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 (test_tsk_need_resched(p))
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))
739 if (strncmp(buf, "NO_", 3) == 0) {
744 for (i = 0; sched_feat_names[i]; i++) {
745 int len = strlen(sched_feat_names[i]);
747 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
749 sysctl_sched_features &= ~(1UL << i);
751 sysctl_sched_features |= (1UL << i);
756 if (!sched_feat_names[i])
764 static int sched_feat_open(struct inode *inode, struct file *filp)
766 return single_open(filp, sched_feat_show, NULL);
769 static const struct file_operations sched_feat_fops = {
770 .open = sched_feat_open,
771 .write = sched_feat_write,
774 .release = single_release,
777 static __init int sched_init_debug(void)
779 debugfs_create_file("sched_features", 0644, NULL, NULL,
784 late_initcall(sched_init_debug);
788 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
791 * Number of tasks to iterate in a single balance run.
792 * Limited because this is done with IRQs disabled.
794 const_debug unsigned int sysctl_sched_nr_migrate = 32;
797 * ratelimit for updating the group shares.
800 unsigned int sysctl_sched_shares_ratelimit = 250000;
801 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
804 * Inject some fuzzyness into changing the per-cpu group shares
805 * this avoids remote rq-locks at the expense of fairness.
808 unsigned int sysctl_sched_shares_thresh = 4;
811 * period over which we average the RT time consumption, measured
816 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
819 * period over which we measure -rt task cpu usage in us.
822 unsigned int sysctl_sched_rt_period = 1000000;
824 static __read_mostly int scheduler_running;
827 * part of the period that we allow rt tasks to run in us.
830 int sysctl_sched_rt_runtime = 950000;
832 static inline u64 global_rt_period(void)
834 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
837 static inline u64 global_rt_runtime(void)
839 if (sysctl_sched_rt_runtime < 0)
842 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
845 #ifndef prepare_arch_switch
846 # define prepare_arch_switch(next) do { } while (0)
848 #ifndef finish_arch_switch
849 # define finish_arch_switch(prev) do { } while (0)
852 static inline int task_current(struct rq *rq, struct task_struct *p)
854 return rq->curr == p;
857 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
858 static inline int task_running(struct rq *rq, struct task_struct *p)
860 return task_current(rq, p);
863 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
867 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
869 #ifdef CONFIG_DEBUG_SPINLOCK
870 /* this is a valid case when another task releases the spinlock */
871 rq->lock.owner = current;
874 * If we are tracking spinlock dependencies then we have to
875 * fix up the runqueue lock - which gets 'carried over' from
878 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
880 raw_spin_unlock_irq(&rq->lock);
883 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
884 static inline int task_running(struct rq *rq, struct task_struct *p)
889 return task_current(rq, p);
893 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
897 * We can optimise this out completely for !SMP, because the
898 * SMP rebalancing from interrupt is the only thing that cares
903 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
904 raw_spin_unlock_irq(&rq->lock);
906 raw_spin_unlock(&rq->lock);
910 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
914 * After ->oncpu is cleared, the task can be moved to a different CPU.
915 * We must ensure this doesn't happen until the switch is completely
921 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
925 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
928 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
931 static inline int task_is_waking(struct task_struct *p)
933 return unlikely(p->state == TASK_WAKING);
937 * __task_rq_lock - lock the runqueue a given task resides on.
938 * Must be called interrupts disabled.
940 static inline struct rq *__task_rq_lock(struct task_struct *p)
947 raw_spin_lock(&rq->lock);
948 if (likely(rq == task_rq(p)))
950 raw_spin_unlock(&rq->lock);
955 * task_rq_lock - lock the runqueue a given task resides on and disable
956 * interrupts. Note the ordering: we can safely lookup the task_rq without
957 * explicitly disabling preemption.
959 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
965 local_irq_save(*flags);
967 raw_spin_lock(&rq->lock);
968 if (likely(rq == task_rq(p)))
970 raw_spin_unlock_irqrestore(&rq->lock, *flags);
974 static void __task_rq_unlock(struct rq *rq)
977 raw_spin_unlock(&rq->lock);
980 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
983 raw_spin_unlock_irqrestore(&rq->lock, *flags);
987 * this_rq_lock - lock this runqueue and disable interrupts.
989 static struct rq *this_rq_lock(void)
996 raw_spin_lock(&rq->lock);
1001 #ifdef CONFIG_SCHED_HRTICK
1003 * Use HR-timers to deliver accurate preemption points.
1005 * Its all a bit involved since we cannot program an hrt while holding the
1006 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1009 * When we get rescheduled we reprogram the hrtick_timer outside of the
1015 * - enabled by features
1016 * - hrtimer is actually high res
1018 static inline int hrtick_enabled(struct rq *rq)
1020 if (!sched_feat(HRTICK))
1022 if (!cpu_active(cpu_of(rq)))
1024 return hrtimer_is_hres_active(&rq->hrtick_timer);
1027 static void hrtick_clear(struct rq *rq)
1029 if (hrtimer_active(&rq->hrtick_timer))
1030 hrtimer_cancel(&rq->hrtick_timer);
1034 * High-resolution timer tick.
1035 * Runs from hardirq context with interrupts disabled.
1037 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1039 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1041 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1043 raw_spin_lock(&rq->lock);
1044 update_rq_clock(rq);
1045 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1046 raw_spin_unlock(&rq->lock);
1048 return HRTIMER_NORESTART;
1053 * called from hardirq (IPI) context
1055 static void __hrtick_start(void *arg)
1057 struct rq *rq = arg;
1059 raw_spin_lock(&rq->lock);
1060 hrtimer_restart(&rq->hrtick_timer);
1061 rq->hrtick_csd_pending = 0;
1062 raw_spin_unlock(&rq->lock);
1066 * Called to set the hrtick timer state.
1068 * called with rq->lock held and irqs disabled
1070 static void hrtick_start(struct rq *rq, u64 delay)
1072 struct hrtimer *timer = &rq->hrtick_timer;
1073 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1075 hrtimer_set_expires(timer, time);
1077 if (rq == this_rq()) {
1078 hrtimer_restart(timer);
1079 } else if (!rq->hrtick_csd_pending) {
1080 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1081 rq->hrtick_csd_pending = 1;
1086 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1088 int cpu = (int)(long)hcpu;
1091 case CPU_UP_CANCELED:
1092 case CPU_UP_CANCELED_FROZEN:
1093 case CPU_DOWN_PREPARE:
1094 case CPU_DOWN_PREPARE_FROZEN:
1096 case CPU_DEAD_FROZEN:
1097 hrtick_clear(cpu_rq(cpu));
1104 static __init void init_hrtick(void)
1106 hotcpu_notifier(hotplug_hrtick, 0);
1110 * Called to set the hrtick timer state.
1112 * called with rq->lock held and irqs disabled
1114 static void hrtick_start(struct rq *rq, u64 delay)
1116 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1117 HRTIMER_MODE_REL_PINNED, 0);
1120 static inline void init_hrtick(void)
1123 #endif /* CONFIG_SMP */
1125 static void init_rq_hrtick(struct rq *rq)
1128 rq->hrtick_csd_pending = 0;
1130 rq->hrtick_csd.flags = 0;
1131 rq->hrtick_csd.func = __hrtick_start;
1132 rq->hrtick_csd.info = rq;
1135 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1136 rq->hrtick_timer.function = hrtick;
1138 #else /* CONFIG_SCHED_HRTICK */
1139 static inline void hrtick_clear(struct rq *rq)
1143 static inline void init_rq_hrtick(struct rq *rq)
1147 static inline void init_hrtick(void)
1150 #endif /* CONFIG_SCHED_HRTICK */
1153 * resched_task - mark a task 'to be rescheduled now'.
1155 * On UP this means the setting of the need_resched flag, on SMP it
1156 * might also involve a cross-CPU call to trigger the scheduler on
1161 #ifndef tsk_is_polling
1162 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1165 static void resched_task(struct task_struct *p)
1169 assert_raw_spin_locked(&task_rq(p)->lock);
1171 if (test_tsk_need_resched(p))
1174 set_tsk_need_resched(p);
1177 if (cpu == smp_processor_id())
1180 /* NEED_RESCHED must be visible before we test polling */
1182 if (!tsk_is_polling(p))
1183 smp_send_reschedule(cpu);
1186 static void resched_cpu(int cpu)
1188 struct rq *rq = cpu_rq(cpu);
1189 unsigned long flags;
1191 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1193 resched_task(cpu_curr(cpu));
1194 raw_spin_unlock_irqrestore(&rq->lock, flags);
1199 * In the semi idle case, use the nearest busy cpu for migrating timers
1200 * from an idle cpu. This is good for power-savings.
1202 * We don't do similar optimization for completely idle system, as
1203 * selecting an idle cpu will add more delays to the timers than intended
1204 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1206 int get_nohz_timer_target(void)
1208 int cpu = smp_processor_id();
1210 struct sched_domain *sd;
1212 for_each_domain(cpu, sd) {
1213 for_each_cpu(i, sched_domain_span(sd))
1220 * When add_timer_on() enqueues a timer into the timer wheel of an
1221 * idle CPU then this timer might expire before the next timer event
1222 * which is scheduled to wake up that CPU. In case of a completely
1223 * idle system the next event might even be infinite time into the
1224 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1225 * leaves the inner idle loop so the newly added timer is taken into
1226 * account when the CPU goes back to idle and evaluates the timer
1227 * wheel for the next timer event.
1229 void wake_up_idle_cpu(int cpu)
1231 struct rq *rq = cpu_rq(cpu);
1233 if (cpu == smp_processor_id())
1237 * This is safe, as this function is called with the timer
1238 * wheel base lock of (cpu) held. When the CPU is on the way
1239 * to idle and has not yet set rq->curr to idle then it will
1240 * be serialized on the timer wheel base lock and take the new
1241 * timer into account automatically.
1243 if (rq->curr != rq->idle)
1247 * We can set TIF_RESCHED on the idle task of the other CPU
1248 * lockless. The worst case is that the other CPU runs the
1249 * idle task through an additional NOOP schedule()
1251 set_tsk_need_resched(rq->idle);
1253 /* NEED_RESCHED must be visible before we test polling */
1255 if (!tsk_is_polling(rq->idle))
1256 smp_send_reschedule(cpu);
1259 #endif /* CONFIG_NO_HZ */
1261 static u64 sched_avg_period(void)
1263 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1266 static void sched_avg_update(struct rq *rq)
1268 s64 period = sched_avg_period();
1270 while ((s64)(rq->clock - rq->age_stamp) > period) {
1272 * Inline assembly required to prevent the compiler
1273 * optimising this loop into a divmod call.
1274 * See __iter_div_u64_rem() for another example of this.
1276 asm("" : "+rm" (rq->age_stamp));
1277 rq->age_stamp += period;
1282 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1284 rq->rt_avg += rt_delta;
1285 sched_avg_update(rq);
1288 #else /* !CONFIG_SMP */
1289 static void resched_task(struct task_struct *p)
1291 assert_raw_spin_locked(&task_rq(p)->lock);
1292 set_tsk_need_resched(p);
1295 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1299 static void sched_avg_update(struct rq *rq)
1302 #endif /* CONFIG_SMP */
1304 #if BITS_PER_LONG == 32
1305 # define WMULT_CONST (~0UL)
1307 # define WMULT_CONST (1UL << 32)
1310 #define WMULT_SHIFT 32
1313 * Shift right and round:
1315 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1318 * delta *= weight / lw
1320 static unsigned long
1321 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1322 struct load_weight *lw)
1326 if (!lw->inv_weight) {
1327 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1330 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1334 tmp = (u64)delta_exec * weight;
1336 * Check whether we'd overflow the 64-bit multiplication:
1338 if (unlikely(tmp > WMULT_CONST))
1339 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1342 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1344 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1347 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1353 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1360 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1361 * of tasks with abnormal "nice" values across CPUs the contribution that
1362 * each task makes to its run queue's load is weighted according to its
1363 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1364 * scaled version of the new time slice allocation that they receive on time
1368 #define WEIGHT_IDLEPRIO 3
1369 #define WMULT_IDLEPRIO 1431655765
1372 * Nice levels are multiplicative, with a gentle 10% change for every
1373 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1374 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1375 * that remained on nice 0.
1377 * The "10% effect" is relative and cumulative: from _any_ nice level,
1378 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1379 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1380 * If a task goes up by ~10% and another task goes down by ~10% then
1381 * the relative distance between them is ~25%.)
1383 static const int prio_to_weight[40] = {
1384 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1385 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1386 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1387 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1388 /* 0 */ 1024, 820, 655, 526, 423,
1389 /* 5 */ 335, 272, 215, 172, 137,
1390 /* 10 */ 110, 87, 70, 56, 45,
1391 /* 15 */ 36, 29, 23, 18, 15,
1395 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1397 * In cases where the weight does not change often, we can use the
1398 * precalculated inverse to speed up arithmetics by turning divisions
1399 * into multiplications:
1401 static const u32 prio_to_wmult[40] = {
1402 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1403 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1404 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1405 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1406 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1407 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1408 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1409 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1412 /* Time spent by the tasks of the cpu accounting group executing in ... */
1413 enum cpuacct_stat_index {
1414 CPUACCT_STAT_USER, /* ... user mode */
1415 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1417 CPUACCT_STAT_NSTATS,
1420 #ifdef CONFIG_CGROUP_CPUACCT
1421 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1422 static void cpuacct_update_stats(struct task_struct *tsk,
1423 enum cpuacct_stat_index idx, cputime_t val);
1425 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1426 static inline void cpuacct_update_stats(struct task_struct *tsk,
1427 enum cpuacct_stat_index idx, cputime_t val) {}
1430 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1432 update_load_add(&rq->load, load);
1435 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1437 update_load_sub(&rq->load, load);
1440 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1441 typedef int (*tg_visitor)(struct task_group *, void *);
1444 * Iterate the full tree, calling @down when first entering a node and @up when
1445 * leaving it for the final time.
1447 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1449 struct task_group *parent, *child;
1453 parent = &root_task_group;
1455 ret = (*down)(parent, data);
1458 list_for_each_entry_rcu(child, &parent->children, siblings) {
1465 ret = (*up)(parent, data);
1470 parent = parent->parent;
1479 static int tg_nop(struct task_group *tg, void *data)
1486 /* Used instead of source_load when we know the type == 0 */
1487 static unsigned long weighted_cpuload(const int cpu)
1489 return cpu_rq(cpu)->load.weight;
1493 * Return a low guess at the load of a migration-source cpu weighted
1494 * according to the scheduling class and "nice" value.
1496 * We want to under-estimate the load of migration sources, to
1497 * balance conservatively.
1499 static unsigned long source_load(int cpu, int type)
1501 struct rq *rq = cpu_rq(cpu);
1502 unsigned long total = weighted_cpuload(cpu);
1504 if (type == 0 || !sched_feat(LB_BIAS))
1507 return min(rq->cpu_load[type-1], total);
1511 * Return a high guess at the load of a migration-target cpu weighted
1512 * according to the scheduling class and "nice" value.
1514 static unsigned long target_load(int cpu, int type)
1516 struct rq *rq = cpu_rq(cpu);
1517 unsigned long total = weighted_cpuload(cpu);
1519 if (type == 0 || !sched_feat(LB_BIAS))
1522 return max(rq->cpu_load[type-1], total);
1525 static unsigned long power_of(int cpu)
1527 return cpu_rq(cpu)->cpu_power;
1530 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1532 static unsigned long cpu_avg_load_per_task(int cpu)
1534 struct rq *rq = cpu_rq(cpu);
1535 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1538 rq->avg_load_per_task = rq->load.weight / nr_running;
1540 rq->avg_load_per_task = 0;
1542 return rq->avg_load_per_task;
1545 #ifdef CONFIG_FAIR_GROUP_SCHED
1547 static __read_mostly unsigned long __percpu *update_shares_data;
1549 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1552 * Calculate and set the cpu's group shares.
1554 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1555 unsigned long sd_shares,
1556 unsigned long sd_rq_weight,
1557 unsigned long *usd_rq_weight)
1559 unsigned long shares, rq_weight;
1562 rq_weight = usd_rq_weight[cpu];
1565 rq_weight = NICE_0_LOAD;
1569 * \Sum_j shares_j * rq_weight_i
1570 * shares_i = -----------------------------
1571 * \Sum_j rq_weight_j
1573 shares = (sd_shares * rq_weight) / sd_rq_weight;
1574 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1576 if (abs(shares - tg->se[cpu]->load.weight) >
1577 sysctl_sched_shares_thresh) {
1578 struct rq *rq = cpu_rq(cpu);
1579 unsigned long flags;
1581 raw_spin_lock_irqsave(&rq->lock, flags);
1582 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1583 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1584 __set_se_shares(tg->se[cpu], shares);
1585 raw_spin_unlock_irqrestore(&rq->lock, flags);
1590 * Re-compute the task group their per cpu shares over the given domain.
1591 * This needs to be done in a bottom-up fashion because the rq weight of a
1592 * parent group depends on the shares of its child groups.
1594 static int tg_shares_up(struct task_group *tg, void *data)
1596 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1597 unsigned long *usd_rq_weight;
1598 struct sched_domain *sd = data;
1599 unsigned long flags;
1605 local_irq_save(flags);
1606 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1608 for_each_cpu(i, sched_domain_span(sd)) {
1609 weight = tg->cfs_rq[i]->load.weight;
1610 usd_rq_weight[i] = weight;
1612 rq_weight += weight;
1614 * If there are currently no tasks on the cpu pretend there
1615 * is one of average load so that when a new task gets to
1616 * run here it will not get delayed by group starvation.
1619 weight = NICE_0_LOAD;
1621 sum_weight += weight;
1622 shares += tg->cfs_rq[i]->shares;
1626 rq_weight = sum_weight;
1628 if ((!shares && rq_weight) || shares > tg->shares)
1629 shares = tg->shares;
1631 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1632 shares = tg->shares;
1634 for_each_cpu(i, sched_domain_span(sd))
1635 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1637 local_irq_restore(flags);
1643 * Compute the cpu's hierarchical load factor for each task group.
1644 * This needs to be done in a top-down fashion because the load of a child
1645 * group is a fraction of its parents load.
1647 static int tg_load_down(struct task_group *tg, void *data)
1650 long cpu = (long)data;
1653 load = cpu_rq(cpu)->load.weight;
1655 load = tg->parent->cfs_rq[cpu]->h_load;
1656 load *= tg->cfs_rq[cpu]->shares;
1657 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1660 tg->cfs_rq[cpu]->h_load = load;
1665 static void update_shares(struct sched_domain *sd)
1670 if (root_task_group_empty())
1673 now = local_clock();
1674 elapsed = now - sd->last_update;
1676 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1677 sd->last_update = now;
1678 walk_tg_tree(tg_nop, tg_shares_up, sd);
1682 static void update_h_load(long cpu)
1684 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1689 static inline void update_shares(struct sched_domain *sd)
1695 #ifdef CONFIG_PREEMPT
1697 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1700 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1701 * way at the expense of forcing extra atomic operations in all
1702 * invocations. This assures that the double_lock is acquired using the
1703 * same underlying policy as the spinlock_t on this architecture, which
1704 * reduces latency compared to the unfair variant below. However, it
1705 * also adds more overhead and therefore may reduce throughput.
1707 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1708 __releases(this_rq->lock)
1709 __acquires(busiest->lock)
1710 __acquires(this_rq->lock)
1712 raw_spin_unlock(&this_rq->lock);
1713 double_rq_lock(this_rq, busiest);
1720 * Unfair double_lock_balance: Optimizes throughput at the expense of
1721 * latency by eliminating extra atomic operations when the locks are
1722 * already in proper order on entry. This favors lower cpu-ids and will
1723 * grant the double lock to lower cpus over higher ids under contention,
1724 * regardless of entry order into the function.
1726 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1727 __releases(this_rq->lock)
1728 __acquires(busiest->lock)
1729 __acquires(this_rq->lock)
1733 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1734 if (busiest < this_rq) {
1735 raw_spin_unlock(&this_rq->lock);
1736 raw_spin_lock(&busiest->lock);
1737 raw_spin_lock_nested(&this_rq->lock,
1738 SINGLE_DEPTH_NESTING);
1741 raw_spin_lock_nested(&busiest->lock,
1742 SINGLE_DEPTH_NESTING);
1747 #endif /* CONFIG_PREEMPT */
1750 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1752 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1754 if (unlikely(!irqs_disabled())) {
1755 /* printk() doesn't work good under rq->lock */
1756 raw_spin_unlock(&this_rq->lock);
1760 return _double_lock_balance(this_rq, busiest);
1763 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1764 __releases(busiest->lock)
1766 raw_spin_unlock(&busiest->lock);
1767 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1771 * double_rq_lock - safely lock two runqueues
1773 * Note this does not disable interrupts like task_rq_lock,
1774 * you need to do so manually before calling.
1776 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1777 __acquires(rq1->lock)
1778 __acquires(rq2->lock)
1780 BUG_ON(!irqs_disabled());
1782 raw_spin_lock(&rq1->lock);
1783 __acquire(rq2->lock); /* Fake it out ;) */
1786 raw_spin_lock(&rq1->lock);
1787 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1789 raw_spin_lock(&rq2->lock);
1790 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1796 * double_rq_unlock - safely unlock two runqueues
1798 * Note this does not restore interrupts like task_rq_unlock,
1799 * you need to do so manually after calling.
1801 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1802 __releases(rq1->lock)
1803 __releases(rq2->lock)
1805 raw_spin_unlock(&rq1->lock);
1807 raw_spin_unlock(&rq2->lock);
1809 __release(rq2->lock);
1814 #ifdef CONFIG_FAIR_GROUP_SCHED
1815 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1818 cfs_rq->shares = shares;
1823 static void calc_load_account_idle(struct rq *this_rq);
1824 static void update_sysctl(void);
1825 static int get_update_sysctl_factor(void);
1826 static void update_cpu_load(struct rq *this_rq);
1828 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1830 set_task_rq(p, cpu);
1833 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1834 * successfuly executed on another CPU. We must ensure that updates of
1835 * per-task data have been completed by this moment.
1838 task_thread_info(p)->cpu = cpu;
1842 static const struct sched_class rt_sched_class;
1844 #define sched_class_highest (&rt_sched_class)
1845 #define for_each_class(class) \
1846 for (class = sched_class_highest; class; class = class->next)
1848 #include "sched_stats.h"
1850 static void inc_nr_running(struct rq *rq)
1855 static void dec_nr_running(struct rq *rq)
1860 static void set_load_weight(struct task_struct *p)
1862 if (task_has_rt_policy(p)) {
1863 p->se.load.weight = 0;
1864 p->se.load.inv_weight = WMULT_CONST;
1869 * SCHED_IDLE tasks get minimal weight:
1871 if (p->policy == SCHED_IDLE) {
1872 p->se.load.weight = WEIGHT_IDLEPRIO;
1873 p->se.load.inv_weight = WMULT_IDLEPRIO;
1877 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1878 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1881 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1883 update_rq_clock(rq);
1884 sched_info_queued(p);
1885 p->sched_class->enqueue_task(rq, p, flags);
1889 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1891 update_rq_clock(rq);
1892 sched_info_dequeued(p);
1893 p->sched_class->dequeue_task(rq, p, flags);
1898 * activate_task - move a task to the runqueue.
1900 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1902 if (task_contributes_to_load(p))
1903 rq->nr_uninterruptible--;
1905 enqueue_task(rq, p, flags);
1910 * deactivate_task - remove a task from the runqueue.
1912 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1914 if (task_contributes_to_load(p))
1915 rq->nr_uninterruptible++;
1917 dequeue_task(rq, p, flags);
1921 #include "sched_idletask.c"
1922 #include "sched_fair.c"
1923 #include "sched_rt.c"
1924 #ifdef CONFIG_SCHED_DEBUG
1925 # include "sched_debug.c"
1929 * __normal_prio - return the priority that is based on the static prio
1931 static inline int __normal_prio(struct task_struct *p)
1933 return p->static_prio;
1937 * Calculate the expected normal priority: i.e. priority
1938 * without taking RT-inheritance into account. Might be
1939 * boosted by interactivity modifiers. Changes upon fork,
1940 * setprio syscalls, and whenever the interactivity
1941 * estimator recalculates.
1943 static inline int normal_prio(struct task_struct *p)
1947 if (task_has_rt_policy(p))
1948 prio = MAX_RT_PRIO-1 - p->rt_priority;
1950 prio = __normal_prio(p);
1955 * Calculate the current priority, i.e. the priority
1956 * taken into account by the scheduler. This value might
1957 * be boosted by RT tasks, or might be boosted by
1958 * interactivity modifiers. Will be RT if the task got
1959 * RT-boosted. If not then it returns p->normal_prio.
1961 static int effective_prio(struct task_struct *p)
1963 p->normal_prio = normal_prio(p);
1965 * If we are RT tasks or we were boosted to RT priority,
1966 * keep the priority unchanged. Otherwise, update priority
1967 * to the normal priority:
1969 if (!rt_prio(p->prio))
1970 return p->normal_prio;
1975 * task_curr - is this task currently executing on a CPU?
1976 * @p: the task in question.
1978 inline int task_curr(const struct task_struct *p)
1980 return cpu_curr(task_cpu(p)) == p;
1983 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1984 const struct sched_class *prev_class,
1985 int oldprio, int running)
1987 if (prev_class != p->sched_class) {
1988 if (prev_class->switched_from)
1989 prev_class->switched_from(rq, p, running);
1990 p->sched_class->switched_to(rq, p, running);
1992 p->sched_class->prio_changed(rq, p, oldprio, running);
1997 * Is this task likely cache-hot:
2000 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2004 if (p->sched_class != &fair_sched_class)
2008 * Buddy candidates are cache hot:
2010 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2011 (&p->se == cfs_rq_of(&p->se)->next ||
2012 &p->se == cfs_rq_of(&p->se)->last))
2015 if (sysctl_sched_migration_cost == -1)
2017 if (sysctl_sched_migration_cost == 0)
2020 delta = now - p->se.exec_start;
2022 return delta < (s64)sysctl_sched_migration_cost;
2025 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2027 #ifdef CONFIG_SCHED_DEBUG
2029 * We should never call set_task_cpu() on a blocked task,
2030 * ttwu() will sort out the placement.
2032 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2033 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2036 trace_sched_migrate_task(p, new_cpu);
2038 if (task_cpu(p) != new_cpu) {
2039 p->se.nr_migrations++;
2040 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2043 __set_task_cpu(p, new_cpu);
2046 struct migration_arg {
2047 struct task_struct *task;
2051 static int migration_cpu_stop(void *data);
2054 * The task's runqueue lock must be held.
2055 * Returns true if you have to wait for migration thread.
2057 static bool migrate_task(struct task_struct *p, int dest_cpu)
2059 struct rq *rq = task_rq(p);
2062 * If the task is not on a runqueue (and not running), then
2063 * the next wake-up will properly place the task.
2065 return p->se.on_rq || task_running(rq, p);
2069 * wait_task_inactive - wait for a thread to unschedule.
2071 * If @match_state is nonzero, it's the @p->state value just checked and
2072 * not expected to change. If it changes, i.e. @p might have woken up,
2073 * then return zero. When we succeed in waiting for @p to be off its CPU,
2074 * we return a positive number (its total switch count). If a second call
2075 * a short while later returns the same number, the caller can be sure that
2076 * @p has remained unscheduled the whole time.
2078 * The caller must ensure that the task *will* unschedule sometime soon,
2079 * else this function might spin for a *long* time. This function can't
2080 * be called with interrupts off, or it may introduce deadlock with
2081 * smp_call_function() if an IPI is sent by the same process we are
2082 * waiting to become inactive.
2084 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2086 unsigned long flags;
2093 * We do the initial early heuristics without holding
2094 * any task-queue locks at all. We'll only try to get
2095 * the runqueue lock when things look like they will
2101 * If the task is actively running on another CPU
2102 * still, just relax and busy-wait without holding
2105 * NOTE! Since we don't hold any locks, it's not
2106 * even sure that "rq" stays as the right runqueue!
2107 * But we don't care, since "task_running()" will
2108 * return false if the runqueue has changed and p
2109 * is actually now running somewhere else!
2111 while (task_running(rq, p)) {
2112 if (match_state && unlikely(p->state != match_state))
2118 * Ok, time to look more closely! We need the rq
2119 * lock now, to be *sure*. If we're wrong, we'll
2120 * just go back and repeat.
2122 rq = task_rq_lock(p, &flags);
2123 trace_sched_wait_task(p);
2124 running = task_running(rq, p);
2125 on_rq = p->se.on_rq;
2127 if (!match_state || p->state == match_state)
2128 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2129 task_rq_unlock(rq, &flags);
2132 * If it changed from the expected state, bail out now.
2134 if (unlikely(!ncsw))
2138 * Was it really running after all now that we
2139 * checked with the proper locks actually held?
2141 * Oops. Go back and try again..
2143 if (unlikely(running)) {
2149 * It's not enough that it's not actively running,
2150 * it must be off the runqueue _entirely_, and not
2153 * So if it was still runnable (but just not actively
2154 * running right now), it's preempted, and we should
2155 * yield - it could be a while.
2157 if (unlikely(on_rq)) {
2158 schedule_timeout_uninterruptible(1);
2163 * Ahh, all good. It wasn't running, and it wasn't
2164 * runnable, which means that it will never become
2165 * running in the future either. We're all done!
2174 * kick_process - kick a running thread to enter/exit the kernel
2175 * @p: the to-be-kicked thread
2177 * Cause a process which is running on another CPU to enter
2178 * kernel-mode, without any delay. (to get signals handled.)
2180 * NOTE: this function doesnt have to take the runqueue lock,
2181 * because all it wants to ensure is that the remote task enters
2182 * the kernel. If the IPI races and the task has been migrated
2183 * to another CPU then no harm is done and the purpose has been
2186 void kick_process(struct task_struct *p)
2192 if ((cpu != smp_processor_id()) && task_curr(p))
2193 smp_send_reschedule(cpu);
2196 EXPORT_SYMBOL_GPL(kick_process);
2197 #endif /* CONFIG_SMP */
2200 * task_oncpu_function_call - call a function on the cpu on which a task runs
2201 * @p: the task to evaluate
2202 * @func: the function to be called
2203 * @info: the function call argument
2205 * Calls the function @func when the task is currently running. This might
2206 * be on the current CPU, which just calls the function directly
2208 void task_oncpu_function_call(struct task_struct *p,
2209 void (*func) (void *info), void *info)
2216 smp_call_function_single(cpu, func, info, 1);
2222 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2224 static int select_fallback_rq(int cpu, struct task_struct *p)
2227 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2229 /* Look for allowed, online CPU in same node. */
2230 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2231 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2234 /* Any allowed, online CPU? */
2235 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2236 if (dest_cpu < nr_cpu_ids)
2239 /* No more Mr. Nice Guy. */
2240 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2241 dest_cpu = cpuset_cpus_allowed_fallback(p);
2243 * Don't tell them about moving exiting tasks or
2244 * kernel threads (both mm NULL), since they never
2247 if (p->mm && printk_ratelimit()) {
2248 printk(KERN_INFO "process %d (%s) no "
2249 "longer affine to cpu%d\n",
2250 task_pid_nr(p), p->comm, cpu);
2258 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2261 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2263 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2266 * In order not to call set_task_cpu() on a blocking task we need
2267 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2270 * Since this is common to all placement strategies, this lives here.
2272 * [ this allows ->select_task() to simply return task_cpu(p) and
2273 * not worry about this generic constraint ]
2275 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2277 cpu = select_fallback_rq(task_cpu(p), p);
2282 static void update_avg(u64 *avg, u64 sample)
2284 s64 diff = sample - *avg;
2289 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2290 bool is_sync, bool is_migrate, bool is_local,
2291 unsigned long en_flags)
2293 schedstat_inc(p, se.statistics.nr_wakeups);
2295 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2297 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2299 schedstat_inc(p, se.statistics.nr_wakeups_local);
2301 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2303 activate_task(rq, p, en_flags);
2306 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2307 int wake_flags, bool success)
2309 trace_sched_wakeup(p, success);
2310 check_preempt_curr(rq, p, wake_flags);
2312 p->state = TASK_RUNNING;
2314 if (p->sched_class->task_woken)
2315 p->sched_class->task_woken(rq, p);
2317 if (unlikely(rq->idle_stamp)) {
2318 u64 delta = rq->clock - rq->idle_stamp;
2319 u64 max = 2*sysctl_sched_migration_cost;
2324 update_avg(&rq->avg_idle, delta);
2328 /* if a worker is waking up, notify workqueue */
2329 if ((p->flags & PF_WQ_WORKER) && success)
2330 wq_worker_waking_up(p, cpu_of(rq));
2334 * try_to_wake_up - wake up a thread
2335 * @p: the thread to be awakened
2336 * @state: the mask of task states that can be woken
2337 * @wake_flags: wake modifier flags (WF_*)
2339 * Put it on the run-queue if it's not already there. The "current"
2340 * thread is always on the run-queue (except when the actual
2341 * re-schedule is in progress), and as such you're allowed to do
2342 * the simpler "current->state = TASK_RUNNING" to mark yourself
2343 * runnable without the overhead of this.
2345 * Returns %true if @p was woken up, %false if it was already running
2346 * or @state didn't match @p's state.
2348 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2351 int cpu, orig_cpu, this_cpu, success = 0;
2352 unsigned long flags;
2353 unsigned long en_flags = ENQUEUE_WAKEUP;
2356 this_cpu = get_cpu();
2359 rq = task_rq_lock(p, &flags);
2360 if (!(p->state & state))
2370 if (unlikely(task_running(rq, p)))
2374 * In order to handle concurrent wakeups and release the rq->lock
2375 * we put the task in TASK_WAKING state.
2377 * First fix up the nr_uninterruptible count:
2379 if (task_contributes_to_load(p)) {
2380 if (likely(cpu_online(orig_cpu)))
2381 rq->nr_uninterruptible--;
2383 this_rq()->nr_uninterruptible--;
2385 p->state = TASK_WAKING;
2387 if (p->sched_class->task_waking) {
2388 p->sched_class->task_waking(rq, p);
2389 en_flags |= ENQUEUE_WAKING;
2392 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2393 if (cpu != orig_cpu)
2394 set_task_cpu(p, cpu);
2395 __task_rq_unlock(rq);
2398 raw_spin_lock(&rq->lock);
2401 * We migrated the task without holding either rq->lock, however
2402 * since the task is not on the task list itself, nobody else
2403 * will try and migrate the task, hence the rq should match the
2404 * cpu we just moved it to.
2406 WARN_ON(task_cpu(p) != cpu);
2407 WARN_ON(p->state != TASK_WAKING);
2409 #ifdef CONFIG_SCHEDSTATS
2410 schedstat_inc(rq, ttwu_count);
2411 if (cpu == this_cpu)
2412 schedstat_inc(rq, ttwu_local);
2414 struct sched_domain *sd;
2415 for_each_domain(this_cpu, sd) {
2416 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2417 schedstat_inc(sd, ttwu_wake_remote);
2422 #endif /* CONFIG_SCHEDSTATS */
2425 #endif /* CONFIG_SMP */
2426 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2427 cpu == this_cpu, en_flags);
2430 ttwu_post_activation(p, rq, wake_flags, success);
2432 task_rq_unlock(rq, &flags);
2439 * try_to_wake_up_local - try to wake up a local task with rq lock held
2440 * @p: the thread to be awakened
2442 * Put @p on the run-queue if it's not alredy there. The caller must
2443 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2444 * the current task. this_rq() stays locked over invocation.
2446 static void try_to_wake_up_local(struct task_struct *p)
2448 struct rq *rq = task_rq(p);
2449 bool success = false;
2451 BUG_ON(rq != this_rq());
2452 BUG_ON(p == current);
2453 lockdep_assert_held(&rq->lock);
2455 if (!(p->state & TASK_NORMAL))
2459 if (likely(!task_running(rq, p))) {
2460 schedstat_inc(rq, ttwu_count);
2461 schedstat_inc(rq, ttwu_local);
2463 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2466 ttwu_post_activation(p, rq, 0, success);
2470 * wake_up_process - Wake up a specific process
2471 * @p: The process to be woken up.
2473 * Attempt to wake up the nominated process and move it to the set of runnable
2474 * processes. Returns 1 if the process was woken up, 0 if it was already
2477 * It may be assumed that this function implies a write memory barrier before
2478 * changing the task state if and only if any tasks are woken up.
2480 int wake_up_process(struct task_struct *p)
2482 return try_to_wake_up(p, TASK_ALL, 0);
2484 EXPORT_SYMBOL(wake_up_process);
2486 int wake_up_state(struct task_struct *p, unsigned int state)
2488 return try_to_wake_up(p, state, 0);
2492 * Perform scheduler related setup for a newly forked process p.
2493 * p is forked by current.
2495 * __sched_fork() is basic setup used by init_idle() too:
2497 static void __sched_fork(struct task_struct *p)
2499 p->se.exec_start = 0;
2500 p->se.sum_exec_runtime = 0;
2501 p->se.prev_sum_exec_runtime = 0;
2502 p->se.nr_migrations = 0;
2504 #ifdef CONFIG_SCHEDSTATS
2505 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2508 INIT_LIST_HEAD(&p->rt.run_list);
2510 INIT_LIST_HEAD(&p->se.group_node);
2512 #ifdef CONFIG_PREEMPT_NOTIFIERS
2513 INIT_HLIST_HEAD(&p->preempt_notifiers);
2518 * fork()/clone()-time setup:
2520 void sched_fork(struct task_struct *p, int clone_flags)
2522 int cpu = get_cpu();
2526 * We mark the process as running here. This guarantees that
2527 * nobody will actually run it, and a signal or other external
2528 * event cannot wake it up and insert it on the runqueue either.
2530 p->state = TASK_RUNNING;
2533 * Revert to default priority/policy on fork if requested.
2535 if (unlikely(p->sched_reset_on_fork)) {
2536 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2537 p->policy = SCHED_NORMAL;
2538 p->normal_prio = p->static_prio;
2541 if (PRIO_TO_NICE(p->static_prio) < 0) {
2542 p->static_prio = NICE_TO_PRIO(0);
2543 p->normal_prio = p->static_prio;
2548 * We don't need the reset flag anymore after the fork. It has
2549 * fulfilled its duty:
2551 p->sched_reset_on_fork = 0;
2555 * Make sure we do not leak PI boosting priority to the child.
2557 p->prio = current->normal_prio;
2559 if (!rt_prio(p->prio))
2560 p->sched_class = &fair_sched_class;
2562 if (p->sched_class->task_fork)
2563 p->sched_class->task_fork(p);
2566 * The child is not yet in the pid-hash so no cgroup attach races,
2567 * and the cgroup is pinned to this child due to cgroup_fork()
2568 * is ran before sched_fork().
2570 * Silence PROVE_RCU.
2573 set_task_cpu(p, cpu);
2576 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2577 if (likely(sched_info_on()))
2578 memset(&p->sched_info, 0, sizeof(p->sched_info));
2580 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2583 #ifdef CONFIG_PREEMPT
2584 /* Want to start with kernel preemption disabled. */
2585 task_thread_info(p)->preempt_count = 1;
2587 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2593 * wake_up_new_task - wake up a newly created task for the first time.
2595 * This function will do some initial scheduler statistics housekeeping
2596 * that must be done for every newly created context, then puts the task
2597 * on the runqueue and wakes it.
2599 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2601 unsigned long flags;
2603 int cpu __maybe_unused = get_cpu();
2606 rq = task_rq_lock(p, &flags);
2607 p->state = TASK_WAKING;
2610 * Fork balancing, do it here and not earlier because:
2611 * - cpus_allowed can change in the fork path
2612 * - any previously selected cpu might disappear through hotplug
2614 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2615 * without people poking at ->cpus_allowed.
2617 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2618 set_task_cpu(p, cpu);
2620 p->state = TASK_RUNNING;
2621 task_rq_unlock(rq, &flags);
2624 rq = task_rq_lock(p, &flags);
2625 activate_task(rq, p, 0);
2626 trace_sched_wakeup_new(p, 1);
2627 check_preempt_curr(rq, p, WF_FORK);
2629 if (p->sched_class->task_woken)
2630 p->sched_class->task_woken(rq, p);
2632 task_rq_unlock(rq, &flags);
2636 #ifdef CONFIG_PREEMPT_NOTIFIERS
2639 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2640 * @notifier: notifier struct to register
2642 void preempt_notifier_register(struct preempt_notifier *notifier)
2644 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2646 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2649 * preempt_notifier_unregister - no longer interested in preemption notifications
2650 * @notifier: notifier struct to unregister
2652 * This is safe to call from within a preemption notifier.
2654 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2656 hlist_del(¬ifier->link);
2658 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2660 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2662 struct preempt_notifier *notifier;
2663 struct hlist_node *node;
2665 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2666 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2670 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2671 struct task_struct *next)
2673 struct preempt_notifier *notifier;
2674 struct hlist_node *node;
2676 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2677 notifier->ops->sched_out(notifier, next);
2680 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2682 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2687 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2688 struct task_struct *next)
2692 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2695 * prepare_task_switch - prepare to switch tasks
2696 * @rq: the runqueue preparing to switch
2697 * @prev: the current task that is being switched out
2698 * @next: the task we are going to switch to.
2700 * This is called with the rq lock held and interrupts off. It must
2701 * be paired with a subsequent finish_task_switch after the context
2704 * prepare_task_switch sets up locking and calls architecture specific
2708 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2709 struct task_struct *next)
2711 fire_sched_out_preempt_notifiers(prev, next);
2712 prepare_lock_switch(rq, next);
2713 prepare_arch_switch(next);
2717 * finish_task_switch - clean up after a task-switch
2718 * @rq: runqueue associated with task-switch
2719 * @prev: the thread we just switched away from.
2721 * finish_task_switch must be called after the context switch, paired
2722 * with a prepare_task_switch call before the context switch.
2723 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2724 * and do any other architecture-specific cleanup actions.
2726 * Note that we may have delayed dropping an mm in context_switch(). If
2727 * so, we finish that here outside of the runqueue lock. (Doing it
2728 * with the lock held can cause deadlocks; see schedule() for
2731 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2732 __releases(rq->lock)
2734 struct mm_struct *mm = rq->prev_mm;
2740 * A task struct has one reference for the use as "current".
2741 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2742 * schedule one last time. The schedule call will never return, and
2743 * the scheduled task must drop that reference.
2744 * The test for TASK_DEAD must occur while the runqueue locks are
2745 * still held, otherwise prev could be scheduled on another cpu, die
2746 * there before we look at prev->state, and then the reference would
2748 * Manfred Spraul <manfred@colorfullife.com>
2750 prev_state = prev->state;
2751 finish_arch_switch(prev);
2752 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2753 local_irq_disable();
2754 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2755 perf_event_task_sched_in(current);
2756 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2758 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2759 finish_lock_switch(rq, prev);
2761 fire_sched_in_preempt_notifiers(current);
2764 if (unlikely(prev_state == TASK_DEAD)) {
2766 * Remove function-return probe instances associated with this
2767 * task and put them back on the free list.
2769 kprobe_flush_task(prev);
2770 put_task_struct(prev);
2776 /* assumes rq->lock is held */
2777 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2779 if (prev->sched_class->pre_schedule)
2780 prev->sched_class->pre_schedule(rq, prev);
2783 /* rq->lock is NOT held, but preemption is disabled */
2784 static inline void post_schedule(struct rq *rq)
2786 if (rq->post_schedule) {
2787 unsigned long flags;
2789 raw_spin_lock_irqsave(&rq->lock, flags);
2790 if (rq->curr->sched_class->post_schedule)
2791 rq->curr->sched_class->post_schedule(rq);
2792 raw_spin_unlock_irqrestore(&rq->lock, flags);
2794 rq->post_schedule = 0;
2800 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2804 static inline void post_schedule(struct rq *rq)
2811 * schedule_tail - first thing a freshly forked thread must call.
2812 * @prev: the thread we just switched away from.
2814 asmlinkage void schedule_tail(struct task_struct *prev)
2815 __releases(rq->lock)
2817 struct rq *rq = this_rq();
2819 finish_task_switch(rq, prev);
2822 * FIXME: do we need to worry about rq being invalidated by the
2827 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2828 /* In this case, finish_task_switch does not reenable preemption */
2831 if (current->set_child_tid)
2832 put_user(task_pid_vnr(current), current->set_child_tid);
2836 * context_switch - switch to the new MM and the new
2837 * thread's register state.
2840 context_switch(struct rq *rq, struct task_struct *prev,
2841 struct task_struct *next)
2843 struct mm_struct *mm, *oldmm;
2845 prepare_task_switch(rq, prev, next);
2846 trace_sched_switch(prev, next);
2848 oldmm = prev->active_mm;
2850 * For paravirt, this is coupled with an exit in switch_to to
2851 * combine the page table reload and the switch backend into
2854 arch_start_context_switch(prev);
2857 next->active_mm = oldmm;
2858 atomic_inc(&oldmm->mm_count);
2859 enter_lazy_tlb(oldmm, next);
2861 switch_mm(oldmm, mm, next);
2863 if (likely(!prev->mm)) {
2864 prev->active_mm = NULL;
2865 rq->prev_mm = oldmm;
2868 * Since the runqueue lock will be released by the next
2869 * task (which is an invalid locking op but in the case
2870 * of the scheduler it's an obvious special-case), so we
2871 * do an early lockdep release here:
2873 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2874 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2877 /* Here we just switch the register state and the stack. */
2878 switch_to(prev, next, prev);
2882 * this_rq must be evaluated again because prev may have moved
2883 * CPUs since it called schedule(), thus the 'rq' on its stack
2884 * frame will be invalid.
2886 finish_task_switch(this_rq(), prev);
2890 * nr_running, nr_uninterruptible and nr_context_switches:
2892 * externally visible scheduler statistics: current number of runnable
2893 * threads, current number of uninterruptible-sleeping threads, total
2894 * number of context switches performed since bootup.
2896 unsigned long nr_running(void)
2898 unsigned long i, sum = 0;
2900 for_each_online_cpu(i)
2901 sum += cpu_rq(i)->nr_running;
2906 unsigned long nr_uninterruptible(void)
2908 unsigned long i, sum = 0;
2910 for_each_possible_cpu(i)
2911 sum += cpu_rq(i)->nr_uninterruptible;
2914 * Since we read the counters lockless, it might be slightly
2915 * inaccurate. Do not allow it to go below zero though:
2917 if (unlikely((long)sum < 0))
2923 unsigned long long nr_context_switches(void)
2926 unsigned long long sum = 0;
2928 for_each_possible_cpu(i)
2929 sum += cpu_rq(i)->nr_switches;
2934 unsigned long nr_iowait(void)
2936 unsigned long i, sum = 0;
2938 for_each_possible_cpu(i)
2939 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2944 unsigned long nr_iowait_cpu(int cpu)
2946 struct rq *this = cpu_rq(cpu);
2947 return atomic_read(&this->nr_iowait);
2950 unsigned long this_cpu_load(void)
2952 struct rq *this = this_rq();
2953 return this->cpu_load[0];
2957 /* Variables and functions for calc_load */
2958 static atomic_long_t calc_load_tasks;
2959 static unsigned long calc_load_update;
2960 unsigned long avenrun[3];
2961 EXPORT_SYMBOL(avenrun);
2963 static long calc_load_fold_active(struct rq *this_rq)
2965 long nr_active, delta = 0;
2967 nr_active = this_rq->nr_running;
2968 nr_active += (long) this_rq->nr_uninterruptible;
2970 if (nr_active != this_rq->calc_load_active) {
2971 delta = nr_active - this_rq->calc_load_active;
2972 this_rq->calc_load_active = nr_active;
2980 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2982 * When making the ILB scale, we should try to pull this in as well.
2984 static atomic_long_t calc_load_tasks_idle;
2986 static void calc_load_account_idle(struct rq *this_rq)
2990 delta = calc_load_fold_active(this_rq);
2992 atomic_long_add(delta, &calc_load_tasks_idle);
2995 static long calc_load_fold_idle(void)
3000 * Its got a race, we don't care...
3002 if (atomic_long_read(&calc_load_tasks_idle))
3003 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3008 static void calc_load_account_idle(struct rq *this_rq)
3012 static inline long calc_load_fold_idle(void)
3019 * get_avenrun - get the load average array
3020 * @loads: pointer to dest load array
3021 * @offset: offset to add
3022 * @shift: shift count to shift the result left
3024 * These values are estimates at best, so no need for locking.
3026 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3028 loads[0] = (avenrun[0] + offset) << shift;
3029 loads[1] = (avenrun[1] + offset) << shift;
3030 loads[2] = (avenrun[2] + offset) << shift;
3033 static unsigned long
3034 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3037 load += active * (FIXED_1 - exp);
3038 return load >> FSHIFT;
3042 * calc_load - update the avenrun load estimates 10 ticks after the
3043 * CPUs have updated calc_load_tasks.
3045 void calc_global_load(void)
3047 unsigned long upd = calc_load_update + 10;
3050 if (time_before(jiffies, upd))
3053 active = atomic_long_read(&calc_load_tasks);
3054 active = active > 0 ? active * FIXED_1 : 0;
3056 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3057 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3058 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3060 calc_load_update += LOAD_FREQ;
3064 * Called from update_cpu_load() to periodically update this CPU's
3067 static void calc_load_account_active(struct rq *this_rq)
3071 if (time_before(jiffies, this_rq->calc_load_update))
3074 delta = calc_load_fold_active(this_rq);
3075 delta += calc_load_fold_idle();
3077 atomic_long_add(delta, &calc_load_tasks);
3079 this_rq->calc_load_update += LOAD_FREQ;
3083 * The exact cpuload at various idx values, calculated at every tick would be
3084 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3086 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3087 * on nth tick when cpu may be busy, then we have:
3088 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3089 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3091 * decay_load_missed() below does efficient calculation of
3092 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3093 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3095 * The calculation is approximated on a 128 point scale.
3096 * degrade_zero_ticks is the number of ticks after which load at any
3097 * particular idx is approximated to be zero.
3098 * degrade_factor is a precomputed table, a row for each load idx.
3099 * Each column corresponds to degradation factor for a power of two ticks,
3100 * based on 128 point scale.
3102 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3103 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3105 * With this power of 2 load factors, we can degrade the load n times
3106 * by looking at 1 bits in n and doing as many mult/shift instead of
3107 * n mult/shifts needed by the exact degradation.
3109 #define DEGRADE_SHIFT 7
3110 static const unsigned char
3111 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3112 static const unsigned char
3113 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3114 {0, 0, 0, 0, 0, 0, 0, 0},
3115 {64, 32, 8, 0, 0, 0, 0, 0},
3116 {96, 72, 40, 12, 1, 0, 0},
3117 {112, 98, 75, 43, 15, 1, 0},
3118 {120, 112, 98, 76, 45, 16, 2} };
3121 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3122 * would be when CPU is idle and so we just decay the old load without
3123 * adding any new load.
3125 static unsigned long
3126 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3130 if (!missed_updates)
3133 if (missed_updates >= degrade_zero_ticks[idx])
3137 return load >> missed_updates;
3139 while (missed_updates) {
3140 if (missed_updates % 2)
3141 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3143 missed_updates >>= 1;
3150 * Update rq->cpu_load[] statistics. This function is usually called every
3151 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3152 * every tick. We fix it up based on jiffies.
3154 static void update_cpu_load(struct rq *this_rq)
3156 unsigned long this_load = this_rq->load.weight;
3157 unsigned long curr_jiffies = jiffies;
3158 unsigned long pending_updates;
3161 this_rq->nr_load_updates++;
3163 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3164 if (curr_jiffies == this_rq->last_load_update_tick)
3167 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3168 this_rq->last_load_update_tick = curr_jiffies;
3170 /* Update our load: */
3171 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3172 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3173 unsigned long old_load, new_load;
3175 /* scale is effectively 1 << i now, and >> i divides by scale */
3177 old_load = this_rq->cpu_load[i];
3178 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3179 new_load = this_load;
3181 * Round up the averaging division if load is increasing. This
3182 * prevents us from getting stuck on 9 if the load is 10, for
3185 if (new_load > old_load)
3186 new_load += scale - 1;
3188 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3191 sched_avg_update(this_rq);
3194 static void update_cpu_load_active(struct rq *this_rq)
3196 update_cpu_load(this_rq);
3198 calc_load_account_active(this_rq);
3204 * sched_exec - execve() is a valuable balancing opportunity, because at
3205 * this point the task has the smallest effective memory and cache footprint.
3207 void sched_exec(void)
3209 struct task_struct *p = current;
3210 unsigned long flags;
3214 rq = task_rq_lock(p, &flags);
3215 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3216 if (dest_cpu == smp_processor_id())
3220 * select_task_rq() can race against ->cpus_allowed
3222 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3223 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3224 struct migration_arg arg = { p, dest_cpu };
3226 task_rq_unlock(rq, &flags);
3227 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3231 task_rq_unlock(rq, &flags);
3236 DEFINE_PER_CPU(struct kernel_stat, kstat);
3238 EXPORT_PER_CPU_SYMBOL(kstat);
3241 * Return any ns on the sched_clock that have not yet been accounted in
3242 * @p in case that task is currently running.
3244 * Called with task_rq_lock() held on @rq.
3246 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3250 if (task_current(rq, p)) {
3251 update_rq_clock(rq);
3252 ns = rq->clock - p->se.exec_start;
3260 unsigned long long task_delta_exec(struct task_struct *p)
3262 unsigned long flags;
3266 rq = task_rq_lock(p, &flags);
3267 ns = do_task_delta_exec(p, rq);
3268 task_rq_unlock(rq, &flags);
3274 * Return accounted runtime for the task.
3275 * In case the task is currently running, return the runtime plus current's
3276 * pending runtime that have not been accounted yet.
3278 unsigned long long task_sched_runtime(struct task_struct *p)
3280 unsigned long flags;
3284 rq = task_rq_lock(p, &flags);
3285 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3286 task_rq_unlock(rq, &flags);
3292 * Return sum_exec_runtime for the thread group.
3293 * In case the task is currently running, return the sum plus current's
3294 * pending runtime that have not been accounted yet.
3296 * Note that the thread group might have other running tasks as well,
3297 * so the return value not includes other pending runtime that other
3298 * running tasks might have.
3300 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3302 struct task_cputime totals;
3303 unsigned long flags;
3307 rq = task_rq_lock(p, &flags);
3308 thread_group_cputime(p, &totals);
3309 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3310 task_rq_unlock(rq, &flags);
3316 * Account user cpu time to a process.
3317 * @p: the process that the cpu time gets accounted to
3318 * @cputime: the cpu time spent in user space since the last update
3319 * @cputime_scaled: cputime scaled by cpu frequency
3321 void account_user_time(struct task_struct *p, cputime_t cputime,
3322 cputime_t cputime_scaled)
3324 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3327 /* Add user time to process. */
3328 p->utime = cputime_add(p->utime, cputime);
3329 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3330 account_group_user_time(p, cputime);
3332 /* Add user time to cpustat. */
3333 tmp = cputime_to_cputime64(cputime);
3334 if (TASK_NICE(p) > 0)
3335 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3337 cpustat->user = cputime64_add(cpustat->user, tmp);
3339 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3340 /* Account for user time used */
3341 acct_update_integrals(p);
3345 * Account guest cpu time to a process.
3346 * @p: the process that the cpu time gets accounted to
3347 * @cputime: the cpu time spent in virtual machine since the last update
3348 * @cputime_scaled: cputime scaled by cpu frequency
3350 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3351 cputime_t cputime_scaled)
3354 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3356 tmp = cputime_to_cputime64(cputime);
3358 /* Add guest time to process. */
3359 p->utime = cputime_add(p->utime, cputime);
3360 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3361 account_group_user_time(p, cputime);
3362 p->gtime = cputime_add(p->gtime, cputime);
3364 /* Add guest time to cpustat. */
3365 if (TASK_NICE(p) > 0) {
3366 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3367 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3369 cpustat->user = cputime64_add(cpustat->user, tmp);
3370 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3375 * Account system cpu time to a process.
3376 * @p: the process that the cpu time gets accounted to
3377 * @hardirq_offset: the offset to subtract from hardirq_count()
3378 * @cputime: the cpu time spent in kernel space since the last update
3379 * @cputime_scaled: cputime scaled by cpu frequency
3381 void account_system_time(struct task_struct *p, int hardirq_offset,
3382 cputime_t cputime, cputime_t cputime_scaled)
3384 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3387 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3388 account_guest_time(p, cputime, cputime_scaled);
3392 /* Add system time to process. */
3393 p->stime = cputime_add(p->stime, cputime);
3394 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3395 account_group_system_time(p, cputime);
3397 /* Add system time to cpustat. */
3398 tmp = cputime_to_cputime64(cputime);
3399 if (hardirq_count() - hardirq_offset)
3400 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3401 else if (softirq_count())
3402 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3404 cpustat->system = cputime64_add(cpustat->system, tmp);
3406 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3408 /* Account for system time used */
3409 acct_update_integrals(p);
3413 * Account for involuntary wait time.
3414 * @steal: the cpu time spent in involuntary wait
3416 void account_steal_time(cputime_t cputime)
3418 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3419 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3421 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3425 * Account for idle time.
3426 * @cputime: the cpu time spent in idle wait
3428 void account_idle_time(cputime_t cputime)
3430 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3431 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3432 struct rq *rq = this_rq();
3434 if (atomic_read(&rq->nr_iowait) > 0)
3435 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3437 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3440 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3443 * Account a single tick of cpu time.
3444 * @p: the process that the cpu time gets accounted to
3445 * @user_tick: indicates if the tick is a user or a system tick
3447 void account_process_tick(struct task_struct *p, int user_tick)
3449 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3450 struct rq *rq = this_rq();
3453 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3454 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3455 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3458 account_idle_time(cputime_one_jiffy);
3462 * Account multiple ticks of steal time.
3463 * @p: the process from which the cpu time has been stolen
3464 * @ticks: number of stolen ticks
3466 void account_steal_ticks(unsigned long ticks)
3468 account_steal_time(jiffies_to_cputime(ticks));
3472 * Account multiple ticks of idle time.
3473 * @ticks: number of stolen ticks
3475 void account_idle_ticks(unsigned long ticks)
3477 account_idle_time(jiffies_to_cputime(ticks));
3483 * Use precise platform statistics if available:
3485 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3486 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3492 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3494 struct task_cputime cputime;
3496 thread_group_cputime(p, &cputime);
3498 *ut = cputime.utime;
3499 *st = cputime.stime;
3503 #ifndef nsecs_to_cputime
3504 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3507 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3509 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3512 * Use CFS's precise accounting:
3514 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3520 do_div(temp, total);
3521 utime = (cputime_t)temp;
3526 * Compare with previous values, to keep monotonicity:
3528 p->prev_utime = max(p->prev_utime, utime);
3529 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3531 *ut = p->prev_utime;
3532 *st = p->prev_stime;
3536 * Must be called with siglock held.
3538 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3540 struct signal_struct *sig = p->signal;
3541 struct task_cputime cputime;
3542 cputime_t rtime, utime, total;
3544 thread_group_cputime(p, &cputime);
3546 total = cputime_add(cputime.utime, cputime.stime);
3547 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3552 temp *= cputime.utime;
3553 do_div(temp, total);
3554 utime = (cputime_t)temp;
3558 sig->prev_utime = max(sig->prev_utime, utime);
3559 sig->prev_stime = max(sig->prev_stime,
3560 cputime_sub(rtime, sig->prev_utime));
3562 *ut = sig->prev_utime;
3563 *st = sig->prev_stime;
3568 * This function gets called by the timer code, with HZ frequency.
3569 * We call it with interrupts disabled.
3571 * It also gets called by the fork code, when changing the parent's
3574 void scheduler_tick(void)
3576 int cpu = smp_processor_id();
3577 struct rq *rq = cpu_rq(cpu);
3578 struct task_struct *curr = rq->curr;
3582 raw_spin_lock(&rq->lock);
3583 update_rq_clock(rq);
3584 update_cpu_load_active(rq);
3585 curr->sched_class->task_tick(rq, curr, 0);
3586 raw_spin_unlock(&rq->lock);
3588 perf_event_task_tick(curr);
3591 rq->idle_at_tick = idle_cpu(cpu);
3592 trigger_load_balance(rq, cpu);
3596 notrace unsigned long get_parent_ip(unsigned long addr)
3598 if (in_lock_functions(addr)) {
3599 addr = CALLER_ADDR2;
3600 if (in_lock_functions(addr))
3601 addr = CALLER_ADDR3;
3606 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3607 defined(CONFIG_PREEMPT_TRACER))
3609 void __kprobes add_preempt_count(int val)
3611 #ifdef CONFIG_DEBUG_PREEMPT
3615 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3618 preempt_count() += val;
3619 #ifdef CONFIG_DEBUG_PREEMPT
3621 * Spinlock count overflowing soon?
3623 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3626 if (preempt_count() == val)
3627 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3629 EXPORT_SYMBOL(add_preempt_count);
3631 void __kprobes sub_preempt_count(int val)
3633 #ifdef CONFIG_DEBUG_PREEMPT
3637 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3640 * Is the spinlock portion underflowing?
3642 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3643 !(preempt_count() & PREEMPT_MASK)))
3647 if (preempt_count() == val)
3648 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3649 preempt_count() -= val;
3651 EXPORT_SYMBOL(sub_preempt_count);
3656 * Print scheduling while atomic bug:
3658 static noinline void __schedule_bug(struct task_struct *prev)
3660 struct pt_regs *regs = get_irq_regs();
3662 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3663 prev->comm, prev->pid, preempt_count());
3665 debug_show_held_locks(prev);
3667 if (irqs_disabled())
3668 print_irqtrace_events(prev);
3677 * Various schedule()-time debugging checks and statistics:
3679 static inline void schedule_debug(struct task_struct *prev)
3682 * Test if we are atomic. Since do_exit() needs to call into
3683 * schedule() atomically, we ignore that path for now.
3684 * Otherwise, whine if we are scheduling when we should not be.
3686 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3687 __schedule_bug(prev);
3689 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3691 schedstat_inc(this_rq(), sched_count);
3692 #ifdef CONFIG_SCHEDSTATS
3693 if (unlikely(prev->lock_depth >= 0)) {
3694 schedstat_inc(this_rq(), bkl_count);
3695 schedstat_inc(prev, sched_info.bkl_count);
3700 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3703 update_rq_clock(rq);
3704 rq->skip_clock_update = 0;
3705 prev->sched_class->put_prev_task(rq, prev);
3709 * Pick up the highest-prio task:
3711 static inline struct task_struct *
3712 pick_next_task(struct rq *rq)
3714 const struct sched_class *class;
3715 struct task_struct *p;
3718 * Optimization: we know that if all tasks are in
3719 * the fair class we can call that function directly:
3721 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3722 p = fair_sched_class.pick_next_task(rq);
3727 class = sched_class_highest;
3729 p = class->pick_next_task(rq);
3733 * Will never be NULL as the idle class always
3734 * returns a non-NULL p:
3736 class = class->next;
3741 * schedule() is the main scheduler function.
3743 asmlinkage void __sched schedule(void)
3745 struct task_struct *prev, *next;
3746 unsigned long *switch_count;
3752 cpu = smp_processor_id();
3754 rcu_note_context_switch(cpu);
3757 release_kernel_lock(prev);
3758 need_resched_nonpreemptible:
3760 schedule_debug(prev);
3762 if (sched_feat(HRTICK))
3765 raw_spin_lock_irq(&rq->lock);
3766 clear_tsk_need_resched(prev);
3768 switch_count = &prev->nivcsw;
3769 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3770 if (unlikely(signal_pending_state(prev->state, prev))) {
3771 prev->state = TASK_RUNNING;
3774 * If a worker is going to sleep, notify and
3775 * ask workqueue whether it wants to wake up a
3776 * task to maintain concurrency. If so, wake
3779 if (prev->flags & PF_WQ_WORKER) {
3780 struct task_struct *to_wakeup;
3782 to_wakeup = wq_worker_sleeping(prev, cpu);
3784 try_to_wake_up_local(to_wakeup);
3786 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3788 switch_count = &prev->nvcsw;
3791 pre_schedule(rq, prev);
3793 if (unlikely(!rq->nr_running))
3794 idle_balance(cpu, rq);
3796 put_prev_task(rq, prev);
3797 next = pick_next_task(rq);
3799 if (likely(prev != next)) {
3800 sched_info_switch(prev, next);
3801 perf_event_task_sched_out(prev, next);
3807 context_switch(rq, prev, next); /* unlocks the rq */
3809 * The context switch have flipped the stack from under us
3810 * and restored the local variables which were saved when
3811 * this task called schedule() in the past. prev == current
3812 * is still correct, but it can be moved to another cpu/rq.
3814 cpu = smp_processor_id();
3817 raw_spin_unlock_irq(&rq->lock);
3821 if (unlikely(reacquire_kernel_lock(prev)))
3822 goto need_resched_nonpreemptible;
3824 preempt_enable_no_resched();
3828 EXPORT_SYMBOL(schedule);
3830 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3832 * Look out! "owner" is an entirely speculative pointer
3833 * access and not reliable.
3835 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3840 if (!sched_feat(OWNER_SPIN))
3843 #ifdef CONFIG_DEBUG_PAGEALLOC
3845 * Need to access the cpu field knowing that
3846 * DEBUG_PAGEALLOC could have unmapped it if
3847 * the mutex owner just released it and exited.
3849 if (probe_kernel_address(&owner->cpu, cpu))
3856 * Even if the access succeeded (likely case),
3857 * the cpu field may no longer be valid.
3859 if (cpu >= nr_cpumask_bits)
3863 * We need to validate that we can do a
3864 * get_cpu() and that we have the percpu area.
3866 if (!cpu_online(cpu))
3873 * Owner changed, break to re-assess state.
3875 if (lock->owner != owner) {
3877 * If the lock has switched to a different owner,
3878 * we likely have heavy contention. Return 0 to quit
3879 * optimistic spinning and not contend further:
3887 * Is that owner really running on that cpu?
3889 if (task_thread_info(rq->curr) != owner || need_resched())
3899 #ifdef CONFIG_PREEMPT
3901 * this is the entry point to schedule() from in-kernel preemption
3902 * off of preempt_enable. Kernel preemptions off return from interrupt
3903 * occur there and call schedule directly.
3905 asmlinkage void __sched notrace preempt_schedule(void)
3907 struct thread_info *ti = current_thread_info();
3910 * If there is a non-zero preempt_count or interrupts are disabled,
3911 * we do not want to preempt the current task. Just return..
3913 if (likely(ti->preempt_count || irqs_disabled()))
3917 add_preempt_count_notrace(PREEMPT_ACTIVE);
3919 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3922 * Check again in case we missed a preemption opportunity
3923 * between schedule and now.
3926 } while (need_resched());
3928 EXPORT_SYMBOL(preempt_schedule);
3931 * this is the entry point to schedule() from kernel preemption
3932 * off of irq context.
3933 * Note, that this is called and return with irqs disabled. This will
3934 * protect us against recursive calling from irq.
3936 asmlinkage void __sched preempt_schedule_irq(void)
3938 struct thread_info *ti = current_thread_info();
3940 /* Catch callers which need to be fixed */
3941 BUG_ON(ti->preempt_count || !irqs_disabled());
3944 add_preempt_count(PREEMPT_ACTIVE);
3947 local_irq_disable();
3948 sub_preempt_count(PREEMPT_ACTIVE);
3951 * Check again in case we missed a preemption opportunity
3952 * between schedule and now.
3955 } while (need_resched());
3958 #endif /* CONFIG_PREEMPT */
3960 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3963 return try_to_wake_up(curr->private, mode, wake_flags);
3965 EXPORT_SYMBOL(default_wake_function);
3968 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3969 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3970 * number) then we wake all the non-exclusive tasks and one exclusive task.
3972 * There are circumstances in which we can try to wake a task which has already
3973 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3974 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3976 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3977 int nr_exclusive, int wake_flags, void *key)
3979 wait_queue_t *curr, *next;
3981 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3982 unsigned flags = curr->flags;
3984 if (curr->func(curr, mode, wake_flags, key) &&
3985 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3991 * __wake_up - wake up threads blocked on a waitqueue.
3993 * @mode: which threads
3994 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3995 * @key: is directly passed to the wakeup function
3997 * It may be assumed that this function implies a write memory barrier before
3998 * changing the task state if and only if any tasks are woken up.
4000 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4001 int nr_exclusive, void *key)
4003 unsigned long flags;
4005 spin_lock_irqsave(&q->lock, flags);
4006 __wake_up_common(q, mode, nr_exclusive, 0, key);
4007 spin_unlock_irqrestore(&q->lock, flags);
4009 EXPORT_SYMBOL(__wake_up);
4012 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4014 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4016 __wake_up_common(q, mode, 1, 0, NULL);
4018 EXPORT_SYMBOL_GPL(__wake_up_locked);
4020 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4022 __wake_up_common(q, mode, 1, 0, key);
4026 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4028 * @mode: which threads
4029 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4030 * @key: opaque value to be passed to wakeup targets
4032 * The sync wakeup differs that the waker knows that it will schedule
4033 * away soon, so while the target thread will be woken up, it will not
4034 * be migrated to another CPU - ie. the two threads are 'synchronized'
4035 * with each other. This can prevent needless bouncing between CPUs.
4037 * On UP it can prevent extra preemption.
4039 * It may be assumed that this function implies a write memory barrier before
4040 * changing the task state if and only if any tasks are woken up.
4042 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4043 int nr_exclusive, void *key)
4045 unsigned long flags;
4046 int wake_flags = WF_SYNC;
4051 if (unlikely(!nr_exclusive))
4054 spin_lock_irqsave(&q->lock, flags);
4055 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4056 spin_unlock_irqrestore(&q->lock, flags);
4058 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4061 * __wake_up_sync - see __wake_up_sync_key()
4063 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4065 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4067 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4070 * complete: - signals a single thread waiting on this completion
4071 * @x: holds the state of this particular completion
4073 * This will wake up a single thread waiting on this completion. Threads will be
4074 * awakened in the same order in which they were queued.
4076 * See also complete_all(), wait_for_completion() and related routines.
4078 * It may be assumed that this function implies a write memory barrier before
4079 * changing the task state if and only if any tasks are woken up.
4081 void complete(struct completion *x)
4083 unsigned long flags;
4085 spin_lock_irqsave(&x->wait.lock, flags);
4087 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4088 spin_unlock_irqrestore(&x->wait.lock, flags);
4090 EXPORT_SYMBOL(complete);
4093 * complete_all: - signals all threads waiting on this completion
4094 * @x: holds the state of this particular completion
4096 * This will wake up all threads waiting on this particular completion event.
4098 * It may be assumed that this function implies a write memory barrier before
4099 * changing the task state if and only if any tasks are woken up.
4101 void complete_all(struct completion *x)
4103 unsigned long flags;
4105 spin_lock_irqsave(&x->wait.lock, flags);
4106 x->done += UINT_MAX/2;
4107 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4108 spin_unlock_irqrestore(&x->wait.lock, flags);
4110 EXPORT_SYMBOL(complete_all);
4112 static inline long __sched
4113 do_wait_for_common(struct completion *x, long timeout, int state)
4116 DECLARE_WAITQUEUE(wait, current);
4118 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4120 if (signal_pending_state(state, current)) {
4121 timeout = -ERESTARTSYS;
4124 __set_current_state(state);
4125 spin_unlock_irq(&x->wait.lock);
4126 timeout = schedule_timeout(timeout);
4127 spin_lock_irq(&x->wait.lock);
4128 } while (!x->done && timeout);
4129 __remove_wait_queue(&x->wait, &wait);
4134 return timeout ?: 1;
4138 wait_for_common(struct completion *x, long timeout, int state)
4142 spin_lock_irq(&x->wait.lock);
4143 timeout = do_wait_for_common(x, timeout, state);
4144 spin_unlock_irq(&x->wait.lock);
4149 * wait_for_completion: - waits for completion of a task
4150 * @x: holds the state of this particular completion
4152 * This waits to be signaled for completion of a specific task. It is NOT
4153 * interruptible and there is no timeout.
4155 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4156 * and interrupt capability. Also see complete().
4158 void __sched wait_for_completion(struct completion *x)
4160 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4162 EXPORT_SYMBOL(wait_for_completion);
4165 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4166 * @x: holds the state of this particular completion
4167 * @timeout: timeout value in jiffies
4169 * This waits for either a completion of a specific task to be signaled or for a
4170 * specified timeout to expire. The timeout is in jiffies. It is not
4173 unsigned long __sched
4174 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4176 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4178 EXPORT_SYMBOL(wait_for_completion_timeout);
4181 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4182 * @x: holds the state of this particular completion
4184 * This waits for completion of a specific task to be signaled. It is
4187 int __sched wait_for_completion_interruptible(struct completion *x)
4189 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4190 if (t == -ERESTARTSYS)
4194 EXPORT_SYMBOL(wait_for_completion_interruptible);
4197 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4198 * @x: holds the state of this particular completion
4199 * @timeout: timeout value in jiffies
4201 * This waits for either a completion of a specific task to be signaled or for a
4202 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4204 unsigned long __sched
4205 wait_for_completion_interruptible_timeout(struct completion *x,
4206 unsigned long timeout)
4208 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4210 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4213 * wait_for_completion_killable: - waits for completion of a task (killable)
4214 * @x: holds the state of this particular completion
4216 * This waits to be signaled for completion of a specific task. It can be
4217 * interrupted by a kill signal.
4219 int __sched wait_for_completion_killable(struct completion *x)
4221 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4222 if (t == -ERESTARTSYS)
4226 EXPORT_SYMBOL(wait_for_completion_killable);
4229 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4230 * @x: holds the state of this particular completion
4231 * @timeout: timeout value in jiffies
4233 * This waits for either a completion of a specific task to be
4234 * signaled or for a specified timeout to expire. It can be
4235 * interrupted by a kill signal. The timeout is in jiffies.
4237 unsigned long __sched
4238 wait_for_completion_killable_timeout(struct completion *x,
4239 unsigned long timeout)
4241 return wait_for_common(x, timeout, TASK_KILLABLE);
4243 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4246 * try_wait_for_completion - try to decrement a completion without blocking
4247 * @x: completion structure
4249 * Returns: 0 if a decrement cannot be done without blocking
4250 * 1 if a decrement succeeded.
4252 * If a completion is being used as a counting completion,
4253 * attempt to decrement the counter without blocking. This
4254 * enables us to avoid waiting if the resource the completion
4255 * is protecting is not available.
4257 bool try_wait_for_completion(struct completion *x)
4259 unsigned long flags;
4262 spin_lock_irqsave(&x->wait.lock, flags);
4267 spin_unlock_irqrestore(&x->wait.lock, flags);
4270 EXPORT_SYMBOL(try_wait_for_completion);
4273 * completion_done - Test to see if a completion has any waiters
4274 * @x: completion structure
4276 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4277 * 1 if there are no waiters.
4280 bool completion_done(struct completion *x)
4282 unsigned long flags;
4285 spin_lock_irqsave(&x->wait.lock, flags);
4288 spin_unlock_irqrestore(&x->wait.lock, flags);
4291 EXPORT_SYMBOL(completion_done);
4294 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4296 unsigned long flags;
4299 init_waitqueue_entry(&wait, current);
4301 __set_current_state(state);
4303 spin_lock_irqsave(&q->lock, flags);
4304 __add_wait_queue(q, &wait);
4305 spin_unlock(&q->lock);
4306 timeout = schedule_timeout(timeout);
4307 spin_lock_irq(&q->lock);
4308 __remove_wait_queue(q, &wait);
4309 spin_unlock_irqrestore(&q->lock, flags);
4314 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4316 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4318 EXPORT_SYMBOL(interruptible_sleep_on);
4321 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4323 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4325 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4327 void __sched sleep_on(wait_queue_head_t *q)
4329 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4331 EXPORT_SYMBOL(sleep_on);
4333 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4335 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4337 EXPORT_SYMBOL(sleep_on_timeout);
4339 #ifdef CONFIG_RT_MUTEXES
4342 * rt_mutex_setprio - set the current priority of a task
4344 * @prio: prio value (kernel-internal form)
4346 * This function changes the 'effective' priority of a task. It does
4347 * not touch ->normal_prio like __setscheduler().
4349 * Used by the rt_mutex code to implement priority inheritance logic.
4351 void rt_mutex_setprio(struct task_struct *p, int prio)
4353 unsigned long flags;
4354 int oldprio, on_rq, running;
4356 const struct sched_class *prev_class;
4358 BUG_ON(prio < 0 || prio > MAX_PRIO);
4360 rq = task_rq_lock(p, &flags);
4363 prev_class = p->sched_class;
4364 on_rq = p->se.on_rq;
4365 running = task_current(rq, p);
4367 dequeue_task(rq, p, 0);
4369 p->sched_class->put_prev_task(rq, p);
4372 p->sched_class = &rt_sched_class;
4374 p->sched_class = &fair_sched_class;
4379 p->sched_class->set_curr_task(rq);
4381 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4383 check_class_changed(rq, p, prev_class, oldprio, running);
4385 task_rq_unlock(rq, &flags);
4390 void set_user_nice(struct task_struct *p, long nice)
4392 int old_prio, delta, on_rq;
4393 unsigned long flags;
4396 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4399 * We have to be careful, if called from sys_setpriority(),
4400 * the task might be in the middle of scheduling on another CPU.
4402 rq = task_rq_lock(p, &flags);
4404 * The RT priorities are set via sched_setscheduler(), but we still
4405 * allow the 'normal' nice value to be set - but as expected
4406 * it wont have any effect on scheduling until the task is
4407 * SCHED_FIFO/SCHED_RR:
4409 if (task_has_rt_policy(p)) {
4410 p->static_prio = NICE_TO_PRIO(nice);
4413 on_rq = p->se.on_rq;
4415 dequeue_task(rq, p, 0);
4417 p->static_prio = NICE_TO_PRIO(nice);
4420 p->prio = effective_prio(p);
4421 delta = p->prio - old_prio;
4424 enqueue_task(rq, p, 0);
4426 * If the task increased its priority or is running and
4427 * lowered its priority, then reschedule its CPU:
4429 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4430 resched_task(rq->curr);
4433 task_rq_unlock(rq, &flags);
4435 EXPORT_SYMBOL(set_user_nice);
4438 * can_nice - check if a task can reduce its nice value
4442 int can_nice(const struct task_struct *p, const int nice)
4444 /* convert nice value [19,-20] to rlimit style value [1,40] */
4445 int nice_rlim = 20 - nice;
4447 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4448 capable(CAP_SYS_NICE));
4451 #ifdef __ARCH_WANT_SYS_NICE
4454 * sys_nice - change the priority of the current process.
4455 * @increment: priority increment
4457 * sys_setpriority is a more generic, but much slower function that
4458 * does similar things.
4460 SYSCALL_DEFINE1(nice, int, increment)
4465 * Setpriority might change our priority at the same moment.
4466 * We don't have to worry. Conceptually one call occurs first
4467 * and we have a single winner.
4469 if (increment < -40)
4474 nice = TASK_NICE(current) + increment;
4480 if (increment < 0 && !can_nice(current, nice))
4483 retval = security_task_setnice(current, nice);
4487 set_user_nice(current, nice);
4494 * task_prio - return the priority value of a given task.
4495 * @p: the task in question.
4497 * This is the priority value as seen by users in /proc.
4498 * RT tasks are offset by -200. Normal tasks are centered
4499 * around 0, value goes from -16 to +15.
4501 int task_prio(const struct task_struct *p)
4503 return p->prio - MAX_RT_PRIO;
4507 * task_nice - return the nice value of a given task.
4508 * @p: the task in question.
4510 int task_nice(const struct task_struct *p)
4512 return TASK_NICE(p);
4514 EXPORT_SYMBOL(task_nice);
4517 * idle_cpu - is a given cpu idle currently?
4518 * @cpu: the processor in question.
4520 int idle_cpu(int cpu)
4522 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4526 * idle_task - return the idle task for a given cpu.
4527 * @cpu: the processor in question.
4529 struct task_struct *idle_task(int cpu)
4531 return cpu_rq(cpu)->idle;
4535 * find_process_by_pid - find a process with a matching PID value.
4536 * @pid: the pid in question.
4538 static struct task_struct *find_process_by_pid(pid_t pid)
4540 return pid ? find_task_by_vpid(pid) : current;
4543 /* Actually do priority change: must hold rq lock. */
4545 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4547 BUG_ON(p->se.on_rq);
4550 p->rt_priority = prio;
4551 p->normal_prio = normal_prio(p);
4552 /* we are holding p->pi_lock already */
4553 p->prio = rt_mutex_getprio(p);
4554 if (rt_prio(p->prio))
4555 p->sched_class = &rt_sched_class;
4557 p->sched_class = &fair_sched_class;
4562 * check the target process has a UID that matches the current process's
4564 static bool check_same_owner(struct task_struct *p)
4566 const struct cred *cred = current_cred(), *pcred;
4570 pcred = __task_cred(p);
4571 match = (cred->euid == pcred->euid ||
4572 cred->euid == pcred->uid);
4577 static int __sched_setscheduler(struct task_struct *p, int policy,
4578 struct sched_param *param, bool user)
4580 int retval, oldprio, oldpolicy = -1, on_rq, running;
4581 unsigned long flags;
4582 const struct sched_class *prev_class;
4586 /* may grab non-irq protected spin_locks */
4587 BUG_ON(in_interrupt());
4589 /* double check policy once rq lock held */
4591 reset_on_fork = p->sched_reset_on_fork;
4592 policy = oldpolicy = p->policy;
4594 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4595 policy &= ~SCHED_RESET_ON_FORK;
4597 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4598 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4599 policy != SCHED_IDLE)
4604 * Valid priorities for SCHED_FIFO and SCHED_RR are
4605 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4606 * SCHED_BATCH and SCHED_IDLE is 0.
4608 if (param->sched_priority < 0 ||
4609 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4610 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4612 if (rt_policy(policy) != (param->sched_priority != 0))
4616 * Allow unprivileged RT tasks to decrease priority:
4618 if (user && !capable(CAP_SYS_NICE)) {
4619 if (rt_policy(policy)) {
4620 unsigned long rlim_rtprio =
4621 task_rlimit(p, RLIMIT_RTPRIO);
4623 /* can't set/change the rt policy */
4624 if (policy != p->policy && !rlim_rtprio)
4627 /* can't increase priority */
4628 if (param->sched_priority > p->rt_priority &&
4629 param->sched_priority > rlim_rtprio)
4633 * Like positive nice levels, dont allow tasks to
4634 * move out of SCHED_IDLE either:
4636 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4639 /* can't change other user's priorities */
4640 if (!check_same_owner(p))
4643 /* Normal users shall not reset the sched_reset_on_fork flag */
4644 if (p->sched_reset_on_fork && !reset_on_fork)
4649 retval = security_task_setscheduler(p, policy, param);
4655 * make sure no PI-waiters arrive (or leave) while we are
4656 * changing the priority of the task:
4658 raw_spin_lock_irqsave(&p->pi_lock, flags);
4660 * To be able to change p->policy safely, the apropriate
4661 * runqueue lock must be held.
4663 rq = __task_rq_lock(p);
4665 #ifdef CONFIG_RT_GROUP_SCHED
4668 * Do not allow realtime tasks into groups that have no runtime
4671 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4672 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4673 __task_rq_unlock(rq);
4674 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4680 /* recheck policy now with rq lock held */
4681 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4682 policy = oldpolicy = -1;
4683 __task_rq_unlock(rq);
4684 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4687 on_rq = p->se.on_rq;
4688 running = task_current(rq, p);
4690 deactivate_task(rq, p, 0);
4692 p->sched_class->put_prev_task(rq, p);
4694 p->sched_reset_on_fork = reset_on_fork;
4697 prev_class = p->sched_class;
4698 __setscheduler(rq, p, policy, param->sched_priority);
4701 p->sched_class->set_curr_task(rq);
4703 activate_task(rq, p, 0);
4705 check_class_changed(rq, p, prev_class, oldprio, running);
4707 __task_rq_unlock(rq);
4708 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4710 rt_mutex_adjust_pi(p);
4716 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4717 * @p: the task in question.
4718 * @policy: new policy.
4719 * @param: structure containing the new RT priority.
4721 * NOTE that the task may be already dead.
4723 int sched_setscheduler(struct task_struct *p, int policy,
4724 struct sched_param *param)
4726 return __sched_setscheduler(p, policy, param, true);
4728 EXPORT_SYMBOL_GPL(sched_setscheduler);
4731 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4732 * @p: the task in question.
4733 * @policy: new policy.
4734 * @param: structure containing the new RT priority.
4736 * Just like sched_setscheduler, only don't bother checking if the
4737 * current context has permission. For example, this is needed in
4738 * stop_machine(): we create temporary high priority worker threads,
4739 * but our caller might not have that capability.
4741 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4742 struct sched_param *param)
4744 return __sched_setscheduler(p, policy, param, false);
4748 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4750 struct sched_param lparam;
4751 struct task_struct *p;
4754 if (!param || pid < 0)
4756 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4761 p = find_process_by_pid(pid);
4763 retval = sched_setscheduler(p, policy, &lparam);
4770 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4771 * @pid: the pid in question.
4772 * @policy: new policy.
4773 * @param: structure containing the new RT priority.
4775 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4776 struct sched_param __user *, param)
4778 /* negative values for policy are not valid */
4782 return do_sched_setscheduler(pid, policy, param);
4786 * sys_sched_setparam - set/change the RT priority of a thread
4787 * @pid: the pid in question.
4788 * @param: structure containing the new RT priority.
4790 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4792 return do_sched_setscheduler(pid, -1, param);
4796 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4797 * @pid: the pid in question.
4799 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4801 struct task_struct *p;
4809 p = find_process_by_pid(pid);
4811 retval = security_task_getscheduler(p);
4814 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4821 * sys_sched_getparam - get the RT priority of a thread
4822 * @pid: the pid in question.
4823 * @param: structure containing the RT priority.
4825 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4827 struct sched_param lp;
4828 struct task_struct *p;
4831 if (!param || pid < 0)
4835 p = find_process_by_pid(pid);
4840 retval = security_task_getscheduler(p);
4844 lp.sched_priority = p->rt_priority;
4848 * This one might sleep, we cannot do it with a spinlock held ...
4850 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4859 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4861 cpumask_var_t cpus_allowed, new_mask;
4862 struct task_struct *p;
4868 p = find_process_by_pid(pid);
4875 /* Prevent p going away */
4879 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4883 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4885 goto out_free_cpus_allowed;
4888 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4891 retval = security_task_setscheduler(p, 0, NULL);
4895 cpuset_cpus_allowed(p, cpus_allowed);
4896 cpumask_and(new_mask, in_mask, cpus_allowed);
4898 retval = set_cpus_allowed_ptr(p, new_mask);
4901 cpuset_cpus_allowed(p, cpus_allowed);
4902 if (!cpumask_subset(new_mask, cpus_allowed)) {
4904 * We must have raced with a concurrent cpuset
4905 * update. Just reset the cpus_allowed to the
4906 * cpuset's cpus_allowed
4908 cpumask_copy(new_mask, cpus_allowed);
4913 free_cpumask_var(new_mask);
4914 out_free_cpus_allowed:
4915 free_cpumask_var(cpus_allowed);
4922 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4923 struct cpumask *new_mask)
4925 if (len < cpumask_size())
4926 cpumask_clear(new_mask);
4927 else if (len > cpumask_size())
4928 len = cpumask_size();
4930 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4934 * sys_sched_setaffinity - set the cpu affinity of a process
4935 * @pid: pid of the process
4936 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4937 * @user_mask_ptr: user-space pointer to the new cpu mask
4939 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4940 unsigned long __user *, user_mask_ptr)
4942 cpumask_var_t new_mask;
4945 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4948 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4950 retval = sched_setaffinity(pid, new_mask);
4951 free_cpumask_var(new_mask);
4955 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4957 struct task_struct *p;
4958 unsigned long flags;
4966 p = find_process_by_pid(pid);
4970 retval = security_task_getscheduler(p);
4974 rq = task_rq_lock(p, &flags);
4975 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4976 task_rq_unlock(rq, &flags);
4986 * sys_sched_getaffinity - get the cpu affinity of a process
4987 * @pid: pid of the process
4988 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4989 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4991 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4992 unsigned long __user *, user_mask_ptr)
4997 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4999 if (len & (sizeof(unsigned long)-1))
5002 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5005 ret = sched_getaffinity(pid, mask);
5007 size_t retlen = min_t(size_t, len, cpumask_size());
5009 if (copy_to_user(user_mask_ptr, mask, retlen))
5014 free_cpumask_var(mask);
5020 * sys_sched_yield - yield the current processor to other threads.
5022 * This function yields the current CPU to other tasks. If there are no
5023 * other threads running on this CPU then this function will return.
5025 SYSCALL_DEFINE0(sched_yield)
5027 struct rq *rq = this_rq_lock();
5029 schedstat_inc(rq, yld_count);
5030 current->sched_class->yield_task(rq);
5033 * Since we are going to call schedule() anyway, there's
5034 * no need to preempt or enable interrupts:
5036 __release(rq->lock);
5037 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5038 do_raw_spin_unlock(&rq->lock);
5039 preempt_enable_no_resched();
5046 static inline int should_resched(void)
5048 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5051 static void __cond_resched(void)
5053 add_preempt_count(PREEMPT_ACTIVE);
5055 sub_preempt_count(PREEMPT_ACTIVE);
5058 int __sched _cond_resched(void)
5060 if (should_resched()) {
5066 EXPORT_SYMBOL(_cond_resched);
5069 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5070 * call schedule, and on return reacquire the lock.
5072 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5073 * operations here to prevent schedule() from being called twice (once via
5074 * spin_unlock(), once by hand).
5076 int __cond_resched_lock(spinlock_t *lock)
5078 int resched = should_resched();
5081 lockdep_assert_held(lock);
5083 if (spin_needbreak(lock) || resched) {
5094 EXPORT_SYMBOL(__cond_resched_lock);
5096 int __sched __cond_resched_softirq(void)
5098 BUG_ON(!in_softirq());
5100 if (should_resched()) {
5108 EXPORT_SYMBOL(__cond_resched_softirq);
5111 * yield - yield the current processor to other threads.
5113 * This is a shortcut for kernel-space yielding - it marks the
5114 * thread runnable and calls sys_sched_yield().
5116 void __sched yield(void)
5118 set_current_state(TASK_RUNNING);
5121 EXPORT_SYMBOL(yield);
5124 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5125 * that process accounting knows that this is a task in IO wait state.
5127 void __sched io_schedule(void)
5129 struct rq *rq = raw_rq();
5131 delayacct_blkio_start();
5132 atomic_inc(&rq->nr_iowait);
5133 current->in_iowait = 1;
5135 current->in_iowait = 0;
5136 atomic_dec(&rq->nr_iowait);
5137 delayacct_blkio_end();
5139 EXPORT_SYMBOL(io_schedule);
5141 long __sched io_schedule_timeout(long timeout)
5143 struct rq *rq = raw_rq();
5146 delayacct_blkio_start();
5147 atomic_inc(&rq->nr_iowait);
5148 current->in_iowait = 1;
5149 ret = schedule_timeout(timeout);
5150 current->in_iowait = 0;
5151 atomic_dec(&rq->nr_iowait);
5152 delayacct_blkio_end();
5157 * sys_sched_get_priority_max - return maximum RT priority.
5158 * @policy: scheduling class.
5160 * this syscall returns the maximum rt_priority that can be used
5161 * by a given scheduling class.
5163 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5170 ret = MAX_USER_RT_PRIO-1;
5182 * sys_sched_get_priority_min - return minimum RT priority.
5183 * @policy: scheduling class.
5185 * this syscall returns the minimum rt_priority that can be used
5186 * by a given scheduling class.
5188 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5206 * sys_sched_rr_get_interval - return the default timeslice of a process.
5207 * @pid: pid of the process.
5208 * @interval: userspace pointer to the timeslice value.
5210 * this syscall writes the default timeslice value of a given process
5211 * into the user-space timespec buffer. A value of '0' means infinity.
5213 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5214 struct timespec __user *, interval)
5216 struct task_struct *p;
5217 unsigned int time_slice;
5218 unsigned long flags;
5228 p = find_process_by_pid(pid);
5232 retval = security_task_getscheduler(p);
5236 rq = task_rq_lock(p, &flags);
5237 time_slice = p->sched_class->get_rr_interval(rq, p);
5238 task_rq_unlock(rq, &flags);
5241 jiffies_to_timespec(time_slice, &t);
5242 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5250 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5252 void sched_show_task(struct task_struct *p)
5254 unsigned long free = 0;
5257 state = p->state ? __ffs(p->state) + 1 : 0;
5258 printk(KERN_INFO "%-15.15s %c", p->comm,
5259 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5260 #if BITS_PER_LONG == 32
5261 if (state == TASK_RUNNING)
5262 printk(KERN_CONT " running ");
5264 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5266 if (state == TASK_RUNNING)
5267 printk(KERN_CONT " running task ");
5269 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5271 #ifdef CONFIG_DEBUG_STACK_USAGE
5272 free = stack_not_used(p);
5274 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5275 task_pid_nr(p), task_pid_nr(p->real_parent),
5276 (unsigned long)task_thread_info(p)->flags);
5278 show_stack(p, NULL);
5281 void show_state_filter(unsigned long state_filter)
5283 struct task_struct *g, *p;
5285 #if BITS_PER_LONG == 32
5287 " task PC stack pid father\n");
5290 " task PC stack pid father\n");
5292 read_lock(&tasklist_lock);
5293 do_each_thread(g, p) {
5295 * reset the NMI-timeout, listing all files on a slow
5296 * console might take alot of time:
5298 touch_nmi_watchdog();
5299 if (!state_filter || (p->state & state_filter))
5301 } while_each_thread(g, p);
5303 touch_all_softlockup_watchdogs();
5305 #ifdef CONFIG_SCHED_DEBUG
5306 sysrq_sched_debug_show();
5308 read_unlock(&tasklist_lock);
5310 * Only show locks if all tasks are dumped:
5313 debug_show_all_locks();
5316 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5318 idle->sched_class = &idle_sched_class;
5322 * init_idle - set up an idle thread for a given CPU
5323 * @idle: task in question
5324 * @cpu: cpu the idle task belongs to
5326 * NOTE: this function does not set the idle thread's NEED_RESCHED
5327 * flag, to make booting more robust.
5329 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5331 struct rq *rq = cpu_rq(cpu);
5332 unsigned long flags;
5334 raw_spin_lock_irqsave(&rq->lock, flags);
5337 idle->state = TASK_RUNNING;
5338 idle->se.exec_start = sched_clock();
5340 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5341 __set_task_cpu(idle, cpu);
5343 rq->curr = rq->idle = idle;
5344 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5347 raw_spin_unlock_irqrestore(&rq->lock, flags);
5349 /* Set the preempt count _outside_ the spinlocks! */
5350 #if defined(CONFIG_PREEMPT)
5351 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5353 task_thread_info(idle)->preempt_count = 0;
5356 * The idle tasks have their own, simple scheduling class:
5358 idle->sched_class = &idle_sched_class;
5359 ftrace_graph_init_task(idle);
5363 * In a system that switches off the HZ timer nohz_cpu_mask
5364 * indicates which cpus entered this state. This is used
5365 * in the rcu update to wait only for active cpus. For system
5366 * which do not switch off the HZ timer nohz_cpu_mask should
5367 * always be CPU_BITS_NONE.
5369 cpumask_var_t nohz_cpu_mask;
5372 * Increase the granularity value when there are more CPUs,
5373 * because with more CPUs the 'effective latency' as visible
5374 * to users decreases. But the relationship is not linear,
5375 * so pick a second-best guess by going with the log2 of the
5378 * This idea comes from the SD scheduler of Con Kolivas:
5380 static int get_update_sysctl_factor(void)
5382 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5383 unsigned int factor;
5385 switch (sysctl_sched_tunable_scaling) {
5386 case SCHED_TUNABLESCALING_NONE:
5389 case SCHED_TUNABLESCALING_LINEAR:
5392 case SCHED_TUNABLESCALING_LOG:
5394 factor = 1 + ilog2(cpus);
5401 static void update_sysctl(void)
5403 unsigned int factor = get_update_sysctl_factor();
5405 #define SET_SYSCTL(name) \
5406 (sysctl_##name = (factor) * normalized_sysctl_##name)
5407 SET_SYSCTL(sched_min_granularity);
5408 SET_SYSCTL(sched_latency);
5409 SET_SYSCTL(sched_wakeup_granularity);
5410 SET_SYSCTL(sched_shares_ratelimit);
5414 static inline void sched_init_granularity(void)
5421 * This is how migration works:
5423 * 1) we invoke migration_cpu_stop() on the target CPU using
5425 * 2) stopper starts to run (implicitly forcing the migrated thread
5427 * 3) it checks whether the migrated task is still in the wrong runqueue.
5428 * 4) if it's in the wrong runqueue then the migration thread removes
5429 * it and puts it into the right queue.
5430 * 5) stopper completes and stop_one_cpu() returns and the migration
5435 * Change a given task's CPU affinity. Migrate the thread to a
5436 * proper CPU and schedule it away if the CPU it's executing on
5437 * is removed from the allowed bitmask.
5439 * NOTE: the caller must have a valid reference to the task, the
5440 * task must not exit() & deallocate itself prematurely. The
5441 * call is not atomic; no spinlocks may be held.
5443 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5445 unsigned long flags;
5447 unsigned int dest_cpu;
5451 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5452 * drop the rq->lock and still rely on ->cpus_allowed.
5455 while (task_is_waking(p))
5457 rq = task_rq_lock(p, &flags);
5458 if (task_is_waking(p)) {
5459 task_rq_unlock(rq, &flags);
5463 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5468 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5469 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5474 if (p->sched_class->set_cpus_allowed)
5475 p->sched_class->set_cpus_allowed(p, new_mask);
5477 cpumask_copy(&p->cpus_allowed, new_mask);
5478 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5481 /* Can the task run on the task's current CPU? If so, we're done */
5482 if (cpumask_test_cpu(task_cpu(p), new_mask))
5485 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5486 if (migrate_task(p, dest_cpu)) {
5487 struct migration_arg arg = { p, dest_cpu };
5488 /* Need help from migration thread: drop lock and wait. */
5489 task_rq_unlock(rq, &flags);
5490 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5491 tlb_migrate_finish(p->mm);
5495 task_rq_unlock(rq, &flags);
5499 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5502 * Move (not current) task off this cpu, onto dest cpu. We're doing
5503 * this because either it can't run here any more (set_cpus_allowed()
5504 * away from this CPU, or CPU going down), or because we're
5505 * attempting to rebalance this task on exec (sched_exec).
5507 * So we race with normal scheduler movements, but that's OK, as long
5508 * as the task is no longer on this CPU.
5510 * Returns non-zero if task was successfully migrated.
5512 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5514 struct rq *rq_dest, *rq_src;
5517 if (unlikely(!cpu_active(dest_cpu)))
5520 rq_src = cpu_rq(src_cpu);
5521 rq_dest = cpu_rq(dest_cpu);
5523 double_rq_lock(rq_src, rq_dest);
5524 /* Already moved. */
5525 if (task_cpu(p) != src_cpu)
5527 /* Affinity changed (again). */
5528 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5532 * If we're not on a rq, the next wake-up will ensure we're
5536 deactivate_task(rq_src, p, 0);
5537 set_task_cpu(p, dest_cpu);
5538 activate_task(rq_dest, p, 0);
5539 check_preempt_curr(rq_dest, p, 0);
5544 double_rq_unlock(rq_src, rq_dest);
5549 * migration_cpu_stop - this will be executed by a highprio stopper thread
5550 * and performs thread migration by bumping thread off CPU then
5551 * 'pushing' onto another runqueue.
5553 static int migration_cpu_stop(void *data)
5555 struct migration_arg *arg = data;
5558 * The original target cpu might have gone down and we might
5559 * be on another cpu but it doesn't matter.
5561 local_irq_disable();
5562 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5567 #ifdef CONFIG_HOTPLUG_CPU
5569 * Figure out where task on dead CPU should go, use force if necessary.
5571 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5573 struct rq *rq = cpu_rq(dead_cpu);
5574 int needs_cpu, uninitialized_var(dest_cpu);
5575 unsigned long flags;
5577 local_irq_save(flags);
5579 raw_spin_lock(&rq->lock);
5580 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5582 dest_cpu = select_fallback_rq(dead_cpu, p);
5583 raw_spin_unlock(&rq->lock);
5585 * It can only fail if we race with set_cpus_allowed(),
5586 * in the racer should migrate the task anyway.
5589 __migrate_task(p, dead_cpu, dest_cpu);
5590 local_irq_restore(flags);
5594 * While a dead CPU has no uninterruptible tasks queued at this point,
5595 * it might still have a nonzero ->nr_uninterruptible counter, because
5596 * for performance reasons the counter is not stricly tracking tasks to
5597 * their home CPUs. So we just add the counter to another CPU's counter,
5598 * to keep the global sum constant after CPU-down:
5600 static void migrate_nr_uninterruptible(struct rq *rq_src)
5602 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5603 unsigned long flags;
5605 local_irq_save(flags);
5606 double_rq_lock(rq_src, rq_dest);
5607 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5608 rq_src->nr_uninterruptible = 0;
5609 double_rq_unlock(rq_src, rq_dest);
5610 local_irq_restore(flags);
5613 /* Run through task list and migrate tasks from the dead cpu. */
5614 static void migrate_live_tasks(int src_cpu)
5616 struct task_struct *p, *t;
5618 read_lock(&tasklist_lock);
5620 do_each_thread(t, p) {
5624 if (task_cpu(p) == src_cpu)
5625 move_task_off_dead_cpu(src_cpu, p);
5626 } while_each_thread(t, p);
5628 read_unlock(&tasklist_lock);
5632 * Schedules idle task to be the next runnable task on current CPU.
5633 * It does so by boosting its priority to highest possible.
5634 * Used by CPU offline code.
5636 void sched_idle_next(void)
5638 int this_cpu = smp_processor_id();
5639 struct rq *rq = cpu_rq(this_cpu);
5640 struct task_struct *p = rq->idle;
5641 unsigned long flags;
5643 /* cpu has to be offline */
5644 BUG_ON(cpu_online(this_cpu));
5647 * Strictly not necessary since rest of the CPUs are stopped by now
5648 * and interrupts disabled on the current cpu.
5650 raw_spin_lock_irqsave(&rq->lock, flags);
5652 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5654 activate_task(rq, p, 0);
5656 raw_spin_unlock_irqrestore(&rq->lock, flags);
5660 * Ensures that the idle task is using init_mm right before its cpu goes
5663 void idle_task_exit(void)
5665 struct mm_struct *mm = current->active_mm;
5667 BUG_ON(cpu_online(smp_processor_id()));
5670 switch_mm(mm, &init_mm, current);
5674 /* called under rq->lock with disabled interrupts */
5675 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5677 struct rq *rq = cpu_rq(dead_cpu);
5679 /* Must be exiting, otherwise would be on tasklist. */
5680 BUG_ON(!p->exit_state);
5682 /* Cannot have done final schedule yet: would have vanished. */
5683 BUG_ON(p->state == TASK_DEAD);
5688 * Drop lock around migration; if someone else moves it,
5689 * that's OK. No task can be added to this CPU, so iteration is
5692 raw_spin_unlock_irq(&rq->lock);
5693 move_task_off_dead_cpu(dead_cpu, p);
5694 raw_spin_lock_irq(&rq->lock);
5699 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5700 static void migrate_dead_tasks(unsigned int dead_cpu)
5702 struct rq *rq = cpu_rq(dead_cpu);
5703 struct task_struct *next;
5706 if (!rq->nr_running)
5708 next = pick_next_task(rq);
5711 next->sched_class->put_prev_task(rq, next);
5712 migrate_dead(dead_cpu, next);
5718 * remove the tasks which were accounted by rq from calc_load_tasks.
5720 static void calc_global_load_remove(struct rq *rq)
5722 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5723 rq->calc_load_active = 0;
5725 #endif /* CONFIG_HOTPLUG_CPU */
5727 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5729 static struct ctl_table sd_ctl_dir[] = {
5731 .procname = "sched_domain",
5737 static struct ctl_table sd_ctl_root[] = {
5739 .procname = "kernel",
5741 .child = sd_ctl_dir,
5746 static struct ctl_table *sd_alloc_ctl_entry(int n)
5748 struct ctl_table *entry =
5749 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5754 static void sd_free_ctl_entry(struct ctl_table **tablep)
5756 struct ctl_table *entry;
5759 * In the intermediate directories, both the child directory and
5760 * procname are dynamically allocated and could fail but the mode
5761 * will always be set. In the lowest directory the names are
5762 * static strings and all have proc handlers.
5764 for (entry = *tablep; entry->mode; entry++) {
5766 sd_free_ctl_entry(&entry->child);
5767 if (entry->proc_handler == NULL)
5768 kfree(entry->procname);
5776 set_table_entry(struct ctl_table *entry,
5777 const char *procname, void *data, int maxlen,
5778 mode_t mode, proc_handler *proc_handler)
5780 entry->procname = procname;
5782 entry->maxlen = maxlen;
5784 entry->proc_handler = proc_handler;
5787 static struct ctl_table *
5788 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5790 struct ctl_table *table = sd_alloc_ctl_entry(13);
5795 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5796 sizeof(long), 0644, proc_doulongvec_minmax);
5797 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5798 sizeof(long), 0644, proc_doulongvec_minmax);
5799 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5800 sizeof(int), 0644, proc_dointvec_minmax);
5801 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5802 sizeof(int), 0644, proc_dointvec_minmax);
5803 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5804 sizeof(int), 0644, proc_dointvec_minmax);
5805 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5806 sizeof(int), 0644, proc_dointvec_minmax);
5807 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5808 sizeof(int), 0644, proc_dointvec_minmax);
5809 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5810 sizeof(int), 0644, proc_dointvec_minmax);
5811 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5812 sizeof(int), 0644, proc_dointvec_minmax);
5813 set_table_entry(&table[9], "cache_nice_tries",
5814 &sd->cache_nice_tries,
5815 sizeof(int), 0644, proc_dointvec_minmax);
5816 set_table_entry(&table[10], "flags", &sd->flags,
5817 sizeof(int), 0644, proc_dointvec_minmax);
5818 set_table_entry(&table[11], "name", sd->name,
5819 CORENAME_MAX_SIZE, 0444, proc_dostring);
5820 /* &table[12] is terminator */
5825 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5827 struct ctl_table *entry, *table;
5828 struct sched_domain *sd;
5829 int domain_num = 0, i;
5832 for_each_domain(cpu, sd)
5834 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5839 for_each_domain(cpu, sd) {
5840 snprintf(buf, 32, "domain%d", i);
5841 entry->procname = kstrdup(buf, GFP_KERNEL);
5843 entry->child = sd_alloc_ctl_domain_table(sd);
5850 static struct ctl_table_header *sd_sysctl_header;
5851 static void register_sched_domain_sysctl(void)
5853 int i, cpu_num = num_possible_cpus();
5854 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5857 WARN_ON(sd_ctl_dir[0].child);
5858 sd_ctl_dir[0].child = entry;
5863 for_each_possible_cpu(i) {
5864 snprintf(buf, 32, "cpu%d", i);
5865 entry->procname = kstrdup(buf, GFP_KERNEL);
5867 entry->child = sd_alloc_ctl_cpu_table(i);
5871 WARN_ON(sd_sysctl_header);
5872 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5875 /* may be called multiple times per register */
5876 static void unregister_sched_domain_sysctl(void)
5878 if (sd_sysctl_header)
5879 unregister_sysctl_table(sd_sysctl_header);
5880 sd_sysctl_header = NULL;
5881 if (sd_ctl_dir[0].child)
5882 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5885 static void register_sched_domain_sysctl(void)
5888 static void unregister_sched_domain_sysctl(void)
5893 static void set_rq_online(struct rq *rq)
5896 const struct sched_class *class;
5898 cpumask_set_cpu(rq->cpu, rq->rd->online);
5901 for_each_class(class) {
5902 if (class->rq_online)
5903 class->rq_online(rq);
5908 static void set_rq_offline(struct rq *rq)
5911 const struct sched_class *class;
5913 for_each_class(class) {
5914 if (class->rq_offline)
5915 class->rq_offline(rq);
5918 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5924 * migration_call - callback that gets triggered when a CPU is added.
5925 * Here we can start up the necessary migration thread for the new CPU.
5927 static int __cpuinit
5928 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5930 int cpu = (long)hcpu;
5931 unsigned long flags;
5932 struct rq *rq = cpu_rq(cpu);
5936 case CPU_UP_PREPARE:
5937 case CPU_UP_PREPARE_FROZEN:
5938 rq->calc_load_update = calc_load_update;
5942 case CPU_ONLINE_FROZEN:
5943 /* Update our root-domain */
5944 raw_spin_lock_irqsave(&rq->lock, flags);
5946 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5950 raw_spin_unlock_irqrestore(&rq->lock, flags);
5953 #ifdef CONFIG_HOTPLUG_CPU
5955 case CPU_DEAD_FROZEN:
5956 migrate_live_tasks(cpu);
5957 /* Idle task back to normal (off runqueue, low prio) */
5958 raw_spin_lock_irq(&rq->lock);
5959 deactivate_task(rq, rq->idle, 0);
5960 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5961 rq->idle->sched_class = &idle_sched_class;
5962 migrate_dead_tasks(cpu);
5963 raw_spin_unlock_irq(&rq->lock);
5964 migrate_nr_uninterruptible(rq);
5965 BUG_ON(rq->nr_running != 0);
5966 calc_global_load_remove(rq);
5970 case CPU_DYING_FROZEN:
5971 /* Update our root-domain */
5972 raw_spin_lock_irqsave(&rq->lock, flags);
5974 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5977 raw_spin_unlock_irqrestore(&rq->lock, flags);
5985 * Register at high priority so that task migration (migrate_all_tasks)
5986 * happens before everything else. This has to be lower priority than
5987 * the notifier in the perf_event subsystem, though.
5989 static struct notifier_block __cpuinitdata migration_notifier = {
5990 .notifier_call = migration_call,
5991 .priority = CPU_PRI_MIGRATION,
5994 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5995 unsigned long action, void *hcpu)
5997 switch (action & ~CPU_TASKS_FROZEN) {
5999 case CPU_DOWN_FAILED:
6000 set_cpu_active((long)hcpu, true);
6007 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6008 unsigned long action, void *hcpu)
6010 switch (action & ~CPU_TASKS_FROZEN) {
6011 case CPU_DOWN_PREPARE:
6012 set_cpu_active((long)hcpu, false);
6019 static int __init migration_init(void)
6021 void *cpu = (void *)(long)smp_processor_id();
6024 /* Initialize migration for the boot CPU */
6025 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6026 BUG_ON(err == NOTIFY_BAD);
6027 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6028 register_cpu_notifier(&migration_notifier);
6030 /* Register cpu active notifiers */
6031 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6032 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6036 early_initcall(migration_init);
6041 #ifdef CONFIG_SCHED_DEBUG
6043 static __read_mostly int sched_domain_debug_enabled;
6045 static int __init sched_domain_debug_setup(char *str)
6047 sched_domain_debug_enabled = 1;
6051 early_param("sched_debug", sched_domain_debug_setup);
6053 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6054 struct cpumask *groupmask)
6056 struct sched_group *group = sd->groups;
6059 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6060 cpumask_clear(groupmask);
6062 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6064 if (!(sd->flags & SD_LOAD_BALANCE)) {
6065 printk("does not load-balance\n");
6067 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6072 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6074 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6075 printk(KERN_ERR "ERROR: domain->span does not contain "
6078 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6079 printk(KERN_ERR "ERROR: domain->groups does not contain"
6083 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6087 printk(KERN_ERR "ERROR: group is NULL\n");
6091 if (!group->cpu_power) {
6092 printk(KERN_CONT "\n");
6093 printk(KERN_ERR "ERROR: domain->cpu_power not "
6098 if (!cpumask_weight(sched_group_cpus(group))) {
6099 printk(KERN_CONT "\n");
6100 printk(KERN_ERR "ERROR: empty group\n");
6104 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6105 printk(KERN_CONT "\n");
6106 printk(KERN_ERR "ERROR: repeated CPUs\n");
6110 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6112 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6114 printk(KERN_CONT " %s", str);
6115 if (group->cpu_power != SCHED_LOAD_SCALE) {
6116 printk(KERN_CONT " (cpu_power = %d)",
6120 group = group->next;
6121 } while (group != sd->groups);
6122 printk(KERN_CONT "\n");
6124 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6125 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6128 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6129 printk(KERN_ERR "ERROR: parent span is not a superset "
6130 "of domain->span\n");
6134 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6136 cpumask_var_t groupmask;
6139 if (!sched_domain_debug_enabled)
6143 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6147 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6149 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6150 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6155 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6162 free_cpumask_var(groupmask);
6164 #else /* !CONFIG_SCHED_DEBUG */
6165 # define sched_domain_debug(sd, cpu) do { } while (0)
6166 #endif /* CONFIG_SCHED_DEBUG */
6168 static int sd_degenerate(struct sched_domain *sd)
6170 if (cpumask_weight(sched_domain_span(sd)) == 1)
6173 /* Following flags need at least 2 groups */
6174 if (sd->flags & (SD_LOAD_BALANCE |
6175 SD_BALANCE_NEWIDLE |
6179 SD_SHARE_PKG_RESOURCES)) {
6180 if (sd->groups != sd->groups->next)
6184 /* Following flags don't use groups */
6185 if (sd->flags & (SD_WAKE_AFFINE))
6192 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6194 unsigned long cflags = sd->flags, pflags = parent->flags;
6196 if (sd_degenerate(parent))
6199 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6202 /* Flags needing groups don't count if only 1 group in parent */
6203 if (parent->groups == parent->groups->next) {
6204 pflags &= ~(SD_LOAD_BALANCE |
6205 SD_BALANCE_NEWIDLE |
6209 SD_SHARE_PKG_RESOURCES);
6210 if (nr_node_ids == 1)
6211 pflags &= ~SD_SERIALIZE;
6213 if (~cflags & pflags)
6219 static void free_rootdomain(struct root_domain *rd)
6221 synchronize_sched();
6223 cpupri_cleanup(&rd->cpupri);
6225 free_cpumask_var(rd->rto_mask);
6226 free_cpumask_var(rd->online);
6227 free_cpumask_var(rd->span);
6231 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6233 struct root_domain *old_rd = NULL;
6234 unsigned long flags;
6236 raw_spin_lock_irqsave(&rq->lock, flags);
6241 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6244 cpumask_clear_cpu(rq->cpu, old_rd->span);
6247 * If we dont want to free the old_rt yet then
6248 * set old_rd to NULL to skip the freeing later
6251 if (!atomic_dec_and_test(&old_rd->refcount))
6255 atomic_inc(&rd->refcount);
6258 cpumask_set_cpu(rq->cpu, rd->span);
6259 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6262 raw_spin_unlock_irqrestore(&rq->lock, flags);
6265 free_rootdomain(old_rd);
6268 static int init_rootdomain(struct root_domain *rd)
6270 memset(rd, 0, sizeof(*rd));
6272 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6274 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6276 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6279 if (cpupri_init(&rd->cpupri) != 0)
6284 free_cpumask_var(rd->rto_mask);
6286 free_cpumask_var(rd->online);
6288 free_cpumask_var(rd->span);
6293 static void init_defrootdomain(void)
6295 init_rootdomain(&def_root_domain);
6297 atomic_set(&def_root_domain.refcount, 1);
6300 static struct root_domain *alloc_rootdomain(void)
6302 struct root_domain *rd;
6304 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6308 if (init_rootdomain(rd) != 0) {
6317 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6318 * hold the hotplug lock.
6321 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6323 struct rq *rq = cpu_rq(cpu);
6324 struct sched_domain *tmp;
6326 for (tmp = sd; tmp; tmp = tmp->parent)
6327 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6329 /* Remove the sched domains which do not contribute to scheduling. */
6330 for (tmp = sd; tmp; ) {
6331 struct sched_domain *parent = tmp->parent;
6335 if (sd_parent_degenerate(tmp, parent)) {
6336 tmp->parent = parent->parent;
6338 parent->parent->child = tmp;
6343 if (sd && sd_degenerate(sd)) {
6349 sched_domain_debug(sd, cpu);
6351 rq_attach_root(rq, rd);
6352 rcu_assign_pointer(rq->sd, sd);
6355 /* cpus with isolated domains */
6356 static cpumask_var_t cpu_isolated_map;
6358 /* Setup the mask of cpus configured for isolated domains */
6359 static int __init isolated_cpu_setup(char *str)
6361 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6362 cpulist_parse(str, cpu_isolated_map);
6366 __setup("isolcpus=", isolated_cpu_setup);
6369 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6370 * to a function which identifies what group(along with sched group) a CPU
6371 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6372 * (due to the fact that we keep track of groups covered with a struct cpumask).
6374 * init_sched_build_groups will build a circular linked list of the groups
6375 * covered by the given span, and will set each group's ->cpumask correctly,
6376 * and ->cpu_power to 0.
6379 init_sched_build_groups(const struct cpumask *span,
6380 const struct cpumask *cpu_map,
6381 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6382 struct sched_group **sg,
6383 struct cpumask *tmpmask),
6384 struct cpumask *covered, struct cpumask *tmpmask)
6386 struct sched_group *first = NULL, *last = NULL;
6389 cpumask_clear(covered);
6391 for_each_cpu(i, span) {
6392 struct sched_group *sg;
6393 int group = group_fn(i, cpu_map, &sg, tmpmask);
6396 if (cpumask_test_cpu(i, covered))
6399 cpumask_clear(sched_group_cpus(sg));
6402 for_each_cpu(j, span) {
6403 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6406 cpumask_set_cpu(j, covered);
6407 cpumask_set_cpu(j, sched_group_cpus(sg));
6418 #define SD_NODES_PER_DOMAIN 16
6423 * find_next_best_node - find the next node to include in a sched_domain
6424 * @node: node whose sched_domain we're building
6425 * @used_nodes: nodes already in the sched_domain
6427 * Find the next node to include in a given scheduling domain. Simply
6428 * finds the closest node not already in the @used_nodes map.
6430 * Should use nodemask_t.
6432 static int find_next_best_node(int node, nodemask_t *used_nodes)
6434 int i, n, val, min_val, best_node = 0;
6438 for (i = 0; i < nr_node_ids; i++) {
6439 /* Start at @node */
6440 n = (node + i) % nr_node_ids;
6442 if (!nr_cpus_node(n))
6445 /* Skip already used nodes */
6446 if (node_isset(n, *used_nodes))
6449 /* Simple min distance search */
6450 val = node_distance(node, n);
6452 if (val < min_val) {
6458 node_set(best_node, *used_nodes);
6463 * sched_domain_node_span - get a cpumask for a node's sched_domain
6464 * @node: node whose cpumask we're constructing
6465 * @span: resulting cpumask
6467 * Given a node, construct a good cpumask for its sched_domain to span. It
6468 * should be one that prevents unnecessary balancing, but also spreads tasks
6471 static void sched_domain_node_span(int node, struct cpumask *span)
6473 nodemask_t used_nodes;
6476 cpumask_clear(span);
6477 nodes_clear(used_nodes);
6479 cpumask_or(span, span, cpumask_of_node(node));
6480 node_set(node, used_nodes);
6482 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6483 int next_node = find_next_best_node(node, &used_nodes);
6485 cpumask_or(span, span, cpumask_of_node(next_node));
6488 #endif /* CONFIG_NUMA */
6490 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6493 * The cpus mask in sched_group and sched_domain hangs off the end.
6495 * ( See the the comments in include/linux/sched.h:struct sched_group
6496 * and struct sched_domain. )
6498 struct static_sched_group {
6499 struct sched_group sg;
6500 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6503 struct static_sched_domain {
6504 struct sched_domain sd;
6505 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6511 cpumask_var_t domainspan;
6512 cpumask_var_t covered;
6513 cpumask_var_t notcovered;
6515 cpumask_var_t nodemask;
6516 cpumask_var_t this_sibling_map;
6517 cpumask_var_t this_core_map;
6518 cpumask_var_t send_covered;
6519 cpumask_var_t tmpmask;
6520 struct sched_group **sched_group_nodes;
6521 struct root_domain *rd;
6525 sa_sched_groups = 0,
6530 sa_this_sibling_map,
6532 sa_sched_group_nodes,
6542 * SMT sched-domains:
6544 #ifdef CONFIG_SCHED_SMT
6545 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6546 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6549 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6550 struct sched_group **sg, struct cpumask *unused)
6553 *sg = &per_cpu(sched_groups, cpu).sg;
6556 #endif /* CONFIG_SCHED_SMT */
6559 * multi-core sched-domains:
6561 #ifdef CONFIG_SCHED_MC
6562 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6563 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6564 #endif /* CONFIG_SCHED_MC */
6566 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6568 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6569 struct sched_group **sg, struct cpumask *mask)
6573 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6574 group = cpumask_first(mask);
6576 *sg = &per_cpu(sched_group_core, group).sg;
6579 #elif defined(CONFIG_SCHED_MC)
6581 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6582 struct sched_group **sg, struct cpumask *unused)
6585 *sg = &per_cpu(sched_group_core, cpu).sg;
6590 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6591 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6594 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6595 struct sched_group **sg, struct cpumask *mask)
6598 #ifdef CONFIG_SCHED_MC
6599 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6600 group = cpumask_first(mask);
6601 #elif defined(CONFIG_SCHED_SMT)
6602 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6603 group = cpumask_first(mask);
6608 *sg = &per_cpu(sched_group_phys, group).sg;
6614 * The init_sched_build_groups can't handle what we want to do with node
6615 * groups, so roll our own. Now each node has its own list of groups which
6616 * gets dynamically allocated.
6618 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6619 static struct sched_group ***sched_group_nodes_bycpu;
6621 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6622 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6624 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6625 struct sched_group **sg,
6626 struct cpumask *nodemask)
6630 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6631 group = cpumask_first(nodemask);
6634 *sg = &per_cpu(sched_group_allnodes, group).sg;
6638 static void init_numa_sched_groups_power(struct sched_group *group_head)
6640 struct sched_group *sg = group_head;
6646 for_each_cpu(j, sched_group_cpus(sg)) {
6647 struct sched_domain *sd;
6649 sd = &per_cpu(phys_domains, j).sd;
6650 if (j != group_first_cpu(sd->groups)) {
6652 * Only add "power" once for each
6658 sg->cpu_power += sd->groups->cpu_power;
6661 } while (sg != group_head);
6664 static int build_numa_sched_groups(struct s_data *d,
6665 const struct cpumask *cpu_map, int num)
6667 struct sched_domain *sd;
6668 struct sched_group *sg, *prev;
6671 cpumask_clear(d->covered);
6672 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6673 if (cpumask_empty(d->nodemask)) {
6674 d->sched_group_nodes[num] = NULL;
6678 sched_domain_node_span(num, d->domainspan);
6679 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6681 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6684 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6688 d->sched_group_nodes[num] = sg;
6690 for_each_cpu(j, d->nodemask) {
6691 sd = &per_cpu(node_domains, j).sd;
6696 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6698 cpumask_or(d->covered, d->covered, d->nodemask);
6701 for (j = 0; j < nr_node_ids; j++) {
6702 n = (num + j) % nr_node_ids;
6703 cpumask_complement(d->notcovered, d->covered);
6704 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6705 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6706 if (cpumask_empty(d->tmpmask))
6708 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6709 if (cpumask_empty(d->tmpmask))
6711 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6715 "Can not alloc domain group for node %d\n", j);
6719 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6720 sg->next = prev->next;
6721 cpumask_or(d->covered, d->covered, d->tmpmask);
6728 #endif /* CONFIG_NUMA */
6731 /* Free memory allocated for various sched_group structures */
6732 static void free_sched_groups(const struct cpumask *cpu_map,
6733 struct cpumask *nodemask)
6737 for_each_cpu(cpu, cpu_map) {
6738 struct sched_group **sched_group_nodes
6739 = sched_group_nodes_bycpu[cpu];
6741 if (!sched_group_nodes)
6744 for (i = 0; i < nr_node_ids; i++) {
6745 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6747 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6748 if (cpumask_empty(nodemask))
6758 if (oldsg != sched_group_nodes[i])
6761 kfree(sched_group_nodes);
6762 sched_group_nodes_bycpu[cpu] = NULL;
6765 #else /* !CONFIG_NUMA */
6766 static void free_sched_groups(const struct cpumask *cpu_map,
6767 struct cpumask *nodemask)
6770 #endif /* CONFIG_NUMA */
6773 * Initialize sched groups cpu_power.
6775 * cpu_power indicates the capacity of sched group, which is used while
6776 * distributing the load between different sched groups in a sched domain.
6777 * Typically cpu_power for all the groups in a sched domain will be same unless
6778 * there are asymmetries in the topology. If there are asymmetries, group
6779 * having more cpu_power will pickup more load compared to the group having
6782 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6784 struct sched_domain *child;
6785 struct sched_group *group;
6789 WARN_ON(!sd || !sd->groups);
6791 if (cpu != group_first_cpu(sd->groups))
6796 sd->groups->cpu_power = 0;
6799 power = SCHED_LOAD_SCALE;
6800 weight = cpumask_weight(sched_domain_span(sd));
6802 * SMT siblings share the power of a single core.
6803 * Usually multiple threads get a better yield out of
6804 * that one core than a single thread would have,
6805 * reflect that in sd->smt_gain.
6807 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6808 power *= sd->smt_gain;
6810 power >>= SCHED_LOAD_SHIFT;
6812 sd->groups->cpu_power += power;
6817 * Add cpu_power of each child group to this groups cpu_power.
6819 group = child->groups;
6821 sd->groups->cpu_power += group->cpu_power;
6822 group = group->next;
6823 } while (group != child->groups);
6827 * Initializers for schedule domains
6828 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6831 #ifdef CONFIG_SCHED_DEBUG
6832 # define SD_INIT_NAME(sd, type) sd->name = #type
6834 # define SD_INIT_NAME(sd, type) do { } while (0)
6837 #define SD_INIT(sd, type) sd_init_##type(sd)
6839 #define SD_INIT_FUNC(type) \
6840 static noinline void sd_init_##type(struct sched_domain *sd) \
6842 memset(sd, 0, sizeof(*sd)); \
6843 *sd = SD_##type##_INIT; \
6844 sd->level = SD_LV_##type; \
6845 SD_INIT_NAME(sd, type); \
6850 SD_INIT_FUNC(ALLNODES)
6853 #ifdef CONFIG_SCHED_SMT
6854 SD_INIT_FUNC(SIBLING)
6856 #ifdef CONFIG_SCHED_MC
6860 static int default_relax_domain_level = -1;
6862 static int __init setup_relax_domain_level(char *str)
6866 val = simple_strtoul(str, NULL, 0);
6867 if (val < SD_LV_MAX)
6868 default_relax_domain_level = val;
6872 __setup("relax_domain_level=", setup_relax_domain_level);
6874 static void set_domain_attribute(struct sched_domain *sd,
6875 struct sched_domain_attr *attr)
6879 if (!attr || attr->relax_domain_level < 0) {
6880 if (default_relax_domain_level < 0)
6883 request = default_relax_domain_level;
6885 request = attr->relax_domain_level;
6886 if (request < sd->level) {
6887 /* turn off idle balance on this domain */
6888 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6890 /* turn on idle balance on this domain */
6891 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6895 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6896 const struct cpumask *cpu_map)
6899 case sa_sched_groups:
6900 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6901 d->sched_group_nodes = NULL;
6903 free_rootdomain(d->rd); /* fall through */
6905 free_cpumask_var(d->tmpmask); /* fall through */
6906 case sa_send_covered:
6907 free_cpumask_var(d->send_covered); /* fall through */
6908 case sa_this_core_map:
6909 free_cpumask_var(d->this_core_map); /* fall through */
6910 case sa_this_sibling_map:
6911 free_cpumask_var(d->this_sibling_map); /* fall through */
6913 free_cpumask_var(d->nodemask); /* fall through */
6914 case sa_sched_group_nodes:
6916 kfree(d->sched_group_nodes); /* fall through */
6918 free_cpumask_var(d->notcovered); /* fall through */
6920 free_cpumask_var(d->covered); /* fall through */
6922 free_cpumask_var(d->domainspan); /* fall through */
6929 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6930 const struct cpumask *cpu_map)
6933 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6935 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6936 return sa_domainspan;
6937 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6939 /* Allocate the per-node list of sched groups */
6940 d->sched_group_nodes = kcalloc(nr_node_ids,
6941 sizeof(struct sched_group *), GFP_KERNEL);
6942 if (!d->sched_group_nodes) {
6943 printk(KERN_WARNING "Can not alloc sched group node list\n");
6944 return sa_notcovered;
6946 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6948 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6949 return sa_sched_group_nodes;
6950 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6952 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6953 return sa_this_sibling_map;
6954 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6955 return sa_this_core_map;
6956 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6957 return sa_send_covered;
6958 d->rd = alloc_rootdomain();
6960 printk(KERN_WARNING "Cannot alloc root domain\n");
6963 return sa_rootdomain;
6966 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6967 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6969 struct sched_domain *sd = NULL;
6971 struct sched_domain *parent;
6974 if (cpumask_weight(cpu_map) >
6975 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6976 sd = &per_cpu(allnodes_domains, i).sd;
6977 SD_INIT(sd, ALLNODES);
6978 set_domain_attribute(sd, attr);
6979 cpumask_copy(sched_domain_span(sd), cpu_map);
6980 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6985 sd = &per_cpu(node_domains, i).sd;
6987 set_domain_attribute(sd, attr);
6988 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6989 sd->parent = parent;
6992 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6997 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6998 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6999 struct sched_domain *parent, int i)
7001 struct sched_domain *sd;
7002 sd = &per_cpu(phys_domains, i).sd;
7004 set_domain_attribute(sd, attr);
7005 cpumask_copy(sched_domain_span(sd), d->nodemask);
7006 sd->parent = parent;
7009 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7013 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7014 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7015 struct sched_domain *parent, int i)
7017 struct sched_domain *sd = parent;
7018 #ifdef CONFIG_SCHED_MC
7019 sd = &per_cpu(core_domains, i).sd;
7021 set_domain_attribute(sd, attr);
7022 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7023 sd->parent = parent;
7025 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7030 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7031 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7032 struct sched_domain *parent, int i)
7034 struct sched_domain *sd = parent;
7035 #ifdef CONFIG_SCHED_SMT
7036 sd = &per_cpu(cpu_domains, i).sd;
7037 SD_INIT(sd, SIBLING);
7038 set_domain_attribute(sd, attr);
7039 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7040 sd->parent = parent;
7042 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7047 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7048 const struct cpumask *cpu_map, int cpu)
7051 #ifdef CONFIG_SCHED_SMT
7052 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7053 cpumask_and(d->this_sibling_map, cpu_map,
7054 topology_thread_cpumask(cpu));
7055 if (cpu == cpumask_first(d->this_sibling_map))
7056 init_sched_build_groups(d->this_sibling_map, cpu_map,
7058 d->send_covered, d->tmpmask);
7061 #ifdef CONFIG_SCHED_MC
7062 case SD_LV_MC: /* set up multi-core groups */
7063 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7064 if (cpu == cpumask_first(d->this_core_map))
7065 init_sched_build_groups(d->this_core_map, cpu_map,
7067 d->send_covered, d->tmpmask);
7070 case SD_LV_CPU: /* set up physical groups */
7071 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7072 if (!cpumask_empty(d->nodemask))
7073 init_sched_build_groups(d->nodemask, cpu_map,
7075 d->send_covered, d->tmpmask);
7078 case SD_LV_ALLNODES:
7079 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7080 d->send_covered, d->tmpmask);
7089 * Build sched domains for a given set of cpus and attach the sched domains
7090 * to the individual cpus
7092 static int __build_sched_domains(const struct cpumask *cpu_map,
7093 struct sched_domain_attr *attr)
7095 enum s_alloc alloc_state = sa_none;
7097 struct sched_domain *sd;
7103 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7104 if (alloc_state != sa_rootdomain)
7106 alloc_state = sa_sched_groups;
7109 * Set up domains for cpus specified by the cpu_map.
7111 for_each_cpu(i, cpu_map) {
7112 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7115 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7116 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7117 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7118 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7121 for_each_cpu(i, cpu_map) {
7122 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7123 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7126 /* Set up physical groups */
7127 for (i = 0; i < nr_node_ids; i++)
7128 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7131 /* Set up node groups */
7133 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7135 for (i = 0; i < nr_node_ids; i++)
7136 if (build_numa_sched_groups(&d, cpu_map, i))
7140 /* Calculate CPU power for physical packages and nodes */
7141 #ifdef CONFIG_SCHED_SMT
7142 for_each_cpu(i, cpu_map) {
7143 sd = &per_cpu(cpu_domains, i).sd;
7144 init_sched_groups_power(i, sd);
7147 #ifdef CONFIG_SCHED_MC
7148 for_each_cpu(i, cpu_map) {
7149 sd = &per_cpu(core_domains, i).sd;
7150 init_sched_groups_power(i, sd);
7154 for_each_cpu(i, cpu_map) {
7155 sd = &per_cpu(phys_domains, i).sd;
7156 init_sched_groups_power(i, sd);
7160 for (i = 0; i < nr_node_ids; i++)
7161 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7163 if (d.sd_allnodes) {
7164 struct sched_group *sg;
7166 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7168 init_numa_sched_groups_power(sg);
7172 /* Attach the domains */
7173 for_each_cpu(i, cpu_map) {
7174 #ifdef CONFIG_SCHED_SMT
7175 sd = &per_cpu(cpu_domains, i).sd;
7176 #elif defined(CONFIG_SCHED_MC)
7177 sd = &per_cpu(core_domains, i).sd;
7179 sd = &per_cpu(phys_domains, i).sd;
7181 cpu_attach_domain(sd, d.rd, i);
7184 d.sched_group_nodes = NULL; /* don't free this we still need it */
7185 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7189 __free_domain_allocs(&d, alloc_state, cpu_map);
7193 static int build_sched_domains(const struct cpumask *cpu_map)
7195 return __build_sched_domains(cpu_map, NULL);
7198 static cpumask_var_t *doms_cur; /* current sched domains */
7199 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7200 static struct sched_domain_attr *dattr_cur;
7201 /* attribues of custom domains in 'doms_cur' */
7204 * Special case: If a kmalloc of a doms_cur partition (array of
7205 * cpumask) fails, then fallback to a single sched domain,
7206 * as determined by the single cpumask fallback_doms.
7208 static cpumask_var_t fallback_doms;
7211 * arch_update_cpu_topology lets virtualized architectures update the
7212 * cpu core maps. It is supposed to return 1 if the topology changed
7213 * or 0 if it stayed the same.
7215 int __attribute__((weak)) arch_update_cpu_topology(void)
7220 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7223 cpumask_var_t *doms;
7225 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7228 for (i = 0; i < ndoms; i++) {
7229 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7230 free_sched_domains(doms, i);
7237 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7240 for (i = 0; i < ndoms; i++)
7241 free_cpumask_var(doms[i]);
7246 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7247 * For now this just excludes isolated cpus, but could be used to
7248 * exclude other special cases in the future.
7250 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7254 arch_update_cpu_topology();
7256 doms_cur = alloc_sched_domains(ndoms_cur);
7258 doms_cur = &fallback_doms;
7259 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7261 err = build_sched_domains(doms_cur[0]);
7262 register_sched_domain_sysctl();
7267 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7268 struct cpumask *tmpmask)
7270 free_sched_groups(cpu_map, tmpmask);
7274 * Detach sched domains from a group of cpus specified in cpu_map
7275 * These cpus will now be attached to the NULL domain
7277 static void detach_destroy_domains(const struct cpumask *cpu_map)
7279 /* Save because hotplug lock held. */
7280 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7283 for_each_cpu(i, cpu_map)
7284 cpu_attach_domain(NULL, &def_root_domain, i);
7285 synchronize_sched();
7286 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7289 /* handle null as "default" */
7290 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7291 struct sched_domain_attr *new, int idx_new)
7293 struct sched_domain_attr tmp;
7300 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7301 new ? (new + idx_new) : &tmp,
7302 sizeof(struct sched_domain_attr));
7306 * Partition sched domains as specified by the 'ndoms_new'
7307 * cpumasks in the array doms_new[] of cpumasks. This compares
7308 * doms_new[] to the current sched domain partitioning, doms_cur[].
7309 * It destroys each deleted domain and builds each new domain.
7311 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7312 * The masks don't intersect (don't overlap.) We should setup one
7313 * sched domain for each mask. CPUs not in any of the cpumasks will
7314 * not be load balanced. If the same cpumask appears both in the
7315 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7318 * The passed in 'doms_new' should be allocated using
7319 * alloc_sched_domains. This routine takes ownership of it and will
7320 * free_sched_domains it when done with it. If the caller failed the
7321 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7322 * and partition_sched_domains() will fallback to the single partition
7323 * 'fallback_doms', it also forces the domains to be rebuilt.
7325 * If doms_new == NULL it will be replaced with cpu_online_mask.
7326 * ndoms_new == 0 is a special case for destroying existing domains,
7327 * and it will not create the default domain.
7329 * Call with hotplug lock held
7331 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7332 struct sched_domain_attr *dattr_new)
7337 mutex_lock(&sched_domains_mutex);
7339 /* always unregister in case we don't destroy any domains */
7340 unregister_sched_domain_sysctl();
7342 /* Let architecture update cpu core mappings. */
7343 new_topology = arch_update_cpu_topology();
7345 n = doms_new ? ndoms_new : 0;
7347 /* Destroy deleted domains */
7348 for (i = 0; i < ndoms_cur; i++) {
7349 for (j = 0; j < n && !new_topology; j++) {
7350 if (cpumask_equal(doms_cur[i], doms_new[j])
7351 && dattrs_equal(dattr_cur, i, dattr_new, j))
7354 /* no match - a current sched domain not in new doms_new[] */
7355 detach_destroy_domains(doms_cur[i]);
7360 if (doms_new == NULL) {
7362 doms_new = &fallback_doms;
7363 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7364 WARN_ON_ONCE(dattr_new);
7367 /* Build new domains */
7368 for (i = 0; i < ndoms_new; i++) {
7369 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7370 if (cpumask_equal(doms_new[i], doms_cur[j])
7371 && dattrs_equal(dattr_new, i, dattr_cur, j))
7374 /* no match - add a new doms_new */
7375 __build_sched_domains(doms_new[i],
7376 dattr_new ? dattr_new + i : NULL);
7381 /* Remember the new sched domains */
7382 if (doms_cur != &fallback_doms)
7383 free_sched_domains(doms_cur, ndoms_cur);
7384 kfree(dattr_cur); /* kfree(NULL) is safe */
7385 doms_cur = doms_new;
7386 dattr_cur = dattr_new;
7387 ndoms_cur = ndoms_new;
7389 register_sched_domain_sysctl();
7391 mutex_unlock(&sched_domains_mutex);
7394 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7395 static void arch_reinit_sched_domains(void)
7399 /* Destroy domains first to force the rebuild */
7400 partition_sched_domains(0, NULL, NULL);
7402 rebuild_sched_domains();
7406 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7408 unsigned int level = 0;
7410 if (sscanf(buf, "%u", &level) != 1)
7414 * level is always be positive so don't check for
7415 * level < POWERSAVINGS_BALANCE_NONE which is 0
7416 * What happens on 0 or 1 byte write,
7417 * need to check for count as well?
7420 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7424 sched_smt_power_savings = level;
7426 sched_mc_power_savings = level;
7428 arch_reinit_sched_domains();
7433 #ifdef CONFIG_SCHED_MC
7434 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7435 struct sysdev_class_attribute *attr,
7438 return sprintf(page, "%u\n", sched_mc_power_savings);
7440 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7441 struct sysdev_class_attribute *attr,
7442 const char *buf, size_t count)
7444 return sched_power_savings_store(buf, count, 0);
7446 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7447 sched_mc_power_savings_show,
7448 sched_mc_power_savings_store);
7451 #ifdef CONFIG_SCHED_SMT
7452 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7453 struct sysdev_class_attribute *attr,
7456 return sprintf(page, "%u\n", sched_smt_power_savings);
7458 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7459 struct sysdev_class_attribute *attr,
7460 const char *buf, size_t count)
7462 return sched_power_savings_store(buf, count, 1);
7464 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7465 sched_smt_power_savings_show,
7466 sched_smt_power_savings_store);
7469 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7473 #ifdef CONFIG_SCHED_SMT
7475 err = sysfs_create_file(&cls->kset.kobj,
7476 &attr_sched_smt_power_savings.attr);
7478 #ifdef CONFIG_SCHED_MC
7479 if (!err && mc_capable())
7480 err = sysfs_create_file(&cls->kset.kobj,
7481 &attr_sched_mc_power_savings.attr);
7485 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7488 * Update cpusets according to cpu_active mask. If cpusets are
7489 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7490 * around partition_sched_domains().
7492 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7495 switch (action & ~CPU_TASKS_FROZEN) {
7497 case CPU_DOWN_FAILED:
7498 cpuset_update_active_cpus();
7505 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7508 switch (action & ~CPU_TASKS_FROZEN) {
7509 case CPU_DOWN_PREPARE:
7510 cpuset_update_active_cpus();
7517 static int update_runtime(struct notifier_block *nfb,
7518 unsigned long action, void *hcpu)
7520 int cpu = (int)(long)hcpu;
7523 case CPU_DOWN_PREPARE:
7524 case CPU_DOWN_PREPARE_FROZEN:
7525 disable_runtime(cpu_rq(cpu));
7528 case CPU_DOWN_FAILED:
7529 case CPU_DOWN_FAILED_FROZEN:
7531 case CPU_ONLINE_FROZEN:
7532 enable_runtime(cpu_rq(cpu));
7540 void __init sched_init_smp(void)
7542 cpumask_var_t non_isolated_cpus;
7544 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7545 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7547 #if defined(CONFIG_NUMA)
7548 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7550 BUG_ON(sched_group_nodes_bycpu == NULL);
7553 mutex_lock(&sched_domains_mutex);
7554 arch_init_sched_domains(cpu_active_mask);
7555 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7556 if (cpumask_empty(non_isolated_cpus))
7557 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7558 mutex_unlock(&sched_domains_mutex);
7561 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7562 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7564 /* RT runtime code needs to handle some hotplug events */
7565 hotcpu_notifier(update_runtime, 0);
7569 /* Move init over to a non-isolated CPU */
7570 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7572 sched_init_granularity();
7573 free_cpumask_var(non_isolated_cpus);
7575 init_sched_rt_class();
7578 void __init sched_init_smp(void)
7580 sched_init_granularity();
7582 #endif /* CONFIG_SMP */
7584 const_debug unsigned int sysctl_timer_migration = 1;
7586 int in_sched_functions(unsigned long addr)
7588 return in_lock_functions(addr) ||
7589 (addr >= (unsigned long)__sched_text_start
7590 && addr < (unsigned long)__sched_text_end);
7593 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7595 cfs_rq->tasks_timeline = RB_ROOT;
7596 INIT_LIST_HEAD(&cfs_rq->tasks);
7597 #ifdef CONFIG_FAIR_GROUP_SCHED
7600 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7603 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7605 struct rt_prio_array *array;
7608 array = &rt_rq->active;
7609 for (i = 0; i < MAX_RT_PRIO; i++) {
7610 INIT_LIST_HEAD(array->queue + i);
7611 __clear_bit(i, array->bitmap);
7613 /* delimiter for bitsearch: */
7614 __set_bit(MAX_RT_PRIO, array->bitmap);
7616 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7617 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7619 rt_rq->highest_prio.next = MAX_RT_PRIO;
7623 rt_rq->rt_nr_migratory = 0;
7624 rt_rq->overloaded = 0;
7625 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7629 rt_rq->rt_throttled = 0;
7630 rt_rq->rt_runtime = 0;
7631 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7633 #ifdef CONFIG_RT_GROUP_SCHED
7634 rt_rq->rt_nr_boosted = 0;
7639 #ifdef CONFIG_FAIR_GROUP_SCHED
7640 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7641 struct sched_entity *se, int cpu, int add,
7642 struct sched_entity *parent)
7644 struct rq *rq = cpu_rq(cpu);
7645 tg->cfs_rq[cpu] = cfs_rq;
7646 init_cfs_rq(cfs_rq, rq);
7649 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7652 /* se could be NULL for init_task_group */
7657 se->cfs_rq = &rq->cfs;
7659 se->cfs_rq = parent->my_q;
7662 se->load.weight = tg->shares;
7663 se->load.inv_weight = 0;
7664 se->parent = parent;
7668 #ifdef CONFIG_RT_GROUP_SCHED
7669 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7670 struct sched_rt_entity *rt_se, int cpu, int add,
7671 struct sched_rt_entity *parent)
7673 struct rq *rq = cpu_rq(cpu);
7675 tg->rt_rq[cpu] = rt_rq;
7676 init_rt_rq(rt_rq, rq);
7678 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7680 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7682 tg->rt_se[cpu] = rt_se;
7687 rt_se->rt_rq = &rq->rt;
7689 rt_se->rt_rq = parent->my_q;
7691 rt_se->my_q = rt_rq;
7692 rt_se->parent = parent;
7693 INIT_LIST_HEAD(&rt_se->run_list);
7697 void __init sched_init(void)
7700 unsigned long alloc_size = 0, ptr;
7702 #ifdef CONFIG_FAIR_GROUP_SCHED
7703 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7705 #ifdef CONFIG_RT_GROUP_SCHED
7706 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7708 #ifdef CONFIG_CPUMASK_OFFSTACK
7709 alloc_size += num_possible_cpus() * cpumask_size();
7712 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7714 #ifdef CONFIG_FAIR_GROUP_SCHED
7715 init_task_group.se = (struct sched_entity **)ptr;
7716 ptr += nr_cpu_ids * sizeof(void **);
7718 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7719 ptr += nr_cpu_ids * sizeof(void **);
7721 #endif /* CONFIG_FAIR_GROUP_SCHED */
7722 #ifdef CONFIG_RT_GROUP_SCHED
7723 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7724 ptr += nr_cpu_ids * sizeof(void **);
7726 init_task_group.rt_rq = (struct rt_rq **)ptr;
7727 ptr += nr_cpu_ids * sizeof(void **);
7729 #endif /* CONFIG_RT_GROUP_SCHED */
7730 #ifdef CONFIG_CPUMASK_OFFSTACK
7731 for_each_possible_cpu(i) {
7732 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7733 ptr += cpumask_size();
7735 #endif /* CONFIG_CPUMASK_OFFSTACK */
7739 init_defrootdomain();
7742 init_rt_bandwidth(&def_rt_bandwidth,
7743 global_rt_period(), global_rt_runtime());
7745 #ifdef CONFIG_RT_GROUP_SCHED
7746 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7747 global_rt_period(), global_rt_runtime());
7748 #endif /* CONFIG_RT_GROUP_SCHED */
7750 #ifdef CONFIG_CGROUP_SCHED
7751 list_add(&init_task_group.list, &task_groups);
7752 INIT_LIST_HEAD(&init_task_group.children);
7754 #endif /* CONFIG_CGROUP_SCHED */
7756 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7757 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7758 __alignof__(unsigned long));
7760 for_each_possible_cpu(i) {
7764 raw_spin_lock_init(&rq->lock);
7766 rq->calc_load_active = 0;
7767 rq->calc_load_update = jiffies + LOAD_FREQ;
7768 init_cfs_rq(&rq->cfs, rq);
7769 init_rt_rq(&rq->rt, rq);
7770 #ifdef CONFIG_FAIR_GROUP_SCHED
7771 init_task_group.shares = init_task_group_load;
7772 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7773 #ifdef CONFIG_CGROUP_SCHED
7775 * How much cpu bandwidth does init_task_group get?
7777 * In case of task-groups formed thr' the cgroup filesystem, it
7778 * gets 100% of the cpu resources in the system. This overall
7779 * system cpu resource is divided among the tasks of
7780 * init_task_group and its child task-groups in a fair manner,
7781 * based on each entity's (task or task-group's) weight
7782 * (se->load.weight).
7784 * In other words, if init_task_group has 10 tasks of weight
7785 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7786 * then A0's share of the cpu resource is:
7788 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7790 * We achieve this by letting init_task_group's tasks sit
7791 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7793 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7795 #endif /* CONFIG_FAIR_GROUP_SCHED */
7797 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7798 #ifdef CONFIG_RT_GROUP_SCHED
7799 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7800 #ifdef CONFIG_CGROUP_SCHED
7801 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7805 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7806 rq->cpu_load[j] = 0;
7808 rq->last_load_update_tick = jiffies;
7813 rq->cpu_power = SCHED_LOAD_SCALE;
7814 rq->post_schedule = 0;
7815 rq->active_balance = 0;
7816 rq->next_balance = jiffies;
7821 rq->avg_idle = 2*sysctl_sched_migration_cost;
7822 rq_attach_root(rq, &def_root_domain);
7824 rq->nohz_balance_kick = 0;
7825 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7829 atomic_set(&rq->nr_iowait, 0);
7832 set_load_weight(&init_task);
7834 #ifdef CONFIG_PREEMPT_NOTIFIERS
7835 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7839 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7842 #ifdef CONFIG_RT_MUTEXES
7843 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7847 * The boot idle thread does lazy MMU switching as well:
7849 atomic_inc(&init_mm.mm_count);
7850 enter_lazy_tlb(&init_mm, current);
7853 * Make us the idle thread. Technically, schedule() should not be
7854 * called from this thread, however somewhere below it might be,
7855 * but because we are the idle thread, we just pick up running again
7856 * when this runqueue becomes "idle".
7858 init_idle(current, smp_processor_id());
7860 calc_load_update = jiffies + LOAD_FREQ;
7863 * During early bootup we pretend to be a normal task:
7865 current->sched_class = &fair_sched_class;
7867 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7868 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7871 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7872 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
7873 atomic_set(&nohz.load_balancer, nr_cpu_ids);
7874 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
7875 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
7877 /* May be allocated at isolcpus cmdline parse time */
7878 if (cpu_isolated_map == NULL)
7879 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7884 scheduler_running = 1;
7887 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7888 static inline int preempt_count_equals(int preempt_offset)
7890 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7892 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7895 static int __might_sleep_init_called;
7896 int __init __might_sleep_init(void)
7898 __might_sleep_init_called = 1;
7901 early_initcall(__might_sleep_init);
7903 void __might_sleep(const char *file, int line, int preempt_offset)
7906 static unsigned long prev_jiffy; /* ratelimiting */
7908 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7911 if (system_state != SYSTEM_RUNNING &&
7912 (!__might_sleep_init_called || system_state != SYSTEM_BOOTING))
7914 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7916 prev_jiffy = jiffies;
7919 "BUG: sleeping function called from invalid context at %s:%d\n",
7922 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7923 in_atomic(), irqs_disabled(),
7924 current->pid, current->comm);
7926 debug_show_held_locks(current);
7927 if (irqs_disabled())
7928 print_irqtrace_events(current);
7932 EXPORT_SYMBOL(__might_sleep);
7935 #ifdef CONFIG_MAGIC_SYSRQ
7936 static void normalize_task(struct rq *rq, struct task_struct *p)
7940 on_rq = p->se.on_rq;
7942 deactivate_task(rq, p, 0);
7943 __setscheduler(rq, p, SCHED_NORMAL, 0);
7945 activate_task(rq, p, 0);
7946 resched_task(rq->curr);
7950 void normalize_rt_tasks(void)
7952 struct task_struct *g, *p;
7953 unsigned long flags;
7956 read_lock_irqsave(&tasklist_lock, flags);
7957 do_each_thread(g, p) {
7959 * Only normalize user tasks:
7964 p->se.exec_start = 0;
7965 #ifdef CONFIG_SCHEDSTATS
7966 p->se.statistics.wait_start = 0;
7967 p->se.statistics.sleep_start = 0;
7968 p->se.statistics.block_start = 0;
7973 * Renice negative nice level userspace
7976 if (TASK_NICE(p) < 0 && p->mm)
7977 set_user_nice(p, 0);
7981 raw_spin_lock(&p->pi_lock);
7982 rq = __task_rq_lock(p);
7984 normalize_task(rq, p);
7986 __task_rq_unlock(rq);
7987 raw_spin_unlock(&p->pi_lock);
7988 } while_each_thread(g, p);
7990 read_unlock_irqrestore(&tasklist_lock, flags);
7993 #endif /* CONFIG_MAGIC_SYSRQ */
7995 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7997 * These functions are only useful for the IA64 MCA handling, or kdb.
7999 * They can only be called when the whole system has been
8000 * stopped - every CPU needs to be quiescent, and no scheduling
8001 * activity can take place. Using them for anything else would
8002 * be a serious bug, and as a result, they aren't even visible
8003 * under any other configuration.
8007 * curr_task - return the current task for a given cpu.
8008 * @cpu: the processor in question.
8010 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8012 struct task_struct *curr_task(int cpu)
8014 return cpu_curr(cpu);
8017 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8021 * set_curr_task - set the current task for a given cpu.
8022 * @cpu: the processor in question.
8023 * @p: the task pointer to set.
8025 * Description: This function must only be used when non-maskable interrupts
8026 * are serviced on a separate stack. It allows the architecture to switch the
8027 * notion of the current task on a cpu in a non-blocking manner. This function
8028 * must be called with all CPU's synchronized, and interrupts disabled, the
8029 * and caller must save the original value of the current task (see
8030 * curr_task() above) and restore that value before reenabling interrupts and
8031 * re-starting the system.
8033 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8035 void set_curr_task(int cpu, struct task_struct *p)
8042 #ifdef CONFIG_FAIR_GROUP_SCHED
8043 static void free_fair_sched_group(struct task_group *tg)
8047 for_each_possible_cpu(i) {
8049 kfree(tg->cfs_rq[i]);
8059 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8061 struct cfs_rq *cfs_rq;
8062 struct sched_entity *se;
8066 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8069 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8073 tg->shares = NICE_0_LOAD;
8075 for_each_possible_cpu(i) {
8078 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8079 GFP_KERNEL, cpu_to_node(i));
8083 se = kzalloc_node(sizeof(struct sched_entity),
8084 GFP_KERNEL, cpu_to_node(i));
8088 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8099 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8101 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8102 &cpu_rq(cpu)->leaf_cfs_rq_list);
8105 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8107 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8109 #else /* !CONFG_FAIR_GROUP_SCHED */
8110 static inline void free_fair_sched_group(struct task_group *tg)
8115 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8120 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8124 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8127 #endif /* CONFIG_FAIR_GROUP_SCHED */
8129 #ifdef CONFIG_RT_GROUP_SCHED
8130 static void free_rt_sched_group(struct task_group *tg)
8134 destroy_rt_bandwidth(&tg->rt_bandwidth);
8136 for_each_possible_cpu(i) {
8138 kfree(tg->rt_rq[i]);
8140 kfree(tg->rt_se[i]);
8148 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8150 struct rt_rq *rt_rq;
8151 struct sched_rt_entity *rt_se;
8155 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8158 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8162 init_rt_bandwidth(&tg->rt_bandwidth,
8163 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8165 for_each_possible_cpu(i) {
8168 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8169 GFP_KERNEL, cpu_to_node(i));
8173 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8174 GFP_KERNEL, cpu_to_node(i));
8178 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8189 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8191 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8192 &cpu_rq(cpu)->leaf_rt_rq_list);
8195 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8197 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8199 #else /* !CONFIG_RT_GROUP_SCHED */
8200 static inline void free_rt_sched_group(struct task_group *tg)
8205 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8210 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8214 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8217 #endif /* CONFIG_RT_GROUP_SCHED */
8219 #ifdef CONFIG_CGROUP_SCHED
8220 static void free_sched_group(struct task_group *tg)
8222 free_fair_sched_group(tg);
8223 free_rt_sched_group(tg);
8227 /* allocate runqueue etc for a new task group */
8228 struct task_group *sched_create_group(struct task_group *parent)
8230 struct task_group *tg;
8231 unsigned long flags;
8234 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8236 return ERR_PTR(-ENOMEM);
8238 if (!alloc_fair_sched_group(tg, parent))
8241 if (!alloc_rt_sched_group(tg, parent))
8244 spin_lock_irqsave(&task_group_lock, flags);
8245 for_each_possible_cpu(i) {
8246 register_fair_sched_group(tg, i);
8247 register_rt_sched_group(tg, i);
8249 list_add_rcu(&tg->list, &task_groups);
8251 WARN_ON(!parent); /* root should already exist */
8253 tg->parent = parent;
8254 INIT_LIST_HEAD(&tg->children);
8255 list_add_rcu(&tg->siblings, &parent->children);
8256 spin_unlock_irqrestore(&task_group_lock, flags);
8261 free_sched_group(tg);
8262 return ERR_PTR(-ENOMEM);
8265 /* rcu callback to free various structures associated with a task group */
8266 static void free_sched_group_rcu(struct rcu_head *rhp)
8268 /* now it should be safe to free those cfs_rqs */
8269 free_sched_group(container_of(rhp, struct task_group, rcu));
8272 /* Destroy runqueue etc associated with a task group */
8273 void sched_destroy_group(struct task_group *tg)
8275 unsigned long flags;
8278 spin_lock_irqsave(&task_group_lock, flags);
8279 for_each_possible_cpu(i) {
8280 unregister_fair_sched_group(tg, i);
8281 unregister_rt_sched_group(tg, i);
8283 list_del_rcu(&tg->list);
8284 list_del_rcu(&tg->siblings);
8285 spin_unlock_irqrestore(&task_group_lock, flags);
8287 /* wait for possible concurrent references to cfs_rqs complete */
8288 call_rcu(&tg->rcu, free_sched_group_rcu);
8291 /* change task's runqueue when it moves between groups.
8292 * The caller of this function should have put the task in its new group
8293 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8294 * reflect its new group.
8296 void sched_move_task(struct task_struct *tsk)
8299 unsigned long flags;
8302 rq = task_rq_lock(tsk, &flags);
8304 running = task_current(rq, tsk);
8305 on_rq = tsk->se.on_rq;
8308 dequeue_task(rq, tsk, 0);
8309 if (unlikely(running))
8310 tsk->sched_class->put_prev_task(rq, tsk);
8312 #ifdef CONFIG_FAIR_GROUP_SCHED
8313 if (tsk->sched_class->prep_move_group)
8314 tsk->sched_class->prep_move_group(tsk, on_rq);
8317 set_task_rq(tsk, task_cpu(tsk));
8319 #ifdef CONFIG_FAIR_GROUP_SCHED
8320 if (tsk->sched_class->moved_group)
8321 tsk->sched_class->moved_group(tsk, on_rq);
8324 if (unlikely(running))
8325 tsk->sched_class->set_curr_task(rq);
8327 enqueue_task(rq, tsk, 0);
8329 task_rq_unlock(rq, &flags);
8331 #endif /* CONFIG_CGROUP_SCHED */
8333 #ifdef CONFIG_FAIR_GROUP_SCHED
8334 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8336 struct cfs_rq *cfs_rq = se->cfs_rq;
8341 dequeue_entity(cfs_rq, se, 0);
8343 se->load.weight = shares;
8344 se->load.inv_weight = 0;
8347 enqueue_entity(cfs_rq, se, 0);
8350 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8352 struct cfs_rq *cfs_rq = se->cfs_rq;
8353 struct rq *rq = cfs_rq->rq;
8354 unsigned long flags;
8356 raw_spin_lock_irqsave(&rq->lock, flags);
8357 __set_se_shares(se, shares);
8358 raw_spin_unlock_irqrestore(&rq->lock, flags);
8361 static DEFINE_MUTEX(shares_mutex);
8363 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8366 unsigned long flags;
8369 * We can't change the weight of the root cgroup.
8374 if (shares < MIN_SHARES)
8375 shares = MIN_SHARES;
8376 else if (shares > MAX_SHARES)
8377 shares = MAX_SHARES;
8379 mutex_lock(&shares_mutex);
8380 if (tg->shares == shares)
8383 spin_lock_irqsave(&task_group_lock, flags);
8384 for_each_possible_cpu(i)
8385 unregister_fair_sched_group(tg, i);
8386 list_del_rcu(&tg->siblings);
8387 spin_unlock_irqrestore(&task_group_lock, flags);
8389 /* wait for any ongoing reference to this group to finish */
8390 synchronize_sched();
8393 * Now we are free to modify the group's share on each cpu
8394 * w/o tripping rebalance_share or load_balance_fair.
8396 tg->shares = shares;
8397 for_each_possible_cpu(i) {
8401 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8402 set_se_shares(tg->se[i], shares);
8406 * Enable load balance activity on this group, by inserting it back on
8407 * each cpu's rq->leaf_cfs_rq_list.
8409 spin_lock_irqsave(&task_group_lock, flags);
8410 for_each_possible_cpu(i)
8411 register_fair_sched_group(tg, i);
8412 list_add_rcu(&tg->siblings, &tg->parent->children);
8413 spin_unlock_irqrestore(&task_group_lock, flags);
8415 mutex_unlock(&shares_mutex);
8419 unsigned long sched_group_shares(struct task_group *tg)
8425 #ifdef CONFIG_RT_GROUP_SCHED
8427 * Ensure that the real time constraints are schedulable.
8429 static DEFINE_MUTEX(rt_constraints_mutex);
8431 static unsigned long to_ratio(u64 period, u64 runtime)
8433 if (runtime == RUNTIME_INF)
8436 return div64_u64(runtime << 20, period);
8439 /* Must be called with tasklist_lock held */
8440 static inline int tg_has_rt_tasks(struct task_group *tg)
8442 struct task_struct *g, *p;
8444 do_each_thread(g, p) {
8445 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8447 } while_each_thread(g, p);
8452 struct rt_schedulable_data {
8453 struct task_group *tg;
8458 static int tg_schedulable(struct task_group *tg, void *data)
8460 struct rt_schedulable_data *d = data;
8461 struct task_group *child;
8462 unsigned long total, sum = 0;
8463 u64 period, runtime;
8465 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8466 runtime = tg->rt_bandwidth.rt_runtime;
8469 period = d->rt_period;
8470 runtime = d->rt_runtime;
8474 * Cannot have more runtime than the period.
8476 if (runtime > period && runtime != RUNTIME_INF)
8480 * Ensure we don't starve existing RT tasks.
8482 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8485 total = to_ratio(period, runtime);
8488 * Nobody can have more than the global setting allows.
8490 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8494 * The sum of our children's runtime should not exceed our own.
8496 list_for_each_entry_rcu(child, &tg->children, siblings) {
8497 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8498 runtime = child->rt_bandwidth.rt_runtime;
8500 if (child == d->tg) {
8501 period = d->rt_period;
8502 runtime = d->rt_runtime;
8505 sum += to_ratio(period, runtime);
8514 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8516 struct rt_schedulable_data data = {
8518 .rt_period = period,
8519 .rt_runtime = runtime,
8522 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8525 static int tg_set_bandwidth(struct task_group *tg,
8526 u64 rt_period, u64 rt_runtime)
8530 mutex_lock(&rt_constraints_mutex);
8531 read_lock(&tasklist_lock);
8532 err = __rt_schedulable(tg, rt_period, rt_runtime);
8536 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8537 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8538 tg->rt_bandwidth.rt_runtime = rt_runtime;
8540 for_each_possible_cpu(i) {
8541 struct rt_rq *rt_rq = tg->rt_rq[i];
8543 raw_spin_lock(&rt_rq->rt_runtime_lock);
8544 rt_rq->rt_runtime = rt_runtime;
8545 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8547 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8549 read_unlock(&tasklist_lock);
8550 mutex_unlock(&rt_constraints_mutex);
8555 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8557 u64 rt_runtime, rt_period;
8559 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8560 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8561 if (rt_runtime_us < 0)
8562 rt_runtime = RUNTIME_INF;
8564 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8567 long sched_group_rt_runtime(struct task_group *tg)
8571 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8574 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8575 do_div(rt_runtime_us, NSEC_PER_USEC);
8576 return rt_runtime_us;
8579 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8581 u64 rt_runtime, rt_period;
8583 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8584 rt_runtime = tg->rt_bandwidth.rt_runtime;
8589 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8592 long sched_group_rt_period(struct task_group *tg)
8596 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8597 do_div(rt_period_us, NSEC_PER_USEC);
8598 return rt_period_us;
8601 static int sched_rt_global_constraints(void)
8603 u64 runtime, period;
8606 if (sysctl_sched_rt_period <= 0)
8609 runtime = global_rt_runtime();
8610 period = global_rt_period();
8613 * Sanity check on the sysctl variables.
8615 if (runtime > period && runtime != RUNTIME_INF)
8618 mutex_lock(&rt_constraints_mutex);
8619 read_lock(&tasklist_lock);
8620 ret = __rt_schedulable(NULL, 0, 0);
8621 read_unlock(&tasklist_lock);
8622 mutex_unlock(&rt_constraints_mutex);
8627 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8629 /* Don't accept realtime tasks when there is no way for them to run */
8630 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8636 #else /* !CONFIG_RT_GROUP_SCHED */
8637 static int sched_rt_global_constraints(void)
8639 unsigned long flags;
8642 if (sysctl_sched_rt_period <= 0)
8646 * There's always some RT tasks in the root group
8647 * -- migration, kstopmachine etc..
8649 if (sysctl_sched_rt_runtime == 0)
8652 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8653 for_each_possible_cpu(i) {
8654 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8656 raw_spin_lock(&rt_rq->rt_runtime_lock);
8657 rt_rq->rt_runtime = global_rt_runtime();
8658 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8660 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8664 #endif /* CONFIG_RT_GROUP_SCHED */
8666 int sched_rt_handler(struct ctl_table *table, int write,
8667 void __user *buffer, size_t *lenp,
8671 int old_period, old_runtime;
8672 static DEFINE_MUTEX(mutex);
8675 old_period = sysctl_sched_rt_period;
8676 old_runtime = sysctl_sched_rt_runtime;
8678 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8680 if (!ret && write) {
8681 ret = sched_rt_global_constraints();
8683 sysctl_sched_rt_period = old_period;
8684 sysctl_sched_rt_runtime = old_runtime;
8686 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8687 def_rt_bandwidth.rt_period =
8688 ns_to_ktime(global_rt_period());
8691 mutex_unlock(&mutex);
8696 #ifdef CONFIG_CGROUP_SCHED
8698 /* return corresponding task_group object of a cgroup */
8699 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8701 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8702 struct task_group, css);
8705 static struct cgroup_subsys_state *
8706 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8708 struct task_group *tg, *parent;
8710 if (!cgrp->parent) {
8711 /* This is early initialization for the top cgroup */
8712 return &init_task_group.css;
8715 parent = cgroup_tg(cgrp->parent);
8716 tg = sched_create_group(parent);
8718 return ERR_PTR(-ENOMEM);
8724 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8726 struct task_group *tg = cgroup_tg(cgrp);
8728 sched_destroy_group(tg);
8732 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8734 if ((current != tsk) && (!capable(CAP_SYS_NICE))) {
8735 const struct cred *cred = current_cred(), *tcred;
8737 tcred = __task_cred(tsk);
8739 if (cred->euid != tcred->uid && cred->euid != tcred->suid)
8743 #ifdef CONFIG_RT_GROUP_SCHED
8744 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8747 /* We don't support RT-tasks being in separate groups */
8748 if (tsk->sched_class != &fair_sched_class)
8755 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8756 struct task_struct *tsk, bool threadgroup)
8758 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8762 struct task_struct *c;
8764 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8765 retval = cpu_cgroup_can_attach_task(cgrp, c);
8777 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8778 struct cgroup *old_cont, struct task_struct *tsk,
8781 sched_move_task(tsk);
8783 struct task_struct *c;
8785 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8792 #ifdef CONFIG_FAIR_GROUP_SCHED
8793 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8796 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8799 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8801 struct task_group *tg = cgroup_tg(cgrp);
8803 return (u64) tg->shares;
8805 #endif /* CONFIG_FAIR_GROUP_SCHED */
8807 #ifdef CONFIG_RT_GROUP_SCHED
8808 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8811 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8814 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8816 return sched_group_rt_runtime(cgroup_tg(cgrp));
8819 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8822 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8825 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8827 return sched_group_rt_period(cgroup_tg(cgrp));
8829 #endif /* CONFIG_RT_GROUP_SCHED */
8831 static struct cftype cpu_files[] = {
8832 #ifdef CONFIG_FAIR_GROUP_SCHED
8835 .read_u64 = cpu_shares_read_u64,
8836 .write_u64 = cpu_shares_write_u64,
8839 #ifdef CONFIG_RT_GROUP_SCHED
8841 .name = "rt_runtime_us",
8842 .read_s64 = cpu_rt_runtime_read,
8843 .write_s64 = cpu_rt_runtime_write,
8846 .name = "rt_period_us",
8847 .read_u64 = cpu_rt_period_read_uint,
8848 .write_u64 = cpu_rt_period_write_uint,
8853 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8855 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8858 struct cgroup_subsys cpu_cgroup_subsys = {
8860 .create = cpu_cgroup_create,
8861 .destroy = cpu_cgroup_destroy,
8862 .can_attach = cpu_cgroup_can_attach,
8863 .attach = cpu_cgroup_attach,
8864 .populate = cpu_cgroup_populate,
8865 .subsys_id = cpu_cgroup_subsys_id,
8869 #endif /* CONFIG_CGROUP_SCHED */
8871 #ifdef CONFIG_CGROUP_CPUACCT
8874 * CPU accounting code for task groups.
8876 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8877 * (balbir@in.ibm.com).
8880 /* track cpu usage of a group of tasks and its child groups */
8882 struct cgroup_subsys_state css;
8883 /* cpuusage holds pointer to a u64-type object on every cpu */
8884 u64 __percpu *cpuusage;
8885 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8886 struct cpuacct *parent;
8887 struct cpuacct_charge_calls *cpufreq_fn;
8891 static struct cpuacct *cpuacct_root;
8893 /* Default calls for cpufreq accounting */
8894 static struct cpuacct_charge_calls *cpuacct_cpufreq;
8895 int cpuacct_register_cpufreq(struct cpuacct_charge_calls *fn)
8897 cpuacct_cpufreq = fn;
8900 * Root node is created before platform can register callbacks,
8903 if (cpuacct_root && fn) {
8904 cpuacct_root->cpufreq_fn = fn;
8906 fn->init(&cpuacct_root->cpuacct_data);
8911 struct cgroup_subsys cpuacct_subsys;
8913 /* return cpu accounting group corresponding to this container */
8914 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8916 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8917 struct cpuacct, css);
8920 /* return cpu accounting group to which this task belongs */
8921 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8923 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8924 struct cpuacct, css);
8927 /* create a new cpu accounting group */
8928 static struct cgroup_subsys_state *cpuacct_create(
8929 struct cgroup_subsys *ss, struct cgroup *cgrp)
8931 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8937 ca->cpuusage = alloc_percpu(u64);
8941 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8942 if (percpu_counter_init(&ca->cpustat[i], 0))
8943 goto out_free_counters;
8945 ca->cpufreq_fn = cpuacct_cpufreq;
8947 /* If available, have platform code initalize cpu frequency table */
8948 if (ca->cpufreq_fn && ca->cpufreq_fn->init)
8949 ca->cpufreq_fn->init(&ca->cpuacct_data);
8952 ca->parent = cgroup_ca(cgrp->parent);
8960 percpu_counter_destroy(&ca->cpustat[i]);
8961 free_percpu(ca->cpuusage);
8965 return ERR_PTR(-ENOMEM);
8968 /* destroy an existing cpu accounting group */
8970 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8972 struct cpuacct *ca = cgroup_ca(cgrp);
8975 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8976 percpu_counter_destroy(&ca->cpustat[i]);
8977 free_percpu(ca->cpuusage);
8981 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8983 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8986 #ifndef CONFIG_64BIT
8988 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8990 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8992 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9000 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9002 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9004 #ifndef CONFIG_64BIT
9006 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9008 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
9010 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
9016 /* return total cpu usage (in nanoseconds) of a group */
9017 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9019 struct cpuacct *ca = cgroup_ca(cgrp);
9020 u64 totalcpuusage = 0;
9023 for_each_present_cpu(i)
9024 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9026 return totalcpuusage;
9029 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9032 struct cpuacct *ca = cgroup_ca(cgrp);
9041 for_each_present_cpu(i)
9042 cpuacct_cpuusage_write(ca, i, 0);
9048 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9051 struct cpuacct *ca = cgroup_ca(cgroup);
9055 for_each_present_cpu(i) {
9056 percpu = cpuacct_cpuusage_read(ca, i);
9057 seq_printf(m, "%llu ", (unsigned long long) percpu);
9059 seq_printf(m, "\n");
9063 static const char *cpuacct_stat_desc[] = {
9064 [CPUACCT_STAT_USER] = "user",
9065 [CPUACCT_STAT_SYSTEM] = "system",
9068 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9069 struct cgroup_map_cb *cb)
9071 struct cpuacct *ca = cgroup_ca(cgrp);
9074 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9075 s64 val = percpu_counter_read(&ca->cpustat[i]);
9076 val = cputime64_to_clock_t(val);
9077 cb->fill(cb, cpuacct_stat_desc[i], val);
9082 static int cpuacct_cpufreq_show(struct cgroup *cgrp, struct cftype *cft,
9083 struct cgroup_map_cb *cb)
9085 struct cpuacct *ca = cgroup_ca(cgrp);
9086 if (ca->cpufreq_fn && ca->cpufreq_fn->cpufreq_show)
9087 ca->cpufreq_fn->cpufreq_show(ca->cpuacct_data, cb);
9092 /* return total cpu power usage (milliWatt second) of a group */
9093 static u64 cpuacct_powerusage_read(struct cgroup *cgrp, struct cftype *cft)
9096 struct cpuacct *ca = cgroup_ca(cgrp);
9099 if (ca->cpufreq_fn && ca->cpufreq_fn->power_usage)
9100 for_each_present_cpu(i) {
9101 totalpower += ca->cpufreq_fn->power_usage(
9108 static struct cftype files[] = {
9111 .read_u64 = cpuusage_read,
9112 .write_u64 = cpuusage_write,
9115 .name = "usage_percpu",
9116 .read_seq_string = cpuacct_percpu_seq_read,
9120 .read_map = cpuacct_stats_show,
9124 .read_map = cpuacct_cpufreq_show,
9128 .read_u64 = cpuacct_powerusage_read
9132 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9134 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9138 * charge this task's execution time to its accounting group.
9140 * called with rq->lock held.
9142 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9147 if (unlikely(!cpuacct_subsys.active))
9150 cpu = task_cpu(tsk);
9156 for (; ca; ca = ca->parent) {
9157 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9158 *cpuusage += cputime;
9160 /* Call back into platform code to account for CPU speeds */
9161 if (ca->cpufreq_fn && ca->cpufreq_fn->charge)
9162 ca->cpufreq_fn->charge(ca->cpuacct_data, cputime, cpu);
9169 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9170 * in cputime_t units. As a result, cpuacct_update_stats calls
9171 * percpu_counter_add with values large enough to always overflow the
9172 * per cpu batch limit causing bad SMP scalability.
9174 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9175 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9176 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9179 #define CPUACCT_BATCH \
9180 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9182 #define CPUACCT_BATCH 0
9186 * Charge the system/user time to the task's accounting group.
9188 static void cpuacct_update_stats(struct task_struct *tsk,
9189 enum cpuacct_stat_index idx, cputime_t val)
9192 int batch = CPUACCT_BATCH;
9194 if (unlikely(!cpuacct_subsys.active))
9201 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9207 struct cgroup_subsys cpuacct_subsys = {
9209 .create = cpuacct_create,
9210 .destroy = cpuacct_destroy,
9211 .populate = cpuacct_populate,
9212 .subsys_id = cpuacct_subsys_id,
9214 #endif /* CONFIG_CGROUP_CPUACCT */
9218 void synchronize_sched_expedited(void)
9222 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9224 #else /* #ifndef CONFIG_SMP */
9226 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9228 static int synchronize_sched_expedited_cpu_stop(void *data)
9231 * There must be a full memory barrier on each affected CPU
9232 * between the time that try_stop_cpus() is called and the
9233 * time that it returns.
9235 * In the current initial implementation of cpu_stop, the
9236 * above condition is already met when the control reaches
9237 * this point and the following smp_mb() is not strictly
9238 * necessary. Do smp_mb() anyway for documentation and
9239 * robustness against future implementation changes.
9241 smp_mb(); /* See above comment block. */
9246 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9247 * approach to force grace period to end quickly. This consumes
9248 * significant time on all CPUs, and is thus not recommended for
9249 * any sort of common-case code.
9251 * Note that it is illegal to call this function while holding any
9252 * lock that is acquired by a CPU-hotplug notifier. Failing to
9253 * observe this restriction will result in deadlock.
9255 void synchronize_sched_expedited(void)
9257 int snap, trycount = 0;
9259 smp_mb(); /* ensure prior mod happens before capturing snap. */
9260 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9262 while (try_stop_cpus(cpu_online_mask,
9263 synchronize_sched_expedited_cpu_stop,
9266 if (trycount++ < 10)
9267 udelay(trycount * num_online_cpus());
9269 synchronize_sched();
9272 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9273 smp_mb(); /* ensure test happens before caller kfree */
9278 atomic_inc(&synchronize_sched_expedited_count);
9279 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9282 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9284 #endif /* #else #ifndef CONFIG_SMP */