destination for @next, with force if needed. */
6363 dest_cpu = select_fallback_rq(dead_cpu, next);
6364 raw_spin_unlock(&rq->lock);
6366 __migrate_task(next, dead_cpu, dest_cpu);
6368 raw_spin_lock(&rq->lock);
6374 #endif /* CONFIG_HOTPLUG_CPU */
6376 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6378 static struct ctl_table sd_ctl_dir[] = {
6380 .procname = "sched_domain",
6386 static struct ctl_table sd_ctl_root[] = {
6388 .procname = "kernel",
6390 .child = sd_ctl_dir,
6395 static struct ctl_table *sd_alloc_ctl_entry(int n)
6397 struct ctl_table *entry =
6398 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6403 static void sd_free_ctl_entry(struct ctl_table **tablep)
6405 struct ctl_table *entry;
6408 * In the intermediate directories, both the child directory and
6409 * procname are dynamically allocated and could fail but the mode
6410 * will always be set. In the lowest directory the names are
6411 * static strings and all have proc handlers.
6413 for (entry = *tablep; entry->mode; entry++) {
6415 sd_free_ctl_entry(&entry->child);
6416 if (entry->proc_handler == NULL)
6417 kfree(entry->procname);
6425 set_table_entry(struct ctl_table *entry,
6426 const char *procname, void *data, int maxlen,
6427 mode_t mode, proc_handler *proc_handler)
6429 entry->procname = procname;
6431 entry->maxlen = maxlen;
6433 entry->proc_handler = proc_handler;
6436 static struct ctl_table *
6437 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6439 struct ctl_table *table = sd_alloc_ctl_entry(13);
6444 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6445 sizeof(long), 0644, proc_doulongvec_minmax);
6446 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6447 sizeof(long), 0644, proc_doulongvec_minmax);
6448 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6449 sizeof(int), 0644, proc_dointvec_minmax);
6450 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6451 sizeof(int), 0644, proc_dointvec_minmax);
6452 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6453 sizeof(int), 0644, proc_dointvec_minmax);
6454 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6455 sizeof(int), 0644, proc_dointvec_minmax);
6456 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6457 sizeof(int), 0644, proc_dointvec_minmax);
6458 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6459 sizeof(int), 0644, proc_dointvec_minmax);
6460 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6461 sizeof(int), 0644, proc_dointvec_minmax);
6462 set_table_entry(&table[9], "cache_nice_tries",
6463 &sd->cache_nice_tries,
6464 sizeof(int), 0644, proc_dointvec_minmax);
6465 set_table_entry(&table[10], "flags", &sd->flags,
6466 sizeof(int), 0644, proc_dointvec_minmax);
6467 set_table_entry(&table[11], "name", sd->name,
6468 CORENAME_MAX_SIZE, 0444, proc_dostring);
6469 /* &table[12] is terminator */
6474 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6476 struct ctl_table *entry, *table;
6477 struct sched_domain *sd;
6478 int domain_num = 0, i;
6481 for_each_domain(cpu, sd)
6483 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6488 for_each_domain(cpu, sd) {
6489 snprintf(buf, 32, "domain%d", i);
6490 entry->procname = kstrdup(buf, GFP_KERNEL);
6492 entry->child = sd_alloc_ctl_domain_table(sd);
6499 static struct ctl_table_header *sd_sysctl_header;
6500 static void register_sched_domain_sysctl(void)
6502 int i, cpu_num = num_possible_cpus();
6503 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6506 WARN_ON(sd_ctl_dir[0].child);
6507 sd_ctl_dir[0].child = entry;
6512 for_each_possible_cpu(i) {
6513 snprintf(buf, 32, "cpu%d", i);
6514 entry->procname = kstrdup(buf, GFP_KERNEL);
6516 entry->child = sd_alloc_ctl_cpu_table(i);
6520 WARN_ON(sd_sysctl_header);
6521 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6524 /* may be called multiple times per register */
6525 static void unregister_sched_domain_sysctl(void)
6527 if (sd_sysctl_header)
6528 unregister_sysctl_table(sd_sysctl_header);
6529 sd_sysctl_header = NULL;
6530 if (sd_ctl_dir[0].child)
6531 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6534 static void register_sched_domain_sysctl(void)
6537 static void unregister_sched_domain_sysctl(void)
6542 static void set_rq_online(struct rq *rq)
6545 const struct sched_class *class;
6547 cpumask_set_cpu(rq->cpu, rq->rd->online);
6550 for_each_class(class) {
6551 if (class->rq_online)
6552 class->rq_online(rq);
6557 static void set_rq_offline(struct rq *rq)
6560 const struct sched_class *class;
6562 for_each_class(class) {
6563 if (class->rq_offline)
6564 class->rq_offline(rq);
6567 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6573 * migration_call - callback that gets triggered when a CPU is added.
6574 * Here we can start up the necessary migration thread for the new CPU.
6576 static int __cpuinit
6577 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6579 int cpu = (long)hcpu;
6580 unsigned long flags;
6581 struct rq *rq = cpu_rq(cpu);
6583 switch (action & ~CPU_TASKS_FROZEN) {
6585 case CPU_UP_PREPARE:
6586 rq->calc_load_update = calc_load_update;
6590 /* Update our root-domain */
6591 raw_spin_lock_irqsave(&rq->lock, flags);
6593 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6597 raw_spin_unlock_irqrestore(&rq->lock, flags);
6600 #ifdef CONFIG_HOTPLUG_CPU
6602 sched_ttwu_pending();
6603 /* Update our root-domain */
6604 raw_spin_lock_irqsave(&rq->lock, flags);
6606 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6610 BUG_ON(rq->nr_running != 1); /* the migration thread */
6611 raw_spin_unlock_irqrestore(&rq->lock, flags);
6613 migrate_nr_uninterruptible(rq);
6614 calc_global_load_remove(rq);
6619 update_max_interval();
6625 * Register at high priority so that task migration (migrate_all_tasks)
6626 * happens before everything else. This has to be lower priority than
6627 * the notifier in the perf_event subsystem, though.
6629 static struct notifier_block __cpuinitdata migration_notifier = {
6630 .notifier_call = migration_call,
6631 .priority = CPU_PRI_MIGRATION,
6634 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6635 unsigned long action, void *hcpu)
6637 switch (action & ~CPU_TASKS_FROZEN) {
6639 case CPU_DOWN_FAILED:
6640 set_cpu_active((long)hcpu, true);
6647 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6648 unsigned long action, void *hcpu)
6650 switch (action & ~CPU_TASKS_FROZEN) {
6651 case CPU_DOWN_PREPARE:
6652 set_cpu_active((long)hcpu, false);
6659 static int __init migration_init(void)
6661 void *cpu = (void *)(long)smp_processor_id();
6664 /* Initialize migration for the boot CPU */
6665 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6666 BUG_ON(err == NOTIFY_BAD);
6667 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6668 register_cpu_notifier(&migration_notifier);
6670 /* Register cpu active notifiers */
6671 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6672 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6676 early_initcall(migration_init);
6681 static cpumask_var_t sched_domains_tmpmask; /* sched_domains_mutex */
6683 #ifdef CONFIG_SCHED_DEBUG
6685 static __read_mostly int sched_domain_debug_enabled;
6687 static int __init sched_domain_debug_setup(char *str)
6689 sched_domain_debug_enabled = 1;
6693 early_param("sched_debug", sched_domain_debug_setup);
6695 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6696 struct cpumask *groupmask)
6698 struct sched_group *group = sd->groups;
6701 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6702 cpumask_clear(groupmask);
6704 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6706 if (!(sd->flags & SD_LOAD_BALANCE)) {
6707 printk("does not load-balance\n");
6709 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6714 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6716 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6717 printk(KERN_ERR "ERROR: domain->span does not contain "
6720 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6721 printk(KERN_ERR "ERROR: domain->groups does not contain"
6725 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6729 printk(KERN_ERR "ERROR: group is NULL\n");
6733 if (!group->sgp->power) {
6734 printk(KERN_CONT "\n");
6735 printk(KERN_ERR "ERROR: domain->cpu_power not "
6740 if (!cpumask_weight(sched_group_cpus(group))) {
6741 printk(KERN_CONT "\n");
6742 printk(KERN_ERR "ERROR: empty group\n");
6746 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6747 printk(KERN_CONT "\n");
6748 printk(KERN_ERR "ERROR: repeated CPUs\n");
6752 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6754 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6756 printk(KERN_CONT " %s", str);
6757 if (group->sgp->power != SCHED_POWER_SCALE) {
6758 printk(KERN_CONT " (cpu_power = %d)",
6762 group = group->next;
6763 } while (group != sd->groups);
6764 printk(KERN_CONT "\n");
6766 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6767 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6770 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6771 printk(KERN_ERR "ERROR: parent span is not a superset "
6772 "of domain->span\n");
6776 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6780 if (!sched_domain_debug_enabled)
6784 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6788 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6791 if (sched_domain_debug_one(sd, cpu, level, sched_domains_tmpmask))
6799 #else /* !CONFIG_SCHED_DEBUG */
6800 # define sched_domain_debug(sd, cpu) do { } while (0)
6801 #endif /* CONFIG_SCHED_DEBUG */
6803 static int sd_degenerate(struct sched_domain *sd)
6805 if (cpumask_weight(sched_domain_span(sd)) == 1)
6808 /* Following flags need at least 2 groups */
6809 if (sd->flags & (SD_LOAD_BALANCE |
6810 SD_BALANCE_NEWIDLE |
6814 SD_SHARE_PKG_RESOURCES)) {
6815 if (sd->groups != sd->groups->next)
6819 /* Following flags don't use groups */
6820 if (sd->flags & (SD_WAKE_AFFINE))
6827 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6829 unsigned long cflags = sd->flags, pflags = parent->flags;
6831 if (sd_degenerate(parent))
6834 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6837 /* Flags needing groups don't count if only 1 group in parent */
6838 if (parent->groups == parent->groups->next) {
6839 pflags &= ~(SD_LOAD_BALANCE |
6840 SD_BALANCE_NEWIDLE |
6844 SD_SHARE_PKG_RESOURCES);
6845 if (nr_node_ids == 1)
6846 pflags &= ~SD_SERIALIZE;
6848 if (~cflags & pflags)
6854 static void free_rootdomain(struct rcu_head *rcu)
6856 struct root_domain *rd = container_of(rcu, struct root_domain, rcu);
6858 cpupri_cleanup(&rd->cpupri);
6859 free_cpumask_var(rd->rto_mask);
6860 free_cpumask_var(rd->online);
6861 free_cpumask_var(rd->span);
6865 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6867 struct root_domain *old_rd = NULL;
6868 unsigned long flags;
6870 raw_spin_lock_irqsave(&rq->lock, flags);
6875 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6878 cpumask_clear_cpu(rq->cpu, old_rd->span);
6881 * If we dont want to free the old_rt yet then
6882 * set old_rd to NULL to skip the freeing later
6885 if (!atomic_dec_and_test(&old_rd->refcount))
6889 atomic_inc(&rd->refcount);
6892 cpumask_set_cpu(rq->cpu, rd->span);
6893 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6896 raw_spin_unlock_irqrestore(&rq->lock, flags);
6899 call_rcu_sched(&old_rd->rcu, free_rootdomain);
6902 static int init_rootdomain(struct root_domain *rd)
6904 memset(rd, 0, sizeof(*rd));
6906 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6908 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6910 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6913 if (cpupri_init(&rd->cpupri) != 0)
6918 free_cpumask_var(rd->rto_mask);
6920 free_cpumask_var(rd->online);
6922 free_cpumask_var(rd->span);
6927 static void init_defrootdomain(void)
6929 init_rootdomain(&def_root_domain);
6931 atomic_set(&def_root_domain.refcount, 1);
6934 static struct root_domain *alloc_rootdomain(void)
6936 struct root_domain *rd;
6938 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6942 if (init_rootdomain(rd) != 0) {
6950 static void free_sched_groups(struct sched_group *sg, int free_sgp)
6952 struct sched_group *tmp, *first;
6961 if (free_sgp && atomic_dec_and_test(&sg->sgp->ref))
6966 } while (sg != first);
6969 static void free_sched_domain(struct rcu_head *rcu)
6971 struct sched_domain *sd = container_of(rcu, struct sched_domain, rcu);
6974 * If its an overlapping domain it has private groups, iterate and
6977 if (sd->flags & SD_OVERLAP) {
6978 free_sched_groups(sd->groups, 1);
6979 } else if (atomic_dec_and_test(&sd->groups->ref)) {
6980 kfree(sd->groups->sgp);
6986 static void destroy_sched_domain(struct sched_domain *sd, int cpu)
6988 call_rcu(&sd->rcu, free_sched_domain);
6991 static void destroy_sched_domains(struct sched_domain *sd, int cpu)
6993 for (; sd; sd = sd->parent)
6994 destroy_sched_domain(sd, cpu);
6998 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6999 * hold the hotplug lock.
7002 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7004 struct rq *rq = cpu_rq(cpu);
7005 struct sched_domain *tmp;
7007 /* Remove the sched domains which do not contribute to scheduling. */
7008 for (tmp = sd; tmp; ) {
7009 struct sched_domain *parent = tmp->parent;
7013 if (sd_parent_degenerate(tmp, parent)) {
7014 tmp->parent = parent->parent;
7016 parent->parent->child = tmp;
7017 destroy_sched_domain(parent, cpu);
7022 if (sd && sd_degenerate(sd)) {
7025 destroy_sched_domain(tmp, cpu);
7030 sched_domain_debug(sd, cpu);
7032 rq_attach_root(rq, rd);
7034 rcu_assign_pointer(rq->sd, sd);
7035 destroy_sched_domains(tmp, cpu);
7038 /* cpus with isolated domains */
7039 static cpumask_var_t cpu_isolated_map;
7041 /* Setup the mask of cpus configured for isolated domains */
7042 static int __init isolated_cpu_setup(char *str)
7044 alloc_bootmem_cpumask_var(&cpu_isolated_map);
7045 cpulist_parse(str, cpu_isolated_map);
7049 __setup("isolcpus=", isolated_cpu_setup);
7051 #define SD_NODES_PER_DOMAIN 16
7056 * find_next_best_node - find the next node to include in a sched_domain
7057 * @node: node whose sched_domain we're building
7058 * @used_nodes: nodes already in the sched_domain
7060 * Find the next node to include in a given scheduling domain. Simply
7061 * finds the closest node not already in the @used_nodes map.
7063 * Should use nodemask_t.
7065 static int find_next_best_node(int node, nodemask_t *used_nodes)
7067 int i, n, val, min_val, best_node = -1;
7071 for (i = 0; i < nr_node_ids; i++) {
7072 /* Start at @node */
7073 n = (node + i) % nr_node_ids;
7075 if (!nr_cpus_node(n))
7078 /* Skip already used nodes */
7079 if (node_isset(n, *used_nodes))
7082 /* Simple min distance search */
7083 val = node_distance(node, n);
7085 if (val < min_val) {
7091 if (best_node != -1)
7092 node_set(best_node, *used_nodes);
7097 * sched_domain_node_span - get a cpumask for a node's sched_domain
7098 * @node: node whose cpumask we're constructing
7099 * @span: resulting cpumask
7101 * Given a node, construct a good cpumask for its sched_domain to span. It
7102 * should be one that prevents unnecessary balancing, but also spreads tasks
7105 static void sched_domain_node_span(int node, struct cpumask *span)
7107 nodemask_t used_nodes;
7110 cpumask_clear(span);
7111 nodes_clear(used_nodes);
7113 cpumask_or(span, span, cpumask_of_node(node));
7114 node_set(node, used_nodes);
7116 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7117 int next_node = find_next_best_node(node, &used_nodes);
7120 cpumask_or(span, span, cpumask_of_node(next_node));
7124 static const struct cpumask *cpu_node_mask(int cpu)
7126 lockdep_assert_held(&sched_domains_mutex);
7128 sched_domain_node_span(cpu_to_node(cpu), sched_domains_tmpmask);
7130 return sched_domains_tmpmask;
7133 static const struct cpumask *cpu_allnodes_mask(int cpu)
7135 return cpu_possible_mask;
7137 #endif /* CONFIG_NUMA */
7139 static const struct cpumask *cpu_cpu_mask(int cpu)
7141 return cpumask_of_node(cpu_to_node(cpu));
7144 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7147 struct sched_domain **__percpu sd;
7148 struct sched_group **__percpu sg;
7149 struct sched_group_power **__percpu sgp;
7153 struct sched_domain ** __percpu sd;
7154 struct root_domain *rd;
7164 struct sched_domain_topology_level;
7166 typedef struct sched_domain *(*sched_domain_init_f)(struct sched_domain_topology_level *tl, int cpu);
7167 typedef const struct cpumask *(*sched_domain_mask_f)(int cpu);
7169 #define SDTL_OVERLAP 0x01
7171 struct sched_domain_topology_level {
7172 sched_domain_init_f init;
7173 sched_domain_mask_f mask;
7175 struct sd_data data;
7179 build_overlap_sched_groups(struct sched_domain *sd, int cpu)
7181 struct sched_group *first = NULL, *last = NULL, *groups = NULL, *sg;
7182 const struct cpumask *span = sched_domain_span(sd);
7183 struct cpumask *covered = sched_domains_tmpmask;
7184 struct sd_data *sdd = sd->private;
7185 struct sched_domain *child;
7188 cpumask_clear(covered);
7190 for_each_cpu(i, span) {
7191 struct cpumask *sg_span;
7193 if (cpumask_test_cpu(i, covered))
7196 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7197 GFP_KERNEL, cpu_to_node(i));
7202 sg_span = sched_group_cpus(sg);
7204 child = *per_cpu_ptr(sdd->sd, i);
7206 child = child->child;
7207 cpumask_copy(sg_span, sched_domain_span(child));
7209 cpumask_set_cpu(i, sg_span);
7211 cpumask_or(covered, covered, sg_span);
7213 sg->sgp = *per_cpu_ptr(sdd->sgp, cpumask_first(sg_span));
7214 atomic_inc(&sg->sgp->ref);
7216 if (cpumask_test_cpu(cpu, sg_span))
7226 sd->groups = groups;
7231 free_sched_groups(first, 0);
7236 static int get_group(int cpu, struct sd_data *sdd, struct sched_group **sg)
7238 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, cpu);
7239 struct sched_domain *child = sd->child;
7242 cpu = cpumask_first(sched_domain_span(child));
7245 *sg = *per_cpu_ptr(sdd->sg, cpu);
7246 (*sg)->sgp = *per_cpu_ptr(sdd->sgp, cpu);
7247 atomic_set(&(*sg)->sgp->ref, 1); /* for claim_allocations */
7254 * build_sched_groups will build a circular linked list of the groups
7255 * covered by the given span, and will set each group's ->cpumask correctly,
7256 * and ->cpu_power to 0.
7258 * Assumes the sched_domain tree is fully constructed
7261 build_sched_groups(struct sched_domain *sd, int cpu)
7263 struct sched_group *first = NULL, *last = NULL;
7264 struct sd_data *sdd = sd->private;
7265 const struct cpumask *span = sched_domain_span(sd);
7266 struct cpumask *covered;
7269 get_group(cpu, sdd, &sd->groups);
7270 atomic_inc(&sd->groups->ref);
7272 if (cpu != cpumask_first(sched_domain_span(sd)))
7275 lockdep_assert_held(&sched_domains_mutex);
7276 covered = sched_domains_tmpmask;
7278 cpumask_clear(covered);
7280 for_each_cpu(i, span) {
7281 struct sched_group *sg;
7282 int group = get_group(i, sdd, &sg);
7285 if (cpumask_test_cpu(i, covered))
7288 cpumask_clear(sched_group_cpus(sg));
7291 for_each_cpu(j, span) {
7292 if (get_group(j, sdd, NULL) != group)
7295 cpumask_set_cpu(j, covered);
7296 cpumask_set_cpu(j, sched_group_cpus(sg));
7311 * Initialize sched groups cpu_power.
7313 * cpu_power indicates the capacity of sched group, which is used while
7314 * distributing the load between different sched groups in a sched domain.
7315 * Typically cpu_power for all the groups in a sched domain will be same unless
7316 * there are asymmetries in the topology. If there are asymmetries, group
7317 * having more cpu_power will pickup more load compared to the group having
7320 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7322 struct sched_group *sg = sd->groups;
7324 WARN_ON(!sd || !sg);
7327 sg->group_weight = cpumask_weight(sched_group_cpus(sg));
7329 } while (sg != sd->groups);
7331 if (cpu != group_first_cpu(sg))
7334 update_group_power(sd, cpu);
7338 * Initializers for schedule domains
7339 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7342 #ifdef CONFIG_SCHED_DEBUG
7343 # define SD_INIT_NAME(sd, type) sd->name = #type
7345 # define SD_INIT_NAME(sd, type) do { } while (0)
7348 #define SD_INIT_FUNC(type) \
7349 static noinline struct sched_domain * \
7350 sd_init_##type(struct sched_domain_topology_level *tl, int cpu) \
7352 struct sched_domain *sd = *per_cpu_ptr(tl->data.sd, cpu); \
7353 *sd = SD_##type##_INIT; \
7354 SD_INIT_NAME(sd, type); \
7355 sd->private = &tl->data; \
7361 SD_INIT_FUNC(ALLNODES)
7364 #ifdef CONFIG_SCHED_SMT
7365 SD_INIT_FUNC(SIBLING)
7367 #ifdef CONFIG_SCHED_MC
7370 #ifdef CONFIG_SCHED_BOOK
7374 static int default_relax_domain_level = -1;
7375 int sched_domain_level_max;
7377 static int __init setup_relax_domain_level(char *str)
7381 val = simple_strtoul(str, NULL, 0);
7382 if (val < sched_domain_level_max)
7383 default_relax_domain_level = val;
7387 __setup("relax_domain_level=", setup_relax_domain_level);
7389 static void set_domain_attribute(struct sched_domain *sd,
7390 struct sched_domain_attr *attr)
7394 if (!attr || attr->relax_domain_level < 0) {
7395 if (default_relax_domain_level < 0)
7398 request = default_relax_domain_level;
7400 request = attr->relax_domain_level;
7401 if (request < sd->level) {
7402 /* turn off idle balance on this domain */
7403 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7405 /* turn on idle balance on this domain */
7406 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7410 static void __sdt_free(const struct cpumask *cpu_map);
7411 static int __sdt_alloc(const struct cpumask *cpu_map);
7413 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7414 const struct cpumask *cpu_map)
7418 if (!atomic_read(&d->rd->refcount))
7419 free_rootdomain(&d->rd->rcu); /* fall through */
7421 free_percpu(d->sd); /* fall through */
7423 __sdt_free(cpu_map); /* fall through */
7429 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7430 const struct cpumask *cpu_map)
7432 memset(d, 0, sizeof(*d));
7434 if (__sdt_alloc(cpu_map))
7435 return sa_sd_storage;
7436 d->sd = alloc_percpu(struct sched_domain *);
7438 return sa_sd_storage;
7439 d->rd = alloc_rootdomain();
7442 return sa_rootdomain;
7446 * NULL the sd_data elements we've used to build the sched_domain and
7447 * sched_group structure so that the subsequent __free_domain_allocs()
7448 * will not free the data we're using.
7450 static void claim_allocations(int cpu, struct sched_domain *sd)
7452 struct sd_data *sdd = sd->private;
7454 WARN_ON_ONCE(*per_cpu_ptr(sdd->sd, cpu) != sd);
7455 *per_cpu_ptr(sdd->sd, cpu) = NULL;
7457 if (atomic_read(&(*per_cpu_ptr(sdd->sg, cpu))->ref))
7458 *per_cpu_ptr(sdd->sg, cpu) = NULL;
7460 if (atomic_read(&(*per_cpu_ptr(sdd->sgp, cpu))->ref))
7461 *per_cpu_ptr(sdd->sgp, cpu) = NULL;
7464 #ifdef CONFIG_SCHED_SMT
7465 static const struct cpumask *cpu_smt_mask(int cpu)
7467 return topology_thread_cpumask(cpu);
7472 * Topology list, bottom-up.
7474 static struct sched_domain_topology_level default_topology[] = {
7475 #ifdef CONFIG_SCHED_SMT
7476 { sd_init_SIBLING, cpu_smt_mask, },
7478 #ifdef CONFIG_SCHED_MC
7479 { sd_init_MC, cpu_coregroup_mask, },
7481 #ifdef CONFIG_SCHED_BOOK
7482 { sd_init_BOOK, cpu_book_mask, },
7484 { sd_init_CPU, cpu_cpu_mask, },
7486 { sd_init_NODE, cpu_node_mask, SDTL_OVERLAP, },
7487 { sd_init_ALLNODES, cpu_allnodes_mask, },
7492 static struct sched_domain_topology_level *sched_domain_topology = default_topology;
7494 static int __sdt_alloc(const struct cpumask *cpu_map)
7496 struct sched_domain_topology_level *tl;
7499 for (tl = sched_domain_topology; tl->init; tl++) {
7500 struct sd_data *sdd = &tl->data;
7502 sdd->sd = alloc_percpu(struct sched_domain *);
7506 sdd->sg = alloc_percpu(struct sched_group *);
7510 sdd->sgp = alloc_percpu(struct sched_group_power *);
7514 for_each_cpu(j, cpu_map) {
7515 struct sched_domain *sd;
7516 struct sched_group *sg;
7517 struct sched_group_power *sgp;
7519 sd = kzalloc_node(sizeof(struct sched_domain) + cpumask_size(),
7520 GFP_KERNEL, cpu_to_node(j));
7524 *per_cpu_ptr(sdd->sd, j) = sd;
7526 sg = kzalloc_node(sizeof(struct sched_group) + cpumask_size(),
7527 GFP_KERNEL, cpu_to_node(j));
7531 *per_cpu_ptr(sdd->sg, j) = sg;
7533 sgp = kzalloc_node(sizeof(struct sched_group_power),
7534 GFP_KERNEL, cpu_to_node(j));
7538 *per_cpu_ptr(sdd->sgp, j) = sgp;
7545 static void __sdt_free(const struct cpumask *cpu_map)
7547 struct sched_domain_topology_level *tl;
7550 for (tl = sched_domain_topology; tl->init; tl++) {
7551 struct sd_data *sdd = &tl->data;
7553 for_each_cpu(j, cpu_map) {
7554 struct sched_domain *sd = *per_cpu_ptr(sdd->sd, j);
7555 if (sd && (sd->flags & SD_OVERLAP))
7556 free_sched_groups(sd->groups, 0);
7557 kfree(*per_cpu_ptr(sdd->sg, j));
7558 kfree(*per_cpu_ptr(sdd->sgp, j));
7560 free_percpu(sdd->sd);
7561 free_percpu(sdd->sg);
7562 free_percpu(sdd->sgp);
7566 struct sched_domain *build_sched_domain(struct sched_domain_topology_level *tl,
7567 struct s_data *d, const struct cpumask *cpu_map,
7568 struct sched_domain_attr *attr, struct sched_domain *child,
7571 struct sched_domain *sd = tl->init(tl, cpu);
7575 set_domain_attribute(sd, attr);
7576 cpumask_and(sched_domain_span(sd), cpu_map, tl->mask(cpu));
7578 sd->level = child->level + 1;
7579 sched_domain_level_max = max(sched_domain_level_max, sd->level);
7588 * Build sched domains for a given set of cpus and attach the sched domains
7589 * to the individual cpus
7591 static int build_sched_domains(const struct cpumask *cpu_map,
7592 struct sched_domain_attr *attr)
7594 enum s_alloc alloc_state = sa_none;
7595 struct sched_domain *sd;
7597 int i, ret = -ENOMEM;
7599 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7600 if (alloc_state != sa_rootdomain)
7603 /* Set up domains for cpus specified by the cpu_map. */
7604 for_each_cpu(i, cpu_map) {
7605 struct sched_domain_topology_level *tl;
7608 for (tl = sched_domain_topology; tl->init; tl++) {
7609 sd = build_sched_domain(tl, &d, cpu_map, attr, sd, i);
7610 if (tl->flags & SDTL_OVERLAP || sched_feat(FORCE_SD_OVERLAP))
7611 sd->flags |= SD_OVERLAP;
7612 if (cpumask_equal(cpu_map, sched_domain_span(sd)))
7619 *per_cpu_ptr(d.sd, i) = sd;
7622 /* Build the groups for the domains */
7623 for_each_cpu(i, cpu_map) {
7624 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7625 sd->span_weight = cpumask_weight(sched_domain_span(sd));
7626 if (sd->flags & SD_OVERLAP) {
7627 if (build_overlap_sched_groups(sd, i))
7630 if (build_sched_groups(sd, i))
7636 /* Calculate CPU power for physical packages and nodes */
7637 for (i = nr_cpumask_bits-1; i >= 0; i--) {
7638 if (!cpumask_test_cpu(i, cpu_map))
7641 for (sd = *per_cpu_ptr(d.sd, i); sd; sd = sd->parent) {
7642 claim_allocations(i, sd);
7643 init_sched_groups_power(i, sd);
7647 /* Attach the domains */
7649 for_each_cpu(i, cpu_map) {
7650 sd = *per_cpu_ptr(d.sd, i);
7651 cpu_attach_domain(sd, d.rd, i);
7657 __free_domain_allocs(&d, alloc_state, cpu_map);
7661 static cpumask_var_t *doms_cur; /* current sched domains */
7662 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7663 static struct sched_domain_attr *dattr_cur;
7664 /* attribues of custom domains in 'doms_cur' */
7667 * Special case: If a kmalloc of a doms_cur partition (array of
7668 * cpumask) fails, then fallback to a single sched domain,
7669 * as determined by the single cpumask fallback_doms.
7671 static cpumask_var_t fallback_doms;
7674 * arch_update_cpu_topology lets virtualized architectures update the
7675 * cpu core maps. It is supposed to return 1 if the topology changed
7676 * or 0 if it stayed the same.
7678 int __attribute__((weak)) arch_update_cpu_topology(void)
7683 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7686 cpumask_var_t *doms;
7688 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7691 for (i = 0; i < ndoms; i++) {
7692 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7693 free_sched_domains(doms, i);
7700 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7703 for (i = 0; i < ndoms; i++)
7704 free_cpumask_var(doms[i]);
7709 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7710 * For now this just excludes isolated cpus, but could be used to
7711 * exclude other special cases in the future.
7713 static int init_sched_domains(const struct cpumask *cpu_map)
7717 arch_update_cpu_topology();
7719 doms_cur = alloc_sched_domains(ndoms_cur);
7721 doms_cur = &fallback_doms;
7722 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7724 err = build_sched_domains(doms_cur[0], NULL);
7725 register_sched_domain_sysctl();
7731 * Detach sched domains from a group of cpus specified in cpu_map
7732 * These cpus will now be attached to the NULL domain
7734 static void detach_destroy_domains(const struct cpumask *cpu_map)
7739 for_each_cpu(i, cpu_map)
7740 cpu_attach_domain(NULL, &def_root_domain, i);
7744 /* handle null as "default" */
7745 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7746 struct sched_domain_attr *new, int idx_new)
7748 struct sched_domain_attr tmp;
7755 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7756 new ? (new + idx_new) : &tmp,
7757 sizeof(struct sched_domain_attr));
7761 * Partition sched domains as specified by the 'ndoms_new'
7762 * cpumasks in the array doms_new[] of cpumasks. This compares
7763 * doms_new[] to the current sched domain partitioning, doms_cur[].
7764 * It destroys each deleted domain and builds each new domain.
7766 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7767 * The masks don't intersect (don't overlap.) We should setup one
7768 * sched domain for each mask. CPUs not in any of the cpumasks will
7769 * not be load balanced. If the same cpumask appears both in the
7770 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7773 * The passed in 'doms_new' should be allocated using
7774 * alloc_sched_domains. This routine takes ownership of it and will
7775 * free_sched_domains it when done with it. If the caller failed the
7776 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7777 * and partition_sched_domains() will fallback to the single partition
7778 * 'fallback_doms', it also forces the domains to be rebuilt.
7780 * If doms_new == NULL it will be replaced with cpu_online_mask.
7781 * ndoms_new == 0 is a special case for destroying existing domains,
7782 * and it will not create the default domain.
7784 * Call with hotplug lock held
7786 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7787 struct sched_domain_attr *dattr_new)
7792 mutex_lock(&sched_domains_mutex);
7794 /* always unregister in case we don't destroy any domains */
7795 unregister_sched_domain_sysctl();
7797 /* Let architecture update cpu core mappings. */
7798 new_topology = arch_update_cpu_topology();
7800 n = doms_new ? ndoms_new : 0;
7802 /* Destroy deleted domains */
7803 for (i = 0; i < ndoms_cur; i++) {
7804 for (j = 0; j < n && !new_topology; j++) {
7805 if (cpumask_equal(doms_cur[i], doms_new[j])
7806 && dattrs_equal(dattr_cur, i, dattr_new, j))
7809 /* no match - a current sched domain not in new doms_new[] */
7810 detach_destroy_domains(doms_cur[i]);
7815 if (doms_new == NULL) {
7817 doms_new = &fallback_doms;
7818 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7819 WARN_ON_ONCE(dattr_new);
7822 /* Build new domains */
7823 for (i = 0; i < ndoms_new; i++) {
7824 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7825 if (cpumask_equal(doms_new[i], doms_cur[j])
7826 && dattrs_equal(dattr_new, i, dattr_cur, j))
7829 /* no match - add a new doms_new */
7830 build_sched_domains(doms_new[i], dattr_new ? dattr_new + i : NULL);
7835 /* Remember the new sched domains */
7836 if (doms_cur != &fallback_doms)
7837 free_sched_domains(doms_cur, ndoms_cur);
7838 kfree(dattr_cur); /* kfree(NULL) is safe */
7839 doms_cur = doms_new;
7840 dattr_cur = dattr_new;
7841 ndoms_cur = ndoms_new;
7843 register_sched_domain_sysctl();
7845 mutex_unlock(&sched_domains_mutex);
7848 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7849 static void reinit_sched_domains(void)
7853 /* Destroy domains first to force the rebuild */
7854 partition_sched_domains(0, NULL, NULL);
7856 rebuild_sched_domains();
7860 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7862 unsigned int level = 0;
7864 if (sscanf(buf, "%u", &level) != 1)
7868 * level is always be positive so don't check for
7869 * level < POWERSAVINGS_BALANCE_NONE which is 0
7870 * What happens on 0 or 1 byte write,
7871 * need to check for count as well?
7874 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7878 sched_smt_power_savings = level;
7880 sched_mc_power_savings = level;
7882 reinit_sched_domains();
7887 #ifdef CONFIG_SCHED_MC
7888 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7889 struct sysdev_class_attribute *attr,
7892 return sprintf(page, "%u\n", sched_mc_power_savings);
7894 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7895 struct sysdev_class_attribute *attr,
7896 const char *buf, size_t count)
7898 return sched_power_savings_store(buf, count, 0);
7900 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7901 sched_mc_power_savings_show,
7902 sched_mc_power_savings_store);
7905 #ifdef CONFIG_SCHED_SMT
7906 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7907 struct sysdev_class_attribute *attr,
7910 return sprintf(page, "%u\n", sched_smt_power_savings);
7912 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7913 struct sysdev_class_attribute *attr,
7914 const char *buf, size_t count)
7916 return sched_power_savings_store(buf, count, 1);
7918 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7919 sched_smt_power_savings_show,
7920 sched_smt_power_savings_store);
7923 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7927 #ifdef CONFIG_SCHED_SMT
7929 err = sysfs_create_file(&cls->kset.kobj,
7930 &attr_sched_smt_power_savings.attr);
7932 #ifdef CONFIG_SCHED_MC
7933 if (!err && mc_capable())
7934 err = sysfs_create_file(&cls->kset.kobj,
7935 &attr_sched_mc_power_savings.attr);
7939 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7942 * Update cpusets according to cpu_active mask. If cpusets are
7943 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7944 * around partition_sched_domains().
7946 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7949 switch (action & ~CPU_TASKS_FROZEN) {
7951 case CPU_DOWN_FAILED:
7952 cpuset_update_active_cpus();
7959 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7962 switch (action & ~CPU_TASKS_FROZEN) {
7963 case CPU_DOWN_PREPARE:
7964 cpuset_update_active_cpus();
7971 static int update_runtime(struct notifier_block *nfb,
7972 unsigned long action, void *hcpu)
7974 int cpu = (int)(long)hcpu;
7977 case CPU_DOWN_PREPARE:
7978 case CPU_DOWN_PREPARE_FROZEN:
7979 disable_runtime(cpu_rq(cpu));
7982 case CPU_DOWN_FAILED:
7983 case CPU_DOWN_FAILED_FROZEN:
7985 case CPU_ONLINE_FROZEN:
7986 enable_runtime(cpu_rq(cpu));
7994 void __init sched_init_smp(void)
7996 cpumask_var_t non_isolated_cpus;
7998 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7999 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8002 mutex_lock(&sched_domains_mutex);
8003 init_sched_domains(cpu_active_mask);
8004 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8005 if (cpumask_empty(non_isolated_cpus))
8006 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8007 mutex_unlock(&sched_domains_mutex);
8010 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
8011 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
8013 /* RT runtime code needs to handle some hotplug events */
8014 hotcpu_notifier(update_runtime, 0);
8018 /* Move init over to a non-isolated CPU */
8019 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8021 sched_init_granularity();
8022 free_cpumask_var(non_isolated_cpus);
8024 init_sched_rt_class();
8027 void __init sched_init_smp(void)
8029 sched_init_granularity();
8031 #endif /* CONFIG_SMP */
8033 const_debug unsigned int sysctl_timer_migration = 1;
8035 int in_sched_functions(unsigned long addr)
8037 return in_lock_functions(addr) ||
8038 (addr >= (unsigned long)__sched_text_start
8039 && addr < (unsigned long)__sched_text_end);
8042 static void init_cfs_rq(struct cfs_rq *cfs_rq)
8044 cfs_rq->tasks_timeline = RB_ROOT;
8045 INIT_LIST_HEAD(&cfs_rq->tasks);
8046 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8047 #ifndef CONFIG_64BIT
8048 cfs_rq->min_vruntime_copy = cfs_rq->min_vruntime;
8052 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8054 struct rt_prio_array *array;
8057 array = &rt_rq->active;
8058 for (i = 0; i < MAX_RT_PRIO; i++) {
8059 INIT_LIST_HEAD(array->queue + i);
8060 __clear_bit(i, array->bitmap);
8062 /* delimiter for bitsearch: */
8063 __set_bit(MAX_RT_PRIO, array->bitmap);
8065 #if defined CONFIG_SMP
8066 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8067 rt_rq->highest_prio.next = MAX_RT_PRIO;
8068 rt_rq->rt_nr_migratory = 0;
8069 rt_rq->overloaded = 0;
8070 plist_head_init(&rt_rq->pushable_tasks);
8074 rt_rq->rt_throttled = 0;
8075 rt_rq->rt_runtime = 0;
8076 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8079 #ifdef CONFIG_FAIR_GROUP_SCHED
8080 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8081 struct sched_entity *se, int cpu,
8082 struct sched_entity *parent)
8084 struct rq *rq = cpu_rq(cpu);
8089 /* allow initial update_cfs_load() to truncate */
8090 cfs_rq->load_stamp = 1;
8092 init_cfs_rq_runtime(cfs_rq);
8094 tg->cfs_rq[cpu] = cfs_rq;
8097 /* se could be NULL for root_task_group */
8102 se->cfs_rq = &rq->cfs;
8104 se->cfs_rq = parent->my_q;
8107 update_load_set(&se->load, 0);
8108 se->parent = parent;
8112 #ifdef CONFIG_RT_GROUP_SCHED
8113 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8114 struct sched_rt_entity *rt_se, int cpu,
8115 struct sched_rt_entity *parent)
8117 struct rq *rq = cpu_rq(cpu);
8119 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8120 rt_rq->rt_nr_boosted = 0;
8124 tg->rt_rq[cpu] = rt_rq;
8125 tg->rt_se[cpu] = rt_se;
8131 rt_se->rt_rq = &rq->rt;
8133 rt_se->rt_rq = parent->my_q;
8135 rt_se->my_q = rt_rq;
8136 rt_se->parent = parent;
8137 INIT_LIST_HEAD(&rt_se->run_list);
8141 void __init sched_init(void)
8144 unsigned long alloc_size = 0, ptr;
8146 #ifdef CONFIG_FAIR_GROUP_SCHED
8147 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8149 #ifdef CONFIG_RT_GROUP_SCHED
8150 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8152 #ifdef CONFIG_CPUMASK_OFFSTACK
8153 alloc_size += num_possible_cpus() * cpumask_size();
8156 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8158 #ifdef CONFIG_FAIR_GROUP_SCHED
8159 root_task_group.se = (struct sched_entity **)ptr;
8160 ptr += nr_cpu_ids * sizeof(void **);
8162 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8163 ptr += nr_cpu_ids * sizeof(void **);
8165 #endif /* CONFIG_FAIR_GROUP_SCHED */
8166 #ifdef CONFIG_RT_GROUP_SCHED
8167 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8168 ptr += nr_cpu_ids * sizeof(void **);
8170 root_task_group.rt_rq = (struct rt_rq **)ptr;
8171 ptr += nr_cpu_ids * sizeof(void **);
8173 #endif /* CONFIG_RT_GROUP_SCHED */
8174 #ifdef CONFIG_CPUMASK_OFFSTACK
8175 for_each_possible_cpu(i) {
8176 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8177 ptr += cpumask_size();
8179 #endif /* CONFIG_CPUMASK_OFFSTACK */
8183 init_defrootdomain();
8186 init_rt_bandwidth(&def_rt_bandwidth,
8187 global_rt_period(), global_rt_runtime());
8189 #ifdef CONFIG_RT_GROUP_SCHED
8190 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8191 global_rt_period(), global_rt_runtime());
8192 #endif /* CONFIG_RT_GROUP_SCHED */
8194 #ifdef CONFIG_CGROUP_SCHED
8195 list_add(&root_task_group.list, &task_groups);
8196 INIT_LIST_HEAD(&root_task_group.children);
8197 autogroup_init(&init_task);
8198 #endif /* CONFIG_CGROUP_SCHED */
8200 for_each_possible_cpu(i) {
8204 raw_spin_lock_init(&rq->lock);
8206 rq->calc_load_active = 0;
8207 rq->calc_load_update = jiffies + LOAD_FREQ;
8208 init_cfs_rq(&rq->cfs);
8209 init_rt_rq(&rq->rt, rq);
8210 #ifdef CONFIG_FAIR_GROUP_SCHED
8211 root_task_group.shares = root_task_group_load;
8212 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8214 * How much cpu bandwidth does root_task_group get?
8216 * In case of task-groups formed thr' the cgroup filesystem, it
8217 * gets 100% of the cpu resources in the system. This overall
8218 * system cpu resource is divided among the tasks of
8219 * root_task_group and its child task-groups in a fair manner,
8220 * based on each entity's (task or task-group's) weight
8221 * (se->load.weight).
8223 * In other words, if root_task_group has 10 tasks of weight
8224 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8225 * then A0's share of the cpu resource is:
8227 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8229 * We achieve this by letting root_task_group's tasks sit
8230 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8232 init_cfs_bandwidth(&root_task_group.cfs_bandwidth);
8233 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8234 #endif /* CONFIG_FAIR_GROUP_SCHED */
8236 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8237 #ifdef CONFIG_RT_GROUP_SCHED
8238 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8239 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8242 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8243 rq->cpu_load[j] = 0;
8245 rq->last_load_update_tick = jiffies;
8250 rq->cpu_power = SCHED_POWER_SCALE;
8251 rq->post_schedule = 0;
8252 rq->active_balance = 0;
8253 rq->next_balance = jiffies;
8258 rq->avg_idle = 2*sysctl_sched_migration_cost;
8259 rq_attach_root(rq, &def_root_domain);
8261 rq->nohz_balance_kick = 0;
8262 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8266 atomic_set(&rq->nr_iowait, 0);
8269 set_load_weight(&init_task);
8271 #ifdef CONFIG_PREEMPT_NOTIFIERS
8272 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8276 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8279 #ifdef CONFIG_RT_MUTEXES
8280 plist_head_init(&init_task.pi_waiters);
8284 * The boot idle thread does lazy MMU switching as well:
8286 atomic_inc(&init_mm.mm_count);
8287 enter_lazy_tlb(&init_mm, current);
8290 * Make us the idle thread. Technically, schedule() should not be
8291 * called from this thread, however somewhere below it might be,
8292 * but because we are the idle thread, we just pick up running again
8293 * when this runqueue becomes "idle".
8295 init_idle(current, smp_processor_id());
8297 calc_load_update = jiffies + LOAD_FREQ;
8300 * During early bootup we pretend to be a normal task:
8302 current->sched_class = &fair_sched_class;
8304 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8305 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8307 zalloc_cpumask_var(&sched_domains_tmpmask, GFP_NOWAIT);
8309 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8310 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8311 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8312 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8313 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8315 /* May be allocated at isolcpus cmdline parse time */
8316 if (cpu_isolated_map == NULL)
8317 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8320 scheduler_running = 1;
8323 #ifdef CONFIG_DEBUG_ATOMIC_SLEEP
8324 static inline int preempt_count_equals(int preempt_offset)
8326 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8328 return (nested == preempt_offset);
8331 void __might_sleep(const char *file, int line, int preempt_offset)
8333 static unsigned long prev_jiffy; /* ratelimiting */
8335 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8336 system_state != SYSTEM_RUNNING || oops_in_progress)
8338 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8340 prev_jiffy = jiffies;
8343 "BUG: sleeping function called from invalid context at %s:%d\n",
8346 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8347 in_atomic(), irqs_disabled(),
8348 current->pid, current->comm);
8350 debug_show_held_locks(current);
8351 if (irqs_disabled())
8352 print_irqtrace_events(current);
8355 EXPORT_SYMBOL(__might_sleep);
8358 #ifdef CONFIG_MAGIC_SYSRQ
8359 static void normalize_task(struct rq *rq, struct task_struct *p)
8361 const struct sched_class *prev_class = p->sched_class;
8362 int old_prio = p->prio;
8367 deactivate_task(rq, p, 0);
8368 __setscheduler(rq, p, SCHED_NORMAL, 0);
8370 activate_task(rq, p, 0);
8371 resched_task(rq->curr);
8374 check_class_changed(rq, p, prev_class, old_prio);
8377 void normalize_rt_tasks(void)
8379 struct task_struct *g, *p;
8380 unsigned long flags;
8383 read_lock_irqsave(&tasklist_lock, flags);
8384 do_each_thread(g, p) {
8386 * Only normalize user tasks:
8391 p->se.exec_start = 0;
8392 #ifdef CONFIG_SCHEDSTATS
8393 p->se.statistics.wait_start = 0;
8394 p->se.statistics.sleep_start = 0;
8395 p->se.statistics.block_start = 0;
8400 * Renice negative nice level userspace
8403 if (TASK_NICE(p) < 0 && p->mm)
8404 set_user_nice(p, 0);
8408 raw_spin_lock(&p->pi_lock);
8409 rq = __task_rq_lock(p);
8411 normalize_task(rq, p);
8413 __task_rq_unlock(rq);
8414 raw_spin_unlock(&p->pi_lock);
8415 } while_each_thread(g, p);
8417 read_unlock_irqrestore(&tasklist_lock, flags);
8420 #endif /* CONFIG_MAGIC_SYSRQ */
8422 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8424 * These functions are only useful for the IA64 MCA handling, or kdb.
8426 * They can only be called when the whole system has been
8427 * stopped - every CPU needs to be quiescent, and no scheduling
8428 * activity can take place. Using them for anything else would
8429 * be a serious bug, and as a result, they aren't even visible
8430 * under any other configuration.
8434 * curr_task - return the current task for a given cpu.
8435 * @cpu: the processor in question.
8437 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8439 struct task_struct *curr_task(int cpu)
8441 return cpu_curr(cpu);
8444 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8448 * set_curr_task - set the current task for a given cpu.
8449 * @cpu: the processor in question.
8450 * @p: the task pointer to set.
8452 * Description: This function must only be used when non-maskable interrupts
8453 * are serviced on a separate stack. It allows the architecture to switch the
8454 * notion of the current task on a cpu in a non-blocking manner. This function
8455 * must be called with all CPU's synchronized, and interrupts disabled, the
8456 * and caller must save the original value of the current task (see
8457 * curr_task() above) and restore that value before reenabling interrupts and
8458 * re-starting the system.
8460 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8462 void set_curr_task(int cpu, struct task_struct *p)
8469 #ifdef CONFIG_FAIR_GROUP_SCHED
8470 static void free_fair_sched_group(struct task_group *tg)
8474 destroy_cfs_bandwidth(tg_cfs_bandwidth(tg));
8476 for_each_possible_cpu(i) {
8478 kfree(tg->cfs_rq[i]);
8488 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8490 struct cfs_rq *cfs_rq;
8491 struct sched_entity *se;
8494 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8497 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8501 tg->shares = NICE_0_LOAD;
8503 init_cfs_bandwidth(tg_cfs_bandwidth(tg));
8505 for_each_possible_cpu(i) {
8506 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8507 GFP_KERNEL, cpu_to_node(i));
8511 se = kzalloc_node(sizeof(struct sched_entity),
8512 GFP_KERNEL, cpu_to_node(i));
8516 init_cfs_rq(cfs_rq);
8517 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8528 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8530 struct rq *rq = cpu_rq(cpu);
8531 unsigned long flags;
8534 * Only empty task groups can be destroyed; so we can speculatively
8535 * check on_list without danger of it being re-added.
8537 if (!tg->cfs_rq[cpu]->on_list)
8540 raw_spin_lock_irqsave(&rq->lock, flags);
8541 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8542 raw_spin_unlock_irqrestore(&rq->lock, flags);
8544 #else /* !CONFIG_FAIR_GROUP_SCHED */
8545 static inline void free_fair_sched_group(struct task_group *tg)
8550 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8555 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8558 #endif /* CONFIG_FAIR_GROUP_SCHED */
8560 #ifdef CONFIG_RT_GROUP_SCHED
8561 static void free_rt_sched_group(struct task_group *tg)
8566 destroy_rt_bandwidth(&tg->rt_bandwidth);
8568 for_each_possible_cpu(i) {
8570 kfree(tg->rt_rq[i]);
8572 kfree(tg->rt_se[i]);
8580 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8582 struct rt_rq *rt_rq;
8583 struct sched_rt_entity *rt_se;
8586 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8589 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8593 init_rt_bandwidth(&tg->rt_bandwidth,
8594 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8596 for_each_possible_cpu(i) {
8597 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8598 GFP_KERNEL, cpu_to_node(i));
8602 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8603 GFP_KERNEL, cpu_to_node(i));
8607 init_rt_rq(rt_rq, cpu_rq(i));
8608 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8609 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8619 #else /* !CONFIG_RT_GROUP_SCHED */
8620 static inline void free_rt_sched_group(struct task_group *tg)
8625 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8629 #endif /* CONFIG_RT_GROUP_SCHED */
8631 #ifdef CONFIG_CGROUP_SCHED
8632 static void free_sched_group(struct task_group *tg)
8634 free_fair_sched_group(tg);
8635 free_rt_sched_group(tg);
8640 /* allocate runqueue etc for a new task group */
8641 struct task_group *sched_create_group(struct task_group *parent)
8643 struct task_group *tg;
8644 unsigned long flags;
8646 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8648 return ERR_PTR(-ENOMEM);
8650 if (!alloc_fair_sched_group(tg, parent))
8653 if (!alloc_rt_sched_group(tg, parent))
8656 spin_lock_irqsave(&task_group_lock, flags);
8657 list_add_rcu(&tg->list, &task_groups);
8659 WARN_ON(!parent); /* root should already exist */
8661 tg->parent = parent;
8662 INIT_LIST_HEAD(&tg->children);
8663 list_add_rcu(&tg->siblings, &parent->children);
8664 spin_unlock_irqrestore(&task_group_lock, flags);
8669 free_sched_group(tg);
8670 return ERR_PTR(-ENOMEM);
8673 /* rcu callback to free various structures associated with a task group */
8674 static void free_sched_group_rcu(struct rcu_head *rhp)
8676 /* now it should be safe to free those cfs_rqs */
8677 free_sched_group(container_of(rhp, struct task_group, rcu));
8680 /* Destroy runqueue etc associated with a task group */
8681 void sched_destroy_group(struct task_group *tg)
8683 unsigned long flags;
8686 /* end participation in shares distribution */
8687 for_each_possible_cpu(i)
8688 unregister_fair_sched_group(tg, i);
8690 spin_lock_irqsave(&task_group_lock, flags);
8691 list_del_rcu(&tg->list);
8692 list_del_rcu(&tg->siblings);
8693 spin_unlock_irqrestore(&task_group_lock, flags);
8695 /* wait for possible concurrent references to cfs_rqs complete */
8696 call_rcu(&tg->rcu, free_sched_group_rcu);
8699 /* change task's runqueue when it moves between groups.
8700 * The caller of this function should have put the task in its new group
8701 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8702 * reflect its new group.
8704 void sched_move_task(struct task_struct *tsk)
8707 unsigned long flags;
8710 rq = task_rq_lock(tsk, &flags);
8712 running = task_current(rq, tsk);
8716 dequeue_task(rq, tsk, 0);
8717 if (unlikely(running))
8718 tsk->sched_class->put_prev_task(rq, tsk);
8720 #ifdef CONFIG_FAIR_GROUP_SCHED
8721 if (tsk->sched_class->task_move_group)
8722 tsk->sched_class->task_move_group(tsk, on_rq);
8725 set_task_rq(tsk, task_cpu(tsk));
8727 if (unlikely(running))
8728 tsk->sched_class->set_curr_task(rq);
8730 enqueue_task(rq, tsk, 0);
8732 task_rq_unlock(rq, tsk, &flags);
8734 #endif /* CONFIG_CGROUP_SCHED */
8736 #ifdef CONFIG_FAIR_GROUP_SCHED
8737 static DEFINE_MUTEX(shares_mutex);
8739 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8742 unsigned long flags;
8745 * We can't change the weight of the root cgroup.
8750 shares = clamp(shares, scale_load(MIN_SHARES), scale_load(MAX_SHARES));
8752 mutex_lock(&shares_mutex);
8753 if (tg->shares == shares)
8756 tg->shares = shares;
8757 for_each_possible_cpu(i) {
8758 struct rq *rq = cpu_rq(i);
8759 struct sched_entity *se;
8762 /* Propagate contribution to hierarchy */
8763 raw_spin_lock_irqsave(&rq->lock, flags);
8764 for_each_sched_entity(se)
8765 update_cfs_shares(group_cfs_rq(se));
8766 raw_spin_unlock_irqrestore(&rq->lock, flags);
8770 mutex_unlock(&shares_mutex);
8774 unsigned long sched_group_shares(struct task_group *tg)
8780 #if defined(CONFIG_RT_GROUP_SCHED) || defined(CONFIG_CFS_BANDWIDTH)
8781 static unsigned long to_ratio(u64 period, u64 runtime)
8783 if (runtime == RUNTIME_INF)
8786 return div64_u64(runtime << 20, period);
8790 #ifdef CONFIG_RT_GROUP_SCHED
8792 * Ensure that the real time constraints are schedulable.
8794 static DEFINE_MUTEX(rt_constraints_mutex);
8796 /* Must be called with tasklist_lock held */
8797 static inline int tg_has_rt_tasks(struct task_group *tg)
8799 struct task_struct *g, *p;
8801 do_each_thread(g, p) {
8802 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8804 } while_each_thread(g, p);
8809 struct rt_schedulable_data {
8810 struct task_group *tg;
8815 static int tg_rt_schedulable(struct task_group *tg, void *data)
8817 struct rt_schedulable_data *d = data;
8818 struct task_group *child;
8819 unsigned long total, sum = 0;
8820 u64 period, runtime;
8822 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8823 runtime = tg->rt_bandwidth.rt_runtime;
8826 period = d->rt_period;
8827 runtime = d->rt_runtime;
8831 * Cannot have more runtime than the period.
8833 if (runtime > period && runtime != RUNTIME_INF)
8837 * Ensure we don't starve existing RT tasks.
8839 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8842 total = to_ratio(period, runtime);
8845 * Nobody can have more than the global setting allows.
8847 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8851 * The sum of our children's runtime should not exceed our own.
8853 list_for_each_entry_rcu(child, &tg->children, siblings) {
8854 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8855 runtime = child->rt_bandwidth.rt_runtime;
8857 if (child == d->tg) {
8858 period = d->rt_period;
8859 runtime = d->rt_runtime;
8862 sum += to_ratio(period, runtime);
8871 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8873 struct rt_schedulable_data data = {
8875 .rt_period = period,
8876 .rt_runtime = runtime,
8879 return walk_tg_tree(tg_rt_schedulable, tg_nop, &data);
8882 static int tg_set_rt_bandwidth(struct task_group *tg,
8883 u64 rt_period, u64 rt_runtime)
8887 mutex_lock(&rt_constraints_mutex);
8888 read_lock(&tasklist_lock);
8889 err = __rt_schedulable(tg, rt_period, rt_runtime);
8893 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8894 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8895 tg->rt_bandwidth.rt_runtime = rt_runtime;
8897 for_each_possible_cpu(i) {
8898 struct rt_rq *rt_rq = tg->rt_rq[i];
8900 raw_spin_lock(&rt_rq->rt_runtime_lock);
8901 rt_rq->rt_runtime = rt_runtime;
8902 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8904 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8906 read_unlock(&tasklist_lock);
8907 mutex_unlock(&rt_constraints_mutex);
8912 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8914 u64 rt_runtime, rt_period;
8916 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8917 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8918 if (rt_runtime_us < 0)
8919 rt_runtime = RUNTIME_INF;
8921 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8924 long sched_group_rt_runtime(struct task_group *tg)
8928 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8931 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8932 do_div(rt_runtime_us, NSEC_PER_USEC);
8933 return rt_runtime_us;
8936 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8938 u64 rt_runtime, rt_period;
8940 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8941 rt_runtime = tg->rt_bandwidth.rt_runtime;
8946 return tg_set_rt_bandwidth(tg, rt_period, rt_runtime);
8949 long sched_group_rt_period(struct task_group *tg)
8953 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8954 do_div(rt_period_us, NSEC_PER_USEC);
8955 return rt_period_us;
8958 static int sched_rt_global_constraints(void)
8960 u64 runtime, period;
8963 if (sysctl_sched_rt_period <= 0)
8966 runtime = global_rt_runtime();
8967 period = global_rt_period();
8970 * Sanity check on the sysctl variables.
8972 if (runtime > period && runtime != RUNTIME_INF)
8975 mutex_lock(&rt_constraints_mutex);
8976 read_lock(&tasklist_lock);
8977 ret = __rt_schedulable(NULL, 0, 0);
8978 read_unlock(&tasklist_lock);
8979 mutex_unlock(&rt_constraints_mutex);
8984 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8986 /* Don't accept realtime tasks when there is no way for them to run */
8987 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8993 #else /* !CONFIG_RT_GROUP_SCHED */
8994 static int sched_rt_global_constraints(void)
8996 unsigned long flags;
8999 if (sysctl_sched_rt_period <= 0)
9003 * There's always some RT tasks in the root group
9004 * -- migration, kstopmachine etc..
9006 if (sysctl_sched_rt_runtime == 0)
9009 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9010 for_each_possible_cpu(i) {
9011 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9013 raw_spin_lock(&rt_rq->rt_runtime_lock);
9014 rt_rq->rt_runtime = global_rt_runtime();
9015 raw_spin_unlock(&rt_rq->rt_runtime_lock);
9017 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9021 #endif /* CONFIG_RT_GROUP_SCHED */
9023 int sched_rt_handler(struct ctl_table *table, int write,
9024 void __user *buffer, size_t *lenp,
9028 int old_period, old_runtime;
9029 static DEFINE_MUTEX(mutex);
9032 old_period = sysctl_sched_rt_period;
9033 old_runtime = sysctl_sched_rt_runtime;
9035 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9037 if (!ret && write) {
9038 ret = sched_rt_global_constraints();
9040 sysctl_sched_rt_period = old_period;
9041 sysctl_sched_rt_runtime = old_runtime;
9043 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9044 def_rt_bandwidth.rt_period =
9045 ns_to_ktime(global_rt_period());
9048 mutex_unlock(&mutex);
9053 #ifdef CONFIG_CGROUP_SCHED
9055 /* return corresponding task_group object of a cgroup */
9056 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9058 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9059 struct task_group, css);
9062 static struct cgroup_subsys_state *
9063 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9065 struct task_group *tg, *parent;
9067 if (!cgrp->parent) {
9068 /* This is early initialization for the top cgroup */
9069 return &root_task_group.css;
9072 parent = cgroup_tg(cgrp->parent);
9073 tg = sched_create_group(parent);
9075 return ERR_PTR(-ENOMEM);
9081 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9083 struct task_group *tg = cgroup_tg(cgrp);
9085 sched_destroy_group(tg);
9089 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9091 #ifdef CONFIG_RT_GROUP_SCHED
9092 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9095 /* We don't support RT-tasks being in separate groups */
9096 if (tsk->sched_class != &fair_sched_class)
9103 cpu_cgroup_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9105 sched_move_task(tsk);
9109 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9110 struct cgroup *old_cgrp, struct task_struct *task)
9113 * cgroup_exit() is called in the copy_process() failure path.
9114 * Ignore this case since the task hasn't ran yet, this avoids
9115 * trying to poke a half freed task state from generic code.
9117 if (!(task->flags & PF_EXITING))
9120 sched_move_task(task);
9123 #ifdef CONFIG_FAIR_GROUP_SCHED
9124 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9127 return sched_group_set_shares(cgroup_tg(cgrp), scale_load(shareval));
9130 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9132 struct task_group *tg = cgroup_tg(cgrp);
9134 return (u64) scale_load_down(tg->shares);
9137 #ifdef CONFIG_CFS_BANDWIDTH
9138 static DEFINE_MUTEX(cfs_constraints_mutex);
9140 const u64 max_cfs_quota_period = 1 * NSEC_PER_SEC; /* 1s */
9141 const u64 min_cfs_quota_period = 1 * NSEC_PER_MSEC; /* 1ms */
9143 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 runtime);
9145 static int tg_set_cfs_bandwidth(struct task_group *tg, u64 period, u64 quota)
9147 int i, ret = 0, runtime_enabled;
9148 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9150 if (tg == &root_task_group)
9154 * Ensure we have at some amount of bandwidth every period. This is
9155 * to prevent reaching a state of large arrears when throttled via
9156 * entity_tick() resulting in prolonged exit starvation.
9158 if (quota < min_cfs_quota_period || period < min_cfs_quota_period)
9162 * Likewise, bound things on the otherside by preventing insane quota
9163 * periods. This also allows us to normalize in computing quota
9166 if (period > max_cfs_quota_period)
9169 mutex_lock(&cfs_constraints_mutex);
9170 ret = __cfs_schedulable(tg, period, quota);
9174 runtime_enabled = quota != RUNTIME_INF;
9175 raw_spin_lock_irq(&cfs_b->lock);
9176 cfs_b->period = ns_to_ktime(period);
9177 cfs_b->quota = quota;
9179 __refill_cfs_bandwidth_runtime(cfs_b);
9180 /* restart the period timer (if active) to handle new period expiry */
9181 if (runtime_enabled && cfs_b->timer_active) {
9182 /* force a reprogram */
9183 cfs_b->timer_active = 0;
9184 __start_cfs_bandwidth(cfs_b);
9186 raw_spin_unlock_irq(&cfs_b->lock);
9188 for_each_possible_cpu(i) {
9189 struct cfs_rq *cfs_rq = tg->cfs_rq[i];
9190 struct rq *rq = rq_of(cfs_rq);
9192 raw_spin_lock_irq(&rq->lock);
9193 cfs_rq->runtime_enabled = runtime_enabled;
9194 cfs_rq->runtime_remaining = 0;
9195 raw_spin_unlock_irq(&rq->lock);
9198 mutex_unlock(&cfs_constraints_mutex);
9203 int tg_set_cfs_quota(struct task_group *tg, long cfs_quota_us)
9207 period = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9208 if (cfs_quota_us < 0)
9209 quota = RUNTIME_INF;
9211 quota = (u64)cfs_quota_us * NSEC_PER_USEC;
9213 return tg_set_cfs_bandwidth(tg, period, quota);
9216 long tg_get_cfs_quota(struct task_group *tg)
9220 if (tg_cfs_bandwidth(tg)->quota == RUNTIME_INF)
9223 quota_us = tg_cfs_bandwidth(tg)->quota;
9224 do_div(quota_us, NSEC_PER_USEC);
9229 int tg_set_cfs_period(struct task_group *tg, long cfs_period_us)
9233 period = (u64)cfs_period_us * NSEC_PER_USEC;
9234 quota = tg_cfs_bandwidth(tg)->quota;
9239 return tg_set_cfs_bandwidth(tg, period, quota);
9242 long tg_get_cfs_period(struct task_group *tg)
9246 cfs_period_us = ktime_to_ns(tg_cfs_bandwidth(tg)->period);
9247 do_div(cfs_period_us, NSEC_PER_USEC);
9249 return cfs_period_us;
9252 static s64 cpu_cfs_quota_read_s64(struct cgroup *cgrp, struct cftype *cft)
9254 return tg_get_cfs_quota(cgroup_tg(cgrp));
9257 static int cpu_cfs_quota_write_s64(struct cgroup *cgrp, struct cftype *cftype,
9260 return tg_set_cfs_quota(cgroup_tg(cgrp), cfs_quota_us);
9263 static u64 cpu_cfs_period_read_u64(struct cgroup *cgrp, struct cftype *cft)
9265 return tg_get_cfs_period(cgroup_tg(cgrp));
9268 static int cpu_cfs_period_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9271 return tg_set_cfs_period(cgroup_tg(cgrp), cfs_period_us);
9274 struct cfs_schedulable_data {
9275 struct task_group *tg;
9280 * normalize group quota/period to be quota/max_period
9281 * note: units are usecs
9283 static u64 normalize_cfs_quota(struct task_group *tg,
9284 struct cfs_schedulable_data *d)
9292 period = tg_get_cfs_period(tg);
9293 quota = tg_get_cfs_quota(tg);
9296 /* note: these should typically be equivalent */
9297 if (quota == RUNTIME_INF || quota == -1)
9300 return to_ratio(period, quota);
9303 static int tg_cfs_schedulable_down(struct task_group *tg, void *data)
9305 struct cfs_schedulable_data *d = data;
9306 struct cfs_bandwidth *cfs_b = tg_cfs_bandwidth(tg);
9307 s64 quota = 0, parent_quota = -1;
9310 quota = RUNTIME_INF;
9312 struct cfs_bandwidth *parent_b = tg_cfs_bandwidth(tg->parent);
9314 quota = normalize_cfs_quota(tg, d);
9315 parent_quota = parent_b->hierarchal_quota;
9318 * ensure max(child_quota) <= parent_quota, inherit when no
9321 if (quota == RUNTIME_INF)
9322 quota = parent_quota;
9323 else if (parent_quota != RUNTIME_INF && quota > parent_quota)
9326 cfs_b->hierarchal_quota = quota;
9331 static int __cfs_schedulable(struct task_group *tg, u64 period, u64 quota)
9333 struct cfs_schedulable_data data = {
9339 if (quota != RUNTIME_INF) {
9340 do_div(data.period, NSEC_PER_USEC);
9341 do_div(data.quota, NSEC_PER_USEC);
9344 return walk_tg_tree(tg_cfs_schedulable_down, tg_nop, &data);
9346 #endif /* CONFIG_CFS_BANDWIDTH */
9347 #endif /* CONFIG_FAIR_GROUP_SCHED */
9349 #ifdef CONFIG_RT_GROUP_SCHED
9350 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9353 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9356 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9358 return sched_group_rt_runtime(cgroup_tg(cgrp));
9361 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9364 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9367 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9369 return sched_group_rt_period(cgroup_tg(cgrp));
9371 #endif /* CONFIG_RT_GROUP_SCHED */
9373 static struct cftype cpu_files[] = {
9374 #ifdef CONFIG_FAIR_GROUP_SCHED
9377 .read_u64 = cpu_shares_read_u64,
9378 .write_u64 = cpu_shares_write_u64,
9381 #ifdef CONFIG_CFS_BANDWIDTH
9383 .name = "cfs_quota_us",
9384 .read_s64 = cpu_cfs_quota_read_s64,
9385 .write_s64 = cpu_cfs_quota_write_s64,
9388 .name = "cfs_period_us",
9389 .read_u64 = cpu_cfs_period_read_u64,
9390 .write_u64 = cpu_cfs_period_write_u64,
9393 #ifdef CONFIG_RT_GROUP_SCHED
9395 .name = "rt_runtime_us",
9396 .read_s64 = cpu_rt_runtime_read,
9397 .write_s64 = cpu_rt_runtime_write,
9400 .name = "rt_period_us",
9401 .read_u64 = cpu_rt_period_read_uint,
9402 .write_u64 = cpu_rt_period_write_uint,
9407 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9409 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9412 struct cgroup_subsys cpu_cgroup_subsys = {
9414 .create = cpu_cgroup_create,
9415 .destroy = cpu_cgroup_destroy,
9416 .can_attach_task = cpu_cgroup_can_attach_task,
9417 .attach_task = cpu_cgroup_attach_task,
9418 .exit = cpu_cgroup_exit,
9419 .populate = cpu_cgroup_populate,
9420 .subsys_id = cpu_cgroup_subsys_id,
9424 #endif /* CONFIG_CGROUP_SCHED */
9426 #ifdef CONFIG_CGROUP_CPUACCT
9429 * CPU accounting code for task groups.
9431 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9432 * (balbir@in.ibm.com).
9435 /* track cpu usage of a group of tasks and its child groups */
9437 struct cgroup_subsys_state css;
9438 /* cpuusage holds pointer to a u64-type object on every cpu */
9439 u64 __percpu *cpuusage;
9440 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9441 struct cpuacct *parent;
9444 struct cgroup_subsys cpuacct_subsys;
9446 /* return cpu accounting group corresponding to this container */
9447 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9449 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9450 struct cpuacct, css);
9453 /* return cpu accounting group to which this task belongs */
9454 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9456 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9457 struct cpuacct, css);
9460 /* create a new cpu accounting group */
9461 static struct cgroup_subsys_state *cpuacct_create(
9462 struct cgroup_subsys *ss, struct cgroup *cgrp)
9464 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9470 ca->cpuusage = alloc_percpu(u64);
9474 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9475 if (percpu_counter_init(&ca->cpustat[i], 0))
9476 goto out_free_counters;
9479 ca->parent = cgroup_ca(cgrp->parent);
9485 percpu_counter_destroy(&ca->cpustat[i]);
9486 free_percpu(ca->cpuusage);
9490 return ERR_PTR(-ENOMEM);
9493 /* destroy an existing cpu accounting group */
9495 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9497 struct cpuacct *ca = cgroup_ca(cgrp);
9500 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9501 percpu_counter_destroy(&ca->cpustat[i]);
9502 free_percpu(ca->cpuusage);
9506 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9508 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9511 #ifndef CONFIG_64BIT
9513 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9515 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9517 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9525 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9527 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9529 #ifndef CONFIG_64BIT
9531 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9533 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9535 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9541 /* return total cpu usage (in nanoseconds) of a group */
9542 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9544 struct cpuacct *ca = cgroup_ca(cgrp);
9545 u64 totalcpuusage = 0;
9548 for_each_present_cpu(i)
9549 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9551 return totalcpuusage;
9554 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9557 struct cpuacct *ca = cgroup_ca(cgrp);
9566 for_each_present_cpu(i)
9567 cpuacct_cpuusage_write(ca, i, 0);
9573 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9576 struct cpuacct *ca = cgroup_ca(cgroup);
9580 for_each_present_cpu(i) {
9581 percpu = cpuacct_cpuusage_read(ca, i);
9582 seq_printf(m, "%llu ", (unsigned long long) percpu);
9584 seq_printf(m, "\n");
9588 static const char *cpuacct_stat_desc[] = {
9589 [CPUACCT_STAT_USER] = "user",
9590 [CPUACCT_STAT_SYSTEM] = "system",
9593 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9594 struct cgroup_map_cb *cb)
9596 struct cpuacct *ca = cgroup_ca(cgrp);
9599 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9600 s64 val = percpu_counter_read(&ca->cpustat[i]);
9601 val = cputime64_to_clock_t(val);
9602 cb->fill(cb, cpuacct_stat_desc[i], val);
9607 static struct cftype files[] = {
9610 .read_u64 = cpuusage_read,
9611 .write_u64 = cpuusage_write,
9614 .name = "usage_percpu",
9615 .read_seq_string = cpuacct_percpu_seq_read,
9619 .read_map = cpuacct_stats_show,
9623 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9625 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9629 * charge this task's execution time to its accounting group.
9631 * called with rq->lock held.
9633 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9638 if (unlikely(!cpuacct_subsys.active))
9641 cpu = task_cpu(tsk);
9647 for (; ca; ca = ca->parent) {
9648 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9649 *cpuusage += cputime;
9656 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9657 * in cputime_t units. As a result, cpuacct_update_stats calls
9658 * percpu_counter_add with values large enough to always overflow the
9659 * per cpu batch limit causing bad SMP scalability.
9661 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9662 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9663 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9666 #define CPUACCT_BATCH \
9667 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9669 #define CPUACCT_BATCH 0
9673 * Charge the system/user time to the task's accounting group.
9675 static void cpuacct_update_stats(struct task_struct *tsk,
9676 enum cpuacct_stat_index idx, cputime_t val)
9679 int batch = CPUACCT_BATCH;
9681 if (unlikely(!cpuacct_subsys.active))
9688 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9694 struct cgroup_subsys cpuacct_subsys = {
9696 .create = cpuacct_create,
9697 .destroy = cpuacct_destroy,
9698 .populate = cpuacct_populate,
9699 .subsys_id = cpuacct_subsys_id,
9701 #endif /* CONFIG_CGROUP_CPUACCT */