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>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
81 #include "sched_autogroup.h"
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/sched.h>
87 * Convert user-nice values [ -20 ... 0 ... 19 ]
88 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
91 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
92 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
93 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
96 * 'User priority' is the nice value converted to something we
97 * can work with better when scaling various scheduler parameters,
98 * it's a [ 0 ... 39 ] range.
100 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
101 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
102 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
105 * Helpers for converting nanosecond timing to jiffy resolution
107 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
109 #define NICE_0_LOAD SCHED_LOAD_SCALE
110 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
113 * These are the 'tuning knobs' of the scheduler:
115 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
116 * Timeslices get refilled after they expire.
118 #define DEF_TIMESLICE (100 * HZ / 1000)
121 * single value that denotes runtime == period, ie unlimited time.
123 #define RUNTIME_INF ((u64)~0ULL)
125 static inline int rt_policy(int policy)
127 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
132 static inline int task_has_rt_policy(struct task_struct *p)
134 return rt_policy(p->policy);
138 * This is the priority-queue data structure of the RT scheduling class:
140 struct rt_prio_array {
141 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
142 struct list_head queue[MAX_RT_PRIO];
145 struct rt_bandwidth {
146 /* nests inside the rq lock: */
147 raw_spinlock_t rt_runtime_lock;
150 struct hrtimer rt_period_timer;
153 static struct rt_bandwidth def_rt_bandwidth;
155 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
157 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
159 struct rt_bandwidth *rt_b =
160 container_of(timer, struct rt_bandwidth, rt_period_timer);
166 now = hrtimer_cb_get_time(timer);
167 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
172 idle = do_sched_rt_period_timer(rt_b, overrun);
175 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
179 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
181 rt_b->rt_period = ns_to_ktime(period);
182 rt_b->rt_runtime = runtime;
184 raw_spin_lock_init(&rt_b->rt_runtime_lock);
186 hrtimer_init(&rt_b->rt_period_timer,
187 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
188 rt_b->rt_period_timer.function = sched_rt_period_timer;
191 static inline int rt_bandwidth_enabled(void)
193 return sysctl_sched_rt_runtime >= 0;
196 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
200 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
203 if (hrtimer_active(&rt_b->rt_period_timer))
206 raw_spin_lock(&rt_b->rt_runtime_lock);
211 if (hrtimer_active(&rt_b->rt_period_timer))
214 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
215 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
217 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
218 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
219 delta = ktime_to_ns(ktime_sub(hard, soft));
220 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
221 HRTIMER_MODE_ABS_PINNED, 0);
223 raw_spin_unlock(&rt_b->rt_runtime_lock);
226 #ifdef CONFIG_RT_GROUP_SCHED
227 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
229 hrtimer_cancel(&rt_b->rt_period_timer);
234 * sched_domains_mutex serializes calls to arch_init_sched_domains,
235 * detach_destroy_domains and partition_sched_domains.
237 static DEFINE_MUTEX(sched_domains_mutex);
239 #ifdef CONFIG_CGROUP_SCHED
241 #include <linux/cgroup.h>
245 static LIST_HEAD(task_groups);
247 /* task group related information */
249 struct cgroup_subsys_state css;
251 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* schedulable entities of this group on each cpu */
253 struct sched_entity **se;
254 /* runqueue "owned" by this group on each cpu */
255 struct cfs_rq **cfs_rq;
256 unsigned long shares;
258 atomic_t load_weight;
261 #ifdef CONFIG_RT_GROUP_SCHED
262 struct sched_rt_entity **rt_se;
263 struct rt_rq **rt_rq;
265 struct rt_bandwidth rt_bandwidth;
269 struct list_head list;
271 struct task_group *parent;
272 struct list_head siblings;
273 struct list_head children;
275 #ifdef CONFIG_SCHED_AUTOGROUP
276 struct autogroup *autogroup;
280 /* task_group_lock serializes the addition/removal of task groups */
281 static DEFINE_SPINLOCK(task_group_lock);
283 #ifdef CONFIG_FAIR_GROUP_SCHED
285 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
288 * A weight of 0 or 1 can cause arithmetics problems.
289 * A weight of a cfs_rq is the sum of weights of which entities
290 * are queued on this cfs_rq, so a weight of a entity should not be
291 * too large, so as the shares value of a task group.
292 * (The default weight is 1024 - so there's no practical
293 * limitation from this.)
296 #define MAX_SHARES (1UL << 18)
298 static int root_task_group_load = ROOT_TASK_GROUP_LOAD;
301 /* Default task group.
302 * Every task in system belong to this group at bootup.
304 struct task_group root_task_group;
306 #endif /* CONFIG_CGROUP_SCHED */
308 /* CFS-related fields in a runqueue */
310 struct load_weight load;
311 unsigned long nr_running;
316 u64 min_vruntime_copy;
319 struct rb_root tasks_timeline;
320 struct rb_node *rb_leftmost;
322 struct list_head tasks;
323 struct list_head *balance_iterator;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity *curr, *next, *last, *skip;
331 unsigned int nr_spread_over;
333 #ifdef CONFIG_FAIR_GROUP_SCHED
334 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
337 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
338 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
339 * (like users, containers etc.)
341 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
342 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list;
346 struct task_group *tg; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load;
363 * Maintaining per-cpu shares distribution for group scheduling
365 * load_stamp is the last time we updated the load average
366 * load_last is the last time we updated the load average and saw load
367 * load_unacc_exec_time is currently unaccounted execution time
371 u64 load_stamp, load_last, load_unacc_exec_time;
373 unsigned long load_contribution;
378 /* Real-Time classes' related field in a runqueue: */
380 struct rt_prio_array active;
381 unsigned long rt_nr_running;
382 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
384 int curr; /* highest queued rt task prio */
386 int next; /* next highest */
391 unsigned long rt_nr_migratory;
392 unsigned long rt_nr_total;
394 struct plist_head pushable_tasks;
399 /* Nests inside the rq lock: */
400 raw_spinlock_t rt_runtime_lock;
402 #ifdef CONFIG_RT_GROUP_SCHED
403 unsigned long rt_nr_boosted;
406 struct list_head leaf_rt_rq_list;
407 struct task_group *tg;
414 * We add the notion of a root-domain which will be used to define per-domain
415 * variables. Each exclusive cpuset essentially defines an island domain by
416 * fully partitioning the member cpus from any other cpuset. Whenever a new
417 * exclusive cpuset is created, we also create and attach a new root-domain
424 cpumask_var_t online;
427 * The "RT overload" flag: it gets set if a CPU has more than
428 * one runnable RT task.
430 cpumask_var_t rto_mask;
432 struct cpupri cpupri;
436 * By default the system creates a single root-domain with all cpus as
437 * members (mimicking the global state we have today).
439 static struct root_domain def_root_domain;
441 #endif /* CONFIG_SMP */
444 * This is the main, per-CPU runqueue data structure.
446 * Locking rule: those places that want to lock multiple runqueues
447 * (such as the load balancing or the thread migration code), lock
448 * acquire operations must be ordered by ascending &runqueue.
455 * nr_running and cpu_load should be in the same cacheline because
456 * remote CPUs use both these fields when doing load calculation.
458 unsigned long nr_running;
459 #define CPU_LOAD_IDX_MAX 5
460 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
461 unsigned long last_load_update_tick;
464 unsigned char nohz_balance_kick;
466 unsigned int skip_clock_update;
468 /* capture load from *all* tasks on this cpu: */
469 struct load_weight load;
470 unsigned long nr_load_updates;
476 #ifdef CONFIG_FAIR_GROUP_SCHED
477 /* list of leaf cfs_rq on this cpu: */
478 struct list_head leaf_cfs_rq_list;
480 #ifdef CONFIG_RT_GROUP_SCHED
481 struct list_head leaf_rt_rq_list;
485 * This is part of a global counter where only the total sum
486 * over all CPUs matters. A task can increase this counter on
487 * one CPU and if it got migrated afterwards it may decrease
488 * it on another CPU. Always updated under the runqueue lock:
490 unsigned long nr_uninterruptible;
492 struct task_struct *curr, *idle, *stop;
493 unsigned long next_balance;
494 struct mm_struct *prev_mm;
502 struct root_domain *rd;
503 struct sched_domain *sd;
505 unsigned long cpu_power;
507 unsigned char idle_at_tick;
508 /* For active balancing */
512 struct cpu_stop_work active_balance_work;
513 /* cpu of this runqueue: */
517 unsigned long avg_load_per_task;
525 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
529 /* calc_load related fields */
530 unsigned long calc_load_update;
531 long calc_load_active;
533 #ifdef CONFIG_SCHED_HRTICK
535 int hrtick_csd_pending;
536 struct call_single_data hrtick_csd;
538 struct hrtimer hrtick_timer;
541 #ifdef CONFIG_SCHEDSTATS
543 struct sched_info rq_sched_info;
544 unsigned long long rq_cpu_time;
545 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
547 /* sys_sched_yield() stats */
548 unsigned int yld_count;
550 /* schedule() stats */
551 unsigned int sched_switch;
552 unsigned int sched_count;
553 unsigned int sched_goidle;
555 /* try_to_wake_up() stats */
556 unsigned int ttwu_count;
557 unsigned int ttwu_local;
561 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
564 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags);
566 static inline int cpu_of(struct rq *rq)
575 #define rcu_dereference_check_sched_domain(p) \
576 rcu_dereference_check((p), \
577 rcu_read_lock_sched_held() || \
578 lockdep_is_held(&sched_domains_mutex))
581 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
582 * See detach_destroy_domains: synchronize_sched for details.
584 * The domain tree of any CPU may only be accessed from within
585 * preempt-disabled sections.
587 #define for_each_domain(cpu, __sd) \
588 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
590 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
591 #define this_rq() (&__get_cpu_var(runqueues))
592 #define task_rq(p) cpu_rq(task_cpu(p))
593 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
594 #define raw_rq() (&__raw_get_cpu_var(runqueues))
596 #ifdef CONFIG_CGROUP_SCHED
599 * Return the group to which this tasks belongs.
601 * We use task_subsys_state_check() and extend the RCU verification
602 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
603 * holds that lock for each task it moves into the cgroup. Therefore
604 * by holding that lock, we pin the task to the current cgroup.
606 static inline struct task_group *task_group(struct task_struct *p)
608 struct task_group *tg;
609 struct cgroup_subsys_state *css;
611 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
612 lockdep_is_held(&task_rq(p)->lock));
613 tg = container_of(css, struct task_group, css);
615 return autogroup_task_group(p, tg);
618 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
619 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
621 #ifdef CONFIG_FAIR_GROUP_SCHED
622 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
623 p->se.parent = task_group(p)->se[cpu];
626 #ifdef CONFIG_RT_GROUP_SCHED
627 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
628 p->rt.parent = task_group(p)->rt_se[cpu];
632 #else /* CONFIG_CGROUP_SCHED */
634 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
635 static inline struct task_group *task_group(struct task_struct *p)
640 #endif /* CONFIG_CGROUP_SCHED */
642 static void update_rq_clock_task(struct rq *rq, s64 delta);
644 static void update_rq_clock(struct rq *rq)
648 if (rq->skip_clock_update)
651 delta = sched_clock_cpu(cpu_of(rq)) - rq->clock;
653 update_rq_clock_task(rq, delta);
657 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
659 #ifdef CONFIG_SCHED_DEBUG
660 # define const_debug __read_mostly
662 # define const_debug static const
666 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
667 * @cpu: the processor in question.
669 * This interface allows printk to be called with the runqueue lock
670 * held and know whether or not it is OK to wake up the klogd.
672 int runqueue_is_locked(int cpu)
674 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
678 * Debugging: various feature bits
681 #define SCHED_FEAT(name, enabled) \
682 __SCHED_FEAT_##name ,
685 #include "sched_features.h"
690 #define SCHED_FEAT(name, enabled) \
691 (1UL << __SCHED_FEAT_##name) * enabled |
693 const_debug unsigned int sysctl_sched_features =
694 #include "sched_features.h"
699 #ifdef CONFIG_SCHED_DEBUG
700 #define SCHED_FEAT(name, enabled) \
703 static __read_mostly char *sched_feat_names[] = {
704 #include "sched_features.h"
710 static int sched_feat_show(struct seq_file *m, void *v)
714 for (i = 0; sched_feat_names[i]; i++) {
715 if (!(sysctl_sched_features & (1UL << i)))
717 seq_printf(m, "%s ", sched_feat_names[i]);
725 sched_feat_write(struct file *filp, const char __user *ubuf,
726 size_t cnt, loff_t *ppos)
736 if (copy_from_user(&buf, ubuf, cnt))
742 if (strncmp(cmp, "NO_", 3) == 0) {
747 for (i = 0; sched_feat_names[i]; i++) {
748 if (strcmp(cmp, sched_feat_names[i]) == 0) {
750 sysctl_sched_features &= ~(1UL << i);
752 sysctl_sched_features |= (1UL << i);
757 if (!sched_feat_names[i])
765 static int sched_feat_open(struct inode *inode, struct file *filp)
767 return single_open(filp, sched_feat_show, NULL);
770 static const struct file_operations sched_feat_fops = {
771 .open = sched_feat_open,
772 .write = sched_feat_write,
775 .release = single_release,
778 static __init int sched_init_debug(void)
780 debugfs_create_file("sched_features", 0644, NULL, NULL,
785 late_initcall(sched_init_debug);
789 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
792 * Number of tasks to iterate in a single balance run.
793 * Limited because this is done with IRQs disabled.
795 const_debug unsigned int sysctl_sched_nr_migrate = 32;
798 * period over which we average the RT time consumption, measured
803 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
806 * period over which we measure -rt task cpu usage in us.
809 unsigned int sysctl_sched_rt_period = 1000000;
811 static __read_mostly int scheduler_running;
814 * part of the period that we allow rt tasks to run in us.
817 int sysctl_sched_rt_runtime = 950000;
819 static inline u64 global_rt_period(void)
821 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
824 static inline u64 global_rt_runtime(void)
826 if (sysctl_sched_rt_runtime < 0)
829 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
832 #ifndef prepare_arch_switch
833 # define prepare_arch_switch(next) do { } while (0)
835 #ifndef finish_arch_switch
836 # define finish_arch_switch(prev) do { } while (0)
839 static inline int task_current(struct rq *rq, struct task_struct *p)
841 return rq->curr == p;
844 static inline int task_running(struct rq *rq, struct task_struct *p)
849 return task_current(rq, p);
853 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
854 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
858 * We can optimise this out completely for !SMP, because the
859 * SMP rebalancing from interrupt is the only thing that cares
866 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
870 * After ->on_cpu is cleared, the task can be moved to a different CPU.
871 * We must ensure this doesn't happen until the switch is completely
877 #ifdef CONFIG_DEBUG_SPINLOCK
878 /* this is a valid case when another task releases the spinlock */
879 rq->lock.owner = current;
882 * If we are tracking spinlock dependencies then we have to
883 * fix up the runqueue lock - which gets 'carried over' from
886 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
888 raw_spin_unlock_irq(&rq->lock);
891 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
892 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
896 * We can optimise this out completely for !SMP, because the
897 * SMP rebalancing from interrupt is the only thing that cares
902 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
903 raw_spin_unlock_irq(&rq->lock);
905 raw_spin_unlock(&rq->lock);
909 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
913 * After ->on_cpu is cleared, the task can be moved to a different CPU.
914 * We must ensure this doesn't happen until the switch is completely
920 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
924 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
927 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
930 static inline int task_is_waking(struct task_struct *p)
932 return unlikely(p->state == TASK_WAKING);
936 * __task_rq_lock - lock the runqueue a given task resides on.
937 * Must be called interrupts disabled.
939 static inline struct rq *__task_rq_lock(struct task_struct *p)
946 raw_spin_lock(&rq->lock);
947 if (likely(rq == task_rq(p)))
949 raw_spin_unlock(&rq->lock);
954 * task_rq_lock - lock the runqueue a given task resides on and disable
955 * interrupts. Note the ordering: we can safely lookup the task_rq without
956 * explicitly disabling preemption.
958 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
964 local_irq_save(*flags);
966 raw_spin_lock(&rq->lock);
967 if (likely(rq == task_rq(p)))
969 raw_spin_unlock_irqrestore(&rq->lock, *flags);
973 static void __task_rq_unlock(struct rq *rq)
976 raw_spin_unlock(&rq->lock);
979 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
982 raw_spin_unlock_irqrestore(&rq->lock, *flags);
986 * this_rq_lock - lock this runqueue and disable interrupts.
988 static struct rq *this_rq_lock(void)
995 raw_spin_lock(&rq->lock);
1000 #ifdef CONFIG_SCHED_HRTICK
1002 * Use HR-timers to deliver accurate preemption points.
1004 * Its all a bit involved since we cannot program an hrt while holding the
1005 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1008 * When we get rescheduled we reprogram the hrtick_timer outside of the
1014 * - enabled by features
1015 * - hrtimer is actually high res
1017 static inline int hrtick_enabled(struct rq *rq)
1019 if (!sched_feat(HRTICK))
1021 if (!cpu_active(cpu_of(rq)))
1023 return hrtimer_is_hres_active(&rq->hrtick_timer);
1026 static void hrtick_clear(struct rq *rq)
1028 if (hrtimer_active(&rq->hrtick_timer))
1029 hrtimer_cancel(&rq->hrtick_timer);
1033 * High-resolution timer tick.
1034 * Runs from hardirq context with interrupts disabled.
1036 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1038 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1040 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1042 raw_spin_lock(&rq->lock);
1043 update_rq_clock(rq);
1044 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1045 raw_spin_unlock(&rq->lock);
1047 return HRTIMER_NORESTART;
1052 * called from hardirq (IPI) context
1054 static void __hrtick_start(void *arg)
1056 struct rq *rq = arg;
1058 raw_spin_lock(&rq->lock);
1059 hrtimer_restart(&rq->hrtick_timer);
1060 rq->hrtick_csd_pending = 0;
1061 raw_spin_unlock(&rq->lock);
1065 * Called to set the hrtick timer state.
1067 * called with rq->lock held and irqs disabled
1069 static void hrtick_start(struct rq *rq, u64 delay)
1071 struct hrtimer *timer = &rq->hrtick_timer;
1072 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1074 hrtimer_set_expires(timer, time);
1076 if (rq == this_rq()) {
1077 hrtimer_restart(timer);
1078 } else if (!rq->hrtick_csd_pending) {
1079 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1080 rq->hrtick_csd_pending = 1;
1085 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1087 int cpu = (int)(long)hcpu;
1090 case CPU_UP_CANCELED:
1091 case CPU_UP_CANCELED_FROZEN:
1092 case CPU_DOWN_PREPARE:
1093 case CPU_DOWN_PREPARE_FROZEN:
1095 case CPU_DEAD_FROZEN:
1096 hrtick_clear(cpu_rq(cpu));
1103 static __init void init_hrtick(void)
1105 hotcpu_notifier(hotplug_hrtick, 0);
1109 * Called to set the hrtick timer state.
1111 * called with rq->lock held and irqs disabled
1113 static void hrtick_start(struct rq *rq, u64 delay)
1115 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1116 HRTIMER_MODE_REL_PINNED, 0);
1119 static inline void init_hrtick(void)
1122 #endif /* CONFIG_SMP */
1124 static void init_rq_hrtick(struct rq *rq)
1127 rq->hrtick_csd_pending = 0;
1129 rq->hrtick_csd.flags = 0;
1130 rq->hrtick_csd.func = __hrtick_start;
1131 rq->hrtick_csd.info = rq;
1134 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1135 rq->hrtick_timer.function = hrtick;
1137 #else /* CONFIG_SCHED_HRTICK */
1138 static inline void hrtick_clear(struct rq *rq)
1142 static inline void init_rq_hrtick(struct rq *rq)
1146 static inline void init_hrtick(void)
1149 #endif /* CONFIG_SCHED_HRTICK */
1152 * resched_task - mark a task 'to be rescheduled now'.
1154 * On UP this means the setting of the need_resched flag, on SMP it
1155 * might also involve a cross-CPU call to trigger the scheduler on
1160 #ifndef tsk_is_polling
1161 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1164 static void resched_task(struct task_struct *p)
1168 assert_raw_spin_locked(&task_rq(p)->lock);
1170 if (test_tsk_need_resched(p))
1173 set_tsk_need_resched(p);
1176 if (cpu == smp_processor_id())
1179 /* NEED_RESCHED must be visible before we test polling */
1181 if (!tsk_is_polling(p))
1182 smp_send_reschedule(cpu);
1185 static void resched_cpu(int cpu)
1187 struct rq *rq = cpu_rq(cpu);
1188 unsigned long flags;
1190 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1192 resched_task(cpu_curr(cpu));
1193 raw_spin_unlock_irqrestore(&rq->lock, flags);
1198 * In the semi idle case, use the nearest busy cpu for migrating timers
1199 * from an idle cpu. This is good for power-savings.
1201 * We don't do similar optimization for completely idle system, as
1202 * selecting an idle cpu will add more delays to the timers than intended
1203 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1205 int get_nohz_timer_target(void)
1207 int cpu = smp_processor_id();
1209 struct sched_domain *sd;
1211 for_each_domain(cpu, sd) {
1212 for_each_cpu(i, sched_domain_span(sd))
1219 * When add_timer_on() enqueues a timer into the timer wheel of an
1220 * idle CPU then this timer might expire before the next timer event
1221 * which is scheduled to wake up that CPU. In case of a completely
1222 * idle system the next event might even be infinite time into the
1223 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1224 * leaves the inner idle loop so the newly added timer is taken into
1225 * account when the CPU goes back to idle and evaluates the timer
1226 * wheel for the next timer event.
1228 void wake_up_idle_cpu(int cpu)
1230 struct rq *rq = cpu_rq(cpu);
1232 if (cpu == smp_processor_id())
1236 * This is safe, as this function is called with the timer
1237 * wheel base lock of (cpu) held. When the CPU is on the way
1238 * to idle and has not yet set rq->curr to idle then it will
1239 * be serialized on the timer wheel base lock and take the new
1240 * timer into account automatically.
1242 if (rq->curr != rq->idle)
1246 * We can set TIF_RESCHED on the idle task of the other CPU
1247 * lockless. The worst case is that the other CPU runs the
1248 * idle task through an additional NOOP schedule()
1250 set_tsk_need_resched(rq->idle);
1252 /* NEED_RESCHED must be visible before we test polling */
1254 if (!tsk_is_polling(rq->idle))
1255 smp_send_reschedule(cpu);
1258 #endif /* CONFIG_NO_HZ */
1260 static u64 sched_avg_period(void)
1262 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1265 static void sched_avg_update(struct rq *rq)
1267 s64 period = sched_avg_period();
1269 while ((s64)(rq->clock - rq->age_stamp) > period) {
1271 * Inline assembly required to prevent the compiler
1272 * optimising this loop into a divmod call.
1273 * See __iter_div_u64_rem() for another example of this.
1275 asm("" : "+rm" (rq->age_stamp));
1276 rq->age_stamp += period;
1281 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1283 rq->rt_avg += rt_delta;
1284 sched_avg_update(rq);
1287 #else /* !CONFIG_SMP */
1288 static void resched_task(struct task_struct *p)
1290 assert_raw_spin_locked(&task_rq(p)->lock);
1291 set_tsk_need_resched(p);
1294 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1298 static void sched_avg_update(struct rq *rq)
1301 #endif /* CONFIG_SMP */
1303 #if BITS_PER_LONG == 32
1304 # define WMULT_CONST (~0UL)
1306 # define WMULT_CONST (1UL << 32)
1309 #define WMULT_SHIFT 32
1312 * Shift right and round:
1314 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1317 * delta *= weight / lw
1319 static unsigned long
1320 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1321 struct load_weight *lw)
1325 if (!lw->inv_weight) {
1326 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1329 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1333 tmp = (u64)delta_exec * weight;
1335 * Check whether we'd overflow the 64-bit multiplication:
1337 if (unlikely(tmp > WMULT_CONST))
1338 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1341 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1343 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1346 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1352 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1358 static inline void update_load_set(struct load_weight *lw, unsigned long w)
1365 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1366 * of tasks with abnormal "nice" values across CPUs the contribution that
1367 * each task makes to its run queue's load is weighted according to its
1368 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1369 * scaled version of the new time slice allocation that they receive on time
1373 #define WEIGHT_IDLEPRIO 3
1374 #define WMULT_IDLEPRIO 1431655765
1377 * Nice levels are multiplicative, with a gentle 10% change for every
1378 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1379 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1380 * that remained on nice 0.
1382 * The "10% effect" is relative and cumulative: from _any_ nice level,
1383 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1384 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1385 * If a task goes up by ~10% and another task goes down by ~10% then
1386 * the relative distance between them is ~25%.)
1388 static const int prio_to_weight[40] = {
1389 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1390 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1391 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1392 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1393 /* 0 */ 1024, 820, 655, 526, 423,
1394 /* 5 */ 335, 272, 215, 172, 137,
1395 /* 10 */ 110, 87, 70, 56, 45,
1396 /* 15 */ 36, 29, 23, 18, 15,
1400 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1402 * In cases where the weight does not change often, we can use the
1403 * precalculated inverse to speed up arithmetics by turning divisions
1404 * into multiplications:
1406 static const u32 prio_to_wmult[40] = {
1407 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1408 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1409 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1410 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1411 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1412 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1413 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1414 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1417 /* Time spent by the tasks of the cpu accounting group executing in ... */
1418 enum cpuacct_stat_index {
1419 CPUACCT_STAT_USER, /* ... user mode */
1420 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1422 CPUACCT_STAT_NSTATS,
1425 #ifdef CONFIG_CGROUP_CPUACCT
1426 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1427 static void cpuacct_update_stats(struct task_struct *tsk,
1428 enum cpuacct_stat_index idx, cputime_t val);
1430 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1431 static inline void cpuacct_update_stats(struct task_struct *tsk,
1432 enum cpuacct_stat_index idx, cputime_t val) {}
1435 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1437 update_load_add(&rq->load, load);
1440 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1442 update_load_sub(&rq->load, load);
1445 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1446 typedef int (*tg_visitor)(struct task_group *, void *);
1449 * Iterate the full tree, calling @down when first entering a node and @up when
1450 * leaving it for the final time.
1452 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1454 struct task_group *parent, *child;
1458 parent = &root_task_group;
1460 ret = (*down)(parent, data);
1463 list_for_each_entry_rcu(child, &parent->children, siblings) {
1470 ret = (*up)(parent, data);
1475 parent = parent->parent;
1484 static int tg_nop(struct task_group *tg, void *data)
1491 /* Used instead of source_load when we know the type == 0 */
1492 static unsigned long weighted_cpuload(const int cpu)
1494 return cpu_rq(cpu)->load.weight;
1498 * Return a low guess at the load of a migration-source cpu weighted
1499 * according to the scheduling class and "nice" value.
1501 * We want to under-estimate the load of migration sources, to
1502 * balance conservatively.
1504 static unsigned long source_load(int cpu, int type)
1506 struct rq *rq = cpu_rq(cpu);
1507 unsigned long total = weighted_cpuload(cpu);
1509 if (type == 0 || !sched_feat(LB_BIAS))
1512 return min(rq->cpu_load[type-1], total);
1516 * Return a high guess at the load of a migration-target cpu weighted
1517 * according to the scheduling class and "nice" value.
1519 static unsigned long target_load(int cpu, int type)
1521 struct rq *rq = cpu_rq(cpu);
1522 unsigned long total = weighted_cpuload(cpu);
1524 if (type == 0 || !sched_feat(LB_BIAS))
1527 return max(rq->cpu_load[type-1], total);
1530 static unsigned long power_of(int cpu)
1532 return cpu_rq(cpu)->cpu_power;
1535 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1537 static unsigned long cpu_avg_load_per_task(int cpu)
1539 struct rq *rq = cpu_rq(cpu);
1540 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1543 rq->avg_load_per_task = rq->load.weight / nr_running;
1545 rq->avg_load_per_task = 0;
1547 return rq->avg_load_per_task;
1550 #ifdef CONFIG_FAIR_GROUP_SCHED
1553 * Compute the cpu's hierarchical load factor for each task group.
1554 * This needs to be done in a top-down fashion because the load of a child
1555 * group is a fraction of its parents load.
1557 static int tg_load_down(struct task_group *tg, void *data)
1560 long cpu = (long)data;
1563 load = cpu_rq(cpu)->load.weight;
1565 load = tg->parent->cfs_rq[cpu]->h_load;
1566 load *= tg->se[cpu]->load.weight;
1567 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1570 tg->cfs_rq[cpu]->h_load = load;
1575 static void update_h_load(long cpu)
1577 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1582 #ifdef CONFIG_PREEMPT
1584 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1587 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1588 * way at the expense of forcing extra atomic operations in all
1589 * invocations. This assures that the double_lock is acquired using the
1590 * same underlying policy as the spinlock_t on this architecture, which
1591 * reduces latency compared to the unfair variant below. However, it
1592 * also adds more overhead and therefore may reduce throughput.
1594 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1595 __releases(this_rq->lock)
1596 __acquires(busiest->lock)
1597 __acquires(this_rq->lock)
1599 raw_spin_unlock(&this_rq->lock);
1600 double_rq_lock(this_rq, busiest);
1607 * Unfair double_lock_balance: Optimizes throughput at the expense of
1608 * latency by eliminating extra atomic operations when the locks are
1609 * already in proper order on entry. This favors lower cpu-ids and will
1610 * grant the double lock to lower cpus over higher ids under contention,
1611 * regardless of entry order into the function.
1613 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1614 __releases(this_rq->lock)
1615 __acquires(busiest->lock)
1616 __acquires(this_rq->lock)
1620 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1621 if (busiest < this_rq) {
1622 raw_spin_unlock(&this_rq->lock);
1623 raw_spin_lock(&busiest->lock);
1624 raw_spin_lock_nested(&this_rq->lock,
1625 SINGLE_DEPTH_NESTING);
1628 raw_spin_lock_nested(&busiest->lock,
1629 SINGLE_DEPTH_NESTING);
1634 #endif /* CONFIG_PREEMPT */
1637 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1639 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1641 if (unlikely(!irqs_disabled())) {
1642 /* printk() doesn't work good under rq->lock */
1643 raw_spin_unlock(&this_rq->lock);
1647 return _double_lock_balance(this_rq, busiest);
1650 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1651 __releases(busiest->lock)
1653 raw_spin_unlock(&busiest->lock);
1654 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1658 * double_rq_lock - safely lock two runqueues
1660 * Note this does not disable interrupts like task_rq_lock,
1661 * you need to do so manually before calling.
1663 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1664 __acquires(rq1->lock)
1665 __acquires(rq2->lock)
1667 BUG_ON(!irqs_disabled());
1669 raw_spin_lock(&rq1->lock);
1670 __acquire(rq2->lock); /* Fake it out ;) */
1673 raw_spin_lock(&rq1->lock);
1674 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1676 raw_spin_lock(&rq2->lock);
1677 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1683 * double_rq_unlock - safely unlock two runqueues
1685 * Note this does not restore interrupts like task_rq_unlock,
1686 * you need to do so manually after calling.
1688 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1689 __releases(rq1->lock)
1690 __releases(rq2->lock)
1692 raw_spin_unlock(&rq1->lock);
1694 raw_spin_unlock(&rq2->lock);
1696 __release(rq2->lock);
1699 #else /* CONFIG_SMP */
1702 * double_rq_lock - safely lock two runqueues
1704 * Note this does not disable interrupts like task_rq_lock,
1705 * you need to do so manually before calling.
1707 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1708 __acquires(rq1->lock)
1709 __acquires(rq2->lock)
1711 BUG_ON(!irqs_disabled());
1713 raw_spin_lock(&rq1->lock);
1714 __acquire(rq2->lock); /* Fake it out ;) */
1718 * double_rq_unlock - safely unlock two runqueues
1720 * Note this does not restore interrupts like task_rq_unlock,
1721 * you need to do so manually after calling.
1723 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1724 __releases(rq1->lock)
1725 __releases(rq2->lock)
1728 raw_spin_unlock(&rq1->lock);
1729 __release(rq2->lock);
1734 static void calc_load_account_idle(struct rq *this_rq);
1735 static void update_sysctl(void);
1736 static int get_update_sysctl_factor(void);
1737 static void update_cpu_load(struct rq *this_rq);
1739 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1741 set_task_rq(p, cpu);
1744 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1745 * successfuly executed on another CPU. We must ensure that updates of
1746 * per-task data have been completed by this moment.
1749 task_thread_info(p)->cpu = cpu;
1753 static const struct sched_class rt_sched_class;
1755 #define sched_class_highest (&stop_sched_class)
1756 #define for_each_class(class) \
1757 for (class = sched_class_highest; class; class = class->next)
1759 #include "sched_stats.h"
1761 static void inc_nr_running(struct rq *rq)
1766 static void dec_nr_running(struct rq *rq)
1771 static void set_load_weight(struct task_struct *p)
1774 * SCHED_IDLE tasks get minimal weight:
1776 if (p->policy == SCHED_IDLE) {
1777 p->se.load.weight = WEIGHT_IDLEPRIO;
1778 p->se.load.inv_weight = WMULT_IDLEPRIO;
1782 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1783 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1786 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1788 update_rq_clock(rq);
1789 sched_info_queued(p);
1790 p->sched_class->enqueue_task(rq, p, flags);
1793 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1795 update_rq_clock(rq);
1796 sched_info_dequeued(p);
1797 p->sched_class->dequeue_task(rq, p, flags);
1801 * activate_task - move a task to the runqueue.
1803 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1805 if (task_contributes_to_load(p))
1806 rq->nr_uninterruptible--;
1808 enqueue_task(rq, p, flags);
1813 * deactivate_task - remove a task from the runqueue.
1815 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1817 if (task_contributes_to_load(p))
1818 rq->nr_uninterruptible++;
1820 dequeue_task(rq, p, flags);
1824 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1827 * There are no locks covering percpu hardirq/softirq time.
1828 * They are only modified in account_system_vtime, on corresponding CPU
1829 * with interrupts disabled. So, writes are safe.
1830 * They are read and saved off onto struct rq in update_rq_clock().
1831 * This may result in other CPU reading this CPU's irq time and can
1832 * race with irq/account_system_vtime on this CPU. We would either get old
1833 * or new value with a side effect of accounting a slice of irq time to wrong
1834 * task when irq is in progress while we read rq->clock. That is a worthy
1835 * compromise in place of having locks on each irq in account_system_time.
1837 static DEFINE_PER_CPU(u64, cpu_hardirq_time);
1838 static DEFINE_PER_CPU(u64, cpu_softirq_time);
1840 static DEFINE_PER_CPU(u64, irq_start_time);
1841 static int sched_clock_irqtime;
1843 void enable_sched_clock_irqtime(void)
1845 sched_clock_irqtime = 1;
1848 void disable_sched_clock_irqtime(void)
1850 sched_clock_irqtime = 0;
1853 #ifndef CONFIG_64BIT
1854 static DEFINE_PER_CPU(seqcount_t, irq_time_seq);
1856 static inline void irq_time_write_begin(void)
1858 __this_cpu_inc(irq_time_seq.sequence);
1862 static inline void irq_time_write_end(void)
1865 __this_cpu_inc(irq_time_seq.sequence);
1868 static inline u64 irq_time_read(int cpu)
1874 seq = read_seqcount_begin(&per_cpu(irq_time_seq, cpu));
1875 irq_time = per_cpu(cpu_softirq_time, cpu) +
1876 per_cpu(cpu_hardirq_time, cpu);
1877 } while (read_seqcount_retry(&per_cpu(irq_time_seq, cpu), seq));
1881 #else /* CONFIG_64BIT */
1882 static inline void irq_time_write_begin(void)
1886 static inline void irq_time_write_end(void)
1890 static inline u64 irq_time_read(int cpu)
1892 return per_cpu(cpu_softirq_time, cpu) + per_cpu(cpu_hardirq_time, cpu);
1894 #endif /* CONFIG_64BIT */
1897 * Called before incrementing preempt_count on {soft,}irq_enter
1898 * and before decrementing preempt_count on {soft,}irq_exit.
1900 void account_system_vtime(struct task_struct *curr)
1902 unsigned long flags;
1906 if (!sched_clock_irqtime)
1909 local_irq_save(flags);
1911 cpu = smp_processor_id();
1912 delta = sched_clock_cpu(cpu) - __this_cpu_read(irq_start_time);
1913 __this_cpu_add(irq_start_time, delta);
1915 irq_time_write_begin();
1917 * We do not account for softirq time from ksoftirqd here.
1918 * We want to continue accounting softirq time to ksoftirqd thread
1919 * in that case, so as not to confuse scheduler with a special task
1920 * that do not consume any time, but still wants to run.
1922 if (hardirq_count())
1923 __this_cpu_add(cpu_hardirq_time, delta);
1924 else if (in_serving_softirq() && curr != this_cpu_ksoftirqd())
1925 __this_cpu_add(cpu_softirq_time, delta);
1927 irq_time_write_end();
1928 local_irq_restore(flags);
1930 EXPORT_SYMBOL_GPL(account_system_vtime);
1932 static void update_rq_clock_task(struct rq *rq, s64 delta)
1936 irq_delta = irq_time_read(cpu_of(rq)) - rq->prev_irq_time;
1939 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1940 * this case when a previous update_rq_clock() happened inside a
1941 * {soft,}irq region.
1943 * When this happens, we stop ->clock_task and only update the
1944 * prev_irq_time stamp to account for the part that fit, so that a next
1945 * update will consume the rest. This ensures ->clock_task is
1948 * It does however cause some slight miss-attribution of {soft,}irq
1949 * time, a more accurate solution would be to update the irq_time using
1950 * the current rq->clock timestamp, except that would require using
1953 if (irq_delta > delta)
1956 rq->prev_irq_time += irq_delta;
1958 rq->clock_task += delta;
1960 if (irq_delta && sched_feat(NONIRQ_POWER))
1961 sched_rt_avg_update(rq, irq_delta);
1964 static int irqtime_account_hi_update(void)
1966 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1967 unsigned long flags;
1971 local_irq_save(flags);
1972 latest_ns = this_cpu_read(cpu_hardirq_time);
1973 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->irq))
1975 local_irq_restore(flags);
1979 static int irqtime_account_si_update(void)
1981 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
1982 unsigned long flags;
1986 local_irq_save(flags);
1987 latest_ns = this_cpu_read(cpu_softirq_time);
1988 if (cputime64_gt(nsecs_to_cputime64(latest_ns), cpustat->softirq))
1990 local_irq_restore(flags);
1994 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
1996 #define sched_clock_irqtime (0)
1998 static void update_rq_clock_task(struct rq *rq, s64 delta)
2000 rq->clock_task += delta;
2003 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2005 #include "sched_idletask.c"
2006 #include "sched_fair.c"
2007 #include "sched_rt.c"
2008 #include "sched_autogroup.c"
2009 #include "sched_stoptask.c"
2010 #ifdef CONFIG_SCHED_DEBUG
2011 # include "sched_debug.c"
2014 void sched_set_stop_task(int cpu, struct task_struct *stop)
2016 struct sched_param param = { .sched_priority = MAX_RT_PRIO - 1 };
2017 struct task_struct *old_stop = cpu_rq(cpu)->stop;
2021 * Make it appear like a SCHED_FIFO task, its something
2022 * userspace knows about and won't get confused about.
2024 * Also, it will make PI more or less work without too
2025 * much confusion -- but then, stop work should not
2026 * rely on PI working anyway.
2028 sched_setscheduler_nocheck(stop, SCHED_FIFO, ¶m);
2030 stop->sched_class = &stop_sched_class;
2033 cpu_rq(cpu)->stop = stop;
2037 * Reset it back to a normal scheduling class so that
2038 * it can die in pieces.
2040 old_stop->sched_class = &rt_sched_class;
2045 * __normal_prio - return the priority that is based on the static prio
2047 static inline int __normal_prio(struct task_struct *p)
2049 return p->static_prio;
2053 * Calculate the expected normal priority: i.e. priority
2054 * without taking RT-inheritance into account. Might be
2055 * boosted by interactivity modifiers. Changes upon fork,
2056 * setprio syscalls, and whenever the interactivity
2057 * estimator recalculates.
2059 static inline int normal_prio(struct task_struct *p)
2063 if (task_has_rt_policy(p))
2064 prio = MAX_RT_PRIO-1 - p->rt_priority;
2066 prio = __normal_prio(p);
2071 * Calculate the current priority, i.e. the priority
2072 * taken into account by the scheduler. This value might
2073 * be boosted by RT tasks, or might be boosted by
2074 * interactivity modifiers. Will be RT if the task got
2075 * RT-boosted. If not then it returns p->normal_prio.
2077 static int effective_prio(struct task_struct *p)
2079 p->normal_prio = normal_prio(p);
2081 * If we are RT tasks or we were boosted to RT priority,
2082 * keep the priority unchanged. Otherwise, update priority
2083 * to the normal priority:
2085 if (!rt_prio(p->prio))
2086 return p->normal_prio;
2091 * task_curr - is this task currently executing on a CPU?
2092 * @p: the task in question.
2094 inline int task_curr(const struct task_struct *p)
2096 return cpu_curr(task_cpu(p)) == p;
2099 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2100 const struct sched_class *prev_class,
2103 if (prev_class != p->sched_class) {
2104 if (prev_class->switched_from)
2105 prev_class->switched_from(rq, p);
2106 p->sched_class->switched_to(rq, p);
2107 } else if (oldprio != p->prio)
2108 p->sched_class->prio_changed(rq, p, oldprio);
2111 static void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
2113 const struct sched_class *class;
2115 if (p->sched_class == rq->curr->sched_class) {
2116 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
2118 for_each_class(class) {
2119 if (class == rq->curr->sched_class)
2121 if (class == p->sched_class) {
2122 resched_task(rq->curr);
2129 * A queue event has occurred, and we're going to schedule. In
2130 * this case, we can save a useless back to back clock update.
2132 if (rq->curr->on_rq && test_tsk_need_resched(rq->curr))
2133 rq->skip_clock_update = 1;
2138 * Is this task likely cache-hot:
2141 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2145 if (p->sched_class != &fair_sched_class)
2148 if (unlikely(p->policy == SCHED_IDLE))
2152 * Buddy candidates are cache hot:
2154 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2155 (&p->se == cfs_rq_of(&p->se)->next ||
2156 &p->se == cfs_rq_of(&p->se)->last))
2159 if (sysctl_sched_migration_cost == -1)
2161 if (sysctl_sched_migration_cost == 0)
2164 delta = now - p->se.exec_start;
2166 return delta < (s64)sysctl_sched_migration_cost;
2169 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2171 #ifdef CONFIG_SCHED_DEBUG
2173 * We should never call set_task_cpu() on a blocked task,
2174 * ttwu() will sort out the placement.
2176 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2177 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2180 trace_sched_migrate_task(p, new_cpu);
2182 if (task_cpu(p) != new_cpu) {
2183 p->se.nr_migrations++;
2184 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2187 __set_task_cpu(p, new_cpu);
2190 struct migration_arg {
2191 struct task_struct *task;
2195 static int migration_cpu_stop(void *data);
2198 * The task's runqueue lock must be held.
2199 * Returns true if you have to wait for migration thread.
2201 static bool need_migrate_task(struct task_struct *p)
2204 * If the task is not on a runqueue (and not running), then
2205 * the next wake-up will properly place the task.
2207 bool running = p->on_rq || p->on_cpu;
2208 smp_rmb(); /* finish_lock_switch() */
2213 * wait_task_inactive - wait for a thread to unschedule.
2215 * If @match_state is nonzero, it's the @p->state value just checked and
2216 * not expected to change. If it changes, i.e. @p might have woken up,
2217 * then return zero. When we succeed in waiting for @p to be off its CPU,
2218 * we return a positive number (its total switch count). If a second call
2219 * a short while later returns the same number, the caller can be sure that
2220 * @p has remained unscheduled the whole time.
2222 * The caller must ensure that the task *will* unschedule sometime soon,
2223 * else this function might spin for a *long* time. This function can't
2224 * be called with interrupts off, or it may introduce deadlock with
2225 * smp_call_function() if an IPI is sent by the same process we are
2226 * waiting to become inactive.
2228 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2230 unsigned long flags;
2237 * We do the initial early heuristics without holding
2238 * any task-queue locks at all. We'll only try to get
2239 * the runqueue lock when things look like they will
2245 * If the task is actively running on another CPU
2246 * still, just relax and busy-wait without holding
2249 * NOTE! Since we don't hold any locks, it's not
2250 * even sure that "rq" stays as the right runqueue!
2251 * But we don't care, since "task_running()" will
2252 * return false if the runqueue has changed and p
2253 * is actually now running somewhere else!
2255 while (task_running(rq, p)) {
2256 if (match_state && unlikely(p->state != match_state))
2262 * Ok, time to look more closely! We need the rq
2263 * lock now, to be *sure*. If we're wrong, we'll
2264 * just go back and repeat.
2266 rq = task_rq_lock(p, &flags);
2267 trace_sched_wait_task(p);
2268 running = task_running(rq, p);
2271 if (!match_state || p->state == match_state)
2272 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2273 task_rq_unlock(rq, &flags);
2276 * If it changed from the expected state, bail out now.
2278 if (unlikely(!ncsw))
2282 * Was it really running after all now that we
2283 * checked with the proper locks actually held?
2285 * Oops. Go back and try again..
2287 if (unlikely(running)) {
2293 * It's not enough that it's not actively running,
2294 * it must be off the runqueue _entirely_, and not
2297 * So if it was still runnable (but just not actively
2298 * running right now), it's preempted, and we should
2299 * yield - it could be a while.
2301 if (unlikely(on_rq)) {
2302 ktime_t to = ktime_set(0, NSEC_PER_SEC/HZ);
2304 set_current_state(TASK_UNINTERRUPTIBLE);
2305 schedule_hrtimeout(&to, HRTIMER_MODE_REL);
2310 * Ahh, all good. It wasn't running, and it wasn't
2311 * runnable, which means that it will never become
2312 * running in the future either. We're all done!
2321 * kick_process - kick a running thread to enter/exit the kernel
2322 * @p: the to-be-kicked thread
2324 * Cause a process which is running on another CPU to enter
2325 * kernel-mode, without any delay. (to get signals handled.)
2327 * NOTE: this function doesn't have to take the runqueue lock,
2328 * because all it wants to ensure is that the remote task enters
2329 * the kernel. If the IPI races and the task has been migrated
2330 * to another CPU then no harm is done and the purpose has been
2333 void kick_process(struct task_struct *p)
2339 if ((cpu != smp_processor_id()) && task_curr(p))
2340 smp_send_reschedule(cpu);
2343 EXPORT_SYMBOL_GPL(kick_process);
2344 #endif /* CONFIG_SMP */
2348 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2350 static int select_fallback_rq(int cpu, struct task_struct *p)
2353 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2355 /* Look for allowed, online CPU in same node. */
2356 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2357 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2360 /* Any allowed, online CPU? */
2361 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2362 if (dest_cpu < nr_cpu_ids)
2365 /* No more Mr. Nice Guy. */
2366 dest_cpu = cpuset_cpus_allowed_fallback(p);
2368 * Don't tell them about moving exiting tasks or
2369 * kernel threads (both mm NULL), since they never
2372 if (p->mm && printk_ratelimit()) {
2373 printk(KERN_INFO "process %d (%s) no longer affine to cpu%d\n",
2374 task_pid_nr(p), p->comm, cpu);
2381 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2384 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2386 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2389 * In order not to call set_task_cpu() on a blocking task we need
2390 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2393 * Since this is common to all placement strategies, this lives here.
2395 * [ this allows ->select_task() to simply return task_cpu(p) and
2396 * not worry about this generic constraint ]
2398 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2400 cpu = select_fallback_rq(task_cpu(p), p);
2405 static void update_avg(u64 *avg, u64 sample)
2407 s64 diff = sample - *avg;
2413 ttwu_stat(struct rq *rq, struct task_struct *p, int cpu, int wake_flags)
2415 #ifdef CONFIG_SCHEDSTATS
2417 int this_cpu = smp_processor_id();
2419 if (cpu == this_cpu) {
2420 schedstat_inc(rq, ttwu_local);
2421 schedstat_inc(p, se.statistics.nr_wakeups_local);
2423 struct sched_domain *sd;
2425 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2426 for_each_domain(this_cpu, sd) {
2427 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2428 schedstat_inc(sd, ttwu_wake_remote);
2433 #endif /* CONFIG_SMP */
2435 schedstat_inc(rq, ttwu_count);
2436 schedstat_inc(p, se.statistics.nr_wakeups);
2438 if (wake_flags & WF_SYNC)
2439 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2441 if (cpu != task_cpu(p))
2442 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2444 #endif /* CONFIG_SCHEDSTATS */
2447 static void ttwu_activate(struct rq *rq, struct task_struct *p, int en_flags)
2449 activate_task(rq, p, en_flags);
2452 /* if a worker is waking up, notify workqueue */
2453 if (p->flags & PF_WQ_WORKER)
2454 wq_worker_waking_up(p, cpu_of(rq));
2458 ttwu_post_activation(struct task_struct *p, struct rq *rq, int wake_flags)
2460 trace_sched_wakeup(p, true);
2461 check_preempt_curr(rq, p, wake_flags);
2463 p->state = TASK_RUNNING;
2465 if (p->sched_class->task_woken)
2466 p->sched_class->task_woken(rq, p);
2468 if (unlikely(rq->idle_stamp)) {
2469 u64 delta = rq->clock - rq->idle_stamp;
2470 u64 max = 2*sysctl_sched_migration_cost;
2475 update_avg(&rq->avg_idle, delta);
2482 * try_to_wake_up - wake up a thread
2483 * @p: the thread to be awakened
2484 * @state: the mask of task states that can be woken
2485 * @wake_flags: wake modifier flags (WF_*)
2487 * Put it on the run-queue if it's not already there. The "current"
2488 * thread is always on the run-queue (except when the actual
2489 * re-schedule is in progress), and as such you're allowed to do
2490 * the simpler "current->state = TASK_RUNNING" to mark yourself
2491 * runnable without the overhead of this.
2493 * Returns %true if @p was woken up, %false if it was already running
2494 * or @state didn't match @p's state.
2496 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2499 int cpu, orig_cpu, this_cpu, success = 0;
2500 unsigned long flags;
2501 unsigned long en_flags = ENQUEUE_WAKEUP;
2504 this_cpu = get_cpu();
2507 raw_spin_lock_irqsave(&p->pi_lock, flags);
2508 rq = __task_rq_lock(p);
2509 if (!(p->state & state))
2519 if (unlikely(task_running(rq, p)))
2522 p->sched_contributes_to_load = !!task_contributes_to_load(p);
2523 p->state = TASK_WAKING;
2525 if (p->sched_class->task_waking) {
2526 p->sched_class->task_waking(p);
2527 en_flags |= ENQUEUE_WAKING;
2530 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2531 if (cpu != orig_cpu)
2532 set_task_cpu(p, cpu);
2533 __task_rq_unlock(rq);
2536 raw_spin_lock(&rq->lock);
2539 * We migrated the task without holding either rq->lock, however
2540 * since the task is not on the task list itself, nobody else
2541 * will try and migrate the task, hence the rq should match the
2542 * cpu we just moved it to.
2544 WARN_ON(task_cpu(p) != cpu);
2545 WARN_ON(p->state != TASK_WAKING);
2547 if (p->sched_contributes_to_load)
2548 rq->nr_uninterruptible--;
2551 #endif /* CONFIG_SMP */
2552 ttwu_activate(rq, p, en_flags);
2554 ttwu_post_activation(p, rq, wake_flags);
2555 ttwu_stat(rq, p, cpu, wake_flags);
2558 __task_rq_unlock(rq);
2559 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
2566 * try_to_wake_up_local - try to wake up a local task with rq lock held
2567 * @p: the thread to be awakened
2569 * Put @p on the run-queue if it's not already there. The caller must
2570 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2571 * the current task. this_rq() stays locked over invocation.
2573 static void try_to_wake_up_local(struct task_struct *p)
2575 struct rq *rq = task_rq(p);
2577 BUG_ON(rq != this_rq());
2578 BUG_ON(p == current);
2579 lockdep_assert_held(&rq->lock);
2581 if (!(p->state & TASK_NORMAL))
2585 ttwu_activate(rq, p, ENQUEUE_WAKEUP);
2587 ttwu_post_activation(p, rq, 0);
2588 ttwu_stat(rq, p, smp_processor_id(), 0);
2592 * wake_up_process - Wake up a specific process
2593 * @p: The process to be woken up.
2595 * Attempt to wake up the nominated process and move it to the set of runnable
2596 * processes. Returns 1 if the process was woken up, 0 if it was already
2599 * It may be assumed that this function implies a write memory barrier before
2600 * changing the task state if and only if any tasks are woken up.
2602 int wake_up_process(struct task_struct *p)
2604 return try_to_wake_up(p, TASK_ALL, 0);
2606 EXPORT_SYMBOL(wake_up_process);
2608 int wake_up_state(struct task_struct *p, unsigned int state)
2610 return try_to_wake_up(p, state, 0);
2614 * Perform scheduler related setup for a newly forked process p.
2615 * p is forked by current.
2617 * __sched_fork() is basic setup used by init_idle() too:
2619 static void __sched_fork(struct task_struct *p)
2624 p->se.exec_start = 0;
2625 p->se.sum_exec_runtime = 0;
2626 p->se.prev_sum_exec_runtime = 0;
2627 p->se.nr_migrations = 0;
2629 INIT_LIST_HEAD(&p->se.group_node);
2631 #ifdef CONFIG_SCHEDSTATS
2632 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2635 INIT_LIST_HEAD(&p->rt.run_list);
2637 #ifdef CONFIG_PREEMPT_NOTIFIERS
2638 INIT_HLIST_HEAD(&p->preempt_notifiers);
2643 * fork()/clone()-time setup:
2645 void sched_fork(struct task_struct *p, int clone_flags)
2647 int cpu = get_cpu();
2651 * We mark the process as running here. This guarantees that
2652 * nobody will actually run it, and a signal or other external
2653 * event cannot wake it up and insert it on the runqueue either.
2655 p->state = TASK_RUNNING;
2658 * Revert to default priority/policy on fork if requested.
2660 if (unlikely(p->sched_reset_on_fork)) {
2661 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2662 p->policy = SCHED_NORMAL;
2663 p->normal_prio = p->static_prio;
2666 if (PRIO_TO_NICE(p->static_prio) < 0) {
2667 p->static_prio = NICE_TO_PRIO(0);
2668 p->normal_prio = p->static_prio;
2673 * We don't need the reset flag anymore after the fork. It has
2674 * fulfilled its duty:
2676 p->sched_reset_on_fork = 0;
2680 * Make sure we do not leak PI boosting priority to the child.
2682 p->prio = current->normal_prio;
2684 if (!rt_prio(p->prio))
2685 p->sched_class = &fair_sched_class;
2687 if (p->sched_class->task_fork)
2688 p->sched_class->task_fork(p);
2691 * The child is not yet in the pid-hash so no cgroup attach races,
2692 * and the cgroup is pinned to this child due to cgroup_fork()
2693 * is ran before sched_fork().
2695 * Silence PROVE_RCU.
2698 set_task_cpu(p, cpu);
2701 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2702 if (likely(sched_info_on()))
2703 memset(&p->sched_info, 0, sizeof(p->sched_info));
2705 #if defined(CONFIG_SMP)
2708 #ifdef CONFIG_PREEMPT
2709 /* Want to start with kernel preemption disabled. */
2710 task_thread_info(p)->preempt_count = 1;
2713 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2720 * wake_up_new_task - wake up a newly created task for the first time.
2722 * This function will do some initial scheduler statistics housekeeping
2723 * that must be done for every newly created context, then puts the task
2724 * on the runqueue and wakes it.
2726 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2728 unsigned long flags;
2730 int cpu __maybe_unused = get_cpu();
2733 rq = task_rq_lock(p, &flags);
2734 p->state = TASK_WAKING;
2737 * Fork balancing, do it here and not earlier because:
2738 * - cpus_allowed can change in the fork path
2739 * - any previously selected cpu might disappear through hotplug
2741 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2742 * without people poking at ->cpus_allowed.
2744 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2745 set_task_cpu(p, cpu);
2747 p->state = TASK_RUNNING;
2748 task_rq_unlock(rq, &flags);
2751 rq = task_rq_lock(p, &flags);
2752 activate_task(rq, p, 0);
2754 trace_sched_wakeup_new(p, true);
2755 check_preempt_curr(rq, p, WF_FORK);
2757 if (p->sched_class->task_woken)
2758 p->sched_class->task_woken(rq, p);
2760 task_rq_unlock(rq, &flags);
2764 #ifdef CONFIG_PREEMPT_NOTIFIERS
2767 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2768 * @notifier: notifier struct to register
2770 void preempt_notifier_register(struct preempt_notifier *notifier)
2772 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2774 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2777 * preempt_notifier_unregister - no longer interested in preemption notifications
2778 * @notifier: notifier struct to unregister
2780 * This is safe to call from within a preemption notifier.
2782 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2784 hlist_del(¬ifier->link);
2786 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2788 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2790 struct preempt_notifier *notifier;
2791 struct hlist_node *node;
2793 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2794 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2798 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2799 struct task_struct *next)
2801 struct preempt_notifier *notifier;
2802 struct hlist_node *node;
2804 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2805 notifier->ops->sched_out(notifier, next);
2808 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2810 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2815 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2816 struct task_struct *next)
2820 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2823 * prepare_task_switch - prepare to switch tasks
2824 * @rq: the runqueue preparing to switch
2825 * @prev: the current task that is being switched out
2826 * @next: the task we are going to switch to.
2828 * This is called with the rq lock held and interrupts off. It must
2829 * be paired with a subsequent finish_task_switch after the context
2832 * prepare_task_switch sets up locking and calls architecture specific
2836 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2837 struct task_struct *next)
2839 sched_info_switch(prev, next);
2840 perf_event_task_sched_out(prev, next);
2841 fire_sched_out_preempt_notifiers(prev, next);
2842 prepare_lock_switch(rq, next);
2843 prepare_arch_switch(next);
2844 trace_sched_switch(prev, next);
2848 * finish_task_switch - clean up after a task-switch
2849 * @rq: runqueue associated with task-switch
2850 * @prev: the thread we just switched away from.
2852 * finish_task_switch must be called after the context switch, paired
2853 * with a prepare_task_switch call before the context switch.
2854 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2855 * and do any other architecture-specific cleanup actions.
2857 * Note that we may have delayed dropping an mm in context_switch(). If
2858 * so, we finish that here outside of the runqueue lock. (Doing it
2859 * with the lock held can cause deadlocks; see schedule() for
2862 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2863 __releases(rq->lock)
2865 struct mm_struct *mm = rq->prev_mm;
2871 * A task struct has one reference for the use as "current".
2872 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2873 * schedule one last time. The schedule call will never return, and
2874 * the scheduled task must drop that reference.
2875 * The test for TASK_DEAD must occur while the runqueue locks are
2876 * still held, otherwise prev could be scheduled on another cpu, die
2877 * there before we look at prev->state, and then the reference would
2879 * Manfred Spraul <manfred@colorfullife.com>
2881 prev_state = prev->state;
2882 finish_arch_switch(prev);
2883 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2884 local_irq_disable();
2885 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2886 perf_event_task_sched_in(current);
2887 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2889 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2890 finish_lock_switch(rq, prev);
2892 fire_sched_in_preempt_notifiers(current);
2895 if (unlikely(prev_state == TASK_DEAD)) {
2897 * Remove function-return probe instances associated with this
2898 * task and put them back on the free list.
2900 kprobe_flush_task(prev);
2901 put_task_struct(prev);
2907 /* assumes rq->lock is held */
2908 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2910 if (prev->sched_class->pre_schedule)
2911 prev->sched_class->pre_schedule(rq, prev);
2914 /* rq->lock is NOT held, but preemption is disabled */
2915 static inline void post_schedule(struct rq *rq)
2917 if (rq->post_schedule) {
2918 unsigned long flags;
2920 raw_spin_lock_irqsave(&rq->lock, flags);
2921 if (rq->curr->sched_class->post_schedule)
2922 rq->curr->sched_class->post_schedule(rq);
2923 raw_spin_unlock_irqrestore(&rq->lock, flags);
2925 rq->post_schedule = 0;
2931 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2935 static inline void post_schedule(struct rq *rq)
2942 * schedule_tail - first thing a freshly forked thread must call.
2943 * @prev: the thread we just switched away from.
2945 asmlinkage void schedule_tail(struct task_struct *prev)
2946 __releases(rq->lock)
2948 struct rq *rq = this_rq();
2950 finish_task_switch(rq, prev);
2953 * FIXME: do we need to worry about rq being invalidated by the
2958 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2959 /* In this case, finish_task_switch does not reenable preemption */
2962 if (current->set_child_tid)
2963 put_user(task_pid_vnr(current), current->set_child_tid);
2967 * context_switch - switch to the new MM and the new
2968 * thread's register state.
2971 context_switch(struct rq *rq, struct task_struct *prev,
2972 struct task_struct *next)
2974 struct mm_struct *mm, *oldmm;
2976 prepare_task_switch(rq, prev, next);
2979 oldmm = prev->active_mm;
2981 * For paravirt, this is coupled with an exit in switch_to to
2982 * combine the page table reload and the switch backend into
2985 arch_start_context_switch(prev);
2988 next->active_mm = oldmm;
2989 atomic_inc(&oldmm->mm_count);
2990 enter_lazy_tlb(oldmm, next);
2992 switch_mm(oldmm, mm, next);
2995 prev->active_mm = NULL;
2996 rq->prev_mm = oldmm;
2999 * Since the runqueue lock will be released by the next
3000 * task (which is an invalid locking op but in the case
3001 * of the scheduler it's an obvious special-case), so we
3002 * do an early lockdep release here:
3004 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3005 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3008 /* Here we just switch the register state and the stack. */
3009 switch_to(prev, next, prev);
3013 * this_rq must be evaluated again because prev may have moved
3014 * CPUs since it called schedule(), thus the 'rq' on its stack
3015 * frame will be invalid.
3017 finish_task_switch(this_rq(), prev);
3021 * nr_running, nr_uninterruptible and nr_context_switches:
3023 * externally visible scheduler statistics: current number of runnable
3024 * threads, current number of uninterruptible-sleeping threads, total
3025 * number of context switches performed since bootup.
3027 unsigned long nr_running(void)
3029 unsigned long i, sum = 0;
3031 for_each_online_cpu(i)
3032 sum += cpu_rq(i)->nr_running;
3037 unsigned long nr_uninterruptible(void)
3039 unsigned long i, sum = 0;
3041 for_each_possible_cpu(i)
3042 sum += cpu_rq(i)->nr_uninterruptible;
3045 * Since we read the counters lockless, it might be slightly
3046 * inaccurate. Do not allow it to go below zero though:
3048 if (unlikely((long)sum < 0))
3054 unsigned long long nr_context_switches(void)
3057 unsigned long long sum = 0;
3059 for_each_possible_cpu(i)
3060 sum += cpu_rq(i)->nr_switches;
3065 unsigned long nr_iowait(void)
3067 unsigned long i, sum = 0;
3069 for_each_possible_cpu(i)
3070 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3075 unsigned long nr_iowait_cpu(int cpu)
3077 struct rq *this = cpu_rq(cpu);
3078 return atomic_read(&this->nr_iowait);
3081 unsigned long this_cpu_load(void)
3083 struct rq *this = this_rq();
3084 return this->cpu_load[0];
3088 /* Variables and functions for calc_load */
3089 static atomic_long_t calc_load_tasks;
3090 static unsigned long calc_load_update;
3091 unsigned long avenrun[3];
3092 EXPORT_SYMBOL(avenrun);
3094 static long calc_load_fold_active(struct rq *this_rq)
3096 long nr_active, delta = 0;
3098 nr_active = this_rq->nr_running;
3099 nr_active += (long) this_rq->nr_uninterruptible;
3101 if (nr_active != this_rq->calc_load_active) {
3102 delta = nr_active - this_rq->calc_load_active;
3103 this_rq->calc_load_active = nr_active;
3109 static unsigned long
3110 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3113 load += active * (FIXED_1 - exp);
3114 load += 1UL << (FSHIFT - 1);
3115 return load >> FSHIFT;
3120 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3122 * When making the ILB scale, we should try to pull this in as well.
3124 static atomic_long_t calc_load_tasks_idle;
3126 static void calc_load_account_idle(struct rq *this_rq)
3130 delta = calc_load_fold_active(this_rq);
3132 atomic_long_add(delta, &calc_load_tasks_idle);
3135 static long calc_load_fold_idle(void)
3140 * Its got a race, we don't care...
3142 if (atomic_long_read(&calc_load_tasks_idle))
3143 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3149 * fixed_power_int - compute: x^n, in O(log n) time
3151 * @x: base of the power
3152 * @frac_bits: fractional bits of @x
3153 * @n: power to raise @x to.
3155 * By exploiting the relation between the definition of the natural power
3156 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3157 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3158 * (where: n_i \elem {0, 1}, the binary vector representing n),
3159 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3160 * of course trivially computable in O(log_2 n), the length of our binary
3163 static unsigned long
3164 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3166 unsigned long result = 1UL << frac_bits;
3171 result += 1UL << (frac_bits - 1);
3172 result >>= frac_bits;
3178 x += 1UL << (frac_bits - 1);
3186 * a1 = a0 * e + a * (1 - e)
3188 * a2 = a1 * e + a * (1 - e)
3189 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3190 * = a0 * e^2 + a * (1 - e) * (1 + e)
3192 * a3 = a2 * e + a * (1 - e)
3193 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3194 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3198 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3199 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3200 * = a0 * e^n + a * (1 - e^n)
3202 * [1] application of the geometric series:
3205 * S_n := \Sum x^i = -------------
3208 static unsigned long
3209 calc_load_n(unsigned long load, unsigned long exp,
3210 unsigned long active, unsigned int n)
3213 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3217 * NO_HZ can leave us missing all per-cpu ticks calling
3218 * calc_load_account_active(), but since an idle CPU folds its delta into
3219 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3220 * in the pending idle delta if our idle period crossed a load cycle boundary.
3222 * Once we've updated the global active value, we need to apply the exponential
3223 * weights adjusted to the number of cycles missed.
3225 static void calc_global_nohz(unsigned long ticks)
3227 long delta, active, n;
3229 if (time_before(jiffies, calc_load_update))
3233 * If we crossed a calc_load_update boundary, make sure to fold
3234 * any pending idle changes, the respective CPUs might have
3235 * missed the tick driven calc_load_account_active() update
3238 delta = calc_load_fold_idle();
3240 atomic_long_add(delta, &calc_load_tasks);
3243 * If we were idle for multiple load cycles, apply them.
3245 if (ticks >= LOAD_FREQ) {
3246 n = ticks / LOAD_FREQ;
3248 active = atomic_long_read(&calc_load_tasks);
3249 active = active > 0 ? active * FIXED_1 : 0;
3251 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3252 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3253 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3255 calc_load_update += n * LOAD_FREQ;
3259 * Its possible the remainder of the above division also crosses
3260 * a LOAD_FREQ period, the regular check in calc_global_load()
3261 * which comes after this will take care of that.
3263 * Consider us being 11 ticks before a cycle completion, and us
3264 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3265 * age us 4 cycles, and the test in calc_global_load() will
3266 * pick up the final one.
3270 static void calc_load_account_idle(struct rq *this_rq)
3274 static inline long calc_load_fold_idle(void)
3279 static void calc_global_nohz(unsigned long ticks)
3285 * get_avenrun - get the load average array
3286 * @loads: pointer to dest load array
3287 * @offset: offset to add
3288 * @shift: shift count to shift the result left
3290 * These values are estimates at best, so no need for locking.
3292 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3294 loads[0] = (avenrun[0] + offset) << shift;
3295 loads[1] = (avenrun[1] + offset) << shift;
3296 loads[2] = (avenrun[2] + offset) << shift;
3300 * calc_load - update the avenrun load estimates 10 ticks after the
3301 * CPUs have updated calc_load_tasks.
3303 void calc_global_load(unsigned long ticks)
3307 calc_global_nohz(ticks);
3309 if (time_before(jiffies, calc_load_update + 10))
3312 active = atomic_long_read(&calc_load_tasks);
3313 active = active > 0 ? active * FIXED_1 : 0;
3315 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3316 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3317 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3319 calc_load_update += LOAD_FREQ;
3323 * Called from update_cpu_load() to periodically update this CPU's
3326 static void calc_load_account_active(struct rq *this_rq)
3330 if (time_before(jiffies, this_rq->calc_load_update))
3333 delta = calc_load_fold_active(this_rq);
3334 delta += calc_load_fold_idle();
3336 atomic_long_add(delta, &calc_load_tasks);
3338 this_rq->calc_load_update += LOAD_FREQ;
3342 * The exact cpuload at various idx values, calculated at every tick would be
3343 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3345 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3346 * on nth tick when cpu may be busy, then we have:
3347 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3348 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3350 * decay_load_missed() below does efficient calculation of
3351 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3352 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3354 * The calculation is approximated on a 128 point scale.
3355 * degrade_zero_ticks is the number of ticks after which load at any
3356 * particular idx is approximated to be zero.
3357 * degrade_factor is a precomputed table, a row for each load idx.
3358 * Each column corresponds to degradation factor for a power of two ticks,
3359 * based on 128 point scale.
3361 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3362 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3364 * With this power of 2 load factors, we can degrade the load n times
3365 * by looking at 1 bits in n and doing as many mult/shift instead of
3366 * n mult/shifts needed by the exact degradation.
3368 #define DEGRADE_SHIFT 7
3369 static const unsigned char
3370 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3371 static const unsigned char
3372 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3373 {0, 0, 0, 0, 0, 0, 0, 0},
3374 {64, 32, 8, 0, 0, 0, 0, 0},
3375 {96, 72, 40, 12, 1, 0, 0},
3376 {112, 98, 75, 43, 15, 1, 0},
3377 {120, 112, 98, 76, 45, 16, 2} };
3380 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3381 * would be when CPU is idle and so we just decay the old load without
3382 * adding any new load.
3384 static unsigned long
3385 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3389 if (!missed_updates)
3392 if (missed_updates >= degrade_zero_ticks[idx])
3396 return load >> missed_updates;
3398 while (missed_updates) {
3399 if (missed_updates % 2)
3400 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3402 missed_updates >>= 1;
3409 * Update rq->cpu_load[] statistics. This function is usually called every
3410 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3411 * every tick. We fix it up based on jiffies.
3413 static void update_cpu_load(struct rq *this_rq)
3415 unsigned long this_load = this_rq->load.weight;
3416 unsigned long curr_jiffies = jiffies;
3417 unsigned long pending_updates;
3420 this_rq->nr_load_updates++;
3422 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3423 if (curr_jiffies == this_rq->last_load_update_tick)
3426 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3427 this_rq->last_load_update_tick = curr_jiffies;
3429 /* Update our load: */
3430 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3431 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3432 unsigned long old_load, new_load;
3434 /* scale is effectively 1 << i now, and >> i divides by scale */
3436 old_load = this_rq->cpu_load[i];
3437 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3438 new_load = this_load;
3440 * Round up the averaging division if load is increasing. This
3441 * prevents us from getting stuck on 9 if the load is 10, for
3444 if (new_load > old_load)
3445 new_load += scale - 1;
3447 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3450 sched_avg_update(this_rq);
3453 static void update_cpu_load_active(struct rq *this_rq)
3455 update_cpu_load(this_rq);
3457 calc_load_account_active(this_rq);
3463 * sched_exec - execve() is a valuable balancing opportunity, because at
3464 * this point the task has the smallest effective memory and cache footprint.
3466 void sched_exec(void)
3468 struct task_struct *p = current;
3469 unsigned long flags;
3473 rq = task_rq_lock(p, &flags);
3474 dest_cpu = p->sched_class->select_task_rq(p, SD_BALANCE_EXEC, 0);
3475 if (dest_cpu == smp_processor_id())
3479 * select_task_rq() can race against ->cpus_allowed
3481 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3482 likely(cpu_active(dest_cpu)) && need_migrate_task(p)) {
3483 struct migration_arg arg = { p, dest_cpu };
3485 task_rq_unlock(rq, &flags);
3486 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3490 task_rq_unlock(rq, &flags);
3495 DEFINE_PER_CPU(struct kernel_stat, kstat);
3497 EXPORT_PER_CPU_SYMBOL(kstat);
3500 * Return any ns on the sched_clock that have not yet been accounted in
3501 * @p in case that task is currently running.
3503 * Called with task_rq_lock() held on @rq.
3505 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3509 if (task_current(rq, p)) {
3510 update_rq_clock(rq);
3511 ns = rq->clock_task - p->se.exec_start;
3519 unsigned long long task_delta_exec(struct task_struct *p)
3521 unsigned long flags;
3525 rq = task_rq_lock(p, &flags);
3526 ns = do_task_delta_exec(p, rq);
3527 task_rq_unlock(rq, &flags);
3533 * Return accounted runtime for the task.
3534 * In case the task is currently running, return the runtime plus current's
3535 * pending runtime that have not been accounted yet.
3537 unsigned long long task_sched_runtime(struct task_struct *p)
3539 unsigned long flags;
3543 rq = task_rq_lock(p, &flags);
3544 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3545 task_rq_unlock(rq, &flags);
3551 * Return sum_exec_runtime for the thread group.
3552 * In case the task is currently running, return the sum plus current's
3553 * pending runtime that have not been accounted yet.
3555 * Note that the thread group might have other running tasks as well,
3556 * so the return value not includes other pending runtime that other
3557 * running tasks might have.
3559 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3561 struct task_cputime totals;
3562 unsigned long flags;
3566 rq = task_rq_lock(p, &flags);
3567 thread_group_cputime(p, &totals);
3568 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3569 task_rq_unlock(rq, &flags);
3575 * Account user cpu time to a process.
3576 * @p: the process that the cpu time gets accounted to
3577 * @cputime: the cpu time spent in user space since the last update
3578 * @cputime_scaled: cputime scaled by cpu frequency
3580 void account_user_time(struct task_struct *p, cputime_t cputime,
3581 cputime_t cputime_scaled)
3583 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3586 /* Add user time to process. */
3587 p->utime = cputime_add(p->utime, cputime);
3588 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3589 account_group_user_time(p, cputime);
3591 /* Add user time to cpustat. */
3592 tmp = cputime_to_cputime64(cputime);
3593 if (TASK_NICE(p) > 0)
3594 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3596 cpustat->user = cputime64_add(cpustat->user, tmp);
3598 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3599 /* Account for user time used */
3600 acct_update_integrals(p);
3604 * Account guest cpu time to a process.
3605 * @p: the process that the cpu time gets accounted to
3606 * @cputime: the cpu time spent in virtual machine since the last update
3607 * @cputime_scaled: cputime scaled by cpu frequency
3609 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3610 cputime_t cputime_scaled)
3613 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3615 tmp = cputime_to_cputime64(cputime);
3617 /* Add guest time to process. */
3618 p->utime = cputime_add(p->utime, cputime);
3619 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3620 account_group_user_time(p, cputime);
3621 p->gtime = cputime_add(p->gtime, cputime);
3623 /* Add guest time to cpustat. */
3624 if (TASK_NICE(p) > 0) {
3625 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3626 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3628 cpustat->user = cputime64_add(cpustat->user, tmp);
3629 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3634 * Account system cpu time to a process and desired cpustat field
3635 * @p: the process that the cpu time gets accounted to
3636 * @cputime: the cpu time spent in kernel space since the last update
3637 * @cputime_scaled: cputime scaled by cpu frequency
3638 * @target_cputime64: pointer to cpustat field that has to be updated
3641 void __account_system_time(struct task_struct *p, cputime_t cputime,
3642 cputime_t cputime_scaled, cputime64_t *target_cputime64)
3644 cputime64_t tmp = cputime_to_cputime64(cputime);
3646 /* Add system time to process. */
3647 p->stime = cputime_add(p->stime, cputime);
3648 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3649 account_group_system_time(p, cputime);
3651 /* Add system time to cpustat. */
3652 *target_cputime64 = cputime64_add(*target_cputime64, tmp);
3653 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3655 /* Account for system time used */
3656 acct_update_integrals(p);
3660 * Account system cpu time to a process.
3661 * @p: the process that the cpu time gets accounted to
3662 * @hardirq_offset: the offset to subtract from hardirq_count()
3663 * @cputime: the cpu time spent in kernel space since the last update
3664 * @cputime_scaled: cputime scaled by cpu frequency
3666 void account_system_time(struct task_struct *p, int hardirq_offset,
3667 cputime_t cputime, cputime_t cputime_scaled)
3669 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3670 cputime64_t *target_cputime64;
3672 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3673 account_guest_time(p, cputime, cputime_scaled);
3677 if (hardirq_count() - hardirq_offset)
3678 target_cputime64 = &cpustat->irq;
3679 else if (in_serving_softirq())
3680 target_cputime64 = &cpustat->softirq;
3682 target_cputime64 = &cpustat->system;
3684 __account_system_time(p, cputime, cputime_scaled, target_cputime64);
3688 * Account for involuntary wait time.
3689 * @cputime: the cpu time spent in involuntary wait
3691 void account_steal_time(cputime_t cputime)
3693 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3694 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3696 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3700 * Account for idle time.
3701 * @cputime: the cpu time spent in idle wait
3703 void account_idle_time(cputime_t cputime)
3705 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3706 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3707 struct rq *rq = this_rq();
3709 if (atomic_read(&rq->nr_iowait) > 0)
3710 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3712 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3715 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3717 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3719 * Account a tick to a process and cpustat
3720 * @p: the process that the cpu time gets accounted to
3721 * @user_tick: is the tick from userspace
3722 * @rq: the pointer to rq
3724 * Tick demultiplexing follows the order
3725 * - pending hardirq update
3726 * - pending softirq update
3730 * - check for guest_time
3731 * - else account as system_time
3733 * Check for hardirq is done both for system and user time as there is
3734 * no timer going off while we are on hardirq and hence we may never get an
3735 * opportunity to update it solely in system time.
3736 * p->stime and friends are only updated on system time and not on irq
3737 * softirq as those do not count in task exec_runtime any more.
3739 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3742 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3743 cputime64_t tmp = cputime_to_cputime64(cputime_one_jiffy);
3744 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3746 if (irqtime_account_hi_update()) {
3747 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3748 } else if (irqtime_account_si_update()) {
3749 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3750 } else if (this_cpu_ksoftirqd() == p) {
3752 * ksoftirqd time do not get accounted in cpu_softirq_time.
3753 * So, we have to handle it separately here.
3754 * Also, p->stime needs to be updated for ksoftirqd.
3756 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3758 } else if (user_tick) {
3759 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3760 } else if (p == rq->idle) {
3761 account_idle_time(cputime_one_jiffy);
3762 } else if (p->flags & PF_VCPU) { /* System time or guest time */
3763 account_guest_time(p, cputime_one_jiffy, one_jiffy_scaled);
3765 __account_system_time(p, cputime_one_jiffy, one_jiffy_scaled,
3770 static void irqtime_account_idle_ticks(int ticks)
3773 struct rq *rq = this_rq();
3775 for (i = 0; i < ticks; i++)
3776 irqtime_account_process_tick(current, 0, rq);
3778 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3779 static void irqtime_account_idle_ticks(int ticks) {}
3780 static void irqtime_account_process_tick(struct task_struct *p, int user_tick,
3782 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3785 * Account a single tick of cpu time.
3786 * @p: the process that the cpu time gets accounted to
3787 * @user_tick: indicates if the tick is a user or a system tick
3789 void account_process_tick(struct task_struct *p, int user_tick)
3791 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3792 struct rq *rq = this_rq();
3794 if (sched_clock_irqtime) {
3795 irqtime_account_process_tick(p, user_tick, rq);
3800 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3801 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3802 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3805 account_idle_time(cputime_one_jiffy);
3809 * Account multiple ticks of steal time.
3810 * @p: the process from which the cpu time has been stolen
3811 * @ticks: number of stolen ticks
3813 void account_steal_ticks(unsigned long ticks)
3815 account_steal_time(jiffies_to_cputime(ticks));
3819 * Account multiple ticks of idle time.
3820 * @ticks: number of stolen ticks
3822 void account_idle_ticks(unsigned long ticks)
3825 if (sched_clock_irqtime) {
3826 irqtime_account_idle_ticks(ticks);
3830 account_idle_time(jiffies_to_cputime(ticks));
3836 * Use precise platform statistics if available:
3838 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3839 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3845 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3847 struct task_cputime cputime;
3849 thread_group_cputime(p, &cputime);
3851 *ut = cputime.utime;
3852 *st = cputime.stime;
3856 #ifndef nsecs_to_cputime
3857 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3860 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3862 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3865 * Use CFS's precise accounting:
3867 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3873 do_div(temp, total);
3874 utime = (cputime_t)temp;
3879 * Compare with previous values, to keep monotonicity:
3881 p->prev_utime = max(p->prev_utime, utime);
3882 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3884 *ut = p->prev_utime;
3885 *st = p->prev_stime;
3889 * Must be called with siglock held.
3891 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3893 struct signal_struct *sig = p->signal;
3894 struct task_cputime cputime;
3895 cputime_t rtime, utime, total;
3897 thread_group_cputime(p, &cputime);
3899 total = cputime_add(cputime.utime, cputime.stime);
3900 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3905 temp *= cputime.utime;
3906 do_div(temp, total);
3907 utime = (cputime_t)temp;
3911 sig->prev_utime = max(sig->prev_utime, utime);
3912 sig->prev_stime = max(sig->prev_stime,
3913 cputime_sub(rtime, sig->prev_utime));
3915 *ut = sig->prev_utime;
3916 *st = sig->prev_stime;
3921 * This function gets called by the timer code, with HZ frequency.
3922 * We call it with interrupts disabled.
3924 * It also gets called by the fork code, when changing the parent's
3927 void scheduler_tick(void)
3929 int cpu = smp_processor_id();
3930 struct rq *rq = cpu_rq(cpu);
3931 struct task_struct *curr = rq->curr;
3935 raw_spin_lock(&rq->lock);
3936 update_rq_clock(rq);
3937 update_cpu_load_active(rq);
3938 curr->sched_class->task_tick(rq, curr, 0);
3939 raw_spin_unlock(&rq->lock);
3941 perf_event_task_tick();
3944 rq->idle_at_tick = idle_cpu(cpu);
3945 trigger_load_balance(rq, cpu);
3949 notrace unsigned long get_parent_ip(unsigned long addr)
3951 if (in_lock_functions(addr)) {
3952 addr = CALLER_ADDR2;
3953 if (in_lock_functions(addr))
3954 addr = CALLER_ADDR3;
3959 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3960 defined(CONFIG_PREEMPT_TRACER))
3962 void __kprobes add_preempt_count(int val)
3964 #ifdef CONFIG_DEBUG_PREEMPT
3968 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3971 preempt_count() += val;
3972 #ifdef CONFIG_DEBUG_PREEMPT
3974 * Spinlock count overflowing soon?
3976 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3979 if (preempt_count() == val)
3980 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3982 EXPORT_SYMBOL(add_preempt_count);
3984 void __kprobes sub_preempt_count(int val)
3986 #ifdef CONFIG_DEBUG_PREEMPT
3990 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3993 * Is the spinlock portion underflowing?
3995 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3996 !(preempt_count() & PREEMPT_MASK)))
4000 if (preempt_count() == val)
4001 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4002 preempt_count() -= val;
4004 EXPORT_SYMBOL(sub_preempt_count);
4009 * Print scheduling while atomic bug:
4011 static noinline void __schedule_bug(struct task_struct *prev)
4013 struct pt_regs *regs = get_irq_regs();
4015 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4016 prev->comm, prev->pid, preempt_count());
4018 debug_show_held_locks(prev);
4020 if (irqs_disabled())
4021 print_irqtrace_events(prev);
4030 * Various schedule()-time debugging checks and statistics:
4032 static inline void schedule_debug(struct task_struct *prev)
4035 * Test if we are atomic. Since do_exit() needs to call into
4036 * schedule() atomically, we ignore that path for now.
4037 * Otherwise, whine if we are scheduling when we should not be.
4039 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4040 __schedule_bug(prev);
4042 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4044 schedstat_inc(this_rq(), sched_count);
4045 #ifdef CONFIG_SCHEDSTATS
4046 if (unlikely(prev->lock_depth >= 0)) {
4047 schedstat_inc(this_rq(), rq_sched_info.bkl_count);
4048 schedstat_inc(prev, sched_info.bkl_count);
4053 static void put_prev_task(struct rq *rq, struct task_struct *prev)
4056 update_rq_clock(rq);
4057 prev->sched_class->put_prev_task(rq, prev);
4061 * Pick up the highest-prio task:
4063 static inline struct task_struct *
4064 pick_next_task(struct rq *rq)
4066 const struct sched_class *class;
4067 struct task_struct *p;
4070 * Optimization: we know that if all tasks are in
4071 * the fair class we can call that function directly:
4073 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4074 p = fair_sched_class.pick_next_task(rq);
4079 for_each_class(class) {
4080 p = class->pick_next_task(rq);
4085 BUG(); /* the idle class will always have a runnable task */
4089 * schedule() is the main scheduler function.
4091 asmlinkage void __sched schedule(void)
4093 struct task_struct *prev, *next;
4094 unsigned long *switch_count;
4100 cpu = smp_processor_id();
4102 rcu_note_context_switch(cpu);
4105 schedule_debug(prev);
4107 if (sched_feat(HRTICK))
4110 raw_spin_lock_irq(&rq->lock);
4112 switch_count = &prev->nivcsw;
4113 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4114 if (unlikely(signal_pending_state(prev->state, prev))) {
4115 prev->state = TASK_RUNNING;
4118 * If a worker is going to sleep, notify and
4119 * ask workqueue whether it wants to wake up a
4120 * task to maintain concurrency. If so, wake
4123 if (prev->flags & PF_WQ_WORKER) {
4124 struct task_struct *to_wakeup;
4126 to_wakeup = wq_worker_sleeping(prev, cpu);
4128 try_to_wake_up_local(to_wakeup);
4131 deactivate_task(rq, prev, DEQUEUE_SLEEP);
4135 * If we are going to sleep and we have plugged IO queued, make
4136 * sure to submit it to avoid deadlocks.
4138 if (blk_needs_flush_plug(prev)) {
4139 raw_spin_unlock(&rq->lock);
4140 blk_flush_plug(prev);
4141 raw_spin_lock(&rq->lock);
4144 switch_count = &prev->nvcsw;
4147 pre_schedule(rq, prev);
4149 if (unlikely(!rq->nr_running))
4150 idle_balance(cpu, rq);
4152 put_prev_task(rq, prev);
4153 next = pick_next_task(rq);
4154 clear_tsk_need_resched(prev);
4155 rq->skip_clock_update = 0;
4157 if (likely(prev != next)) {
4162 context_switch(rq, prev, next); /* unlocks the rq */
4164 * The context switch have flipped the stack from under us
4165 * and restored the local variables which were saved when
4166 * this task called schedule() in the past. prev == current
4167 * is still correct, but it can be moved to another cpu/rq.
4169 cpu = smp_processor_id();
4172 raw_spin_unlock_irq(&rq->lock);
4176 preempt_enable_no_resched();
4180 EXPORT_SYMBOL(schedule);
4182 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4184 static inline bool owner_running(struct mutex *lock, struct task_struct *owner)
4189 if (lock->owner != owner)
4193 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4194 * lock->owner still matches owner, if that fails, owner might
4195 * point to free()d memory, if it still matches, the rcu_read_lock()
4196 * ensures the memory stays valid.
4200 ret = owner->on_cpu;
4208 * Look out! "owner" is an entirely speculative pointer
4209 * access and not reliable.
4211 int mutex_spin_on_owner(struct mutex *lock, struct task_struct *owner)
4213 if (!sched_feat(OWNER_SPIN))
4216 while (owner_running(lock, owner)) {
4220 arch_mutex_cpu_relax();
4224 * If the owner changed to another task there is likely
4225 * heavy contention, stop spinning.
4234 #ifdef CONFIG_PREEMPT
4236 * this is the entry point to schedule() from in-kernel preemption
4237 * off of preempt_enable. Kernel preemptions off return from interrupt
4238 * occur there and call schedule directly.
4240 asmlinkage void __sched notrace preempt_schedule(void)
4242 struct thread_info *ti = current_thread_info();
4245 * If there is a non-zero preempt_count or interrupts are disabled,
4246 * we do not want to preempt the current task. Just return..
4248 if (likely(ti->preempt_count || irqs_disabled()))
4252 add_preempt_count_notrace(PREEMPT_ACTIVE);
4254 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4257 * Check again in case we missed a preemption opportunity
4258 * between schedule and now.
4261 } while (need_resched());
4263 EXPORT_SYMBOL(preempt_schedule);
4266 * this is the entry point to schedule() from kernel preemption
4267 * off of irq context.
4268 * Note, that this is called and return with irqs disabled. This will
4269 * protect us against recursive calling from irq.
4271 asmlinkage void __sched preempt_schedule_irq(void)
4273 struct thread_info *ti = current_thread_info();
4275 /* Catch callers which need to be fixed */
4276 BUG_ON(ti->preempt_count || !irqs_disabled());
4279 add_preempt_count(PREEMPT_ACTIVE);
4282 local_irq_disable();
4283 sub_preempt_count(PREEMPT_ACTIVE);
4286 * Check again in case we missed a preemption opportunity
4287 * between schedule and now.
4290 } while (need_resched());
4293 #endif /* CONFIG_PREEMPT */
4295 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4298 return try_to_wake_up(curr->private, mode, wake_flags);
4300 EXPORT_SYMBOL(default_wake_function);
4303 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4304 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4305 * number) then we wake all the non-exclusive tasks and one exclusive task.
4307 * There are circumstances in which we can try to wake a task which has already
4308 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4309 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4311 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4312 int nr_exclusive, int wake_flags, void *key)
4314 wait_queue_t *curr, *next;
4316 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4317 unsigned flags = curr->flags;
4319 if (curr->func(curr, mode, wake_flags, key) &&
4320 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4326 * __wake_up - wake up threads blocked on a waitqueue.
4328 * @mode: which threads
4329 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4330 * @key: is directly passed to the wakeup function
4332 * It may be assumed that this function implies a write memory barrier before
4333 * changing the task state if and only if any tasks are woken up.
4335 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4336 int nr_exclusive, void *key)
4338 unsigned long flags;
4340 spin_lock_irqsave(&q->lock, flags);
4341 __wake_up_common(q, mode, nr_exclusive, 0, key);
4342 spin_unlock_irqrestore(&q->lock, flags);
4344 EXPORT_SYMBOL(__wake_up);
4347 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4349 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4351 __wake_up_common(q, mode, 1, 0, NULL);
4353 EXPORT_SYMBOL_GPL(__wake_up_locked);
4355 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4357 __wake_up_common(q, mode, 1, 0, key);
4359 EXPORT_SYMBOL_GPL(__wake_up_locked_key);
4362 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4364 * @mode: which threads
4365 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4366 * @key: opaque value to be passed to wakeup targets
4368 * The sync wakeup differs that the waker knows that it will schedule
4369 * away soon, so while the target thread will be woken up, it will not
4370 * be migrated to another CPU - ie. the two threads are 'synchronized'
4371 * with each other. This can prevent needless bouncing between CPUs.
4373 * On UP it can prevent extra preemption.
4375 * It may be assumed that this function implies a write memory barrier before
4376 * changing the task state if and only if any tasks are woken up.
4378 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4379 int nr_exclusive, void *key)
4381 unsigned long flags;
4382 int wake_flags = WF_SYNC;
4387 if (unlikely(!nr_exclusive))
4390 spin_lock_irqsave(&q->lock, flags);
4391 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4392 spin_unlock_irqrestore(&q->lock, flags);
4394 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4397 * __wake_up_sync - see __wake_up_sync_key()
4399 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4401 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4403 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4406 * complete: - signals a single thread waiting on this completion
4407 * @x: holds the state of this particular completion
4409 * This will wake up a single thread waiting on this completion. Threads will be
4410 * awakened in the same order in which they were queued.
4412 * See also complete_all(), wait_for_completion() and related routines.
4414 * It may be assumed that this function implies a write memory barrier before
4415 * changing the task state if and only if any tasks are woken up.
4417 void complete(struct completion *x)
4419 unsigned long flags;
4421 spin_lock_irqsave(&x->wait.lock, flags);
4423 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4424 spin_unlock_irqrestore(&x->wait.lock, flags);
4426 EXPORT_SYMBOL(complete);
4429 * complete_all: - signals all threads waiting on this completion
4430 * @x: holds the state of this particular completion
4432 * This will wake up all threads waiting on this particular completion event.
4434 * It may be assumed that this function implies a write memory barrier before
4435 * changing the task state if and only if any tasks are woken up.
4437 void complete_all(struct completion *x)
4439 unsigned long flags;
4441 spin_lock_irqsave(&x->wait.lock, flags);
4442 x->done += UINT_MAX/2;
4443 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4444 spin_unlock_irqrestore(&x->wait.lock, flags);
4446 EXPORT_SYMBOL(complete_all);
4448 static inline long __sched
4449 do_wait_for_common(struct completion *x, long timeout, int state)
4452 DECLARE_WAITQUEUE(wait, current);
4454 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4456 if (signal_pending_state(state, current)) {
4457 timeout = -ERESTARTSYS;
4460 __set_current_state(state);
4461 spin_unlock_irq(&x->wait.lock);
4462 timeout = schedule_timeout(timeout);
4463 spin_lock_irq(&x->wait.lock);
4464 } while (!x->done && timeout);
4465 __remove_wait_queue(&x->wait, &wait);
4470 return timeout ?: 1;
4474 wait_for_common(struct completion *x, long timeout, int state)
4478 spin_lock_irq(&x->wait.lock);
4479 timeout = do_wait_for_common(x, timeout, state);
4480 spin_unlock_irq(&x->wait.lock);
4485 * wait_for_completion: - waits for completion of a task
4486 * @x: holds the state of this particular completion
4488 * This waits to be signaled for completion of a specific task. It is NOT
4489 * interruptible and there is no timeout.
4491 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4492 * and interrupt capability. Also see complete().
4494 void __sched wait_for_completion(struct completion *x)
4496 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4498 EXPORT_SYMBOL(wait_for_completion);
4501 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4502 * @x: holds the state of this particular completion
4503 * @timeout: timeout value in jiffies
4505 * This waits for either a completion of a specific task to be signaled or for a
4506 * specified timeout to expire. The timeout is in jiffies. It is not
4509 unsigned long __sched
4510 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4512 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4514 EXPORT_SYMBOL(wait_for_completion_timeout);
4517 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4518 * @x: holds the state of this particular completion
4520 * This waits for completion of a specific task to be signaled. It is
4523 int __sched wait_for_completion_interruptible(struct completion *x)
4525 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4526 if (t == -ERESTARTSYS)
4530 EXPORT_SYMBOL(wait_for_completion_interruptible);
4533 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4534 * @x: holds the state of this particular completion
4535 * @timeout: timeout value in jiffies
4537 * This waits for either a completion of a specific task to be signaled or for a
4538 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4541 wait_for_completion_interruptible_timeout(struct completion *x,
4542 unsigned long timeout)
4544 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4546 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4549 * wait_for_completion_killable: - waits for completion of a task (killable)
4550 * @x: holds the state of this particular completion
4552 * This waits to be signaled for completion of a specific task. It can be
4553 * interrupted by a kill signal.
4555 int __sched wait_for_completion_killable(struct completion *x)
4557 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4558 if (t == -ERESTARTSYS)
4562 EXPORT_SYMBOL(wait_for_completion_killable);
4565 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4566 * @x: holds the state of this particular completion
4567 * @timeout: timeout value in jiffies
4569 * This waits for either a completion of a specific task to be
4570 * signaled or for a specified timeout to expire. It can be
4571 * interrupted by a kill signal. The timeout is in jiffies.
4574 wait_for_completion_killable_timeout(struct completion *x,
4575 unsigned long timeout)
4577 return wait_for_common(x, timeout, TASK_KILLABLE);
4579 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4582 * try_wait_for_completion - try to decrement a completion without blocking
4583 * @x: completion structure
4585 * Returns: 0 if a decrement cannot be done without blocking
4586 * 1 if a decrement succeeded.
4588 * If a completion is being used as a counting completion,
4589 * attempt to decrement the counter without blocking. This
4590 * enables us to avoid waiting if the resource the completion
4591 * is protecting is not available.
4593 bool try_wait_for_completion(struct completion *x)
4595 unsigned long flags;
4598 spin_lock_irqsave(&x->wait.lock, flags);
4603 spin_unlock_irqrestore(&x->wait.lock, flags);
4606 EXPORT_SYMBOL(try_wait_for_completion);
4609 * completion_done - Test to see if a completion has any waiters
4610 * @x: completion structure
4612 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4613 * 1 if there are no waiters.
4616 bool completion_done(struct completion *x)
4618 unsigned long flags;
4621 spin_lock_irqsave(&x->wait.lock, flags);
4624 spin_unlock_irqrestore(&x->wait.lock, flags);
4627 EXPORT_SYMBOL(completion_done);
4630 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4632 unsigned long flags;
4635 init_waitqueue_entry(&wait, current);
4637 __set_current_state(state);
4639 spin_lock_irqsave(&q->lock, flags);
4640 __add_wait_queue(q, &wait);
4641 spin_unlock(&q->lock);
4642 timeout = schedule_timeout(timeout);
4643 spin_lock_irq(&q->lock);
4644 __remove_wait_queue(q, &wait);
4645 spin_unlock_irqrestore(&q->lock, flags);
4650 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4652 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4654 EXPORT_SYMBOL(interruptible_sleep_on);
4657 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4659 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4661 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4663 void __sched sleep_on(wait_queue_head_t *q)
4665 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4667 EXPORT_SYMBOL(sleep_on);
4669 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4671 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4673 EXPORT_SYMBOL(sleep_on_timeout);
4675 #ifdef CONFIG_RT_MUTEXES
4678 * rt_mutex_setprio - set the current priority of a task
4680 * @prio: prio value (kernel-internal form)
4682 * This function changes the 'effective' priority of a task. It does
4683 * not touch ->normal_prio like __setscheduler().
4685 * Used by the rt_mutex code to implement priority inheritance logic.
4687 void rt_mutex_setprio(struct task_struct *p, int prio)
4689 unsigned long flags;
4690 int oldprio, on_rq, running;
4692 const struct sched_class *prev_class;
4694 BUG_ON(prio < 0 || prio > MAX_PRIO);
4696 lockdep_assert_held(&p->pi_lock);
4698 rq = task_rq_lock(p, &flags);
4700 trace_sched_pi_setprio(p, prio);
4702 prev_class = p->sched_class;
4704 running = task_current(rq, p);
4706 dequeue_task(rq, p, 0);
4708 p->sched_class->put_prev_task(rq, p);
4711 p->sched_class = &rt_sched_class;
4713 p->sched_class = &fair_sched_class;
4718 p->sched_class->set_curr_task(rq);
4720 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4722 check_class_changed(rq, p, prev_class, oldprio);
4723 task_rq_unlock(rq, &flags);
4728 void set_user_nice(struct task_struct *p, long nice)
4730 int old_prio, delta, on_rq;
4731 unsigned long flags;
4734 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4737 * We have to be careful, if called from sys_setpriority(),
4738 * the task might be in the middle of scheduling on another CPU.
4740 rq = task_rq_lock(p, &flags);
4742 * The RT priorities are set via sched_setscheduler(), but we still
4743 * allow the 'normal' nice value to be set - but as expected
4744 * it wont have any effect on scheduling until the task is
4745 * SCHED_FIFO/SCHED_RR:
4747 if (task_has_rt_policy(p)) {
4748 p->static_prio = NICE_TO_PRIO(nice);
4753 dequeue_task(rq, p, 0);
4755 p->static_prio = NICE_TO_PRIO(nice);
4758 p->prio = effective_prio(p);
4759 delta = p->prio - old_prio;
4762 enqueue_task(rq, p, 0);
4764 * If the task increased its priority or is running and
4765 * lowered its priority, then reschedule its CPU:
4767 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4768 resched_task(rq->curr);
4771 task_rq_unlock(rq, &flags);
4773 EXPORT_SYMBOL(set_user_nice);
4776 * can_nice - check if a task can reduce its nice value
4780 int can_nice(const struct task_struct *p, const int nice)
4782 /* convert nice value [19,-20] to rlimit style value [1,40] */
4783 int nice_rlim = 20 - nice;
4785 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4786 capable(CAP_SYS_NICE));
4789 #ifdef __ARCH_WANT_SYS_NICE
4792 * sys_nice - change the priority of the current process.
4793 * @increment: priority increment
4795 * sys_setpriority is a more generic, but much slower function that
4796 * does similar things.
4798 SYSCALL_DEFINE1(nice, int, increment)
4803 * Setpriority might change our priority at the same moment.
4804 * We don't have to worry. Conceptually one call occurs first
4805 * and we have a single winner.
4807 if (increment < -40)
4812 nice = TASK_NICE(current) + increment;
4818 if (increment < 0 && !can_nice(current, nice))
4821 retval = security_task_setnice(current, nice);
4825 set_user_nice(current, nice);
4832 * task_prio - return the priority value of a given task.
4833 * @p: the task in question.
4835 * This is the priority value as seen by users in /proc.
4836 * RT tasks are offset by -200. Normal tasks are centered
4837 * around 0, value goes from -16 to +15.
4839 int task_prio(const struct task_struct *p)
4841 return p->prio - MAX_RT_PRIO;
4845 * task_nice - return the nice value of a given task.
4846 * @p: the task in question.
4848 int task_nice(const struct task_struct *p)
4850 return TASK_NICE(p);
4852 EXPORT_SYMBOL(task_nice);
4855 * idle_cpu - is a given cpu idle currently?
4856 * @cpu: the processor in question.
4858 int idle_cpu(int cpu)
4860 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4864 * idle_task - return the idle task for a given cpu.
4865 * @cpu: the processor in question.
4867 struct task_struct *idle_task(int cpu)
4869 return cpu_rq(cpu)->idle;
4873 * find_process_by_pid - find a process with a matching PID value.
4874 * @pid: the pid in question.
4876 static struct task_struct *find_process_by_pid(pid_t pid)
4878 return pid ? find_task_by_vpid(pid) : current;
4881 /* Actually do priority change: must hold rq lock. */
4883 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4886 p->rt_priority = prio;
4887 p->normal_prio = normal_prio(p);
4888 /* we are holding p->pi_lock already */
4889 p->prio = rt_mutex_getprio(p);
4890 if (rt_prio(p->prio))
4891 p->sched_class = &rt_sched_class;
4893 p->sched_class = &fair_sched_class;
4898 * check the target process has a UID that matches the current process's
4900 static bool check_same_owner(struct task_struct *p)
4902 const struct cred *cred = current_cred(), *pcred;
4906 pcred = __task_cred(p);
4907 if (cred->user->user_ns == pcred->user->user_ns)
4908 match = (cred->euid == pcred->euid ||
4909 cred->euid == pcred->uid);
4916 static int __sched_setscheduler(struct task_struct *p, int policy,
4917 const struct sched_param *param, bool user)
4919 int retval, oldprio, oldpolicy = -1, on_rq, running;
4920 unsigned long flags;
4921 const struct sched_class *prev_class;
4925 /* may grab non-irq protected spin_locks */
4926 BUG_ON(in_interrupt());
4928 /* double check policy once rq lock held */
4930 reset_on_fork = p->sched_reset_on_fork;
4931 policy = oldpolicy = p->policy;
4933 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4934 policy &= ~SCHED_RESET_ON_FORK;
4936 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4937 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4938 policy != SCHED_IDLE)
4943 * Valid priorities for SCHED_FIFO and SCHED_RR are
4944 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4945 * SCHED_BATCH and SCHED_IDLE is 0.
4947 if (param->sched_priority < 0 ||
4948 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4949 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4951 if (rt_policy(policy) != (param->sched_priority != 0))
4955 * Allow unprivileged RT tasks to decrease priority:
4957 if (user && !capable(CAP_SYS_NICE)) {
4958 if (rt_policy(policy)) {
4959 unsigned long rlim_rtprio =
4960 task_rlimit(p, RLIMIT_RTPRIO);
4962 /* can't set/change the rt policy */
4963 if (policy != p->policy && !rlim_rtprio)
4966 /* can't increase priority */
4967 if (param->sched_priority > p->rt_priority &&
4968 param->sched_priority > rlim_rtprio)
4973 * Treat SCHED_IDLE as nice 20. Only allow a switch to
4974 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
4976 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE) {
4977 if (!can_nice(p, TASK_NICE(p)))
4981 /* can't change other user's priorities */
4982 if (!check_same_owner(p))
4985 /* Normal users shall not reset the sched_reset_on_fork flag */
4986 if (p->sched_reset_on_fork && !reset_on_fork)
4991 retval = security_task_setscheduler(p);
4997 * make sure no PI-waiters arrive (or leave) while we are
4998 * changing the priority of the task:
5000 raw_spin_lock_irqsave(&p->pi_lock, flags);
5002 * To be able to change p->policy safely, the appropriate
5003 * runqueue lock must be held.
5005 rq = __task_rq_lock(p);
5008 * Changing the policy of the stop threads its a very bad idea
5010 if (p == rq->stop) {
5011 __task_rq_unlock(rq);
5012 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5017 * If not changing anything there's no need to proceed further:
5019 if (unlikely(policy == p->policy && (!rt_policy(policy) ||
5020 param->sched_priority == p->rt_priority))) {
5022 __task_rq_unlock(rq);
5023 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5027 #ifdef CONFIG_RT_GROUP_SCHED
5030 * Do not allow realtime tasks into groups that have no runtime
5033 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5034 task_group(p)->rt_bandwidth.rt_runtime == 0 &&
5035 !task_group_is_autogroup(task_group(p))) {
5036 __task_rq_unlock(rq);
5037 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5043 /* recheck policy now with rq lock held */
5044 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5045 policy = oldpolicy = -1;
5046 __task_rq_unlock(rq);
5047 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5051 running = task_current(rq, p);
5053 deactivate_task(rq, p, 0);
5055 p->sched_class->put_prev_task(rq, p);
5057 p->sched_reset_on_fork = reset_on_fork;
5060 prev_class = p->sched_class;
5061 __setscheduler(rq, p, policy, param->sched_priority);
5064 p->sched_class->set_curr_task(rq);
5066 activate_task(rq, p, 0);
5068 check_class_changed(rq, p, prev_class, oldprio);
5069 __task_rq_unlock(rq);
5070 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5072 rt_mutex_adjust_pi(p);
5078 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5079 * @p: the task in question.
5080 * @policy: new policy.
5081 * @param: structure containing the new RT priority.
5083 * NOTE that the task may be already dead.
5085 int sched_setscheduler(struct task_struct *p, int policy,
5086 const struct sched_param *param)
5088 return __sched_setscheduler(p, policy, param, true);
5090 EXPORT_SYMBOL_GPL(sched_setscheduler);
5093 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5094 * @p: the task in question.
5095 * @policy: new policy.
5096 * @param: structure containing the new RT priority.
5098 * Just like sched_setscheduler, only don't bother checking if the
5099 * current context has permission. For example, this is needed in
5100 * stop_machine(): we create temporary high priority worker threads,
5101 * but our caller might not have that capability.
5103 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5104 const struct sched_param *param)
5106 return __sched_setscheduler(p, policy, param, false);
5110 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5112 struct sched_param lparam;
5113 struct task_struct *p;
5116 if (!param || pid < 0)
5118 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5123 p = find_process_by_pid(pid);
5125 retval = sched_setscheduler(p, policy, &lparam);
5132 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5133 * @pid: the pid in question.
5134 * @policy: new policy.
5135 * @param: structure containing the new RT priority.
5137 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5138 struct sched_param __user *, param)
5140 /* negative values for policy are not valid */
5144 return do_sched_setscheduler(pid, policy, param);
5148 * sys_sched_setparam - set/change the RT priority of a thread
5149 * @pid: the pid in question.
5150 * @param: structure containing the new RT priority.
5152 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5154 return do_sched_setscheduler(pid, -1, param);
5158 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5159 * @pid: the pid in question.
5161 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5163 struct task_struct *p;
5171 p = find_process_by_pid(pid);
5173 retval = security_task_getscheduler(p);
5176 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
5183 * sys_sched_getparam - get the RT priority of a thread
5184 * @pid: the pid in question.
5185 * @param: structure containing the RT priority.
5187 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5189 struct sched_param lp;
5190 struct task_struct *p;
5193 if (!param || pid < 0)
5197 p = find_process_by_pid(pid);
5202 retval = security_task_getscheduler(p);
5206 lp.sched_priority = p->rt_priority;
5210 * This one might sleep, we cannot do it with a spinlock held ...
5212 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5221 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5223 cpumask_var_t cpus_allowed, new_mask;
5224 struct task_struct *p;
5230 p = find_process_by_pid(pid);
5237 /* Prevent p going away */
5241 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5245 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5247 goto out_free_cpus_allowed;
5250 if (!check_same_owner(p) && !task_ns_capable(p, CAP_SYS_NICE))
5253 retval = security_task_setscheduler(p);
5257 cpuset_cpus_allowed(p, cpus_allowed);
5258 cpumask_and(new_mask, in_mask, cpus_allowed);
5260 retval = set_cpus_allowed_ptr(p, new_mask);
5263 cpuset_cpus_allowed(p, cpus_allowed);
5264 if (!cpumask_subset(new_mask, cpus_allowed)) {
5266 * We must have raced with a concurrent cpuset
5267 * update. Just reset the cpus_allowed to the
5268 * cpuset's cpus_allowed
5270 cpumask_copy(new_mask, cpus_allowed);
5275 free_cpumask_var(new_mask);
5276 out_free_cpus_allowed:
5277 free_cpumask_var(cpus_allowed);
5284 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5285 struct cpumask *new_mask)
5287 if (len < cpumask_size())
5288 cpumask_clear(new_mask);
5289 else if (len > cpumask_size())
5290 len = cpumask_size();
5292 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5296 * sys_sched_setaffinity - set the cpu affinity of a process
5297 * @pid: pid of the process
5298 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5299 * @user_mask_ptr: user-space pointer to the new cpu mask
5301 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5302 unsigned long __user *, user_mask_ptr)
5304 cpumask_var_t new_mask;
5307 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5310 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5312 retval = sched_setaffinity(pid, new_mask);
5313 free_cpumask_var(new_mask);
5317 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5319 struct task_struct *p;
5320 unsigned long flags;
5327 p = find_process_by_pid(pid);
5331 retval = security_task_getscheduler(p);
5335 raw_spin_lock_irqsave(&p->pi_lock, flags);
5336 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5337 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5347 * sys_sched_getaffinity - get the cpu affinity of a process
5348 * @pid: pid of the process
5349 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5350 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5352 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5353 unsigned long __user *, user_mask_ptr)
5358 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5360 if (len & (sizeof(unsigned long)-1))
5363 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5366 ret = sched_getaffinity(pid, mask);
5368 size_t retlen = min_t(size_t, len, cpumask_size());
5370 if (copy_to_user(user_mask_ptr, mask, retlen))
5375 free_cpumask_var(mask);
5381 * sys_sched_yield - yield the current processor to other threads.
5383 * This function yields the current CPU to other tasks. If there are no
5384 * other threads running on this CPU then this function will return.
5386 SYSCALL_DEFINE0(sched_yield)
5388 struct rq *rq = this_rq_lock();
5390 schedstat_inc(rq, yld_count);
5391 current->sched_class->yield_task(rq);
5394 * Since we are going to call schedule() anyway, there's
5395 * no need to preempt or enable interrupts:
5397 __release(rq->lock);
5398 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5399 do_raw_spin_unlock(&rq->lock);
5400 preempt_enable_no_resched();
5407 static inline int should_resched(void)
5409 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5412 static void __cond_resched(void)
5414 add_preempt_count(PREEMPT_ACTIVE);
5416 sub_preempt_count(PREEMPT_ACTIVE);
5419 int __sched _cond_resched(void)
5421 if (should_resched()) {
5427 EXPORT_SYMBOL(_cond_resched);
5430 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5431 * call schedule, and on return reacquire the lock.
5433 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5434 * operations here to prevent schedule() from being called twice (once via
5435 * spin_unlock(), once by hand).
5437 int __cond_resched_lock(spinlock_t *lock)
5439 int resched = should_resched();
5442 lockdep_assert_held(lock);
5444 if (spin_needbreak(lock) || resched) {
5455 EXPORT_SYMBOL(__cond_resched_lock);
5457 int __sched __cond_resched_softirq(void)
5459 BUG_ON(!in_softirq());
5461 if (should_resched()) {
5469 EXPORT_SYMBOL(__cond_resched_softirq);
5472 * yield - yield the current processor to other threads.
5474 * This is a shortcut for kernel-space yielding - it marks the
5475 * thread runnable and calls sys_sched_yield().
5477 void __sched yield(void)
5479 set_current_state(TASK_RUNNING);
5482 EXPORT_SYMBOL(yield);
5485 * yield_to - yield the current processor to another thread in
5486 * your thread group, or accelerate that thread toward the
5487 * processor it's on.
5489 * @preempt: whether task preemption is allowed or not
5491 * It's the caller's job to ensure that the target task struct
5492 * can't go away on us before we can do any checks.
5494 * Returns true if we indeed boosted the target task.
5496 bool __sched yield_to(struct task_struct *p, bool preempt)
5498 struct task_struct *curr = current;
5499 struct rq *rq, *p_rq;
5500 unsigned long flags;
5503 local_irq_save(flags);
5508 double_rq_lock(rq, p_rq);
5509 while (task_rq(p) != p_rq) {
5510 double_rq_unlock(rq, p_rq);
5514 if (!curr->sched_class->yield_to_task)
5517 if (curr->sched_class != p->sched_class)
5520 if (task_running(p_rq, p) || p->state)
5523 yielded = curr->sched_class->yield_to_task(rq, p, preempt);
5525 schedstat_inc(rq, yld_count);
5527 * Make p's CPU reschedule; pick_next_entity takes care of
5530 if (preempt && rq != p_rq)
5531 resched_task(p_rq->curr);
5535 double_rq_unlock(rq, p_rq);
5536 local_irq_restore(flags);
5543 EXPORT_SYMBOL_GPL(yield_to);
5546 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5547 * that process accounting knows that this is a task in IO wait state.
5549 void __sched io_schedule(void)
5551 struct rq *rq = raw_rq();
5553 delayacct_blkio_start();
5554 atomic_inc(&rq->nr_iowait);
5555 blk_flush_plug(current);
5556 current->in_iowait = 1;
5558 current->in_iowait = 0;
5559 atomic_dec(&rq->nr_iowait);
5560 delayacct_blkio_end();
5562 EXPORT_SYMBOL(io_schedule);
5564 long __sched io_schedule_timeout(long timeout)
5566 struct rq *rq = raw_rq();
5569 delayacct_blkio_start();
5570 atomic_inc(&rq->nr_iowait);
5571 blk_flush_plug(current);
5572 current->in_iowait = 1;
5573 ret = schedule_timeout(timeout);
5574 current->in_iowait = 0;
5575 atomic_dec(&rq->nr_iowait);
5576 delayacct_blkio_end();
5581 * sys_sched_get_priority_max - return maximum RT priority.
5582 * @policy: scheduling class.
5584 * this syscall returns the maximum rt_priority that can be used
5585 * by a given scheduling class.
5587 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5594 ret = MAX_USER_RT_PRIO-1;
5606 * sys_sched_get_priority_min - return minimum RT priority.
5607 * @policy: scheduling class.
5609 * this syscall returns the minimum rt_priority that can be used
5610 * by a given scheduling class.
5612 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5630 * sys_sched_rr_get_interval - return the default timeslice of a process.
5631 * @pid: pid of the process.
5632 * @interval: userspace pointer to the timeslice value.
5634 * this syscall writes the default timeslice value of a given process
5635 * into the user-space timespec buffer. A value of '0' means infinity.
5637 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5638 struct timespec __user *, interval)
5640 struct task_struct *p;
5641 unsigned int time_slice;
5642 unsigned long flags;
5652 p = find_process_by_pid(pid);
5656 retval = security_task_getscheduler(p);
5660 rq = task_rq_lock(p, &flags);
5661 time_slice = p->sched_class->get_rr_interval(rq, p);
5662 task_rq_unlock(rq, &flags);
5665 jiffies_to_timespec(time_slice, &t);
5666 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5674 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5676 void sched_show_task(struct task_struct *p)
5678 unsigned long free = 0;
5681 state = p->state ? __ffs(p->state) + 1 : 0;
5682 printk(KERN_INFO "%-15.15s %c", p->comm,
5683 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5684 #if BITS_PER_LONG == 32
5685 if (state == TASK_RUNNING)
5686 printk(KERN_CONT " running ");
5688 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5690 if (state == TASK_RUNNING)
5691 printk(KERN_CONT " running task ");
5693 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5695 #ifdef CONFIG_DEBUG_STACK_USAGE
5696 free = stack_not_used(p);
5698 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5699 task_pid_nr(p), task_pid_nr(p->real_parent),
5700 (unsigned long)task_thread_info(p)->flags);
5702 show_stack(p, NULL);
5705 void show_state_filter(unsigned long state_filter)
5707 struct task_struct *g, *p;
5709 #if BITS_PER_LONG == 32
5711 " task PC stack pid father\n");
5714 " task PC stack pid father\n");
5716 read_lock(&tasklist_lock);
5717 do_each_thread(g, p) {
5719 * reset the NMI-timeout, listing all files on a slow
5720 * console might take a lot of time:
5722 touch_nmi_watchdog();
5723 if (!state_filter || (p->state & state_filter))
5725 } while_each_thread(g, p);
5727 touch_all_softlockup_watchdogs();
5729 #ifdef CONFIG_SCHED_DEBUG
5730 sysrq_sched_debug_show();
5732 read_unlock(&tasklist_lock);
5734 * Only show locks if all tasks are dumped:
5737 debug_show_all_locks();
5740 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5742 idle->sched_class = &idle_sched_class;
5746 * init_idle - set up an idle thread for a given CPU
5747 * @idle: task in question
5748 * @cpu: cpu the idle task belongs to
5750 * NOTE: this function does not set the idle thread's NEED_RESCHED
5751 * flag, to make booting more robust.
5753 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5755 struct rq *rq = cpu_rq(cpu);
5756 unsigned long flags;
5758 raw_spin_lock_irqsave(&rq->lock, flags);
5761 idle->state = TASK_RUNNING;
5762 idle->se.exec_start = sched_clock();
5764 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5766 * We're having a chicken and egg problem, even though we are
5767 * holding rq->lock, the cpu isn't yet set to this cpu so the
5768 * lockdep check in task_group() will fail.
5770 * Similar case to sched_fork(). / Alternatively we could
5771 * use task_rq_lock() here and obtain the other rq->lock.
5776 __set_task_cpu(idle, cpu);
5779 rq->curr = rq->idle = idle;
5780 #if defined(CONFIG_SMP)
5783 raw_spin_unlock_irqrestore(&rq->lock, flags);
5785 /* Set the preempt count _outside_ the spinlocks! */
5786 #if defined(CONFIG_PREEMPT)
5787 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5789 task_thread_info(idle)->preempt_count = 0;
5792 * The idle tasks have their own, simple scheduling class:
5794 idle->sched_class = &idle_sched_class;
5795 ftrace_graph_init_idle_task(idle, cpu);
5799 * In a system that switches off the HZ timer nohz_cpu_mask
5800 * indicates which cpus entered this state. This is used
5801 * in the rcu update to wait only for active cpus. For system
5802 * which do not switch off the HZ timer nohz_cpu_mask should
5803 * always be CPU_BITS_NONE.
5805 cpumask_var_t nohz_cpu_mask;
5808 * Increase the granularity value when there are more CPUs,
5809 * because with more CPUs the 'effective latency' as visible
5810 * to users decreases. But the relationship is not linear,
5811 * so pick a second-best guess by going with the log2 of the
5814 * This idea comes from the SD scheduler of Con Kolivas:
5816 static int get_update_sysctl_factor(void)
5818 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5819 unsigned int factor;
5821 switch (sysctl_sched_tunable_scaling) {
5822 case SCHED_TUNABLESCALING_NONE:
5825 case SCHED_TUNABLESCALING_LINEAR:
5828 case SCHED_TUNABLESCALING_LOG:
5830 factor = 1 + ilog2(cpus);
5837 static void update_sysctl(void)
5839 unsigned int factor = get_update_sysctl_factor();
5841 #define SET_SYSCTL(name) \
5842 (sysctl_##name = (factor) * normalized_sysctl_##name)
5843 SET_SYSCTL(sched_min_granularity);
5844 SET_SYSCTL(sched_latency);
5845 SET_SYSCTL(sched_wakeup_granularity);
5849 static inline void sched_init_granularity(void)
5856 * This is how migration works:
5858 * 1) we invoke migration_cpu_stop() on the target CPU using
5860 * 2) stopper starts to run (implicitly forcing the migrated thread
5862 * 3) it checks whether the migrated task is still in the wrong runqueue.
5863 * 4) if it's in the wrong runqueue then the migration thread removes
5864 * it and puts it into the right queue.
5865 * 5) stopper completes and stop_one_cpu() returns and the migration
5870 * Change a given task's CPU affinity. Migrate the thread to a
5871 * proper CPU and schedule it away if the CPU it's executing on
5872 * is removed from the allowed bitmask.
5874 * NOTE: the caller must have a valid reference to the task, the
5875 * task must not exit() & deallocate itself prematurely. The
5876 * call is not atomic; no spinlocks may be held.
5878 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5880 unsigned long flags;
5882 unsigned int dest_cpu;
5885 raw_spin_lock_irqsave(&p->pi_lock, flags);
5886 rq = __task_rq_lock(p);
5888 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5893 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5894 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5899 if (p->sched_class->set_cpus_allowed)
5900 p->sched_class->set_cpus_allowed(p, new_mask);
5902 cpumask_copy(&p->cpus_allowed, new_mask);
5903 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5906 /* Can the task run on the task's current CPU? If so, we're done */
5907 if (cpumask_test_cpu(task_cpu(p), new_mask))
5910 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5911 if (need_migrate_task(p)) {
5912 struct migration_arg arg = { p, dest_cpu };
5913 /* Need help from migration thread: drop lock and wait. */
5914 __task_rq_unlock(rq);
5915 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5916 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5917 tlb_migrate_finish(p->mm);
5921 __task_rq_unlock(rq);
5922 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
5926 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5929 * Move (not current) task off this cpu, onto dest cpu. We're doing
5930 * this because either it can't run here any more (set_cpus_allowed()
5931 * away from this CPU, or CPU going down), or because we're
5932 * attempting to rebalance this task on exec (sched_exec).
5934 * So we race with normal scheduler movements, but that's OK, as long
5935 * as the task is no longer on this CPU.
5937 * Returns non-zero if task was successfully migrated.
5939 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5941 struct rq *rq_dest, *rq_src;
5944 if (unlikely(!cpu_active(dest_cpu)))
5947 rq_src = cpu_rq(src_cpu);
5948 rq_dest = cpu_rq(dest_cpu);
5950 double_rq_lock(rq_src, rq_dest);
5951 /* Already moved. */
5952 if (task_cpu(p) != src_cpu)
5954 /* Affinity changed (again). */
5955 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5959 * If we're not on a rq, the next wake-up will ensure we're
5963 deactivate_task(rq_src, p, 0);
5964 set_task_cpu(p, dest_cpu);
5965 activate_task(rq_dest, p, 0);
5966 check_preempt_curr(rq_dest, p, 0);
5971 double_rq_unlock(rq_src, rq_dest);
5976 * migration_cpu_stop - this will be executed by a highprio stopper thread
5977 * and performs thread migration by bumping thread off CPU then
5978 * 'pushing' onto another runqueue.
5980 static int migration_cpu_stop(void *data)
5982 struct migration_arg *arg = data;
5985 * The original target cpu might have gone down and we might
5986 * be on another cpu but it doesn't matter.
5988 local_irq_disable();
5989 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5994 #ifdef CONFIG_HOTPLUG_CPU
5997 * Ensures that the idle task is using init_mm right before its cpu goes
6000 void idle_task_exit(void)
6002 struct mm_struct *mm = current->active_mm;
6004 BUG_ON(cpu_online(smp_processor_id()));
6007 switch_mm(mm, &init_mm, current);
6012 * While a dead CPU has no uninterruptible tasks queued at this point,
6013 * it might still have a nonzero ->nr_uninterruptible counter, because
6014 * for performance reasons the counter is not stricly tracking tasks to
6015 * their home CPUs. So we just add the counter to another CPU's counter,
6016 * to keep the global sum constant after CPU-down:
6018 static void migrate_nr_uninterruptible(struct rq *rq_src)
6020 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
6022 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6023 rq_src->nr_uninterruptible = 0;
6027 * remove the tasks which were accounted by rq from calc_load_tasks.
6029 static void calc_global_load_remove(struct rq *rq)
6031 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
6032 rq->calc_load_active = 0;
6036 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6037 * try_to_wake_up()->select_task_rq().
6039 * Called with rq->lock held even though we'er in stop_machine() and
6040 * there's no concurrency possible, we hold the required locks anyway
6041 * because of lock validation efforts.
6043 static void migrate_tasks(unsigned int dead_cpu)
6045 struct rq *rq = cpu_rq(dead_cpu);
6046 struct task_struct *next, *stop = rq->stop;
6050 * Fudge the rq selection such that the below task selection loop
6051 * doesn't get stuck on the currently eligible stop task.
6053 * We're currently inside stop_machine() and the rq is either stuck
6054 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6055 * either way we should never end up calling schedule() until we're
6062 * There's this thread running, bail when that's the only
6065 if (rq->nr_running == 1)
6068 next = pick_next_task(rq);
6070 next->sched_class->put_prev_task(rq, next);
6072 /* Find suitable destination for @next, with force if needed. */
6073 dest_cpu = select_fallback_rq(dead_cpu, next);
6074 raw_spin_unlock(&rq->lock);
6076 __migrate_task(next, dead_cpu, dest_cpu);
6078 raw_spin_lock(&rq->lock);
6084 #endif /* CONFIG_HOTPLUG_CPU */
6086 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6088 static struct ctl_table sd_ctl_dir[] = {
6090 .procname = "sched_domain",
6096 static struct ctl_table sd_ctl_root[] = {
6098 .procname = "kernel",
6100 .child = sd_ctl_dir,
6105 static struct ctl_table *sd_alloc_ctl_entry(int n)
6107 struct ctl_table *entry =
6108 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6113 static void sd_free_ctl_entry(struct ctl_table **tablep)
6115 struct ctl_table *entry;
6118 * In the intermediate directories, both the child directory and
6119 * procname are dynamically allocated and could fail but the mode
6120 * will always be set. In the lowest directory the names are
6121 * static strings and all have proc handlers.
6123 for (entry = *tablep; entry->mode; entry++) {
6125 sd_free_ctl_entry(&entry->child);
6126 if (entry->proc_handler == NULL)
6127 kfree(entry->procname);
6135 set_table_entry(struct ctl_table *entry,
6136 const char *procname, void *data, int maxlen,
6137 mode_t mode, proc_handler *proc_handler)
6139 entry->procname = procname;
6141 entry->maxlen = maxlen;
6143 entry->proc_handler = proc_handler;
6146 static struct ctl_table *
6147 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6149 struct ctl_table *table = sd_alloc_ctl_entry(13);
6154 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6155 sizeof(long), 0644, proc_doulongvec_minmax);
6156 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6157 sizeof(long), 0644, proc_doulongvec_minmax);
6158 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6159 sizeof(int), 0644, proc_dointvec_minmax);
6160 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6161 sizeof(int), 0644, proc_dointvec_minmax);
6162 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6163 sizeof(int), 0644, proc_dointvec_minmax);
6164 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6165 sizeof(int), 0644, proc_dointvec_minmax);
6166 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6167 sizeof(int), 0644, proc_dointvec_minmax);
6168 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6169 sizeof(int), 0644, proc_dointvec_minmax);
6170 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6171 sizeof(int), 0644, proc_dointvec_minmax);
6172 set_table_entry(&table[9], "cache_nice_tries",
6173 &sd->cache_nice_tries,
6174 sizeof(int), 0644, proc_dointvec_minmax);
6175 set_table_entry(&table[10], "flags", &sd->flags,
6176 sizeof(int), 0644, proc_dointvec_minmax);
6177 set_table_entry(&table[11], "name", sd->name,
6178 CORENAME_MAX_SIZE, 0444, proc_dostring);
6179 /* &table[12] is terminator */
6184 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6186 struct ctl_table *entry, *table;
6187 struct sched_domain *sd;
6188 int domain_num = 0, i;
6191 for_each_domain(cpu, sd)
6193 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6198 for_each_domain(cpu, sd) {
6199 snprintf(buf, 32, "domain%d", i);
6200 entry->procname = kstrdup(buf, GFP_KERNEL);
6202 entry->child = sd_alloc_ctl_domain_table(sd);
6209 static struct ctl_table_header *sd_sysctl_header;
6210 static void register_sched_domain_sysctl(void)
6212 int i, cpu_num = num_possible_cpus();
6213 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6216 WARN_ON(sd_ctl_dir[0].child);
6217 sd_ctl_dir[0].child = entry;
6222 for_each_possible_cpu(i) {
6223 snprintf(buf, 32, "cpu%d", i);
6224 entry->procname = kstrdup(buf, GFP_KERNEL);
6226 entry->child = sd_alloc_ctl_cpu_table(i);
6230 WARN_ON(sd_sysctl_header);
6231 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6234 /* may be called multiple times per register */
6235 static void unregister_sched_domain_sysctl(void)
6237 if (sd_sysctl_header)
6238 unregister_sysctl_table(sd_sysctl_header);
6239 sd_sysctl_header = NULL;
6240 if (sd_ctl_dir[0].child)
6241 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6244 static void register_sched_domain_sysctl(void)
6247 static void unregister_sched_domain_sysctl(void)
6252 static void set_rq_online(struct rq *rq)
6255 const struct sched_class *class;
6257 cpumask_set_cpu(rq->cpu, rq->rd->online);
6260 for_each_class(class) {
6261 if (class->rq_online)
6262 class->rq_online(rq);
6267 static void set_rq_offline(struct rq *rq)
6270 const struct sched_class *class;
6272 for_each_class(class) {
6273 if (class->rq_offline)
6274 class->rq_offline(rq);
6277 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6283 * migration_call - callback that gets triggered when a CPU is added.
6284 * Here we can start up the necessary migration thread for the new CPU.
6286 static int __cpuinit
6287 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6289 int cpu = (long)hcpu;
6290 unsigned long flags;
6291 struct rq *rq = cpu_rq(cpu);
6293 switch (action & ~CPU_TASKS_FROZEN) {
6295 case CPU_UP_PREPARE:
6296 rq->calc_load_update = calc_load_update;
6300 /* Update our root-domain */
6301 raw_spin_lock_irqsave(&rq->lock, flags);
6303 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6307 raw_spin_unlock_irqrestore(&rq->lock, flags);
6310 #ifdef CONFIG_HOTPLUG_CPU
6312 /* Update our root-domain */
6313 raw_spin_lock_irqsave(&rq->lock, flags);
6315 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6319 BUG_ON(rq->nr_running != 1); /* the migration thread */
6320 raw_spin_unlock_irqrestore(&rq->lock, flags);
6322 migrate_nr_uninterruptible(rq);
6323 calc_global_load_remove(rq);
6328 update_max_interval();
6334 * Register at high priority so that task migration (migrate_all_tasks)
6335 * happens before everything else. This has to be lower priority than
6336 * the notifier in the perf_event subsystem, though.
6338 static struct notifier_block __cpuinitdata migration_notifier = {
6339 .notifier_call = migration_call,
6340 .priority = CPU_PRI_MIGRATION,
6343 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6344 unsigned long action, void *hcpu)
6346 switch (action & ~CPU_TASKS_FROZEN) {
6348 case CPU_DOWN_FAILED:
6349 set_cpu_active((long)hcpu, true);
6356 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6357 unsigned long action, void *hcpu)
6359 switch (action & ~CPU_TASKS_FROZEN) {
6360 case CPU_DOWN_PREPARE:
6361 set_cpu_active((long)hcpu, false);
6368 static int __init migration_init(void)
6370 void *cpu = (void *)(long)smp_processor_id();
6373 /* Initialize migration for the boot CPU */
6374 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6375 BUG_ON(err == NOTIFY_BAD);
6376 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6377 register_cpu_notifier(&migration_notifier);
6379 /* Register cpu active notifiers */
6380 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6381 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6385 early_initcall(migration_init);
6390 #ifdef CONFIG_SCHED_DEBUG
6392 static __read_mostly int sched_domain_debug_enabled;
6394 static int __init sched_domain_debug_setup(char *str)
6396 sched_domain_debug_enabled = 1;
6400 early_param("sched_debug", sched_domain_debug_setup);
6402 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6403 struct cpumask *groupmask)
6405 struct sched_group *group = sd->groups;
6408 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6409 cpumask_clear(groupmask);
6411 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6413 if (!(sd->flags & SD_LOAD_BALANCE)) {
6414 printk("does not load-balance\n");
6416 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6421 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6423 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6424 printk(KERN_ERR "ERROR: domain->span does not contain "
6427 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6428 printk(KERN_ERR "ERROR: domain->groups does not contain"
6432 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6436 printk(KERN_ERR "ERROR: group is NULL\n");
6440 if (!group->cpu_power) {
6441 printk(KERN_CONT "\n");
6442 printk(KERN_ERR "ERROR: domain->cpu_power not "
6447 if (!cpumask_weight(sched_group_cpus(group))) {
6448 printk(KERN_CONT "\n");
6449 printk(KERN_ERR "ERROR: empty group\n");
6453 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6454 printk(KERN_CONT "\n");
6455 printk(KERN_ERR "ERROR: repeated CPUs\n");
6459 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6461 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6463 printk(KERN_CONT " %s", str);
6464 if (group->cpu_power != SCHED_LOAD_SCALE) {
6465 printk(KERN_CONT " (cpu_power = %d)",
6469 group = group->next;
6470 } while (group != sd->groups);
6471 printk(KERN_CONT "\n");
6473 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6474 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6477 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6478 printk(KERN_ERR "ERROR: parent span is not a superset "
6479 "of domain->span\n");
6483 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6485 cpumask_var_t groupmask;
6488 if (!sched_domain_debug_enabled)
6492 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6496 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6498 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6499 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6504 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6511 free_cpumask_var(groupmask);
6513 #else /* !CONFIG_SCHED_DEBUG */
6514 # define sched_domain_debug(sd, cpu) do { } while (0)
6515 #endif /* CONFIG_SCHED_DEBUG */
6517 static int sd_degenerate(struct sched_domain *sd)
6519 if (cpumask_weight(sched_domain_span(sd)) == 1)
6522 /* Following flags need at least 2 groups */
6523 if (sd->flags & (SD_LOAD_BALANCE |
6524 SD_BALANCE_NEWIDLE |
6528 SD_SHARE_PKG_RESOURCES)) {
6529 if (sd->groups != sd->groups->next)
6533 /* Following flags don't use groups */
6534 if (sd->flags & (SD_WAKE_AFFINE))
6541 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6543 unsigned long cflags = sd->flags, pflags = parent->flags;
6545 if (sd_degenerate(parent))
6548 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6551 /* Flags needing groups don't count if only 1 group in parent */
6552 if (parent->groups == parent->groups->next) {
6553 pflags &= ~(SD_LOAD_BALANCE |
6554 SD_BALANCE_NEWIDLE |
6558 SD_SHARE_PKG_RESOURCES);
6559 if (nr_node_ids == 1)
6560 pflags &= ~SD_SERIALIZE;
6562 if (~cflags & pflags)
6568 static void free_rootdomain(struct root_domain *rd)
6570 synchronize_sched();
6572 cpupri_cleanup(&rd->cpupri);
6574 free_cpumask_var(rd->rto_mask);
6575 free_cpumask_var(rd->online);
6576 free_cpumask_var(rd->span);
6580 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6582 struct root_domain *old_rd = NULL;
6583 unsigned long flags;
6585 raw_spin_lock_irqsave(&rq->lock, flags);
6590 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6593 cpumask_clear_cpu(rq->cpu, old_rd->span);
6596 * If we dont want to free the old_rt yet then
6597 * set old_rd to NULL to skip the freeing later
6600 if (!atomic_dec_and_test(&old_rd->refcount))
6604 atomic_inc(&rd->refcount);
6607 cpumask_set_cpu(rq->cpu, rd->span);
6608 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6611 raw_spin_unlock_irqrestore(&rq->lock, flags);
6614 free_rootdomain(old_rd);
6617 static int init_rootdomain(struct root_domain *rd)
6619 memset(rd, 0, sizeof(*rd));
6621 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6623 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6625 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6628 if (cpupri_init(&rd->cpupri) != 0)
6633 free_cpumask_var(rd->rto_mask);
6635 free_cpumask_var(rd->online);
6637 free_cpumask_var(rd->span);
6642 static void init_defrootdomain(void)
6644 init_rootdomain(&def_root_domain);
6646 atomic_set(&def_root_domain.refcount, 1);
6649 static struct root_domain *alloc_rootdomain(void)
6651 struct root_domain *rd;
6653 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6657 if (init_rootdomain(rd) != 0) {
6666 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6667 * hold the hotplug lock.
6670 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6672 struct rq *rq = cpu_rq(cpu);
6673 struct sched_domain *tmp;
6675 for (tmp = sd; tmp; tmp = tmp->parent)
6676 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6678 /* Remove the sched domains which do not contribute to scheduling. */
6679 for (tmp = sd; tmp; ) {
6680 struct sched_domain *parent = tmp->parent;
6684 if (sd_parent_degenerate(tmp, parent)) {
6685 tmp->parent = parent->parent;
6687 parent->parent->child = tmp;
6692 if (sd && sd_degenerate(sd)) {
6698 sched_domain_debug(sd, cpu);
6700 rq_attach_root(rq, rd);
6701 rcu_assign_pointer(rq->sd, sd);
6704 /* cpus with isolated domains */
6705 static cpumask_var_t cpu_isolated_map;
6707 /* Setup the mask of cpus configured for isolated domains */
6708 static int __init isolated_cpu_setup(char *str)
6710 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6711 cpulist_parse(str, cpu_isolated_map);
6715 __setup("isolcpus=", isolated_cpu_setup);
6718 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6719 * to a function which identifies what group(along with sched group) a CPU
6720 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6721 * (due to the fact that we keep track of groups covered with a struct cpumask).
6723 * init_sched_build_groups will build a circular linked list of the groups
6724 * covered by the given span, and will set each group's ->cpumask correctly,
6725 * and ->cpu_power to 0.
6728 init_sched_build_groups(const struct cpumask *span,
6729 const struct cpumask *cpu_map,
6730 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6731 struct sched_group **sg,
6732 struct cpumask *tmpmask),
6733 struct cpumask *covered, struct cpumask *tmpmask)
6735 struct sched_group *first = NULL, *last = NULL;
6738 cpumask_clear(covered);
6740 for_each_cpu(i, span) {
6741 struct sched_group *sg;
6742 int group = group_fn(i, cpu_map, &sg, tmpmask);
6745 if (cpumask_test_cpu(i, covered))
6748 cpumask_clear(sched_group_cpus(sg));
6751 for_each_cpu(j, span) {
6752 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6755 cpumask_set_cpu(j, covered);
6756 cpumask_set_cpu(j, sched_group_cpus(sg));
6767 #define SD_NODES_PER_DOMAIN 16
6772 * find_next_best_node - find the next node to include in a sched_domain
6773 * @node: node whose sched_domain we're building
6774 * @used_nodes: nodes already in the sched_domain
6776 * Find the next node to include in a given scheduling domain. Simply
6777 * finds the closest node not already in the @used_nodes map.
6779 * Should use nodemask_t.
6781 static int find_next_best_node(int node, nodemask_t *used_nodes)
6783 int i, n, val, min_val, best_node = 0;
6787 for (i = 0; i < nr_node_ids; i++) {
6788 /* Start at @node */
6789 n = (node + i) % nr_node_ids;
6791 if (!nr_cpus_node(n))
6794 /* Skip already used nodes */
6795 if (node_isset(n, *used_nodes))
6798 /* Simple min distance search */
6799 val = node_distance(node, n);
6801 if (val < min_val) {
6807 node_set(best_node, *used_nodes);
6812 * sched_domain_node_span - get a cpumask for a node's sched_domain
6813 * @node: node whose cpumask we're constructing
6814 * @span: resulting cpumask
6816 * Given a node, construct a good cpumask for its sched_domain to span. It
6817 * should be one that prevents unnecessary balancing, but also spreads tasks
6820 static void sched_domain_node_span(int node, struct cpumask *span)
6822 nodemask_t used_nodes;
6825 cpumask_clear(span);
6826 nodes_clear(used_nodes);
6828 cpumask_or(span, span, cpumask_of_node(node));
6829 node_set(node, used_nodes);
6831 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6832 int next_node = find_next_best_node(node, &used_nodes);
6834 cpumask_or(span, span, cpumask_of_node(next_node));
6837 #endif /* CONFIG_NUMA */
6839 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6842 * The cpus mask in sched_group and sched_domain hangs off the end.
6844 * ( See the the comments in include/linux/sched.h:struct sched_group
6845 * and struct sched_domain. )
6847 struct static_sched_group {
6848 struct sched_group sg;
6849 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6852 struct static_sched_domain {
6853 struct sched_domain sd;
6854 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6860 cpumask_var_t domainspan;
6861 cpumask_var_t covered;
6862 cpumask_var_t notcovered;
6864 cpumask_var_t nodemask;
6865 cpumask_var_t this_sibling_map;
6866 cpumask_var_t this_core_map;
6867 cpumask_var_t this_book_map;
6868 cpumask_var_t send_covered;
6869 cpumask_var_t tmpmask;
6870 struct sched_group **sched_group_nodes;
6871 struct root_domain *rd;
6875 sa_sched_groups = 0,
6881 sa_this_sibling_map,
6883 sa_sched_group_nodes,
6893 * SMT sched-domains:
6895 #ifdef CONFIG_SCHED_SMT
6896 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6897 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6900 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6901 struct sched_group **sg, struct cpumask *unused)
6904 *sg = &per_cpu(sched_groups, cpu).sg;
6907 #endif /* CONFIG_SCHED_SMT */
6910 * multi-core sched-domains:
6912 #ifdef CONFIG_SCHED_MC
6913 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6914 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6917 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6918 struct sched_group **sg, struct cpumask *mask)
6921 #ifdef CONFIG_SCHED_SMT
6922 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6923 group = cpumask_first(mask);
6928 *sg = &per_cpu(sched_group_core, group).sg;
6931 #endif /* CONFIG_SCHED_MC */
6934 * book sched-domains:
6936 #ifdef CONFIG_SCHED_BOOK
6937 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6938 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6941 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6942 struct sched_group **sg, struct cpumask *mask)
6945 #ifdef CONFIG_SCHED_MC
6946 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6947 group = cpumask_first(mask);
6948 #elif defined(CONFIG_SCHED_SMT)
6949 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6950 group = cpumask_first(mask);
6953 *sg = &per_cpu(sched_group_book, group).sg;
6956 #endif /* CONFIG_SCHED_BOOK */
6958 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6959 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6962 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6963 struct sched_group **sg, struct cpumask *mask)
6966 #ifdef CONFIG_SCHED_BOOK
6967 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6968 group = cpumask_first(mask);
6969 #elif defined(CONFIG_SCHED_MC)
6970 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6971 group = cpumask_first(mask);
6972 #elif defined(CONFIG_SCHED_SMT)
6973 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6974 group = cpumask_first(mask);
6979 *sg = &per_cpu(sched_group_phys, group).sg;
6985 * The init_sched_build_groups can't handle what we want to do with node
6986 * groups, so roll our own. Now each node has its own list of groups which
6987 * gets dynamically allocated.
6989 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6990 static struct sched_group ***sched_group_nodes_bycpu;
6992 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6993 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6995 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6996 struct sched_group **sg,
6997 struct cpumask *nodemask)
7001 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7002 group = cpumask_first(nodemask);
7005 *sg = &per_cpu(sched_group_allnodes, group).sg;
7009 static void init_numa_sched_groups_power(struct sched_group *group_head)
7011 struct sched_group *sg = group_head;
7017 for_each_cpu(j, sched_group_cpus(sg)) {
7018 struct sched_domain *sd;
7020 sd = &per_cpu(phys_domains, j).sd;
7021 if (j != group_first_cpu(sd->groups)) {
7023 * Only add "power" once for each
7029 sg->cpu_power += sd->groups->cpu_power;
7032 } while (sg != group_head);
7035 static int build_numa_sched_groups(struct s_data *d,
7036 const struct cpumask *cpu_map, int num)
7038 struct sched_domain *sd;
7039 struct sched_group *sg, *prev;
7042 cpumask_clear(d->covered);
7043 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
7044 if (cpumask_empty(d->nodemask)) {
7045 d->sched_group_nodes[num] = NULL;
7049 sched_domain_node_span(num, d->domainspan);
7050 cpumask_and(d->domainspan, d->domainspan, cpu_map);
7052 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7055 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
7059 d->sched_group_nodes[num] = sg;
7061 for_each_cpu(j, d->nodemask) {
7062 sd = &per_cpu(node_domains, j).sd;
7067 cpumask_copy(sched_group_cpus(sg), d->nodemask);
7069 cpumask_or(d->covered, d->covered, d->nodemask);
7072 for (j = 0; j < nr_node_ids; j++) {
7073 n = (num + j) % nr_node_ids;
7074 cpumask_complement(d->notcovered, d->covered);
7075 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
7076 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
7077 if (cpumask_empty(d->tmpmask))
7079 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
7080 if (cpumask_empty(d->tmpmask))
7082 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7086 "Can not alloc domain group for node %d\n", j);
7090 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
7091 sg->next = prev->next;
7092 cpumask_or(d->covered, d->covered, d->tmpmask);
7099 #endif /* CONFIG_NUMA */
7102 /* Free memory allocated for various sched_group structures */
7103 static void free_sched_groups(const struct cpumask *cpu_map,
7104 struct cpumask *nodemask)
7108 for_each_cpu(cpu, cpu_map) {
7109 struct sched_group **sched_group_nodes
7110 = sched_group_nodes_bycpu[cpu];
7112 if (!sched_group_nodes)
7115 for (i = 0; i < nr_node_ids; i++) {
7116 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7118 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7119 if (cpumask_empty(nodemask))
7129 if (oldsg != sched_group_nodes[i])
7132 kfree(sched_group_nodes);
7133 sched_group_nodes_bycpu[cpu] = NULL;
7136 #else /* !CONFIG_NUMA */
7137 static void free_sched_groups(const struct cpumask *cpu_map,
7138 struct cpumask *nodemask)
7141 #endif /* CONFIG_NUMA */
7144 * Initialize sched groups cpu_power.
7146 * cpu_power indicates the capacity of sched group, which is used while
7147 * distributing the load between different sched groups in a sched domain.
7148 * Typically cpu_power for all the groups in a sched domain will be same unless
7149 * there are asymmetries in the topology. If there are asymmetries, group
7150 * having more cpu_power will pickup more load compared to the group having
7153 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7155 struct sched_domain *child;
7156 struct sched_group *group;
7160 WARN_ON(!sd || !sd->groups);
7162 if (cpu != group_first_cpu(sd->groups))
7165 sd->groups->group_weight = cpumask_weight(sched_group_cpus(sd->groups));
7169 sd->groups->cpu_power = 0;
7172 power = SCHED_LOAD_SCALE;
7173 weight = cpumask_weight(sched_domain_span(sd));
7175 * SMT siblings share the power of a single core.
7176 * Usually multiple threads get a better yield out of
7177 * that one core than a single thread would have,
7178 * reflect that in sd->smt_gain.
7180 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
7181 power *= sd->smt_gain;
7183 power >>= SCHED_LOAD_SHIFT;
7185 sd->groups->cpu_power += power;
7190 * Add cpu_power of each child group to this groups cpu_power.
7192 group = child->groups;
7194 sd->groups->cpu_power += group->cpu_power;
7195 group = group->next;
7196 } while (group != child->groups);
7200 * Initializers for schedule domains
7201 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7204 #ifdef CONFIG_SCHED_DEBUG
7205 # define SD_INIT_NAME(sd, type) sd->name = #type
7207 # define SD_INIT_NAME(sd, type) do { } while (0)
7210 #define SD_INIT(sd, type) sd_init_##type(sd)
7212 #define SD_INIT_FUNC(type) \
7213 static noinline void sd_init_##type(struct sched_domain *sd) \
7215 memset(sd, 0, sizeof(*sd)); \
7216 *sd = SD_##type##_INIT; \
7217 sd->level = SD_LV_##type; \
7218 SD_INIT_NAME(sd, type); \
7223 SD_INIT_FUNC(ALLNODES)
7226 #ifdef CONFIG_SCHED_SMT
7227 SD_INIT_FUNC(SIBLING)
7229 #ifdef CONFIG_SCHED_MC
7232 #ifdef CONFIG_SCHED_BOOK
7236 static int default_relax_domain_level = -1;
7238 static int __init setup_relax_domain_level(char *str)
7242 val = simple_strtoul(str, NULL, 0);
7243 if (val < SD_LV_MAX)
7244 default_relax_domain_level = val;
7248 __setup("relax_domain_level=", setup_relax_domain_level);
7250 static void set_domain_attribute(struct sched_domain *sd,
7251 struct sched_domain_attr *attr)
7255 if (!attr || attr->relax_domain_level < 0) {
7256 if (default_relax_domain_level < 0)
7259 request = default_relax_domain_level;
7261 request = attr->relax_domain_level;
7262 if (request < sd->level) {
7263 /* turn off idle balance on this domain */
7264 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7266 /* turn on idle balance on this domain */
7267 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7271 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7272 const struct cpumask *cpu_map)
7275 case sa_sched_groups:
7276 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7277 d->sched_group_nodes = NULL;
7279 free_rootdomain(d->rd); /* fall through */
7281 free_cpumask_var(d->tmpmask); /* fall through */
7282 case sa_send_covered:
7283 free_cpumask_var(d->send_covered); /* fall through */
7284 case sa_this_book_map:
7285 free_cpumask_var(d->this_book_map); /* fall through */
7286 case sa_this_core_map:
7287 free_cpumask_var(d->this_core_map); /* fall through */
7288 case sa_this_sibling_map:
7289 free_cpumask_var(d->this_sibling_map); /* fall through */
7291 free_cpumask_var(d->nodemask); /* fall through */
7292 case sa_sched_group_nodes:
7294 kfree(d->sched_group_nodes); /* fall through */
7296 free_cpumask_var(d->notcovered); /* fall through */
7298 free_cpumask_var(d->covered); /* fall through */
7300 free_cpumask_var(d->domainspan); /* fall through */
7307 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7308 const struct cpumask *cpu_map)
7311 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7313 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7314 return sa_domainspan;
7315 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7317 /* Allocate the per-node list of sched groups */
7318 d->sched_group_nodes = kcalloc(nr_node_ids,
7319 sizeof(struct sched_group *), GFP_KERNEL);
7320 if (!d->sched_group_nodes) {
7321 printk(KERN_WARNING "Can not alloc sched group node list\n");
7322 return sa_notcovered;
7324 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7326 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7327 return sa_sched_group_nodes;
7328 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7330 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7331 return sa_this_sibling_map;
7332 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
7333 return sa_this_core_map;
7334 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7335 return sa_this_book_map;
7336 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7337 return sa_send_covered;
7338 d->rd = alloc_rootdomain();
7340 printk(KERN_WARNING "Cannot alloc root domain\n");
7343 return sa_rootdomain;
7346 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7347 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7349 struct sched_domain *sd = NULL;
7351 struct sched_domain *parent;
7354 if (cpumask_weight(cpu_map) >
7355 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7356 sd = &per_cpu(allnodes_domains, i).sd;
7357 SD_INIT(sd, ALLNODES);
7358 set_domain_attribute(sd, attr);
7359 cpumask_copy(sched_domain_span(sd), cpu_map);
7360 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7365 sd = &per_cpu(node_domains, i).sd;
7367 set_domain_attribute(sd, attr);
7368 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7369 sd->parent = parent;
7372 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7377 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7378 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7379 struct sched_domain *parent, int i)
7381 struct sched_domain *sd;
7382 sd = &per_cpu(phys_domains, i).sd;
7384 set_domain_attribute(sd, attr);
7385 cpumask_copy(sched_domain_span(sd), d->nodemask);
7386 sd->parent = parent;
7389 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7393 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7394 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7395 struct sched_domain *parent, int i)
7397 struct sched_domain *sd = parent;
7398 #ifdef CONFIG_SCHED_BOOK
7399 sd = &per_cpu(book_domains, i).sd;
7401 set_domain_attribute(sd, attr);
7402 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7403 sd->parent = parent;
7405 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7410 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7411 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7412 struct sched_domain *parent, int i)
7414 struct sched_domain *sd = parent;
7415 #ifdef CONFIG_SCHED_MC
7416 sd = &per_cpu(core_domains, i).sd;
7418 set_domain_attribute(sd, attr);
7419 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7420 sd->parent = parent;
7422 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7427 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7428 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7429 struct sched_domain *parent, int i)
7431 struct sched_domain *sd = parent;
7432 #ifdef CONFIG_SCHED_SMT
7433 sd = &per_cpu(cpu_domains, i).sd;
7434 SD_INIT(sd, SIBLING);
7435 set_domain_attribute(sd, attr);
7436 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7437 sd->parent = parent;
7439 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7444 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7445 const struct cpumask *cpu_map, int cpu)
7448 #ifdef CONFIG_SCHED_SMT
7449 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7450 cpumask_and(d->this_sibling_map, cpu_map,
7451 topology_thread_cpumask(cpu));
7452 if (cpu == cpumask_first(d->this_sibling_map))
7453 init_sched_build_groups(d->this_sibling_map, cpu_map,
7455 d->send_covered, d->tmpmask);
7458 #ifdef CONFIG_SCHED_MC
7459 case SD_LV_MC: /* set up multi-core groups */
7460 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7461 if (cpu == cpumask_first(d->this_core_map))
7462 init_sched_build_groups(d->this_core_map, cpu_map,
7464 d->send_covered, d->tmpmask);
7467 #ifdef CONFIG_SCHED_BOOK
7468 case SD_LV_BOOK: /* set up book groups */
7469 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7470 if (cpu == cpumask_first(d->this_book_map))
7471 init_sched_build_groups(d->this_book_map, cpu_map,
7473 d->send_covered, d->tmpmask);
7476 case SD_LV_CPU: /* set up physical groups */
7477 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7478 if (!cpumask_empty(d->nodemask))
7479 init_sched_build_groups(d->nodemask, cpu_map,
7481 d->send_covered, d->tmpmask);
7484 case SD_LV_ALLNODES:
7485 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7486 d->send_covered, d->tmpmask);
7495 * Build sched domains for a given set of cpus and attach the sched domains
7496 * to the individual cpus
7498 static int __build_sched_domains(const struct cpumask *cpu_map,
7499 struct sched_domain_attr *attr)
7501 enum s_alloc alloc_state = sa_none;
7503 struct sched_domain *sd;
7509 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7510 if (alloc_state != sa_rootdomain)
7512 alloc_state = sa_sched_groups;
7515 * Set up domains for cpus specified by the cpu_map.
7517 for_each_cpu(i, cpu_map) {
7518 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7521 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7522 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7523 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7524 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7525 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7528 for_each_cpu(i, cpu_map) {
7529 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7530 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7531 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7534 /* Set up physical groups */
7535 for (i = 0; i < nr_node_ids; i++)
7536 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7539 /* Set up node groups */
7541 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7543 for (i = 0; i < nr_node_ids; i++)
7544 if (build_numa_sched_groups(&d, cpu_map, i))
7548 /* Calculate CPU power for physical packages and nodes */
7549 #ifdef CONFIG_SCHED_SMT
7550 for_each_cpu(i, cpu_map) {
7551 sd = &per_cpu(cpu_domains, i).sd;
7552 init_sched_groups_power(i, sd);
7555 #ifdef CONFIG_SCHED_MC
7556 for_each_cpu(i, cpu_map) {
7557 sd = &per_cpu(core_domains, i).sd;
7558 init_sched_groups_power(i, sd);
7561 #ifdef CONFIG_SCHED_BOOK
7562 for_each_cpu(i, cpu_map) {
7563 sd = &per_cpu(book_domains, i).sd;
7564 init_sched_groups_power(i, sd);
7568 for_each_cpu(i, cpu_map) {
7569 sd = &per_cpu(phys_domains, i).sd;
7570 init_sched_groups_power(i, sd);
7574 for (i = 0; i < nr_node_ids; i++)
7575 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7577 if (d.sd_allnodes) {
7578 struct sched_group *sg;
7580 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7582 init_numa_sched_groups_power(sg);
7586 /* Attach the domains */
7587 for_each_cpu(i, cpu_map) {
7588 #ifdef CONFIG_SCHED_SMT
7589 sd = &per_cpu(cpu_domains, i).sd;
7590 #elif defined(CONFIG_SCHED_MC)
7591 sd = &per_cpu(core_domains, i).sd;
7592 #elif defined(CONFIG_SCHED_BOOK)
7593 sd = &per_cpu(book_domains, i).sd;
7595 sd = &per_cpu(phys_domains, i).sd;
7597 cpu_attach_domain(sd, d.rd, i);
7600 d.sched_group_nodes = NULL; /* don't free this we still need it */
7601 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7605 __free_domain_allocs(&d, alloc_state, cpu_map);
7609 static int build_sched_domains(const struct cpumask *cpu_map)
7611 return __build_sched_domains(cpu_map, NULL);
7614 static cpumask_var_t *doms_cur; /* current sched domains */
7615 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7616 static struct sched_domain_attr *dattr_cur;
7617 /* attribues of custom domains in 'doms_cur' */
7620 * Special case: If a kmalloc of a doms_cur partition (array of
7621 * cpumask) fails, then fallback to a single sched domain,
7622 * as determined by the single cpumask fallback_doms.
7624 static cpumask_var_t fallback_doms;
7627 * arch_update_cpu_topology lets virtualized architectures update the
7628 * cpu core maps. It is supposed to return 1 if the topology changed
7629 * or 0 if it stayed the same.
7631 int __attribute__((weak)) arch_update_cpu_topology(void)
7636 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7639 cpumask_var_t *doms;
7641 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7644 for (i = 0; i < ndoms; i++) {
7645 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7646 free_sched_domains(doms, i);
7653 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7656 for (i = 0; i < ndoms; i++)
7657 free_cpumask_var(doms[i]);
7662 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7663 * For now this just excludes isolated cpus, but could be used to
7664 * exclude other special cases in the future.
7666 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7670 arch_update_cpu_topology();
7672 doms_cur = alloc_sched_domains(ndoms_cur);
7674 doms_cur = &fallback_doms;
7675 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7677 err = build_sched_domains(doms_cur[0]);
7678 register_sched_domain_sysctl();
7683 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7684 struct cpumask *tmpmask)
7686 free_sched_groups(cpu_map, tmpmask);
7690 * Detach sched domains from a group of cpus specified in cpu_map
7691 * These cpus will now be attached to the NULL domain
7693 static void detach_destroy_domains(const struct cpumask *cpu_map)
7695 /* Save because hotplug lock held. */
7696 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7699 for_each_cpu(i, cpu_map)
7700 cpu_attach_domain(NULL, &def_root_domain, i);
7701 synchronize_sched();
7702 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7705 /* handle null as "default" */
7706 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7707 struct sched_domain_attr *new, int idx_new)
7709 struct sched_domain_attr tmp;
7716 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7717 new ? (new + idx_new) : &tmp,
7718 sizeof(struct sched_domain_attr));
7722 * Partition sched domains as specified by the 'ndoms_new'
7723 * cpumasks in the array doms_new[] of cpumasks. This compares
7724 * doms_new[] to the current sched domain partitioning, doms_cur[].
7725 * It destroys each deleted domain and builds each new domain.
7727 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7728 * The masks don't intersect (don't overlap.) We should setup one
7729 * sched domain for each mask. CPUs not in any of the cpumasks will
7730 * not be load balanced. If the same cpumask appears both in the
7731 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7734 * The passed in 'doms_new' should be allocated using
7735 * alloc_sched_domains. This routine takes ownership of it and will
7736 * free_sched_domains it when done with it. If the caller failed the
7737 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7738 * and partition_sched_domains() will fallback to the single partition
7739 * 'fallback_doms', it also forces the domains to be rebuilt.
7741 * If doms_new == NULL it will be replaced with cpu_online_mask.
7742 * ndoms_new == 0 is a special case for destroying existing domains,
7743 * and it will not create the default domain.
7745 * Call with hotplug lock held
7747 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7748 struct sched_domain_attr *dattr_new)
7753 mutex_lock(&sched_domains_mutex);
7755 /* always unregister in case we don't destroy any domains */
7756 unregister_sched_domain_sysctl();
7758 /* Let architecture update cpu core mappings. */
7759 new_topology = arch_update_cpu_topology();
7761 n = doms_new ? ndoms_new : 0;
7763 /* Destroy deleted domains */
7764 for (i = 0; i < ndoms_cur; i++) {
7765 for (j = 0; j < n && !new_topology; j++) {
7766 if (cpumask_equal(doms_cur[i], doms_new[j])
7767 && dattrs_equal(dattr_cur, i, dattr_new, j))
7770 /* no match - a current sched domain not in new doms_new[] */
7771 detach_destroy_domains(doms_cur[i]);
7776 if (doms_new == NULL) {
7778 doms_new = &fallback_doms;
7779 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7780 WARN_ON_ONCE(dattr_new);
7783 /* Build new domains */
7784 for (i = 0; i < ndoms_new; i++) {
7785 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7786 if (cpumask_equal(doms_new[i], doms_cur[j])
7787 && dattrs_equal(dattr_new, i, dattr_cur, j))
7790 /* no match - add a new doms_new */
7791 __build_sched_domains(doms_new[i],
7792 dattr_new ? dattr_new + i : NULL);
7797 /* Remember the new sched domains */
7798 if (doms_cur != &fallback_doms)
7799 free_sched_domains(doms_cur, ndoms_cur);
7800 kfree(dattr_cur); /* kfree(NULL) is safe */
7801 doms_cur = doms_new;
7802 dattr_cur = dattr_new;
7803 ndoms_cur = ndoms_new;
7805 register_sched_domain_sysctl();
7807 mutex_unlock(&sched_domains_mutex);
7810 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7811 static void arch_reinit_sched_domains(void)
7815 /* Destroy domains first to force the rebuild */
7816 partition_sched_domains(0, NULL, NULL);
7818 rebuild_sched_domains();
7822 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7824 unsigned int level = 0;
7826 if (sscanf(buf, "%u", &level) != 1)
7830 * level is always be positive so don't check for
7831 * level < POWERSAVINGS_BALANCE_NONE which is 0
7832 * What happens on 0 or 1 byte write,
7833 * need to check for count as well?
7836 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7840 sched_smt_power_savings = level;
7842 sched_mc_power_savings = level;
7844 arch_reinit_sched_domains();
7849 #ifdef CONFIG_SCHED_MC
7850 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7851 struct sysdev_class_attribute *attr,
7854 return sprintf(page, "%u\n", sched_mc_power_savings);
7856 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7857 struct sysdev_class_attribute *attr,
7858 const char *buf, size_t count)
7860 return sched_power_savings_store(buf, count, 0);
7862 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7863 sched_mc_power_savings_show,
7864 sched_mc_power_savings_store);
7867 #ifdef CONFIG_SCHED_SMT
7868 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7869 struct sysdev_class_attribute *attr,
7872 return sprintf(page, "%u\n", sched_smt_power_savings);
7874 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7875 struct sysdev_class_attribute *attr,
7876 const char *buf, size_t count)
7878 return sched_power_savings_store(buf, count, 1);
7880 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7881 sched_smt_power_savings_show,
7882 sched_smt_power_savings_store);
7885 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7889 #ifdef CONFIG_SCHED_SMT
7891 err = sysfs_create_file(&cls->kset.kobj,
7892 &attr_sched_smt_power_savings.attr);
7894 #ifdef CONFIG_SCHED_MC
7895 if (!err && mc_capable())
7896 err = sysfs_create_file(&cls->kset.kobj,
7897 &attr_sched_mc_power_savings.attr);
7901 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7904 * Update cpusets according to cpu_active mask. If cpusets are
7905 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7906 * around partition_sched_domains().
7908 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7911 switch (action & ~CPU_TASKS_FROZEN) {
7913 case CPU_DOWN_FAILED:
7914 cpuset_update_active_cpus();
7921 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7924 switch (action & ~CPU_TASKS_FROZEN) {
7925 case CPU_DOWN_PREPARE:
7926 cpuset_update_active_cpus();
7933 static int update_runtime(struct notifier_block *nfb,
7934 unsigned long action, void *hcpu)
7936 int cpu = (int)(long)hcpu;
7939 case CPU_DOWN_PREPARE:
7940 case CPU_DOWN_PREPARE_FROZEN:
7941 disable_runtime(cpu_rq(cpu));
7944 case CPU_DOWN_FAILED:
7945 case CPU_DOWN_FAILED_FROZEN:
7947 case CPU_ONLINE_FROZEN:
7948 enable_runtime(cpu_rq(cpu));
7956 void __init sched_init_smp(void)
7958 cpumask_var_t non_isolated_cpus;
7960 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7961 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7963 #if defined(CONFIG_NUMA)
7964 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7966 BUG_ON(sched_group_nodes_bycpu == NULL);
7969 mutex_lock(&sched_domains_mutex);
7970 arch_init_sched_domains(cpu_active_mask);
7971 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7972 if (cpumask_empty(non_isolated_cpus))
7973 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7974 mutex_unlock(&sched_domains_mutex);
7977 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7978 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7980 /* RT runtime code needs to handle some hotplug events */
7981 hotcpu_notifier(update_runtime, 0);
7985 /* Move init over to a non-isolated CPU */
7986 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7988 sched_init_granularity();
7989 free_cpumask_var(non_isolated_cpus);
7991 init_sched_rt_class();
7994 void __init sched_init_smp(void)
7996 sched_init_granularity();
7998 #endif /* CONFIG_SMP */
8000 const_debug unsigned int sysctl_timer_migration = 1;
8002 int in_sched_functions(unsigned long addr)
8004 return in_lock_functions(addr) ||
8005 (addr >= (unsigned long)__sched_text_start
8006 && addr < (unsigned long)__sched_text_end);
8009 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8011 cfs_rq->tasks_timeline = RB_ROOT;
8012 INIT_LIST_HEAD(&cfs_rq->tasks);
8013 #ifdef CONFIG_FAIR_GROUP_SCHED
8015 /* allow initial update_cfs_load() to truncate */
8017 cfs_rq->load_stamp = 1;
8020 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8023 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8025 struct rt_prio_array *array;
8028 array = &rt_rq->active;
8029 for (i = 0; i < MAX_RT_PRIO; i++) {
8030 INIT_LIST_HEAD(array->queue + i);
8031 __clear_bit(i, array->bitmap);
8033 /* delimiter for bitsearch: */
8034 __set_bit(MAX_RT_PRIO, array->bitmap);
8036 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8037 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8039 rt_rq->highest_prio.next = MAX_RT_PRIO;
8043 rt_rq->rt_nr_migratory = 0;
8044 rt_rq->overloaded = 0;
8045 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
8049 rt_rq->rt_throttled = 0;
8050 rt_rq->rt_runtime = 0;
8051 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
8053 #ifdef CONFIG_RT_GROUP_SCHED
8054 rt_rq->rt_nr_boosted = 0;
8059 #ifdef CONFIG_FAIR_GROUP_SCHED
8060 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8061 struct sched_entity *se, int cpu,
8062 struct sched_entity *parent)
8064 struct rq *rq = cpu_rq(cpu);
8065 tg->cfs_rq[cpu] = cfs_rq;
8066 init_cfs_rq(cfs_rq, rq);
8070 /* se could be NULL for root_task_group */
8075 se->cfs_rq = &rq->cfs;
8077 se->cfs_rq = parent->my_q;
8080 update_load_set(&se->load, 0);
8081 se->parent = parent;
8085 #ifdef CONFIG_RT_GROUP_SCHED
8086 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8087 struct sched_rt_entity *rt_se, int cpu,
8088 struct sched_rt_entity *parent)
8090 struct rq *rq = cpu_rq(cpu);
8092 tg->rt_rq[cpu] = rt_rq;
8093 init_rt_rq(rt_rq, rq);
8095 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8097 tg->rt_se[cpu] = rt_se;
8102 rt_se->rt_rq = &rq->rt;
8104 rt_se->rt_rq = parent->my_q;
8106 rt_se->my_q = rt_rq;
8107 rt_se->parent = parent;
8108 INIT_LIST_HEAD(&rt_se->run_list);
8112 void __init sched_init(void)
8115 unsigned long alloc_size = 0, ptr;
8117 #ifdef CONFIG_FAIR_GROUP_SCHED
8118 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8120 #ifdef CONFIG_RT_GROUP_SCHED
8121 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8123 #ifdef CONFIG_CPUMASK_OFFSTACK
8124 alloc_size += num_possible_cpus() * cpumask_size();
8127 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
8129 #ifdef CONFIG_FAIR_GROUP_SCHED
8130 root_task_group.se = (struct sched_entity **)ptr;
8131 ptr += nr_cpu_ids * sizeof(void **);
8133 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8134 ptr += nr_cpu_ids * sizeof(void **);
8136 #endif /* CONFIG_FAIR_GROUP_SCHED */
8137 #ifdef CONFIG_RT_GROUP_SCHED
8138 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8139 ptr += nr_cpu_ids * sizeof(void **);
8141 root_task_group.rt_rq = (struct rt_rq **)ptr;
8142 ptr += nr_cpu_ids * sizeof(void **);
8144 #endif /* CONFIG_RT_GROUP_SCHED */
8145 #ifdef CONFIG_CPUMASK_OFFSTACK
8146 for_each_possible_cpu(i) {
8147 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
8148 ptr += cpumask_size();
8150 #endif /* CONFIG_CPUMASK_OFFSTACK */
8154 init_defrootdomain();
8157 init_rt_bandwidth(&def_rt_bandwidth,
8158 global_rt_period(), global_rt_runtime());
8160 #ifdef CONFIG_RT_GROUP_SCHED
8161 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8162 global_rt_period(), global_rt_runtime());
8163 #endif /* CONFIG_RT_GROUP_SCHED */
8165 #ifdef CONFIG_CGROUP_SCHED
8166 list_add(&root_task_group.list, &task_groups);
8167 INIT_LIST_HEAD(&root_task_group.children);
8168 autogroup_init(&init_task);
8169 #endif /* CONFIG_CGROUP_SCHED */
8171 for_each_possible_cpu(i) {
8175 raw_spin_lock_init(&rq->lock);
8177 rq->calc_load_active = 0;
8178 rq->calc_load_update = jiffies + LOAD_FREQ;
8179 init_cfs_rq(&rq->cfs, rq);
8180 init_rt_rq(&rq->rt, rq);
8181 #ifdef CONFIG_FAIR_GROUP_SCHED
8182 root_task_group.shares = root_task_group_load;
8183 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8185 * How much cpu bandwidth does root_task_group get?
8187 * In case of task-groups formed thr' the cgroup filesystem, it
8188 * gets 100% of the cpu resources in the system. This overall
8189 * system cpu resource is divided among the tasks of
8190 * root_task_group and its child task-groups in a fair manner,
8191 * based on each entity's (task or task-group's) weight
8192 * (se->load.weight).
8194 * In other words, if root_task_group has 10 tasks of weight
8195 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8196 * then A0's share of the cpu resource is:
8198 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8200 * We achieve this by letting root_task_group's tasks sit
8201 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8203 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, NULL);
8204 #endif /* CONFIG_FAIR_GROUP_SCHED */
8206 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8207 #ifdef CONFIG_RT_GROUP_SCHED
8208 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8209 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, NULL);
8212 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8213 rq->cpu_load[j] = 0;
8215 rq->last_load_update_tick = jiffies;
8220 rq->cpu_power = SCHED_LOAD_SCALE;
8221 rq->post_schedule = 0;
8222 rq->active_balance = 0;
8223 rq->next_balance = jiffies;
8228 rq->avg_idle = 2*sysctl_sched_migration_cost;
8229 rq_attach_root(rq, &def_root_domain);
8231 rq->nohz_balance_kick = 0;
8232 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
8236 atomic_set(&rq->nr_iowait, 0);
8239 set_load_weight(&init_task);
8241 #ifdef CONFIG_PREEMPT_NOTIFIERS
8242 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8246 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8249 #ifdef CONFIG_RT_MUTEXES
8250 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
8254 * The boot idle thread does lazy MMU switching as well:
8256 atomic_inc(&init_mm.mm_count);
8257 enter_lazy_tlb(&init_mm, current);
8260 * Make us the idle thread. Technically, schedule() should not be
8261 * called from this thread, however somewhere below it might be,
8262 * but because we are the idle thread, we just pick up running again
8263 * when this runqueue becomes "idle".
8265 init_idle(current, smp_processor_id());
8267 calc_load_update = jiffies + LOAD_FREQ;
8270 * During early bootup we pretend to be a normal task:
8272 current->sched_class = &fair_sched_class;
8274 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8275 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8278 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8279 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8280 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8281 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8282 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8284 /* May be allocated at isolcpus cmdline parse time */
8285 if (cpu_isolated_map == NULL)
8286 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8289 scheduler_running = 1;
8292 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8293 static inline int preempt_count_equals(int preempt_offset)
8295 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8297 return (nested == preempt_offset);
8300 void __might_sleep(const char *file, int line, int preempt_offset)
8303 static unsigned long prev_jiffy; /* ratelimiting */
8305 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8306 system_state != SYSTEM_RUNNING || oops_in_progress)
8308 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8310 prev_jiffy = jiffies;
8313 "BUG: sleeping function called from invalid context at %s:%d\n",
8316 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8317 in_atomic(), irqs_disabled(),
8318 current->pid, current->comm);
8320 debug_show_held_locks(current);
8321 if (irqs_disabled())
8322 print_irqtrace_events(current);
8326 EXPORT_SYMBOL(__might_sleep);
8329 #ifdef CONFIG_MAGIC_SYSRQ
8330 static void normalize_task(struct rq *rq, struct task_struct *p)
8332 const struct sched_class *prev_class = p->sched_class;
8333 int old_prio = p->prio;
8338 deactivate_task(rq, p, 0);
8339 __setscheduler(rq, p, SCHED_NORMAL, 0);
8341 activate_task(rq, p, 0);
8342 resched_task(rq->curr);
8345 check_class_changed(rq, p, prev_class, old_prio);
8348 void normalize_rt_tasks(void)
8350 struct task_struct *g, *p;
8351 unsigned long flags;
8354 read_lock_irqsave(&tasklist_lock, flags);
8355 do_each_thread(g, p) {
8357 * Only normalize user tasks:
8362 p->se.exec_start = 0;
8363 #ifdef CONFIG_SCHEDSTATS
8364 p->se.statistics.wait_start = 0;
8365 p->se.statistics.sleep_start = 0;
8366 p->se.statistics.block_start = 0;
8371 * Renice negative nice level userspace
8374 if (TASK_NICE(p) < 0 && p->mm)
8375 set_user_nice(p, 0);
8379 raw_spin_lock(&p->pi_lock);
8380 rq = __task_rq_lock(p);
8382 normalize_task(rq, p);
8384 __task_rq_unlock(rq);
8385 raw_spin_unlock(&p->pi_lock);
8386 } while_each_thread(g, p);
8388 read_unlock_irqrestore(&tasklist_lock, flags);
8391 #endif /* CONFIG_MAGIC_SYSRQ */
8393 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8395 * These functions are only useful for the IA64 MCA handling, or kdb.
8397 * They can only be called when the whole system has been
8398 * stopped - every CPU needs to be quiescent, and no scheduling
8399 * activity can take place. Using them for anything else would
8400 * be a serious bug, and as a result, they aren't even visible
8401 * under any other configuration.
8405 * curr_task - return the current task for a given cpu.
8406 * @cpu: the processor in question.
8408 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8410 struct task_struct *curr_task(int cpu)
8412 return cpu_curr(cpu);
8415 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8419 * set_curr_task - set the current task for a given cpu.
8420 * @cpu: the processor in question.
8421 * @p: the task pointer to set.
8423 * Description: This function must only be used when non-maskable interrupts
8424 * are serviced on a separate stack. It allows the architecture to switch the
8425 * notion of the current task on a cpu in a non-blocking manner. This function
8426 * must be called with all CPU's synchronized, and interrupts disabled, the
8427 * and caller must save the original value of the current task (see
8428 * curr_task() above) and restore that value before reenabling interrupts and
8429 * re-starting the system.
8431 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8433 void set_curr_task(int cpu, struct task_struct *p)
8440 #ifdef CONFIG_FAIR_GROUP_SCHED
8441 static void free_fair_sched_group(struct task_group *tg)
8445 for_each_possible_cpu(i) {
8447 kfree(tg->cfs_rq[i]);
8457 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8459 struct cfs_rq *cfs_rq;
8460 struct sched_entity *se;
8463 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8466 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8470 tg->shares = NICE_0_LOAD;
8472 for_each_possible_cpu(i) {
8473 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8474 GFP_KERNEL, cpu_to_node(i));
8478 se = kzalloc_node(sizeof(struct sched_entity),
8479 GFP_KERNEL, cpu_to_node(i));
8483 init_tg_cfs_entry(tg, cfs_rq, se, i, parent->se[i]);
8494 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8496 struct rq *rq = cpu_rq(cpu);
8497 unsigned long flags;
8500 * Only empty task groups can be destroyed; so we can speculatively
8501 * check on_list without danger of it being re-added.
8503 if (!tg->cfs_rq[cpu]->on_list)
8506 raw_spin_lock_irqsave(&rq->lock, flags);
8507 list_del_leaf_cfs_rq(tg->cfs_rq[cpu]);
8508 raw_spin_unlock_irqrestore(&rq->lock, flags);
8510 #else /* !CONFG_FAIR_GROUP_SCHED */
8511 static inline void free_fair_sched_group(struct task_group *tg)
8516 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8521 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8524 #endif /* CONFIG_FAIR_GROUP_SCHED */
8526 #ifdef CONFIG_RT_GROUP_SCHED
8527 static void free_rt_sched_group(struct task_group *tg)
8531 destroy_rt_bandwidth(&tg->rt_bandwidth);
8533 for_each_possible_cpu(i) {
8535 kfree(tg->rt_rq[i]);
8537 kfree(tg->rt_se[i]);
8545 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8547 struct rt_rq *rt_rq;
8548 struct sched_rt_entity *rt_se;
8552 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8555 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8559 init_rt_bandwidth(&tg->rt_bandwidth,
8560 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8562 for_each_possible_cpu(i) {
8565 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8566 GFP_KERNEL, cpu_to_node(i));
8570 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8571 GFP_KERNEL, cpu_to_node(i));
8575 init_tg_rt_entry(tg, rt_rq, rt_se, i, parent->rt_se[i]);
8585 #else /* !CONFIG_RT_GROUP_SCHED */
8586 static inline void free_rt_sched_group(struct task_group *tg)
8591 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8595 #endif /* CONFIG_RT_GROUP_SCHED */
8597 #ifdef CONFIG_CGROUP_SCHED
8598 static void free_sched_group(struct task_group *tg)
8600 free_fair_sched_group(tg);
8601 free_rt_sched_group(tg);
8606 /* allocate runqueue etc for a new task group */
8607 struct task_group *sched_create_group(struct task_group *parent)
8609 struct task_group *tg;
8610 unsigned long flags;
8612 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8614 return ERR_PTR(-ENOMEM);
8616 if (!alloc_fair_sched_group(tg, parent))
8619 if (!alloc_rt_sched_group(tg, parent))
8622 spin_lock_irqsave(&task_group_lock, flags);
8623 list_add_rcu(&tg->list, &task_groups);
8625 WARN_ON(!parent); /* root should already exist */
8627 tg->parent = parent;
8628 INIT_LIST_HEAD(&tg->children);
8629 list_add_rcu(&tg->siblings, &parent->children);
8630 spin_unlock_irqrestore(&task_group_lock, flags);
8635 free_sched_group(tg);
8636 return ERR_PTR(-ENOMEM);
8639 /* rcu callback to free various structures associated with a task group */
8640 static void free_sched_group_rcu(struct rcu_head *rhp)
8642 /* now it should be safe to free those cfs_rqs */
8643 free_sched_group(container_of(rhp, struct task_group, rcu));
8646 /* Destroy runqueue etc associated with a task group */
8647 void sched_destroy_group(struct task_group *tg)
8649 unsigned long flags;
8652 /* end participation in shares distribution */
8653 for_each_possible_cpu(i)
8654 unregister_fair_sched_group(tg, i);
8656 spin_lock_irqsave(&task_group_lock, flags);
8657 list_del_rcu(&tg->list);
8658 list_del_rcu(&tg->siblings);
8659 spin_unlock_irqrestore(&task_group_lock, flags);
8661 /* wait for possible concurrent references to cfs_rqs complete */
8662 call_rcu(&tg->rcu, free_sched_group_rcu);
8665 /* change task's runqueue when it moves between groups.
8666 * The caller of this function should have put the task in its new group
8667 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8668 * reflect its new group.
8670 void sched_move_task(struct task_struct *tsk)
8673 unsigned long flags;
8676 rq = task_rq_lock(tsk, &flags);
8678 running = task_current(rq, tsk);
8682 dequeue_task(rq, tsk, 0);
8683 if (unlikely(running))
8684 tsk->sched_class->put_prev_task(rq, tsk);
8686 #ifdef CONFIG_FAIR_GROUP_SCHED
8687 if (tsk->sched_class->task_move_group)
8688 tsk->sched_class->task_move_group(tsk, on_rq);
8691 set_task_rq(tsk, task_cpu(tsk));
8693 if (unlikely(running))
8694 tsk->sched_class->set_curr_task(rq);
8696 enqueue_task(rq, tsk, 0);
8698 task_rq_unlock(rq, &flags);
8700 #endif /* CONFIG_CGROUP_SCHED */
8702 #ifdef CONFIG_FAIR_GROUP_SCHED
8703 static DEFINE_MUTEX(shares_mutex);
8705 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8708 unsigned long flags;
8711 * We can't change the weight of the root cgroup.
8716 if (shares < MIN_SHARES)
8717 shares = MIN_SHARES;
8718 else if (shares > MAX_SHARES)
8719 shares = MAX_SHARES;
8721 mutex_lock(&shares_mutex);
8722 if (tg->shares == shares)
8725 tg->shares = shares;
8726 for_each_possible_cpu(i) {
8727 struct rq *rq = cpu_rq(i);
8728 struct sched_entity *se;
8731 /* Propagate contribution to hierarchy */
8732 raw_spin_lock_irqsave(&rq->lock, flags);
8733 for_each_sched_entity(se)
8734 update_cfs_shares(group_cfs_rq(se));
8735 raw_spin_unlock_irqrestore(&rq->lock, flags);
8739 mutex_unlock(&shares_mutex);
8743 unsigned long sched_group_shares(struct task_group *tg)
8749 #ifdef CONFIG_RT_GROUP_SCHED
8751 * Ensure that the real time constraints are schedulable.
8753 static DEFINE_MUTEX(rt_constraints_mutex);
8755 static unsigned long to_ratio(u64 period, u64 runtime)
8757 if (runtime == RUNTIME_INF)
8760 return div64_u64(runtime << 20, period);
8763 /* Must be called with tasklist_lock held */
8764 static inline int tg_has_rt_tasks(struct task_group *tg)
8766 struct task_struct *g, *p;
8768 do_each_thread(g, p) {
8769 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8771 } while_each_thread(g, p);
8776 struct rt_schedulable_data {
8777 struct task_group *tg;
8782 static int tg_schedulable(struct task_group *tg, void *data)
8784 struct rt_schedulable_data *d = data;
8785 struct task_group *child;
8786 unsigned long total, sum = 0;
8787 u64 period, runtime;
8789 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8790 runtime = tg->rt_bandwidth.rt_runtime;
8793 period = d->rt_period;
8794 runtime = d->rt_runtime;
8798 * Cannot have more runtime than the period.
8800 if (runtime > period && runtime != RUNTIME_INF)
8804 * Ensure we don't starve existing RT tasks.
8806 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8809 total = to_ratio(period, runtime);
8812 * Nobody can have more than the global setting allows.
8814 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8818 * The sum of our children's runtime should not exceed our own.
8820 list_for_each_entry_rcu(child, &tg->children, siblings) {
8821 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8822 runtime = child->rt_bandwidth.rt_runtime;
8824 if (child == d->tg) {
8825 period = d->rt_period;
8826 runtime = d->rt_runtime;
8829 sum += to_ratio(period, runtime);
8838 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8840 struct rt_schedulable_data data = {
8842 .rt_period = period,
8843 .rt_runtime = runtime,
8846 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8849 static int tg_set_bandwidth(struct task_group *tg,
8850 u64 rt_period, u64 rt_runtime)
8854 mutex_lock(&rt_constraints_mutex);
8855 read_lock(&tasklist_lock);
8856 err = __rt_schedulable(tg, rt_period, rt_runtime);
8860 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8861 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8862 tg->rt_bandwidth.rt_runtime = rt_runtime;
8864 for_each_possible_cpu(i) {
8865 struct rt_rq *rt_rq = tg->rt_rq[i];
8867 raw_spin_lock(&rt_rq->rt_runtime_lock);
8868 rt_rq->rt_runtime = rt_runtime;
8869 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8871 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8873 read_unlock(&tasklist_lock);
8874 mutex_unlock(&rt_constraints_mutex);
8879 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8881 u64 rt_runtime, rt_period;
8883 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8884 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8885 if (rt_runtime_us < 0)
8886 rt_runtime = RUNTIME_INF;
8888 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8891 long sched_group_rt_runtime(struct task_group *tg)
8895 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8898 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8899 do_div(rt_runtime_us, NSEC_PER_USEC);
8900 return rt_runtime_us;
8903 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8905 u64 rt_runtime, rt_period;
8907 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8908 rt_runtime = tg->rt_bandwidth.rt_runtime;
8913 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8916 long sched_group_rt_period(struct task_group *tg)
8920 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8921 do_div(rt_period_us, NSEC_PER_USEC);
8922 return rt_period_us;
8925 static int sched_rt_global_constraints(void)
8927 u64 runtime, period;
8930 if (sysctl_sched_rt_period <= 0)
8933 runtime = global_rt_runtime();
8934 period = global_rt_period();
8937 * Sanity check on the sysctl variables.
8939 if (runtime > period && runtime != RUNTIME_INF)
8942 mutex_lock(&rt_constraints_mutex);
8943 read_lock(&tasklist_lock);
8944 ret = __rt_schedulable(NULL, 0, 0);
8945 read_unlock(&tasklist_lock);
8946 mutex_unlock(&rt_constraints_mutex);
8951 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8953 /* Don't accept realtime tasks when there is no way for them to run */
8954 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8960 #else /* !CONFIG_RT_GROUP_SCHED */
8961 static int sched_rt_global_constraints(void)
8963 unsigned long flags;
8966 if (sysctl_sched_rt_period <= 0)
8970 * There's always some RT tasks in the root group
8971 * -- migration, kstopmachine etc..
8973 if (sysctl_sched_rt_runtime == 0)
8976 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8977 for_each_possible_cpu(i) {
8978 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8980 raw_spin_lock(&rt_rq->rt_runtime_lock);
8981 rt_rq->rt_runtime = global_rt_runtime();
8982 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8984 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8988 #endif /* CONFIG_RT_GROUP_SCHED */
8990 int sched_rt_handler(struct ctl_table *table, int write,
8991 void __user *buffer, size_t *lenp,
8995 int old_period, old_runtime;
8996 static DEFINE_MUTEX(mutex);
8999 old_period = sysctl_sched_rt_period;
9000 old_runtime = sysctl_sched_rt_runtime;
9002 ret = proc_dointvec(table, write, buffer, lenp, ppos);
9004 if (!ret && write) {
9005 ret = sched_rt_global_constraints();
9007 sysctl_sched_rt_period = old_period;
9008 sysctl_sched_rt_runtime = old_runtime;
9010 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9011 def_rt_bandwidth.rt_period =
9012 ns_to_ktime(global_rt_period());
9015 mutex_unlock(&mutex);
9020 #ifdef CONFIG_CGROUP_SCHED
9022 /* return corresponding task_group object of a cgroup */
9023 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9025 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9026 struct task_group, css);
9029 static struct cgroup_subsys_state *
9030 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9032 struct task_group *tg, *parent;
9034 if (!cgrp->parent) {
9035 /* This is early initialization for the top cgroup */
9036 return &root_task_group.css;
9039 parent = cgroup_tg(cgrp->parent);
9040 tg = sched_create_group(parent);
9042 return ERR_PTR(-ENOMEM);
9048 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9050 struct task_group *tg = cgroup_tg(cgrp);
9052 sched_destroy_group(tg);
9056 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
9058 #ifdef CONFIG_RT_GROUP_SCHED
9059 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
9062 /* We don't support RT-tasks being in separate groups */
9063 if (tsk->sched_class != &fair_sched_class)
9070 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9071 struct task_struct *tsk, bool threadgroup)
9073 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
9077 struct task_struct *c;
9079 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9080 retval = cpu_cgroup_can_attach_task(cgrp, c);
9092 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9093 struct cgroup *old_cont, struct task_struct *tsk,
9096 sched_move_task(tsk);
9098 struct task_struct *c;
9100 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
9108 cpu_cgroup_exit(struct cgroup_subsys *ss, struct cgroup *cgrp,
9109 struct cgroup *old_cgrp, struct task_struct *task)
9112 * cgroup_exit() is called in the copy_process() failure path.
9113 * Ignore this case since the task hasn't ran yet, this avoids
9114 * trying to poke a half freed task state from generic code.
9116 if (!(task->flags & PF_EXITING))
9119 sched_move_task(task);
9122 #ifdef CONFIG_FAIR_GROUP_SCHED
9123 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9126 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9129 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9131 struct task_group *tg = cgroup_tg(cgrp);
9133 return (u64) tg->shares;
9135 #endif /* CONFIG_FAIR_GROUP_SCHED */
9137 #ifdef CONFIG_RT_GROUP_SCHED
9138 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9141 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9144 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9146 return sched_group_rt_runtime(cgroup_tg(cgrp));
9149 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9152 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9155 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9157 return sched_group_rt_period(cgroup_tg(cgrp));
9159 #endif /* CONFIG_RT_GROUP_SCHED */
9161 static struct cftype cpu_files[] = {
9162 #ifdef CONFIG_FAIR_GROUP_SCHED
9165 .read_u64 = cpu_shares_read_u64,
9166 .write_u64 = cpu_shares_write_u64,
9169 #ifdef CONFIG_RT_GROUP_SCHED
9171 .name = "rt_runtime_us",
9172 .read_s64 = cpu_rt_runtime_read,
9173 .write_s64 = cpu_rt_runtime_write,
9176 .name = "rt_period_us",
9177 .read_u64 = cpu_rt_period_read_uint,
9178 .write_u64 = cpu_rt_period_write_uint,
9183 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9185 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9188 struct cgroup_subsys cpu_cgroup_subsys = {
9190 .create = cpu_cgroup_create,
9191 .destroy = cpu_cgroup_destroy,
9192 .can_attach = cpu_cgroup_can_attach,
9193 .attach = cpu_cgroup_attach,
9194 .exit = cpu_cgroup_exit,
9195 .populate = cpu_cgroup_populate,
9196 .subsys_id = cpu_cgroup_subsys_id,
9200 #endif /* CONFIG_CGROUP_SCHED */
9202 #ifdef CONFIG_CGROUP_CPUACCT
9205 * CPU accounting code for task groups.
9207 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9208 * (balbir@in.ibm.com).
9211 /* track cpu usage of a group of tasks and its child groups */
9213 struct cgroup_subsys_state css;
9214 /* cpuusage holds pointer to a u64-type object on every cpu */
9215 u64 __percpu *cpuusage;
9216 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9217 struct cpuacct *parent;
9220 struct cgroup_subsys cpuacct_subsys;
9222 /* return cpu accounting group corresponding to this container */
9223 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9225 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9226 struct cpuacct, css);
9229 /* return cpu accounting group to which this task belongs */
9230 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9232 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9233 struct cpuacct, css);
9236 /* create a new cpu accounting group */
9237 static struct cgroup_subsys_state *cpuacct_create(
9238 struct cgroup_subsys *ss, struct cgroup *cgrp)
9240 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9246 ca->cpuusage = alloc_percpu(u64);
9250 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9251 if (percpu_counter_init(&ca->cpustat[i], 0))
9252 goto out_free_counters;
9255 ca->parent = cgroup_ca(cgrp->parent);
9261 percpu_counter_destroy(&ca->cpustat[i]);
9262 free_percpu(ca->cpuusage);
9266 return ERR_PTR(-ENOMEM);
9269 /* destroy an existing cpu accounting group */
9271 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9273 struct cpuacct *ca = cgroup_ca(cgrp);
9276 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9277 percpu_counter_destroy(&ca->cpustat[i]);
9278 free_percpu(ca->cpuusage);
9282 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9284 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9287 #ifndef CONFIG_64BIT
9289 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9291 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9293 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9301 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9303 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9305 #ifndef CONFIG_64BIT
9307 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9309 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9311 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9317 /* return total cpu usage (in nanoseconds) of a group */
9318 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9320 struct cpuacct *ca = cgroup_ca(cgrp);
9321 u64 totalcpuusage = 0;
9324 for_each_present_cpu(i)
9325 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9327 return totalcpuusage;
9330 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9333 struct cpuacct *ca = cgroup_ca(cgrp);
9342 for_each_present_cpu(i)
9343 cpuacct_cpuusage_write(ca, i, 0);
9349 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9352 struct cpuacct *ca = cgroup_ca(cgroup);
9356 for_each_present_cpu(i) {
9357 percpu = cpuacct_cpuusage_read(ca, i);
9358 seq_printf(m, "%llu ", (unsigned long long) percpu);
9360 seq_printf(m, "\n");
9364 static const char *cpuacct_stat_desc[] = {
9365 [CPUACCT_STAT_USER] = "user",
9366 [CPUACCT_STAT_SYSTEM] = "system",
9369 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9370 struct cgroup_map_cb *cb)
9372 struct cpuacct *ca = cgroup_ca(cgrp);
9375 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9376 s64 val = percpu_counter_read(&ca->cpustat[i]);
9377 val = cputime64_to_clock_t(val);
9378 cb->fill(cb, cpuacct_stat_desc[i], val);
9383 static struct cftype files[] = {
9386 .read_u64 = cpuusage_read,
9387 .write_u64 = cpuusage_write,
9390 .name = "usage_percpu",
9391 .read_seq_string = cpuacct_percpu_seq_read,
9395 .read_map = cpuacct_stats_show,
9399 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9401 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9405 * charge this task's execution time to its accounting group.
9407 * called with rq->lock held.
9409 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9414 if (unlikely(!cpuacct_subsys.active))
9417 cpu = task_cpu(tsk);
9423 for (; ca; ca = ca->parent) {
9424 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9425 *cpuusage += cputime;
9432 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9433 * in cputime_t units. As a result, cpuacct_update_stats calls
9434 * percpu_counter_add with values large enough to always overflow the
9435 * per cpu batch limit causing bad SMP scalability.
9437 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9438 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9439 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9442 #define CPUACCT_BATCH \
9443 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9445 #define CPUACCT_BATCH 0
9449 * Charge the system/user time to the task's accounting group.
9451 static void cpuacct_update_stats(struct task_struct *tsk,
9452 enum cpuacct_stat_index idx, cputime_t val)
9455 int batch = CPUACCT_BATCH;
9457 if (unlikely(!cpuacct_subsys.active))
9464 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9470 struct cgroup_subsys cpuacct_subsys = {
9472 .create = cpuacct_create,
9473 .destroy = cpuacct_destroy,
9474 .populate = cpuacct_populate,
9475 .subsys_id = cpuacct_subsys_id,
9477 #endif /* CONFIG_CGROUP_CPUACCT */