4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/sched.h>
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
124 static inline int rt_policy(int policy)
126 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
131 static inline int task_has_rt_policy(struct task_struct *p)
133 return rt_policy(p->policy);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array {
140 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
141 struct list_head queue[MAX_RT_PRIO];
144 struct rt_bandwidth {
145 /* nests inside the rq lock: */
146 raw_spinlock_t rt_runtime_lock;
149 struct hrtimer rt_period_timer;
152 static struct rt_bandwidth def_rt_bandwidth;
154 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
156 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
158 struct rt_bandwidth *rt_b =
159 container_of(timer, struct rt_bandwidth, rt_period_timer);
165 now = hrtimer_cb_get_time(timer);
166 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
171 idle = do_sched_rt_period_timer(rt_b, overrun);
174 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
180 rt_b->rt_period = ns_to_ktime(period);
181 rt_b->rt_runtime = runtime;
183 raw_spin_lock_init(&rt_b->rt_runtime_lock);
185 hrtimer_init(&rt_b->rt_period_timer,
186 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
187 rt_b->rt_period_timer.function = sched_rt_period_timer;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime >= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
199 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
202 if (hrtimer_active(&rt_b->rt_period_timer))
205 raw_spin_lock(&rt_b->rt_runtime_lock);
210 if (hrtimer_active(&rt_b->rt_period_timer))
213 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
214 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
216 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
217 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
218 delta = ktime_to_ns(ktime_sub(hard, soft));
219 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
220 HRTIMER_MODE_ABS_PINNED, 0);
222 raw_spin_unlock(&rt_b->rt_runtime_lock);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
228 hrtimer_cancel(&rt_b->rt_period_timer);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex);
238 #ifdef CONFIG_CGROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups);
246 /* task group related information */
248 struct cgroup_subsys_state css;
250 #ifdef CONFIG_FAIR_GROUP_SCHED
251 /* schedulable entities of this group on each cpu */
252 struct sched_entity **se;
253 /* runqueue "owned" by this group on each cpu */
254 struct cfs_rq **cfs_rq;
255 unsigned long shares;
258 #ifdef CONFIG_RT_GROUP_SCHED
259 struct sched_rt_entity **rt_se;
260 struct rt_rq **rt_rq;
262 struct rt_bandwidth rt_bandwidth;
266 struct list_head list;
268 struct task_group *parent;
269 struct list_head siblings;
270 struct list_head children;
273 #define root_task_group init_task_group
275 /* task_group_lock serializes add/remove of task groups and also changes to
276 * a task group's cpu shares.
278 static DEFINE_SPINLOCK(task_group_lock);
280 #ifdef CONFIG_FAIR_GROUP_SCHED
283 static int root_task_group_empty(void)
285 return list_empty(&root_task_group.children);
289 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
292 * A weight of 0 or 1 can cause arithmetics problems.
293 * A weight of a cfs_rq is the sum of weights of which entities
294 * are queued on this cfs_rq, so a weight of a entity should not be
295 * too large, so as the shares value of a task group.
296 * (The default weight is 1024 - so there's no practical
297 * limitation from this.)
300 #define MAX_SHARES (1UL << 18)
302 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
305 /* Default task group.
306 * Every task in system belong to this group at bootup.
308 struct task_group init_task_group;
310 #endif /* CONFIG_CGROUP_SCHED */
312 /* CFS-related fields in a runqueue */
314 struct load_weight load;
315 unsigned long nr_running;
320 struct rb_root tasks_timeline;
321 struct rb_node *rb_leftmost;
323 struct list_head tasks;
324 struct list_head *balance_iterator;
327 * 'curr' points to currently running entity on this cfs_rq.
328 * It is set to NULL otherwise (i.e when none are currently running).
330 struct sched_entity *curr, *next, *last;
332 unsigned int nr_spread_over;
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
338 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
339 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
340 * (like users, containers etc.)
342 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
343 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list;
346 struct task_group *tg; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load;
363 * this cpu's part of tg->shares
365 unsigned long shares;
368 * load.weight at the time we set shares
370 unsigned long rq_weight;
375 /* Real-Time classes' related field in a runqueue: */
377 struct rt_prio_array active;
378 unsigned long rt_nr_running;
379 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
381 int curr; /* highest queued rt task prio */
383 int next; /* next highest */
388 unsigned long rt_nr_migratory;
389 unsigned long rt_nr_total;
391 struct plist_head pushable_tasks;
396 /* Nests inside the rq lock: */
397 raw_spinlock_t rt_runtime_lock;
399 #ifdef CONFIG_RT_GROUP_SCHED
400 unsigned long rt_nr_boosted;
403 struct list_head leaf_rt_rq_list;
404 struct task_group *tg;
411 * We add the notion of a root-domain which will be used to define per-domain
412 * variables. Each exclusive cpuset essentially defines an island domain by
413 * fully partitioning the member cpus from any other cpuset. Whenever a new
414 * exclusive cpuset is created, we also create and attach a new root-domain
421 cpumask_var_t online;
424 * The "RT overload" flag: it gets set if a CPU has more than
425 * one runnable RT task.
427 cpumask_var_t rto_mask;
429 struct cpupri cpupri;
433 * By default the system creates a single root-domain with all cpus as
434 * members (mimicking the global state we have today).
436 static struct root_domain def_root_domain;
438 #endif /* CONFIG_SMP */
441 * This is the main, per-CPU runqueue data structure.
443 * Locking rule: those places that want to lock multiple runqueues
444 * (such as the load balancing or the thread migration code), lock
445 * acquire operations must be ordered by ascending &runqueue.
452 * nr_running and cpu_load should be in the same cacheline because
453 * remote CPUs use both these fields when doing load calculation.
455 unsigned long nr_running;
456 #define CPU_LOAD_IDX_MAX 5
457 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
458 unsigned long last_load_update_tick;
461 unsigned char nohz_balance_kick;
463 unsigned int skip_clock_update;
465 /* capture load from *all* tasks on this cpu: */
466 struct load_weight load;
467 unsigned long nr_load_updates;
473 #ifdef CONFIG_FAIR_GROUP_SCHED
474 /* list of leaf cfs_rq on this cpu: */
475 struct list_head leaf_cfs_rq_list;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 struct list_head leaf_rt_rq_list;
482 * This is part of a global counter where only the total sum
483 * over all CPUs matters. A task can increase this counter on
484 * one CPU and if it got migrated afterwards it may decrease
485 * it on another CPU. Always updated under the runqueue lock:
487 unsigned long nr_uninterruptible;
489 struct task_struct *curr, *idle;
490 unsigned long next_balance;
491 struct mm_struct *prev_mm;
498 struct root_domain *rd;
499 struct sched_domain *sd;
501 unsigned long cpu_power;
503 unsigned char idle_at_tick;
504 /* For active balancing */
508 struct cpu_stop_work active_balance_work;
509 /* cpu of this runqueue: */
513 unsigned long avg_load_per_task;
521 /* calc_load related fields */
522 unsigned long calc_load_update;
523 long calc_load_active;
525 #ifdef CONFIG_SCHED_HRTICK
527 int hrtick_csd_pending;
528 struct call_single_data hrtick_csd;
530 struct hrtimer hrtick_timer;
533 #ifdef CONFIG_SCHEDSTATS
535 struct sched_info rq_sched_info;
536 unsigned long long rq_cpu_time;
537 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
539 /* sys_sched_yield() stats */
540 unsigned int yld_count;
542 /* schedule() stats */
543 unsigned int sched_switch;
544 unsigned int sched_count;
545 unsigned int sched_goidle;
547 /* try_to_wake_up() stats */
548 unsigned int ttwu_count;
549 unsigned int ttwu_local;
552 unsigned int bkl_count;
556 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
559 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
561 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
564 * A queue event has occurred, and we're going to schedule. In
565 * this case, we can save a useless back to back clock update.
567 if (test_tsk_need_resched(p))
568 rq->skip_clock_update = 1;
571 static inline int cpu_of(struct rq *rq)
580 #define rcu_dereference_check_sched_domain(p) \
581 rcu_dereference_check((p), \
582 rcu_read_lock_sched_held() || \
583 lockdep_is_held(&sched_domains_mutex))
586 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
587 * See detach_destroy_domains: synchronize_sched for details.
589 * The domain tree of any CPU may only be accessed from within
590 * preempt-disabled sections.
592 #define for_each_domain(cpu, __sd) \
593 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
595 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
596 #define this_rq() (&__get_cpu_var(runqueues))
597 #define task_rq(p) cpu_rq(task_cpu(p))
598 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
599 #define raw_rq() (&__raw_get_cpu_var(runqueues))
601 #ifdef CONFIG_CGROUP_SCHED
604 * Return the group to which this tasks belongs.
606 * We use task_subsys_state_check() and extend the RCU verification
607 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
608 * holds that lock for each task it moves into the cgroup. Therefore
609 * by holding that lock, we pin the task to the current cgroup.
611 static inline struct task_group *task_group(struct task_struct *p)
613 struct cgroup_subsys_state *css;
615 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
616 lockdep_is_held(&task_rq(p)->lock));
617 return container_of(css, struct task_group, css);
620 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
621 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
623 #ifdef CONFIG_FAIR_GROUP_SCHED
624 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
625 p->se.parent = task_group(p)->se[cpu];
628 #ifdef CONFIG_RT_GROUP_SCHED
629 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
630 p->rt.parent = task_group(p)->rt_se[cpu];
634 #else /* CONFIG_CGROUP_SCHED */
636 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
637 static inline struct task_group *task_group(struct task_struct *p)
642 #endif /* CONFIG_CGROUP_SCHED */
644 inline void update_rq_clock(struct rq *rq)
646 if (!rq->skip_clock_update)
647 rq->clock = sched_clock_cpu(cpu_of(rq));
651 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
653 #ifdef CONFIG_SCHED_DEBUG
654 # define const_debug __read_mostly
656 # define const_debug static const
661 * @cpu: the processor in question.
663 * Returns true if the current cpu runqueue is locked.
664 * This interface allows printk to be called with the runqueue lock
665 * held and know whether or not it is OK to wake up the klogd.
667 int runqueue_is_locked(int cpu)
669 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
673 * Debugging: various feature bits
676 #define SCHED_FEAT(name, enabled) \
677 __SCHED_FEAT_##name ,
680 #include "sched_features.h"
685 #define SCHED_FEAT(name, enabled) \
686 (1UL << __SCHED_FEAT_##name) * enabled |
688 const_debug unsigned int sysctl_sched_features =
689 #include "sched_features.h"
694 #ifdef CONFIG_SCHED_DEBUG
695 #define SCHED_FEAT(name, enabled) \
698 static __read_mostly char *sched_feat_names[] = {
699 #include "sched_features.h"
705 static int sched_feat_show(struct seq_file *m, void *v)
709 for (i = 0; sched_feat_names[i]; i++) {
710 if (!(sysctl_sched_features & (1UL << i)))
712 seq_printf(m, "%s ", sched_feat_names[i]);
720 sched_feat_write(struct file *filp, const char __user *ubuf,
721 size_t cnt, loff_t *ppos)
731 if (copy_from_user(&buf, ubuf, cnt))
737 if (strncmp(buf, "NO_", 3) == 0) {
742 for (i = 0; sched_feat_names[i]; i++) {
743 if (strcmp(cmp, sched_feat_names[i]) == 0) {
745 sysctl_sched_features &= ~(1UL << i);
747 sysctl_sched_features |= (1UL << i);
752 if (!sched_feat_names[i])
760 static int sched_feat_open(struct inode *inode, struct file *filp)
762 return single_open(filp, sched_feat_show, NULL);
765 static const struct file_operations sched_feat_fops = {
766 .open = sched_feat_open,
767 .write = sched_feat_write,
770 .release = single_release,
773 static __init int sched_init_debug(void)
775 debugfs_create_file("sched_features", 0644, NULL, NULL,
780 late_initcall(sched_init_debug);
784 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
787 * Number of tasks to iterate in a single balance run.
788 * Limited because this is done with IRQs disabled.
790 const_debug unsigned int sysctl_sched_nr_migrate = 32;
793 * ratelimit for updating the group shares.
796 unsigned int sysctl_sched_shares_ratelimit = 250000;
797 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
800 * Inject some fuzzyness into changing the per-cpu group shares
801 * this avoids remote rq-locks at the expense of fairness.
804 unsigned int sysctl_sched_shares_thresh = 4;
807 * period over which we average the RT time consumption, measured
812 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
815 * period over which we measure -rt task cpu usage in us.
818 unsigned int sysctl_sched_rt_period = 1000000;
820 static __read_mostly int scheduler_running;
823 * part of the period that we allow rt tasks to run in us.
826 int sysctl_sched_rt_runtime = 950000;
828 static inline u64 global_rt_period(void)
830 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
833 static inline u64 global_rt_runtime(void)
835 if (sysctl_sched_rt_runtime < 0)
838 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
841 #ifndef prepare_arch_switch
842 # define prepare_arch_switch(next) do { } while (0)
844 #ifndef finish_arch_switch
845 # define finish_arch_switch(prev) do { } while (0)
848 static inline int task_current(struct rq *rq, struct task_struct *p)
850 return rq->curr == p;
853 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
854 static inline int task_running(struct rq *rq, struct task_struct *p)
856 return task_current(rq, p);
859 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
863 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
865 #ifdef CONFIG_DEBUG_SPINLOCK
866 /* this is a valid case when another task releases the spinlock */
867 rq->lock.owner = current;
870 * If we are tracking spinlock dependencies then we have to
871 * fix up the runqueue lock - which gets 'carried over' from
874 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
876 raw_spin_unlock_irq(&rq->lock);
879 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
880 static inline int task_running(struct rq *rq, struct task_struct *p)
885 return task_current(rq, p);
889 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
893 * We can optimise this out completely for !SMP, because the
894 * SMP rebalancing from interrupt is the only thing that cares
899 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
900 raw_spin_unlock_irq(&rq->lock);
902 raw_spin_unlock(&rq->lock);
906 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
910 * After ->oncpu is cleared, the task can be moved to a different CPU.
911 * We must ensure this doesn't happen until the switch is completely
917 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
924 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
927 static inline int task_is_waking(struct task_struct *p)
929 return unlikely(p->state == TASK_WAKING);
933 * __task_rq_lock - lock the runqueue a given task resides on.
934 * Must be called interrupts disabled.
936 static inline struct rq *__task_rq_lock(struct task_struct *p)
943 raw_spin_lock(&rq->lock);
944 if (likely(rq == task_rq(p)))
946 raw_spin_unlock(&rq->lock);
951 * task_rq_lock - lock the runqueue a given task resides on and disable
952 * interrupts. Note the ordering: we can safely lookup the task_rq without
953 * explicitly disabling preemption.
955 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
961 local_irq_save(*flags);
963 raw_spin_lock(&rq->lock);
964 if (likely(rq == task_rq(p)))
966 raw_spin_unlock_irqrestore(&rq->lock, *flags);
970 static void __task_rq_unlock(struct rq *rq)
973 raw_spin_unlock(&rq->lock);
976 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
979 raw_spin_unlock_irqrestore(&rq->lock, *flags);
983 * this_rq_lock - lock this runqueue and disable interrupts.
985 static struct rq *this_rq_lock(void)
992 raw_spin_lock(&rq->lock);
997 #ifdef CONFIG_SCHED_HRTICK
999 * Use HR-timers to deliver accurate preemption points.
1001 * Its all a bit involved since we cannot program an hrt while holding the
1002 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1005 * When we get rescheduled we reprogram the hrtick_timer outside of the
1011 * - enabled by features
1012 * - hrtimer is actually high res
1014 static inline int hrtick_enabled(struct rq *rq)
1016 if (!sched_feat(HRTICK))
1018 if (!cpu_active(cpu_of(rq)))
1020 return hrtimer_is_hres_active(&rq->hrtick_timer);
1023 static void hrtick_clear(struct rq *rq)
1025 if (hrtimer_active(&rq->hrtick_timer))
1026 hrtimer_cancel(&rq->hrtick_timer);
1030 * High-resolution timer tick.
1031 * Runs from hardirq context with interrupts disabled.
1033 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1035 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1037 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1039 raw_spin_lock(&rq->lock);
1040 update_rq_clock(rq);
1041 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1042 raw_spin_unlock(&rq->lock);
1044 return HRTIMER_NORESTART;
1049 * called from hardirq (IPI) context
1051 static void __hrtick_start(void *arg)
1053 struct rq *rq = arg;
1055 raw_spin_lock(&rq->lock);
1056 hrtimer_restart(&rq->hrtick_timer);
1057 rq->hrtick_csd_pending = 0;
1058 raw_spin_unlock(&rq->lock);
1062 * Called to set the hrtick timer state.
1064 * called with rq->lock held and irqs disabled
1066 static void hrtick_start(struct rq *rq, u64 delay)
1068 struct hrtimer *timer = &rq->hrtick_timer;
1069 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1071 hrtimer_set_expires(timer, time);
1073 if (rq == this_rq()) {
1074 hrtimer_restart(timer);
1075 } else if (!rq->hrtick_csd_pending) {
1076 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1077 rq->hrtick_csd_pending = 1;
1082 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1084 int cpu = (int)(long)hcpu;
1087 case CPU_UP_CANCELED:
1088 case CPU_UP_CANCELED_FROZEN:
1089 case CPU_DOWN_PREPARE:
1090 case CPU_DOWN_PREPARE_FROZEN:
1092 case CPU_DEAD_FROZEN:
1093 hrtick_clear(cpu_rq(cpu));
1100 static __init void init_hrtick(void)
1102 hotcpu_notifier(hotplug_hrtick, 0);
1106 * Called to set the hrtick timer state.
1108 * called with rq->lock held and irqs disabled
1110 static void hrtick_start(struct rq *rq, u64 delay)
1112 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1113 HRTIMER_MODE_REL_PINNED, 0);
1116 static inline void init_hrtick(void)
1119 #endif /* CONFIG_SMP */
1121 static void init_rq_hrtick(struct rq *rq)
1124 rq->hrtick_csd_pending = 0;
1126 rq->hrtick_csd.flags = 0;
1127 rq->hrtick_csd.func = __hrtick_start;
1128 rq->hrtick_csd.info = rq;
1131 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1132 rq->hrtick_timer.function = hrtick;
1134 #else /* CONFIG_SCHED_HRTICK */
1135 static inline void hrtick_clear(struct rq *rq)
1139 static inline void init_rq_hrtick(struct rq *rq)
1143 static inline void init_hrtick(void)
1146 #endif /* CONFIG_SCHED_HRTICK */
1149 * resched_task - mark a task 'to be rescheduled now'.
1151 * On UP this means the setting of the need_resched flag, on SMP it
1152 * might also involve a cross-CPU call to trigger the scheduler on
1157 #ifndef tsk_is_polling
1158 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1161 static void resched_task(struct task_struct *p)
1165 assert_raw_spin_locked(&task_rq(p)->lock);
1167 if (test_tsk_need_resched(p))
1170 set_tsk_need_resched(p);
1173 if (cpu == smp_processor_id())
1176 /* NEED_RESCHED must be visible before we test polling */
1178 if (!tsk_is_polling(p))
1179 smp_send_reschedule(cpu);
1182 static void resched_cpu(int cpu)
1184 struct rq *rq = cpu_rq(cpu);
1185 unsigned long flags;
1187 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1189 resched_task(cpu_curr(cpu));
1190 raw_spin_unlock_irqrestore(&rq->lock, flags);
1195 * In the semi idle case, use the nearest busy cpu for migrating timers
1196 * from an idle cpu. This is good for power-savings.
1198 * We don't do similar optimization for completely idle system, as
1199 * selecting an idle cpu will add more delays to the timers than intended
1200 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1202 int get_nohz_timer_target(void)
1204 int cpu = smp_processor_id();
1206 struct sched_domain *sd;
1208 for_each_domain(cpu, sd) {
1209 for_each_cpu(i, sched_domain_span(sd))
1216 * When add_timer_on() enqueues a timer into the timer wheel of an
1217 * idle CPU then this timer might expire before the next timer event
1218 * which is scheduled to wake up that CPU. In case of a completely
1219 * idle system the next event might even be infinite time into the
1220 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1221 * leaves the inner idle loop so the newly added timer is taken into
1222 * account when the CPU goes back to idle and evaluates the timer
1223 * wheel for the next timer event.
1225 void wake_up_idle_cpu(int cpu)
1227 struct rq *rq = cpu_rq(cpu);
1229 if (cpu == smp_processor_id())
1233 * This is safe, as this function is called with the timer
1234 * wheel base lock of (cpu) held. When the CPU is on the way
1235 * to idle and has not yet set rq->curr to idle then it will
1236 * be serialized on the timer wheel base lock and take the new
1237 * timer into account automatically.
1239 if (rq->curr != rq->idle)
1243 * We can set TIF_RESCHED on the idle task of the other CPU
1244 * lockless. The worst case is that the other CPU runs the
1245 * idle task through an additional NOOP schedule()
1247 set_tsk_need_resched(rq->idle);
1249 /* NEED_RESCHED must be visible before we test polling */
1251 if (!tsk_is_polling(rq->idle))
1252 smp_send_reschedule(cpu);
1255 #endif /* CONFIG_NO_HZ */
1257 static u64 sched_avg_period(void)
1259 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1262 static void sched_avg_update(struct rq *rq)
1264 s64 period = sched_avg_period();
1266 while ((s64)(rq->clock - rq->age_stamp) > period) {
1268 * Inline assembly required to prevent the compiler
1269 * optimising this loop into a divmod call.
1270 * See __iter_div_u64_rem() for another example of this.
1272 asm("" : "+rm" (rq->age_stamp));
1273 rq->age_stamp += period;
1278 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1280 rq->rt_avg += rt_delta;
1281 sched_avg_update(rq);
1284 #else /* !CONFIG_SMP */
1285 static void resched_task(struct task_struct *p)
1287 assert_raw_spin_locked(&task_rq(p)->lock);
1288 set_tsk_need_resched(p);
1291 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1294 #endif /* CONFIG_SMP */
1296 #if BITS_PER_LONG == 32
1297 # define WMULT_CONST (~0UL)
1299 # define WMULT_CONST (1UL << 32)
1302 #define WMULT_SHIFT 32
1305 * Shift right and round:
1307 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1310 * delta *= weight / lw
1312 static unsigned long
1313 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1314 struct load_weight *lw)
1318 if (!lw->inv_weight) {
1319 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1322 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1326 tmp = (u64)delta_exec * weight;
1328 * Check whether we'd overflow the 64-bit multiplication:
1330 if (unlikely(tmp > WMULT_CONST))
1331 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1334 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1336 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1339 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1345 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1352 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1353 * of tasks with abnormal "nice" values across CPUs the contribution that
1354 * each task makes to its run queue's load is weighted according to its
1355 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1356 * scaled version of the new time slice allocation that they receive on time
1360 #define WEIGHT_IDLEPRIO 3
1361 #define WMULT_IDLEPRIO 1431655765
1364 * Nice levels are multiplicative, with a gentle 10% change for every
1365 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1366 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1367 * that remained on nice 0.
1369 * The "10% effect" is relative and cumulative: from _any_ nice level,
1370 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1371 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1372 * If a task goes up by ~10% and another task goes down by ~10% then
1373 * the relative distance between them is ~25%.)
1375 static const int prio_to_weight[40] = {
1376 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1377 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1378 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1379 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1380 /* 0 */ 1024, 820, 655, 526, 423,
1381 /* 5 */ 335, 272, 215, 172, 137,
1382 /* 10 */ 110, 87, 70, 56, 45,
1383 /* 15 */ 36, 29, 23, 18, 15,
1387 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1389 * In cases where the weight does not change often, we can use the
1390 * precalculated inverse to speed up arithmetics by turning divisions
1391 * into multiplications:
1393 static const u32 prio_to_wmult[40] = {
1394 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1395 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1396 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1397 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1398 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1399 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1400 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1401 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1404 /* Time spent by the tasks of the cpu accounting group executing in ... */
1405 enum cpuacct_stat_index {
1406 CPUACCT_STAT_USER, /* ... user mode */
1407 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1409 CPUACCT_STAT_NSTATS,
1412 #ifdef CONFIG_CGROUP_CPUACCT
1413 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1414 static void cpuacct_update_stats(struct task_struct *tsk,
1415 enum cpuacct_stat_index idx, cputime_t val);
1417 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1418 static inline void cpuacct_update_stats(struct task_struct *tsk,
1419 enum cpuacct_stat_index idx, cputime_t val) {}
1422 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1424 update_load_add(&rq->load, load);
1427 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1429 update_load_sub(&rq->load, load);
1432 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1433 typedef int (*tg_visitor)(struct task_group *, void *);
1436 * Iterate the full tree, calling @down when first entering a node and @up when
1437 * leaving it for the final time.
1439 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1441 struct task_group *parent, *child;
1445 parent = &root_task_group;
1447 ret = (*down)(parent, data);
1450 list_for_each_entry_rcu(child, &parent->children, siblings) {
1457 ret = (*up)(parent, data);
1462 parent = parent->parent;
1471 static int tg_nop(struct task_group *tg, void *data)
1478 /* Used instead of source_load when we know the type == 0 */
1479 static unsigned long weighted_cpuload(const int cpu)
1481 return cpu_rq(cpu)->load.weight;
1485 * Return a low guess at the load of a migration-source cpu weighted
1486 * according to the scheduling class and "nice" value.
1488 * We want to under-estimate the load of migration sources, to
1489 * balance conservatively.
1491 static unsigned long source_load(int cpu, int type)
1493 struct rq *rq = cpu_rq(cpu);
1494 unsigned long total = weighted_cpuload(cpu);
1496 if (type == 0 || !sched_feat(LB_BIAS))
1499 return min(rq->cpu_load[type-1], total);
1503 * Return a high guess at the load of a migration-target cpu weighted
1504 * according to the scheduling class and "nice" value.
1506 static unsigned long target_load(int cpu, int type)
1508 struct rq *rq = cpu_rq(cpu);
1509 unsigned long total = weighted_cpuload(cpu);
1511 if (type == 0 || !sched_feat(LB_BIAS))
1514 return max(rq->cpu_load[type-1], total);
1517 static unsigned long power_of(int cpu)
1519 return cpu_rq(cpu)->cpu_power;
1522 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1524 static unsigned long cpu_avg_load_per_task(int cpu)
1526 struct rq *rq = cpu_rq(cpu);
1527 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1530 rq->avg_load_per_task = rq->load.weight / nr_running;
1532 rq->avg_load_per_task = 0;
1534 return rq->avg_load_per_task;
1537 #ifdef CONFIG_FAIR_GROUP_SCHED
1539 static __read_mostly unsigned long __percpu *update_shares_data;
1541 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1544 * Calculate and set the cpu's group shares.
1546 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1547 unsigned long sd_shares,
1548 unsigned long sd_rq_weight,
1549 unsigned long *usd_rq_weight)
1551 unsigned long shares, rq_weight;
1554 rq_weight = usd_rq_weight[cpu];
1557 rq_weight = NICE_0_LOAD;
1561 * \Sum_j shares_j * rq_weight_i
1562 * shares_i = -----------------------------
1563 * \Sum_j rq_weight_j
1565 shares = (sd_shares * rq_weight) / sd_rq_weight;
1566 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1568 if (abs(shares - tg->se[cpu]->load.weight) >
1569 sysctl_sched_shares_thresh) {
1570 struct rq *rq = cpu_rq(cpu);
1571 unsigned long flags;
1573 raw_spin_lock_irqsave(&rq->lock, flags);
1574 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1575 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1576 __set_se_shares(tg->se[cpu], shares);
1577 raw_spin_unlock_irqrestore(&rq->lock, flags);
1582 * Re-compute the task group their per cpu shares over the given domain.
1583 * This needs to be done in a bottom-up fashion because the rq weight of a
1584 * parent group depends on the shares of its child groups.
1586 static int tg_shares_up(struct task_group *tg, void *data)
1588 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1589 unsigned long *usd_rq_weight;
1590 struct sched_domain *sd = data;
1591 unsigned long flags;
1597 local_irq_save(flags);
1598 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1600 for_each_cpu(i, sched_domain_span(sd)) {
1601 weight = tg->cfs_rq[i]->load.weight;
1602 usd_rq_weight[i] = weight;
1604 rq_weight += weight;
1606 * If there are currently no tasks on the cpu pretend there
1607 * is one of average load so that when a new task gets to
1608 * run here it will not get delayed by group starvation.
1611 weight = NICE_0_LOAD;
1613 sum_weight += weight;
1614 shares += tg->cfs_rq[i]->shares;
1618 rq_weight = sum_weight;
1620 if ((!shares && rq_weight) || shares > tg->shares)
1621 shares = tg->shares;
1623 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1624 shares = tg->shares;
1626 for_each_cpu(i, sched_domain_span(sd))
1627 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1629 local_irq_restore(flags);
1635 * Compute the cpu's hierarchical load factor for each task group.
1636 * This needs to be done in a top-down fashion because the load of a child
1637 * group is a fraction of its parents load.
1639 static int tg_load_down(struct task_group *tg, void *data)
1642 long cpu = (long)data;
1645 load = cpu_rq(cpu)->load.weight;
1647 load = tg->parent->cfs_rq[cpu]->h_load;
1648 load *= tg->cfs_rq[cpu]->shares;
1649 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1652 tg->cfs_rq[cpu]->h_load = load;
1657 static void update_shares(struct sched_domain *sd)
1662 if (root_task_group_empty())
1665 now = local_clock();
1666 elapsed = now - sd->last_update;
1668 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1669 sd->last_update = now;
1670 walk_tg_tree(tg_nop, tg_shares_up, sd);
1674 static void update_h_load(long cpu)
1676 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1681 static inline void update_shares(struct sched_domain *sd)
1687 #ifdef CONFIG_PREEMPT
1689 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1692 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1693 * way at the expense of forcing extra atomic operations in all
1694 * invocations. This assures that the double_lock is acquired using the
1695 * same underlying policy as the spinlock_t on this architecture, which
1696 * reduces latency compared to the unfair variant below. However, it
1697 * also adds more overhead and therefore may reduce throughput.
1699 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1700 __releases(this_rq->lock)
1701 __acquires(busiest->lock)
1702 __acquires(this_rq->lock)
1704 raw_spin_unlock(&this_rq->lock);
1705 double_rq_lock(this_rq, busiest);
1712 * Unfair double_lock_balance: Optimizes throughput at the expense of
1713 * latency by eliminating extra atomic operations when the locks are
1714 * already in proper order on entry. This favors lower cpu-ids and will
1715 * grant the double lock to lower cpus over higher ids under contention,
1716 * regardless of entry order into the function.
1718 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1719 __releases(this_rq->lock)
1720 __acquires(busiest->lock)
1721 __acquires(this_rq->lock)
1725 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1726 if (busiest < this_rq) {
1727 raw_spin_unlock(&this_rq->lock);
1728 raw_spin_lock(&busiest->lock);
1729 raw_spin_lock_nested(&this_rq->lock,
1730 SINGLE_DEPTH_NESTING);
1733 raw_spin_lock_nested(&busiest->lock,
1734 SINGLE_DEPTH_NESTING);
1739 #endif /* CONFIG_PREEMPT */
1742 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1744 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1746 if (unlikely(!irqs_disabled())) {
1747 /* printk() doesn't work good under rq->lock */
1748 raw_spin_unlock(&this_rq->lock);
1752 return _double_lock_balance(this_rq, busiest);
1755 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1756 __releases(busiest->lock)
1758 raw_spin_unlock(&busiest->lock);
1759 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1763 * double_rq_lock - safely lock two runqueues
1765 * Note this does not disable interrupts like task_rq_lock,
1766 * you need to do so manually before calling.
1768 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1769 __acquires(rq1->lock)
1770 __acquires(rq2->lock)
1772 BUG_ON(!irqs_disabled());
1774 raw_spin_lock(&rq1->lock);
1775 __acquire(rq2->lock); /* Fake it out ;) */
1778 raw_spin_lock(&rq1->lock);
1779 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1781 raw_spin_lock(&rq2->lock);
1782 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1788 * double_rq_unlock - safely unlock two runqueues
1790 * Note this does not restore interrupts like task_rq_unlock,
1791 * you need to do so manually after calling.
1793 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1794 __releases(rq1->lock)
1795 __releases(rq2->lock)
1797 raw_spin_unlock(&rq1->lock);
1799 raw_spin_unlock(&rq2->lock);
1801 __release(rq2->lock);
1806 #ifdef CONFIG_FAIR_GROUP_SCHED
1807 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1810 cfs_rq->shares = shares;
1815 static void calc_load_account_idle(struct rq *this_rq);
1816 static void update_sysctl(void);
1817 static int get_update_sysctl_factor(void);
1818 static void update_cpu_load(struct rq *this_rq);
1820 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1822 set_task_rq(p, cpu);
1825 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1826 * successfuly executed on another CPU. We must ensure that updates of
1827 * per-task data have been completed by this moment.
1830 task_thread_info(p)->cpu = cpu;
1834 static const struct sched_class rt_sched_class;
1836 #define sched_class_highest (&rt_sched_class)
1837 #define for_each_class(class) \
1838 for (class = sched_class_highest; class; class = class->next)
1840 #include "sched_stats.h"
1842 static void inc_nr_running(struct rq *rq)
1847 static void dec_nr_running(struct rq *rq)
1852 static void set_load_weight(struct task_struct *p)
1854 if (task_has_rt_policy(p)) {
1855 p->se.load.weight = 0;
1856 p->se.load.inv_weight = WMULT_CONST;
1861 * SCHED_IDLE tasks get minimal weight:
1863 if (p->policy == SCHED_IDLE) {
1864 p->se.load.weight = WEIGHT_IDLEPRIO;
1865 p->se.load.inv_weight = WMULT_IDLEPRIO;
1869 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1870 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1873 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1875 update_rq_clock(rq);
1876 sched_info_queued(p);
1877 p->sched_class->enqueue_task(rq, p, flags);
1881 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1883 update_rq_clock(rq);
1884 sched_info_dequeued(p);
1885 p->sched_class->dequeue_task(rq, p, flags);
1890 * activate_task - move a task to the runqueue.
1892 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1894 if (task_contributes_to_load(p))
1895 rq->nr_uninterruptible--;
1897 enqueue_task(rq, p, flags);
1902 * deactivate_task - remove a task from the runqueue.
1904 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1906 if (task_contributes_to_load(p))
1907 rq->nr_uninterruptible++;
1909 dequeue_task(rq, p, flags);
1913 #include "sched_idletask.c"
1914 #include "sched_fair.c"
1915 #include "sched_rt.c"
1916 #ifdef CONFIG_SCHED_DEBUG
1917 # include "sched_debug.c"
1921 * __normal_prio - return the priority that is based on the static prio
1923 static inline int __normal_prio(struct task_struct *p)
1925 return p->static_prio;
1929 * Calculate the expected normal priority: i.e. priority
1930 * without taking RT-inheritance into account. Might be
1931 * boosted by interactivity modifiers. Changes upon fork,
1932 * setprio syscalls, and whenever the interactivity
1933 * estimator recalculates.
1935 static inline int normal_prio(struct task_struct *p)
1939 if (task_has_rt_policy(p))
1940 prio = MAX_RT_PRIO-1 - p->rt_priority;
1942 prio = __normal_prio(p);
1947 * Calculate the current priority, i.e. the priority
1948 * taken into account by the scheduler. This value might
1949 * be boosted by RT tasks, or might be boosted by
1950 * interactivity modifiers. Will be RT if the task got
1951 * RT-boosted. If not then it returns p->normal_prio.
1953 static int effective_prio(struct task_struct *p)
1955 p->normal_prio = normal_prio(p);
1957 * If we are RT tasks or we were boosted to RT priority,
1958 * keep the priority unchanged. Otherwise, update priority
1959 * to the normal priority:
1961 if (!rt_prio(p->prio))
1962 return p->normal_prio;
1967 * task_curr - is this task currently executing on a CPU?
1968 * @p: the task in question.
1970 inline int task_curr(const struct task_struct *p)
1972 return cpu_curr(task_cpu(p)) == p;
1975 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1976 const struct sched_class *prev_class,
1977 int oldprio, int running)
1979 if (prev_class != p->sched_class) {
1980 if (prev_class->switched_from)
1981 prev_class->switched_from(rq, p, running);
1982 p->sched_class->switched_to(rq, p, running);
1984 p->sched_class->prio_changed(rq, p, oldprio, running);
1989 * Is this task likely cache-hot:
1992 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1996 if (p->sched_class != &fair_sched_class)
2000 * Buddy candidates are cache hot:
2002 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2003 (&p->se == cfs_rq_of(&p->se)->next ||
2004 &p->se == cfs_rq_of(&p->se)->last))
2007 if (sysctl_sched_migration_cost == -1)
2009 if (sysctl_sched_migration_cost == 0)
2012 delta = now - p->se.exec_start;
2014 return delta < (s64)sysctl_sched_migration_cost;
2017 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2019 #ifdef CONFIG_SCHED_DEBUG
2021 * We should never call set_task_cpu() on a blocked task,
2022 * ttwu() will sort out the placement.
2024 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2025 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2028 trace_sched_migrate_task(p, new_cpu);
2030 if (task_cpu(p) != new_cpu) {
2031 p->se.nr_migrations++;
2032 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2035 __set_task_cpu(p, new_cpu);
2038 struct migration_arg {
2039 struct task_struct *task;
2043 static int migration_cpu_stop(void *data);
2046 * The task's runqueue lock must be held.
2047 * Returns true if you have to wait for migration thread.
2049 static bool migrate_task(struct task_struct *p, int dest_cpu)
2051 struct rq *rq = task_rq(p);
2054 * If the task is not on a runqueue (and not running), then
2055 * the next wake-up will properly place the task.
2057 return p->se.on_rq || task_running(rq, p);
2061 * wait_task_inactive - wait for a thread to unschedule.
2063 * If @match_state is nonzero, it's the @p->state value just checked and
2064 * not expected to change. If it changes, i.e. @p might have woken up,
2065 * then return zero. When we succeed in waiting for @p to be off its CPU,
2066 * we return a positive number (its total switch count). If a second call
2067 * a short while later returns the same number, the caller can be sure that
2068 * @p has remained unscheduled the whole time.
2070 * The caller must ensure that the task *will* unschedule sometime soon,
2071 * else this function might spin for a *long* time. This function can't
2072 * be called with interrupts off, or it may introduce deadlock with
2073 * smp_call_function() if an IPI is sent by the same process we are
2074 * waiting to become inactive.
2076 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2078 unsigned long flags;
2085 * We do the initial early heuristics without holding
2086 * any task-queue locks at all. We'll only try to get
2087 * the runqueue lock when things look like they will
2093 * If the task is actively running on another CPU
2094 * still, just relax and busy-wait without holding
2097 * NOTE! Since we don't hold any locks, it's not
2098 * even sure that "rq" stays as the right runqueue!
2099 * But we don't care, since "task_running()" will
2100 * return false if the runqueue has changed and p
2101 * is actually now running somewhere else!
2103 while (task_running(rq, p)) {
2104 if (match_state && unlikely(p->state != match_state))
2110 * Ok, time to look more closely! We need the rq
2111 * lock now, to be *sure*. If we're wrong, we'll
2112 * just go back and repeat.
2114 rq = task_rq_lock(p, &flags);
2115 trace_sched_wait_task(p);
2116 running = task_running(rq, p);
2117 on_rq = p->se.on_rq;
2119 if (!match_state || p->state == match_state)
2120 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2121 task_rq_unlock(rq, &flags);
2124 * If it changed from the expected state, bail out now.
2126 if (unlikely(!ncsw))
2130 * Was it really running after all now that we
2131 * checked with the proper locks actually held?
2133 * Oops. Go back and try again..
2135 if (unlikely(running)) {
2141 * It's not enough that it's not actively running,
2142 * it must be off the runqueue _entirely_, and not
2145 * So if it was still runnable (but just not actively
2146 * running right now), it's preempted, and we should
2147 * yield - it could be a while.
2149 if (unlikely(on_rq)) {
2150 schedule_timeout_uninterruptible(1);
2155 * Ahh, all good. It wasn't running, and it wasn't
2156 * runnable, which means that it will never become
2157 * running in the future either. We're all done!
2166 * kick_process - kick a running thread to enter/exit the kernel
2167 * @p: the to-be-kicked thread
2169 * Cause a process which is running on another CPU to enter
2170 * kernel-mode, without any delay. (to get signals handled.)
2172 * NOTE: this function doesnt have to take the runqueue lock,
2173 * because all it wants to ensure is that the remote task enters
2174 * the kernel. If the IPI races and the task has been migrated
2175 * to another CPU then no harm is done and the purpose has been
2178 void kick_process(struct task_struct *p)
2184 if ((cpu != smp_processor_id()) && task_curr(p))
2185 smp_send_reschedule(cpu);
2188 EXPORT_SYMBOL_GPL(kick_process);
2189 #endif /* CONFIG_SMP */
2192 * task_oncpu_function_call - call a function on the cpu on which a task runs
2193 * @p: the task to evaluate
2194 * @func: the function to be called
2195 * @info: the function call argument
2197 * Calls the function @func when the task is currently running. This might
2198 * be on the current CPU, which just calls the function directly
2200 void task_oncpu_function_call(struct task_struct *p,
2201 void (*func) (void *info), void *info)
2208 smp_call_function_single(cpu, func, info, 1);
2214 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2216 static int select_fallback_rq(int cpu, struct task_struct *p)
2219 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2221 /* Look for allowed, online CPU in same node. */
2222 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2223 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2226 /* Any allowed, online CPU? */
2227 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2228 if (dest_cpu < nr_cpu_ids)
2231 /* No more Mr. Nice Guy. */
2232 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2233 dest_cpu = cpuset_cpus_allowed_fallback(p);
2235 * Don't tell them about moving exiting tasks or
2236 * kernel threads (both mm NULL), since they never
2239 if (p->mm && printk_ratelimit()) {
2240 printk(KERN_INFO "process %d (%s) no "
2241 "longer affine to cpu%d\n",
2242 task_pid_nr(p), p->comm, cpu);
2250 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2253 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2255 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2258 * In order not to call set_task_cpu() on a blocking task we need
2259 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2262 * Since this is common to all placement strategies, this lives here.
2264 * [ this allows ->select_task() to simply return task_cpu(p) and
2265 * not worry about this generic constraint ]
2267 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2269 cpu = select_fallback_rq(task_cpu(p), p);
2274 static void update_avg(u64 *avg, u64 sample)
2276 s64 diff = sample - *avg;
2281 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2282 bool is_sync, bool is_migrate, bool is_local,
2283 unsigned long en_flags)
2285 schedstat_inc(p, se.statistics.nr_wakeups);
2287 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2289 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2291 schedstat_inc(p, se.statistics.nr_wakeups_local);
2293 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2295 activate_task(rq, p, en_flags);
2298 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2299 int wake_flags, bool success)
2301 trace_sched_wakeup(p, success);
2302 check_preempt_curr(rq, p, wake_flags);
2304 p->state = TASK_RUNNING;
2306 if (p->sched_class->task_woken)
2307 p->sched_class->task_woken(rq, p);
2309 if (unlikely(rq->idle_stamp)) {
2310 u64 delta = rq->clock - rq->idle_stamp;
2311 u64 max = 2*sysctl_sched_migration_cost;
2316 update_avg(&rq->avg_idle, delta);
2320 /* if a worker is waking up, notify workqueue */
2321 if ((p->flags & PF_WQ_WORKER) && success)
2322 wq_worker_waking_up(p, cpu_of(rq));
2326 * try_to_wake_up - wake up a thread
2327 * @p: the thread to be awakened
2328 * @state: the mask of task states that can be woken
2329 * @wake_flags: wake modifier flags (WF_*)
2331 * Put it on the run-queue if it's not already there. The "current"
2332 * thread is always on the run-queue (except when the actual
2333 * re-schedule is in progress), and as such you're allowed to do
2334 * the simpler "current->state = TASK_RUNNING" to mark yourself
2335 * runnable without the overhead of this.
2337 * Returns %true if @p was woken up, %false if it was already running
2338 * or @state didn't match @p's state.
2340 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2343 int cpu, orig_cpu, this_cpu, success = 0;
2344 unsigned long flags;
2345 unsigned long en_flags = ENQUEUE_WAKEUP;
2348 this_cpu = get_cpu();
2351 rq = task_rq_lock(p, &flags);
2352 if (!(p->state & state))
2362 if (unlikely(task_running(rq, p)))
2366 * In order to handle concurrent wakeups and release the rq->lock
2367 * we put the task in TASK_WAKING state.
2369 * First fix up the nr_uninterruptible count:
2371 if (task_contributes_to_load(p)) {
2372 if (likely(cpu_online(orig_cpu)))
2373 rq->nr_uninterruptible--;
2375 this_rq()->nr_uninterruptible--;
2377 p->state = TASK_WAKING;
2379 if (p->sched_class->task_waking) {
2380 p->sched_class->task_waking(rq, p);
2381 en_flags |= ENQUEUE_WAKING;
2384 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2385 if (cpu != orig_cpu)
2386 set_task_cpu(p, cpu);
2387 __task_rq_unlock(rq);
2390 raw_spin_lock(&rq->lock);
2393 * We migrated the task without holding either rq->lock, however
2394 * since the task is not on the task list itself, nobody else
2395 * will try and migrate the task, hence the rq should match the
2396 * cpu we just moved it to.
2398 WARN_ON(task_cpu(p) != cpu);
2399 WARN_ON(p->state != TASK_WAKING);
2401 #ifdef CONFIG_SCHEDSTATS
2402 schedstat_inc(rq, ttwu_count);
2403 if (cpu == this_cpu)
2404 schedstat_inc(rq, ttwu_local);
2406 struct sched_domain *sd;
2407 for_each_domain(this_cpu, sd) {
2408 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2409 schedstat_inc(sd, ttwu_wake_remote);
2414 #endif /* CONFIG_SCHEDSTATS */
2417 #endif /* CONFIG_SMP */
2418 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2419 cpu == this_cpu, en_flags);
2422 ttwu_post_activation(p, rq, wake_flags, success);
2424 task_rq_unlock(rq, &flags);
2431 * try_to_wake_up_local - try to wake up a local task with rq lock held
2432 * @p: the thread to be awakened
2434 * Put @p on the run-queue if it's not alredy there. The caller must
2435 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2436 * the current task. this_rq() stays locked over invocation.
2438 static void try_to_wake_up_local(struct task_struct *p)
2440 struct rq *rq = task_rq(p);
2441 bool success = false;
2443 BUG_ON(rq != this_rq());
2444 BUG_ON(p == current);
2445 lockdep_assert_held(&rq->lock);
2447 if (!(p->state & TASK_NORMAL))
2451 if (likely(!task_running(rq, p))) {
2452 schedstat_inc(rq, ttwu_count);
2453 schedstat_inc(rq, ttwu_local);
2455 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2458 ttwu_post_activation(p, rq, 0, success);
2462 * wake_up_process - Wake up a specific process
2463 * @p: The process to be woken up.
2465 * Attempt to wake up the nominated process and move it to the set of runnable
2466 * processes. Returns 1 if the process was woken up, 0 if it was already
2469 * It may be assumed that this function implies a write memory barrier before
2470 * changing the task state if and only if any tasks are woken up.
2472 int wake_up_process(struct task_struct *p)
2474 return try_to_wake_up(p, TASK_ALL, 0);
2476 EXPORT_SYMBOL(wake_up_process);
2478 int wake_up_state(struct task_struct *p, unsigned int state)
2480 return try_to_wake_up(p, state, 0);
2484 * Perform scheduler related setup for a newly forked process p.
2485 * p is forked by current.
2487 * __sched_fork() is basic setup used by init_idle() too:
2489 static void __sched_fork(struct task_struct *p)
2491 p->se.exec_start = 0;
2492 p->se.sum_exec_runtime = 0;
2493 p->se.prev_sum_exec_runtime = 0;
2494 p->se.nr_migrations = 0;
2496 #ifdef CONFIG_SCHEDSTATS
2497 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2500 INIT_LIST_HEAD(&p->rt.run_list);
2502 INIT_LIST_HEAD(&p->se.group_node);
2504 #ifdef CONFIG_PREEMPT_NOTIFIERS
2505 INIT_HLIST_HEAD(&p->preempt_notifiers);
2510 * fork()/clone()-time setup:
2512 void sched_fork(struct task_struct *p, int clone_flags)
2514 int cpu = get_cpu();
2518 * We mark the process as running here. This guarantees that
2519 * nobody will actually run it, and a signal or other external
2520 * event cannot wake it up and insert it on the runqueue either.
2522 p->state = TASK_RUNNING;
2525 * Revert to default priority/policy on fork if requested.
2527 if (unlikely(p->sched_reset_on_fork)) {
2528 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2529 p->policy = SCHED_NORMAL;
2530 p->normal_prio = p->static_prio;
2533 if (PRIO_TO_NICE(p->static_prio) < 0) {
2534 p->static_prio = NICE_TO_PRIO(0);
2535 p->normal_prio = p->static_prio;
2540 * We don't need the reset flag anymore after the fork. It has
2541 * fulfilled its duty:
2543 p->sched_reset_on_fork = 0;
2547 * Make sure we do not leak PI boosting priority to the child.
2549 p->prio = current->normal_prio;
2551 if (!rt_prio(p->prio))
2552 p->sched_class = &fair_sched_class;
2554 if (p->sched_class->task_fork)
2555 p->sched_class->task_fork(p);
2558 * The child is not yet in the pid-hash so no cgroup attach races,
2559 * and the cgroup is pinned to this child due to cgroup_fork()
2560 * is ran before sched_fork().
2562 * Silence PROVE_RCU.
2565 set_task_cpu(p, cpu);
2568 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2569 if (likely(sched_info_on()))
2570 memset(&p->sched_info, 0, sizeof(p->sched_info));
2572 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2575 #ifdef CONFIG_PREEMPT
2576 /* Want to start with kernel preemption disabled. */
2577 task_thread_info(p)->preempt_count = 1;
2579 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2585 * wake_up_new_task - wake up a newly created task for the first time.
2587 * This function will do some initial scheduler statistics housekeeping
2588 * that must be done for every newly created context, then puts the task
2589 * on the runqueue and wakes it.
2591 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2593 unsigned long flags;
2595 int cpu __maybe_unused = get_cpu();
2598 rq = task_rq_lock(p, &flags);
2599 p->state = TASK_WAKING;
2602 * Fork balancing, do it here and not earlier because:
2603 * - cpus_allowed can change in the fork path
2604 * - any previously selected cpu might disappear through hotplug
2606 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2607 * without people poking at ->cpus_allowed.
2609 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2610 set_task_cpu(p, cpu);
2612 p->state = TASK_RUNNING;
2613 task_rq_unlock(rq, &flags);
2616 rq = task_rq_lock(p, &flags);
2617 activate_task(rq, p, 0);
2618 trace_sched_wakeup_new(p, 1);
2619 check_preempt_curr(rq, p, WF_FORK);
2621 if (p->sched_class->task_woken)
2622 p->sched_class->task_woken(rq, p);
2624 task_rq_unlock(rq, &flags);
2628 #ifdef CONFIG_PREEMPT_NOTIFIERS
2631 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2632 * @notifier: notifier struct to register
2634 void preempt_notifier_register(struct preempt_notifier *notifier)
2636 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2638 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2641 * preempt_notifier_unregister - no longer interested in preemption notifications
2642 * @notifier: notifier struct to unregister
2644 * This is safe to call from within a preemption notifier.
2646 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2648 hlist_del(¬ifier->link);
2650 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2652 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2654 struct preempt_notifier *notifier;
2655 struct hlist_node *node;
2657 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2658 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2662 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2663 struct task_struct *next)
2665 struct preempt_notifier *notifier;
2666 struct hlist_node *node;
2668 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2669 notifier->ops->sched_out(notifier, next);
2672 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2674 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2679 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2680 struct task_struct *next)
2684 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2687 * prepare_task_switch - prepare to switch tasks
2688 * @rq: the runqueue preparing to switch
2689 * @prev: the current task that is being switched out
2690 * @next: the task we are going to switch to.
2692 * This is called with the rq lock held and interrupts off. It must
2693 * be paired with a subsequent finish_task_switch after the context
2696 * prepare_task_switch sets up locking and calls architecture specific
2700 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2701 struct task_struct *next)
2703 fire_sched_out_preempt_notifiers(prev, next);
2704 prepare_lock_switch(rq, next);
2705 prepare_arch_switch(next);
2709 * finish_task_switch - clean up after a task-switch
2710 * @rq: runqueue associated with task-switch
2711 * @prev: the thread we just switched away from.
2713 * finish_task_switch must be called after the context switch, paired
2714 * with a prepare_task_switch call before the context switch.
2715 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2716 * and do any other architecture-specific cleanup actions.
2718 * Note that we may have delayed dropping an mm in context_switch(). If
2719 * so, we finish that here outside of the runqueue lock. (Doing it
2720 * with the lock held can cause deadlocks; see schedule() for
2723 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2724 __releases(rq->lock)
2726 struct mm_struct *mm = rq->prev_mm;
2732 * A task struct has one reference for the use as "current".
2733 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2734 * schedule one last time. The schedule call will never return, and
2735 * the scheduled task must drop that reference.
2736 * The test for TASK_DEAD must occur while the runqueue locks are
2737 * still held, otherwise prev could be scheduled on another cpu, die
2738 * there before we look at prev->state, and then the reference would
2740 * Manfred Spraul <manfred@colorfullife.com>
2742 prev_state = prev->state;
2743 finish_arch_switch(prev);
2744 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2745 local_irq_disable();
2746 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2747 perf_event_task_sched_in(current);
2748 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2750 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2751 finish_lock_switch(rq, prev);
2753 fire_sched_in_preempt_notifiers(current);
2756 if (unlikely(prev_state == TASK_DEAD)) {
2758 * Remove function-return probe instances associated with this
2759 * task and put them back on the free list.
2761 kprobe_flush_task(prev);
2762 put_task_struct(prev);
2768 /* assumes rq->lock is held */
2769 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2771 if (prev->sched_class->pre_schedule)
2772 prev->sched_class->pre_schedule(rq, prev);
2775 /* rq->lock is NOT held, but preemption is disabled */
2776 static inline void post_schedule(struct rq *rq)
2778 if (rq->post_schedule) {
2779 unsigned long flags;
2781 raw_spin_lock_irqsave(&rq->lock, flags);
2782 if (rq->curr->sched_class->post_schedule)
2783 rq->curr->sched_class->post_schedule(rq);
2784 raw_spin_unlock_irqrestore(&rq->lock, flags);
2786 rq->post_schedule = 0;
2792 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2796 static inline void post_schedule(struct rq *rq)
2803 * schedule_tail - first thing a freshly forked thread must call.
2804 * @prev: the thread we just switched away from.
2806 asmlinkage void schedule_tail(struct task_struct *prev)
2807 __releases(rq->lock)
2809 struct rq *rq = this_rq();
2811 finish_task_switch(rq, prev);
2814 * FIXME: do we need to worry about rq being invalidated by the
2819 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2820 /* In this case, finish_task_switch does not reenable preemption */
2823 if (current->set_child_tid)
2824 put_user(task_pid_vnr(current), current->set_child_tid);
2828 * context_switch - switch to the new MM and the new
2829 * thread's register state.
2832 context_switch(struct rq *rq, struct task_struct *prev,
2833 struct task_struct *next)
2835 struct mm_struct *mm, *oldmm;
2837 prepare_task_switch(rq, prev, next);
2838 trace_sched_switch(prev, next);
2840 oldmm = prev->active_mm;
2842 * For paravirt, this is coupled with an exit in switch_to to
2843 * combine the page table reload and the switch backend into
2846 arch_start_context_switch(prev);
2849 next->active_mm = oldmm;
2850 atomic_inc(&oldmm->mm_count);
2851 enter_lazy_tlb(oldmm, next);
2853 switch_mm(oldmm, mm, next);
2856 prev->active_mm = NULL;
2857 rq->prev_mm = oldmm;
2860 * Since the runqueue lock will be released by the next
2861 * task (which is an invalid locking op but in the case
2862 * of the scheduler it's an obvious special-case), so we
2863 * do an early lockdep release here:
2865 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2866 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2869 /* Here we just switch the register state and the stack. */
2870 switch_to(prev, next, prev);
2874 * this_rq must be evaluated again because prev may have moved
2875 * CPUs since it called schedule(), thus the 'rq' on its stack
2876 * frame will be invalid.
2878 finish_task_switch(this_rq(), prev);
2882 * nr_running, nr_uninterruptible and nr_context_switches:
2884 * externally visible scheduler statistics: current number of runnable
2885 * threads, current number of uninterruptible-sleeping threads, total
2886 * number of context switches performed since bootup.
2888 unsigned long nr_running(void)
2890 unsigned long i, sum = 0;
2892 for_each_online_cpu(i)
2893 sum += cpu_rq(i)->nr_running;
2898 unsigned long nr_uninterruptible(void)
2900 unsigned long i, sum = 0;
2902 for_each_possible_cpu(i)
2903 sum += cpu_rq(i)->nr_uninterruptible;
2906 * Since we read the counters lockless, it might be slightly
2907 * inaccurate. Do not allow it to go below zero though:
2909 if (unlikely((long)sum < 0))
2915 unsigned long long nr_context_switches(void)
2918 unsigned long long sum = 0;
2920 for_each_possible_cpu(i)
2921 sum += cpu_rq(i)->nr_switches;
2926 unsigned long nr_iowait(void)
2928 unsigned long i, sum = 0;
2930 for_each_possible_cpu(i)
2931 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2936 unsigned long nr_iowait_cpu(int cpu)
2938 struct rq *this = cpu_rq(cpu);
2939 return atomic_read(&this->nr_iowait);
2942 unsigned long this_cpu_load(void)
2944 struct rq *this = this_rq();
2945 return this->cpu_load[0];
2949 /* Variables and functions for calc_load */
2950 static atomic_long_t calc_load_tasks;
2951 static unsigned long calc_load_update;
2952 unsigned long avenrun[3];
2953 EXPORT_SYMBOL(avenrun);
2955 static long calc_load_fold_active(struct rq *this_rq)
2957 long nr_active, delta = 0;
2959 nr_active = this_rq->nr_running;
2960 nr_active += (long) this_rq->nr_uninterruptible;
2962 if (nr_active != this_rq->calc_load_active) {
2963 delta = nr_active - this_rq->calc_load_active;
2964 this_rq->calc_load_active = nr_active;
2972 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2974 * When making the ILB scale, we should try to pull this in as well.
2976 static atomic_long_t calc_load_tasks_idle;
2978 static void calc_load_account_idle(struct rq *this_rq)
2982 delta = calc_load_fold_active(this_rq);
2984 atomic_long_add(delta, &calc_load_tasks_idle);
2987 static long calc_load_fold_idle(void)
2992 * Its got a race, we don't care...
2994 if (atomic_long_read(&calc_load_tasks_idle))
2995 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3000 static void calc_load_account_idle(struct rq *this_rq)
3004 static inline long calc_load_fold_idle(void)
3011 * get_avenrun - get the load average array
3012 * @loads: pointer to dest load array
3013 * @offset: offset to add
3014 * @shift: shift count to shift the result left
3016 * These values are estimates at best, so no need for locking.
3018 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3020 loads[0] = (avenrun[0] + offset) << shift;
3021 loads[1] = (avenrun[1] + offset) << shift;
3022 loads[2] = (avenrun[2] + offset) << shift;
3025 static unsigned long
3026 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3029 load += active * (FIXED_1 - exp);
3030 return load >> FSHIFT;
3034 * calc_load - update the avenrun load estimates 10 ticks after the
3035 * CPUs have updated calc_load_tasks.
3037 void calc_global_load(void)
3039 unsigned long upd = calc_load_update + 10;
3042 if (time_before(jiffies, upd))
3045 active = atomic_long_read(&calc_load_tasks);
3046 active = active > 0 ? active * FIXED_1 : 0;
3048 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3049 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3050 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3052 calc_load_update += LOAD_FREQ;
3056 * Called from update_cpu_load() to periodically update this CPU's
3059 static void calc_load_account_active(struct rq *this_rq)
3063 if (time_before(jiffies, this_rq->calc_load_update))
3066 delta = calc_load_fold_active(this_rq);
3067 delta += calc_load_fold_idle();
3069 atomic_long_add(delta, &calc_load_tasks);
3071 this_rq->calc_load_update += LOAD_FREQ;
3075 * The exact cpuload at various idx values, calculated at every tick would be
3076 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3078 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3079 * on nth tick when cpu may be busy, then we have:
3080 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3081 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3083 * decay_load_missed() below does efficient calculation of
3084 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3085 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3087 * The calculation is approximated on a 128 point scale.
3088 * degrade_zero_ticks is the number of ticks after which load at any
3089 * particular idx is approximated to be zero.
3090 * degrade_factor is a precomputed table, a row for each load idx.
3091 * Each column corresponds to degradation factor for a power of two ticks,
3092 * based on 128 point scale.
3094 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3095 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3097 * With this power of 2 load factors, we can degrade the load n times
3098 * by looking at 1 bits in n and doing as many mult/shift instead of
3099 * n mult/shifts needed by the exact degradation.
3101 #define DEGRADE_SHIFT 7
3102 static const unsigned char
3103 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3104 static const unsigned char
3105 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3106 {0, 0, 0, 0, 0, 0, 0, 0},
3107 {64, 32, 8, 0, 0, 0, 0, 0},
3108 {96, 72, 40, 12, 1, 0, 0},
3109 {112, 98, 75, 43, 15, 1, 0},
3110 {120, 112, 98, 76, 45, 16, 2} };
3113 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3114 * would be when CPU is idle and so we just decay the old load without
3115 * adding any new load.
3117 static unsigned long
3118 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3122 if (!missed_updates)
3125 if (missed_updates >= degrade_zero_ticks[idx])
3129 return load >> missed_updates;
3131 while (missed_updates) {
3132 if (missed_updates % 2)
3133 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3135 missed_updates >>= 1;
3142 * Update rq->cpu_load[] statistics. This function is usually called every
3143 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3144 * every tick. We fix it up based on jiffies.
3146 static void update_cpu_load(struct rq *this_rq)
3148 unsigned long this_load = this_rq->load.weight;
3149 unsigned long curr_jiffies = jiffies;
3150 unsigned long pending_updates;
3153 this_rq->nr_load_updates++;
3155 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3156 if (curr_jiffies == this_rq->last_load_update_tick)
3159 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3160 this_rq->last_load_update_tick = curr_jiffies;
3162 /* Update our load: */
3163 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3164 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3165 unsigned long old_load, new_load;
3167 /* scale is effectively 1 << i now, and >> i divides by scale */
3169 old_load = this_rq->cpu_load[i];
3170 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3171 new_load = this_load;
3173 * Round up the averaging division if load is increasing. This
3174 * prevents us from getting stuck on 9 if the load is 10, for
3177 if (new_load > old_load)
3178 new_load += scale - 1;
3180 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3184 static void update_cpu_load_active(struct rq *this_rq)
3186 update_cpu_load(this_rq);
3188 calc_load_account_active(this_rq);
3194 * sched_exec - execve() is a valuable balancing opportunity, because at
3195 * this point the task has the smallest effective memory and cache footprint.
3197 void sched_exec(void)
3199 struct task_struct *p = current;
3200 unsigned long flags;
3204 rq = task_rq_lock(p, &flags);
3205 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3206 if (dest_cpu == smp_processor_id())
3210 * select_task_rq() can race against ->cpus_allowed
3212 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3213 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3214 struct migration_arg arg = { p, dest_cpu };
3216 task_rq_unlock(rq, &flags);
3217 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3221 task_rq_unlock(rq, &flags);
3226 DEFINE_PER_CPU(struct kernel_stat, kstat);
3228 EXPORT_PER_CPU_SYMBOL(kstat);
3231 * Return any ns on the sched_clock that have not yet been accounted in
3232 * @p in case that task is currently running.
3234 * Called with task_rq_lock() held on @rq.
3236 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3240 if (task_current(rq, p)) {
3241 update_rq_clock(rq);
3242 ns = rq->clock - p->se.exec_start;
3250 unsigned long long task_delta_exec(struct task_struct *p)
3252 unsigned long flags;
3256 rq = task_rq_lock(p, &flags);
3257 ns = do_task_delta_exec(p, rq);
3258 task_rq_unlock(rq, &flags);
3264 * Return accounted runtime for the task.
3265 * In case the task is currently running, return the runtime plus current's
3266 * pending runtime that have not been accounted yet.
3268 unsigned long long task_sched_runtime(struct task_struct *p)
3270 unsigned long flags;
3274 rq = task_rq_lock(p, &flags);
3275 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3276 task_rq_unlock(rq, &flags);
3282 * Return sum_exec_runtime for the thread group.
3283 * In case the task is currently running, return the sum plus current's
3284 * pending runtime that have not been accounted yet.
3286 * Note that the thread group might have other running tasks as well,
3287 * so the return value not includes other pending runtime that other
3288 * running tasks might have.
3290 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3292 struct task_cputime totals;
3293 unsigned long flags;
3297 rq = task_rq_lock(p, &flags);
3298 thread_group_cputime(p, &totals);
3299 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3300 task_rq_unlock(rq, &flags);
3306 * Account user cpu time to a process.
3307 * @p: the process that the cpu time gets accounted to
3308 * @cputime: the cpu time spent in user space since the last update
3309 * @cputime_scaled: cputime scaled by cpu frequency
3311 void account_user_time(struct task_struct *p, cputime_t cputime,
3312 cputime_t cputime_scaled)
3314 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3317 /* Add user time to process. */
3318 p->utime = cputime_add(p->utime, cputime);
3319 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3320 account_group_user_time(p, cputime);
3322 /* Add user time to cpustat. */
3323 tmp = cputime_to_cputime64(cputime);
3324 if (TASK_NICE(p) > 0)
3325 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3327 cpustat->user = cputime64_add(cpustat->user, tmp);
3329 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3330 /* Account for user time used */
3331 acct_update_integrals(p);
3335 * Account guest cpu time to a process.
3336 * @p: the process that the cpu time gets accounted to
3337 * @cputime: the cpu time spent in virtual machine since the last update
3338 * @cputime_scaled: cputime scaled by cpu frequency
3340 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3341 cputime_t cputime_scaled)
3344 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3346 tmp = cputime_to_cputime64(cputime);
3348 /* Add guest time to process. */
3349 p->utime = cputime_add(p->utime, cputime);
3350 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3351 account_group_user_time(p, cputime);
3352 p->gtime = cputime_add(p->gtime, cputime);
3354 /* Add guest time to cpustat. */
3355 if (TASK_NICE(p) > 0) {
3356 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3357 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3359 cpustat->user = cputime64_add(cpustat->user, tmp);
3360 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3365 * Account system cpu time to a process.
3366 * @p: the process that the cpu time gets accounted to
3367 * @hardirq_offset: the offset to subtract from hardirq_count()
3368 * @cputime: the cpu time spent in kernel space since the last update
3369 * @cputime_scaled: cputime scaled by cpu frequency
3371 void account_system_time(struct task_struct *p, int hardirq_offset,
3372 cputime_t cputime, cputime_t cputime_scaled)
3374 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3377 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3378 account_guest_time(p, cputime, cputime_scaled);
3382 /* Add system time to process. */
3383 p->stime = cputime_add(p->stime, cputime);
3384 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3385 account_group_system_time(p, cputime);
3387 /* Add system time to cpustat. */
3388 tmp = cputime_to_cputime64(cputime);
3389 if (hardirq_count() - hardirq_offset)
3390 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3391 else if (softirq_count())
3392 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3394 cpustat->system = cputime64_add(cpustat->system, tmp);
3396 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3398 /* Account for system time used */
3399 acct_update_integrals(p);
3403 * Account for involuntary wait time.
3404 * @steal: the cpu time spent in involuntary wait
3406 void account_steal_time(cputime_t cputime)
3408 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3409 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3411 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3415 * Account for idle time.
3416 * @cputime: the cpu time spent in idle wait
3418 void account_idle_time(cputime_t cputime)
3420 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3421 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3422 struct rq *rq = this_rq();
3424 if (atomic_read(&rq->nr_iowait) > 0)
3425 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3427 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3430 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3433 * Account a single tick of cpu time.
3434 * @p: the process that the cpu time gets accounted to
3435 * @user_tick: indicates if the tick is a user or a system tick
3437 void account_process_tick(struct task_struct *p, int user_tick)
3439 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3440 struct rq *rq = this_rq();
3443 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3444 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3445 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3448 account_idle_time(cputime_one_jiffy);
3452 * Account multiple ticks of steal time.
3453 * @p: the process from which the cpu time has been stolen
3454 * @ticks: number of stolen ticks
3456 void account_steal_ticks(unsigned long ticks)
3458 account_steal_time(jiffies_to_cputime(ticks));
3462 * Account multiple ticks of idle time.
3463 * @ticks: number of stolen ticks
3465 void account_idle_ticks(unsigned long ticks)
3467 account_idle_time(jiffies_to_cputime(ticks));
3473 * Use precise platform statistics if available:
3475 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3476 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3482 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3484 struct task_cputime cputime;
3486 thread_group_cputime(p, &cputime);
3488 *ut = cputime.utime;
3489 *st = cputime.stime;
3493 #ifndef nsecs_to_cputime
3494 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3497 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3499 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3502 * Use CFS's precise accounting:
3504 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3509 temp = (u64)(rtime * utime);
3510 do_div(temp, total);
3511 utime = (cputime_t)temp;
3516 * Compare with previous values, to keep monotonicity:
3518 p->prev_utime = max(p->prev_utime, utime);
3519 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3521 *ut = p->prev_utime;
3522 *st = p->prev_stime;
3526 * Must be called with siglock held.
3528 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3530 struct signal_struct *sig = p->signal;
3531 struct task_cputime cputime;
3532 cputime_t rtime, utime, total;
3534 thread_group_cputime(p, &cputime);
3536 total = cputime_add(cputime.utime, cputime.stime);
3537 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3542 temp = (u64)(rtime * cputime.utime);
3543 do_div(temp, total);
3544 utime = (cputime_t)temp;
3548 sig->prev_utime = max(sig->prev_utime, utime);
3549 sig->prev_stime = max(sig->prev_stime,
3550 cputime_sub(rtime, sig->prev_utime));
3552 *ut = sig->prev_utime;
3553 *st = sig->prev_stime;
3558 * This function gets called by the timer code, with HZ frequency.
3559 * We call it with interrupts disabled.
3561 * It also gets called by the fork code, when changing the parent's
3564 void scheduler_tick(void)
3566 int cpu = smp_processor_id();
3567 struct rq *rq = cpu_rq(cpu);
3568 struct task_struct *curr = rq->curr;
3572 raw_spin_lock(&rq->lock);
3573 update_rq_clock(rq);
3574 update_cpu_load_active(rq);
3575 curr->sched_class->task_tick(rq, curr, 0);
3576 raw_spin_unlock(&rq->lock);
3578 perf_event_task_tick(curr);
3581 rq->idle_at_tick = idle_cpu(cpu);
3582 trigger_load_balance(rq, cpu);
3586 notrace unsigned long get_parent_ip(unsigned long addr)
3588 if (in_lock_functions(addr)) {
3589 addr = CALLER_ADDR2;
3590 if (in_lock_functions(addr))
3591 addr = CALLER_ADDR3;
3596 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3597 defined(CONFIG_PREEMPT_TRACER))
3599 void __kprobes add_preempt_count(int val)
3601 #ifdef CONFIG_DEBUG_PREEMPT
3605 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3608 preempt_count() += val;
3609 #ifdef CONFIG_DEBUG_PREEMPT
3611 * Spinlock count overflowing soon?
3613 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3616 if (preempt_count() == val)
3617 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3619 EXPORT_SYMBOL(add_preempt_count);
3621 void __kprobes sub_preempt_count(int val)
3623 #ifdef CONFIG_DEBUG_PREEMPT
3627 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3630 * Is the spinlock portion underflowing?
3632 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3633 !(preempt_count() & PREEMPT_MASK)))
3637 if (preempt_count() == val)
3638 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3639 preempt_count() -= val;
3641 EXPORT_SYMBOL(sub_preempt_count);
3646 * Print scheduling while atomic bug:
3648 static noinline void __schedule_bug(struct task_struct *prev)
3650 struct pt_regs *regs = get_irq_regs();
3652 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3653 prev->comm, prev->pid, preempt_count());
3655 debug_show_held_locks(prev);
3657 if (irqs_disabled())
3658 print_irqtrace_events(prev);
3667 * Various schedule()-time debugging checks and statistics:
3669 static inline void schedule_debug(struct task_struct *prev)
3672 * Test if we are atomic. Since do_exit() needs to call into
3673 * schedule() atomically, we ignore that path for now.
3674 * Otherwise, whine if we are scheduling when we should not be.
3676 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3677 __schedule_bug(prev);
3679 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3681 schedstat_inc(this_rq(), sched_count);
3682 #ifdef CONFIG_SCHEDSTATS
3683 if (unlikely(prev->lock_depth >= 0)) {
3684 schedstat_inc(this_rq(), bkl_count);
3685 schedstat_inc(prev, sched_info.bkl_count);
3690 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3693 update_rq_clock(rq);
3694 rq->skip_clock_update = 0;
3695 prev->sched_class->put_prev_task(rq, prev);
3699 * Pick up the highest-prio task:
3701 static inline struct task_struct *
3702 pick_next_task(struct rq *rq)
3704 const struct sched_class *class;
3705 struct task_struct *p;
3708 * Optimization: we know that if all tasks are in
3709 * the fair class we can call that function directly:
3711 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3712 p = fair_sched_class.pick_next_task(rq);
3717 class = sched_class_highest;
3719 p = class->pick_next_task(rq);
3723 * Will never be NULL as the idle class always
3724 * returns a non-NULL p:
3726 class = class->next;
3731 * schedule() is the main scheduler function.
3733 asmlinkage void __sched schedule(void)
3735 struct task_struct *prev, *next;
3736 unsigned long *switch_count;
3742 cpu = smp_processor_id();
3744 rcu_note_context_switch(cpu);
3747 release_kernel_lock(prev);
3748 need_resched_nonpreemptible:
3750 schedule_debug(prev);
3752 if (sched_feat(HRTICK))
3755 raw_spin_lock_irq(&rq->lock);
3756 clear_tsk_need_resched(prev);
3758 switch_count = &prev->nivcsw;
3759 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3760 if (unlikely(signal_pending_state(prev->state, prev))) {
3761 prev->state = TASK_RUNNING;
3764 * If a worker is going to sleep, notify and
3765 * ask workqueue whether it wants to wake up a
3766 * task to maintain concurrency. If so, wake
3769 if (prev->flags & PF_WQ_WORKER) {
3770 struct task_struct *to_wakeup;
3772 to_wakeup = wq_worker_sleeping(prev, cpu);
3774 try_to_wake_up_local(to_wakeup);
3776 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3778 switch_count = &prev->nvcsw;
3781 pre_schedule(rq, prev);
3783 if (unlikely(!rq->nr_running))
3784 idle_balance(cpu, rq);
3786 put_prev_task(rq, prev);
3787 next = pick_next_task(rq);
3789 if (likely(prev != next)) {
3790 sched_info_switch(prev, next);
3791 perf_event_task_sched_out(prev, next);
3797 context_switch(rq, prev, next); /* unlocks the rq */
3799 * The context switch have flipped the stack from under us
3800 * and restored the local variables which were saved when
3801 * this task called schedule() in the past. prev == current
3802 * is still correct, but it can be moved to another cpu/rq.
3804 cpu = smp_processor_id();
3807 raw_spin_unlock_irq(&rq->lock);
3811 if (unlikely(reacquire_kernel_lock(prev)))
3812 goto need_resched_nonpreemptible;
3814 preempt_enable_no_resched();
3818 EXPORT_SYMBOL(schedule);
3820 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3822 * Look out! "owner" is an entirely speculative pointer
3823 * access and not reliable.
3825 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3830 if (!sched_feat(OWNER_SPIN))
3833 #ifdef CONFIG_DEBUG_PAGEALLOC
3835 * Need to access the cpu field knowing that
3836 * DEBUG_PAGEALLOC could have unmapped it if
3837 * the mutex owner just released it and exited.
3839 if (probe_kernel_address(&owner->cpu, cpu))
3846 * Even if the access succeeded (likely case),
3847 * the cpu field may no longer be valid.
3849 if (cpu >= nr_cpumask_bits)
3853 * We need to validate that we can do a
3854 * get_cpu() and that we have the percpu area.
3856 if (!cpu_online(cpu))
3863 * Owner changed, break to re-assess state.
3865 if (lock->owner != owner) {
3867 * If the lock has switched to a different owner,
3868 * we likely have heavy contention. Return 0 to quit
3869 * optimistic spinning and not contend further:
3877 * Is that owner really running on that cpu?
3879 if (task_thread_info(rq->curr) != owner || need_resched())
3889 #ifdef CONFIG_PREEMPT
3891 * this is the entry point to schedule() from in-kernel preemption
3892 * off of preempt_enable. Kernel preemptions off return from interrupt
3893 * occur there and call schedule directly.
3895 asmlinkage void __sched notrace preempt_schedule(void)
3897 struct thread_info *ti = current_thread_info();
3900 * If there is a non-zero preempt_count or interrupts are disabled,
3901 * we do not want to preempt the current task. Just return..
3903 if (likely(ti->preempt_count || irqs_disabled()))
3907 add_preempt_count_notrace(PREEMPT_ACTIVE);
3909 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3912 * Check again in case we missed a preemption opportunity
3913 * between schedule and now.
3916 } while (need_resched());
3918 EXPORT_SYMBOL(preempt_schedule);
3921 * this is the entry point to schedule() from kernel preemption
3922 * off of irq context.
3923 * Note, that this is called and return with irqs disabled. This will
3924 * protect us against recursive calling from irq.
3926 asmlinkage void __sched preempt_schedule_irq(void)
3928 struct thread_info *ti = current_thread_info();
3930 /* Catch callers which need to be fixed */
3931 BUG_ON(ti->preempt_count || !irqs_disabled());
3934 add_preempt_count(PREEMPT_ACTIVE);
3937 local_irq_disable();
3938 sub_preempt_count(PREEMPT_ACTIVE);
3941 * Check again in case we missed a preemption opportunity
3942 * between schedule and now.
3945 } while (need_resched());
3948 #endif /* CONFIG_PREEMPT */
3950 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3953 return try_to_wake_up(curr->private, mode, wake_flags);
3955 EXPORT_SYMBOL(default_wake_function);
3958 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3959 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3960 * number) then we wake all the non-exclusive tasks and one exclusive task.
3962 * There are circumstances in which we can try to wake a task which has already
3963 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3964 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3966 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3967 int nr_exclusive, int wake_flags, void *key)
3969 wait_queue_t *curr, *next;
3971 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3972 unsigned flags = curr->flags;
3974 if (curr->func(curr, mode, wake_flags, key) &&
3975 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3981 * __wake_up - wake up threads blocked on a waitqueue.
3983 * @mode: which threads
3984 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3985 * @key: is directly passed to the wakeup function
3987 * It may be assumed that this function implies a write memory barrier before
3988 * changing the task state if and only if any tasks are woken up.
3990 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3991 int nr_exclusive, void *key)
3993 unsigned long flags;
3995 spin_lock_irqsave(&q->lock, flags);
3996 __wake_up_common(q, mode, nr_exclusive, 0, key);
3997 spin_unlock_irqrestore(&q->lock, flags);
3999 EXPORT_SYMBOL(__wake_up);
4002 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4004 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4006 __wake_up_common(q, mode, 1, 0, NULL);
4008 EXPORT_SYMBOL_GPL(__wake_up_locked);
4010 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4012 __wake_up_common(q, mode, 1, 0, key);
4016 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4018 * @mode: which threads
4019 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4020 * @key: opaque value to be passed to wakeup targets
4022 * The sync wakeup differs that the waker knows that it will schedule
4023 * away soon, so while the target thread will be woken up, it will not
4024 * be migrated to another CPU - ie. the two threads are 'synchronized'
4025 * with each other. This can prevent needless bouncing between CPUs.
4027 * On UP it can prevent extra preemption.
4029 * It may be assumed that this function implies a write memory barrier before
4030 * changing the task state if and only if any tasks are woken up.
4032 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4033 int nr_exclusive, void *key)
4035 unsigned long flags;
4036 int wake_flags = WF_SYNC;
4041 if (unlikely(!nr_exclusive))
4044 spin_lock_irqsave(&q->lock, flags);
4045 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4046 spin_unlock_irqrestore(&q->lock, flags);
4048 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4051 * __wake_up_sync - see __wake_up_sync_key()
4053 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4055 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4057 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4060 * complete: - signals a single thread waiting on this completion
4061 * @x: holds the state of this particular completion
4063 * This will wake up a single thread waiting on this completion. Threads will be
4064 * awakened in the same order in which they were queued.
4066 * See also complete_all(), wait_for_completion() and related routines.
4068 * It may be assumed that this function implies a write memory barrier before
4069 * changing the task state if and only if any tasks are woken up.
4071 void complete(struct completion *x)
4073 unsigned long flags;
4075 spin_lock_irqsave(&x->wait.lock, flags);
4077 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4078 spin_unlock_irqrestore(&x->wait.lock, flags);
4080 EXPORT_SYMBOL(complete);
4083 * complete_all: - signals all threads waiting on this completion
4084 * @x: holds the state of this particular completion
4086 * This will wake up all threads waiting on this particular completion event.
4088 * It may be assumed that this function implies a write memory barrier before
4089 * changing the task state if and only if any tasks are woken up.
4091 void complete_all(struct completion *x)
4093 unsigned long flags;
4095 spin_lock_irqsave(&x->wait.lock, flags);
4096 x->done += UINT_MAX/2;
4097 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4098 spin_unlock_irqrestore(&x->wait.lock, flags);
4100 EXPORT_SYMBOL(complete_all);
4102 static inline long __sched
4103 do_wait_for_common(struct completion *x, long timeout, int state)
4106 DECLARE_WAITQUEUE(wait, current);
4108 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4110 if (signal_pending_state(state, current)) {
4111 timeout = -ERESTARTSYS;
4114 __set_current_state(state);
4115 spin_unlock_irq(&x->wait.lock);
4116 timeout = schedule_timeout(timeout);
4117 spin_lock_irq(&x->wait.lock);
4118 } while (!x->done && timeout);
4119 __remove_wait_queue(&x->wait, &wait);
4124 return timeout ?: 1;
4128 wait_for_common(struct completion *x, long timeout, int state)
4132 spin_lock_irq(&x->wait.lock);
4133 timeout = do_wait_for_common(x, timeout, state);
4134 spin_unlock_irq(&x->wait.lock);
4139 * wait_for_completion: - waits for completion of a task
4140 * @x: holds the state of this particular completion
4142 * This waits to be signaled for completion of a specific task. It is NOT
4143 * interruptible and there is no timeout.
4145 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4146 * and interrupt capability. Also see complete().
4148 void __sched wait_for_completion(struct completion *x)
4150 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4152 EXPORT_SYMBOL(wait_for_completion);
4155 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4156 * @x: holds the state of this particular completion
4157 * @timeout: timeout value in jiffies
4159 * This waits for either a completion of a specific task to be signaled or for a
4160 * specified timeout to expire. The timeout is in jiffies. It is not
4163 unsigned long __sched
4164 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4166 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4168 EXPORT_SYMBOL(wait_for_completion_timeout);
4171 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4172 * @x: holds the state of this particular completion
4174 * This waits for completion of a specific task to be signaled. It is
4177 int __sched wait_for_completion_interruptible(struct completion *x)
4179 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4180 if (t == -ERESTARTSYS)
4184 EXPORT_SYMBOL(wait_for_completion_interruptible);
4187 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4188 * @x: holds the state of this particular completion
4189 * @timeout: timeout value in jiffies
4191 * This waits for either a completion of a specific task to be signaled or for a
4192 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4194 unsigned long __sched
4195 wait_for_completion_interruptible_timeout(struct completion *x,
4196 unsigned long timeout)
4198 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4200 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4203 * wait_for_completion_killable: - waits for completion of a task (killable)
4204 * @x: holds the state of this particular completion
4206 * This waits to be signaled for completion of a specific task. It can be
4207 * interrupted by a kill signal.
4209 int __sched wait_for_completion_killable(struct completion *x)
4211 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4212 if (t == -ERESTARTSYS)
4216 EXPORT_SYMBOL(wait_for_completion_killable);
4219 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4220 * @x: holds the state of this particular completion
4221 * @timeout: timeout value in jiffies
4223 * This waits for either a completion of a specific task to be
4224 * signaled or for a specified timeout to expire. It can be
4225 * interrupted by a kill signal. The timeout is in jiffies.
4227 unsigned long __sched
4228 wait_for_completion_killable_timeout(struct completion *x,
4229 unsigned long timeout)
4231 return wait_for_common(x, timeout, TASK_KILLABLE);
4233 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4236 * try_wait_for_completion - try to decrement a completion without blocking
4237 * @x: completion structure
4239 * Returns: 0 if a decrement cannot be done without blocking
4240 * 1 if a decrement succeeded.
4242 * If a completion is being used as a counting completion,
4243 * attempt to decrement the counter without blocking. This
4244 * enables us to avoid waiting if the resource the completion
4245 * is protecting is not available.
4247 bool try_wait_for_completion(struct completion *x)
4249 unsigned long flags;
4252 spin_lock_irqsave(&x->wait.lock, flags);
4257 spin_unlock_irqrestore(&x->wait.lock, flags);
4260 EXPORT_SYMBOL(try_wait_for_completion);
4263 * completion_done - Test to see if a completion has any waiters
4264 * @x: completion structure
4266 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4267 * 1 if there are no waiters.
4270 bool completion_done(struct completion *x)
4272 unsigned long flags;
4275 spin_lock_irqsave(&x->wait.lock, flags);
4278 spin_unlock_irqrestore(&x->wait.lock, flags);
4281 EXPORT_SYMBOL(completion_done);
4284 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4286 unsigned long flags;
4289 init_waitqueue_entry(&wait, current);
4291 __set_current_state(state);
4293 spin_lock_irqsave(&q->lock, flags);
4294 __add_wait_queue(q, &wait);
4295 spin_unlock(&q->lock);
4296 timeout = schedule_timeout(timeout);
4297 spin_lock_irq(&q->lock);
4298 __remove_wait_queue(q, &wait);
4299 spin_unlock_irqrestore(&q->lock, flags);
4304 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4306 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4308 EXPORT_SYMBOL(interruptible_sleep_on);
4311 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4313 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4315 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4317 void __sched sleep_on(wait_queue_head_t *q)
4319 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4321 EXPORT_SYMBOL(sleep_on);
4323 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4325 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4327 EXPORT_SYMBOL(sleep_on_timeout);
4329 #ifdef CONFIG_RT_MUTEXES
4332 * rt_mutex_setprio - set the current priority of a task
4334 * @prio: prio value (kernel-internal form)
4336 * This function changes the 'effective' priority of a task. It does
4337 * not touch ->normal_prio like __setscheduler().
4339 * Used by the rt_mutex code to implement priority inheritance logic.
4341 void rt_mutex_setprio(struct task_struct *p, int prio)
4343 unsigned long flags;
4344 int oldprio, on_rq, running;
4346 const struct sched_class *prev_class;
4348 BUG_ON(prio < 0 || prio > MAX_PRIO);
4350 rq = task_rq_lock(p, &flags);
4353 prev_class = p->sched_class;
4354 on_rq = p->se.on_rq;
4355 running = task_current(rq, p);
4357 dequeue_task(rq, p, 0);
4359 p->sched_class->put_prev_task(rq, p);
4362 p->sched_class = &rt_sched_class;
4364 p->sched_class = &fair_sched_class;
4369 p->sched_class->set_curr_task(rq);
4371 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4373 check_class_changed(rq, p, prev_class, oldprio, running);
4375 task_rq_unlock(rq, &flags);
4380 void set_user_nice(struct task_struct *p, long nice)
4382 int old_prio, delta, on_rq;
4383 unsigned long flags;
4386 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4389 * We have to be careful, if called from sys_setpriority(),
4390 * the task might be in the middle of scheduling on another CPU.
4392 rq = task_rq_lock(p, &flags);
4394 * The RT priorities are set via sched_setscheduler(), but we still
4395 * allow the 'normal' nice value to be set - but as expected
4396 * it wont have any effect on scheduling until the task is
4397 * SCHED_FIFO/SCHED_RR:
4399 if (task_has_rt_policy(p)) {
4400 p->static_prio = NICE_TO_PRIO(nice);
4403 on_rq = p->se.on_rq;
4405 dequeue_task(rq, p, 0);
4407 p->static_prio = NICE_TO_PRIO(nice);
4410 p->prio = effective_prio(p);
4411 delta = p->prio - old_prio;
4414 enqueue_task(rq, p, 0);
4416 * If the task increased its priority or is running and
4417 * lowered its priority, then reschedule its CPU:
4419 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4420 resched_task(rq->curr);
4423 task_rq_unlock(rq, &flags);
4425 EXPORT_SYMBOL(set_user_nice);
4428 * can_nice - check if a task can reduce its nice value
4432 int can_nice(const struct task_struct *p, const int nice)
4434 /* convert nice value [19,-20] to rlimit style value [1,40] */
4435 int nice_rlim = 20 - nice;
4437 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4438 capable(CAP_SYS_NICE));
4441 #ifdef __ARCH_WANT_SYS_NICE
4444 * sys_nice - change the priority of the current process.
4445 * @increment: priority increment
4447 * sys_setpriority is a more generic, but much slower function that
4448 * does similar things.
4450 SYSCALL_DEFINE1(nice, int, increment)
4455 * Setpriority might change our priority at the same moment.
4456 * We don't have to worry. Conceptually one call occurs first
4457 * and we have a single winner.
4459 if (increment < -40)
4464 nice = TASK_NICE(current) + increment;
4470 if (increment < 0 && !can_nice(current, nice))
4473 retval = security_task_setnice(current, nice);
4477 set_user_nice(current, nice);
4484 * task_prio - return the priority value of a given task.
4485 * @p: the task in question.
4487 * This is the priority value as seen by users in /proc.
4488 * RT tasks are offset by -200. Normal tasks are centered
4489 * around 0, value goes from -16 to +15.
4491 int task_prio(const struct task_struct *p)
4493 return p->prio - MAX_RT_PRIO;
4497 * task_nice - return the nice value of a given task.
4498 * @p: the task in question.
4500 int task_nice(const struct task_struct *p)
4502 return TASK_NICE(p);
4504 EXPORT_SYMBOL(task_nice);
4507 * idle_cpu - is a given cpu idle currently?
4508 * @cpu: the processor in question.
4510 int idle_cpu(int cpu)
4512 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4516 * idle_task - return the idle task for a given cpu.
4517 * @cpu: the processor in question.
4519 struct task_struct *idle_task(int cpu)
4521 return cpu_rq(cpu)->idle;
4525 * find_process_by_pid - find a process with a matching PID value.
4526 * @pid: the pid in question.
4528 static struct task_struct *find_process_by_pid(pid_t pid)
4530 return pid ? find_task_by_vpid(pid) : current;
4533 /* Actually do priority change: must hold rq lock. */
4535 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4537 BUG_ON(p->se.on_rq);
4540 p->rt_priority = prio;
4541 p->normal_prio = normal_prio(p);
4542 /* we are holding p->pi_lock already */
4543 p->prio = rt_mutex_getprio(p);
4544 if (rt_prio(p->prio))
4545 p->sched_class = &rt_sched_class;
4547 p->sched_class = &fair_sched_class;
4552 * check the target process has a UID that matches the current process's
4554 static bool check_same_owner(struct task_struct *p)
4556 const struct cred *cred = current_cred(), *pcred;
4560 pcred = __task_cred(p);
4561 match = (cred->euid == pcred->euid ||
4562 cred->euid == pcred->uid);
4567 static int __sched_setscheduler(struct task_struct *p, int policy,
4568 struct sched_param *param, bool user)
4570 int retval, oldprio, oldpolicy = -1, on_rq, running;
4571 unsigned long flags;
4572 const struct sched_class *prev_class;
4576 /* may grab non-irq protected spin_locks */
4577 BUG_ON(in_interrupt());
4579 /* double check policy once rq lock held */
4581 reset_on_fork = p->sched_reset_on_fork;
4582 policy = oldpolicy = p->policy;
4584 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4585 policy &= ~SCHED_RESET_ON_FORK;
4587 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4588 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4589 policy != SCHED_IDLE)
4594 * Valid priorities for SCHED_FIFO and SCHED_RR are
4595 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4596 * SCHED_BATCH and SCHED_IDLE is 0.
4598 if (param->sched_priority < 0 ||
4599 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4600 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4602 if (rt_policy(policy) != (param->sched_priority != 0))
4606 * Allow unprivileged RT tasks to decrease priority:
4608 if (user && !capable(CAP_SYS_NICE)) {
4609 if (rt_policy(policy)) {
4610 unsigned long rlim_rtprio =
4611 task_rlimit(p, RLIMIT_RTPRIO);
4613 /* can't set/change the rt policy */
4614 if (policy != p->policy && !rlim_rtprio)
4617 /* can't increase priority */
4618 if (param->sched_priority > p->rt_priority &&
4619 param->sched_priority > rlim_rtprio)
4623 * Like positive nice levels, dont allow tasks to
4624 * move out of SCHED_IDLE either:
4626 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4629 /* can't change other user's priorities */
4630 if (!check_same_owner(p))
4633 /* Normal users shall not reset the sched_reset_on_fork flag */
4634 if (p->sched_reset_on_fork && !reset_on_fork)
4639 retval = security_task_setscheduler(p, policy, param);
4645 * make sure no PI-waiters arrive (or leave) while we are
4646 * changing the priority of the task:
4648 raw_spin_lock_irqsave(&p->pi_lock, flags);
4650 * To be able to change p->policy safely, the apropriate
4651 * runqueue lock must be held.
4653 rq = __task_rq_lock(p);
4655 #ifdef CONFIG_RT_GROUP_SCHED
4658 * Do not allow realtime tasks into groups that have no runtime
4661 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4662 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4663 __task_rq_unlock(rq);
4664 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4670 /* recheck policy now with rq lock held */
4671 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4672 policy = oldpolicy = -1;
4673 __task_rq_unlock(rq);
4674 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4677 on_rq = p->se.on_rq;
4678 running = task_current(rq, p);
4680 deactivate_task(rq, p, 0);
4682 p->sched_class->put_prev_task(rq, p);
4684 p->sched_reset_on_fork = reset_on_fork;
4687 prev_class = p->sched_class;
4688 __setscheduler(rq, p, policy, param->sched_priority);
4691 p->sched_class->set_curr_task(rq);
4693 activate_task(rq, p, 0);
4695 check_class_changed(rq, p, prev_class, oldprio, running);
4697 __task_rq_unlock(rq);
4698 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4700 rt_mutex_adjust_pi(p);
4706 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4707 * @p: the task in question.
4708 * @policy: new policy.
4709 * @param: structure containing the new RT priority.
4711 * NOTE that the task may be already dead.
4713 int sched_setscheduler(struct task_struct *p, int policy,
4714 struct sched_param *param)
4716 return __sched_setscheduler(p, policy, param, true);
4718 EXPORT_SYMBOL_GPL(sched_setscheduler);
4721 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4722 * @p: the task in question.
4723 * @policy: new policy.
4724 * @param: structure containing the new RT priority.
4726 * Just like sched_setscheduler, only don't bother checking if the
4727 * current context has permission. For example, this is needed in
4728 * stop_machine(): we create temporary high priority worker threads,
4729 * but our caller might not have that capability.
4731 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4732 struct sched_param *param)
4734 return __sched_setscheduler(p, policy, param, false);
4738 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4740 struct sched_param lparam;
4741 struct task_struct *p;
4744 if (!param || pid < 0)
4746 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4751 p = find_process_by_pid(pid);
4753 retval = sched_setscheduler(p, policy, &lparam);
4760 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4761 * @pid: the pid in question.
4762 * @policy: new policy.
4763 * @param: structure containing the new RT priority.
4765 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4766 struct sched_param __user *, param)
4768 /* negative values for policy are not valid */
4772 return do_sched_setscheduler(pid, policy, param);
4776 * sys_sched_setparam - set/change the RT priority of a thread
4777 * @pid: the pid in question.
4778 * @param: structure containing the new RT priority.
4780 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4782 return do_sched_setscheduler(pid, -1, param);
4786 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4787 * @pid: the pid in question.
4789 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4791 struct task_struct *p;
4799 p = find_process_by_pid(pid);
4801 retval = security_task_getscheduler(p);
4804 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4811 * sys_sched_getparam - get the RT priority of a thread
4812 * @pid: the pid in question.
4813 * @param: structure containing the RT priority.
4815 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4817 struct sched_param lp;
4818 struct task_struct *p;
4821 if (!param || pid < 0)
4825 p = find_process_by_pid(pid);
4830 retval = security_task_getscheduler(p);
4834 lp.sched_priority = p->rt_priority;
4838 * This one might sleep, we cannot do it with a spinlock held ...
4840 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4849 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4851 cpumask_var_t cpus_allowed, new_mask;
4852 struct task_struct *p;
4858 p = find_process_by_pid(pid);
4865 /* Prevent p going away */
4869 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4873 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4875 goto out_free_cpus_allowed;
4878 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4881 retval = security_task_setscheduler(p, 0, NULL);
4885 cpuset_cpus_allowed(p, cpus_allowed);
4886 cpumask_and(new_mask, in_mask, cpus_allowed);
4888 retval = set_cpus_allowed_ptr(p, new_mask);
4891 cpuset_cpus_allowed(p, cpus_allowed);
4892 if (!cpumask_subset(new_mask, cpus_allowed)) {
4894 * We must have raced with a concurrent cpuset
4895 * update. Just reset the cpus_allowed to the
4896 * cpuset's cpus_allowed
4898 cpumask_copy(new_mask, cpus_allowed);
4903 free_cpumask_var(new_mask);
4904 out_free_cpus_allowed:
4905 free_cpumask_var(cpus_allowed);
4912 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4913 struct cpumask *new_mask)
4915 if (len < cpumask_size())
4916 cpumask_clear(new_mask);
4917 else if (len > cpumask_size())
4918 len = cpumask_size();
4920 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4924 * sys_sched_setaffinity - set the cpu affinity of a process
4925 * @pid: pid of the process
4926 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4927 * @user_mask_ptr: user-space pointer to the new cpu mask
4929 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4930 unsigned long __user *, user_mask_ptr)
4932 cpumask_var_t new_mask;
4935 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4938 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4940 retval = sched_setaffinity(pid, new_mask);
4941 free_cpumask_var(new_mask);
4945 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4947 struct task_struct *p;
4948 unsigned long flags;
4956 p = find_process_by_pid(pid);
4960 retval = security_task_getscheduler(p);
4964 rq = task_rq_lock(p, &flags);
4965 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4966 task_rq_unlock(rq, &flags);
4976 * sys_sched_getaffinity - get the cpu affinity of a process
4977 * @pid: pid of the process
4978 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4979 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4981 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4982 unsigned long __user *, user_mask_ptr)
4987 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4989 if (len & (sizeof(unsigned long)-1))
4992 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4995 ret = sched_getaffinity(pid, mask);
4997 size_t retlen = min_t(size_t, len, cpumask_size());
4999 if (copy_to_user(user_mask_ptr, mask, retlen))
5004 free_cpumask_var(mask);
5010 * sys_sched_yield - yield the current processor to other threads.
5012 * This function yields the current CPU to other tasks. If there are no
5013 * other threads running on this CPU then this function will return.
5015 SYSCALL_DEFINE0(sched_yield)
5017 struct rq *rq = this_rq_lock();
5019 schedstat_inc(rq, yld_count);
5020 current->sched_class->yield_task(rq);
5023 * Since we are going to call schedule() anyway, there's
5024 * no need to preempt or enable interrupts:
5026 __release(rq->lock);
5027 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5028 do_raw_spin_unlock(&rq->lock);
5029 preempt_enable_no_resched();
5036 static inline int should_resched(void)
5038 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5041 static void __cond_resched(void)
5043 add_preempt_count(PREEMPT_ACTIVE);
5045 sub_preempt_count(PREEMPT_ACTIVE);
5048 int __sched _cond_resched(void)
5050 if (should_resched()) {
5056 EXPORT_SYMBOL(_cond_resched);
5059 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5060 * call schedule, and on return reacquire the lock.
5062 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5063 * operations here to prevent schedule() from being called twice (once via
5064 * spin_unlock(), once by hand).
5066 int __cond_resched_lock(spinlock_t *lock)
5068 int resched = should_resched();
5071 lockdep_assert_held(lock);
5073 if (spin_needbreak(lock) || resched) {
5084 EXPORT_SYMBOL(__cond_resched_lock);
5086 int __sched __cond_resched_softirq(void)
5088 BUG_ON(!in_softirq());
5090 if (should_resched()) {
5098 EXPORT_SYMBOL(__cond_resched_softirq);
5101 * yield - yield the current processor to other threads.
5103 * This is a shortcut for kernel-space yielding - it marks the
5104 * thread runnable and calls sys_sched_yield().
5106 void __sched yield(void)
5108 set_current_state(TASK_RUNNING);
5111 EXPORT_SYMBOL(yield);
5114 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5115 * that process accounting knows that this is a task in IO wait state.
5117 void __sched io_schedule(void)
5119 struct rq *rq = raw_rq();
5121 delayacct_blkio_start();
5122 atomic_inc(&rq->nr_iowait);
5123 current->in_iowait = 1;
5125 current->in_iowait = 0;
5126 atomic_dec(&rq->nr_iowait);
5127 delayacct_blkio_end();
5129 EXPORT_SYMBOL(io_schedule);
5131 long __sched io_schedule_timeout(long timeout)
5133 struct rq *rq = raw_rq();
5136 delayacct_blkio_start();
5137 atomic_inc(&rq->nr_iowait);
5138 current->in_iowait = 1;
5139 ret = schedule_timeout(timeout);
5140 current->in_iowait = 0;
5141 atomic_dec(&rq->nr_iowait);
5142 delayacct_blkio_end();
5147 * sys_sched_get_priority_max - return maximum RT priority.
5148 * @policy: scheduling class.
5150 * this syscall returns the maximum rt_priority that can be used
5151 * by a given scheduling class.
5153 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5160 ret = MAX_USER_RT_PRIO-1;
5172 * sys_sched_get_priority_min - return minimum RT priority.
5173 * @policy: scheduling class.
5175 * this syscall returns the minimum rt_priority that can be used
5176 * by a given scheduling class.
5178 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5196 * sys_sched_rr_get_interval - return the default timeslice of a process.
5197 * @pid: pid of the process.
5198 * @interval: userspace pointer to the timeslice value.
5200 * this syscall writes the default timeslice value of a given process
5201 * into the user-space timespec buffer. A value of '0' means infinity.
5203 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5204 struct timespec __user *, interval)
5206 struct task_struct *p;
5207 unsigned int time_slice;
5208 unsigned long flags;
5218 p = find_process_by_pid(pid);
5222 retval = security_task_getscheduler(p);
5226 rq = task_rq_lock(p, &flags);
5227 time_slice = p->sched_class->get_rr_interval(rq, p);
5228 task_rq_unlock(rq, &flags);
5231 jiffies_to_timespec(time_slice, &t);
5232 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5240 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5242 void sched_show_task(struct task_struct *p)
5244 unsigned long free = 0;
5247 state = p->state ? __ffs(p->state) + 1 : 0;
5248 printk(KERN_INFO "%-13.13s %c", p->comm,
5249 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5250 #if BITS_PER_LONG == 32
5251 if (state == TASK_RUNNING)
5252 printk(KERN_CONT " running ");
5254 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5256 if (state == TASK_RUNNING)
5257 printk(KERN_CONT " running task ");
5259 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5261 #ifdef CONFIG_DEBUG_STACK_USAGE
5262 free = stack_not_used(p);
5264 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5265 task_pid_nr(p), task_pid_nr(p->real_parent),
5266 (unsigned long)task_thread_info(p)->flags);
5268 show_stack(p, NULL);
5271 void show_state_filter(unsigned long state_filter)
5273 struct task_struct *g, *p;
5275 #if BITS_PER_LONG == 32
5277 " task PC stack pid father\n");
5280 " task PC stack pid father\n");
5282 read_lock(&tasklist_lock);
5283 do_each_thread(g, p) {
5285 * reset the NMI-timeout, listing all files on a slow
5286 * console might take alot of time:
5288 touch_nmi_watchdog();
5289 if (!state_filter || (p->state & state_filter))
5291 } while_each_thread(g, p);
5293 touch_all_softlockup_watchdogs();
5295 #ifdef CONFIG_SCHED_DEBUG
5296 sysrq_sched_debug_show();
5298 read_unlock(&tasklist_lock);
5300 * Only show locks if all tasks are dumped:
5303 debug_show_all_locks();
5306 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5308 idle->sched_class = &idle_sched_class;
5312 * init_idle - set up an idle thread for a given CPU
5313 * @idle: task in question
5314 * @cpu: cpu the idle task belongs to
5316 * NOTE: this function does not set the idle thread's NEED_RESCHED
5317 * flag, to make booting more robust.
5319 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5321 struct rq *rq = cpu_rq(cpu);
5322 unsigned long flags;
5324 raw_spin_lock_irqsave(&rq->lock, flags);
5327 idle->state = TASK_RUNNING;
5328 idle->se.exec_start = sched_clock();
5330 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5331 __set_task_cpu(idle, cpu);
5333 rq->curr = rq->idle = idle;
5334 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5337 raw_spin_unlock_irqrestore(&rq->lock, flags);
5339 /* Set the preempt count _outside_ the spinlocks! */
5340 #if defined(CONFIG_PREEMPT)
5341 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5343 task_thread_info(idle)->preempt_count = 0;
5346 * The idle tasks have their own, simple scheduling class:
5348 idle->sched_class = &idle_sched_class;
5349 ftrace_graph_init_task(idle);
5353 * In a system that switches off the HZ timer nohz_cpu_mask
5354 * indicates which cpus entered this state. This is used
5355 * in the rcu update to wait only for active cpus. For system
5356 * which do not switch off the HZ timer nohz_cpu_mask should
5357 * always be CPU_BITS_NONE.
5359 cpumask_var_t nohz_cpu_mask;
5362 * Increase the granularity value when there are more CPUs,
5363 * because with more CPUs the 'effective latency' as visible
5364 * to users decreases. But the relationship is not linear,
5365 * so pick a second-best guess by going with the log2 of the
5368 * This idea comes from the SD scheduler of Con Kolivas:
5370 static int get_update_sysctl_factor(void)
5372 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5373 unsigned int factor;
5375 switch (sysctl_sched_tunable_scaling) {
5376 case SCHED_TUNABLESCALING_NONE:
5379 case SCHED_TUNABLESCALING_LINEAR:
5382 case SCHED_TUNABLESCALING_LOG:
5384 factor = 1 + ilog2(cpus);
5391 static void update_sysctl(void)
5393 unsigned int factor = get_update_sysctl_factor();
5395 #define SET_SYSCTL(name) \
5396 (sysctl_##name = (factor) * normalized_sysctl_##name)
5397 SET_SYSCTL(sched_min_granularity);
5398 SET_SYSCTL(sched_latency);
5399 SET_SYSCTL(sched_wakeup_granularity);
5400 SET_SYSCTL(sched_shares_ratelimit);
5404 static inline void sched_init_granularity(void)
5411 * This is how migration works:
5413 * 1) we invoke migration_cpu_stop() on the target CPU using
5415 * 2) stopper starts to run (implicitly forcing the migrated thread
5417 * 3) it checks whether the migrated task is still in the wrong runqueue.
5418 * 4) if it's in the wrong runqueue then the migration thread removes
5419 * it and puts it into the right queue.
5420 * 5) stopper completes and stop_one_cpu() returns and the migration
5425 * Change a given task's CPU affinity. Migrate the thread to a
5426 * proper CPU and schedule it away if the CPU it's executing on
5427 * is removed from the allowed bitmask.
5429 * NOTE: the caller must have a valid reference to the task, the
5430 * task must not exit() & deallocate itself prematurely. The
5431 * call is not atomic; no spinlocks may be held.
5433 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5435 unsigned long flags;
5437 unsigned int dest_cpu;
5441 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5442 * drop the rq->lock and still rely on ->cpus_allowed.
5445 while (task_is_waking(p))
5447 rq = task_rq_lock(p, &flags);
5448 if (task_is_waking(p)) {
5449 task_rq_unlock(rq, &flags);
5453 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5458 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5459 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5464 if (p->sched_class->set_cpus_allowed)
5465 p->sched_class->set_cpus_allowed(p, new_mask);
5467 cpumask_copy(&p->cpus_allowed, new_mask);
5468 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5471 /* Can the task run on the task's current CPU? If so, we're done */
5472 if (cpumask_test_cpu(task_cpu(p), new_mask))
5475 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5476 if (migrate_task(p, dest_cpu)) {
5477 struct migration_arg arg = { p, dest_cpu };
5478 /* Need help from migration thread: drop lock and wait. */
5479 task_rq_unlock(rq, &flags);
5480 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5481 tlb_migrate_finish(p->mm);
5485 task_rq_unlock(rq, &flags);
5489 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5492 * Move (not current) task off this cpu, onto dest cpu. We're doing
5493 * this because either it can't run here any more (set_cpus_allowed()
5494 * away from this CPU, or CPU going down), or because we're
5495 * attempting to rebalance this task on exec (sched_exec).
5497 * So we race with normal scheduler movements, but that's OK, as long
5498 * as the task is no longer on this CPU.
5500 * Returns non-zero if task was successfully migrated.
5502 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5504 struct rq *rq_dest, *rq_src;
5507 if (unlikely(!cpu_active(dest_cpu)))
5510 rq_src = cpu_rq(src_cpu);
5511 rq_dest = cpu_rq(dest_cpu);
5513 double_rq_lock(rq_src, rq_dest);
5514 /* Already moved. */
5515 if (task_cpu(p) != src_cpu)
5517 /* Affinity changed (again). */
5518 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5522 * If we're not on a rq, the next wake-up will ensure we're
5526 deactivate_task(rq_src, p, 0);
5527 set_task_cpu(p, dest_cpu);
5528 activate_task(rq_dest, p, 0);
5529 check_preempt_curr(rq_dest, p, 0);
5534 double_rq_unlock(rq_src, rq_dest);
5539 * migration_cpu_stop - this will be executed by a highprio stopper thread
5540 * and performs thread migration by bumping thread off CPU then
5541 * 'pushing' onto another runqueue.
5543 static int migration_cpu_stop(void *data)
5545 struct migration_arg *arg = data;
5548 * The original target cpu might have gone down and we might
5549 * be on another cpu but it doesn't matter.
5551 local_irq_disable();
5552 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5557 #ifdef CONFIG_HOTPLUG_CPU
5559 * Figure out where task on dead CPU should go, use force if necessary.
5561 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5563 struct rq *rq = cpu_rq(dead_cpu);
5564 int needs_cpu, uninitialized_var(dest_cpu);
5565 unsigned long flags;
5567 local_irq_save(flags);
5569 raw_spin_lock(&rq->lock);
5570 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5572 dest_cpu = select_fallback_rq(dead_cpu, p);
5573 raw_spin_unlock(&rq->lock);
5575 * It can only fail if we race with set_cpus_allowed(),
5576 * in the racer should migrate the task anyway.
5579 __migrate_task(p, dead_cpu, dest_cpu);
5580 local_irq_restore(flags);
5584 * While a dead CPU has no uninterruptible tasks queued at this point,
5585 * it might still have a nonzero ->nr_uninterruptible counter, because
5586 * for performance reasons the counter is not stricly tracking tasks to
5587 * their home CPUs. So we just add the counter to another CPU's counter,
5588 * to keep the global sum constant after CPU-down:
5590 static void migrate_nr_uninterruptible(struct rq *rq_src)
5592 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5593 unsigned long flags;
5595 local_irq_save(flags);
5596 double_rq_lock(rq_src, rq_dest);
5597 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5598 rq_src->nr_uninterruptible = 0;
5599 double_rq_unlock(rq_src, rq_dest);
5600 local_irq_restore(flags);
5603 /* Run through task list and migrate tasks from the dead cpu. */
5604 static void migrate_live_tasks(int src_cpu)
5606 struct task_struct *p, *t;
5608 read_lock(&tasklist_lock);
5610 do_each_thread(t, p) {
5614 if (task_cpu(p) == src_cpu)
5615 move_task_off_dead_cpu(src_cpu, p);
5616 } while_each_thread(t, p);
5618 read_unlock(&tasklist_lock);
5622 * Schedules idle task to be the next runnable task on current CPU.
5623 * It does so by boosting its priority to highest possible.
5624 * Used by CPU offline code.
5626 void sched_idle_next(void)
5628 int this_cpu = smp_processor_id();
5629 struct rq *rq = cpu_rq(this_cpu);
5630 struct task_struct *p = rq->idle;
5631 unsigned long flags;
5633 /* cpu has to be offline */
5634 BUG_ON(cpu_online(this_cpu));
5637 * Strictly not necessary since rest of the CPUs are stopped by now
5638 * and interrupts disabled on the current cpu.
5640 raw_spin_lock_irqsave(&rq->lock, flags);
5642 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5644 activate_task(rq, p, 0);
5646 raw_spin_unlock_irqrestore(&rq->lock, flags);
5650 * Ensures that the idle task is using init_mm right before its cpu goes
5653 void idle_task_exit(void)
5655 struct mm_struct *mm = current->active_mm;
5657 BUG_ON(cpu_online(smp_processor_id()));
5660 switch_mm(mm, &init_mm, current);
5664 /* called under rq->lock with disabled interrupts */
5665 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5667 struct rq *rq = cpu_rq(dead_cpu);
5669 /* Must be exiting, otherwise would be on tasklist. */
5670 BUG_ON(!p->exit_state);
5672 /* Cannot have done final schedule yet: would have vanished. */
5673 BUG_ON(p->state == TASK_DEAD);
5678 * Drop lock around migration; if someone else moves it,
5679 * that's OK. No task can be added to this CPU, so iteration is
5682 raw_spin_unlock_irq(&rq->lock);
5683 move_task_off_dead_cpu(dead_cpu, p);
5684 raw_spin_lock_irq(&rq->lock);
5689 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5690 static void migrate_dead_tasks(unsigned int dead_cpu)
5692 struct rq *rq = cpu_rq(dead_cpu);
5693 struct task_struct *next;
5696 if (!rq->nr_running)
5698 next = pick_next_task(rq);
5701 next->sched_class->put_prev_task(rq, next);
5702 migrate_dead(dead_cpu, next);
5708 * remove the tasks which were accounted by rq from calc_load_tasks.
5710 static void calc_global_load_remove(struct rq *rq)
5712 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5713 rq->calc_load_active = 0;
5715 #endif /* CONFIG_HOTPLUG_CPU */
5717 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5719 static struct ctl_table sd_ctl_dir[] = {
5721 .procname = "sched_domain",
5727 static struct ctl_table sd_ctl_root[] = {
5729 .procname = "kernel",
5731 .child = sd_ctl_dir,
5736 static struct ctl_table *sd_alloc_ctl_entry(int n)
5738 struct ctl_table *entry =
5739 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5744 static void sd_free_ctl_entry(struct ctl_table **tablep)
5746 struct ctl_table *entry;
5749 * In the intermediate directories, both the child directory and
5750 * procname are dynamically allocated and could fail but the mode
5751 * will always be set. In the lowest directory the names are
5752 * static strings and all have proc handlers.
5754 for (entry = *tablep; entry->mode; entry++) {
5756 sd_free_ctl_entry(&entry->child);
5757 if (entry->proc_handler == NULL)
5758 kfree(entry->procname);
5766 set_table_entry(struct ctl_table *entry,
5767 const char *procname, void *data, int maxlen,
5768 mode_t mode, proc_handler *proc_handler)
5770 entry->procname = procname;
5772 entry->maxlen = maxlen;
5774 entry->proc_handler = proc_handler;
5777 static struct ctl_table *
5778 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5780 struct ctl_table *table = sd_alloc_ctl_entry(13);
5785 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5786 sizeof(long), 0644, proc_doulongvec_minmax);
5787 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5788 sizeof(long), 0644, proc_doulongvec_minmax);
5789 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5790 sizeof(int), 0644, proc_dointvec_minmax);
5791 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5792 sizeof(int), 0644, proc_dointvec_minmax);
5793 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5794 sizeof(int), 0644, proc_dointvec_minmax);
5795 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5796 sizeof(int), 0644, proc_dointvec_minmax);
5797 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5798 sizeof(int), 0644, proc_dointvec_minmax);
5799 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5800 sizeof(int), 0644, proc_dointvec_minmax);
5801 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5802 sizeof(int), 0644, proc_dointvec_minmax);
5803 set_table_entry(&table[9], "cache_nice_tries",
5804 &sd->cache_nice_tries,
5805 sizeof(int), 0644, proc_dointvec_minmax);
5806 set_table_entry(&table[10], "flags", &sd->flags,
5807 sizeof(int), 0644, proc_dointvec_minmax);
5808 set_table_entry(&table[11], "name", sd->name,
5809 CORENAME_MAX_SIZE, 0444, proc_dostring);
5810 /* &table[12] is terminator */
5815 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5817 struct ctl_table *entry, *table;
5818 struct sched_domain *sd;
5819 int domain_num = 0, i;
5822 for_each_domain(cpu, sd)
5824 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5829 for_each_domain(cpu, sd) {
5830 snprintf(buf, 32, "domain%d", i);
5831 entry->procname = kstrdup(buf, GFP_KERNEL);
5833 entry->child = sd_alloc_ctl_domain_table(sd);
5840 static struct ctl_table_header *sd_sysctl_header;
5841 static void register_sched_domain_sysctl(void)
5843 int i, cpu_num = num_possible_cpus();
5844 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5847 WARN_ON(sd_ctl_dir[0].child);
5848 sd_ctl_dir[0].child = entry;
5853 for_each_possible_cpu(i) {
5854 snprintf(buf, 32, "cpu%d", i);
5855 entry->procname = kstrdup(buf, GFP_KERNEL);
5857 entry->child = sd_alloc_ctl_cpu_table(i);
5861 WARN_ON(sd_sysctl_header);
5862 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5865 /* may be called multiple times per register */
5866 static void unregister_sched_domain_sysctl(void)
5868 if (sd_sysctl_header)
5869 unregister_sysctl_table(sd_sysctl_header);
5870 sd_sysctl_header = NULL;
5871 if (sd_ctl_dir[0].child)
5872 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5875 static void register_sched_domain_sysctl(void)
5878 static void unregister_sched_domain_sysctl(void)
5883 static void set_rq_online(struct rq *rq)
5886 const struct sched_class *class;
5888 cpumask_set_cpu(rq->cpu, rq->rd->online);
5891 for_each_class(class) {
5892 if (class->rq_online)
5893 class->rq_online(rq);
5898 static void set_rq_offline(struct rq *rq)
5901 const struct sched_class *class;
5903 for_each_class(class) {
5904 if (class->rq_offline)
5905 class->rq_offline(rq);
5908 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5914 * migration_call - callback that gets triggered when a CPU is added.
5915 * Here we can start up the necessary migration thread for the new CPU.
5917 static int __cpuinit
5918 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5920 int cpu = (long)hcpu;
5921 unsigned long flags;
5922 struct rq *rq = cpu_rq(cpu);
5926 case CPU_UP_PREPARE:
5927 case CPU_UP_PREPARE_FROZEN:
5928 rq->calc_load_update = calc_load_update;
5932 case CPU_ONLINE_FROZEN:
5933 /* Update our root-domain */
5934 raw_spin_lock_irqsave(&rq->lock, flags);
5936 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5940 raw_spin_unlock_irqrestore(&rq->lock, flags);
5943 #ifdef CONFIG_HOTPLUG_CPU
5945 case CPU_DEAD_FROZEN:
5946 migrate_live_tasks(cpu);
5947 /* Idle task back to normal (off runqueue, low prio) */
5948 raw_spin_lock_irq(&rq->lock);
5949 deactivate_task(rq, rq->idle, 0);
5950 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5951 rq->idle->sched_class = &idle_sched_class;
5952 migrate_dead_tasks(cpu);
5953 raw_spin_unlock_irq(&rq->lock);
5954 migrate_nr_uninterruptible(rq);
5955 BUG_ON(rq->nr_running != 0);
5956 calc_global_load_remove(rq);
5960 case CPU_DYING_FROZEN:
5961 /* Update our root-domain */
5962 raw_spin_lock_irqsave(&rq->lock, flags);
5964 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5967 raw_spin_unlock_irqrestore(&rq->lock, flags);
5975 * Register at high priority so that task migration (migrate_all_tasks)
5976 * happens before everything else. This has to be lower priority than
5977 * the notifier in the perf_event subsystem, though.
5979 static struct notifier_block __cpuinitdata migration_notifier = {
5980 .notifier_call = migration_call,
5981 .priority = CPU_PRI_MIGRATION,
5984 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5985 unsigned long action, void *hcpu)
5987 switch (action & ~CPU_TASKS_FROZEN) {
5989 case CPU_DOWN_FAILED:
5990 set_cpu_active((long)hcpu, true);
5997 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5998 unsigned long action, void *hcpu)
6000 switch (action & ~CPU_TASKS_FROZEN) {
6001 case CPU_DOWN_PREPARE:
6002 set_cpu_active((long)hcpu, false);
6009 static int __init migration_init(void)
6011 void *cpu = (void *)(long)smp_processor_id();
6014 /* Initialize migration for the boot CPU */
6015 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6016 BUG_ON(err == NOTIFY_BAD);
6017 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6018 register_cpu_notifier(&migration_notifier);
6020 /* Register cpu active notifiers */
6021 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6022 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6026 early_initcall(migration_init);
6031 #ifdef CONFIG_SCHED_DEBUG
6033 static __read_mostly int sched_domain_debug_enabled;
6035 static int __init sched_domain_debug_setup(char *str)
6037 sched_domain_debug_enabled = 1;
6041 early_param("sched_debug", sched_domain_debug_setup);
6043 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6044 struct cpumask *groupmask)
6046 struct sched_group *group = sd->groups;
6049 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6050 cpumask_clear(groupmask);
6052 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6054 if (!(sd->flags & SD_LOAD_BALANCE)) {
6055 printk("does not load-balance\n");
6057 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6062 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6064 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6065 printk(KERN_ERR "ERROR: domain->span does not contain "
6068 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6069 printk(KERN_ERR "ERROR: domain->groups does not contain"
6073 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6077 printk(KERN_ERR "ERROR: group is NULL\n");
6081 if (!group->cpu_power) {
6082 printk(KERN_CONT "\n");
6083 printk(KERN_ERR "ERROR: domain->cpu_power not "
6088 if (!cpumask_weight(sched_group_cpus(group))) {
6089 printk(KERN_CONT "\n");
6090 printk(KERN_ERR "ERROR: empty group\n");
6094 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6095 printk(KERN_CONT "\n");
6096 printk(KERN_ERR "ERROR: repeated CPUs\n");
6100 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6102 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6104 printk(KERN_CONT " %s", str);
6105 if (group->cpu_power != SCHED_LOAD_SCALE) {
6106 printk(KERN_CONT " (cpu_power = %d)",
6110 group = group->next;
6111 } while (group != sd->groups);
6112 printk(KERN_CONT "\n");
6114 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6115 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6118 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6119 printk(KERN_ERR "ERROR: parent span is not a superset "
6120 "of domain->span\n");
6124 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6126 cpumask_var_t groupmask;
6129 if (!sched_domain_debug_enabled)
6133 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6137 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6139 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6140 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6145 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6152 free_cpumask_var(groupmask);
6154 #else /* !CONFIG_SCHED_DEBUG */
6155 # define sched_domain_debug(sd, cpu) do { } while (0)
6156 #endif /* CONFIG_SCHED_DEBUG */
6158 static int sd_degenerate(struct sched_domain *sd)
6160 if (cpumask_weight(sched_domain_span(sd)) == 1)
6163 /* Following flags need at least 2 groups */
6164 if (sd->flags & (SD_LOAD_BALANCE |
6165 SD_BALANCE_NEWIDLE |
6169 SD_SHARE_PKG_RESOURCES)) {
6170 if (sd->groups != sd->groups->next)
6174 /* Following flags don't use groups */
6175 if (sd->flags & (SD_WAKE_AFFINE))
6182 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6184 unsigned long cflags = sd->flags, pflags = parent->flags;
6186 if (sd_degenerate(parent))
6189 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6192 /* Flags needing groups don't count if only 1 group in parent */
6193 if (parent->groups == parent->groups->next) {
6194 pflags &= ~(SD_LOAD_BALANCE |
6195 SD_BALANCE_NEWIDLE |
6199 SD_SHARE_PKG_RESOURCES);
6200 if (nr_node_ids == 1)
6201 pflags &= ~SD_SERIALIZE;
6203 if (~cflags & pflags)
6209 static void free_rootdomain(struct root_domain *rd)
6211 synchronize_sched();
6213 cpupri_cleanup(&rd->cpupri);
6215 free_cpumask_var(rd->rto_mask);
6216 free_cpumask_var(rd->online);
6217 free_cpumask_var(rd->span);
6221 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6223 struct root_domain *old_rd = NULL;
6224 unsigned long flags;
6226 raw_spin_lock_irqsave(&rq->lock, flags);
6231 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6234 cpumask_clear_cpu(rq->cpu, old_rd->span);
6237 * If we dont want to free the old_rt yet then
6238 * set old_rd to NULL to skip the freeing later
6241 if (!atomic_dec_and_test(&old_rd->refcount))
6245 atomic_inc(&rd->refcount);
6248 cpumask_set_cpu(rq->cpu, rd->span);
6249 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6252 raw_spin_unlock_irqrestore(&rq->lock, flags);
6255 free_rootdomain(old_rd);
6258 static int init_rootdomain(struct root_domain *rd)
6260 memset(rd, 0, sizeof(*rd));
6262 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6264 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6266 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6269 if (cpupri_init(&rd->cpupri) != 0)
6274 free_cpumask_var(rd->rto_mask);
6276 free_cpumask_var(rd->online);
6278 free_cpumask_var(rd->span);
6283 static void init_defrootdomain(void)
6285 init_rootdomain(&def_root_domain);
6287 atomic_set(&def_root_domain.refcount, 1);
6290 static struct root_domain *alloc_rootdomain(void)
6292 struct root_domain *rd;
6294 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6298 if (init_rootdomain(rd) != 0) {
6307 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6308 * hold the hotplug lock.
6311 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6313 struct rq *rq = cpu_rq(cpu);
6314 struct sched_domain *tmp;
6316 for (tmp = sd; tmp; tmp = tmp->parent)
6317 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6319 /* Remove the sched domains which do not contribute to scheduling. */
6320 for (tmp = sd; tmp; ) {
6321 struct sched_domain *parent = tmp->parent;
6325 if (sd_parent_degenerate(tmp, parent)) {
6326 tmp->parent = parent->parent;
6328 parent->parent->child = tmp;
6333 if (sd && sd_degenerate(sd)) {
6339 sched_domain_debug(sd, cpu);
6341 rq_attach_root(rq, rd);
6342 rcu_assign_pointer(rq->sd, sd);
6345 /* cpus with isolated domains */
6346 static cpumask_var_t cpu_isolated_map;
6348 /* Setup the mask of cpus configured for isolated domains */
6349 static int __init isolated_cpu_setup(char *str)
6351 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6352 cpulist_parse(str, cpu_isolated_map);
6356 __setup("isolcpus=", isolated_cpu_setup);
6359 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6360 * to a function which identifies what group(along with sched group) a CPU
6361 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6362 * (due to the fact that we keep track of groups covered with a struct cpumask).
6364 * init_sched_build_groups will build a circular linked list of the groups
6365 * covered by the given span, and will set each group's ->cpumask correctly,
6366 * and ->cpu_power to 0.
6369 init_sched_build_groups(const struct cpumask *span,
6370 const struct cpumask *cpu_map,
6371 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6372 struct sched_group **sg,
6373 struct cpumask *tmpmask),
6374 struct cpumask *covered, struct cpumask *tmpmask)
6376 struct sched_group *first = NULL, *last = NULL;
6379 cpumask_clear(covered);
6381 for_each_cpu(i, span) {
6382 struct sched_group *sg;
6383 int group = group_fn(i, cpu_map, &sg, tmpmask);
6386 if (cpumask_test_cpu(i, covered))
6389 cpumask_clear(sched_group_cpus(sg));
6392 for_each_cpu(j, span) {
6393 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6396 cpumask_set_cpu(j, covered);
6397 cpumask_set_cpu(j, sched_group_cpus(sg));
6408 #define SD_NODES_PER_DOMAIN 16
6413 * find_next_best_node - find the next node to include in a sched_domain
6414 * @node: node whose sched_domain we're building
6415 * @used_nodes: nodes already in the sched_domain
6417 * Find the next node to include in a given scheduling domain. Simply
6418 * finds the closest node not already in the @used_nodes map.
6420 * Should use nodemask_t.
6422 static int find_next_best_node(int node, nodemask_t *used_nodes)
6424 int i, n, val, min_val, best_node = 0;
6428 for (i = 0; i < nr_node_ids; i++) {
6429 /* Start at @node */
6430 n = (node + i) % nr_node_ids;
6432 if (!nr_cpus_node(n))
6435 /* Skip already used nodes */
6436 if (node_isset(n, *used_nodes))
6439 /* Simple min distance search */
6440 val = node_distance(node, n);
6442 if (val < min_val) {
6448 node_set(best_node, *used_nodes);
6453 * sched_domain_node_span - get a cpumask for a node's sched_domain
6454 * @node: node whose cpumask we're constructing
6455 * @span: resulting cpumask
6457 * Given a node, construct a good cpumask for its sched_domain to span. It
6458 * should be one that prevents unnecessary balancing, but also spreads tasks
6461 static void sched_domain_node_span(int node, struct cpumask *span)
6463 nodemask_t used_nodes;
6466 cpumask_clear(span);
6467 nodes_clear(used_nodes);
6469 cpumask_or(span, span, cpumask_of_node(node));
6470 node_set(node, used_nodes);
6472 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6473 int next_node = find_next_best_node(node, &used_nodes);
6475 cpumask_or(span, span, cpumask_of_node(next_node));
6478 #endif /* CONFIG_NUMA */
6480 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6483 * The cpus mask in sched_group and sched_domain hangs off the end.
6485 * ( See the the comments in include/linux/sched.h:struct sched_group
6486 * and struct sched_domain. )
6488 struct static_sched_group {
6489 struct sched_group sg;
6490 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6493 struct static_sched_domain {
6494 struct sched_domain sd;
6495 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6501 cpumask_var_t domainspan;
6502 cpumask_var_t covered;
6503 cpumask_var_t notcovered;
6505 cpumask_var_t nodemask;
6506 cpumask_var_t this_sibling_map;
6507 cpumask_var_t this_core_map;
6508 cpumask_var_t this_book_map;
6509 cpumask_var_t send_covered;
6510 cpumask_var_t tmpmask;
6511 struct sched_group **sched_group_nodes;
6512 struct root_domain *rd;
6516 sa_sched_groups = 0,
6522 sa_this_sibling_map,
6524 sa_sched_group_nodes,
6534 * SMT sched-domains:
6536 #ifdef CONFIG_SCHED_SMT
6537 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6538 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6541 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6542 struct sched_group **sg, struct cpumask *unused)
6545 *sg = &per_cpu(sched_groups, cpu).sg;
6548 #endif /* CONFIG_SCHED_SMT */
6551 * multi-core sched-domains:
6553 #ifdef CONFIG_SCHED_MC
6554 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6555 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6558 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6559 struct sched_group **sg, struct cpumask *mask)
6562 #ifdef CONFIG_SCHED_SMT
6563 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6564 group = cpumask_first(mask);
6569 *sg = &per_cpu(sched_group_core, group).sg;
6572 #endif /* CONFIG_SCHED_MC */
6575 * book sched-domains:
6577 #ifdef CONFIG_SCHED_BOOK
6578 static DEFINE_PER_CPU(struct static_sched_domain, book_domains);
6579 static DEFINE_PER_CPU(struct static_sched_group, sched_group_book);
6582 cpu_to_book_group(int cpu, const struct cpumask *cpu_map,
6583 struct sched_group **sg, struct cpumask *mask)
6586 #ifdef CONFIG_SCHED_MC
6587 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6588 group = cpumask_first(mask);
6589 #elif defined(CONFIG_SCHED_SMT)
6590 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6591 group = cpumask_first(mask);
6594 *sg = &per_cpu(sched_group_book, group).sg;
6597 #endif /* CONFIG_SCHED_BOOK */
6599 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6600 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6603 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6604 struct sched_group **sg, struct cpumask *mask)
6607 #ifdef CONFIG_SCHED_BOOK
6608 cpumask_and(mask, cpu_book_mask(cpu), cpu_map);
6609 group = cpumask_first(mask);
6610 #elif defined(CONFIG_SCHED_MC)
6611 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6612 group = cpumask_first(mask);
6613 #elif defined(CONFIG_SCHED_SMT)
6614 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6615 group = cpumask_first(mask);
6620 *sg = &per_cpu(sched_group_phys, group).sg;
6626 * The init_sched_build_groups can't handle what we want to do with node
6627 * groups, so roll our own. Now each node has its own list of groups which
6628 * gets dynamically allocated.
6630 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6631 static struct sched_group ***sched_group_nodes_bycpu;
6633 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6634 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6636 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6637 struct sched_group **sg,
6638 struct cpumask *nodemask)
6642 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6643 group = cpumask_first(nodemask);
6646 *sg = &per_cpu(sched_group_allnodes, group).sg;
6650 static void init_numa_sched_groups_power(struct sched_group *group_head)
6652 struct sched_group *sg = group_head;
6658 for_each_cpu(j, sched_group_cpus(sg)) {
6659 struct sched_domain *sd;
6661 sd = &per_cpu(phys_domains, j).sd;
6662 if (j != group_first_cpu(sd->groups)) {
6664 * Only add "power" once for each
6670 sg->cpu_power += sd->groups->cpu_power;
6673 } while (sg != group_head);
6676 static int build_numa_sched_groups(struct s_data *d,
6677 const struct cpumask *cpu_map, int num)
6679 struct sched_domain *sd;
6680 struct sched_group *sg, *prev;
6683 cpumask_clear(d->covered);
6684 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6685 if (cpumask_empty(d->nodemask)) {
6686 d->sched_group_nodes[num] = NULL;
6690 sched_domain_node_span(num, d->domainspan);
6691 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6693 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6696 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6700 d->sched_group_nodes[num] = sg;
6702 for_each_cpu(j, d->nodemask) {
6703 sd = &per_cpu(node_domains, j).sd;
6708 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6710 cpumask_or(d->covered, d->covered, d->nodemask);
6713 for (j = 0; j < nr_node_ids; j++) {
6714 n = (num + j) % nr_node_ids;
6715 cpumask_complement(d->notcovered, d->covered);
6716 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6717 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6718 if (cpumask_empty(d->tmpmask))
6720 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6721 if (cpumask_empty(d->tmpmask))
6723 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6727 "Can not alloc domain group for node %d\n", j);
6731 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6732 sg->next = prev->next;
6733 cpumask_or(d->covered, d->covered, d->tmpmask);
6740 #endif /* CONFIG_NUMA */
6743 /* Free memory allocated for various sched_group structures */
6744 static void free_sched_groups(const struct cpumask *cpu_map,
6745 struct cpumask *nodemask)
6749 for_each_cpu(cpu, cpu_map) {
6750 struct sched_group **sched_group_nodes
6751 = sched_group_nodes_bycpu[cpu];
6753 if (!sched_group_nodes)
6756 for (i = 0; i < nr_node_ids; i++) {
6757 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6759 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6760 if (cpumask_empty(nodemask))
6770 if (oldsg != sched_group_nodes[i])
6773 kfree(sched_group_nodes);
6774 sched_group_nodes_bycpu[cpu] = NULL;
6777 #else /* !CONFIG_NUMA */
6778 static void free_sched_groups(const struct cpumask *cpu_map,
6779 struct cpumask *nodemask)
6782 #endif /* CONFIG_NUMA */
6785 * Initialize sched groups cpu_power.
6787 * cpu_power indicates the capacity of sched group, which is used while
6788 * distributing the load between different sched groups in a sched domain.
6789 * Typically cpu_power for all the groups in a sched domain will be same unless
6790 * there are asymmetries in the topology. If there are asymmetries, group
6791 * having more cpu_power will pickup more load compared to the group having
6794 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6796 struct sched_domain *child;
6797 struct sched_group *group;
6801 WARN_ON(!sd || !sd->groups);
6803 if (cpu != group_first_cpu(sd->groups))
6808 sd->groups->cpu_power = 0;
6811 power = SCHED_LOAD_SCALE;
6812 weight = cpumask_weight(sched_domain_span(sd));
6814 * SMT siblings share the power of a single core.
6815 * Usually multiple threads get a better yield out of
6816 * that one core than a single thread would have,
6817 * reflect that in sd->smt_gain.
6819 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6820 power *= sd->smt_gain;
6822 power >>= SCHED_LOAD_SHIFT;
6824 sd->groups->cpu_power += power;
6829 * Add cpu_power of each child group to this groups cpu_power.
6831 group = child->groups;
6833 sd->groups->cpu_power += group->cpu_power;
6834 group = group->next;
6835 } while (group != child->groups);
6839 * Initializers for schedule domains
6840 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6843 #ifdef CONFIG_SCHED_DEBUG
6844 # define SD_INIT_NAME(sd, type) sd->name = #type
6846 # define SD_INIT_NAME(sd, type) do { } while (0)
6849 #define SD_INIT(sd, type) sd_init_##type(sd)
6851 #define SD_INIT_FUNC(type) \
6852 static noinline void sd_init_##type(struct sched_domain *sd) \
6854 memset(sd, 0, sizeof(*sd)); \
6855 *sd = SD_##type##_INIT; \
6856 sd->level = SD_LV_##type; \
6857 SD_INIT_NAME(sd, type); \
6862 SD_INIT_FUNC(ALLNODES)
6865 #ifdef CONFIG_SCHED_SMT
6866 SD_INIT_FUNC(SIBLING)
6868 #ifdef CONFIG_SCHED_MC
6871 #ifdef CONFIG_SCHED_BOOK
6875 static int default_relax_domain_level = -1;
6877 static int __init setup_relax_domain_level(char *str)
6881 val = simple_strtoul(str, NULL, 0);
6882 if (val < SD_LV_MAX)
6883 default_relax_domain_level = val;
6887 __setup("relax_domain_level=", setup_relax_domain_level);
6889 static void set_domain_attribute(struct sched_domain *sd,
6890 struct sched_domain_attr *attr)
6894 if (!attr || attr->relax_domain_level < 0) {
6895 if (default_relax_domain_level < 0)
6898 request = default_relax_domain_level;
6900 request = attr->relax_domain_level;
6901 if (request < sd->level) {
6902 /* turn off idle balance on this domain */
6903 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6905 /* turn on idle balance on this domain */
6906 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6910 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6911 const struct cpumask *cpu_map)
6914 case sa_sched_groups:
6915 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6916 d->sched_group_nodes = NULL;
6918 free_rootdomain(d->rd); /* fall through */
6920 free_cpumask_var(d->tmpmask); /* fall through */
6921 case sa_send_covered:
6922 free_cpumask_var(d->send_covered); /* fall through */
6923 case sa_this_book_map:
6924 free_cpumask_var(d->this_book_map); /* fall through */
6925 case sa_this_core_map:
6926 free_cpumask_var(d->this_core_map); /* fall through */
6927 case sa_this_sibling_map:
6928 free_cpumask_var(d->this_sibling_map); /* fall through */
6930 free_cpumask_var(d->nodemask); /* fall through */
6931 case sa_sched_group_nodes:
6933 kfree(d->sched_group_nodes); /* fall through */
6935 free_cpumask_var(d->notcovered); /* fall through */
6937 free_cpumask_var(d->covered); /* fall through */
6939 free_cpumask_var(d->domainspan); /* fall through */
6946 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6947 const struct cpumask *cpu_map)
6950 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6952 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6953 return sa_domainspan;
6954 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6956 /* Allocate the per-node list of sched groups */
6957 d->sched_group_nodes = kcalloc(nr_node_ids,
6958 sizeof(struct sched_group *), GFP_KERNEL);
6959 if (!d->sched_group_nodes) {
6960 printk(KERN_WARNING "Can not alloc sched group node list\n");
6961 return sa_notcovered;
6963 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6965 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6966 return sa_sched_group_nodes;
6967 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6969 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6970 return sa_this_sibling_map;
6971 if (!alloc_cpumask_var(&d->this_book_map, GFP_KERNEL))
6972 return sa_this_core_map;
6973 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6974 return sa_this_book_map;
6975 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6976 return sa_send_covered;
6977 d->rd = alloc_rootdomain();
6979 printk(KERN_WARNING "Cannot alloc root domain\n");
6982 return sa_rootdomain;
6985 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6986 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6988 struct sched_domain *sd = NULL;
6990 struct sched_domain *parent;
6993 if (cpumask_weight(cpu_map) >
6994 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6995 sd = &per_cpu(allnodes_domains, i).sd;
6996 SD_INIT(sd, ALLNODES);
6997 set_domain_attribute(sd, attr);
6998 cpumask_copy(sched_domain_span(sd), cpu_map);
6999 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
7004 sd = &per_cpu(node_domains, i).sd;
7006 set_domain_attribute(sd, attr);
7007 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7008 sd->parent = parent;
7011 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7016 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7017 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7018 struct sched_domain *parent, int i)
7020 struct sched_domain *sd;
7021 sd = &per_cpu(phys_domains, i).sd;
7023 set_domain_attribute(sd, attr);
7024 cpumask_copy(sched_domain_span(sd), d->nodemask);
7025 sd->parent = parent;
7028 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7032 static struct sched_domain *__build_book_sched_domain(struct s_data *d,
7033 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7034 struct sched_domain *parent, int i)
7036 struct sched_domain *sd = parent;
7037 #ifdef CONFIG_SCHED_BOOK
7038 sd = &per_cpu(book_domains, i).sd;
7040 set_domain_attribute(sd, attr);
7041 cpumask_and(sched_domain_span(sd), cpu_map, cpu_book_mask(i));
7042 sd->parent = parent;
7044 cpu_to_book_group(i, cpu_map, &sd->groups, d->tmpmask);
7049 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7050 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7051 struct sched_domain *parent, int i)
7053 struct sched_domain *sd = parent;
7054 #ifdef CONFIG_SCHED_MC
7055 sd = &per_cpu(core_domains, i).sd;
7057 set_domain_attribute(sd, attr);
7058 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7059 sd->parent = parent;
7061 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7066 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7067 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7068 struct sched_domain *parent, int i)
7070 struct sched_domain *sd = parent;
7071 #ifdef CONFIG_SCHED_SMT
7072 sd = &per_cpu(cpu_domains, i).sd;
7073 SD_INIT(sd, SIBLING);
7074 set_domain_attribute(sd, attr);
7075 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7076 sd->parent = parent;
7078 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7083 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7084 const struct cpumask *cpu_map, int cpu)
7087 #ifdef CONFIG_SCHED_SMT
7088 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7089 cpumask_and(d->this_sibling_map, cpu_map,
7090 topology_thread_cpumask(cpu));
7091 if (cpu == cpumask_first(d->this_sibling_map))
7092 init_sched_build_groups(d->this_sibling_map, cpu_map,
7094 d->send_covered, d->tmpmask);
7097 #ifdef CONFIG_SCHED_MC
7098 case SD_LV_MC: /* set up multi-core groups */
7099 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7100 if (cpu == cpumask_first(d->this_core_map))
7101 init_sched_build_groups(d->this_core_map, cpu_map,
7103 d->send_covered, d->tmpmask);
7106 #ifdef CONFIG_SCHED_BOOK
7107 case SD_LV_BOOK: /* set up book groups */
7108 cpumask_and(d->this_book_map, cpu_map, cpu_book_mask(cpu));
7109 if (cpu == cpumask_first(d->this_book_map))
7110 init_sched_build_groups(d->this_book_map, cpu_map,
7112 d->send_covered, d->tmpmask);
7115 case SD_LV_CPU: /* set up physical groups */
7116 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7117 if (!cpumask_empty(d->nodemask))
7118 init_sched_build_groups(d->nodemask, cpu_map,
7120 d->send_covered, d->tmpmask);
7123 case SD_LV_ALLNODES:
7124 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7125 d->send_covered, d->tmpmask);
7134 * Build sched domains for a given set of cpus and attach the sched domains
7135 * to the individual cpus
7137 static int __build_sched_domains(const struct cpumask *cpu_map,
7138 struct sched_domain_attr *attr)
7140 enum s_alloc alloc_state = sa_none;
7142 struct sched_domain *sd;
7148 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7149 if (alloc_state != sa_rootdomain)
7151 alloc_state = sa_sched_groups;
7154 * Set up domains for cpus specified by the cpu_map.
7156 for_each_cpu(i, cpu_map) {
7157 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7160 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7161 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7162 sd = __build_book_sched_domain(&d, cpu_map, attr, sd, i);
7163 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7164 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7167 for_each_cpu(i, cpu_map) {
7168 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7169 build_sched_groups(&d, SD_LV_BOOK, cpu_map, i);
7170 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7173 /* Set up physical groups */
7174 for (i = 0; i < nr_node_ids; i++)
7175 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7178 /* Set up node groups */
7180 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7182 for (i = 0; i < nr_node_ids; i++)
7183 if (build_numa_sched_groups(&d, cpu_map, i))
7187 /* Calculate CPU power for physical packages and nodes */
7188 #ifdef CONFIG_SCHED_SMT
7189 for_each_cpu(i, cpu_map) {
7190 sd = &per_cpu(cpu_domains, i).sd;
7191 init_sched_groups_power(i, sd);
7194 #ifdef CONFIG_SCHED_MC
7195 for_each_cpu(i, cpu_map) {
7196 sd = &per_cpu(core_domains, i).sd;
7197 init_sched_groups_power(i, sd);
7200 #ifdef CONFIG_SCHED_BOOK
7201 for_each_cpu(i, cpu_map) {
7202 sd = &per_cpu(book_domains, i).sd;
7203 init_sched_groups_power(i, sd);
7207 for_each_cpu(i, cpu_map) {
7208 sd = &per_cpu(phys_domains, i).sd;
7209 init_sched_groups_power(i, sd);
7213 for (i = 0; i < nr_node_ids; i++)
7214 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7216 if (d.sd_allnodes) {
7217 struct sched_group *sg;
7219 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7221 init_numa_sched_groups_power(sg);
7225 /* Attach the domains */
7226 for_each_cpu(i, cpu_map) {
7227 #ifdef CONFIG_SCHED_SMT
7228 sd = &per_cpu(cpu_domains, i).sd;
7229 #elif defined(CONFIG_SCHED_MC)
7230 sd = &per_cpu(core_domains, i).sd;
7231 #elif defined(CONFIG_SCHED_BOOK)
7232 sd = &per_cpu(book_domains, i).sd;
7234 sd = &per_cpu(phys_domains, i).sd;
7236 cpu_attach_domain(sd, d.rd, i);
7239 d.sched_group_nodes = NULL; /* don't free this we still need it */
7240 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7244 __free_domain_allocs(&d, alloc_state, cpu_map);
7248 static int build_sched_domains(const struct cpumask *cpu_map)
7250 return __build_sched_domains(cpu_map, NULL);
7253 static cpumask_var_t *doms_cur; /* current sched domains */
7254 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7255 static struct sched_domain_attr *dattr_cur;
7256 /* attribues of custom domains in 'doms_cur' */
7259 * Special case: If a kmalloc of a doms_cur partition (array of
7260 * cpumask) fails, then fallback to a single sched domain,
7261 * as determined by the single cpumask fallback_doms.
7263 static cpumask_var_t fallback_doms;
7266 * arch_update_cpu_topology lets virtualized architectures update the
7267 * cpu core maps. It is supposed to return 1 if the topology changed
7268 * or 0 if it stayed the same.
7270 int __attribute__((weak)) arch_update_cpu_topology(void)
7275 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7278 cpumask_var_t *doms;
7280 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7283 for (i = 0; i < ndoms; i++) {
7284 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7285 free_sched_domains(doms, i);
7292 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7295 for (i = 0; i < ndoms; i++)
7296 free_cpumask_var(doms[i]);
7301 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7302 * For now this just excludes isolated cpus, but could be used to
7303 * exclude other special cases in the future.
7305 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7309 arch_update_cpu_topology();
7311 doms_cur = alloc_sched_domains(ndoms_cur);
7313 doms_cur = &fallback_doms;
7314 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7316 err = build_sched_domains(doms_cur[0]);
7317 register_sched_domain_sysctl();
7322 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7323 struct cpumask *tmpmask)
7325 free_sched_groups(cpu_map, tmpmask);
7329 * Detach sched domains from a group of cpus specified in cpu_map
7330 * These cpus will now be attached to the NULL domain
7332 static void detach_destroy_domains(const struct cpumask *cpu_map)
7334 /* Save because hotplug lock held. */
7335 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7338 for_each_cpu(i, cpu_map)
7339 cpu_attach_domain(NULL, &def_root_domain, i);
7340 synchronize_sched();
7341 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7344 /* handle null as "default" */
7345 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7346 struct sched_domain_attr *new, int idx_new)
7348 struct sched_domain_attr tmp;
7355 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7356 new ? (new + idx_new) : &tmp,
7357 sizeof(struct sched_domain_attr));
7361 * Partition sched domains as specified by the 'ndoms_new'
7362 * cpumasks in the array doms_new[] of cpumasks. This compares
7363 * doms_new[] to the current sched domain partitioning, doms_cur[].
7364 * It destroys each deleted domain and builds each new domain.
7366 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7367 * The masks don't intersect (don't overlap.) We should setup one
7368 * sched domain for each mask. CPUs not in any of the cpumasks will
7369 * not be load balanced. If the same cpumask appears both in the
7370 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7373 * The passed in 'doms_new' should be allocated using
7374 * alloc_sched_domains. This routine takes ownership of it and will
7375 * free_sched_domains it when done with it. If the caller failed the
7376 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7377 * and partition_sched_domains() will fallback to the single partition
7378 * 'fallback_doms', it also forces the domains to be rebuilt.
7380 * If doms_new == NULL it will be replaced with cpu_online_mask.
7381 * ndoms_new == 0 is a special case for destroying existing domains,
7382 * and it will not create the default domain.
7384 * Call with hotplug lock held
7386 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7387 struct sched_domain_attr *dattr_new)
7392 mutex_lock(&sched_domains_mutex);
7394 /* always unregister in case we don't destroy any domains */
7395 unregister_sched_domain_sysctl();
7397 /* Let architecture update cpu core mappings. */
7398 new_topology = arch_update_cpu_topology();
7400 n = doms_new ? ndoms_new : 0;
7402 /* Destroy deleted domains */
7403 for (i = 0; i < ndoms_cur; i++) {
7404 for (j = 0; j < n && !new_topology; j++) {
7405 if (cpumask_equal(doms_cur[i], doms_new[j])
7406 && dattrs_equal(dattr_cur, i, dattr_new, j))
7409 /* no match - a current sched domain not in new doms_new[] */
7410 detach_destroy_domains(doms_cur[i]);
7415 if (doms_new == NULL) {
7417 doms_new = &fallback_doms;
7418 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7419 WARN_ON_ONCE(dattr_new);
7422 /* Build new domains */
7423 for (i = 0; i < ndoms_new; i++) {
7424 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7425 if (cpumask_equal(doms_new[i], doms_cur[j])
7426 && dattrs_equal(dattr_new, i, dattr_cur, j))
7429 /* no match - add a new doms_new */
7430 __build_sched_domains(doms_new[i],
7431 dattr_new ? dattr_new + i : NULL);
7436 /* Remember the new sched domains */
7437 if (doms_cur != &fallback_doms)
7438 free_sched_domains(doms_cur, ndoms_cur);
7439 kfree(dattr_cur); /* kfree(NULL) is safe */
7440 doms_cur = doms_new;
7441 dattr_cur = dattr_new;
7442 ndoms_cur = ndoms_new;
7444 register_sched_domain_sysctl();
7446 mutex_unlock(&sched_domains_mutex);
7449 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7450 static void arch_reinit_sched_domains(void)
7454 /* Destroy domains first to force the rebuild */
7455 partition_sched_domains(0, NULL, NULL);
7457 rebuild_sched_domains();
7461 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7463 unsigned int level = 0;
7465 if (sscanf(buf, "%u", &level) != 1)
7469 * level is always be positive so don't check for
7470 * level < POWERSAVINGS_BALANCE_NONE which is 0
7471 * What happens on 0 or 1 byte write,
7472 * need to check for count as well?
7475 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7479 sched_smt_power_savings = level;
7481 sched_mc_power_savings = level;
7483 arch_reinit_sched_domains();
7488 #ifdef CONFIG_SCHED_MC
7489 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7490 struct sysdev_class_attribute *attr,
7493 return sprintf(page, "%u\n", sched_mc_power_savings);
7495 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7496 struct sysdev_class_attribute *attr,
7497 const char *buf, size_t count)
7499 return sched_power_savings_store(buf, count, 0);
7501 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7502 sched_mc_power_savings_show,
7503 sched_mc_power_savings_store);
7506 #ifdef CONFIG_SCHED_SMT
7507 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7508 struct sysdev_class_attribute *attr,
7511 return sprintf(page, "%u\n", sched_smt_power_savings);
7513 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7514 struct sysdev_class_attribute *attr,
7515 const char *buf, size_t count)
7517 return sched_power_savings_store(buf, count, 1);
7519 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7520 sched_smt_power_savings_show,
7521 sched_smt_power_savings_store);
7524 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7528 #ifdef CONFIG_SCHED_SMT
7530 err = sysfs_create_file(&cls->kset.kobj,
7531 &attr_sched_smt_power_savings.attr);
7533 #ifdef CONFIG_SCHED_MC
7534 if (!err && mc_capable())
7535 err = sysfs_create_file(&cls->kset.kobj,
7536 &attr_sched_mc_power_savings.attr);
7540 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7543 * Update cpusets according to cpu_active mask. If cpusets are
7544 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7545 * around partition_sched_domains().
7547 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7550 switch (action & ~CPU_TASKS_FROZEN) {
7552 case CPU_DOWN_FAILED:
7553 cpuset_update_active_cpus();
7560 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7563 switch (action & ~CPU_TASKS_FROZEN) {
7564 case CPU_DOWN_PREPARE:
7565 cpuset_update_active_cpus();
7572 static int update_runtime(struct notifier_block *nfb,
7573 unsigned long action, void *hcpu)
7575 int cpu = (int)(long)hcpu;
7578 case CPU_DOWN_PREPARE:
7579 case CPU_DOWN_PREPARE_FROZEN:
7580 disable_runtime(cpu_rq(cpu));
7583 case CPU_DOWN_FAILED:
7584 case CPU_DOWN_FAILED_FROZEN:
7586 case CPU_ONLINE_FROZEN:
7587 enable_runtime(cpu_rq(cpu));
7595 void __init sched_init_smp(void)
7597 cpumask_var_t non_isolated_cpus;
7599 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7600 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7602 #if defined(CONFIG_NUMA)
7603 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7605 BUG_ON(sched_group_nodes_bycpu == NULL);
7608 mutex_lock(&sched_domains_mutex);
7609 arch_init_sched_domains(cpu_active_mask);
7610 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7611 if (cpumask_empty(non_isolated_cpus))
7612 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7613 mutex_unlock(&sched_domains_mutex);
7616 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7617 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7619 /* RT runtime code needs to handle some hotplug events */
7620 hotcpu_notifier(update_runtime, 0);
7624 /* Move init over to a non-isolated CPU */
7625 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7627 sched_init_granularity();
7628 free_cpumask_var(non_isolated_cpus);
7630 init_sched_rt_class();
7633 void __init sched_init_smp(void)
7635 sched_init_granularity();
7637 #endif /* CONFIG_SMP */
7639 const_debug unsigned int sysctl_timer_migration = 1;
7641 int in_sched_functions(unsigned long addr)
7643 return in_lock_functions(addr) ||
7644 (addr >= (unsigned long)__sched_text_start
7645 && addr < (unsigned long)__sched_text_end);
7648 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7650 cfs_rq->tasks_timeline = RB_ROOT;
7651 INIT_LIST_HEAD(&cfs_rq->tasks);
7652 #ifdef CONFIG_FAIR_GROUP_SCHED
7655 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7658 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7660 struct rt_prio_array *array;
7663 array = &rt_rq->active;
7664 for (i = 0; i < MAX_RT_PRIO; i++) {
7665 INIT_LIST_HEAD(array->queue + i);
7666 __clear_bit(i, array->bitmap);
7668 /* delimiter for bitsearch: */
7669 __set_bit(MAX_RT_PRIO, array->bitmap);
7671 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7672 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7674 rt_rq->highest_prio.next = MAX_RT_PRIO;
7678 rt_rq->rt_nr_migratory = 0;
7679 rt_rq->overloaded = 0;
7680 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7684 rt_rq->rt_throttled = 0;
7685 rt_rq->rt_runtime = 0;
7686 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7688 #ifdef CONFIG_RT_GROUP_SCHED
7689 rt_rq->rt_nr_boosted = 0;
7694 #ifdef CONFIG_FAIR_GROUP_SCHED
7695 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7696 struct sched_entity *se, int cpu, int add,
7697 struct sched_entity *parent)
7699 struct rq *rq = cpu_rq(cpu);
7700 tg->cfs_rq[cpu] = cfs_rq;
7701 init_cfs_rq(cfs_rq, rq);
7704 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7707 /* se could be NULL for init_task_group */
7712 se->cfs_rq = &rq->cfs;
7714 se->cfs_rq = parent->my_q;
7717 se->load.weight = tg->shares;
7718 se->load.inv_weight = 0;
7719 se->parent = parent;
7723 #ifdef CONFIG_RT_GROUP_SCHED
7724 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7725 struct sched_rt_entity *rt_se, int cpu, int add,
7726 struct sched_rt_entity *parent)
7728 struct rq *rq = cpu_rq(cpu);
7730 tg->rt_rq[cpu] = rt_rq;
7731 init_rt_rq(rt_rq, rq);
7733 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7735 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7737 tg->rt_se[cpu] = rt_se;
7742 rt_se->rt_rq = &rq->rt;
7744 rt_se->rt_rq = parent->my_q;
7746 rt_se->my_q = rt_rq;
7747 rt_se->parent = parent;
7748 INIT_LIST_HEAD(&rt_se->run_list);
7752 void __init sched_init(void)
7755 unsigned long alloc_size = 0, ptr;
7757 #ifdef CONFIG_FAIR_GROUP_SCHED
7758 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7760 #ifdef CONFIG_RT_GROUP_SCHED
7761 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7763 #ifdef CONFIG_CPUMASK_OFFSTACK
7764 alloc_size += num_possible_cpus() * cpumask_size();
7767 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7769 #ifdef CONFIG_FAIR_GROUP_SCHED
7770 init_task_group.se = (struct sched_entity **)ptr;
7771 ptr += nr_cpu_ids * sizeof(void **);
7773 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7774 ptr += nr_cpu_ids * sizeof(void **);
7776 #endif /* CONFIG_FAIR_GROUP_SCHED */
7777 #ifdef CONFIG_RT_GROUP_SCHED
7778 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7779 ptr += nr_cpu_ids * sizeof(void **);
7781 init_task_group.rt_rq = (struct rt_rq **)ptr;
7782 ptr += nr_cpu_ids * sizeof(void **);
7784 #endif /* CONFIG_RT_GROUP_SCHED */
7785 #ifdef CONFIG_CPUMASK_OFFSTACK
7786 for_each_possible_cpu(i) {
7787 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7788 ptr += cpumask_size();
7790 #endif /* CONFIG_CPUMASK_OFFSTACK */
7794 init_defrootdomain();
7797 init_rt_bandwidth(&def_rt_bandwidth,
7798 global_rt_period(), global_rt_runtime());
7800 #ifdef CONFIG_RT_GROUP_SCHED
7801 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7802 global_rt_period(), global_rt_runtime());
7803 #endif /* CONFIG_RT_GROUP_SCHED */
7805 #ifdef CONFIG_CGROUP_SCHED
7806 list_add(&init_task_group.list, &task_groups);
7807 INIT_LIST_HEAD(&init_task_group.children);
7809 #endif /* CONFIG_CGROUP_SCHED */
7811 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7812 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7813 __alignof__(unsigned long));
7815 for_each_possible_cpu(i) {
7819 raw_spin_lock_init(&rq->lock);
7821 rq->calc_load_active = 0;
7822 rq->calc_load_update = jiffies + LOAD_FREQ;
7823 init_cfs_rq(&rq->cfs, rq);
7824 init_rt_rq(&rq->rt, rq);
7825 #ifdef CONFIG_FAIR_GROUP_SCHED
7826 init_task_group.shares = init_task_group_load;
7827 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7828 #ifdef CONFIG_CGROUP_SCHED
7830 * How much cpu bandwidth does init_task_group get?
7832 * In case of task-groups formed thr' the cgroup filesystem, it
7833 * gets 100% of the cpu resources in the system. This overall
7834 * system cpu resource is divided among the tasks of
7835 * init_task_group and its child task-groups in a fair manner,
7836 * based on each entity's (task or task-group's) weight
7837 * (se->load.weight).
7839 * In other words, if init_task_group has 10 tasks of weight
7840 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7841 * then A0's share of the cpu resource is:
7843 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7845 * We achieve this by letting init_task_group's tasks sit
7846 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7848 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7850 #endif /* CONFIG_FAIR_GROUP_SCHED */
7852 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7853 #ifdef CONFIG_RT_GROUP_SCHED
7854 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7855 #ifdef CONFIG_CGROUP_SCHED
7856 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7860 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7861 rq->cpu_load[j] = 0;
7863 rq->last_load_update_tick = jiffies;
7868 rq->cpu_power = SCHED_LOAD_SCALE;
7869 rq->post_schedule = 0;
7870 rq->active_balance = 0;
7871 rq->next_balance = jiffies;
7876 rq->avg_idle = 2*sysctl_sched_migration_cost;
7877 rq_attach_root(rq, &def_root_domain);
7879 rq->nohz_balance_kick = 0;
7880 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7884 atomic_set(&rq->nr_iowait, 0);
7887 set_load_weight(&init_task);
7889 #ifdef CONFIG_PREEMPT_NOTIFIERS
7890 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7894 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7897 #ifdef CONFIG_RT_MUTEXES
7898 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7902 * The boot idle thread does lazy MMU switching as well:
7904 atomic_inc(&init_mm.mm_count);
7905 enter_lazy_tlb(&init_mm, current);
7908 * Make us the idle thread. Technically, schedule() should not be
7909 * called from this thread, however somewhere below it might be,
7910 * but because we are the idle thread, we just pick up running again
7911 * when this runqueue becomes "idle".
7913 init_idle(current, smp_processor_id());
7915 calc_load_update = jiffies + LOAD_FREQ;
7918 * During early bootup we pretend to be a normal task:
7920 current->sched_class = &fair_sched_class;
7922 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7923 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7926 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7927 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
7928 atomic_set(&nohz.load_balancer, nr_cpu_ids);
7929 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
7930 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
7932 /* May be allocated at isolcpus cmdline parse time */
7933 if (cpu_isolated_map == NULL)
7934 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7939 scheduler_running = 1;
7942 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7943 static inline int preempt_count_equals(int preempt_offset)
7945 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7947 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7950 void __might_sleep(const char *file, int line, int preempt_offset)
7953 static unsigned long prev_jiffy; /* ratelimiting */
7955 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7956 system_state != SYSTEM_RUNNING || oops_in_progress)
7958 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7960 prev_jiffy = jiffies;
7963 "BUG: sleeping function called from invalid context at %s:%d\n",
7966 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7967 in_atomic(), irqs_disabled(),
7968 current->pid, current->comm);
7970 debug_show_held_locks(current);
7971 if (irqs_disabled())
7972 print_irqtrace_events(current);
7976 EXPORT_SYMBOL(__might_sleep);
7979 #ifdef CONFIG_MAGIC_SYSRQ
7980 static void normalize_task(struct rq *rq, struct task_struct *p)
7984 on_rq = p->se.on_rq;
7986 deactivate_task(rq, p, 0);
7987 __setscheduler(rq, p, SCHED_NORMAL, 0);
7989 activate_task(rq, p, 0);
7990 resched_task(rq->curr);
7994 void normalize_rt_tasks(void)
7996 struct task_struct *g, *p;
7997 unsigned long flags;
8000 read_lock_irqsave(&tasklist_lock, flags);
8001 do_each_thread(g, p) {
8003 * Only normalize user tasks:
8008 p->se.exec_start = 0;
8009 #ifdef CONFIG_SCHEDSTATS
8010 p->se.statistics.wait_start = 0;
8011 p->se.statistics.sleep_start = 0;
8012 p->se.statistics.block_start = 0;
8017 * Renice negative nice level userspace
8020 if (TASK_NICE(p) < 0 && p->mm)
8021 set_user_nice(p, 0);
8025 raw_spin_lock(&p->pi_lock);
8026 rq = __task_rq_lock(p);
8028 normalize_task(rq, p);
8030 __task_rq_unlock(rq);
8031 raw_spin_unlock(&p->pi_lock);
8032 } while_each_thread(g, p);
8034 read_unlock_irqrestore(&tasklist_lock, flags);
8037 #endif /* CONFIG_MAGIC_SYSRQ */
8039 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8041 * These functions are only useful for the IA64 MCA handling, or kdb.
8043 * They can only be called when the whole system has been
8044 * stopped - every CPU needs to be quiescent, and no scheduling
8045 * activity can take place. Using them for anything else would
8046 * be a serious bug, and as a result, they aren't even visible
8047 * under any other configuration.
8051 * curr_task - return the current task for a given cpu.
8052 * @cpu: the processor in question.
8054 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8056 struct task_struct *curr_task(int cpu)
8058 return cpu_curr(cpu);
8061 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8065 * set_curr_task - set the current task for a given cpu.
8066 * @cpu: the processor in question.
8067 * @p: the task pointer to set.
8069 * Description: This function must only be used when non-maskable interrupts
8070 * are serviced on a separate stack. It allows the architecture to switch the
8071 * notion of the current task on a cpu in a non-blocking manner. This function
8072 * must be called with all CPU's synchronized, and interrupts disabled, the
8073 * and caller must save the original value of the current task (see
8074 * curr_task() above) and restore that value before reenabling interrupts and
8075 * re-starting the system.
8077 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8079 void set_curr_task(int cpu, struct task_struct *p)
8086 #ifdef CONFIG_FAIR_GROUP_SCHED
8087 static void free_fair_sched_group(struct task_group *tg)
8091 for_each_possible_cpu(i) {
8093 kfree(tg->cfs_rq[i]);
8103 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8105 struct cfs_rq *cfs_rq;
8106 struct sched_entity *se;
8110 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8113 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8117 tg->shares = NICE_0_LOAD;
8119 for_each_possible_cpu(i) {
8122 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8123 GFP_KERNEL, cpu_to_node(i));
8127 se = kzalloc_node(sizeof(struct sched_entity),
8128 GFP_KERNEL, cpu_to_node(i));
8132 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8143 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8145 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8146 &cpu_rq(cpu)->leaf_cfs_rq_list);
8149 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8151 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8153 #else /* !CONFG_FAIR_GROUP_SCHED */
8154 static inline void free_fair_sched_group(struct task_group *tg)
8159 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8164 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8168 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8171 #endif /* CONFIG_FAIR_GROUP_SCHED */
8173 #ifdef CONFIG_RT_GROUP_SCHED
8174 static void free_rt_sched_group(struct task_group *tg)
8178 destroy_rt_bandwidth(&tg->rt_bandwidth);
8180 for_each_possible_cpu(i) {
8182 kfree(tg->rt_rq[i]);
8184 kfree(tg->rt_se[i]);
8192 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8194 struct rt_rq *rt_rq;
8195 struct sched_rt_entity *rt_se;
8199 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8202 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8206 init_rt_bandwidth(&tg->rt_bandwidth,
8207 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8209 for_each_possible_cpu(i) {
8212 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8213 GFP_KERNEL, cpu_to_node(i));
8217 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8218 GFP_KERNEL, cpu_to_node(i));
8222 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8233 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8235 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8236 &cpu_rq(cpu)->leaf_rt_rq_list);
8239 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8241 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8243 #else /* !CONFIG_RT_GROUP_SCHED */
8244 static inline void free_rt_sched_group(struct task_group *tg)
8249 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8254 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8258 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8261 #endif /* CONFIG_RT_GROUP_SCHED */
8263 #ifdef CONFIG_CGROUP_SCHED
8264 static void free_sched_group(struct task_group *tg)
8266 free_fair_sched_group(tg);
8267 free_rt_sched_group(tg);
8271 /* allocate runqueue etc for a new task group */
8272 struct task_group *sched_create_group(struct task_group *parent)
8274 struct task_group *tg;
8275 unsigned long flags;
8278 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8280 return ERR_PTR(-ENOMEM);
8282 if (!alloc_fair_sched_group(tg, parent))
8285 if (!alloc_rt_sched_group(tg, parent))
8288 spin_lock_irqsave(&task_group_lock, flags);
8289 for_each_possible_cpu(i) {
8290 register_fair_sched_group(tg, i);
8291 register_rt_sched_group(tg, i);
8293 list_add_rcu(&tg->list, &task_groups);
8295 WARN_ON(!parent); /* root should already exist */
8297 tg->parent = parent;
8298 INIT_LIST_HEAD(&tg->children);
8299 list_add_rcu(&tg->siblings, &parent->children);
8300 spin_unlock_irqrestore(&task_group_lock, flags);
8305 free_sched_group(tg);
8306 return ERR_PTR(-ENOMEM);
8309 /* rcu callback to free various structures associated with a task group */
8310 static void free_sched_group_rcu(struct rcu_head *rhp)
8312 /* now it should be safe to free those cfs_rqs */
8313 free_sched_group(container_of(rhp, struct task_group, rcu));
8316 /* Destroy runqueue etc associated with a task group */
8317 void sched_destroy_group(struct task_group *tg)
8319 unsigned long flags;
8322 spin_lock_irqsave(&task_group_lock, flags);
8323 for_each_possible_cpu(i) {
8324 unregister_fair_sched_group(tg, i);
8325 unregister_rt_sched_group(tg, i);
8327 list_del_rcu(&tg->list);
8328 list_del_rcu(&tg->siblings);
8329 spin_unlock_irqrestore(&task_group_lock, flags);
8331 /* wait for possible concurrent references to cfs_rqs complete */
8332 call_rcu(&tg->rcu, free_sched_group_rcu);
8335 /* change task's runqueue when it moves between groups.
8336 * The caller of this function should have put the task in its new group
8337 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8338 * reflect its new group.
8340 void sched_move_task(struct task_struct *tsk)
8343 unsigned long flags;
8346 rq = task_rq_lock(tsk, &flags);
8348 running = task_current(rq, tsk);
8349 on_rq = tsk->se.on_rq;
8352 dequeue_task(rq, tsk, 0);
8353 if (unlikely(running))
8354 tsk->sched_class->put_prev_task(rq, tsk);
8356 set_task_rq(tsk, task_cpu(tsk));
8358 #ifdef CONFIG_FAIR_GROUP_SCHED
8359 if (tsk->sched_class->moved_group)
8360 tsk->sched_class->moved_group(tsk, on_rq);
8363 if (unlikely(running))
8364 tsk->sched_class->set_curr_task(rq);
8366 enqueue_task(rq, tsk, 0);
8368 task_rq_unlock(rq, &flags);
8370 #endif /* CONFIG_CGROUP_SCHED */
8372 #ifdef CONFIG_FAIR_GROUP_SCHED
8373 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8375 struct cfs_rq *cfs_rq = se->cfs_rq;
8380 dequeue_entity(cfs_rq, se, 0);
8382 se->load.weight = shares;
8383 se->load.inv_weight = 0;
8386 enqueue_entity(cfs_rq, se, 0);
8389 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8391 struct cfs_rq *cfs_rq = se->cfs_rq;
8392 struct rq *rq = cfs_rq->rq;
8393 unsigned long flags;
8395 raw_spin_lock_irqsave(&rq->lock, flags);
8396 __set_se_shares(se, shares);
8397 raw_spin_unlock_irqrestore(&rq->lock, flags);
8400 static DEFINE_MUTEX(shares_mutex);
8402 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8405 unsigned long flags;
8408 * We can't change the weight of the root cgroup.
8413 if (shares < MIN_SHARES)
8414 shares = MIN_SHARES;
8415 else if (shares > MAX_SHARES)
8416 shares = MAX_SHARES;
8418 mutex_lock(&shares_mutex);
8419 if (tg->shares == shares)
8422 spin_lock_irqsave(&task_group_lock, flags);
8423 for_each_possible_cpu(i)
8424 unregister_fair_sched_group(tg, i);
8425 list_del_rcu(&tg->siblings);
8426 spin_unlock_irqrestore(&task_group_lock, flags);
8428 /* wait for any ongoing reference to this group to finish */
8429 synchronize_sched();
8432 * Now we are free to modify the group's share on each cpu
8433 * w/o tripping rebalance_share or load_balance_fair.
8435 tg->shares = shares;
8436 for_each_possible_cpu(i) {
8440 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8441 set_se_shares(tg->se[i], shares);
8445 * Enable load balance activity on this group, by inserting it back on
8446 * each cpu's rq->leaf_cfs_rq_list.
8448 spin_lock_irqsave(&task_group_lock, flags);
8449 for_each_possible_cpu(i)
8450 register_fair_sched_group(tg, i);
8451 list_add_rcu(&tg->siblings, &tg->parent->children);
8452 spin_unlock_irqrestore(&task_group_lock, flags);
8454 mutex_unlock(&shares_mutex);
8458 unsigned long sched_group_shares(struct task_group *tg)
8464 #ifdef CONFIG_RT_GROUP_SCHED
8466 * Ensure that the real time constraints are schedulable.
8468 static DEFINE_MUTEX(rt_constraints_mutex);
8470 static unsigned long to_ratio(u64 period, u64 runtime)
8472 if (runtime == RUNTIME_INF)
8475 return div64_u64(runtime << 20, period);
8478 /* Must be called with tasklist_lock held */
8479 static inline int tg_has_rt_tasks(struct task_group *tg)
8481 struct task_struct *g, *p;
8483 do_each_thread(g, p) {
8484 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8486 } while_each_thread(g, p);
8491 struct rt_schedulable_data {
8492 struct task_group *tg;
8497 static int tg_schedulable(struct task_group *tg, void *data)
8499 struct rt_schedulable_data *d = data;
8500 struct task_group *child;
8501 unsigned long total, sum = 0;
8502 u64 period, runtime;
8504 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8505 runtime = tg->rt_bandwidth.rt_runtime;
8508 period = d->rt_period;
8509 runtime = d->rt_runtime;
8513 * Cannot have more runtime than the period.
8515 if (runtime > period && runtime != RUNTIME_INF)
8519 * Ensure we don't starve existing RT tasks.
8521 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8524 total = to_ratio(period, runtime);
8527 * Nobody can have more than the global setting allows.
8529 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8533 * The sum of our children's runtime should not exceed our own.
8535 list_for_each_entry_rcu(child, &tg->children, siblings) {
8536 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8537 runtime = child->rt_bandwidth.rt_runtime;
8539 if (child == d->tg) {
8540 period = d->rt_period;
8541 runtime = d->rt_runtime;
8544 sum += to_ratio(period, runtime);
8553 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8555 struct rt_schedulable_data data = {
8557 .rt_period = period,
8558 .rt_runtime = runtime,
8561 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8564 static int tg_set_bandwidth(struct task_group *tg,
8565 u64 rt_period, u64 rt_runtime)
8569 mutex_lock(&rt_constraints_mutex);
8570 read_lock(&tasklist_lock);
8571 err = __rt_schedulable(tg, rt_period, rt_runtime);
8575 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8576 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8577 tg->rt_bandwidth.rt_runtime = rt_runtime;
8579 for_each_possible_cpu(i) {
8580 struct rt_rq *rt_rq = tg->rt_rq[i];
8582 raw_spin_lock(&rt_rq->rt_runtime_lock);
8583 rt_rq->rt_runtime = rt_runtime;
8584 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8586 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8588 read_unlock(&tasklist_lock);
8589 mutex_unlock(&rt_constraints_mutex);
8594 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8596 u64 rt_runtime, rt_period;
8598 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8599 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8600 if (rt_runtime_us < 0)
8601 rt_runtime = RUNTIME_INF;
8603 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8606 long sched_group_rt_runtime(struct task_group *tg)
8610 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8613 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8614 do_div(rt_runtime_us, NSEC_PER_USEC);
8615 return rt_runtime_us;
8618 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8620 u64 rt_runtime, rt_period;
8622 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8623 rt_runtime = tg->rt_bandwidth.rt_runtime;
8628 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8631 long sched_group_rt_period(struct task_group *tg)
8635 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8636 do_div(rt_period_us, NSEC_PER_USEC);
8637 return rt_period_us;
8640 static int sched_rt_global_constraints(void)
8642 u64 runtime, period;
8645 if (sysctl_sched_rt_period <= 0)
8648 runtime = global_rt_runtime();
8649 period = global_rt_period();
8652 * Sanity check on the sysctl variables.
8654 if (runtime > period && runtime != RUNTIME_INF)
8657 mutex_lock(&rt_constraints_mutex);
8658 read_lock(&tasklist_lock);
8659 ret = __rt_schedulable(NULL, 0, 0);
8660 read_unlock(&tasklist_lock);
8661 mutex_unlock(&rt_constraints_mutex);
8666 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8668 /* Don't accept realtime tasks when there is no way for them to run */
8669 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8675 #else /* !CONFIG_RT_GROUP_SCHED */
8676 static int sched_rt_global_constraints(void)
8678 unsigned long flags;
8681 if (sysctl_sched_rt_period <= 0)
8685 * There's always some RT tasks in the root group
8686 * -- migration, kstopmachine etc..
8688 if (sysctl_sched_rt_runtime == 0)
8691 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8692 for_each_possible_cpu(i) {
8693 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8695 raw_spin_lock(&rt_rq->rt_runtime_lock);
8696 rt_rq->rt_runtime = global_rt_runtime();
8697 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8699 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8703 #endif /* CONFIG_RT_GROUP_SCHED */
8705 int sched_rt_handler(struct ctl_table *table, int write,
8706 void __user *buffer, size_t *lenp,
8710 int old_period, old_runtime;
8711 static DEFINE_MUTEX(mutex);
8714 old_period = sysctl_sched_rt_period;
8715 old_runtime = sysctl_sched_rt_runtime;
8717 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8719 if (!ret && write) {
8720 ret = sched_rt_global_constraints();
8722 sysctl_sched_rt_period = old_period;
8723 sysctl_sched_rt_runtime = old_runtime;
8725 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8726 def_rt_bandwidth.rt_period =
8727 ns_to_ktime(global_rt_period());
8730 mutex_unlock(&mutex);
8735 #ifdef CONFIG_CGROUP_SCHED
8737 /* return corresponding task_group object of a cgroup */
8738 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8740 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8741 struct task_group, css);
8744 static struct cgroup_subsys_state *
8745 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8747 struct task_group *tg, *parent;
8749 if (!cgrp->parent) {
8750 /* This is early initialization for the top cgroup */
8751 return &init_task_group.css;
8754 parent = cgroup_tg(cgrp->parent);
8755 tg = sched_create_group(parent);
8757 return ERR_PTR(-ENOMEM);
8763 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8765 struct task_group *tg = cgroup_tg(cgrp);
8767 sched_destroy_group(tg);
8771 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8773 #ifdef CONFIG_RT_GROUP_SCHED
8774 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8777 /* We don't support RT-tasks being in separate groups */
8778 if (tsk->sched_class != &fair_sched_class)
8785 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8786 struct task_struct *tsk, bool threadgroup)
8788 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8792 struct task_struct *c;
8794 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8795 retval = cpu_cgroup_can_attach_task(cgrp, c);
8807 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8808 struct cgroup *old_cont, struct task_struct *tsk,
8811 sched_move_task(tsk);
8813 struct task_struct *c;
8815 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8822 #ifdef CONFIG_FAIR_GROUP_SCHED
8823 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8826 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8829 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8831 struct task_group *tg = cgroup_tg(cgrp);
8833 return (u64) tg->shares;
8835 #endif /* CONFIG_FAIR_GROUP_SCHED */
8837 #ifdef CONFIG_RT_GROUP_SCHED
8838 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8841 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8844 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8846 return sched_group_rt_runtime(cgroup_tg(cgrp));
8849 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8852 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8855 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8857 return sched_group_rt_period(cgroup_tg(cgrp));
8859 #endif /* CONFIG_RT_GROUP_SCHED */
8861 static struct cftype cpu_files[] = {
8862 #ifdef CONFIG_FAIR_GROUP_SCHED
8865 .read_u64 = cpu_shares_read_u64,
8866 .write_u64 = cpu_shares_write_u64,
8869 #ifdef CONFIG_RT_GROUP_SCHED
8871 .name = "rt_runtime_us",
8872 .read_s64 = cpu_rt_runtime_read,
8873 .write_s64 = cpu_rt_runtime_write,
8876 .name = "rt_period_us",
8877 .read_u64 = cpu_rt_period_read_uint,
8878 .write_u64 = cpu_rt_period_write_uint,
8883 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8885 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8888 struct cgroup_subsys cpu_cgroup_subsys = {
8890 .create = cpu_cgroup_create,
8891 .destroy = cpu_cgroup_destroy,
8892 .can_attach = cpu_cgroup_can_attach,
8893 .attach = cpu_cgroup_attach,
8894 .populate = cpu_cgroup_populate,
8895 .subsys_id = cpu_cgroup_subsys_id,
8899 #endif /* CONFIG_CGROUP_SCHED */
8901 #ifdef CONFIG_CGROUP_CPUACCT
8904 * CPU accounting code for task groups.
8906 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8907 * (balbir@in.ibm.com).
8910 /* track cpu usage of a group of tasks and its child groups */
8912 struct cgroup_subsys_state css;
8913 /* cpuusage holds pointer to a u64-type object on every cpu */
8914 u64 __percpu *cpuusage;
8915 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8916 struct cpuacct *parent;
8919 struct cgroup_subsys cpuacct_subsys;
8921 /* return cpu accounting group corresponding to this container */
8922 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8924 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8925 struct cpuacct, css);
8928 /* return cpu accounting group to which this task belongs */
8929 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8931 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8932 struct cpuacct, css);
8935 /* create a new cpu accounting group */
8936 static struct cgroup_subsys_state *cpuacct_create(
8937 struct cgroup_subsys *ss, struct cgroup *cgrp)
8939 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8945 ca->cpuusage = alloc_percpu(u64);
8949 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8950 if (percpu_counter_init(&ca->cpustat[i], 0))
8951 goto out_free_counters;
8954 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 struct cftype files[] = {
9085 .read_u64 = cpuusage_read,
9086 .write_u64 = cpuusage_write,
9089 .name = "usage_percpu",
9090 .read_seq_string = cpuacct_percpu_seq_read,
9094 .read_map = cpuacct_stats_show,
9098 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9100 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9104 * charge this task's execution time to its accounting group.
9106 * called with rq->lock held.
9108 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9113 if (unlikely(!cpuacct_subsys.active))
9116 cpu = task_cpu(tsk);
9122 for (; ca; ca = ca->parent) {
9123 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9124 *cpuusage += cputime;
9131 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9132 * in cputime_t units. As a result, cpuacct_update_stats calls
9133 * percpu_counter_add with values large enough to always overflow the
9134 * per cpu batch limit causing bad SMP scalability.
9136 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9137 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9138 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9141 #define CPUACCT_BATCH \
9142 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9144 #define CPUACCT_BATCH 0
9148 * Charge the system/user time to the task's accounting group.
9150 static void cpuacct_update_stats(struct task_struct *tsk,
9151 enum cpuacct_stat_index idx, cputime_t val)
9154 int batch = CPUACCT_BATCH;
9156 if (unlikely(!cpuacct_subsys.active))
9163 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9169 struct cgroup_subsys cpuacct_subsys = {
9171 .create = cpuacct_create,
9172 .destroy = cpuacct_destroy,
9173 .populate = cpuacct_populate,
9174 .subsys_id = cpuacct_subsys_id,
9176 #endif /* CONFIG_CGROUP_CPUACCT */
9180 void synchronize_sched_expedited(void)
9184 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9186 #else /* #ifndef CONFIG_SMP */
9188 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9190 static int synchronize_sched_expedited_cpu_stop(void *data)
9193 * There must be a full memory barrier on each affected CPU
9194 * between the time that try_stop_cpus() is called and the
9195 * time that it returns.
9197 * In the current initial implementation of cpu_stop, the
9198 * above condition is already met when the control reaches
9199 * this point and the following smp_mb() is not strictly
9200 * necessary. Do smp_mb() anyway for documentation and
9201 * robustness against future implementation changes.
9203 smp_mb(); /* See above comment block. */
9208 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9209 * approach to force grace period to end quickly. This consumes
9210 * significant time on all CPUs, and is thus not recommended for
9211 * any sort of common-case code.
9213 * Note that it is illegal to call this function while holding any
9214 * lock that is acquired by a CPU-hotplug notifier. Failing to
9215 * observe this restriction will result in deadlock.
9217 void synchronize_sched_expedited(void)
9219 int snap, trycount = 0;
9221 smp_mb(); /* ensure prior mod happens before capturing snap. */
9222 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9224 while (try_stop_cpus(cpu_online_mask,
9225 synchronize_sched_expedited_cpu_stop,
9228 if (trycount++ < 10)
9229 udelay(trycount * num_online_cpus());
9231 synchronize_sched();
9234 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9235 smp_mb(); /* ensure test happens before caller kfree */
9240 atomic_inc(&synchronize_sched_expedited_count);
9241 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9244 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9246 #endif /* #else #ifndef CONFIG_SMP */