4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
75 #include <linux/cpuacct.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
81 #include "workqueue_sched.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;
259 #ifdef CONFIG_RT_GROUP_SCHED
260 struct sched_rt_entity **rt_se;
261 struct rt_rq **rt_rq;
263 struct rt_bandwidth rt_bandwidth;
267 struct list_head list;
269 struct task_group *parent;
270 struct list_head siblings;
271 struct list_head children;
274 #define root_task_group init_task_group
276 /* task_group_lock serializes add/remove of task groups and also changes to
277 * a task group's cpu shares.
279 static DEFINE_SPINLOCK(task_group_lock);
281 #ifdef CONFIG_FAIR_GROUP_SCHED
284 static int root_task_group_empty(void)
286 return list_empty(&root_task_group.children);
290 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
293 * A weight of 0 or 1 can cause arithmetics problems.
294 * A weight of a cfs_rq is the sum of weights of which entities
295 * are queued on this cfs_rq, so a weight of a entity should not be
296 * too large, so as the shares value of a task group.
297 * (The default weight is 1024 - so there's no practical
298 * limitation from this.)
301 #define MAX_SHARES (1UL << 18)
303 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
306 /* Default task group.
307 * Every task in system belong to this group at bootup.
309 struct task_group init_task_group;
311 #endif /* CONFIG_CGROUP_SCHED */
313 /* CFS-related fields in a runqueue */
315 struct load_weight load;
316 unsigned long nr_running;
321 struct rb_root tasks_timeline;
322 struct rb_node *rb_leftmost;
324 struct list_head tasks;
325 struct list_head *balance_iterator;
328 * 'curr' points to currently running entity on this cfs_rq.
329 * It is set to NULL otherwise (i.e when none are currently running).
331 struct sched_entity *curr, *next, *last;
333 unsigned int nr_spread_over;
335 #ifdef CONFIG_FAIR_GROUP_SCHED
336 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
339 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
340 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
341 * (like users, containers etc.)
343 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
344 * list is used during load balance.
346 struct list_head leaf_cfs_rq_list;
347 struct task_group *tg; /* group that "owns" this runqueue */
351 * the part of load.weight contributed by tasks
353 unsigned long task_weight;
356 * h_load = weight * f(tg)
358 * Where f(tg) is the recursive weight fraction assigned to
361 unsigned long h_load;
364 * this cpu's part of tg->shares
366 unsigned long shares;
369 * load.weight at the time we set shares
371 unsigned long rq_weight;
376 /* Real-Time classes' related field in a runqueue: */
378 struct rt_prio_array active;
379 unsigned long rt_nr_running;
380 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
382 int curr; /* highest queued rt task prio */
384 int next; /* next highest */
389 unsigned long rt_nr_migratory;
390 unsigned long rt_nr_total;
392 struct plist_head pushable_tasks;
397 /* Nests inside the rq lock: */
398 raw_spinlock_t rt_runtime_lock;
400 #ifdef CONFIG_RT_GROUP_SCHED
401 unsigned long rt_nr_boosted;
404 struct list_head leaf_rt_rq_list;
405 struct task_group *tg;
412 * We add the notion of a root-domain which will be used to define per-domain
413 * variables. Each exclusive cpuset essentially defines an island domain by
414 * fully partitioning the member cpus from any other cpuset. Whenever a new
415 * exclusive cpuset is created, we also create and attach a new root-domain
422 cpumask_var_t online;
425 * The "RT overload" flag: it gets set if a CPU has more than
426 * one runnable RT task.
428 cpumask_var_t rto_mask;
431 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;
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;
493 unsigned long next_balance;
494 struct mm_struct *prev_mm;
501 struct root_domain *rd;
502 struct sched_domain *sd;
504 unsigned long cpu_power;
506 unsigned char idle_at_tick;
507 /* For active balancing */
511 struct cpu_stop_work active_balance_work;
512 /* cpu of this runqueue: */
516 unsigned long avg_load_per_task;
524 /* calc_load related fields */
525 unsigned long calc_load_update;
526 long calc_load_active;
528 #ifdef CONFIG_SCHED_HRTICK
530 int hrtick_csd_pending;
531 struct call_single_data hrtick_csd;
533 struct hrtimer hrtick_timer;
536 #ifdef CONFIG_SCHEDSTATS
538 struct sched_info rq_sched_info;
539 unsigned long long rq_cpu_time;
540 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
542 /* sys_sched_yield() stats */
543 unsigned int yld_count;
545 /* schedule() stats */
546 unsigned int sched_switch;
547 unsigned int sched_count;
548 unsigned int sched_goidle;
550 /* try_to_wake_up() stats */
551 unsigned int ttwu_count;
552 unsigned int ttwu_local;
555 unsigned int bkl_count;
559 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
562 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
564 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
567 * A queue event has occurred, and we're going to schedule. In
568 * this case, we can save a useless back to back clock update.
570 if (rq->curr->se.on_rq && test_tsk_need_resched(rq->curr))
571 rq->skip_clock_update = 1;
574 static inline int cpu_of(struct rq *rq)
583 #define rcu_dereference_check_sched_domain(p) \
584 rcu_dereference_check((p), \
585 rcu_read_lock_sched_held() || \
586 lockdep_is_held(&sched_domains_mutex))
589 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
590 * See detach_destroy_domains: synchronize_sched for details.
592 * The domain tree of any CPU may only be accessed from within
593 * preempt-disabled sections.
595 #define for_each_domain(cpu, __sd) \
596 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
598 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
599 #define this_rq() (&__get_cpu_var(runqueues))
600 #define task_rq(p) cpu_rq(task_cpu(p))
601 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
602 #define raw_rq() (&__raw_get_cpu_var(runqueues))
604 #ifdef CONFIG_CGROUP_SCHED
607 * Return the group to which this tasks belongs.
609 * We use task_subsys_state_check() and extend the RCU verification
610 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
611 * holds that lock for each task it moves into the cgroup. Therefore
612 * by holding that lock, we pin the task to the current cgroup.
614 static inline struct task_group *task_group(struct task_struct *p)
616 struct cgroup_subsys_state *css;
618 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
619 lockdep_is_held(&task_rq(p)->lock));
620 return container_of(css, struct task_group, css);
623 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
624 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
626 #ifdef CONFIG_FAIR_GROUP_SCHED
627 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
628 p->se.parent = task_group(p)->se[cpu];
631 #ifdef CONFIG_RT_GROUP_SCHED
632 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
633 p->rt.parent = task_group(p)->rt_se[cpu];
637 #else /* CONFIG_CGROUP_SCHED */
639 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
640 static inline struct task_group *task_group(struct task_struct *p)
645 #endif /* CONFIG_CGROUP_SCHED */
647 inline void update_rq_clock(struct rq *rq)
649 if (!rq->skip_clock_update)
650 rq->clock = sched_clock_cpu(cpu_of(rq));
654 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
656 #ifdef CONFIG_SCHED_DEBUG
657 # define const_debug __read_mostly
659 # define const_debug static const
664 * @cpu: the processor in question.
666 * Returns true if the current cpu runqueue is locked.
667 * This interface allows printk to be called with the runqueue lock
668 * held and know whether or not it is OK to wake up the klogd.
670 int runqueue_is_locked(int cpu)
672 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
676 * Debugging: various feature bits
679 #define SCHED_FEAT(name, enabled) \
680 __SCHED_FEAT_##name ,
683 #include "sched_features.h"
688 #define SCHED_FEAT(name, enabled) \
689 (1UL << __SCHED_FEAT_##name) * enabled |
691 const_debug unsigned int sysctl_sched_features =
692 #include "sched_features.h"
697 #ifdef CONFIG_SCHED_DEBUG
698 #define SCHED_FEAT(name, enabled) \
701 static __read_mostly char *sched_feat_names[] = {
702 #include "sched_features.h"
708 static int sched_feat_show(struct seq_file *m, void *v)
712 for (i = 0; sched_feat_names[i]; i++) {
713 if (!(sysctl_sched_features & (1UL << i)))
715 seq_printf(m, "%s ", sched_feat_names[i]);
723 sched_feat_write(struct file *filp, const char __user *ubuf,
724 size_t cnt, loff_t *ppos)
734 if (copy_from_user(&buf, ubuf, cnt))
740 if (strncmp(buf, "NO_", 3) == 0) {
745 for (i = 0; sched_feat_names[i]; i++) {
746 if (strcmp(cmp, sched_feat_names[i]) == 0) {
748 sysctl_sched_features &= ~(1UL << i);
750 sysctl_sched_features |= (1UL << i);
755 if (!sched_feat_names[i])
763 static int sched_feat_open(struct inode *inode, struct file *filp)
765 return single_open(filp, sched_feat_show, NULL);
768 static const struct file_operations sched_feat_fops = {
769 .open = sched_feat_open,
770 .write = sched_feat_write,
773 .release = single_release,
776 static __init int sched_init_debug(void)
778 debugfs_create_file("sched_features", 0644, NULL, NULL,
783 late_initcall(sched_init_debug);
787 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
790 * Number of tasks to iterate in a single balance run.
791 * Limited because this is done with IRQs disabled.
793 const_debug unsigned int sysctl_sched_nr_migrate = 32;
796 * ratelimit for updating the group shares.
799 unsigned int sysctl_sched_shares_ratelimit = 250000;
800 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
803 * Inject some fuzzyness into changing the per-cpu group shares
804 * this avoids remote rq-locks at the expense of fairness.
807 unsigned int sysctl_sched_shares_thresh = 4;
810 * period over which we average the RT time consumption, measured
815 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
818 * period over which we measure -rt task cpu usage in us.
821 unsigned int sysctl_sched_rt_period = 1000000;
823 static __read_mostly int scheduler_running;
826 * part of the period that we allow rt tasks to run in us.
829 int sysctl_sched_rt_runtime = 950000;
831 static inline u64 global_rt_period(void)
833 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
836 static inline u64 global_rt_runtime(void)
838 if (sysctl_sched_rt_runtime < 0)
841 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
844 #ifndef prepare_arch_switch
845 # define prepare_arch_switch(next) do { } while (0)
847 #ifndef finish_arch_switch
848 # define finish_arch_switch(prev) do { } while (0)
851 static inline int task_current(struct rq *rq, struct task_struct *p)
853 return rq->curr == p;
856 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
857 static inline int task_running(struct rq *rq, struct task_struct *p)
859 return task_current(rq, p);
862 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
866 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
868 #ifdef CONFIG_DEBUG_SPINLOCK
869 /* this is a valid case when another task releases the spinlock */
870 rq->lock.owner = current;
873 * If we are tracking spinlock dependencies then we have to
874 * fix up the runqueue lock - which gets 'carried over' from
877 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
879 raw_spin_unlock_irq(&rq->lock);
882 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
883 static inline int task_running(struct rq *rq, struct task_struct *p)
888 return task_current(rq, p);
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 ->oncpu 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)
1359 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1360 * of tasks with abnormal "nice" values across CPUs the contribution that
1361 * each task makes to its run queue's load is weighted according to its
1362 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1363 * scaled version of the new time slice allocation that they receive on time
1367 #define WEIGHT_IDLEPRIO 3
1368 #define WMULT_IDLEPRIO 1431655765
1371 * Nice levels are multiplicative, with a gentle 10% change for every
1372 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1373 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1374 * that remained on nice 0.
1376 * The "10% effect" is relative and cumulative: from _any_ nice level,
1377 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1378 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1379 * If a task goes up by ~10% and another task goes down by ~10% then
1380 * the relative distance between them is ~25%.)
1382 static const int prio_to_weight[40] = {
1383 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1384 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1385 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1386 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1387 /* 0 */ 1024, 820, 655, 526, 423,
1388 /* 5 */ 335, 272, 215, 172, 137,
1389 /* 10 */ 110, 87, 70, 56, 45,
1390 /* 15 */ 36, 29, 23, 18, 15,
1394 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1396 * In cases where the weight does not change often, we can use the
1397 * precalculated inverse to speed up arithmetics by turning divisions
1398 * into multiplications:
1400 static const u32 prio_to_wmult[40] = {
1401 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1402 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1403 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1404 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1405 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1406 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1407 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1408 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1411 /* Time spent by the tasks of the cpu accounting group executing in ... */
1412 enum cpuacct_stat_index {
1413 CPUACCT_STAT_USER, /* ... user mode */
1414 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1416 CPUACCT_STAT_NSTATS,
1419 #ifdef CONFIG_CGROUP_CPUACCT
1420 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1421 static void cpuacct_update_stats(struct task_struct *tsk,
1422 enum cpuacct_stat_index idx, cputime_t val);
1424 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1425 static inline void cpuacct_update_stats(struct task_struct *tsk,
1426 enum cpuacct_stat_index idx, cputime_t val) {}
1429 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1431 update_load_add(&rq->load, load);
1434 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1436 update_load_sub(&rq->load, load);
1439 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1440 typedef int (*tg_visitor)(struct task_group *, void *);
1443 * Iterate the full tree, calling @down when first entering a node and @up when
1444 * leaving it for the final time.
1446 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1448 struct task_group *parent, *child;
1452 parent = &root_task_group;
1454 ret = (*down)(parent, data);
1457 list_for_each_entry_rcu(child, &parent->children, siblings) {
1464 ret = (*up)(parent, data);
1469 parent = parent->parent;
1478 static int tg_nop(struct task_group *tg, void *data)
1485 /* Used instead of source_load when we know the type == 0 */
1486 static unsigned long weighted_cpuload(const int cpu)
1488 return cpu_rq(cpu)->load.weight;
1492 * Return a low guess at the load of a migration-source cpu weighted
1493 * according to the scheduling class and "nice" value.
1495 * We want to under-estimate the load of migration sources, to
1496 * balance conservatively.
1498 static unsigned long source_load(int cpu, int type)
1500 struct rq *rq = cpu_rq(cpu);
1501 unsigned long total = weighted_cpuload(cpu);
1503 if (type == 0 || !sched_feat(LB_BIAS))
1506 return min(rq->cpu_load[type-1], total);
1510 * Return a high guess at the load of a migration-target cpu weighted
1511 * according to the scheduling class and "nice" value.
1513 static unsigned long target_load(int cpu, int type)
1515 struct rq *rq = cpu_rq(cpu);
1516 unsigned long total = weighted_cpuload(cpu);
1518 if (type == 0 || !sched_feat(LB_BIAS))
1521 return max(rq->cpu_load[type-1], total);
1524 static unsigned long power_of(int cpu)
1526 return cpu_rq(cpu)->cpu_power;
1529 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1531 static unsigned long cpu_avg_load_per_task(int cpu)
1533 struct rq *rq = cpu_rq(cpu);
1534 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1537 rq->avg_load_per_task = rq->load.weight / nr_running;
1539 rq->avg_load_per_task = 0;
1541 return rq->avg_load_per_task;
1544 #ifdef CONFIG_FAIR_GROUP_SCHED
1546 static __read_mostly unsigned long __percpu *update_shares_data;
1548 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1551 * Calculate and set the cpu's group shares.
1553 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1554 unsigned long sd_shares,
1555 unsigned long sd_rq_weight,
1556 unsigned long *usd_rq_weight)
1558 unsigned long shares, rq_weight;
1561 rq_weight = usd_rq_weight[cpu];
1564 rq_weight = NICE_0_LOAD;
1568 * \Sum_j shares_j * rq_weight_i
1569 * shares_i = -----------------------------
1570 * \Sum_j rq_weight_j
1572 shares = (sd_shares * rq_weight) / sd_rq_weight;
1573 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1575 if (abs(shares - tg->se[cpu]->load.weight) >
1576 sysctl_sched_shares_thresh) {
1577 struct rq *rq = cpu_rq(cpu);
1578 unsigned long flags;
1580 raw_spin_lock_irqsave(&rq->lock, flags);
1581 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1582 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1583 __set_se_shares(tg->se[cpu], shares);
1584 raw_spin_unlock_irqrestore(&rq->lock, flags);
1589 * Re-compute the task group their per cpu shares over the given domain.
1590 * This needs to be done in a bottom-up fashion because the rq weight of a
1591 * parent group depends on the shares of its child groups.
1593 static int tg_shares_up(struct task_group *tg, void *data)
1595 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1596 unsigned long *usd_rq_weight;
1597 struct sched_domain *sd = data;
1598 unsigned long flags;
1604 local_irq_save(flags);
1605 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1607 for_each_cpu(i, sched_domain_span(sd)) {
1608 weight = tg->cfs_rq[i]->load.weight;
1609 usd_rq_weight[i] = weight;
1611 rq_weight += weight;
1613 * If there are currently no tasks on the cpu pretend there
1614 * is one of average load so that when a new task gets to
1615 * run here it will not get delayed by group starvation.
1618 weight = NICE_0_LOAD;
1620 sum_weight += weight;
1621 shares += tg->cfs_rq[i]->shares;
1625 rq_weight = sum_weight;
1627 if ((!shares && rq_weight) || shares > tg->shares)
1628 shares = tg->shares;
1630 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1631 shares = tg->shares;
1633 for_each_cpu(i, sched_domain_span(sd))
1634 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1636 local_irq_restore(flags);
1642 * Compute the cpu's hierarchical load factor for each task group.
1643 * This needs to be done in a top-down fashion because the load of a child
1644 * group is a fraction of its parents load.
1646 static int tg_load_down(struct task_group *tg, void *data)
1649 long cpu = (long)data;
1652 load = cpu_rq(cpu)->load.weight;
1654 load = tg->parent->cfs_rq[cpu]->h_load;
1655 load *= tg->cfs_rq[cpu]->shares;
1656 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1659 tg->cfs_rq[cpu]->h_load = load;
1664 static void update_shares(struct sched_domain *sd)
1669 if (root_task_group_empty())
1672 now = local_clock();
1673 elapsed = now - sd->last_update;
1675 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1676 sd->last_update = now;
1677 walk_tg_tree(tg_nop, tg_shares_up, sd);
1681 static void update_h_load(long cpu)
1683 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1688 static inline void update_shares(struct sched_domain *sd)
1694 #ifdef CONFIG_PREEMPT
1696 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1699 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1700 * way at the expense of forcing extra atomic operations in all
1701 * invocations. This assures that the double_lock is acquired using the
1702 * same underlying policy as the spinlock_t on this architecture, which
1703 * reduces latency compared to the unfair variant below. However, it
1704 * also adds more overhead and therefore may reduce throughput.
1706 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1707 __releases(this_rq->lock)
1708 __acquires(busiest->lock)
1709 __acquires(this_rq->lock)
1711 raw_spin_unlock(&this_rq->lock);
1712 double_rq_lock(this_rq, busiest);
1719 * Unfair double_lock_balance: Optimizes throughput at the expense of
1720 * latency by eliminating extra atomic operations when the locks are
1721 * already in proper order on entry. This favors lower cpu-ids and will
1722 * grant the double lock to lower cpus over higher ids under contention,
1723 * regardless of entry order into the function.
1725 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1726 __releases(this_rq->lock)
1727 __acquires(busiest->lock)
1728 __acquires(this_rq->lock)
1732 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1733 if (busiest < this_rq) {
1734 raw_spin_unlock(&this_rq->lock);
1735 raw_spin_lock(&busiest->lock);
1736 raw_spin_lock_nested(&this_rq->lock,
1737 SINGLE_DEPTH_NESTING);
1740 raw_spin_lock_nested(&busiest->lock,
1741 SINGLE_DEPTH_NESTING);
1746 #endif /* CONFIG_PREEMPT */
1749 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1751 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1753 if (unlikely(!irqs_disabled())) {
1754 /* printk() doesn't work good under rq->lock */
1755 raw_spin_unlock(&this_rq->lock);
1759 return _double_lock_balance(this_rq, busiest);
1762 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1763 __releases(busiest->lock)
1765 raw_spin_unlock(&busiest->lock);
1766 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1770 * double_rq_lock - safely lock two runqueues
1772 * Note this does not disable interrupts like task_rq_lock,
1773 * you need to do so manually before calling.
1775 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1776 __acquires(rq1->lock)
1777 __acquires(rq2->lock)
1779 BUG_ON(!irqs_disabled());
1781 raw_spin_lock(&rq1->lock);
1782 __acquire(rq2->lock); /* Fake it out ;) */
1785 raw_spin_lock(&rq1->lock);
1786 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1788 raw_spin_lock(&rq2->lock);
1789 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1795 * double_rq_unlock - safely unlock two runqueues
1797 * Note this does not restore interrupts like task_rq_unlock,
1798 * you need to do so manually after calling.
1800 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1801 __releases(rq1->lock)
1802 __releases(rq2->lock)
1804 raw_spin_unlock(&rq1->lock);
1806 raw_spin_unlock(&rq2->lock);
1808 __release(rq2->lock);
1813 #ifdef CONFIG_FAIR_GROUP_SCHED
1814 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1817 cfs_rq->shares = shares;
1822 static void calc_load_account_idle(struct rq *this_rq);
1823 static void update_sysctl(void);
1824 static int get_update_sysctl_factor(void);
1825 static void update_cpu_load(struct rq *this_rq);
1827 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1829 set_task_rq(p, cpu);
1832 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1833 * successfuly executed on another CPU. We must ensure that updates of
1834 * per-task data have been completed by this moment.
1837 task_thread_info(p)->cpu = cpu;
1841 static const struct sched_class rt_sched_class;
1843 #define sched_class_highest (&rt_sched_class)
1844 #define for_each_class(class) \
1845 for (class = sched_class_highest; class; class = class->next)
1847 #include "sched_stats.h"
1849 static void inc_nr_running(struct rq *rq)
1854 static void dec_nr_running(struct rq *rq)
1859 static void set_load_weight(struct task_struct *p)
1862 * SCHED_IDLE tasks get minimal weight:
1864 if (p->policy == SCHED_IDLE) {
1865 p->se.load.weight = WEIGHT_IDLEPRIO;
1866 p->se.load.inv_weight = WMULT_IDLEPRIO;
1870 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1871 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1874 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1876 update_rq_clock(rq);
1877 sched_info_queued(p);
1878 p->sched_class->enqueue_task(rq, p, flags);
1882 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1884 update_rq_clock(rq);
1885 sched_info_dequeued(p);
1886 p->sched_class->dequeue_task(rq, p, flags);
1891 * activate_task - move a task to the runqueue.
1893 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1895 if (task_contributes_to_load(p))
1896 rq->nr_uninterruptible--;
1898 enqueue_task(rq, p, flags);
1903 * deactivate_task - remove a task from the runqueue.
1905 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1907 if (task_contributes_to_load(p))
1908 rq->nr_uninterruptible++;
1910 dequeue_task(rq, p, flags);
1914 #include "sched_idletask.c"
1915 #include "sched_fair.c"
1916 #include "sched_rt.c"
1917 #ifdef CONFIG_SCHED_DEBUG
1918 # include "sched_debug.c"
1922 * __normal_prio - return the priority that is based on the static prio
1924 static inline int __normal_prio(struct task_struct *p)
1926 return p->static_prio;
1930 * Calculate the expected normal priority: i.e. priority
1931 * without taking RT-inheritance into account. Might be
1932 * boosted by interactivity modifiers. Changes upon fork,
1933 * setprio syscalls, and whenever the interactivity
1934 * estimator recalculates.
1936 static inline int normal_prio(struct task_struct *p)
1940 if (task_has_rt_policy(p))
1941 prio = MAX_RT_PRIO-1 - p->rt_priority;
1943 prio = __normal_prio(p);
1948 * Calculate the current priority, i.e. the priority
1949 * taken into account by the scheduler. This value might
1950 * be boosted by RT tasks, or might be boosted by
1951 * interactivity modifiers. Will be RT if the task got
1952 * RT-boosted. If not then it returns p->normal_prio.
1954 static int effective_prio(struct task_struct *p)
1956 p->normal_prio = normal_prio(p);
1958 * If we are RT tasks or we were boosted to RT priority,
1959 * keep the priority unchanged. Otherwise, update priority
1960 * to the normal priority:
1962 if (!rt_prio(p->prio))
1963 return p->normal_prio;
1968 * task_curr - is this task currently executing on a CPU?
1969 * @p: the task in question.
1971 inline int task_curr(const struct task_struct *p)
1973 return cpu_curr(task_cpu(p)) == p;
1976 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1977 const struct sched_class *prev_class,
1978 int oldprio, int running)
1980 if (prev_class != p->sched_class) {
1981 if (prev_class->switched_from)
1982 prev_class->switched_from(rq, p, running);
1983 p->sched_class->switched_to(rq, p, running);
1985 p->sched_class->prio_changed(rq, p, oldprio, running);
1990 * Is this task likely cache-hot:
1993 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1997 if (p->sched_class != &fair_sched_class)
2001 * Buddy candidates are cache hot:
2003 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2004 (&p->se == cfs_rq_of(&p->se)->next ||
2005 &p->se == cfs_rq_of(&p->se)->last))
2008 if (sysctl_sched_migration_cost == -1)
2010 if (sysctl_sched_migration_cost == 0)
2013 delta = now - p->se.exec_start;
2015 return delta < (s64)sysctl_sched_migration_cost;
2018 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2020 #ifdef CONFIG_SCHED_DEBUG
2022 * We should never call set_task_cpu() on a blocked task,
2023 * ttwu() will sort out the placement.
2025 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2026 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2029 trace_sched_migrate_task(p, new_cpu);
2031 if (task_cpu(p) != new_cpu) {
2032 p->se.nr_migrations++;
2033 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2036 __set_task_cpu(p, new_cpu);
2039 struct migration_arg {
2040 struct task_struct *task;
2044 static int migration_cpu_stop(void *data);
2047 * The task's runqueue lock must be held.
2048 * Returns true if you have to wait for migration thread.
2050 static bool migrate_task(struct task_struct *p, int dest_cpu)
2052 struct rq *rq = task_rq(p);
2055 * If the task is not on a runqueue (and not running), then
2056 * the next wake-up will properly place the task.
2058 return p->se.on_rq || task_running(rq, p);
2062 * wait_task_inactive - wait for a thread to unschedule.
2064 * If @match_state is nonzero, it's the @p->state value just checked and
2065 * not expected to change. If it changes, i.e. @p might have woken up,
2066 * then return zero. When we succeed in waiting for @p to be off its CPU,
2067 * we return a positive number (its total switch count). If a second call
2068 * a short while later returns the same number, the caller can be sure that
2069 * @p has remained unscheduled the whole time.
2071 * The caller must ensure that the task *will* unschedule sometime soon,
2072 * else this function might spin for a *long* time. This function can't
2073 * be called with interrupts off, or it may introduce deadlock with
2074 * smp_call_function() if an IPI is sent by the same process we are
2075 * waiting to become inactive.
2077 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2079 unsigned long flags;
2086 * We do the initial early heuristics without holding
2087 * any task-queue locks at all. We'll only try to get
2088 * the runqueue lock when things look like they will
2094 * If the task is actively running on another CPU
2095 * still, just relax and busy-wait without holding
2098 * NOTE! Since we don't hold any locks, it's not
2099 * even sure that "rq" stays as the right runqueue!
2100 * But we don't care, since "task_running()" will
2101 * return false if the runqueue has changed and p
2102 * is actually now running somewhere else!
2104 while (task_running(rq, p)) {
2105 if (match_state && unlikely(p->state != match_state))
2111 * Ok, time to look more closely! We need the rq
2112 * lock now, to be *sure*. If we're wrong, we'll
2113 * just go back and repeat.
2115 rq = task_rq_lock(p, &flags);
2116 trace_sched_wait_task(p);
2117 running = task_running(rq, p);
2118 on_rq = p->se.on_rq;
2120 if (!match_state || p->state == match_state)
2121 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2122 task_rq_unlock(rq, &flags);
2125 * If it changed from the expected state, bail out now.
2127 if (unlikely(!ncsw))
2131 * Was it really running after all now that we
2132 * checked with the proper locks actually held?
2134 * Oops. Go back and try again..
2136 if (unlikely(running)) {
2142 * It's not enough that it's not actively running,
2143 * it must be off the runqueue _entirely_, and not
2146 * So if it was still runnable (but just not actively
2147 * running right now), it's preempted, and we should
2148 * yield - it could be a while.
2150 if (unlikely(on_rq)) {
2151 schedule_timeout_uninterruptible(1);
2156 * Ahh, all good. It wasn't running, and it wasn't
2157 * runnable, which means that it will never become
2158 * running in the future either. We're all done!
2167 * kick_process - kick a running thread to enter/exit the kernel
2168 * @p: the to-be-kicked thread
2170 * Cause a process which is running on another CPU to enter
2171 * kernel-mode, without any delay. (to get signals handled.)
2173 * NOTE: this function doesnt have to take the runqueue lock,
2174 * because all it wants to ensure is that the remote task enters
2175 * the kernel. If the IPI races and the task has been migrated
2176 * to another CPU then no harm is done and the purpose has been
2179 void kick_process(struct task_struct *p)
2185 if ((cpu != smp_processor_id()) && task_curr(p))
2186 smp_send_reschedule(cpu);
2189 EXPORT_SYMBOL_GPL(kick_process);
2190 #endif /* CONFIG_SMP */
2193 * task_oncpu_function_call - call a function on the cpu on which a task runs
2194 * @p: the task to evaluate
2195 * @func: the function to be called
2196 * @info: the function call argument
2198 * Calls the function @func when the task is currently running. This might
2199 * be on the current CPU, which just calls the function directly
2201 void task_oncpu_function_call(struct task_struct *p,
2202 void (*func) (void *info), void *info)
2209 smp_call_function_single(cpu, func, info, 1);
2215 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2217 static int select_fallback_rq(int cpu, struct task_struct *p)
2220 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2222 /* Look for allowed, online CPU in same node. */
2223 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2224 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2227 /* Any allowed, online CPU? */
2228 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2229 if (dest_cpu < nr_cpu_ids)
2232 /* No more Mr. Nice Guy. */
2233 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2234 dest_cpu = cpuset_cpus_allowed_fallback(p);
2236 * Don't tell them about moving exiting tasks or
2237 * kernel threads (both mm NULL), since they never
2240 if (p->mm && printk_ratelimit()) {
2241 printk(KERN_INFO "process %d (%s) no "
2242 "longer affine to cpu%d\n",
2243 task_pid_nr(p), p->comm, cpu);
2251 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2254 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2256 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2259 * In order not to call set_task_cpu() on a blocking task we need
2260 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2263 * Since this is common to all placement strategies, this lives here.
2265 * [ this allows ->select_task() to simply return task_cpu(p) and
2266 * not worry about this generic constraint ]
2268 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2270 cpu = select_fallback_rq(task_cpu(p), p);
2275 static void update_avg(u64 *avg, u64 sample)
2277 s64 diff = sample - *avg;
2282 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2283 bool is_sync, bool is_migrate, bool is_local,
2284 unsigned long en_flags)
2286 schedstat_inc(p, se.statistics.nr_wakeups);
2288 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2290 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2292 schedstat_inc(p, se.statistics.nr_wakeups_local);
2294 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2296 activate_task(rq, p, en_flags);
2299 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2300 int wake_flags, bool success)
2302 trace_sched_wakeup(p, success);
2303 check_preempt_curr(rq, p, wake_flags);
2305 p->state = TASK_RUNNING;
2307 if (p->sched_class->task_woken)
2308 p->sched_class->task_woken(rq, p);
2310 if (unlikely(rq->idle_stamp)) {
2311 u64 delta = rq->clock - rq->idle_stamp;
2312 u64 max = 2*sysctl_sched_migration_cost;
2317 update_avg(&rq->avg_idle, delta);
2321 /* if a worker is waking up, notify workqueue */
2322 if ((p->flags & PF_WQ_WORKER) && success)
2323 wq_worker_waking_up(p, cpu_of(rq));
2327 * try_to_wake_up - wake up a thread
2328 * @p: the thread to be awakened
2329 * @state: the mask of task states that can be woken
2330 * @wake_flags: wake modifier flags (WF_*)
2332 * Put it on the run-queue if it's not already there. The "current"
2333 * thread is always on the run-queue (except when the actual
2334 * re-schedule is in progress), and as such you're allowed to do
2335 * the simpler "current->state = TASK_RUNNING" to mark yourself
2336 * runnable without the overhead of this.
2338 * Returns %true if @p was woken up, %false if it was already running
2339 * or @state didn't match @p's state.
2341 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2344 int cpu, orig_cpu, this_cpu, success = 0;
2345 unsigned long flags;
2346 unsigned long en_flags = ENQUEUE_WAKEUP;
2349 this_cpu = get_cpu();
2352 rq = task_rq_lock(p, &flags);
2353 if (!(p->state & state))
2363 if (unlikely(task_running(rq, p)))
2367 * In order to handle concurrent wakeups and release the rq->lock
2368 * we put the task in TASK_WAKING state.
2370 * First fix up the nr_uninterruptible count:
2372 if (task_contributes_to_load(p)) {
2373 if (likely(cpu_online(orig_cpu)))
2374 rq->nr_uninterruptible--;
2376 this_rq()->nr_uninterruptible--;
2378 p->state = TASK_WAKING;
2380 if (p->sched_class->task_waking) {
2381 p->sched_class->task_waking(rq, p);
2382 en_flags |= ENQUEUE_WAKING;
2385 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2386 if (cpu != orig_cpu)
2387 set_task_cpu(p, cpu);
2388 __task_rq_unlock(rq);
2391 raw_spin_lock(&rq->lock);
2394 * We migrated the task without holding either rq->lock, however
2395 * since the task is not on the task list itself, nobody else
2396 * will try and migrate the task, hence the rq should match the
2397 * cpu we just moved it to.
2399 WARN_ON(task_cpu(p) != cpu);
2400 WARN_ON(p->state != TASK_WAKING);
2402 #ifdef CONFIG_SCHEDSTATS
2403 schedstat_inc(rq, ttwu_count);
2404 if (cpu == this_cpu)
2405 schedstat_inc(rq, ttwu_local);
2407 struct sched_domain *sd;
2408 for_each_domain(this_cpu, sd) {
2409 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2410 schedstat_inc(sd, ttwu_wake_remote);
2415 #endif /* CONFIG_SCHEDSTATS */
2418 #endif /* CONFIG_SMP */
2419 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2420 cpu == this_cpu, en_flags);
2423 ttwu_post_activation(p, rq, wake_flags, success);
2425 task_rq_unlock(rq, &flags);
2432 * try_to_wake_up_local - try to wake up a local task with rq lock held
2433 * @p: the thread to be awakened
2435 * Put @p on the run-queue if it's not alredy there. The caller must
2436 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2437 * the current task. this_rq() stays locked over invocation.
2439 static void try_to_wake_up_local(struct task_struct *p)
2441 struct rq *rq = task_rq(p);
2442 bool success = false;
2444 BUG_ON(rq != this_rq());
2445 BUG_ON(p == current);
2446 lockdep_assert_held(&rq->lock);
2448 if (!(p->state & TASK_NORMAL))
2452 if (likely(!task_running(rq, p))) {
2453 schedstat_inc(rq, ttwu_count);
2454 schedstat_inc(rq, ttwu_local);
2456 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2459 ttwu_post_activation(p, rq, 0, success);
2463 * wake_up_process - Wake up a specific process
2464 * @p: The process to be woken up.
2466 * Attempt to wake up the nominated process and move it to the set of runnable
2467 * processes. Returns 1 if the process was woken up, 0 if it was already
2470 * It may be assumed that this function implies a write memory barrier before
2471 * changing the task state if and only if any tasks are woken up.
2473 int wake_up_process(struct task_struct *p)
2475 return try_to_wake_up(p, TASK_ALL, 0);
2477 EXPORT_SYMBOL(wake_up_process);
2479 int wake_up_state(struct task_struct *p, unsigned int state)
2481 return try_to_wake_up(p, state, 0);
2485 * Perform scheduler related setup for a newly forked process p.
2486 * p is forked by current.
2488 * __sched_fork() is basic setup used by init_idle() too:
2490 static void __sched_fork(struct task_struct *p)
2492 p->se.exec_start = 0;
2493 p->se.sum_exec_runtime = 0;
2494 p->se.prev_sum_exec_runtime = 0;
2495 p->se.nr_migrations = 0;
2497 #ifdef CONFIG_SCHEDSTATS
2498 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2501 INIT_LIST_HEAD(&p->rt.run_list);
2503 INIT_LIST_HEAD(&p->se.group_node);
2505 #ifdef CONFIG_PREEMPT_NOTIFIERS
2506 INIT_HLIST_HEAD(&p->preempt_notifiers);
2511 * fork()/clone()-time setup:
2513 void sched_fork(struct task_struct *p, int clone_flags)
2515 int cpu = get_cpu();
2519 * We mark the process as running here. This guarantees that
2520 * nobody will actually run it, and a signal or other external
2521 * event cannot wake it up and insert it on the runqueue either.
2523 p->state = TASK_RUNNING;
2526 * Revert to default priority/policy on fork if requested.
2528 if (unlikely(p->sched_reset_on_fork)) {
2529 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2530 p->policy = SCHED_NORMAL;
2531 p->normal_prio = p->static_prio;
2534 if (PRIO_TO_NICE(p->static_prio) < 0) {
2535 p->static_prio = NICE_TO_PRIO(0);
2536 p->normal_prio = p->static_prio;
2541 * We don't need the reset flag anymore after the fork. It has
2542 * fulfilled its duty:
2544 p->sched_reset_on_fork = 0;
2548 * Make sure we do not leak PI boosting priority to the child.
2550 p->prio = current->normal_prio;
2552 if (!rt_prio(p->prio))
2553 p->sched_class = &fair_sched_class;
2555 if (p->sched_class->task_fork)
2556 p->sched_class->task_fork(p);
2559 * The child is not yet in the pid-hash so no cgroup attach races,
2560 * and the cgroup is pinned to this child due to cgroup_fork()
2561 * is ran before sched_fork().
2563 * Silence PROVE_RCU.
2566 set_task_cpu(p, cpu);
2569 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2570 if (likely(sched_info_on()))
2571 memset(&p->sched_info, 0, sizeof(p->sched_info));
2573 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2576 #ifdef CONFIG_PREEMPT
2577 /* Want to start with kernel preemption disabled. */
2578 task_thread_info(p)->preempt_count = 1;
2580 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2586 * wake_up_new_task - wake up a newly created task for the first time.
2588 * This function will do some initial scheduler statistics housekeeping
2589 * that must be done for every newly created context, then puts the task
2590 * on the runqueue and wakes it.
2592 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2594 unsigned long flags;
2596 int cpu __maybe_unused = get_cpu();
2599 rq = task_rq_lock(p, &flags);
2600 p->state = TASK_WAKING;
2603 * Fork balancing, do it here and not earlier because:
2604 * - cpus_allowed can change in the fork path
2605 * - any previously selected cpu might disappear through hotplug
2607 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2608 * without people poking at ->cpus_allowed.
2610 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2611 set_task_cpu(p, cpu);
2613 p->state = TASK_RUNNING;
2614 task_rq_unlock(rq, &flags);
2617 rq = task_rq_lock(p, &flags);
2618 activate_task(rq, p, 0);
2619 trace_sched_wakeup_new(p, 1);
2620 check_preempt_curr(rq, p, WF_FORK);
2622 if (p->sched_class->task_woken)
2623 p->sched_class->task_woken(rq, p);
2625 task_rq_unlock(rq, &flags);
2629 #ifdef CONFIG_PREEMPT_NOTIFIERS
2632 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2633 * @notifier: notifier struct to register
2635 void preempt_notifier_register(struct preempt_notifier *notifier)
2637 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2639 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2642 * preempt_notifier_unregister - no longer interested in preemption notifications
2643 * @notifier: notifier struct to unregister
2645 * This is safe to call from within a preemption notifier.
2647 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2649 hlist_del(¬ifier->link);
2651 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2653 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2655 struct preempt_notifier *notifier;
2656 struct hlist_node *node;
2658 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2659 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2663 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2664 struct task_struct *next)
2666 struct preempt_notifier *notifier;
2667 struct hlist_node *node;
2669 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2670 notifier->ops->sched_out(notifier, next);
2673 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2675 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2680 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2681 struct task_struct *next)
2685 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2688 * prepare_task_switch - prepare to switch tasks
2689 * @rq: the runqueue preparing to switch
2690 * @prev: the current task that is being switched out
2691 * @next: the task we are going to switch to.
2693 * This is called with the rq lock held and interrupts off. It must
2694 * be paired with a subsequent finish_task_switch after the context
2697 * prepare_task_switch sets up locking and calls architecture specific
2701 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2702 struct task_struct *next)
2704 fire_sched_out_preempt_notifiers(prev, next);
2705 prepare_lock_switch(rq, next);
2706 prepare_arch_switch(next);
2710 * finish_task_switch - clean up after a task-switch
2711 * @rq: runqueue associated with task-switch
2712 * @prev: the thread we just switched away from.
2714 * finish_task_switch must be called after the context switch, paired
2715 * with a prepare_task_switch call before the context switch.
2716 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2717 * and do any other architecture-specific cleanup actions.
2719 * Note that we may have delayed dropping an mm in context_switch(). If
2720 * so, we finish that here outside of the runqueue lock. (Doing it
2721 * with the lock held can cause deadlocks; see schedule() for
2724 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2725 __releases(rq->lock)
2727 struct mm_struct *mm = rq->prev_mm;
2733 * A task struct has one reference for the use as "current".
2734 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2735 * schedule one last time. The schedule call will never return, and
2736 * the scheduled task must drop that reference.
2737 * The test for TASK_DEAD must occur while the runqueue locks are
2738 * still held, otherwise prev could be scheduled on another cpu, die
2739 * there before we look at prev->state, and then the reference would
2741 * Manfred Spraul <manfred@colorfullife.com>
2743 prev_state = prev->state;
2744 finish_arch_switch(prev);
2745 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2746 local_irq_disable();
2747 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2748 perf_event_task_sched_in(current);
2749 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2751 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2752 finish_lock_switch(rq, prev);
2754 fire_sched_in_preempt_notifiers(current);
2757 if (unlikely(prev_state == TASK_DEAD)) {
2759 * Remove function-return probe instances associated with this
2760 * task and put them back on the free list.
2762 kprobe_flush_task(prev);
2763 put_task_struct(prev);
2769 /* assumes rq->lock is held */
2770 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2772 if (prev->sched_class->pre_schedule)
2773 prev->sched_class->pre_schedule(rq, prev);
2776 /* rq->lock is NOT held, but preemption is disabled */
2777 static inline void post_schedule(struct rq *rq)
2779 if (rq->post_schedule) {
2780 unsigned long flags;
2782 raw_spin_lock_irqsave(&rq->lock, flags);
2783 if (rq->curr->sched_class->post_schedule)
2784 rq->curr->sched_class->post_schedule(rq);
2785 raw_spin_unlock_irqrestore(&rq->lock, flags);
2787 rq->post_schedule = 0;
2793 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2797 static inline void post_schedule(struct rq *rq)
2804 * schedule_tail - first thing a freshly forked thread must call.
2805 * @prev: the thread we just switched away from.
2807 asmlinkage void schedule_tail(struct task_struct *prev)
2808 __releases(rq->lock)
2810 struct rq *rq = this_rq();
2812 finish_task_switch(rq, prev);
2815 * FIXME: do we need to worry about rq being invalidated by the
2820 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2821 /* In this case, finish_task_switch does not reenable preemption */
2824 if (current->set_child_tid)
2825 put_user(task_pid_vnr(current), current->set_child_tid);
2829 * context_switch - switch to the new MM and the new
2830 * thread's register state.
2833 context_switch(struct rq *rq, struct task_struct *prev,
2834 struct task_struct *next)
2836 struct mm_struct *mm, *oldmm;
2838 prepare_task_switch(rq, prev, next);
2839 trace_sched_switch(prev, next);
2841 oldmm = prev->active_mm;
2843 * For paravirt, this is coupled with an exit in switch_to to
2844 * combine the page table reload and the switch backend into
2847 arch_start_context_switch(prev);
2850 next->active_mm = oldmm;
2851 atomic_inc(&oldmm->mm_count);
2852 enter_lazy_tlb(oldmm, next);
2854 switch_mm(oldmm, mm, next);
2856 if (likely(!prev->mm)) {
2857 prev->active_mm = NULL;
2858 rq->prev_mm = oldmm;
2861 * Since the runqueue lock will be released by the next
2862 * task (which is an invalid locking op but in the case
2863 * of the scheduler it's an obvious special-case), so we
2864 * do an early lockdep release here:
2866 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2867 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2870 /* Here we just switch the register state and the stack. */
2871 switch_to(prev, next, prev);
2875 * this_rq must be evaluated again because prev may have moved
2876 * CPUs since it called schedule(), thus the 'rq' on its stack
2877 * frame will be invalid.
2879 finish_task_switch(this_rq(), prev);
2883 * nr_running, nr_uninterruptible and nr_context_switches:
2885 * externally visible scheduler statistics: current number of runnable
2886 * threads, current number of uninterruptible-sleeping threads, total
2887 * number of context switches performed since bootup.
2889 unsigned long nr_running(void)
2891 unsigned long i, sum = 0;
2893 for_each_online_cpu(i)
2894 sum += cpu_rq(i)->nr_running;
2899 unsigned long nr_uninterruptible(void)
2901 unsigned long i, sum = 0;
2903 for_each_possible_cpu(i)
2904 sum += cpu_rq(i)->nr_uninterruptible;
2907 * Since we read the counters lockless, it might be slightly
2908 * inaccurate. Do not allow it to go below zero though:
2910 if (unlikely((long)sum < 0))
2916 unsigned long long nr_context_switches(void)
2919 unsigned long long sum = 0;
2921 for_each_possible_cpu(i)
2922 sum += cpu_rq(i)->nr_switches;
2927 unsigned long nr_iowait(void)
2929 unsigned long i, sum = 0;
2931 for_each_possible_cpu(i)
2932 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2937 unsigned long nr_iowait_cpu(int cpu)
2939 struct rq *this = cpu_rq(cpu);
2940 return atomic_read(&this->nr_iowait);
2943 unsigned long this_cpu_load(void)
2945 struct rq *this = this_rq();
2946 return this->cpu_load[0];
2950 /* Variables and functions for calc_load */
2951 static atomic_long_t calc_load_tasks;
2952 static unsigned long calc_load_update;
2953 unsigned long avenrun[3];
2954 EXPORT_SYMBOL(avenrun);
2956 static long calc_load_fold_active(struct rq *this_rq)
2958 long nr_active, delta = 0;
2960 nr_active = this_rq->nr_running;
2961 nr_active += (long) this_rq->nr_uninterruptible;
2963 if (nr_active != this_rq->calc_load_active) {
2964 delta = nr_active - this_rq->calc_load_active;
2965 this_rq->calc_load_active = nr_active;
2971 static unsigned long
2972 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2975 load += active * (FIXED_1 - exp);
2976 load += 1UL << (FSHIFT - 1);
2977 return load >> FSHIFT;
2982 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2984 * When making the ILB scale, we should try to pull this in as well.
2986 static atomic_long_t calc_load_tasks_idle;
2988 static void calc_load_account_idle(struct rq *this_rq)
2992 delta = calc_load_fold_active(this_rq);
2994 atomic_long_add(delta, &calc_load_tasks_idle);
2997 static long calc_load_fold_idle(void)
3002 * Its got a race, we don't care...
3004 if (atomic_long_read(&calc_load_tasks_idle))
3005 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3011 * fixed_power_int - compute: x^n, in O(log n) time
3013 * @x: base of the power
3014 * @frac_bits: fractional bits of @x
3015 * @n: power to raise @x to.
3017 * By exploiting the relation between the definition of the natural power
3018 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3019 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3020 * (where: n_i \elem {0, 1}, the binary vector representing n),
3021 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3022 * of course trivially computable in O(log_2 n), the length of our binary
3025 static unsigned long
3026 fixed_power_int(unsigned long x, unsigned int frac_bits, unsigned int n)
3028 unsigned long result = 1UL << frac_bits;
3033 result += 1UL << (frac_bits - 1);
3034 result >>= frac_bits;
3040 x += 1UL << (frac_bits - 1);
3048 * a1 = a0 * e + a * (1 - e)
3050 * a2 = a1 * e + a * (1 - e)
3051 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3052 * = a0 * e^2 + a * (1 - e) * (1 + e)
3054 * a3 = a2 * e + a * (1 - e)
3055 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3056 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3060 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3061 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3062 * = a0 * e^n + a * (1 - e^n)
3064 * [1] application of the geometric series:
3067 * S_n := \Sum x^i = -------------
3070 static unsigned long
3071 calc_load_n(unsigned long load, unsigned long exp,
3072 unsigned long active, unsigned int n)
3075 return calc_load(load, fixed_power_int(exp, FSHIFT, n), active);
3079 * NO_HZ can leave us missing all per-cpu ticks calling
3080 * calc_load_account_active(), but since an idle CPU folds its delta into
3081 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3082 * in the pending idle delta if our idle period crossed a load cycle boundary.
3084 * Once we've updated the global active value, we need to apply the exponential
3085 * weights adjusted to the number of cycles missed.
3087 static void calc_global_nohz(unsigned long ticks)
3089 long delta, active, n;
3091 if (time_before(jiffies, calc_load_update))
3095 * If we crossed a calc_load_update boundary, make sure to fold
3096 * any pending idle changes, the respective CPUs might have
3097 * missed the tick driven calc_load_account_active() update
3100 delta = calc_load_fold_idle();
3102 atomic_long_add(delta, &calc_load_tasks);
3105 * If we were idle for multiple load cycles, apply them.
3107 if (ticks >= LOAD_FREQ) {
3108 n = ticks / LOAD_FREQ;
3110 active = atomic_long_read(&calc_load_tasks);
3111 active = active > 0 ? active * FIXED_1 : 0;
3113 avenrun[0] = calc_load_n(avenrun[0], EXP_1, active, n);
3114 avenrun[1] = calc_load_n(avenrun[1], EXP_5, active, n);
3115 avenrun[2] = calc_load_n(avenrun[2], EXP_15, active, n);
3117 calc_load_update += n * LOAD_FREQ;
3121 * Its possible the remainder of the above division also crosses
3122 * a LOAD_FREQ period, the regular check in calc_global_load()
3123 * which comes after this will take care of that.
3125 * Consider us being 11 ticks before a cycle completion, and us
3126 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3127 * age us 4 cycles, and the test in calc_global_load() will
3128 * pick up the final one.
3132 static void calc_load_account_idle(struct rq *this_rq)
3136 static inline long calc_load_fold_idle(void)
3141 static void calc_global_nohz(unsigned long ticks)
3147 * get_avenrun - get the load average array
3148 * @loads: pointer to dest load array
3149 * @offset: offset to add
3150 * @shift: shift count to shift the result left
3152 * These values are estimates at best, so no need for locking.
3154 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3156 loads[0] = (avenrun[0] + offset) << shift;
3157 loads[1] = (avenrun[1] + offset) << shift;
3158 loads[2] = (avenrun[2] + offset) << shift;
3162 * calc_load - update the avenrun load estimates 10 ticks after the
3163 * CPUs have updated calc_load_tasks.
3165 void calc_global_load(unsigned long ticks)
3169 calc_global_nohz(ticks);
3171 if (time_before(jiffies, calc_load_update + 10))
3174 active = atomic_long_read(&calc_load_tasks);
3175 active = active > 0 ? active * FIXED_1 : 0;
3177 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3178 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3179 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3181 calc_load_update += LOAD_FREQ;
3185 * Called from update_cpu_load() to periodically update this CPU's
3188 static void calc_load_account_active(struct rq *this_rq)
3192 if (time_before(jiffies, this_rq->calc_load_update))
3195 delta = calc_load_fold_active(this_rq);
3196 delta += calc_load_fold_idle();
3198 atomic_long_add(delta, &calc_load_tasks);
3200 this_rq->calc_load_update += LOAD_FREQ;
3204 * The exact cpuload at various idx values, calculated at every tick would be
3205 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3207 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3208 * on nth tick when cpu may be busy, then we have:
3209 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3210 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3212 * decay_load_missed() below does efficient calculation of
3213 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3214 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3216 * The calculation is approximated on a 128 point scale.
3217 * degrade_zero_ticks is the number of ticks after which load at any
3218 * particular idx is approximated to be zero.
3219 * degrade_factor is a precomputed table, a row for each load idx.
3220 * Each column corresponds to degradation factor for a power of two ticks,
3221 * based on 128 point scale.
3223 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3224 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3226 * With this power of 2 load factors, we can degrade the load n times
3227 * by looking at 1 bits in n and doing as many mult/shift instead of
3228 * n mult/shifts needed by the exact degradation.
3230 #define DEGRADE_SHIFT 7
3231 static const unsigned char
3232 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3233 static const unsigned char
3234 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3235 {0, 0, 0, 0, 0, 0, 0, 0},
3236 {64, 32, 8, 0, 0, 0, 0, 0},
3237 {96, 72, 40, 12, 1, 0, 0},
3238 {112, 98, 75, 43, 15, 1, 0},
3239 {120, 112, 98, 76, 45, 16, 2} };
3242 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3243 * would be when CPU is idle and so we just decay the old load without
3244 * adding any new load.
3246 static unsigned long
3247 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3251 if (!missed_updates)
3254 if (missed_updates >= degrade_zero_ticks[idx])
3258 return load >> missed_updates;
3260 while (missed_updates) {
3261 if (missed_updates % 2)
3262 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3264 missed_updates >>= 1;
3271 * Update rq->cpu_load[] statistics. This function is usually called every
3272 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3273 * every tick. We fix it up based on jiffies.
3275 static void update_cpu_load(struct rq *this_rq)
3277 unsigned long this_load = this_rq->load.weight;
3278 unsigned long curr_jiffies = jiffies;
3279 unsigned long pending_updates;
3282 this_rq->nr_load_updates++;
3284 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3285 if (curr_jiffies == this_rq->last_load_update_tick)
3288 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3289 this_rq->last_load_update_tick = curr_jiffies;
3291 /* Update our load: */
3292 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3293 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3294 unsigned long old_load, new_load;
3296 /* scale is effectively 1 << i now, and >> i divides by scale */
3298 old_load = this_rq->cpu_load[i];
3299 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3300 new_load = this_load;
3302 * Round up the averaging division if load is increasing. This
3303 * prevents us from getting stuck on 9 if the load is 10, for
3306 if (new_load > old_load)
3307 new_load += scale - 1;
3309 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3312 sched_avg_update(this_rq);
3315 static void update_cpu_load_active(struct rq *this_rq)
3317 update_cpu_load(this_rq);
3319 calc_load_account_active(this_rq);
3325 * sched_exec - execve() is a valuable balancing opportunity, because at
3326 * this point the task has the smallest effective memory and cache footprint.
3328 void sched_exec(void)
3330 struct task_struct *p = current;
3331 unsigned long flags;
3335 rq = task_rq_lock(p, &flags);
3336 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3337 if (dest_cpu == smp_processor_id())
3341 * select_task_rq() can race against ->cpus_allowed
3343 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3344 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3345 struct migration_arg arg = { p, dest_cpu };
3347 task_rq_unlock(rq, &flags);
3348 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3352 task_rq_unlock(rq, &flags);
3357 DEFINE_PER_CPU(struct kernel_stat, kstat);
3359 EXPORT_PER_CPU_SYMBOL(kstat);
3362 * Return any ns on the sched_clock that have not yet been accounted in
3363 * @p in case that task is currently running.
3365 * Called with task_rq_lock() held on @rq.
3367 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3371 if (task_current(rq, p)) {
3372 update_rq_clock(rq);
3373 ns = rq->clock - p->se.exec_start;
3381 unsigned long long task_delta_exec(struct task_struct *p)
3383 unsigned long flags;
3387 rq = task_rq_lock(p, &flags);
3388 ns = do_task_delta_exec(p, rq);
3389 task_rq_unlock(rq, &flags);
3395 * Return accounted runtime for the task.
3396 * In case the task is currently running, return the runtime plus current's
3397 * pending runtime that have not been accounted yet.
3399 unsigned long long task_sched_runtime(struct task_struct *p)
3401 unsigned long flags;
3405 rq = task_rq_lock(p, &flags);
3406 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3407 task_rq_unlock(rq, &flags);
3413 * Return sum_exec_runtime for the thread group.
3414 * In case the task is currently running, return the sum plus current's
3415 * pending runtime that have not been accounted yet.
3417 * Note that the thread group might have other running tasks as well,
3418 * so the return value not includes other pending runtime that other
3419 * running tasks might have.
3421 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3423 struct task_cputime totals;
3424 unsigned long flags;
3428 rq = task_rq_lock(p, &flags);
3429 thread_group_cputime(p, &totals);
3430 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3431 task_rq_unlock(rq, &flags);
3437 * Account user cpu time to a process.
3438 * @p: the process that the cpu time gets accounted to
3439 * @cputime: the cpu time spent in user space since the last update
3440 * @cputime_scaled: cputime scaled by cpu frequency
3442 void account_user_time(struct task_struct *p, cputime_t cputime,
3443 cputime_t cputime_scaled)
3445 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3448 /* Add user time to process. */
3449 p->utime = cputime_add(p->utime, cputime);
3450 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3451 account_group_user_time(p, cputime);
3453 /* Add user time to cpustat. */
3454 tmp = cputime_to_cputime64(cputime);
3455 if (TASK_NICE(p) > 0)
3456 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3458 cpustat->user = cputime64_add(cpustat->user, tmp);
3460 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3461 /* Account for user time used */
3462 acct_update_integrals(p);
3466 * Account guest cpu time to a process.
3467 * @p: the process that the cpu time gets accounted to
3468 * @cputime: the cpu time spent in virtual machine since the last update
3469 * @cputime_scaled: cputime scaled by cpu frequency
3471 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3472 cputime_t cputime_scaled)
3475 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3477 tmp = cputime_to_cputime64(cputime);
3479 /* Add guest time to process. */
3480 p->utime = cputime_add(p->utime, cputime);
3481 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3482 account_group_user_time(p, cputime);
3483 p->gtime = cputime_add(p->gtime, cputime);
3485 /* Add guest time to cpustat. */
3486 if (TASK_NICE(p) > 0) {
3487 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3488 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3490 cpustat->user = cputime64_add(cpustat->user, tmp);
3491 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3496 * Account system cpu time to a process.
3497 * @p: the process that the cpu time gets accounted to
3498 * @hardirq_offset: the offset to subtract from hardirq_count()
3499 * @cputime: the cpu time spent in kernel space since the last update
3500 * @cputime_scaled: cputime scaled by cpu frequency
3502 void account_system_time(struct task_struct *p, int hardirq_offset,
3503 cputime_t cputime, cputime_t cputime_scaled)
3505 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3508 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3509 account_guest_time(p, cputime, cputime_scaled);
3513 /* Add system time to process. */
3514 p->stime = cputime_add(p->stime, cputime);
3515 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3516 account_group_system_time(p, cputime);
3518 /* Add system time to cpustat. */
3519 tmp = cputime_to_cputime64(cputime);
3520 if (hardirq_count() - hardirq_offset)
3521 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3522 else if (softirq_count())
3523 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3525 cpustat->system = cputime64_add(cpustat->system, tmp);
3527 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3529 /* Account for system time used */
3530 acct_update_integrals(p);
3534 * Account for involuntary wait time.
3535 * @steal: the cpu time spent in involuntary wait
3537 void account_steal_time(cputime_t cputime)
3539 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3540 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3542 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3546 * Account for idle time.
3547 * @cputime: the cpu time spent in idle wait
3549 void account_idle_time(cputime_t cputime)
3551 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3552 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3553 struct rq *rq = this_rq();
3555 if (atomic_read(&rq->nr_iowait) > 0)
3556 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3558 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3561 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3564 * Account a single tick of cpu time.
3565 * @p: the process that the cpu time gets accounted to
3566 * @user_tick: indicates if the tick is a user or a system tick
3568 void account_process_tick(struct task_struct *p, int user_tick)
3570 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3571 struct rq *rq = this_rq();
3574 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3575 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3576 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3579 account_idle_time(cputime_one_jiffy);
3583 * Account multiple ticks of steal time.
3584 * @p: the process from which the cpu time has been stolen
3585 * @ticks: number of stolen ticks
3587 void account_steal_ticks(unsigned long ticks)
3589 account_steal_time(jiffies_to_cputime(ticks));
3593 * Account multiple ticks of idle time.
3594 * @ticks: number of stolen ticks
3596 void account_idle_ticks(unsigned long ticks)
3598 account_idle_time(jiffies_to_cputime(ticks));
3604 * Use precise platform statistics if available:
3606 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3607 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3613 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3615 struct task_cputime cputime;
3617 thread_group_cputime(p, &cputime);
3619 *ut = cputime.utime;
3620 *st = cputime.stime;
3624 #ifndef nsecs_to_cputime
3625 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3628 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3630 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3633 * Use CFS's precise accounting:
3635 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3641 do_div(temp, total);
3642 utime = (cputime_t)temp;
3647 * Compare with previous values, to keep monotonicity:
3649 p->prev_utime = max(p->prev_utime, utime);
3650 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3652 *ut = p->prev_utime;
3653 *st = p->prev_stime;
3657 * Must be called with siglock held.
3659 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3661 struct signal_struct *sig = p->signal;
3662 struct task_cputime cputime;
3663 cputime_t rtime, utime, total;
3665 thread_group_cputime(p, &cputime);
3667 total = cputime_add(cputime.utime, cputime.stime);
3668 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3673 temp *= cputime.utime;
3674 do_div(temp, total);
3675 utime = (cputime_t)temp;
3679 sig->prev_utime = max(sig->prev_utime, utime);
3680 sig->prev_stime = max(sig->prev_stime,
3681 cputime_sub(rtime, sig->prev_utime));
3683 *ut = sig->prev_utime;
3684 *st = sig->prev_stime;
3689 * This function gets called by the timer code, with HZ frequency.
3690 * We call it with interrupts disabled.
3692 * It also gets called by the fork code, when changing the parent's
3695 void scheduler_tick(void)
3697 int cpu = smp_processor_id();
3698 struct rq *rq = cpu_rq(cpu);
3699 struct task_struct *curr = rq->curr;
3703 raw_spin_lock(&rq->lock);
3704 update_rq_clock(rq);
3705 update_cpu_load_active(rq);
3706 curr->sched_class->task_tick(rq, curr, 0);
3707 raw_spin_unlock(&rq->lock);
3709 perf_event_task_tick(curr);
3712 rq->idle_at_tick = idle_cpu(cpu);
3713 trigger_load_balance(rq, cpu);
3717 notrace unsigned long get_parent_ip(unsigned long addr)
3719 if (in_lock_functions(addr)) {
3720 addr = CALLER_ADDR2;
3721 if (in_lock_functions(addr))
3722 addr = CALLER_ADDR3;
3727 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3728 defined(CONFIG_PREEMPT_TRACER))
3730 void __kprobes add_preempt_count(int val)
3732 #ifdef CONFIG_DEBUG_PREEMPT
3736 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3739 preempt_count() += val;
3740 #ifdef CONFIG_DEBUG_PREEMPT
3742 * Spinlock count overflowing soon?
3744 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3747 if (preempt_count() == val)
3748 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3750 EXPORT_SYMBOL(add_preempt_count);
3752 void __kprobes sub_preempt_count(int val)
3754 #ifdef CONFIG_DEBUG_PREEMPT
3758 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3761 * Is the spinlock portion underflowing?
3763 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3764 !(preempt_count() & PREEMPT_MASK)))
3768 if (preempt_count() == val)
3769 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3770 preempt_count() -= val;
3772 EXPORT_SYMBOL(sub_preempt_count);
3777 * Print scheduling while atomic bug:
3779 static noinline void __schedule_bug(struct task_struct *prev)
3781 struct pt_regs *regs = get_irq_regs();
3783 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3784 prev->comm, prev->pid, preempt_count());
3786 debug_show_held_locks(prev);
3788 if (irqs_disabled())
3789 print_irqtrace_events(prev);
3798 * Various schedule()-time debugging checks and statistics:
3800 static inline void schedule_debug(struct task_struct *prev)
3803 * Test if we are atomic. Since do_exit() needs to call into
3804 * schedule() atomically, we ignore that path for now.
3805 * Otherwise, whine if we are scheduling when we should not be.
3807 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3808 __schedule_bug(prev);
3810 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3812 schedstat_inc(this_rq(), sched_count);
3813 #ifdef CONFIG_SCHEDSTATS
3814 if (unlikely(prev->lock_depth >= 0)) {
3815 schedstat_inc(this_rq(), bkl_count);
3816 schedstat_inc(prev, sched_info.bkl_count);
3821 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3824 update_rq_clock(rq);
3825 prev->sched_class->put_prev_task(rq, prev);
3829 * Pick up the highest-prio task:
3831 static inline struct task_struct *
3832 pick_next_task(struct rq *rq)
3834 const struct sched_class *class;
3835 struct task_struct *p;
3838 * Optimization: we know that if all tasks are in
3839 * the fair class we can call that function directly:
3841 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3842 p = fair_sched_class.pick_next_task(rq);
3847 class = sched_class_highest;
3849 p = class->pick_next_task(rq);
3853 * Will never be NULL as the idle class always
3854 * returns a non-NULL p:
3856 class = class->next;
3861 * schedule() is the main scheduler function.
3863 asmlinkage void __sched schedule(void)
3865 struct task_struct *prev, *next;
3866 unsigned long *switch_count;
3872 cpu = smp_processor_id();
3874 rcu_note_context_switch(cpu);
3877 release_kernel_lock(prev);
3878 need_resched_nonpreemptible:
3880 schedule_debug(prev);
3882 if (sched_feat(HRTICK))
3885 raw_spin_lock_irq(&rq->lock);
3887 switch_count = &prev->nivcsw;
3888 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3889 if (unlikely(signal_pending_state(prev->state, prev))) {
3890 prev->state = TASK_RUNNING;
3893 * If a worker is going to sleep, notify and
3894 * ask workqueue whether it wants to wake up a
3895 * task to maintain concurrency. If so, wake
3898 if (prev->flags & PF_WQ_WORKER) {
3899 struct task_struct *to_wakeup;
3901 to_wakeup = wq_worker_sleeping(prev, cpu);
3903 try_to_wake_up_local(to_wakeup);
3905 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3907 switch_count = &prev->nvcsw;
3910 pre_schedule(rq, prev);
3912 if (unlikely(!rq->nr_running))
3913 idle_balance(cpu, rq);
3915 put_prev_task(rq, prev);
3916 next = pick_next_task(rq);
3917 clear_tsk_need_resched(prev);
3918 rq->skip_clock_update = 0;
3920 if (likely(prev != next)) {
3921 sched_info_switch(prev, next);
3922 perf_event_task_sched_out(prev, next);
3928 context_switch(rq, prev, next); /* unlocks the rq */
3930 * The context switch have flipped the stack from under us
3931 * and restored the local variables which were saved when
3932 * this task called schedule() in the past. prev == current
3933 * is still correct, but it can be moved to another cpu/rq.
3935 cpu = smp_processor_id();
3938 raw_spin_unlock_irq(&rq->lock);
3942 if (unlikely(reacquire_kernel_lock(prev)))
3943 goto need_resched_nonpreemptible;
3945 preempt_enable_no_resched();
3949 EXPORT_SYMBOL(schedule);
3951 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3953 * Look out! "owner" is an entirely speculative pointer
3954 * access and not reliable.
3956 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3961 if (!sched_feat(OWNER_SPIN))
3964 #ifdef CONFIG_DEBUG_PAGEALLOC
3966 * Need to access the cpu field knowing that
3967 * DEBUG_PAGEALLOC could have unmapped it if
3968 * the mutex owner just released it and exited.
3970 if (probe_kernel_address(&owner->cpu, cpu))
3977 * Even if the access succeeded (likely case),
3978 * the cpu field may no longer be valid.
3980 if (cpu >= nr_cpumask_bits)
3984 * We need to validate that we can do a
3985 * get_cpu() and that we have the percpu area.
3987 if (!cpu_online(cpu))
3994 * Owner changed, break to re-assess state.
3996 if (lock->owner != owner) {
3998 * If the lock has switched to a different owner,
3999 * we likely have heavy contention. Return 0 to quit
4000 * optimistic spinning and not contend further:
4008 * Is that owner really running on that cpu?
4010 if (task_thread_info(rq->curr) != owner || need_resched())
4020 #ifdef CONFIG_PREEMPT
4022 * this is the entry point to schedule() from in-kernel preemption
4023 * off of preempt_enable. Kernel preemptions off return from interrupt
4024 * occur there and call schedule directly.
4026 asmlinkage void __sched notrace preempt_schedule(void)
4028 struct thread_info *ti = current_thread_info();
4031 * If there is a non-zero preempt_count or interrupts are disabled,
4032 * we do not want to preempt the current task. Just return..
4034 if (likely(ti->preempt_count || irqs_disabled()))
4038 add_preempt_count_notrace(PREEMPT_ACTIVE);
4040 sub_preempt_count_notrace(PREEMPT_ACTIVE);
4043 * Check again in case we missed a preemption opportunity
4044 * between schedule and now.
4047 } while (need_resched());
4049 EXPORT_SYMBOL(preempt_schedule);
4052 * this is the entry point to schedule() from kernel preemption
4053 * off of irq context.
4054 * Note, that this is called and return with irqs disabled. This will
4055 * protect us against recursive calling from irq.
4057 asmlinkage void __sched preempt_schedule_irq(void)
4059 struct thread_info *ti = current_thread_info();
4061 /* Catch callers which need to be fixed */
4062 BUG_ON(ti->preempt_count || !irqs_disabled());
4065 add_preempt_count(PREEMPT_ACTIVE);
4068 local_irq_disable();
4069 sub_preempt_count(PREEMPT_ACTIVE);
4072 * Check again in case we missed a preemption opportunity
4073 * between schedule and now.
4076 } while (need_resched());
4079 #endif /* CONFIG_PREEMPT */
4081 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
4084 return try_to_wake_up(curr->private, mode, wake_flags);
4086 EXPORT_SYMBOL(default_wake_function);
4089 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4090 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4091 * number) then we wake all the non-exclusive tasks and one exclusive task.
4093 * There are circumstances in which we can try to wake a task which has already
4094 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4095 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4097 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4098 int nr_exclusive, int wake_flags, void *key)
4100 wait_queue_t *curr, *next;
4102 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4103 unsigned flags = curr->flags;
4105 if (curr->func(curr, mode, wake_flags, key) &&
4106 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4112 * __wake_up - wake up threads blocked on a waitqueue.
4114 * @mode: which threads
4115 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4116 * @key: is directly passed to the wakeup function
4118 * It may be assumed that this function implies a write memory barrier before
4119 * changing the task state if and only if any tasks are woken up.
4121 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4122 int nr_exclusive, void *key)
4124 unsigned long flags;
4126 spin_lock_irqsave(&q->lock, flags);
4127 __wake_up_common(q, mode, nr_exclusive, 0, key);
4128 spin_unlock_irqrestore(&q->lock, flags);
4130 EXPORT_SYMBOL(__wake_up);
4133 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4135 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4137 __wake_up_common(q, mode, 1, 0, NULL);
4139 EXPORT_SYMBOL_GPL(__wake_up_locked);
4141 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4143 __wake_up_common(q, mode, 1, 0, key);
4147 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4149 * @mode: which threads
4150 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4151 * @key: opaque value to be passed to wakeup targets
4153 * The sync wakeup differs that the waker knows that it will schedule
4154 * away soon, so while the target thread will be woken up, it will not
4155 * be migrated to another CPU - ie. the two threads are 'synchronized'
4156 * with each other. This can prevent needless bouncing between CPUs.
4158 * On UP it can prevent extra preemption.
4160 * It may be assumed that this function implies a write memory barrier before
4161 * changing the task state if and only if any tasks are woken up.
4163 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4164 int nr_exclusive, void *key)
4166 unsigned long flags;
4167 int wake_flags = WF_SYNC;
4172 if (unlikely(!nr_exclusive))
4175 spin_lock_irqsave(&q->lock, flags);
4176 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4177 spin_unlock_irqrestore(&q->lock, flags);
4179 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4182 * __wake_up_sync - see __wake_up_sync_key()
4184 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4186 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4188 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4191 * complete: - signals a single thread waiting on this completion
4192 * @x: holds the state of this particular completion
4194 * This will wake up a single thread waiting on this completion. Threads will be
4195 * awakened in the same order in which they were queued.
4197 * See also complete_all(), wait_for_completion() and related routines.
4199 * It may be assumed that this function implies a write memory barrier before
4200 * changing the task state if and only if any tasks are woken up.
4202 void complete(struct completion *x)
4204 unsigned long flags;
4206 spin_lock_irqsave(&x->wait.lock, flags);
4208 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4209 spin_unlock_irqrestore(&x->wait.lock, flags);
4211 EXPORT_SYMBOL(complete);
4214 * complete_all: - signals all threads waiting on this completion
4215 * @x: holds the state of this particular completion
4217 * This will wake up all threads waiting on this particular completion event.
4219 * It may be assumed that this function implies a write memory barrier before
4220 * changing the task state if and only if any tasks are woken up.
4222 void complete_all(struct completion *x)
4224 unsigned long flags;
4226 spin_lock_irqsave(&x->wait.lock, flags);
4227 x->done += UINT_MAX/2;
4228 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4229 spin_unlock_irqrestore(&x->wait.lock, flags);
4231 EXPORT_SYMBOL(complete_all);
4233 static inline long __sched
4234 do_wait_for_common(struct completion *x, long timeout, int state)
4237 DECLARE_WAITQUEUE(wait, current);
4239 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4241 if (signal_pending_state(state, current)) {
4242 timeout = -ERESTARTSYS;
4245 __set_current_state(state);
4246 spin_unlock_irq(&x->wait.lock);
4247 timeout = schedule_timeout(timeout);
4248 spin_lock_irq(&x->wait.lock);
4249 } while (!x->done && timeout);
4250 __remove_wait_queue(&x->wait, &wait);
4255 return timeout ?: 1;
4259 wait_for_common(struct completion *x, long timeout, int state)
4263 spin_lock_irq(&x->wait.lock);
4264 timeout = do_wait_for_common(x, timeout, state);
4265 spin_unlock_irq(&x->wait.lock);
4270 * wait_for_completion: - waits for completion of a task
4271 * @x: holds the state of this particular completion
4273 * This waits to be signaled for completion of a specific task. It is NOT
4274 * interruptible and there is no timeout.
4276 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4277 * and interrupt capability. Also see complete().
4279 void __sched wait_for_completion(struct completion *x)
4281 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4283 EXPORT_SYMBOL(wait_for_completion);
4286 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4287 * @x: holds the state of this particular completion
4288 * @timeout: timeout value in jiffies
4290 * This waits for either a completion of a specific task to be signaled or for a
4291 * specified timeout to expire. The timeout is in jiffies. It is not
4294 unsigned long __sched
4295 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4297 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4299 EXPORT_SYMBOL(wait_for_completion_timeout);
4302 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4303 * @x: holds the state of this particular completion
4305 * This waits for completion of a specific task to be signaled. It is
4308 int __sched wait_for_completion_interruptible(struct completion *x)
4310 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4311 if (t == -ERESTARTSYS)
4315 EXPORT_SYMBOL(wait_for_completion_interruptible);
4318 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4319 * @x: holds the state of this particular completion
4320 * @timeout: timeout value in jiffies
4322 * This waits for either a completion of a specific task to be signaled or for a
4323 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4325 unsigned long __sched
4326 wait_for_completion_interruptible_timeout(struct completion *x,
4327 unsigned long timeout)
4329 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4331 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4334 * wait_for_completion_killable: - waits for completion of a task (killable)
4335 * @x: holds the state of this particular completion
4337 * This waits to be signaled for completion of a specific task. It can be
4338 * interrupted by a kill signal.
4340 int __sched wait_for_completion_killable(struct completion *x)
4342 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4343 if (t == -ERESTARTSYS)
4347 EXPORT_SYMBOL(wait_for_completion_killable);
4350 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4351 * @x: holds the state of this particular completion
4352 * @timeout: timeout value in jiffies
4354 * This waits for either a completion of a specific task to be
4355 * signaled or for a specified timeout to expire. It can be
4356 * interrupted by a kill signal. The timeout is in jiffies.
4358 unsigned long __sched
4359 wait_for_completion_killable_timeout(struct completion *x,
4360 unsigned long timeout)
4362 return wait_for_common(x, timeout, TASK_KILLABLE);
4364 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4367 * try_wait_for_completion - try to decrement a completion without blocking
4368 * @x: completion structure
4370 * Returns: 0 if a decrement cannot be done without blocking
4371 * 1 if a decrement succeeded.
4373 * If a completion is being used as a counting completion,
4374 * attempt to decrement the counter without blocking. This
4375 * enables us to avoid waiting if the resource the completion
4376 * is protecting is not available.
4378 bool try_wait_for_completion(struct completion *x)
4380 unsigned long flags;
4383 spin_lock_irqsave(&x->wait.lock, flags);
4388 spin_unlock_irqrestore(&x->wait.lock, flags);
4391 EXPORT_SYMBOL(try_wait_for_completion);
4394 * completion_done - Test to see if a completion has any waiters
4395 * @x: completion structure
4397 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4398 * 1 if there are no waiters.
4401 bool completion_done(struct completion *x)
4403 unsigned long flags;
4406 spin_lock_irqsave(&x->wait.lock, flags);
4409 spin_unlock_irqrestore(&x->wait.lock, flags);
4412 EXPORT_SYMBOL(completion_done);
4415 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4417 unsigned long flags;
4420 init_waitqueue_entry(&wait, current);
4422 __set_current_state(state);
4424 spin_lock_irqsave(&q->lock, flags);
4425 __add_wait_queue(q, &wait);
4426 spin_unlock(&q->lock);
4427 timeout = schedule_timeout(timeout);
4428 spin_lock_irq(&q->lock);
4429 __remove_wait_queue(q, &wait);
4430 spin_unlock_irqrestore(&q->lock, flags);
4435 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4437 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4439 EXPORT_SYMBOL(interruptible_sleep_on);
4442 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4444 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4446 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4448 void __sched sleep_on(wait_queue_head_t *q)
4450 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4452 EXPORT_SYMBOL(sleep_on);
4454 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4456 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4458 EXPORT_SYMBOL(sleep_on_timeout);
4460 #ifdef CONFIG_RT_MUTEXES
4463 * rt_mutex_setprio - set the current priority of a task
4465 * @prio: prio value (kernel-internal form)
4467 * This function changes the 'effective' priority of a task. It does
4468 * not touch ->normal_prio like __setscheduler().
4470 * Used by the rt_mutex code to implement priority inheritance logic.
4472 void rt_mutex_setprio(struct task_struct *p, int prio)
4474 unsigned long flags;
4475 int oldprio, on_rq, running;
4477 const struct sched_class *prev_class;
4479 BUG_ON(prio < 0 || prio > MAX_PRIO);
4481 rq = task_rq_lock(p, &flags);
4484 prev_class = p->sched_class;
4485 on_rq = p->se.on_rq;
4486 running = task_current(rq, p);
4488 dequeue_task(rq, p, 0);
4490 p->sched_class->put_prev_task(rq, p);
4493 p->sched_class = &rt_sched_class;
4495 p->sched_class = &fair_sched_class;
4500 p->sched_class->set_curr_task(rq);
4502 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4504 check_class_changed(rq, p, prev_class, oldprio, running);
4506 task_rq_unlock(rq, &flags);
4511 void set_user_nice(struct task_struct *p, long nice)
4513 int old_prio, delta, on_rq;
4514 unsigned long flags;
4517 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4520 * We have to be careful, if called from sys_setpriority(),
4521 * the task might be in the middle of scheduling on another CPU.
4523 rq = task_rq_lock(p, &flags);
4525 * The RT priorities are set via sched_setscheduler(), but we still
4526 * allow the 'normal' nice value to be set - but as expected
4527 * it wont have any effect on scheduling until the task is
4528 * SCHED_FIFO/SCHED_RR:
4530 if (task_has_rt_policy(p)) {
4531 p->static_prio = NICE_TO_PRIO(nice);
4534 on_rq = p->se.on_rq;
4536 dequeue_task(rq, p, 0);
4538 p->static_prio = NICE_TO_PRIO(nice);
4541 p->prio = effective_prio(p);
4542 delta = p->prio - old_prio;
4545 enqueue_task(rq, p, 0);
4547 * If the task increased its priority or is running and
4548 * lowered its priority, then reschedule its CPU:
4550 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4551 resched_task(rq->curr);
4554 task_rq_unlock(rq, &flags);
4556 EXPORT_SYMBOL(set_user_nice);
4559 * can_nice - check if a task can reduce its nice value
4563 int can_nice(const struct task_struct *p, const int nice)
4565 /* convert nice value [19,-20] to rlimit style value [1,40] */
4566 int nice_rlim = 20 - nice;
4568 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4569 capable(CAP_SYS_NICE));
4572 #ifdef __ARCH_WANT_SYS_NICE
4575 * sys_nice - change the priority of the current process.
4576 * @increment: priority increment
4578 * sys_setpriority is a more generic, but much slower function that
4579 * does similar things.
4581 SYSCALL_DEFINE1(nice, int, increment)
4586 * Setpriority might change our priority at the same moment.
4587 * We don't have to worry. Conceptually one call occurs first
4588 * and we have a single winner.
4590 if (increment < -40)
4595 nice = TASK_NICE(current) + increment;
4601 if (increment < 0 && !can_nice(current, nice))
4604 retval = security_task_setnice(current, nice);
4608 set_user_nice(current, nice);
4615 * task_prio - return the priority value of a given task.
4616 * @p: the task in question.
4618 * This is the priority value as seen by users in /proc.
4619 * RT tasks are offset by -200. Normal tasks are centered
4620 * around 0, value goes from -16 to +15.
4622 int task_prio(const struct task_struct *p)
4624 return p->prio - MAX_RT_PRIO;
4628 * task_nice - return the nice value of a given task.
4629 * @p: the task in question.
4631 int task_nice(const struct task_struct *p)
4633 return TASK_NICE(p);
4635 EXPORT_SYMBOL(task_nice);
4638 * idle_cpu - is a given cpu idle currently?
4639 * @cpu: the processor in question.
4641 int idle_cpu(int cpu)
4643 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4647 * idle_task - return the idle task for a given cpu.
4648 * @cpu: the processor in question.
4650 struct task_struct *idle_task(int cpu)
4652 return cpu_rq(cpu)->idle;
4656 * find_process_by_pid - find a process with a matching PID value.
4657 * @pid: the pid in question.
4659 static struct task_struct *find_process_by_pid(pid_t pid)
4661 return pid ? find_task_by_vpid(pid) : current;
4664 /* Actually do priority change: must hold rq lock. */
4666 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4668 BUG_ON(p->se.on_rq);
4671 p->rt_priority = prio;
4672 p->normal_prio = normal_prio(p);
4673 /* we are holding p->pi_lock already */
4674 p->prio = rt_mutex_getprio(p);
4675 if (rt_prio(p->prio))
4676 p->sched_class = &rt_sched_class;
4678 p->sched_class = &fair_sched_class;
4683 * check the target process has a UID that matches the current process's
4685 static bool check_same_owner(struct task_struct *p)
4687 const struct cred *cred = current_cred(), *pcred;
4691 pcred = __task_cred(p);
4692 match = (cred->euid == pcred->euid ||
4693 cred->euid == pcred->uid);
4698 static int __sched_setscheduler(struct task_struct *p, int policy,
4699 struct sched_param *param, bool user)
4701 int retval, oldprio, oldpolicy = -1, on_rq, running;
4702 unsigned long flags;
4703 const struct sched_class *prev_class;
4707 /* may grab non-irq protected spin_locks */
4708 BUG_ON(in_interrupt());
4710 /* double check policy once rq lock held */
4712 reset_on_fork = p->sched_reset_on_fork;
4713 policy = oldpolicy = p->policy;
4715 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4716 policy &= ~SCHED_RESET_ON_FORK;
4718 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4719 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4720 policy != SCHED_IDLE)
4725 * Valid priorities for SCHED_FIFO and SCHED_RR are
4726 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4727 * SCHED_BATCH and SCHED_IDLE is 0.
4729 if (param->sched_priority < 0 ||
4730 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4731 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4733 if (rt_policy(policy) != (param->sched_priority != 0))
4737 * Allow unprivileged RT tasks to decrease priority:
4739 if (user && !capable(CAP_SYS_NICE)) {
4740 if (rt_policy(policy)) {
4741 unsigned long rlim_rtprio =
4742 task_rlimit(p, RLIMIT_RTPRIO);
4744 /* can't set/change the rt policy */
4745 if (policy != p->policy && !rlim_rtprio)
4748 /* can't increase priority */
4749 if (param->sched_priority > p->rt_priority &&
4750 param->sched_priority > rlim_rtprio)
4754 * Like positive nice levels, dont allow tasks to
4755 * move out of SCHED_IDLE either:
4757 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4760 /* can't change other user's priorities */
4761 if (!check_same_owner(p))
4764 /* Normal users shall not reset the sched_reset_on_fork flag */
4765 if (p->sched_reset_on_fork && !reset_on_fork)
4770 retval = security_task_setscheduler(p, policy, param);
4776 * make sure no PI-waiters arrive (or leave) while we are
4777 * changing the priority of the task:
4779 raw_spin_lock_irqsave(&p->pi_lock, flags);
4781 * To be able to change p->policy safely, the apropriate
4782 * runqueue lock must be held.
4784 rq = __task_rq_lock(p);
4786 #ifdef CONFIG_RT_GROUP_SCHED
4789 * Do not allow realtime tasks into groups that have no runtime
4792 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4793 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4794 __task_rq_unlock(rq);
4795 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4801 /* recheck policy now with rq lock held */
4802 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4803 policy = oldpolicy = -1;
4804 __task_rq_unlock(rq);
4805 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4808 on_rq = p->se.on_rq;
4809 running = task_current(rq, p);
4811 deactivate_task(rq, p, 0);
4813 p->sched_class->put_prev_task(rq, p);
4815 p->sched_reset_on_fork = reset_on_fork;
4818 prev_class = p->sched_class;
4819 __setscheduler(rq, p, policy, param->sched_priority);
4822 p->sched_class->set_curr_task(rq);
4824 activate_task(rq, p, 0);
4826 check_class_changed(rq, p, prev_class, oldprio, running);
4828 __task_rq_unlock(rq);
4829 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4831 rt_mutex_adjust_pi(p);
4837 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4838 * @p: the task in question.
4839 * @policy: new policy.
4840 * @param: structure containing the new RT priority.
4842 * NOTE that the task may be already dead.
4844 int sched_setscheduler(struct task_struct *p, int policy,
4845 struct sched_param *param)
4847 return __sched_setscheduler(p, policy, param, true);
4849 EXPORT_SYMBOL_GPL(sched_setscheduler);
4852 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4853 * @p: the task in question.
4854 * @policy: new policy.
4855 * @param: structure containing the new RT priority.
4857 * Just like sched_setscheduler, only don't bother checking if the
4858 * current context has permission. For example, this is needed in
4859 * stop_machine(): we create temporary high priority worker threads,
4860 * but our caller might not have that capability.
4862 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4863 struct sched_param *param)
4865 return __sched_setscheduler(p, policy, param, false);
4869 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4871 struct sched_param lparam;
4872 struct task_struct *p;
4875 if (!param || pid < 0)
4877 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4882 p = find_process_by_pid(pid);
4884 retval = sched_setscheduler(p, policy, &lparam);
4891 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4892 * @pid: the pid in question.
4893 * @policy: new policy.
4894 * @param: structure containing the new RT priority.
4896 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4897 struct sched_param __user *, param)
4899 /* negative values for policy are not valid */
4903 return do_sched_setscheduler(pid, policy, param);
4907 * sys_sched_setparam - set/change the RT priority of a thread
4908 * @pid: the pid in question.
4909 * @param: structure containing the new RT priority.
4911 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4913 return do_sched_setscheduler(pid, -1, param);
4917 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4918 * @pid: the pid in question.
4920 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4922 struct task_struct *p;
4930 p = find_process_by_pid(pid);
4932 retval = security_task_getscheduler(p);
4935 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4942 * sys_sched_getparam - get the RT priority of a thread
4943 * @pid: the pid in question.
4944 * @param: structure containing the RT priority.
4946 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4948 struct sched_param lp;
4949 struct task_struct *p;
4952 if (!param || pid < 0)
4956 p = find_process_by_pid(pid);
4961 retval = security_task_getscheduler(p);
4965 lp.sched_priority = p->rt_priority;
4969 * This one might sleep, we cannot do it with a spinlock held ...
4971 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4980 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4982 cpumask_var_t cpus_allowed, new_mask;
4983 struct task_struct *p;
4989 p = find_process_by_pid(pid);
4996 /* Prevent p going away */
5000 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5004 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5006 goto out_free_cpus_allowed;
5009 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5012 retval = security_task_setscheduler(p, 0, NULL);
5016 cpuset_cpus_allowed(p, cpus_allowed);
5017 cpumask_and(new_mask, in_mask, cpus_allowed);
5019 retval = set_cpus_allowed_ptr(p, new_mask);
5022 cpuset_cpus_allowed(p, cpus_allowed);
5023 if (!cpumask_subset(new_mask, cpus_allowed)) {
5025 * We must have raced with a concurrent cpuset
5026 * update. Just reset the cpus_allowed to the
5027 * cpuset's cpus_allowed
5029 cpumask_copy(new_mask, cpus_allowed);
5034 free_cpumask_var(new_mask);
5035 out_free_cpus_allowed:
5036 free_cpumask_var(cpus_allowed);
5043 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5044 struct cpumask *new_mask)
5046 if (len < cpumask_size())
5047 cpumask_clear(new_mask);
5048 else if (len > cpumask_size())
5049 len = cpumask_size();
5051 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5055 * sys_sched_setaffinity - set the cpu affinity of a process
5056 * @pid: pid of the process
5057 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5058 * @user_mask_ptr: user-space pointer to the new cpu mask
5060 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5061 unsigned long __user *, user_mask_ptr)
5063 cpumask_var_t new_mask;
5066 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5069 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5071 retval = sched_setaffinity(pid, new_mask);
5072 free_cpumask_var(new_mask);
5076 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5078 struct task_struct *p;
5079 unsigned long flags;
5087 p = find_process_by_pid(pid);
5091 retval = security_task_getscheduler(p);
5095 rq = task_rq_lock(p, &flags);
5096 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5097 task_rq_unlock(rq, &flags);
5107 * sys_sched_getaffinity - get the cpu affinity of a process
5108 * @pid: pid of the process
5109 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5110 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5112 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5113 unsigned long __user *, user_mask_ptr)
5118 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
5120 if (len & (sizeof(unsigned long)-1))
5123 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5126 ret = sched_getaffinity(pid, mask);
5128 size_t retlen = min_t(size_t, len, cpumask_size());
5130 if (copy_to_user(user_mask_ptr, mask, retlen))
5135 free_cpumask_var(mask);
5141 * sys_sched_yield - yield the current processor to other threads.
5143 * This function yields the current CPU to other tasks. If there are no
5144 * other threads running on this CPU then this function will return.
5146 SYSCALL_DEFINE0(sched_yield)
5148 struct rq *rq = this_rq_lock();
5150 schedstat_inc(rq, yld_count);
5151 current->sched_class->yield_task(rq);
5154 * Since we are going to call schedule() anyway, there's
5155 * no need to preempt or enable interrupts:
5157 __release(rq->lock);
5158 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5159 do_raw_spin_unlock(&rq->lock);
5160 preempt_enable_no_resched();
5167 static inline int should_resched(void)
5169 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5172 static void __cond_resched(void)
5174 add_preempt_count(PREEMPT_ACTIVE);
5176 sub_preempt_count(PREEMPT_ACTIVE);
5179 int __sched _cond_resched(void)
5181 if (should_resched()) {
5187 EXPORT_SYMBOL(_cond_resched);
5190 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5191 * call schedule, and on return reacquire the lock.
5193 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5194 * operations here to prevent schedule() from being called twice (once via
5195 * spin_unlock(), once by hand).
5197 int __cond_resched_lock(spinlock_t *lock)
5199 int resched = should_resched();
5202 lockdep_assert_held(lock);
5204 if (spin_needbreak(lock) || resched) {
5215 EXPORT_SYMBOL(__cond_resched_lock);
5217 int __sched __cond_resched_softirq(void)
5219 BUG_ON(!in_softirq());
5221 if (should_resched()) {
5229 EXPORT_SYMBOL(__cond_resched_softirq);
5232 * yield - yield the current processor to other threads.
5234 * This is a shortcut for kernel-space yielding - it marks the
5235 * thread runnable and calls sys_sched_yield().
5237 void __sched yield(void)
5239 set_current_state(TASK_RUNNING);
5242 EXPORT_SYMBOL(yield);
5245 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5246 * that process accounting knows that this is a task in IO wait state.
5248 void __sched io_schedule(void)
5250 struct rq *rq = raw_rq();
5252 delayacct_blkio_start();
5253 atomic_inc(&rq->nr_iowait);
5254 current->in_iowait = 1;
5256 current->in_iowait = 0;
5257 atomic_dec(&rq->nr_iowait);
5258 delayacct_blkio_end();
5260 EXPORT_SYMBOL(io_schedule);
5262 long __sched io_schedule_timeout(long timeout)
5264 struct rq *rq = raw_rq();
5267 delayacct_blkio_start();
5268 atomic_inc(&rq->nr_iowait);
5269 current->in_iowait = 1;
5270 ret = schedule_timeout(timeout);
5271 current->in_iowait = 0;
5272 atomic_dec(&rq->nr_iowait);
5273 delayacct_blkio_end();
5278 * sys_sched_get_priority_max - return maximum RT priority.
5279 * @policy: scheduling class.
5281 * this syscall returns the maximum rt_priority that can be used
5282 * by a given scheduling class.
5284 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5291 ret = MAX_USER_RT_PRIO-1;
5303 * sys_sched_get_priority_min - return minimum RT priority.
5304 * @policy: scheduling class.
5306 * this syscall returns the minimum rt_priority that can be used
5307 * by a given scheduling class.
5309 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5327 * sys_sched_rr_get_interval - return the default timeslice of a process.
5328 * @pid: pid of the process.
5329 * @interval: userspace pointer to the timeslice value.
5331 * this syscall writes the default timeslice value of a given process
5332 * into the user-space timespec buffer. A value of '0' means infinity.
5334 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5335 struct timespec __user *, interval)
5337 struct task_struct *p;
5338 unsigned int time_slice;
5339 unsigned long flags;
5349 p = find_process_by_pid(pid);
5353 retval = security_task_getscheduler(p);
5357 rq = task_rq_lock(p, &flags);
5358 time_slice = p->sched_class->get_rr_interval(rq, p);
5359 task_rq_unlock(rq, &flags);
5362 jiffies_to_timespec(time_slice, &t);
5363 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5371 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5373 void sched_show_task(struct task_struct *p)
5375 unsigned long free = 0;
5378 state = p->state ? __ffs(p->state) + 1 : 0;
5379 printk(KERN_INFO "%-15.15s %c", p->comm,
5380 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5381 #if BITS_PER_LONG == 32
5382 if (state == TASK_RUNNING)
5383 printk(KERN_CONT " running ");
5385 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5387 if (state == TASK_RUNNING)
5388 printk(KERN_CONT " running task ");
5390 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5392 #ifdef CONFIG_DEBUG_STACK_USAGE
5393 free = stack_not_used(p);
5395 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5396 task_pid_nr(p), task_pid_nr(p->real_parent),
5397 (unsigned long)task_thread_info(p)->flags);
5399 show_stack(p, NULL);
5402 void show_state_filter(unsigned long state_filter)
5404 struct task_struct *g, *p;
5406 #if BITS_PER_LONG == 32
5408 " task PC stack pid father\n");
5411 " task PC stack pid father\n");
5413 read_lock(&tasklist_lock);
5414 do_each_thread(g, p) {
5416 * reset the NMI-timeout, listing all files on a slow
5417 * console might take alot of time:
5419 touch_nmi_watchdog();
5420 if (!state_filter || (p->state & state_filter))
5422 } while_each_thread(g, p);
5424 touch_all_softlockup_watchdogs();
5426 #ifdef CONFIG_SCHED_DEBUG
5427 sysrq_sched_debug_show();
5429 read_unlock(&tasklist_lock);
5431 * Only show locks if all tasks are dumped:
5434 debug_show_all_locks();
5437 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5439 idle->sched_class = &idle_sched_class;
5443 * init_idle - set up an idle thread for a given CPU
5444 * @idle: task in question
5445 * @cpu: cpu the idle task belongs to
5447 * NOTE: this function does not set the idle thread's NEED_RESCHED
5448 * flag, to make booting more robust.
5450 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5452 struct rq *rq = cpu_rq(cpu);
5453 unsigned long flags;
5455 raw_spin_lock_irqsave(&rq->lock, flags);
5458 idle->state = TASK_RUNNING;
5459 idle->se.exec_start = sched_clock();
5461 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5463 * We're having a chicken and egg problem, even though we are
5464 * holding rq->lock, the cpu isn't yet set to this cpu so the
5465 * lockdep check in task_group() will fail.
5467 * Similar case to sched_fork(). / Alternatively we could
5468 * use task_rq_lock() here and obtain the other rq->lock.
5473 __set_task_cpu(idle, cpu);
5476 rq->curr = rq->idle = idle;
5477 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5480 raw_spin_unlock_irqrestore(&rq->lock, flags);
5482 /* Set the preempt count _outside_ the spinlocks! */
5483 #if defined(CONFIG_PREEMPT)
5484 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5486 task_thread_info(idle)->preempt_count = 0;
5489 * The idle tasks have their own, simple scheduling class:
5491 idle->sched_class = &idle_sched_class;
5492 ftrace_graph_init_task(idle);
5496 * In a system that switches off the HZ timer nohz_cpu_mask
5497 * indicates which cpus entered this state. This is used
5498 * in the rcu update to wait only for active cpus. For system
5499 * which do not switch off the HZ timer nohz_cpu_mask should
5500 * always be CPU_BITS_NONE.
5502 cpumask_var_t nohz_cpu_mask;
5505 * Increase the granularity value when there are more CPUs,
5506 * because with more CPUs the 'effective latency' as visible
5507 * to users decreases. But the relationship is not linear,
5508 * so pick a second-best guess by going with the log2 of the
5511 * This idea comes from the SD scheduler of Con Kolivas:
5513 static int get_update_sysctl_factor(void)
5515 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5516 unsigned int factor;
5518 switch (sysctl_sched_tunable_scaling) {
5519 case SCHED_TUNABLESCALING_NONE:
5522 case SCHED_TUNABLESCALING_LINEAR:
5525 case SCHED_TUNABLESCALING_LOG:
5527 factor = 1 + ilog2(cpus);
5534 static void update_sysctl(void)
5536 unsigned int factor = get_update_sysctl_factor();
5538 #define SET_SYSCTL(name) \
5539 (sysctl_##name = (factor) * normalized_sysctl_##name)
5540 SET_SYSCTL(sched_min_granularity);
5541 SET_SYSCTL(sched_latency);
5542 SET_SYSCTL(sched_wakeup_granularity);
5543 SET_SYSCTL(sched_shares_ratelimit);
5547 static inline void sched_init_granularity(void)
5554 * This is how migration works:
5556 * 1) we invoke migration_cpu_stop() on the target CPU using
5558 * 2) stopper starts to run (implicitly forcing the migrated thread
5560 * 3) it checks whether the migrated task is still in the wrong runqueue.
5561 * 4) if it's in the wrong runqueue then the migration thread removes
5562 * it and puts it into the right queue.
5563 * 5) stopper completes and stop_one_cpu() returns and the migration
5568 * Change a given task's CPU affinity. Migrate the thread to a
5569 * proper CPU and schedule it away if the CPU it's executing on
5570 * is removed from the allowed bitmask.
5572 * NOTE: the caller must have a valid reference to the task, the
5573 * task must not exit() & deallocate itself prematurely. The
5574 * call is not atomic; no spinlocks may be held.
5576 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5578 unsigned long flags;
5580 unsigned int dest_cpu;
5584 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5585 * drop the rq->lock and still rely on ->cpus_allowed.
5588 while (task_is_waking(p))
5590 rq = task_rq_lock(p, &flags);
5591 if (task_is_waking(p)) {
5592 task_rq_unlock(rq, &flags);
5596 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5601 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5602 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5607 if (p->sched_class->set_cpus_allowed)
5608 p->sched_class->set_cpus_allowed(p, new_mask);
5610 cpumask_copy(&p->cpus_allowed, new_mask);
5611 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5614 /* Can the task run on the task's current CPU? If so, we're done */
5615 if (cpumask_test_cpu(task_cpu(p), new_mask))
5618 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5619 if (migrate_task(p, dest_cpu)) {
5620 struct migration_arg arg = { p, dest_cpu };
5621 /* Need help from migration thread: drop lock and wait. */
5622 task_rq_unlock(rq, &flags);
5623 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5624 tlb_migrate_finish(p->mm);
5628 task_rq_unlock(rq, &flags);
5632 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5635 * Move (not current) task off this cpu, onto dest cpu. We're doing
5636 * this because either it can't run here any more (set_cpus_allowed()
5637 * away from this CPU, or CPU going down), or because we're
5638 * attempting to rebalance this task on exec (sched_exec).
5640 * So we race with normal scheduler movements, but that's OK, as long
5641 * as the task is no longer on this CPU.
5643 * Returns non-zero if task was successfully migrated.
5645 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5647 struct rq *rq_dest, *rq_src;
5650 if (unlikely(!cpu_active(dest_cpu)))
5653 rq_src = cpu_rq(src_cpu);
5654 rq_dest = cpu_rq(dest_cpu);
5656 double_rq_lock(rq_src, rq_dest);
5657 /* Already moved. */
5658 if (task_cpu(p) != src_cpu)
5660 /* Affinity changed (again). */
5661 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5665 * If we're not on a rq, the next wake-up will ensure we're
5669 deactivate_task(rq_src, p, 0);
5670 set_task_cpu(p, dest_cpu);
5671 activate_task(rq_dest, p, 0);
5672 check_preempt_curr(rq_dest, p, 0);
5677 double_rq_unlock(rq_src, rq_dest);
5682 * migration_cpu_stop - this will be executed by a highprio stopper thread
5683 * and performs thread migration by bumping thread off CPU then
5684 * 'pushing' onto another runqueue.
5686 static int migration_cpu_stop(void *data)
5688 struct migration_arg *arg = data;
5691 * The original target cpu might have gone down and we might
5692 * be on another cpu but it doesn't matter.
5694 local_irq_disable();
5695 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5700 #ifdef CONFIG_HOTPLUG_CPU
5702 * Figure out where task on dead CPU should go, use force if necessary.
5704 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5706 struct rq *rq = cpu_rq(dead_cpu);
5707 int needs_cpu, uninitialized_var(dest_cpu);
5708 unsigned long flags;
5710 local_irq_save(flags);
5712 raw_spin_lock(&rq->lock);
5713 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5715 dest_cpu = select_fallback_rq(dead_cpu, p);
5716 raw_spin_unlock(&rq->lock);
5718 * It can only fail if we race with set_cpus_allowed(),
5719 * in the racer should migrate the task anyway.
5722 __migrate_task(p, dead_cpu, dest_cpu);
5723 local_irq_restore(flags);
5727 * While a dead CPU has no uninterruptible tasks queued at this point,
5728 * it might still have a nonzero ->nr_uninterruptible counter, because
5729 * for performance reasons the counter is not stricly tracking tasks to
5730 * their home CPUs. So we just add the counter to another CPU's counter,
5731 * to keep the global sum constant after CPU-down:
5733 static void migrate_nr_uninterruptible(struct rq *rq_src)
5735 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5736 unsigned long flags;
5738 local_irq_save(flags);
5739 double_rq_lock(rq_src, rq_dest);
5740 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5741 rq_src->nr_uninterruptible = 0;
5742 double_rq_unlock(rq_src, rq_dest);
5743 local_irq_restore(flags);
5746 /* Run through task list and migrate tasks from the dead cpu. */
5747 static void migrate_live_tasks(int src_cpu)
5749 struct task_struct *p, *t;
5751 read_lock(&tasklist_lock);
5753 do_each_thread(t, p) {
5757 if (task_cpu(p) == src_cpu)
5758 move_task_off_dead_cpu(src_cpu, p);
5759 } while_each_thread(t, p);
5761 read_unlock(&tasklist_lock);
5765 * Schedules idle task to be the next runnable task on current CPU.
5766 * It does so by boosting its priority to highest possible.
5767 * Used by CPU offline code.
5769 void sched_idle_next(void)
5771 int this_cpu = smp_processor_id();
5772 struct rq *rq = cpu_rq(this_cpu);
5773 struct task_struct *p = rq->idle;
5774 unsigned long flags;
5776 /* cpu has to be offline */
5777 BUG_ON(cpu_online(this_cpu));
5780 * Strictly not necessary since rest of the CPUs are stopped by now
5781 * and interrupts disabled on the current cpu.
5783 raw_spin_lock_irqsave(&rq->lock, flags);
5785 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5787 activate_task(rq, p, 0);
5789 raw_spin_unlock_irqrestore(&rq->lock, flags);
5793 * Ensures that the idle task is using init_mm right before its cpu goes
5796 void idle_task_exit(void)
5798 struct mm_struct *mm = current->active_mm;
5800 BUG_ON(cpu_online(smp_processor_id()));
5803 switch_mm(mm, &init_mm, current);
5807 /* called under rq->lock with disabled interrupts */
5808 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5810 struct rq *rq = cpu_rq(dead_cpu);
5812 /* Must be exiting, otherwise would be on tasklist. */
5813 BUG_ON(!p->exit_state);
5815 /* Cannot have done final schedule yet: would have vanished. */
5816 BUG_ON(p->state == TASK_DEAD);
5821 * Drop lock around migration; if someone else moves it,
5822 * that's OK. No task can be added to this CPU, so iteration is
5825 raw_spin_unlock_irq(&rq->lock);
5826 move_task_off_dead_cpu(dead_cpu, p);
5827 raw_spin_lock_irq(&rq->lock);
5832 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5833 static void migrate_dead_tasks(unsigned int dead_cpu)
5835 struct rq *rq = cpu_rq(dead_cpu);
5836 struct task_struct *next;
5839 if (!rq->nr_running)
5841 next = pick_next_task(rq);
5844 next->sched_class->put_prev_task(rq, next);
5845 migrate_dead(dead_cpu, next);
5851 * remove the tasks which were accounted by rq from calc_load_tasks.
5853 static void calc_global_load_remove(struct rq *rq)
5855 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5856 rq->calc_load_active = 0;
5858 #endif /* CONFIG_HOTPLUG_CPU */
5860 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5862 static struct ctl_table sd_ctl_dir[] = {
5864 .procname = "sched_domain",
5870 static struct ctl_table sd_ctl_root[] = {
5872 .procname = "kernel",
5874 .child = sd_ctl_dir,
5879 static struct ctl_table *sd_alloc_ctl_entry(int n)
5881 struct ctl_table *entry =
5882 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5887 static void sd_free_ctl_entry(struct ctl_table **tablep)
5889 struct ctl_table *entry;
5892 * In the intermediate directories, both the child directory and
5893 * procname are dynamically allocated and could fail but the mode
5894 * will always be set. In the lowest directory the names are
5895 * static strings and all have proc handlers.
5897 for (entry = *tablep; entry->mode; entry++) {
5899 sd_free_ctl_entry(&entry->child);
5900 if (entry->proc_handler == NULL)
5901 kfree(entry->procname);
5909 set_table_entry(struct ctl_table *entry,
5910 const char *procname, void *data, int maxlen,
5911 mode_t mode, proc_handler *proc_handler)
5913 entry->procname = procname;
5915 entry->maxlen = maxlen;
5917 entry->proc_handler = proc_handler;
5920 static struct ctl_table *
5921 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5923 struct ctl_table *table = sd_alloc_ctl_entry(13);
5928 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5929 sizeof(long), 0644, proc_doulongvec_minmax);
5930 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5931 sizeof(long), 0644, proc_doulongvec_minmax);
5932 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5933 sizeof(int), 0644, proc_dointvec_minmax);
5934 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5935 sizeof(int), 0644, proc_dointvec_minmax);
5936 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5937 sizeof(int), 0644, proc_dointvec_minmax);
5938 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5939 sizeof(int), 0644, proc_dointvec_minmax);
5940 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5941 sizeof(int), 0644, proc_dointvec_minmax);
5942 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5943 sizeof(int), 0644, proc_dointvec_minmax);
5944 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5945 sizeof(int), 0644, proc_dointvec_minmax);
5946 set_table_entry(&table[9], "cache_nice_tries",
5947 &sd->cache_nice_tries,
5948 sizeof(int), 0644, proc_dointvec_minmax);
5949 set_table_entry(&table[10], "flags", &sd->flags,
5950 sizeof(int), 0644, proc_dointvec_minmax);
5951 set_table_entry(&table[11], "name", sd->name,
5952 CORENAME_MAX_SIZE, 0444, proc_dostring);
5953 /* &table[12] is terminator */
5958 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5960 struct ctl_table *entry, *table;
5961 struct sched_domain *sd;
5962 int domain_num = 0, i;
5965 for_each_domain(cpu, sd)
5967 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5972 for_each_domain(cpu, sd) {
5973 snprintf(buf, 32, "domain%d", i);
5974 entry->procname = kstrdup(buf, GFP_KERNEL);
5976 entry->child = sd_alloc_ctl_domain_table(sd);
5983 static struct ctl_table_header *sd_sysctl_header;
5984 static void register_sched_domain_sysctl(void)
5986 int i, cpu_num = num_possible_cpus();
5987 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5990 WARN_ON(sd_ctl_dir[0].child);
5991 sd_ctl_dir[0].child = entry;
5996 for_each_possible_cpu(i) {
5997 snprintf(buf, 32, "cpu%d", i);
5998 entry->procname = kstrdup(buf, GFP_KERNEL);
6000 entry->child = sd_alloc_ctl_cpu_table(i);
6004 WARN_ON(sd_sysctl_header);
6005 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6008 /* may be called multiple times per register */
6009 static void unregister_sched_domain_sysctl(void)
6011 if (sd_sysctl_header)
6012 unregister_sysctl_table(sd_sysctl_header);
6013 sd_sysctl_header = NULL;
6014 if (sd_ctl_dir[0].child)
6015 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6018 static void register_sched_domain_sysctl(void)
6021 static void unregister_sched_domain_sysctl(void)
6026 static void set_rq_online(struct rq *rq)
6029 const struct sched_class *class;
6031 cpumask_set_cpu(rq->cpu, rq->rd->online);
6034 for_each_class(class) {
6035 if (class->rq_online)
6036 class->rq_online(rq);
6041 static void set_rq_offline(struct rq *rq)
6044 const struct sched_class *class;
6046 for_each_class(class) {
6047 if (class->rq_offline)
6048 class->rq_offline(rq);
6051 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6057 * migration_call - callback that gets triggered when a CPU is added.
6058 * Here we can start up the necessary migration thread for the new CPU.
6060 static int __cpuinit
6061 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6063 int cpu = (long)hcpu;
6064 unsigned long flags;
6065 struct rq *rq = cpu_rq(cpu);
6069 case CPU_UP_PREPARE:
6070 case CPU_UP_PREPARE_FROZEN:
6071 rq->calc_load_update = calc_load_update;
6075 case CPU_ONLINE_FROZEN:
6076 /* Update our root-domain */
6077 raw_spin_lock_irqsave(&rq->lock, flags);
6079 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6083 raw_spin_unlock_irqrestore(&rq->lock, flags);
6086 #ifdef CONFIG_HOTPLUG_CPU
6088 case CPU_DEAD_FROZEN:
6089 migrate_live_tasks(cpu);
6090 /* Idle task back to normal (off runqueue, low prio) */
6091 raw_spin_lock_irq(&rq->lock);
6092 deactivate_task(rq, rq->idle, 0);
6093 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6094 rq->idle->sched_class = &idle_sched_class;
6095 migrate_dead_tasks(cpu);
6096 raw_spin_unlock_irq(&rq->lock);
6097 migrate_nr_uninterruptible(rq);
6098 BUG_ON(rq->nr_running != 0);
6099 calc_global_load_remove(rq);
6103 case CPU_DYING_FROZEN:
6104 /* Update our root-domain */
6105 raw_spin_lock_irqsave(&rq->lock, flags);
6107 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6110 raw_spin_unlock_irqrestore(&rq->lock, flags);
6118 * Register at high priority so that task migration (migrate_all_tasks)
6119 * happens before everything else. This has to be lower priority than
6120 * the notifier in the perf_event subsystem, though.
6122 static struct notifier_block __cpuinitdata migration_notifier = {
6123 .notifier_call = migration_call,
6124 .priority = CPU_PRI_MIGRATION,
6127 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6128 unsigned long action, void *hcpu)
6130 switch (action & ~CPU_TASKS_FROZEN) {
6132 case CPU_DOWN_FAILED:
6133 set_cpu_active((long)hcpu, true);
6140 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6141 unsigned long action, void *hcpu)
6143 switch (action & ~CPU_TASKS_FROZEN) {
6144 case CPU_DOWN_PREPARE:
6145 set_cpu_active((long)hcpu, false);
6152 static int __init migration_init(void)
6154 void *cpu = (void *)(long)smp_processor_id();
6157 /* Initialize migration for the boot CPU */
6158 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6159 BUG_ON(err == NOTIFY_BAD);
6160 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6161 register_cpu_notifier(&migration_notifier);
6163 /* Register cpu active notifiers */
6164 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6165 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6169 early_initcall(migration_init);
6174 #ifdef CONFIG_SCHED_DEBUG
6176 static __read_mostly int sched_domain_debug_enabled;
6178 static int __init sched_domain_debug_setup(char *str)
6180 sched_domain_debug_enabled = 1;
6184 early_param("sched_debug", sched_domain_debug_setup);
6186 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6187 struct cpumask *groupmask)
6189 struct sched_group *group = sd->groups;
6192 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6193 cpumask_clear(groupmask);
6195 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6197 if (!(sd->flags & SD_LOAD_BALANCE)) {
6198 printk("does not load-balance\n");
6200 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6205 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6207 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6208 printk(KERN_ERR "ERROR: domain->span does not contain "
6211 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6212 printk(KERN_ERR "ERROR: domain->groups does not contain"
6216 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6220 printk(KERN_ERR "ERROR: group is NULL\n");
6224 if (!group->cpu_power) {
6225 printk(KERN_CONT "\n");
6226 printk(KERN_ERR "ERROR: domain->cpu_power not "
6231 if (!cpumask_weight(sched_group_cpus(group))) {
6232 printk(KERN_CONT "\n");
6233 printk(KERN_ERR "ERROR: empty group\n");
6237 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6238 printk(KERN_CONT "\n");
6239 printk(KERN_ERR "ERROR: repeated CPUs\n");
6243 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6245 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6247 printk(KERN_CONT " %s", str);
6248 if (group->cpu_power != SCHED_LOAD_SCALE) {
6249 printk(KERN_CONT " (cpu_power = %d)",
6253 group = group->next;
6254 } while (group != sd->groups);
6255 printk(KERN_CONT "\n");
6257 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6258 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6261 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6262 printk(KERN_ERR "ERROR: parent span is not a superset "
6263 "of domain->span\n");
6267 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6269 cpumask_var_t groupmask;
6272 if (!sched_domain_debug_enabled)
6276 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6280 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6282 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6283 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6288 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6295 free_cpumask_var(groupmask);
6297 #else /* !CONFIG_SCHED_DEBUG */
6298 # define sched_domain_debug(sd, cpu) do { } while (0)
6299 #endif /* CONFIG_SCHED_DEBUG */
6301 static int sd_degenerate(struct sched_domain *sd)
6303 if (cpumask_weight(sched_domain_span(sd)) == 1)
6306 /* Following flags need at least 2 groups */
6307 if (sd->flags & (SD_LOAD_BALANCE |
6308 SD_BALANCE_NEWIDLE |
6312 SD_SHARE_PKG_RESOURCES)) {
6313 if (sd->groups != sd->groups->next)
6317 /* Following flags don't use groups */
6318 if (sd->flags & (SD_WAKE_AFFINE))
6325 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6327 unsigned long cflags = sd->flags, pflags = parent->flags;
6329 if (sd_degenerate(parent))
6332 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6335 /* Flags needing groups don't count if only 1 group in parent */
6336 if (parent->groups == parent->groups->next) {
6337 pflags &= ~(SD_LOAD_BALANCE |
6338 SD_BALANCE_NEWIDLE |
6342 SD_SHARE_PKG_RESOURCES);
6343 if (nr_node_ids == 1)
6344 pflags &= ~SD_SERIALIZE;
6346 if (~cflags & pflags)
6352 static void free_rootdomain(struct root_domain *rd)
6354 synchronize_sched();
6356 cpupri_cleanup(&rd->cpupri);
6358 free_cpumask_var(rd->rto_mask);
6359 free_cpumask_var(rd->online);
6360 free_cpumask_var(rd->span);
6364 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6366 struct root_domain *old_rd = NULL;
6367 unsigned long flags;
6369 raw_spin_lock_irqsave(&rq->lock, flags);
6374 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6377 cpumask_clear_cpu(rq->cpu, old_rd->span);
6380 * If we dont want to free the old_rt yet then
6381 * set old_rd to NULL to skip the freeing later
6384 if (!atomic_dec_and_test(&old_rd->refcount))
6388 atomic_inc(&rd->refcount);
6391 cpumask_set_cpu(rq->cpu, rd->span);
6392 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6395 raw_spin_unlock_irqrestore(&rq->lock, flags);
6398 free_rootdomain(old_rd);
6401 static int init_rootdomain(struct root_domain *rd)
6403 memset(rd, 0, sizeof(*rd));
6405 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6407 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6409 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6412 if (cpupri_init(&rd->cpupri) != 0)
6417 free_cpumask_var(rd->rto_mask);
6419 free_cpumask_var(rd->online);
6421 free_cpumask_var(rd->span);
6426 static void init_defrootdomain(void)
6428 init_rootdomain(&def_root_domain);
6430 atomic_set(&def_root_domain.refcount, 1);
6433 static struct root_domain *alloc_rootdomain(void)
6435 struct root_domain *rd;
6437 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6441 if (init_rootdomain(rd) != 0) {
6450 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6451 * hold the hotplug lock.
6454 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6456 struct rq *rq = cpu_rq(cpu);
6457 struct sched_domain *tmp;
6459 for (tmp = sd; tmp; tmp = tmp->parent)
6460 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6462 /* Remove the sched domains which do not contribute to scheduling. */
6463 for (tmp = sd; tmp; ) {
6464 struct sched_domain *parent = tmp->parent;
6468 if (sd_parent_degenerate(tmp, parent)) {
6469 tmp->parent = parent->parent;
6471 parent->parent->child = tmp;
6476 if (sd && sd_degenerate(sd)) {
6482 sched_domain_debug(sd, cpu);
6484 rq_attach_root(rq, rd);
6485 rcu_assign_pointer(rq->sd, sd);
6488 /* cpus with isolated domains */
6489 static cpumask_var_t cpu_isolated_map;
6491 /* Setup the mask of cpus configured for isolated domains */
6492 static int __init isolated_cpu_setup(char *str)
6494 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6495 cpulist_parse(str, cpu_isolated_map);
6499 __setup("isolcpus=", isolated_cpu_setup);
6502 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6503 * to a function which identifies what group(along with sched group) a CPU
6504 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6505 * (due to the fact that we keep track of groups covered with a struct cpumask).
6507 * init_sched_build_groups will build a circular linked list of the groups
6508 * covered by the given span, and will set each group's ->cpumask correctly,
6509 * and ->cpu_power to 0.
6512 init_sched_build_groups(const struct cpumask *span,
6513 const struct cpumask *cpu_map,
6514 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6515 struct sched_group **sg,
6516 struct cpumask *tmpmask),
6517 struct cpumask *covered, struct cpumask *tmpmask)
6519 struct sched_group *first = NULL, *last = NULL;
6522 cpumask_clear(covered);
6524 for_each_cpu(i, span) {
6525 struct sched_group *sg;
6526 int group = group_fn(i, cpu_map, &sg, tmpmask);
6529 if (cpumask_test_cpu(i, covered))
6532 cpumask_clear(sched_group_cpus(sg));
6535 for_each_cpu(j, span) {
6536 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6539 cpumask_set_cpu(j, covered);
6540 cpumask_set_cpu(j, sched_group_cpus(sg));
6551 #define SD_NODES_PER_DOMAIN 16
6556 * find_next_best_node - find the next node to include in a sched_domain
6557 * @node: node whose sched_domain we're building
6558 * @used_nodes: nodes already in the sched_domain
6560 * Find the next node to include in a given scheduling domain. Simply
6561 * finds the closest node not already in the @used_nodes map.
6563 * Should use nodemask_t.
6565 static int find_next_best_node(int node, nodemask_t *used_nodes)
6567 int i, n, val, min_val, best_node = 0;
6571 for (i = 0; i < nr_node_ids; i++) {
6572 /* Start at @node */
6573 n = (node + i) % nr_node_ids;
6575 if (!nr_cpus_node(n))
6578 /* Skip already used nodes */
6579 if (node_isset(n, *used_nodes))
6582 /* Simple min distance search */
6583 val = node_distance(node, n);
6585 if (val < min_val) {
6591 node_set(best_node, *used_nodes);
6596 * sched_domain_node_span - get a cpumask for a node's sched_domain
6597 * @node: node whose cpumask we're constructing
6598 * @span: resulting cpumask
6600 * Given a node, construct a good cpumask for its sched_domain to span. It
6601 * should be one that prevents unnecessary balancing, but also spreads tasks
6604 static void sched_domain_node_span(int node, struct cpumask *span)
6606 nodemask_t used_nodes;
6609 cpumask_clear(span);
6610 nodes_clear(used_nodes);
6612 cpumask_or(span, span, cpumask_of_node(node));
6613 node_set(node, used_nodes);
6615 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6616 int next_node = find_next_best_node(node, &used_nodes);
6618 cpumask_or(span, span, cpumask_of_node(next_node));
6621 #endif /* CONFIG_NUMA */
6623 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6626 * The cpus mask in sched_group and sched_domain hangs off the end.
6628 * ( See the the comments in include/linux/sched.h:struct sched_group
6629 * and struct sched_domain. )
6631 struct static_sched_group {
6632 struct sched_group sg;
6633 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6636 struct static_sched_domain {
6637 struct sched_domain sd;
6638 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6644 cpumask_var_t domainspan;
6645 cpumask_var_t covered;
6646 cpumask_var_t notcovered;
6648 cpumask_var_t nodemask;
6649 cpumask_var_t this_sibling_map;
6650 cpumask_var_t this_core_map;
6651 cpumask_var_t send_covered;
6652 cpumask_var_t tmpmask;
6653 struct sched_group **sched_group_nodes;
6654 struct root_domain *rd;
6658 sa_sched_groups = 0,
6663 sa_this_sibling_map,
6665 sa_sched_group_nodes,
6675 * SMT sched-domains:
6677 #ifdef CONFIG_SCHED_SMT
6678 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6679 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6682 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6683 struct sched_group **sg, struct cpumask *unused)
6686 *sg = &per_cpu(sched_groups, cpu).sg;
6689 #endif /* CONFIG_SCHED_SMT */
6692 * multi-core sched-domains:
6694 #ifdef CONFIG_SCHED_MC
6695 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6696 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6697 #endif /* CONFIG_SCHED_MC */
6699 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6701 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6702 struct sched_group **sg, struct cpumask *mask)
6706 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6707 group = cpumask_first(mask);
6709 *sg = &per_cpu(sched_group_core, group).sg;
6712 #elif defined(CONFIG_SCHED_MC)
6714 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6715 struct sched_group **sg, struct cpumask *unused)
6718 *sg = &per_cpu(sched_group_core, cpu).sg;
6723 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6724 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6727 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6728 struct sched_group **sg, struct cpumask *mask)
6731 #ifdef CONFIG_SCHED_MC
6732 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6733 group = cpumask_first(mask);
6734 #elif defined(CONFIG_SCHED_SMT)
6735 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6736 group = cpumask_first(mask);
6741 *sg = &per_cpu(sched_group_phys, group).sg;
6747 * The init_sched_build_groups can't handle what we want to do with node
6748 * groups, so roll our own. Now each node has its own list of groups which
6749 * gets dynamically allocated.
6751 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6752 static struct sched_group ***sched_group_nodes_bycpu;
6754 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6755 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6757 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6758 struct sched_group **sg,
6759 struct cpumask *nodemask)
6763 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6764 group = cpumask_first(nodemask);
6767 *sg = &per_cpu(sched_group_allnodes, group).sg;
6771 static void init_numa_sched_groups_power(struct sched_group *group_head)
6773 struct sched_group *sg = group_head;
6779 for_each_cpu(j, sched_group_cpus(sg)) {
6780 struct sched_domain *sd;
6782 sd = &per_cpu(phys_domains, j).sd;
6783 if (j != group_first_cpu(sd->groups)) {
6785 * Only add "power" once for each
6791 sg->cpu_power += sd->groups->cpu_power;
6794 } while (sg != group_head);
6797 static int build_numa_sched_groups(struct s_data *d,
6798 const struct cpumask *cpu_map, int num)
6800 struct sched_domain *sd;
6801 struct sched_group *sg, *prev;
6804 cpumask_clear(d->covered);
6805 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6806 if (cpumask_empty(d->nodemask)) {
6807 d->sched_group_nodes[num] = NULL;
6811 sched_domain_node_span(num, d->domainspan);
6812 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6814 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6817 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6821 d->sched_group_nodes[num] = sg;
6823 for_each_cpu(j, d->nodemask) {
6824 sd = &per_cpu(node_domains, j).sd;
6829 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6831 cpumask_or(d->covered, d->covered, d->nodemask);
6834 for (j = 0; j < nr_node_ids; j++) {
6835 n = (num + j) % nr_node_ids;
6836 cpumask_complement(d->notcovered, d->covered);
6837 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6838 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6839 if (cpumask_empty(d->tmpmask))
6841 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6842 if (cpumask_empty(d->tmpmask))
6844 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6848 "Can not alloc domain group for node %d\n", j);
6852 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6853 sg->next = prev->next;
6854 cpumask_or(d->covered, d->covered, d->tmpmask);
6861 #endif /* CONFIG_NUMA */
6864 /* Free memory allocated for various sched_group structures */
6865 static void free_sched_groups(const struct cpumask *cpu_map,
6866 struct cpumask *nodemask)
6870 for_each_cpu(cpu, cpu_map) {
6871 struct sched_group **sched_group_nodes
6872 = sched_group_nodes_bycpu[cpu];
6874 if (!sched_group_nodes)
6877 for (i = 0; i < nr_node_ids; i++) {
6878 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6880 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6881 if (cpumask_empty(nodemask))
6891 if (oldsg != sched_group_nodes[i])
6894 kfree(sched_group_nodes);
6895 sched_group_nodes_bycpu[cpu] = NULL;
6898 #else /* !CONFIG_NUMA */
6899 static void free_sched_groups(const struct cpumask *cpu_map,
6900 struct cpumask *nodemask)
6903 #endif /* CONFIG_NUMA */
6906 * Initialize sched groups cpu_power.
6908 * cpu_power indicates the capacity of sched group, which is used while
6909 * distributing the load between different sched groups in a sched domain.
6910 * Typically cpu_power for all the groups in a sched domain will be same unless
6911 * there are asymmetries in the topology. If there are asymmetries, group
6912 * having more cpu_power will pickup more load compared to the group having
6915 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6917 struct sched_domain *child;
6918 struct sched_group *group;
6922 WARN_ON(!sd || !sd->groups);
6924 if (cpu != group_first_cpu(sd->groups))
6929 sd->groups->cpu_power = 0;
6932 power = SCHED_LOAD_SCALE;
6933 weight = cpumask_weight(sched_domain_span(sd));
6935 * SMT siblings share the power of a single core.
6936 * Usually multiple threads get a better yield out of
6937 * that one core than a single thread would have,
6938 * reflect that in sd->smt_gain.
6940 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6941 power *= sd->smt_gain;
6943 power >>= SCHED_LOAD_SHIFT;
6945 sd->groups->cpu_power += power;
6950 * Add cpu_power of each child group to this groups cpu_power.
6952 group = child->groups;
6954 sd->groups->cpu_power += group->cpu_power;
6955 group = group->next;
6956 } while (group != child->groups);
6960 * Initializers for schedule domains
6961 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6964 #ifdef CONFIG_SCHED_DEBUG
6965 # define SD_INIT_NAME(sd, type) sd->name = #type
6967 # define SD_INIT_NAME(sd, type) do { } while (0)
6970 #define SD_INIT(sd, type) sd_init_##type(sd)
6972 #define SD_INIT_FUNC(type) \
6973 static noinline void sd_init_##type(struct sched_domain *sd) \
6975 memset(sd, 0, sizeof(*sd)); \
6976 *sd = SD_##type##_INIT; \
6977 sd->level = SD_LV_##type; \
6978 SD_INIT_NAME(sd, type); \
6983 SD_INIT_FUNC(ALLNODES)
6986 #ifdef CONFIG_SCHED_SMT
6987 SD_INIT_FUNC(SIBLING)
6989 #ifdef CONFIG_SCHED_MC
6993 static int default_relax_domain_level = -1;
6995 static int __init setup_relax_domain_level(char *str)
6999 val = simple_strtoul(str, NULL, 0);
7000 if (val < SD_LV_MAX)
7001 default_relax_domain_level = val;
7005 __setup("relax_domain_level=", setup_relax_domain_level);
7007 static void set_domain_attribute(struct sched_domain *sd,
7008 struct sched_domain_attr *attr)
7012 if (!attr || attr->relax_domain_level < 0) {
7013 if (default_relax_domain_level < 0)
7016 request = default_relax_domain_level;
7018 request = attr->relax_domain_level;
7019 if (request < sd->level) {
7020 /* turn off idle balance on this domain */
7021 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7023 /* turn on idle balance on this domain */
7024 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
7028 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
7029 const struct cpumask *cpu_map)
7032 case sa_sched_groups:
7033 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
7034 d->sched_group_nodes = NULL;
7036 free_rootdomain(d->rd); /* fall through */
7038 free_cpumask_var(d->tmpmask); /* fall through */
7039 case sa_send_covered:
7040 free_cpumask_var(d->send_covered); /* fall through */
7041 case sa_this_core_map:
7042 free_cpumask_var(d->this_core_map); /* fall through */
7043 case sa_this_sibling_map:
7044 free_cpumask_var(d->this_sibling_map); /* fall through */
7046 free_cpumask_var(d->nodemask); /* fall through */
7047 case sa_sched_group_nodes:
7049 kfree(d->sched_group_nodes); /* fall through */
7051 free_cpumask_var(d->notcovered); /* fall through */
7053 free_cpumask_var(d->covered); /* fall through */
7055 free_cpumask_var(d->domainspan); /* fall through */
7062 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
7063 const struct cpumask *cpu_map)
7066 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
7068 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
7069 return sa_domainspan;
7070 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
7072 /* Allocate the per-node list of sched groups */
7073 d->sched_group_nodes = kcalloc(nr_node_ids,
7074 sizeof(struct sched_group *), GFP_KERNEL);
7075 if (!d->sched_group_nodes) {
7076 printk(KERN_WARNING "Can not alloc sched group node list\n");
7077 return sa_notcovered;
7079 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
7081 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
7082 return sa_sched_group_nodes;
7083 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
7085 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
7086 return sa_this_sibling_map;
7087 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
7088 return sa_this_core_map;
7089 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
7090 return sa_send_covered;
7091 d->rd = alloc_rootdomain();
7093 printk(KERN_WARNING "Cannot alloc root domain\n");
7096 return sa_rootdomain;
7099 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
7100 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
7102 struct sched_domain *sd = NULL;
7104 struct sched_domain *parent;
7107 if (cpumask_weight(cpu_map) >
7108 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
7109 sd = &per_cpu(allnodes_domains, i).sd;
7110 SD_INIT(sd, ALLNODES);
7111 set_domain_attribute(sd, attr);
7112 cpumask_copy(sched_domain_span(sd), cpu_map);
7113 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7118 sd = &per_cpu(node_domains, i).sd;
7120 set_domain_attribute(sd, attr);
7121 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7122 sd->parent = parent;
7125 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7130 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7131 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7132 struct sched_domain *parent, int i)
7134 struct sched_domain *sd;
7135 sd = &per_cpu(phys_domains, i).sd;
7137 set_domain_attribute(sd, attr);
7138 cpumask_copy(sched_domain_span(sd), d->nodemask);
7139 sd->parent = parent;
7142 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7146 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7147 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7148 struct sched_domain *parent, int i)
7150 struct sched_domain *sd = parent;
7151 #ifdef CONFIG_SCHED_MC
7152 sd = &per_cpu(core_domains, i).sd;
7154 set_domain_attribute(sd, attr);
7155 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7156 sd->parent = parent;
7158 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7163 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7164 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7165 struct sched_domain *parent, int i)
7167 struct sched_domain *sd = parent;
7168 #ifdef CONFIG_SCHED_SMT
7169 sd = &per_cpu(cpu_domains, i).sd;
7170 SD_INIT(sd, SIBLING);
7171 set_domain_attribute(sd, attr);
7172 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7173 sd->parent = parent;
7175 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7180 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7181 const struct cpumask *cpu_map, int cpu)
7184 #ifdef CONFIG_SCHED_SMT
7185 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7186 cpumask_and(d->this_sibling_map, cpu_map,
7187 topology_thread_cpumask(cpu));
7188 if (cpu == cpumask_first(d->this_sibling_map))
7189 init_sched_build_groups(d->this_sibling_map, cpu_map,
7191 d->send_covered, d->tmpmask);
7194 #ifdef CONFIG_SCHED_MC
7195 case SD_LV_MC: /* set up multi-core groups */
7196 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7197 if (cpu == cpumask_first(d->this_core_map))
7198 init_sched_build_groups(d->this_core_map, cpu_map,
7200 d->send_covered, d->tmpmask);
7203 case SD_LV_CPU: /* set up physical groups */
7204 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7205 if (!cpumask_empty(d->nodemask))
7206 init_sched_build_groups(d->nodemask, cpu_map,
7208 d->send_covered, d->tmpmask);
7211 case SD_LV_ALLNODES:
7212 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7213 d->send_covered, d->tmpmask);
7222 * Build sched domains for a given set of cpus and attach the sched domains
7223 * to the individual cpus
7225 static int __build_sched_domains(const struct cpumask *cpu_map,
7226 struct sched_domain_attr *attr)
7228 enum s_alloc alloc_state = sa_none;
7230 struct sched_domain *sd;
7236 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7237 if (alloc_state != sa_rootdomain)
7239 alloc_state = sa_sched_groups;
7242 * Set up domains for cpus specified by the cpu_map.
7244 for_each_cpu(i, cpu_map) {
7245 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7248 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7249 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7250 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7251 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7254 for_each_cpu(i, cpu_map) {
7255 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7256 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7259 /* Set up physical groups */
7260 for (i = 0; i < nr_node_ids; i++)
7261 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7264 /* Set up node groups */
7266 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7268 for (i = 0; i < nr_node_ids; i++)
7269 if (build_numa_sched_groups(&d, cpu_map, i))
7273 /* Calculate CPU power for physical packages and nodes */
7274 #ifdef CONFIG_SCHED_SMT
7275 for_each_cpu(i, cpu_map) {
7276 sd = &per_cpu(cpu_domains, i).sd;
7277 init_sched_groups_power(i, sd);
7280 #ifdef CONFIG_SCHED_MC
7281 for_each_cpu(i, cpu_map) {
7282 sd = &per_cpu(core_domains, i).sd;
7283 init_sched_groups_power(i, sd);
7287 for_each_cpu(i, cpu_map) {
7288 sd = &per_cpu(phys_domains, i).sd;
7289 init_sched_groups_power(i, sd);
7293 for (i = 0; i < nr_node_ids; i++)
7294 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7296 if (d.sd_allnodes) {
7297 struct sched_group *sg;
7299 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7301 init_numa_sched_groups_power(sg);
7305 /* Attach the domains */
7306 for_each_cpu(i, cpu_map) {
7307 #ifdef CONFIG_SCHED_SMT
7308 sd = &per_cpu(cpu_domains, i).sd;
7309 #elif defined(CONFIG_SCHED_MC)
7310 sd = &per_cpu(core_domains, i).sd;
7312 sd = &per_cpu(phys_domains, i).sd;
7314 cpu_attach_domain(sd, d.rd, i);
7317 d.sched_group_nodes = NULL; /* don't free this we still need it */
7318 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7322 __free_domain_allocs(&d, alloc_state, cpu_map);
7326 static int build_sched_domains(const struct cpumask *cpu_map)
7328 return __build_sched_domains(cpu_map, NULL);
7331 static cpumask_var_t *doms_cur; /* current sched domains */
7332 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7333 static struct sched_domain_attr *dattr_cur;
7334 /* attribues of custom domains in 'doms_cur' */
7337 * Special case: If a kmalloc of a doms_cur partition (array of
7338 * cpumask) fails, then fallback to a single sched domain,
7339 * as determined by the single cpumask fallback_doms.
7341 static cpumask_var_t fallback_doms;
7344 * arch_update_cpu_topology lets virtualized architectures update the
7345 * cpu core maps. It is supposed to return 1 if the topology changed
7346 * or 0 if it stayed the same.
7348 int __attribute__((weak)) arch_update_cpu_topology(void)
7353 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7356 cpumask_var_t *doms;
7358 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7361 for (i = 0; i < ndoms; i++) {
7362 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7363 free_sched_domains(doms, i);
7370 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7373 for (i = 0; i < ndoms; i++)
7374 free_cpumask_var(doms[i]);
7379 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7380 * For now this just excludes isolated cpus, but could be used to
7381 * exclude other special cases in the future.
7383 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7387 arch_update_cpu_topology();
7389 doms_cur = alloc_sched_domains(ndoms_cur);
7391 doms_cur = &fallback_doms;
7392 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7394 err = build_sched_domains(doms_cur[0]);
7395 register_sched_domain_sysctl();
7400 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7401 struct cpumask *tmpmask)
7403 free_sched_groups(cpu_map, tmpmask);
7407 * Detach sched domains from a group of cpus specified in cpu_map
7408 * These cpus will now be attached to the NULL domain
7410 static void detach_destroy_domains(const struct cpumask *cpu_map)
7412 /* Save because hotplug lock held. */
7413 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7416 for_each_cpu(i, cpu_map)
7417 cpu_attach_domain(NULL, &def_root_domain, i);
7418 synchronize_sched();
7419 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7422 /* handle null as "default" */
7423 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7424 struct sched_domain_attr *new, int idx_new)
7426 struct sched_domain_attr tmp;
7433 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7434 new ? (new + idx_new) : &tmp,
7435 sizeof(struct sched_domain_attr));
7439 * Partition sched domains as specified by the 'ndoms_new'
7440 * cpumasks in the array doms_new[] of cpumasks. This compares
7441 * doms_new[] to the current sched domain partitioning, doms_cur[].
7442 * It destroys each deleted domain and builds each new domain.
7444 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7445 * The masks don't intersect (don't overlap.) We should setup one
7446 * sched domain for each mask. CPUs not in any of the cpumasks will
7447 * not be load balanced. If the same cpumask appears both in the
7448 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7451 * The passed in 'doms_new' should be allocated using
7452 * alloc_sched_domains. This routine takes ownership of it and will
7453 * free_sched_domains it when done with it. If the caller failed the
7454 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7455 * and partition_sched_domains() will fallback to the single partition
7456 * 'fallback_doms', it also forces the domains to be rebuilt.
7458 * If doms_new == NULL it will be replaced with cpu_online_mask.
7459 * ndoms_new == 0 is a special case for destroying existing domains,
7460 * and it will not create the default domain.
7462 * Call with hotplug lock held
7464 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7465 struct sched_domain_attr *dattr_new)
7470 mutex_lock(&sched_domains_mutex);
7472 /* always unregister in case we don't destroy any domains */
7473 unregister_sched_domain_sysctl();
7475 /* Let architecture update cpu core mappings. */
7476 new_topology = arch_update_cpu_topology();
7478 n = doms_new ? ndoms_new : 0;
7480 /* Destroy deleted domains */
7481 for (i = 0; i < ndoms_cur; i++) {
7482 for (j = 0; j < n && !new_topology; j++) {
7483 if (cpumask_equal(doms_cur[i], doms_new[j])
7484 && dattrs_equal(dattr_cur, i, dattr_new, j))
7487 /* no match - a current sched domain not in new doms_new[] */
7488 detach_destroy_domains(doms_cur[i]);
7493 if (doms_new == NULL) {
7495 doms_new = &fallback_doms;
7496 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7497 WARN_ON_ONCE(dattr_new);
7500 /* Build new domains */
7501 for (i = 0; i < ndoms_new; i++) {
7502 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7503 if (cpumask_equal(doms_new[i], doms_cur[j])
7504 && dattrs_equal(dattr_new, i, dattr_cur, j))
7507 /* no match - add a new doms_new */
7508 __build_sched_domains(doms_new[i],
7509 dattr_new ? dattr_new + i : NULL);
7514 /* Remember the new sched domains */
7515 if (doms_cur != &fallback_doms)
7516 free_sched_domains(doms_cur, ndoms_cur);
7517 kfree(dattr_cur); /* kfree(NULL) is safe */
7518 doms_cur = doms_new;
7519 dattr_cur = dattr_new;
7520 ndoms_cur = ndoms_new;
7522 register_sched_domain_sysctl();
7524 mutex_unlock(&sched_domains_mutex);
7527 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7528 static void arch_reinit_sched_domains(void)
7532 /* Destroy domains first to force the rebuild */
7533 partition_sched_domains(0, NULL, NULL);
7535 rebuild_sched_domains();
7539 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7541 unsigned int level = 0;
7543 if (sscanf(buf, "%u", &level) != 1)
7547 * level is always be positive so don't check for
7548 * level < POWERSAVINGS_BALANCE_NONE which is 0
7549 * What happens on 0 or 1 byte write,
7550 * need to check for count as well?
7553 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7557 sched_smt_power_savings = level;
7559 sched_mc_power_savings = level;
7561 arch_reinit_sched_domains();
7566 #ifdef CONFIG_SCHED_MC
7567 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7568 struct sysdev_class_attribute *attr,
7571 return sprintf(page, "%u\n", sched_mc_power_savings);
7573 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7574 struct sysdev_class_attribute *attr,
7575 const char *buf, size_t count)
7577 return sched_power_savings_store(buf, count, 0);
7579 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7580 sched_mc_power_savings_show,
7581 sched_mc_power_savings_store);
7584 #ifdef CONFIG_SCHED_SMT
7585 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7586 struct sysdev_class_attribute *attr,
7589 return sprintf(page, "%u\n", sched_smt_power_savings);
7591 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7592 struct sysdev_class_attribute *attr,
7593 const char *buf, size_t count)
7595 return sched_power_savings_store(buf, count, 1);
7597 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7598 sched_smt_power_savings_show,
7599 sched_smt_power_savings_store);
7602 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7606 #ifdef CONFIG_SCHED_SMT
7608 err = sysfs_create_file(&cls->kset.kobj,
7609 &attr_sched_smt_power_savings.attr);
7611 #ifdef CONFIG_SCHED_MC
7612 if (!err && mc_capable())
7613 err = sysfs_create_file(&cls->kset.kobj,
7614 &attr_sched_mc_power_savings.attr);
7618 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7621 * Update cpusets according to cpu_active mask. If cpusets are
7622 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7623 * around partition_sched_domains().
7625 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7628 switch (action & ~CPU_TASKS_FROZEN) {
7630 case CPU_DOWN_FAILED:
7631 cpuset_update_active_cpus();
7638 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7641 switch (action & ~CPU_TASKS_FROZEN) {
7642 case CPU_DOWN_PREPARE:
7643 cpuset_update_active_cpus();
7650 static int update_runtime(struct notifier_block *nfb,
7651 unsigned long action, void *hcpu)
7653 int cpu = (int)(long)hcpu;
7656 case CPU_DOWN_PREPARE:
7657 case CPU_DOWN_PREPARE_FROZEN:
7658 disable_runtime(cpu_rq(cpu));
7661 case CPU_DOWN_FAILED:
7662 case CPU_DOWN_FAILED_FROZEN:
7664 case CPU_ONLINE_FROZEN:
7665 enable_runtime(cpu_rq(cpu));
7673 void __init sched_init_smp(void)
7675 cpumask_var_t non_isolated_cpus;
7677 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7678 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7680 #if defined(CONFIG_NUMA)
7681 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7683 BUG_ON(sched_group_nodes_bycpu == NULL);
7686 mutex_lock(&sched_domains_mutex);
7687 arch_init_sched_domains(cpu_active_mask);
7688 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7689 if (cpumask_empty(non_isolated_cpus))
7690 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7691 mutex_unlock(&sched_domains_mutex);
7694 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7695 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7697 /* RT runtime code needs to handle some hotplug events */
7698 hotcpu_notifier(update_runtime, 0);
7702 /* Move init over to a non-isolated CPU */
7703 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7705 sched_init_granularity();
7706 free_cpumask_var(non_isolated_cpus);
7708 init_sched_rt_class();
7711 void __init sched_init_smp(void)
7713 sched_init_granularity();
7715 #endif /* CONFIG_SMP */
7717 const_debug unsigned int sysctl_timer_migration = 1;
7719 int in_sched_functions(unsigned long addr)
7721 return in_lock_functions(addr) ||
7722 (addr >= (unsigned long)__sched_text_start
7723 && addr < (unsigned long)__sched_text_end);
7726 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7728 cfs_rq->tasks_timeline = RB_ROOT;
7729 INIT_LIST_HEAD(&cfs_rq->tasks);
7730 #ifdef CONFIG_FAIR_GROUP_SCHED
7733 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7736 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7738 struct rt_prio_array *array;
7741 array = &rt_rq->active;
7742 for (i = 0; i < MAX_RT_PRIO; i++) {
7743 INIT_LIST_HEAD(array->queue + i);
7744 __clear_bit(i, array->bitmap);
7746 /* delimiter for bitsearch: */
7747 __set_bit(MAX_RT_PRIO, array->bitmap);
7749 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7750 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7752 rt_rq->highest_prio.next = MAX_RT_PRIO;
7756 rt_rq->rt_nr_migratory = 0;
7757 rt_rq->overloaded = 0;
7758 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7762 rt_rq->rt_throttled = 0;
7763 rt_rq->rt_runtime = 0;
7764 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7766 #ifdef CONFIG_RT_GROUP_SCHED
7767 rt_rq->rt_nr_boosted = 0;
7772 #ifdef CONFIG_FAIR_GROUP_SCHED
7773 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7774 struct sched_entity *se, int cpu, int add,
7775 struct sched_entity *parent)
7777 struct rq *rq = cpu_rq(cpu);
7778 tg->cfs_rq[cpu] = cfs_rq;
7779 init_cfs_rq(cfs_rq, rq);
7782 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7785 /* se could be NULL for init_task_group */
7790 se->cfs_rq = &rq->cfs;
7792 se->cfs_rq = parent->my_q;
7795 se->load.weight = tg->shares;
7796 se->load.inv_weight = 0;
7797 se->parent = parent;
7801 #ifdef CONFIG_RT_GROUP_SCHED
7802 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7803 struct sched_rt_entity *rt_se, int cpu, int add,
7804 struct sched_rt_entity *parent)
7806 struct rq *rq = cpu_rq(cpu);
7808 tg->rt_rq[cpu] = rt_rq;
7809 init_rt_rq(rt_rq, rq);
7811 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7813 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7815 tg->rt_se[cpu] = rt_se;
7820 rt_se->rt_rq = &rq->rt;
7822 rt_se->rt_rq = parent->my_q;
7824 rt_se->my_q = rt_rq;
7825 rt_se->parent = parent;
7826 INIT_LIST_HEAD(&rt_se->run_list);
7830 void __init sched_init(void)
7833 unsigned long alloc_size = 0, ptr;
7835 #ifdef CONFIG_FAIR_GROUP_SCHED
7836 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7838 #ifdef CONFIG_RT_GROUP_SCHED
7839 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7841 #ifdef CONFIG_CPUMASK_OFFSTACK
7842 alloc_size += num_possible_cpus() * cpumask_size();
7845 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7847 #ifdef CONFIG_FAIR_GROUP_SCHED
7848 init_task_group.se = (struct sched_entity **)ptr;
7849 ptr += nr_cpu_ids * sizeof(void **);
7851 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7852 ptr += nr_cpu_ids * sizeof(void **);
7854 #endif /* CONFIG_FAIR_GROUP_SCHED */
7855 #ifdef CONFIG_RT_GROUP_SCHED
7856 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7857 ptr += nr_cpu_ids * sizeof(void **);
7859 init_task_group.rt_rq = (struct rt_rq **)ptr;
7860 ptr += nr_cpu_ids * sizeof(void **);
7862 #endif /* CONFIG_RT_GROUP_SCHED */
7863 #ifdef CONFIG_CPUMASK_OFFSTACK
7864 for_each_possible_cpu(i) {
7865 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7866 ptr += cpumask_size();
7868 #endif /* CONFIG_CPUMASK_OFFSTACK */
7872 init_defrootdomain();
7875 init_rt_bandwidth(&def_rt_bandwidth,
7876 global_rt_period(), global_rt_runtime());
7878 #ifdef CONFIG_RT_GROUP_SCHED
7879 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7880 global_rt_period(), global_rt_runtime());
7881 #endif /* CONFIG_RT_GROUP_SCHED */
7883 #ifdef CONFIG_CGROUP_SCHED
7884 list_add(&init_task_group.list, &task_groups);
7885 INIT_LIST_HEAD(&init_task_group.children);
7887 #endif /* CONFIG_CGROUP_SCHED */
7889 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7890 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7891 __alignof__(unsigned long));
7893 for_each_possible_cpu(i) {
7897 raw_spin_lock_init(&rq->lock);
7899 rq->calc_load_active = 0;
7900 rq->calc_load_update = jiffies + LOAD_FREQ;
7901 init_cfs_rq(&rq->cfs, rq);
7902 init_rt_rq(&rq->rt, rq);
7903 #ifdef CONFIG_FAIR_GROUP_SCHED
7904 init_task_group.shares = init_task_group_load;
7905 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7906 #ifdef CONFIG_CGROUP_SCHED
7908 * How much cpu bandwidth does init_task_group get?
7910 * In case of task-groups formed thr' the cgroup filesystem, it
7911 * gets 100% of the cpu resources in the system. This overall
7912 * system cpu resource is divided among the tasks of
7913 * init_task_group and its child task-groups in a fair manner,
7914 * based on each entity's (task or task-group's) weight
7915 * (se->load.weight).
7917 * In other words, if init_task_group has 10 tasks of weight
7918 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7919 * then A0's share of the cpu resource is:
7921 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7923 * We achieve this by letting init_task_group's tasks sit
7924 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7926 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7928 #endif /* CONFIG_FAIR_GROUP_SCHED */
7930 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7931 #ifdef CONFIG_RT_GROUP_SCHED
7932 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7933 #ifdef CONFIG_CGROUP_SCHED
7934 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7938 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7939 rq->cpu_load[j] = 0;
7941 rq->last_load_update_tick = jiffies;
7946 rq->cpu_power = SCHED_LOAD_SCALE;
7947 rq->post_schedule = 0;
7948 rq->active_balance = 0;
7949 rq->next_balance = jiffies;
7954 rq->avg_idle = 2*sysctl_sched_migration_cost;
7955 rq_attach_root(rq, &def_root_domain);
7957 rq->nohz_balance_kick = 0;
7958 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7962 atomic_set(&rq->nr_iowait, 0);
7965 set_load_weight(&init_task);
7967 #ifdef CONFIG_PREEMPT_NOTIFIERS
7968 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7972 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7975 #ifdef CONFIG_RT_MUTEXES
7976 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7980 * The boot idle thread does lazy MMU switching as well:
7982 atomic_inc(&init_mm.mm_count);
7983 enter_lazy_tlb(&init_mm, current);
7986 * Make us the idle thread. Technically, schedule() should not be
7987 * called from this thread, however somewhere below it might be,
7988 * but because we are the idle thread, we just pick up running again
7989 * when this runqueue becomes "idle".
7991 init_idle(current, smp_processor_id());
7993 calc_load_update = jiffies + LOAD_FREQ;
7996 * During early bootup we pretend to be a normal task:
7998 current->sched_class = &fair_sched_class;
8000 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8001 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
8004 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
8005 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
8006 atomic_set(&nohz.load_balancer, nr_cpu_ids);
8007 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
8008 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
8010 /* May be allocated at isolcpus cmdline parse time */
8011 if (cpu_isolated_map == NULL)
8012 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
8017 scheduler_running = 1;
8020 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8021 static inline int preempt_count_equals(int preempt_offset)
8023 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
8025 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
8028 static int __might_sleep_init_called;
8029 int __init __might_sleep_init(void)
8031 __might_sleep_init_called = 1;
8034 early_initcall(__might_sleep_init);
8036 void __might_sleep(const char *file, int line, int preempt_offset)
8039 static unsigned long prev_jiffy; /* ratelimiting */
8041 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
8044 if (system_state != SYSTEM_RUNNING &&
8045 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
8047 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8049 prev_jiffy = jiffies;
8052 "BUG: sleeping function called from invalid context at %s:%d\n",
8055 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8056 in_atomic(), irqs_disabled(),
8057 current->pid, current->comm);
8059 debug_show_held_locks(current);
8060 if (irqs_disabled())
8061 print_irqtrace_events(current);
8065 EXPORT_SYMBOL(__might_sleep);
8068 #ifdef CONFIG_MAGIC_SYSRQ
8069 static void normalize_task(struct rq *rq, struct task_struct *p)
8073 on_rq = p->se.on_rq;
8075 deactivate_task(rq, p, 0);
8076 __setscheduler(rq, p, SCHED_NORMAL, 0);
8078 activate_task(rq, p, 0);
8079 resched_task(rq->curr);
8083 void normalize_rt_tasks(void)
8085 struct task_struct *g, *p;
8086 unsigned long flags;
8089 read_lock_irqsave(&tasklist_lock, flags);
8090 do_each_thread(g, p) {
8092 * Only normalize user tasks:
8097 p->se.exec_start = 0;
8098 #ifdef CONFIG_SCHEDSTATS
8099 p->se.statistics.wait_start = 0;
8100 p->se.statistics.sleep_start = 0;
8101 p->se.statistics.block_start = 0;
8106 * Renice negative nice level userspace
8109 if (TASK_NICE(p) < 0 && p->mm)
8110 set_user_nice(p, 0);
8114 raw_spin_lock(&p->pi_lock);
8115 rq = __task_rq_lock(p);
8117 normalize_task(rq, p);
8119 __task_rq_unlock(rq);
8120 raw_spin_unlock(&p->pi_lock);
8121 } while_each_thread(g, p);
8123 read_unlock_irqrestore(&tasklist_lock, flags);
8126 #endif /* CONFIG_MAGIC_SYSRQ */
8128 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8130 * These functions are only useful for the IA64 MCA handling, or kdb.
8132 * They can only be called when the whole system has been
8133 * stopped - every CPU needs to be quiescent, and no scheduling
8134 * activity can take place. Using them for anything else would
8135 * be a serious bug, and as a result, they aren't even visible
8136 * under any other configuration.
8140 * curr_task - return the current task for a given cpu.
8141 * @cpu: the processor in question.
8143 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8145 struct task_struct *curr_task(int cpu)
8147 return cpu_curr(cpu);
8150 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8154 * set_curr_task - set the current task for a given cpu.
8155 * @cpu: the processor in question.
8156 * @p: the task pointer to set.
8158 * Description: This function must only be used when non-maskable interrupts
8159 * are serviced on a separate stack. It allows the architecture to switch the
8160 * notion of the current task on a cpu in a non-blocking manner. This function
8161 * must be called with all CPU's synchronized, and interrupts disabled, the
8162 * and caller must save the original value of the current task (see
8163 * curr_task() above) and restore that value before reenabling interrupts and
8164 * re-starting the system.
8166 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8168 void set_curr_task(int cpu, struct task_struct *p)
8175 #ifdef CONFIG_FAIR_GROUP_SCHED
8176 static void free_fair_sched_group(struct task_group *tg)
8180 for_each_possible_cpu(i) {
8182 kfree(tg->cfs_rq[i]);
8192 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8194 struct cfs_rq *cfs_rq;
8195 struct sched_entity *se;
8199 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8202 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8206 tg->shares = NICE_0_LOAD;
8208 for_each_possible_cpu(i) {
8211 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8212 GFP_KERNEL, cpu_to_node(i));
8216 se = kzalloc_node(sizeof(struct sched_entity),
8217 GFP_KERNEL, cpu_to_node(i));
8221 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8232 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8234 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8235 &cpu_rq(cpu)->leaf_cfs_rq_list);
8238 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8240 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8242 #else /* !CONFG_FAIR_GROUP_SCHED */
8243 static inline void free_fair_sched_group(struct task_group *tg)
8248 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8253 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8257 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8260 #endif /* CONFIG_FAIR_GROUP_SCHED */
8262 #ifdef CONFIG_RT_GROUP_SCHED
8263 static void free_rt_sched_group(struct task_group *tg)
8267 destroy_rt_bandwidth(&tg->rt_bandwidth);
8269 for_each_possible_cpu(i) {
8271 kfree(tg->rt_rq[i]);
8273 kfree(tg->rt_se[i]);
8281 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8283 struct rt_rq *rt_rq;
8284 struct sched_rt_entity *rt_se;
8288 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8291 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8295 init_rt_bandwidth(&tg->rt_bandwidth,
8296 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8298 for_each_possible_cpu(i) {
8301 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8302 GFP_KERNEL, cpu_to_node(i));
8306 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8307 GFP_KERNEL, cpu_to_node(i));
8311 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8322 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8324 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8325 &cpu_rq(cpu)->leaf_rt_rq_list);
8328 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8330 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8332 #else /* !CONFIG_RT_GROUP_SCHED */
8333 static inline void free_rt_sched_group(struct task_group *tg)
8338 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8343 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8347 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8350 #endif /* CONFIG_RT_GROUP_SCHED */
8352 #ifdef CONFIG_CGROUP_SCHED
8353 static void free_sched_group(struct task_group *tg)
8355 free_fair_sched_group(tg);
8356 free_rt_sched_group(tg);
8360 /* allocate runqueue etc for a new task group */
8361 struct task_group *sched_create_group(struct task_group *parent)
8363 struct task_group *tg;
8364 unsigned long flags;
8367 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8369 return ERR_PTR(-ENOMEM);
8371 if (!alloc_fair_sched_group(tg, parent))
8374 if (!alloc_rt_sched_group(tg, parent))
8377 spin_lock_irqsave(&task_group_lock, flags);
8378 for_each_possible_cpu(i) {
8379 register_fair_sched_group(tg, i);
8380 register_rt_sched_group(tg, i);
8382 list_add_rcu(&tg->list, &task_groups);
8384 WARN_ON(!parent); /* root should already exist */
8386 tg->parent = parent;
8387 INIT_LIST_HEAD(&tg->children);
8388 list_add_rcu(&tg->siblings, &parent->children);
8389 spin_unlock_irqrestore(&task_group_lock, flags);
8394 free_sched_group(tg);
8395 return ERR_PTR(-ENOMEM);
8398 /* rcu callback to free various structures associated with a task group */
8399 static void free_sched_group_rcu(struct rcu_head *rhp)
8401 /* now it should be safe to free those cfs_rqs */
8402 free_sched_group(container_of(rhp, struct task_group, rcu));
8405 /* Destroy runqueue etc associated with a task group */
8406 void sched_destroy_group(struct task_group *tg)
8408 unsigned long flags;
8411 spin_lock_irqsave(&task_group_lock, flags);
8412 for_each_possible_cpu(i) {
8413 unregister_fair_sched_group(tg, i);
8414 unregister_rt_sched_group(tg, i);
8416 list_del_rcu(&tg->list);
8417 list_del_rcu(&tg->siblings);
8418 spin_unlock_irqrestore(&task_group_lock, flags);
8420 /* wait for possible concurrent references to cfs_rqs complete */
8421 call_rcu(&tg->rcu, free_sched_group_rcu);
8424 /* change task's runqueue when it moves between groups.
8425 * The caller of this function should have put the task in its new group
8426 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8427 * reflect its new group.
8429 void sched_move_task(struct task_struct *tsk)
8432 unsigned long flags;
8435 rq = task_rq_lock(tsk, &flags);
8437 running = task_current(rq, tsk);
8438 on_rq = tsk->se.on_rq;
8441 dequeue_task(rq, tsk, 0);
8442 if (unlikely(running))
8443 tsk->sched_class->put_prev_task(rq, tsk);
8445 #ifdef CONFIG_FAIR_GROUP_SCHED
8446 if (tsk->sched_class->prep_move_group)
8447 tsk->sched_class->prep_move_group(tsk, on_rq);
8450 set_task_rq(tsk, task_cpu(tsk));
8452 #ifdef CONFIG_FAIR_GROUP_SCHED
8453 if (tsk->sched_class->moved_group)
8454 tsk->sched_class->moved_group(tsk, on_rq);
8457 if (unlikely(running))
8458 tsk->sched_class->set_curr_task(rq);
8460 enqueue_task(rq, tsk, 0);
8462 task_rq_unlock(rq, &flags);
8464 #endif /* CONFIG_CGROUP_SCHED */
8466 #ifdef CONFIG_FAIR_GROUP_SCHED
8467 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8469 struct cfs_rq *cfs_rq = se->cfs_rq;
8474 dequeue_entity(cfs_rq, se, 0);
8476 se->load.weight = shares;
8477 se->load.inv_weight = 0;
8480 enqueue_entity(cfs_rq, se, 0);
8483 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8485 struct cfs_rq *cfs_rq = se->cfs_rq;
8486 struct rq *rq = cfs_rq->rq;
8487 unsigned long flags;
8489 raw_spin_lock_irqsave(&rq->lock, flags);
8490 __set_se_shares(se, shares);
8491 raw_spin_unlock_irqrestore(&rq->lock, flags);
8494 static DEFINE_MUTEX(shares_mutex);
8496 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8499 unsigned long flags;
8502 * We can't change the weight of the root cgroup.
8507 if (shares < MIN_SHARES)
8508 shares = MIN_SHARES;
8509 else if (shares > MAX_SHARES)
8510 shares = MAX_SHARES;
8512 mutex_lock(&shares_mutex);
8513 if (tg->shares == shares)
8516 spin_lock_irqsave(&task_group_lock, flags);
8517 for_each_possible_cpu(i)
8518 unregister_fair_sched_group(tg, i);
8519 list_del_rcu(&tg->siblings);
8520 spin_unlock_irqrestore(&task_group_lock, flags);
8522 /* wait for any ongoing reference to this group to finish */
8523 synchronize_sched();
8526 * Now we are free to modify the group's share on each cpu
8527 * w/o tripping rebalance_share or load_balance_fair.
8529 tg->shares = shares;
8530 for_each_possible_cpu(i) {
8534 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8535 set_se_shares(tg->se[i], shares);
8539 * Enable load balance activity on this group, by inserting it back on
8540 * each cpu's rq->leaf_cfs_rq_list.
8542 spin_lock_irqsave(&task_group_lock, flags);
8543 for_each_possible_cpu(i)
8544 register_fair_sched_group(tg, i);
8545 list_add_rcu(&tg->siblings, &tg->parent->children);
8546 spin_unlock_irqrestore(&task_group_lock, flags);
8548 mutex_unlock(&shares_mutex);
8552 unsigned long sched_group_shares(struct task_group *tg)
8558 #ifdef CONFIG_RT_GROUP_SCHED
8560 * Ensure that the real time constraints are schedulable.
8562 static DEFINE_MUTEX(rt_constraints_mutex);
8564 static unsigned long to_ratio(u64 period, u64 runtime)
8566 if (runtime == RUNTIME_INF)
8569 return div64_u64(runtime << 20, period);
8572 /* Must be called with tasklist_lock held */
8573 static inline int tg_has_rt_tasks(struct task_group *tg)
8575 struct task_struct *g, *p;
8577 do_each_thread(g, p) {
8578 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8580 } while_each_thread(g, p);
8585 struct rt_schedulable_data {
8586 struct task_group *tg;
8591 static int tg_schedulable(struct task_group *tg, void *data)
8593 struct rt_schedulable_data *d = data;
8594 struct task_group *child;
8595 unsigned long total, sum = 0;
8596 u64 period, runtime;
8598 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8599 runtime = tg->rt_bandwidth.rt_runtime;
8602 period = d->rt_period;
8603 runtime = d->rt_runtime;
8607 * Cannot have more runtime than the period.
8609 if (runtime > period && runtime != RUNTIME_INF)
8613 * Ensure we don't starve existing RT tasks.
8615 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8618 total = to_ratio(period, runtime);
8621 * Nobody can have more than the global setting allows.
8623 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8627 * The sum of our children's runtime should not exceed our own.
8629 list_for_each_entry_rcu(child, &tg->children, siblings) {
8630 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8631 runtime = child->rt_bandwidth.rt_runtime;
8633 if (child == d->tg) {
8634 period = d->rt_period;
8635 runtime = d->rt_runtime;
8638 sum += to_ratio(period, runtime);
8647 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8649 struct rt_schedulable_data data = {
8651 .rt_period = period,
8652 .rt_runtime = runtime,
8655 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8658 static int tg_set_bandwidth(struct task_group *tg,
8659 u64 rt_period, u64 rt_runtime)
8663 mutex_lock(&rt_constraints_mutex);
8664 read_lock(&tasklist_lock);
8665 err = __rt_schedulable(tg, rt_period, rt_runtime);
8669 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8670 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8671 tg->rt_bandwidth.rt_runtime = rt_runtime;
8673 for_each_possible_cpu(i) {
8674 struct rt_rq *rt_rq = tg->rt_rq[i];
8676 raw_spin_lock(&rt_rq->rt_runtime_lock);
8677 rt_rq->rt_runtime = rt_runtime;
8678 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8680 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8682 read_unlock(&tasklist_lock);
8683 mutex_unlock(&rt_constraints_mutex);
8688 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8690 u64 rt_runtime, rt_period;
8692 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8693 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8694 if (rt_runtime_us < 0)
8695 rt_runtime = RUNTIME_INF;
8697 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8700 long sched_group_rt_runtime(struct task_group *tg)
8704 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8707 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8708 do_div(rt_runtime_us, NSEC_PER_USEC);
8709 return rt_runtime_us;
8712 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8714 u64 rt_runtime, rt_period;
8716 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8717 rt_runtime = tg->rt_bandwidth.rt_runtime;
8722 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8725 long sched_group_rt_period(struct task_group *tg)
8729 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8730 do_div(rt_period_us, NSEC_PER_USEC);
8731 return rt_period_us;
8734 static int sched_rt_global_constraints(void)
8736 u64 runtime, period;
8739 if (sysctl_sched_rt_period <= 0)
8742 runtime = global_rt_runtime();
8743 period = global_rt_period();
8746 * Sanity check on the sysctl variables.
8748 if (runtime > period && runtime != RUNTIME_INF)
8751 mutex_lock(&rt_constraints_mutex);
8752 read_lock(&tasklist_lock);
8753 ret = __rt_schedulable(NULL, 0, 0);
8754 read_unlock(&tasklist_lock);
8755 mutex_unlock(&rt_constraints_mutex);
8760 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8762 /* Don't accept realtime tasks when there is no way for them to run */
8763 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8769 #else /* !CONFIG_RT_GROUP_SCHED */
8770 static int sched_rt_global_constraints(void)
8772 unsigned long flags;
8775 if (sysctl_sched_rt_period <= 0)
8779 * There's always some RT tasks in the root group
8780 * -- migration, kstopmachine etc..
8782 if (sysctl_sched_rt_runtime == 0)
8785 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8786 for_each_possible_cpu(i) {
8787 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8789 raw_spin_lock(&rt_rq->rt_runtime_lock);
8790 rt_rq->rt_runtime = global_rt_runtime();
8791 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8793 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8797 #endif /* CONFIG_RT_GROUP_SCHED */
8799 int sched_rt_handler(struct ctl_table *table, int write,
8800 void __user *buffer, size_t *lenp,
8804 int old_period, old_runtime;
8805 static DEFINE_MUTEX(mutex);
8808 old_period = sysctl_sched_rt_period;
8809 old_runtime = sysctl_sched_rt_runtime;
8811 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8813 if (!ret && write) {
8814 ret = sched_rt_global_constraints();
8816 sysctl_sched_rt_period = old_period;
8817 sysctl_sched_rt_runtime = old_runtime;
8819 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8820 def_rt_bandwidth.rt_period =
8821 ns_to_ktime(global_rt_period());
8824 mutex_unlock(&mutex);
8829 #ifdef CONFIG_CGROUP_SCHED
8831 /* return corresponding task_group object of a cgroup */
8832 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8834 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8835 struct task_group, css);
8838 static struct cgroup_subsys_state *
8839 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8841 struct task_group *tg, *parent;
8843 if (!cgrp->parent) {
8844 /* This is early initialization for the top cgroup */
8845 return &init_task_group.css;
8848 parent = cgroup_tg(cgrp->parent);
8849 tg = sched_create_group(parent);
8851 return ERR_PTR(-ENOMEM);
8857 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8859 struct task_group *tg = cgroup_tg(cgrp);
8861 sched_destroy_group(tg);
8865 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8867 if ((current != tsk) && (!capable(CAP_SYS_NICE))) {
8868 const struct cred *cred = current_cred(), *tcred;
8870 tcred = __task_cred(tsk);
8872 if (cred->euid != tcred->uid && cred->euid != tcred->suid)
8876 #ifdef CONFIG_RT_GROUP_SCHED
8877 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8880 /* We don't support RT-tasks being in separate groups */
8881 if (tsk->sched_class != &fair_sched_class)
8888 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8889 struct task_struct *tsk, bool threadgroup)
8891 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8895 struct task_struct *c;
8897 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8898 retval = cpu_cgroup_can_attach_task(cgrp, c);
8910 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8911 struct cgroup *old_cont, struct task_struct *tsk,
8914 sched_move_task(tsk);
8916 struct task_struct *c;
8918 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8925 #ifdef CONFIG_FAIR_GROUP_SCHED
8926 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8929 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8932 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8934 struct task_group *tg = cgroup_tg(cgrp);
8936 return (u64) tg->shares;
8938 #endif /* CONFIG_FAIR_GROUP_SCHED */
8940 #ifdef CONFIG_RT_GROUP_SCHED
8941 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8944 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8947 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8949 return sched_group_rt_runtime(cgroup_tg(cgrp));
8952 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8955 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8958 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8960 return sched_group_rt_period(cgroup_tg(cgrp));
8962 #endif /* CONFIG_RT_GROUP_SCHED */
8964 static struct cftype cpu_files[] = {
8965 #ifdef CONFIG_FAIR_GROUP_SCHED
8968 .read_u64 = cpu_shares_read_u64,
8969 .write_u64 = cpu_shares_write_u64,
8972 #ifdef CONFIG_RT_GROUP_SCHED
8974 .name = "rt_runtime_us",
8975 .read_s64 = cpu_rt_runtime_read,
8976 .write_s64 = cpu_rt_runtime_write,
8979 .name = "rt_period_us",
8980 .read_u64 = cpu_rt_period_read_uint,
8981 .write_u64 = cpu_rt_period_write_uint,
8986 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8988 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8991 struct cgroup_subsys cpu_cgroup_subsys = {
8993 .create = cpu_cgroup_create,
8994 .destroy = cpu_cgroup_destroy,
8995 .can_attach = cpu_cgroup_can_attach,
8996 .attach = cpu_cgroup_attach,
8997 .populate = cpu_cgroup_populate,
8998 .subsys_id = cpu_cgroup_subsys_id,
9002 #endif /* CONFIG_CGROUP_SCHED */
9004 #ifdef CONFIG_CGROUP_CPUACCT
9007 * CPU accounting code for task groups.
9009 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9010 * (balbir@in.ibm.com).
9013 /* track cpu usage of a group of tasks and its child groups */
9015 struct cgroup_subsys_state css;
9016 /* cpuusage holds pointer to a u64-type object on every cpu */
9017 u64 __percpu *cpuusage;
9018 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9019 struct cpuacct *parent;
9020 struct cpuacct_charge_calls *cpufreq_fn;
9024 static struct cpuacct *cpuacct_root;
9026 /* Default calls for cpufreq accounting */
9027 static struct cpuacct_charge_calls *cpuacct_cpufreq;
9028 int cpuacct_register_cpufreq(struct cpuacct_charge_calls *fn)
9030 cpuacct_cpufreq = fn;
9033 * Root node is created before platform can register callbacks,
9036 if (cpuacct_root && fn) {
9037 cpuacct_root->cpufreq_fn = fn;
9039 fn->init(&cpuacct_root->cpuacct_data);
9044 struct cgroup_subsys cpuacct_subsys;
9046 /* return cpu accounting group corresponding to this container */
9047 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9049 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9050 struct cpuacct, css);
9053 /* return cpu accounting group to which this task belongs */
9054 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9056 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9057 struct cpuacct, css);
9060 /* create a new cpu accounting group */
9061 static struct cgroup_subsys_state *cpuacct_create(
9062 struct cgroup_subsys *ss, struct cgroup *cgrp)
9064 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9070 ca->cpuusage = alloc_percpu(u64);
9074 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9075 if (percpu_counter_init(&ca->cpustat[i], 0))
9076 goto out_free_counters;
9078 ca->cpufreq_fn = cpuacct_cpufreq;
9080 /* If available, have platform code initalize cpu frequency table */
9081 if (ca->cpufreq_fn && ca->cpufreq_fn->init)
9082 ca->cpufreq_fn->init(&ca->cpuacct_data);
9085 ca->parent = cgroup_ca(cgrp->parent);
9093 percpu_counter_destroy(&ca->cpustat[i]);
9094 free_percpu(ca->cpuusage);
9098 return ERR_PTR(-ENOMEM);
9101 /* destroy an existing cpu accounting group */
9103 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9105 struct cpuacct *ca = cgroup_ca(cgrp);
9108 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9109 percpu_counter_destroy(&ca->cpustat[i]);
9110 free_percpu(ca->cpuusage);
9114 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9116 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9119 #ifndef CONFIG_64BIT
9121 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9123 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9125 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9133 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9135 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9137 #ifndef CONFIG_64BIT
9139 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9141 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9143 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9149 /* return total cpu usage (in nanoseconds) of a group */
9150 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9152 struct cpuacct *ca = cgroup_ca(cgrp);
9153 u64 totalcpuusage = 0;
9156 for_each_present_cpu(i)
9157 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9159 return totalcpuusage;
9162 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9165 struct cpuacct *ca = cgroup_ca(cgrp);
9174 for_each_present_cpu(i)
9175 cpuacct_cpuusage_write(ca, i, 0);
9181 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9184 struct cpuacct *ca = cgroup_ca(cgroup);
9188 for_each_present_cpu(i) {
9189 percpu = cpuacct_cpuusage_read(ca, i);
9190 seq_printf(m, "%llu ", (unsigned long long) percpu);
9192 seq_printf(m, "\n");
9196 static const char *cpuacct_stat_desc[] = {
9197 [CPUACCT_STAT_USER] = "user",
9198 [CPUACCT_STAT_SYSTEM] = "system",
9201 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9202 struct cgroup_map_cb *cb)
9204 struct cpuacct *ca = cgroup_ca(cgrp);
9207 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9208 s64 val = percpu_counter_read(&ca->cpustat[i]);
9209 val = cputime64_to_clock_t(val);
9210 cb->fill(cb, cpuacct_stat_desc[i], val);
9215 static int cpuacct_cpufreq_show(struct cgroup *cgrp, struct cftype *cft,
9216 struct cgroup_map_cb *cb)
9218 struct cpuacct *ca = cgroup_ca(cgrp);
9219 if (ca->cpufreq_fn && ca->cpufreq_fn->cpufreq_show)
9220 ca->cpufreq_fn->cpufreq_show(ca->cpuacct_data, cb);
9225 /* return total cpu power usage (milliWatt second) of a group */
9226 static u64 cpuacct_powerusage_read(struct cgroup *cgrp, struct cftype *cft)
9229 struct cpuacct *ca = cgroup_ca(cgrp);
9232 if (ca->cpufreq_fn && ca->cpufreq_fn->power_usage)
9233 for_each_present_cpu(i) {
9234 totalpower += ca->cpufreq_fn->power_usage(
9241 static struct cftype files[] = {
9244 .read_u64 = cpuusage_read,
9245 .write_u64 = cpuusage_write,
9248 .name = "usage_percpu",
9249 .read_seq_string = cpuacct_percpu_seq_read,
9253 .read_map = cpuacct_stats_show,
9257 .read_map = cpuacct_cpufreq_show,
9261 .read_u64 = cpuacct_powerusage_read
9265 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9267 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9271 * charge this task's execution time to its accounting group.
9273 * called with rq->lock held.
9275 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9280 if (unlikely(!cpuacct_subsys.active))
9283 cpu = task_cpu(tsk);
9289 for (; ca; ca = ca->parent) {
9290 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9291 *cpuusage += cputime;
9293 /* Call back into platform code to account for CPU speeds */
9294 if (ca->cpufreq_fn && ca->cpufreq_fn->charge)
9295 ca->cpufreq_fn->charge(ca->cpuacct_data, cputime, cpu);
9302 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9303 * in cputime_t units. As a result, cpuacct_update_stats calls
9304 * percpu_counter_add with values large enough to always overflow the
9305 * per cpu batch limit causing bad SMP scalability.
9307 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9308 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9309 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9312 #define CPUACCT_BATCH \
9313 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9315 #define CPUACCT_BATCH 0
9319 * Charge the system/user time to the task's accounting group.
9321 static void cpuacct_update_stats(struct task_struct *tsk,
9322 enum cpuacct_stat_index idx, cputime_t val)
9325 int batch = CPUACCT_BATCH;
9327 if (unlikely(!cpuacct_subsys.active))
9334 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9340 struct cgroup_subsys cpuacct_subsys = {
9342 .create = cpuacct_create,
9343 .destroy = cpuacct_destroy,
9344 .populate = cpuacct_populate,
9345 .subsys_id = cpuacct_subsys_id,
9347 #endif /* CONFIG_CGROUP_CPUACCT */
9351 void synchronize_sched_expedited(void)
9355 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9357 #else /* #ifndef CONFIG_SMP */
9359 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9361 static int synchronize_sched_expedited_cpu_stop(void *data)
9364 * There must be a full memory barrier on each affected CPU
9365 * between the time that try_stop_cpus() is called and the
9366 * time that it returns.
9368 * In the current initial implementation of cpu_stop, the
9369 * above condition is already met when the control reaches
9370 * this point and the following smp_mb() is not strictly
9371 * necessary. Do smp_mb() anyway for documentation and
9372 * robustness against future implementation changes.
9374 smp_mb(); /* See above comment block. */
9379 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9380 * approach to force grace period to end quickly. This consumes
9381 * significant time on all CPUs, and is thus not recommended for
9382 * any sort of common-case code.
9384 * Note that it is illegal to call this function while holding any
9385 * lock that is acquired by a CPU-hotplug notifier. Failing to
9386 * observe this restriction will result in deadlock.
9388 void synchronize_sched_expedited(void)
9390 int snap, trycount = 0;
9392 smp_mb(); /* ensure prior mod happens before capturing snap. */
9393 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9395 while (try_stop_cpus(cpu_online_mask,
9396 synchronize_sched_expedited_cpu_stop,
9399 if (trycount++ < 10)
9400 udelay(trycount * num_online_cpus());
9402 synchronize_sched();
9405 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9406 smp_mb(); /* ensure test happens before caller kfree */
9411 atomic_inc(&synchronize_sched_expedited_count);
9412 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9415 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9417 #endif /* #else #ifndef CONFIG_SMP */