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;
430 struct cpupri cpupri;
435 * By default the system creates a single root-domain with all cpus as
436 * members (mimicking the global state we have today).
438 static struct root_domain def_root_domain;
443 * This is the main, per-CPU runqueue data structure.
445 * Locking rule: those places that want to lock multiple runqueues
446 * (such as the load balancing or the thread migration code), lock
447 * acquire operations must be ordered by ascending &runqueue.
454 * nr_running and cpu_load should be in the same cacheline because
455 * remote CPUs use both these fields when doing load calculation.
457 unsigned long nr_running;
458 #define CPU_LOAD_IDX_MAX 5
459 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
460 unsigned long last_load_update_tick;
463 unsigned char in_nohz_recently;
465 unsigned int skip_clock_update;
467 /* capture load from *all* tasks on this cpu: */
468 struct load_weight load;
469 unsigned long nr_load_updates;
475 #ifdef CONFIG_FAIR_GROUP_SCHED
476 /* list of leaf cfs_rq on this cpu: */
477 struct list_head leaf_cfs_rq_list;
479 #ifdef CONFIG_RT_GROUP_SCHED
480 struct list_head leaf_rt_rq_list;
484 * This is part of a global counter where only the total sum
485 * over all CPUs matters. A task can increase this counter on
486 * one CPU and if it got migrated afterwards it may decrease
487 * it on another CPU. Always updated under the runqueue lock:
489 unsigned long nr_uninterruptible;
491 struct task_struct *curr, *idle;
492 unsigned long next_balance;
493 struct mm_struct *prev_mm;
500 struct root_domain *rd;
501 struct sched_domain *sd;
503 unsigned long cpu_power;
505 unsigned char idle_at_tick;
506 /* For active balancing */
510 struct cpu_stop_work active_balance_work;
511 /* cpu of this runqueue: */
515 unsigned long avg_load_per_task;
523 /* calc_load related fields */
524 unsigned long calc_load_update;
525 long calc_load_active;
527 #ifdef CONFIG_SCHED_HRTICK
529 int hrtick_csd_pending;
530 struct call_single_data hrtick_csd;
532 struct hrtimer hrtick_timer;
535 #ifdef CONFIG_SCHEDSTATS
537 struct sched_info rq_sched_info;
538 unsigned long long rq_cpu_time;
539 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
541 /* sys_sched_yield() stats */
542 unsigned int yld_count;
544 /* schedule() stats */
545 unsigned int sched_switch;
546 unsigned int sched_count;
547 unsigned int sched_goidle;
549 /* try_to_wake_up() stats */
550 unsigned int ttwu_count;
551 unsigned int ttwu_local;
554 unsigned int bkl_count;
558 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
561 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
563 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
566 * A queue event has occurred, and we're going to schedule. In
567 * this case, we can save a useless back to back clock update.
569 if (test_tsk_need_resched(p))
570 rq->skip_clock_update = 1;
573 static inline int cpu_of(struct rq *rq)
582 #define rcu_dereference_check_sched_domain(p) \
583 rcu_dereference_check((p), \
584 rcu_read_lock_sched_held() || \
585 lockdep_is_held(&sched_domains_mutex))
588 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
589 * See detach_destroy_domains: synchronize_sched for details.
591 * The domain tree of any CPU may only be accessed from within
592 * preempt-disabled sections.
594 #define for_each_domain(cpu, __sd) \
595 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
597 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
598 #define this_rq() (&__get_cpu_var(runqueues))
599 #define task_rq(p) cpu_rq(task_cpu(p))
600 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
601 #define raw_rq() (&__raw_get_cpu_var(runqueues))
603 #ifdef CONFIG_CGROUP_SCHED
606 * Return the group to which this tasks belongs.
608 * We use task_subsys_state_check() and extend the RCU verification
609 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
610 * holds that lock for each task it moves into the cgroup. Therefore
611 * by holding that lock, we pin the task to the current cgroup.
613 static inline struct task_group *task_group(struct task_struct *p)
615 struct cgroup_subsys_state *css;
617 css = task_subsys_state_check(p, cpu_cgroup_subsys_id,
618 lockdep_is_held(&task_rq(p)->lock));
619 return container_of(css, struct task_group, css);
622 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
623 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
625 #ifdef CONFIG_FAIR_GROUP_SCHED
626 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
627 p->se.parent = task_group(p)->se[cpu];
630 #ifdef CONFIG_RT_GROUP_SCHED
631 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
632 p->rt.parent = task_group(p)->rt_se[cpu];
636 #else /* CONFIG_CGROUP_SCHED */
638 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
639 static inline struct task_group *task_group(struct task_struct *p)
644 #endif /* CONFIG_CGROUP_SCHED */
646 inline void update_rq_clock(struct rq *rq)
648 if (!rq->skip_clock_update)
649 rq->clock = sched_clock_cpu(cpu_of(rq));
653 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
655 #ifdef CONFIG_SCHED_DEBUG
656 # define const_debug __read_mostly
658 # define const_debug static const
663 * @cpu: the processor in question.
665 * Returns true if the current cpu runqueue is locked.
666 * This interface allows printk to be called with the runqueue lock
667 * held and know whether or not it is OK to wake up the klogd.
669 int runqueue_is_locked(int cpu)
671 return raw_spin_is_locked(&cpu_rq(cpu)->lock);
675 * Debugging: various feature bits
678 #define SCHED_FEAT(name, enabled) \
679 __SCHED_FEAT_##name ,
682 #include "sched_features.h"
687 #define SCHED_FEAT(name, enabled) \
688 (1UL << __SCHED_FEAT_##name) * enabled |
690 const_debug unsigned int sysctl_sched_features =
691 #include "sched_features.h"
696 #ifdef CONFIG_SCHED_DEBUG
697 #define SCHED_FEAT(name, enabled) \
700 static __read_mostly char *sched_feat_names[] = {
701 #include "sched_features.h"
707 static int sched_feat_show(struct seq_file *m, void *v)
711 for (i = 0; sched_feat_names[i]; i++) {
712 if (!(sysctl_sched_features & (1UL << i)))
714 seq_printf(m, "%s ", sched_feat_names[i]);
722 sched_feat_write(struct file *filp, const char __user *ubuf,
723 size_t cnt, loff_t *ppos)
733 if (copy_from_user(&buf, ubuf, cnt))
738 if (strncmp(buf, "NO_", 3) == 0) {
743 for (i = 0; sched_feat_names[i]; i++) {
744 int len = strlen(sched_feat_names[i]);
746 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
748 sysctl_sched_features &= ~(1UL << i);
750 sysctl_sched_features |= (1UL << i);
755 if (!sched_feat_names[i])
763 static int sched_feat_open(struct inode *inode, struct file *filp)
765 return single_open(filp, sched_feat_show, NULL);
768 static const struct file_operations sched_feat_fops = {
769 .open = sched_feat_open,
770 .write = sched_feat_write,
773 .release = single_release,
776 static __init int sched_init_debug(void)
778 debugfs_create_file("sched_features", 0644, NULL, NULL,
783 late_initcall(sched_init_debug);
787 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
790 * Number of tasks to iterate in a single balance run.
791 * Limited because this is done with IRQs disabled.
793 const_debug unsigned int sysctl_sched_nr_migrate = 32;
796 * ratelimit for updating the group shares.
799 unsigned int sysctl_sched_shares_ratelimit = 250000;
800 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
803 * Inject some fuzzyness into changing the per-cpu group shares
804 * this avoids remote rq-locks at the expense of fairness.
807 unsigned int sysctl_sched_shares_thresh = 4;
810 * period over which we average the RT time consumption, measured
815 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
818 * period over which we measure -rt task cpu usage in us.
821 unsigned int sysctl_sched_rt_period = 1000000;
823 static __read_mostly int scheduler_running;
826 * part of the period that we allow rt tasks to run in us.
829 int sysctl_sched_rt_runtime = 950000;
831 static inline u64 global_rt_period(void)
833 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
836 static inline u64 global_rt_runtime(void)
838 if (sysctl_sched_rt_runtime < 0)
841 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
844 #ifndef prepare_arch_switch
845 # define prepare_arch_switch(next) do { } while (0)
847 #ifndef finish_arch_switch
848 # define finish_arch_switch(prev) do { } while (0)
851 static inline int task_current(struct rq *rq, struct task_struct *p)
853 return rq->curr == p;
856 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
857 static inline int task_running(struct rq *rq, struct task_struct *p)
859 return task_current(rq, p);
862 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
866 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
868 #ifdef CONFIG_DEBUG_SPINLOCK
869 /* this is a valid case when another task releases the spinlock */
870 rq->lock.owner = current;
873 * If we are tracking spinlock dependencies then we have to
874 * fix up the runqueue lock - which gets 'carried over' from
877 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
879 raw_spin_unlock_irq(&rq->lock);
882 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
883 static inline int task_running(struct rq *rq, struct task_struct *p)
888 return task_current(rq, p);
892 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
896 * We can optimise this out completely for !SMP, because the
897 * SMP rebalancing from interrupt is the only thing that cares
902 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
903 raw_spin_unlock_irq(&rq->lock);
905 raw_spin_unlock(&rq->lock);
909 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
913 * After ->oncpu is cleared, the task can be moved to a different CPU.
914 * We must ensure this doesn't happen until the switch is completely
920 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
924 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
927 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
930 static inline int task_is_waking(struct task_struct *p)
932 return unlikely(p->state == TASK_WAKING);
936 * __task_rq_lock - lock the runqueue a given task resides on.
937 * Must be called interrupts disabled.
939 static inline struct rq *__task_rq_lock(struct task_struct *p)
946 raw_spin_lock(&rq->lock);
947 if (likely(rq == task_rq(p)))
949 raw_spin_unlock(&rq->lock);
954 * task_rq_lock - lock the runqueue a given task resides on and disable
955 * interrupts. Note the ordering: we can safely lookup the task_rq without
956 * explicitly disabling preemption.
958 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
964 local_irq_save(*flags);
966 raw_spin_lock(&rq->lock);
967 if (likely(rq == task_rq(p)))
969 raw_spin_unlock_irqrestore(&rq->lock, *flags);
973 static void __task_rq_unlock(struct rq *rq)
976 raw_spin_unlock(&rq->lock);
979 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
982 raw_spin_unlock_irqrestore(&rq->lock, *flags);
986 * this_rq_lock - lock this runqueue and disable interrupts.
988 static struct rq *this_rq_lock(void)
995 raw_spin_lock(&rq->lock);
1000 #ifdef CONFIG_SCHED_HRTICK
1002 * Use HR-timers to deliver accurate preemption points.
1004 * Its all a bit involved since we cannot program an hrt while holding the
1005 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1008 * When we get rescheduled we reprogram the hrtick_timer outside of the
1014 * - enabled by features
1015 * - hrtimer is actually high res
1017 static inline int hrtick_enabled(struct rq *rq)
1019 if (!sched_feat(HRTICK))
1021 if (!cpu_active(cpu_of(rq)))
1023 return hrtimer_is_hres_active(&rq->hrtick_timer);
1026 static void hrtick_clear(struct rq *rq)
1028 if (hrtimer_active(&rq->hrtick_timer))
1029 hrtimer_cancel(&rq->hrtick_timer);
1033 * High-resolution timer tick.
1034 * Runs from hardirq context with interrupts disabled.
1036 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1038 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1040 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1042 raw_spin_lock(&rq->lock);
1043 update_rq_clock(rq);
1044 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1045 raw_spin_unlock(&rq->lock);
1047 return HRTIMER_NORESTART;
1052 * called from hardirq (IPI) context
1054 static void __hrtick_start(void *arg)
1056 struct rq *rq = arg;
1058 raw_spin_lock(&rq->lock);
1059 hrtimer_restart(&rq->hrtick_timer);
1060 rq->hrtick_csd_pending = 0;
1061 raw_spin_unlock(&rq->lock);
1065 * Called to set the hrtick timer state.
1067 * called with rq->lock held and irqs disabled
1069 static void hrtick_start(struct rq *rq, u64 delay)
1071 struct hrtimer *timer = &rq->hrtick_timer;
1072 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1074 hrtimer_set_expires(timer, time);
1076 if (rq == this_rq()) {
1077 hrtimer_restart(timer);
1078 } else if (!rq->hrtick_csd_pending) {
1079 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1080 rq->hrtick_csd_pending = 1;
1085 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1087 int cpu = (int)(long)hcpu;
1090 case CPU_UP_CANCELED:
1091 case CPU_UP_CANCELED_FROZEN:
1092 case CPU_DOWN_PREPARE:
1093 case CPU_DOWN_PREPARE_FROZEN:
1095 case CPU_DEAD_FROZEN:
1096 hrtick_clear(cpu_rq(cpu));
1103 static __init void init_hrtick(void)
1105 hotcpu_notifier(hotplug_hrtick, 0);
1109 * Called to set the hrtick timer state.
1111 * called with rq->lock held and irqs disabled
1113 static void hrtick_start(struct rq *rq, u64 delay)
1115 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1116 HRTIMER_MODE_REL_PINNED, 0);
1119 static inline void init_hrtick(void)
1122 #endif /* CONFIG_SMP */
1124 static void init_rq_hrtick(struct rq *rq)
1127 rq->hrtick_csd_pending = 0;
1129 rq->hrtick_csd.flags = 0;
1130 rq->hrtick_csd.func = __hrtick_start;
1131 rq->hrtick_csd.info = rq;
1134 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1135 rq->hrtick_timer.function = hrtick;
1137 #else /* CONFIG_SCHED_HRTICK */
1138 static inline void hrtick_clear(struct rq *rq)
1142 static inline void init_rq_hrtick(struct rq *rq)
1146 static inline void init_hrtick(void)
1149 #endif /* CONFIG_SCHED_HRTICK */
1152 * resched_task - mark a task 'to be rescheduled now'.
1154 * On UP this means the setting of the need_resched flag, on SMP it
1155 * might also involve a cross-CPU call to trigger the scheduler on
1160 #ifndef tsk_is_polling
1161 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1164 static void resched_task(struct task_struct *p)
1168 assert_raw_spin_locked(&task_rq(p)->lock);
1170 if (test_tsk_need_resched(p))
1173 set_tsk_need_resched(p);
1176 if (cpu == smp_processor_id())
1179 /* NEED_RESCHED must be visible before we test polling */
1181 if (!tsk_is_polling(p))
1182 smp_send_reschedule(cpu);
1185 static void resched_cpu(int cpu)
1187 struct rq *rq = cpu_rq(cpu);
1188 unsigned long flags;
1190 if (!raw_spin_trylock_irqsave(&rq->lock, flags))
1192 resched_task(cpu_curr(cpu));
1193 raw_spin_unlock_irqrestore(&rq->lock, flags);
1198 * When add_timer_on() enqueues a timer into the timer wheel of an
1199 * idle CPU then this timer might expire before the next timer event
1200 * which is scheduled to wake up that CPU. In case of a completely
1201 * idle system the next event might even be infinite time into the
1202 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1203 * leaves the inner idle loop so the newly added timer is taken into
1204 * account when the CPU goes back to idle and evaluates the timer
1205 * wheel for the next timer event.
1207 void wake_up_idle_cpu(int cpu)
1209 struct rq *rq = cpu_rq(cpu);
1211 if (cpu == smp_processor_id())
1215 * This is safe, as this function is called with the timer
1216 * wheel base lock of (cpu) held. When the CPU is on the way
1217 * to idle and has not yet set rq->curr to idle then it will
1218 * be serialized on the timer wheel base lock and take the new
1219 * timer into account automatically.
1221 if (rq->curr != rq->idle)
1225 * We can set TIF_RESCHED on the idle task of the other CPU
1226 * lockless. The worst case is that the other CPU runs the
1227 * idle task through an additional NOOP schedule()
1229 set_tsk_need_resched(rq->idle);
1231 /* NEED_RESCHED must be visible before we test polling */
1233 if (!tsk_is_polling(rq->idle))
1234 smp_send_reschedule(cpu);
1237 int nohz_ratelimit(int cpu)
1239 struct rq *rq = cpu_rq(cpu);
1240 u64 diff = rq->clock - rq->nohz_stamp;
1242 rq->nohz_stamp = rq->clock;
1244 return diff < (NSEC_PER_SEC / HZ) >> 1;
1247 #endif /* CONFIG_NO_HZ */
1249 static u64 sched_avg_period(void)
1251 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1254 static void sched_avg_update(struct rq *rq)
1256 s64 period = sched_avg_period();
1258 while ((s64)(rq->clock - rq->age_stamp) > period) {
1259 rq->age_stamp += period;
1264 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1266 rq->rt_avg += rt_delta;
1267 sched_avg_update(rq);
1270 #else /* !CONFIG_SMP */
1271 static void resched_task(struct task_struct *p)
1273 assert_raw_spin_locked(&task_rq(p)->lock);
1274 set_tsk_need_resched(p);
1277 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1280 #endif /* CONFIG_SMP */
1282 #if BITS_PER_LONG == 32
1283 # define WMULT_CONST (~0UL)
1285 # define WMULT_CONST (1UL << 32)
1288 #define WMULT_SHIFT 32
1291 * Shift right and round:
1293 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1296 * delta *= weight / lw
1298 static unsigned long
1299 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1300 struct load_weight *lw)
1304 if (!lw->inv_weight) {
1305 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1308 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1312 tmp = (u64)delta_exec * weight;
1314 * Check whether we'd overflow the 64-bit multiplication:
1316 if (unlikely(tmp > WMULT_CONST))
1317 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1320 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1322 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1325 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1331 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1338 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1339 * of tasks with abnormal "nice" values across CPUs the contribution that
1340 * each task makes to its run queue's load is weighted according to its
1341 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1342 * scaled version of the new time slice allocation that they receive on time
1346 #define WEIGHT_IDLEPRIO 3
1347 #define WMULT_IDLEPRIO 1431655765
1350 * Nice levels are multiplicative, with a gentle 10% change for every
1351 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1352 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1353 * that remained on nice 0.
1355 * The "10% effect" is relative and cumulative: from _any_ nice level,
1356 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1357 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1358 * If a task goes up by ~10% and another task goes down by ~10% then
1359 * the relative distance between them is ~25%.)
1361 static const int prio_to_weight[40] = {
1362 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1363 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1364 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1365 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1366 /* 0 */ 1024, 820, 655, 526, 423,
1367 /* 5 */ 335, 272, 215, 172, 137,
1368 /* 10 */ 110, 87, 70, 56, 45,
1369 /* 15 */ 36, 29, 23, 18, 15,
1373 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1375 * In cases where the weight does not change often, we can use the
1376 * precalculated inverse to speed up arithmetics by turning divisions
1377 * into multiplications:
1379 static const u32 prio_to_wmult[40] = {
1380 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1381 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1382 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1383 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1384 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1385 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1386 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1387 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1390 /* Time spent by the tasks of the cpu accounting group executing in ... */
1391 enum cpuacct_stat_index {
1392 CPUACCT_STAT_USER, /* ... user mode */
1393 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1395 CPUACCT_STAT_NSTATS,
1398 #ifdef CONFIG_CGROUP_CPUACCT
1399 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1400 static void cpuacct_update_stats(struct task_struct *tsk,
1401 enum cpuacct_stat_index idx, cputime_t val);
1403 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1404 static inline void cpuacct_update_stats(struct task_struct *tsk,
1405 enum cpuacct_stat_index idx, cputime_t val) {}
1408 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1410 update_load_add(&rq->load, load);
1413 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1415 update_load_sub(&rq->load, load);
1418 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1419 typedef int (*tg_visitor)(struct task_group *, void *);
1422 * Iterate the full tree, calling @down when first entering a node and @up when
1423 * leaving it for the final time.
1425 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1427 struct task_group *parent, *child;
1431 parent = &root_task_group;
1433 ret = (*down)(parent, data);
1436 list_for_each_entry_rcu(child, &parent->children, siblings) {
1443 ret = (*up)(parent, data);
1448 parent = parent->parent;
1457 static int tg_nop(struct task_group *tg, void *data)
1464 /* Used instead of source_load when we know the type == 0 */
1465 static unsigned long weighted_cpuload(const int cpu)
1467 return cpu_rq(cpu)->load.weight;
1471 * Return a low guess at the load of a migration-source cpu weighted
1472 * according to the scheduling class and "nice" value.
1474 * We want to under-estimate the load of migration sources, to
1475 * balance conservatively.
1477 static unsigned long source_load(int cpu, int type)
1479 struct rq *rq = cpu_rq(cpu);
1480 unsigned long total = weighted_cpuload(cpu);
1482 if (type == 0 || !sched_feat(LB_BIAS))
1485 return min(rq->cpu_load[type-1], total);
1489 * Return a high guess at the load of a migration-target cpu weighted
1490 * according to the scheduling class and "nice" value.
1492 static unsigned long target_load(int cpu, int type)
1494 struct rq *rq = cpu_rq(cpu);
1495 unsigned long total = weighted_cpuload(cpu);
1497 if (type == 0 || !sched_feat(LB_BIAS))
1500 return max(rq->cpu_load[type-1], total);
1503 static unsigned long power_of(int cpu)
1505 return cpu_rq(cpu)->cpu_power;
1508 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1510 static unsigned long cpu_avg_load_per_task(int cpu)
1512 struct rq *rq = cpu_rq(cpu);
1513 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1516 rq->avg_load_per_task = rq->load.weight / nr_running;
1518 rq->avg_load_per_task = 0;
1520 return rq->avg_load_per_task;
1523 #ifdef CONFIG_FAIR_GROUP_SCHED
1525 static __read_mostly unsigned long __percpu *update_shares_data;
1527 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1530 * Calculate and set the cpu's group shares.
1532 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1533 unsigned long sd_shares,
1534 unsigned long sd_rq_weight,
1535 unsigned long *usd_rq_weight)
1537 unsigned long shares, rq_weight;
1540 rq_weight = usd_rq_weight[cpu];
1543 rq_weight = NICE_0_LOAD;
1547 * \Sum_j shares_j * rq_weight_i
1548 * shares_i = -----------------------------
1549 * \Sum_j rq_weight_j
1551 shares = (sd_shares * rq_weight) / sd_rq_weight;
1552 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1554 if (abs(shares - tg->se[cpu]->load.weight) >
1555 sysctl_sched_shares_thresh) {
1556 struct rq *rq = cpu_rq(cpu);
1557 unsigned long flags;
1559 raw_spin_lock_irqsave(&rq->lock, flags);
1560 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1561 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1562 __set_se_shares(tg->se[cpu], shares);
1563 raw_spin_unlock_irqrestore(&rq->lock, flags);
1568 * Re-compute the task group their per cpu shares over the given domain.
1569 * This needs to be done in a bottom-up fashion because the rq weight of a
1570 * parent group depends on the shares of its child groups.
1572 static int tg_shares_up(struct task_group *tg, void *data)
1574 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1575 unsigned long *usd_rq_weight;
1576 struct sched_domain *sd = data;
1577 unsigned long flags;
1583 local_irq_save(flags);
1584 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1586 for_each_cpu(i, sched_domain_span(sd)) {
1587 weight = tg->cfs_rq[i]->load.weight;
1588 usd_rq_weight[i] = weight;
1590 rq_weight += weight;
1592 * If there are currently no tasks on the cpu pretend there
1593 * is one of average load so that when a new task gets to
1594 * run here it will not get delayed by group starvation.
1597 weight = NICE_0_LOAD;
1599 sum_weight += weight;
1600 shares += tg->cfs_rq[i]->shares;
1604 rq_weight = sum_weight;
1606 if ((!shares && rq_weight) || shares > tg->shares)
1607 shares = tg->shares;
1609 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1610 shares = tg->shares;
1612 for_each_cpu(i, sched_domain_span(sd))
1613 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1615 local_irq_restore(flags);
1621 * Compute the cpu's hierarchical load factor for each task group.
1622 * This needs to be done in a top-down fashion because the load of a child
1623 * group is a fraction of its parents load.
1625 static int tg_load_down(struct task_group *tg, void *data)
1628 long cpu = (long)data;
1631 load = cpu_rq(cpu)->load.weight;
1633 load = tg->parent->cfs_rq[cpu]->h_load;
1634 load *= tg->cfs_rq[cpu]->shares;
1635 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1638 tg->cfs_rq[cpu]->h_load = load;
1643 static void update_shares(struct sched_domain *sd)
1648 if (root_task_group_empty())
1651 now = local_clock();
1652 elapsed = now - sd->last_update;
1654 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1655 sd->last_update = now;
1656 walk_tg_tree(tg_nop, tg_shares_up, sd);
1660 static void update_h_load(long cpu)
1662 if (root_task_group_empty())
1665 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1670 static inline void update_shares(struct sched_domain *sd)
1676 #ifdef CONFIG_PREEMPT
1678 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1681 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1682 * way at the expense of forcing extra atomic operations in all
1683 * invocations. This assures that the double_lock is acquired using the
1684 * same underlying policy as the spinlock_t on this architecture, which
1685 * reduces latency compared to the unfair variant below. However, it
1686 * also adds more overhead and therefore may reduce throughput.
1688 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1689 __releases(this_rq->lock)
1690 __acquires(busiest->lock)
1691 __acquires(this_rq->lock)
1693 raw_spin_unlock(&this_rq->lock);
1694 double_rq_lock(this_rq, busiest);
1701 * Unfair double_lock_balance: Optimizes throughput at the expense of
1702 * latency by eliminating extra atomic operations when the locks are
1703 * already in proper order on entry. This favors lower cpu-ids and will
1704 * grant the double lock to lower cpus over higher ids under contention,
1705 * regardless of entry order into the function.
1707 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1708 __releases(this_rq->lock)
1709 __acquires(busiest->lock)
1710 __acquires(this_rq->lock)
1714 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1715 if (busiest < this_rq) {
1716 raw_spin_unlock(&this_rq->lock);
1717 raw_spin_lock(&busiest->lock);
1718 raw_spin_lock_nested(&this_rq->lock,
1719 SINGLE_DEPTH_NESTING);
1722 raw_spin_lock_nested(&busiest->lock,
1723 SINGLE_DEPTH_NESTING);
1728 #endif /* CONFIG_PREEMPT */
1731 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1733 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1735 if (unlikely(!irqs_disabled())) {
1736 /* printk() doesn't work good under rq->lock */
1737 raw_spin_unlock(&this_rq->lock);
1741 return _double_lock_balance(this_rq, busiest);
1744 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1745 __releases(busiest->lock)
1747 raw_spin_unlock(&busiest->lock);
1748 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1752 * double_rq_lock - safely lock two runqueues
1754 * Note this does not disable interrupts like task_rq_lock,
1755 * you need to do so manually before calling.
1757 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1758 __acquires(rq1->lock)
1759 __acquires(rq2->lock)
1761 BUG_ON(!irqs_disabled());
1763 raw_spin_lock(&rq1->lock);
1764 __acquire(rq2->lock); /* Fake it out ;) */
1767 raw_spin_lock(&rq1->lock);
1768 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1770 raw_spin_lock(&rq2->lock);
1771 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1777 * double_rq_unlock - safely unlock two runqueues
1779 * Note this does not restore interrupts like task_rq_unlock,
1780 * you need to do so manually after calling.
1782 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1783 __releases(rq1->lock)
1784 __releases(rq2->lock)
1786 raw_spin_unlock(&rq1->lock);
1788 raw_spin_unlock(&rq2->lock);
1790 __release(rq2->lock);
1795 #ifdef CONFIG_FAIR_GROUP_SCHED
1796 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1799 cfs_rq->shares = shares;
1804 static void calc_load_account_idle(struct rq *this_rq);
1805 static void update_sysctl(void);
1806 static int get_update_sysctl_factor(void);
1807 static void update_cpu_load(struct rq *this_rq);
1809 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1811 set_task_rq(p, cpu);
1814 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1815 * successfuly executed on another CPU. We must ensure that updates of
1816 * per-task data have been completed by this moment.
1819 task_thread_info(p)->cpu = cpu;
1823 static const struct sched_class rt_sched_class;
1825 #define sched_class_highest (&rt_sched_class)
1826 #define for_each_class(class) \
1827 for (class = sched_class_highest; class; class = class->next)
1829 #include "sched_stats.h"
1831 static void inc_nr_running(struct rq *rq)
1836 static void dec_nr_running(struct rq *rq)
1841 static void set_load_weight(struct task_struct *p)
1843 if (task_has_rt_policy(p)) {
1844 p->se.load.weight = 0;
1845 p->se.load.inv_weight = WMULT_CONST;
1850 * SCHED_IDLE tasks get minimal weight:
1852 if (p->policy == SCHED_IDLE) {
1853 p->se.load.weight = WEIGHT_IDLEPRIO;
1854 p->se.load.inv_weight = WMULT_IDLEPRIO;
1858 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1859 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1862 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1864 update_rq_clock(rq);
1865 sched_info_queued(p);
1866 p->sched_class->enqueue_task(rq, p, flags);
1870 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1872 update_rq_clock(rq);
1873 sched_info_dequeued(p);
1874 p->sched_class->dequeue_task(rq, p, flags);
1879 * activate_task - move a task to the runqueue.
1881 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1883 if (task_contributes_to_load(p))
1884 rq->nr_uninterruptible--;
1886 enqueue_task(rq, p, flags);
1891 * deactivate_task - remove a task from the runqueue.
1893 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1895 if (task_contributes_to_load(p))
1896 rq->nr_uninterruptible++;
1898 dequeue_task(rq, p, flags);
1902 #include "sched_idletask.c"
1903 #include "sched_fair.c"
1904 #include "sched_rt.c"
1905 #ifdef CONFIG_SCHED_DEBUG
1906 # include "sched_debug.c"
1910 * __normal_prio - return the priority that is based on the static prio
1912 static inline int __normal_prio(struct task_struct *p)
1914 return p->static_prio;
1918 * Calculate the expected normal priority: i.e. priority
1919 * without taking RT-inheritance into account. Might be
1920 * boosted by interactivity modifiers. Changes upon fork,
1921 * setprio syscalls, and whenever the interactivity
1922 * estimator recalculates.
1924 static inline int normal_prio(struct task_struct *p)
1928 if (task_has_rt_policy(p))
1929 prio = MAX_RT_PRIO-1 - p->rt_priority;
1931 prio = __normal_prio(p);
1936 * Calculate the current priority, i.e. the priority
1937 * taken into account by the scheduler. This value might
1938 * be boosted by RT tasks, or might be boosted by
1939 * interactivity modifiers. Will be RT if the task got
1940 * RT-boosted. If not then it returns p->normal_prio.
1942 static int effective_prio(struct task_struct *p)
1944 p->normal_prio = normal_prio(p);
1946 * If we are RT tasks or we were boosted to RT priority,
1947 * keep the priority unchanged. Otherwise, update priority
1948 * to the normal priority:
1950 if (!rt_prio(p->prio))
1951 return p->normal_prio;
1956 * task_curr - is this task currently executing on a CPU?
1957 * @p: the task in question.
1959 inline int task_curr(const struct task_struct *p)
1961 return cpu_curr(task_cpu(p)) == p;
1964 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1965 const struct sched_class *prev_class,
1966 int oldprio, int running)
1968 if (prev_class != p->sched_class) {
1969 if (prev_class->switched_from)
1970 prev_class->switched_from(rq, p, running);
1971 p->sched_class->switched_to(rq, p, running);
1973 p->sched_class->prio_changed(rq, p, oldprio, running);
1978 * Is this task likely cache-hot:
1981 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1985 if (p->sched_class != &fair_sched_class)
1989 * Buddy candidates are cache hot:
1991 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
1992 (&p->se == cfs_rq_of(&p->se)->next ||
1993 &p->se == cfs_rq_of(&p->se)->last))
1996 if (sysctl_sched_migration_cost == -1)
1998 if (sysctl_sched_migration_cost == 0)
2001 delta = now - p->se.exec_start;
2003 return delta < (s64)sysctl_sched_migration_cost;
2006 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2008 #ifdef CONFIG_SCHED_DEBUG
2010 * We should never call set_task_cpu() on a blocked task,
2011 * ttwu() will sort out the placement.
2013 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2014 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2017 trace_sched_migrate_task(p, new_cpu);
2019 if (task_cpu(p) != new_cpu) {
2020 p->se.nr_migrations++;
2021 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2024 __set_task_cpu(p, new_cpu);
2027 struct migration_arg {
2028 struct task_struct *task;
2032 static int migration_cpu_stop(void *data);
2035 * The task's runqueue lock must be held.
2036 * Returns true if you have to wait for migration thread.
2038 static bool migrate_task(struct task_struct *p, int dest_cpu)
2040 struct rq *rq = task_rq(p);
2043 * If the task is not on a runqueue (and not running), then
2044 * the next wake-up will properly place the task.
2046 return p->se.on_rq || task_running(rq, p);
2050 * wait_task_inactive - wait for a thread to unschedule.
2052 * If @match_state is nonzero, it's the @p->state value just checked and
2053 * not expected to change. If it changes, i.e. @p might have woken up,
2054 * then return zero. When we succeed in waiting for @p to be off its CPU,
2055 * we return a positive number (its total switch count). If a second call
2056 * a short while later returns the same number, the caller can be sure that
2057 * @p has remained unscheduled the whole time.
2059 * The caller must ensure that the task *will* unschedule sometime soon,
2060 * else this function might spin for a *long* time. This function can't
2061 * be called with interrupts off, or it may introduce deadlock with
2062 * smp_call_function() if an IPI is sent by the same process we are
2063 * waiting to become inactive.
2065 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2067 unsigned long flags;
2074 * We do the initial early heuristics without holding
2075 * any task-queue locks at all. We'll only try to get
2076 * the runqueue lock when things look like they will
2082 * If the task is actively running on another CPU
2083 * still, just relax and busy-wait without holding
2086 * NOTE! Since we don't hold any locks, it's not
2087 * even sure that "rq" stays as the right runqueue!
2088 * But we don't care, since "task_running()" will
2089 * return false if the runqueue has changed and p
2090 * is actually now running somewhere else!
2092 while (task_running(rq, p)) {
2093 if (match_state && unlikely(p->state != match_state))
2099 * Ok, time to look more closely! We need the rq
2100 * lock now, to be *sure*. If we're wrong, we'll
2101 * just go back and repeat.
2103 rq = task_rq_lock(p, &flags);
2104 trace_sched_wait_task(p);
2105 running = task_running(rq, p);
2106 on_rq = p->se.on_rq;
2108 if (!match_state || p->state == match_state)
2109 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2110 task_rq_unlock(rq, &flags);
2113 * If it changed from the expected state, bail out now.
2115 if (unlikely(!ncsw))
2119 * Was it really running after all now that we
2120 * checked with the proper locks actually held?
2122 * Oops. Go back and try again..
2124 if (unlikely(running)) {
2130 * It's not enough that it's not actively running,
2131 * it must be off the runqueue _entirely_, and not
2134 * So if it was still runnable (but just not actively
2135 * running right now), it's preempted, and we should
2136 * yield - it could be a while.
2138 if (unlikely(on_rq)) {
2139 schedule_timeout_uninterruptible(1);
2144 * Ahh, all good. It wasn't running, and it wasn't
2145 * runnable, which means that it will never become
2146 * running in the future either. We're all done!
2155 * kick_process - kick a running thread to enter/exit the kernel
2156 * @p: the to-be-kicked thread
2158 * Cause a process which is running on another CPU to enter
2159 * kernel-mode, without any delay. (to get signals handled.)
2161 * NOTE: this function doesnt have to take the runqueue lock,
2162 * because all it wants to ensure is that the remote task enters
2163 * the kernel. If the IPI races and the task has been migrated
2164 * to another CPU then no harm is done and the purpose has been
2167 void kick_process(struct task_struct *p)
2173 if ((cpu != smp_processor_id()) && task_curr(p))
2174 smp_send_reschedule(cpu);
2177 EXPORT_SYMBOL_GPL(kick_process);
2178 #endif /* CONFIG_SMP */
2181 * task_oncpu_function_call - call a function on the cpu on which a task runs
2182 * @p: the task to evaluate
2183 * @func: the function to be called
2184 * @info: the function call argument
2186 * Calls the function @func when the task is currently running. This might
2187 * be on the current CPU, which just calls the function directly
2189 void task_oncpu_function_call(struct task_struct *p,
2190 void (*func) (void *info), void *info)
2197 smp_call_function_single(cpu, func, info, 1);
2203 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2205 static int select_fallback_rq(int cpu, struct task_struct *p)
2208 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2210 /* Look for allowed, online CPU in same node. */
2211 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2212 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2215 /* Any allowed, online CPU? */
2216 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2217 if (dest_cpu < nr_cpu_ids)
2220 /* No more Mr. Nice Guy. */
2221 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2222 dest_cpu = cpuset_cpus_allowed_fallback(p);
2224 * Don't tell them about moving exiting tasks or
2225 * kernel threads (both mm NULL), since they never
2228 if (p->mm && printk_ratelimit()) {
2229 printk(KERN_INFO "process %d (%s) no "
2230 "longer affine to cpu%d\n",
2231 task_pid_nr(p), p->comm, cpu);
2239 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2242 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2244 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2247 * In order not to call set_task_cpu() on a blocking task we need
2248 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2251 * Since this is common to all placement strategies, this lives here.
2253 * [ this allows ->select_task() to simply return task_cpu(p) and
2254 * not worry about this generic constraint ]
2256 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2258 cpu = select_fallback_rq(task_cpu(p), p);
2263 static void update_avg(u64 *avg, u64 sample)
2265 s64 diff = sample - *avg;
2270 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2271 bool is_sync, bool is_migrate, bool is_local,
2272 unsigned long en_flags)
2274 schedstat_inc(p, se.statistics.nr_wakeups);
2276 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2278 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2280 schedstat_inc(p, se.statistics.nr_wakeups_local);
2282 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2284 activate_task(rq, p, en_flags);
2287 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2288 int wake_flags, bool success)
2290 trace_sched_wakeup(p, success);
2291 check_preempt_curr(rq, p, wake_flags);
2293 p->state = TASK_RUNNING;
2295 if (p->sched_class->task_woken)
2296 p->sched_class->task_woken(rq, p);
2298 if (unlikely(rq->idle_stamp)) {
2299 u64 delta = rq->clock - rq->idle_stamp;
2300 u64 max = 2*sysctl_sched_migration_cost;
2305 update_avg(&rq->avg_idle, delta);
2309 /* if a worker is waking up, notify workqueue */
2310 if ((p->flags & PF_WQ_WORKER) && success)
2311 wq_worker_waking_up(p, cpu_of(rq));
2315 * try_to_wake_up - wake up a thread
2316 * @p: the thread to be awakened
2317 * @state: the mask of task states that can be woken
2318 * @wake_flags: wake modifier flags (WF_*)
2320 * Put it on the run-queue if it's not already there. The "current"
2321 * thread is always on the run-queue (except when the actual
2322 * re-schedule is in progress), and as such you're allowed to do
2323 * the simpler "current->state = TASK_RUNNING" to mark yourself
2324 * runnable without the overhead of this.
2326 * Returns %true if @p was woken up, %false if it was already running
2327 * or @state didn't match @p's state.
2329 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2332 int cpu, orig_cpu, this_cpu, success = 0;
2333 unsigned long flags;
2334 unsigned long en_flags = ENQUEUE_WAKEUP;
2337 this_cpu = get_cpu();
2340 rq = task_rq_lock(p, &flags);
2341 if (!(p->state & state))
2351 if (unlikely(task_running(rq, p)))
2355 * In order to handle concurrent wakeups and release the rq->lock
2356 * we put the task in TASK_WAKING state.
2358 * First fix up the nr_uninterruptible count:
2360 if (task_contributes_to_load(p)) {
2361 if (likely(cpu_online(orig_cpu)))
2362 rq->nr_uninterruptible--;
2364 this_rq()->nr_uninterruptible--;
2366 p->state = TASK_WAKING;
2368 if (p->sched_class->task_waking) {
2369 p->sched_class->task_waking(rq, p);
2370 en_flags |= ENQUEUE_WAKING;
2373 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2374 if (cpu != orig_cpu)
2375 set_task_cpu(p, cpu);
2376 __task_rq_unlock(rq);
2379 raw_spin_lock(&rq->lock);
2382 * We migrated the task without holding either rq->lock, however
2383 * since the task is not on the task list itself, nobody else
2384 * will try and migrate the task, hence the rq should match the
2385 * cpu we just moved it to.
2387 WARN_ON(task_cpu(p) != cpu);
2388 WARN_ON(p->state != TASK_WAKING);
2390 #ifdef CONFIG_SCHEDSTATS
2391 schedstat_inc(rq, ttwu_count);
2392 if (cpu == this_cpu)
2393 schedstat_inc(rq, ttwu_local);
2395 struct sched_domain *sd;
2396 for_each_domain(this_cpu, sd) {
2397 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2398 schedstat_inc(sd, ttwu_wake_remote);
2403 #endif /* CONFIG_SCHEDSTATS */
2406 #endif /* CONFIG_SMP */
2407 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2408 cpu == this_cpu, en_flags);
2411 ttwu_post_activation(p, rq, wake_flags, success);
2413 task_rq_unlock(rq, &flags);
2420 * try_to_wake_up_local - try to wake up a local task with rq lock held
2421 * @p: the thread to be awakened
2423 * Put @p on the run-queue if it's not alredy there. The caller must
2424 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2425 * the current task. this_rq() stays locked over invocation.
2427 static void try_to_wake_up_local(struct task_struct *p)
2429 struct rq *rq = task_rq(p);
2430 bool success = false;
2432 BUG_ON(rq != this_rq());
2433 BUG_ON(p == current);
2434 lockdep_assert_held(&rq->lock);
2436 if (!(p->state & TASK_NORMAL))
2440 if (likely(!task_running(rq, p))) {
2441 schedstat_inc(rq, ttwu_count);
2442 schedstat_inc(rq, ttwu_local);
2444 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2447 ttwu_post_activation(p, rq, 0, success);
2451 * wake_up_process - Wake up a specific process
2452 * @p: The process to be woken up.
2454 * Attempt to wake up the nominated process and move it to the set of runnable
2455 * processes. Returns 1 if the process was woken up, 0 if it was already
2458 * It may be assumed that this function implies a write memory barrier before
2459 * changing the task state if and only if any tasks are woken up.
2461 int wake_up_process(struct task_struct *p)
2463 return try_to_wake_up(p, TASK_ALL, 0);
2465 EXPORT_SYMBOL(wake_up_process);
2467 int wake_up_state(struct task_struct *p, unsigned int state)
2469 return try_to_wake_up(p, state, 0);
2473 * Perform scheduler related setup for a newly forked process p.
2474 * p is forked by current.
2476 * __sched_fork() is basic setup used by init_idle() too:
2478 static void __sched_fork(struct task_struct *p)
2480 p->se.exec_start = 0;
2481 p->se.sum_exec_runtime = 0;
2482 p->se.prev_sum_exec_runtime = 0;
2483 p->se.nr_migrations = 0;
2485 #ifdef CONFIG_SCHEDSTATS
2486 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2489 INIT_LIST_HEAD(&p->rt.run_list);
2491 INIT_LIST_HEAD(&p->se.group_node);
2493 #ifdef CONFIG_PREEMPT_NOTIFIERS
2494 INIT_HLIST_HEAD(&p->preempt_notifiers);
2499 * fork()/clone()-time setup:
2501 void sched_fork(struct task_struct *p, int clone_flags)
2503 int cpu = get_cpu();
2507 * We mark the process as running here. This guarantees that
2508 * nobody will actually run it, and a signal or other external
2509 * event cannot wake it up and insert it on the runqueue either.
2511 p->state = TASK_RUNNING;
2514 * Revert to default priority/policy on fork if requested.
2516 if (unlikely(p->sched_reset_on_fork)) {
2517 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2518 p->policy = SCHED_NORMAL;
2519 p->normal_prio = p->static_prio;
2522 if (PRIO_TO_NICE(p->static_prio) < 0) {
2523 p->static_prio = NICE_TO_PRIO(0);
2524 p->normal_prio = p->static_prio;
2529 * We don't need the reset flag anymore after the fork. It has
2530 * fulfilled its duty:
2532 p->sched_reset_on_fork = 0;
2536 * Make sure we do not leak PI boosting priority to the child.
2538 p->prio = current->normal_prio;
2540 if (!rt_prio(p->prio))
2541 p->sched_class = &fair_sched_class;
2543 if (p->sched_class->task_fork)
2544 p->sched_class->task_fork(p);
2546 set_task_cpu(p, cpu);
2548 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2549 if (likely(sched_info_on()))
2550 memset(&p->sched_info, 0, sizeof(p->sched_info));
2552 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2555 #ifdef CONFIG_PREEMPT
2556 /* Want to start with kernel preemption disabled. */
2557 task_thread_info(p)->preempt_count = 1;
2559 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2565 * wake_up_new_task - wake up a newly created task for the first time.
2567 * This function will do some initial scheduler statistics housekeeping
2568 * that must be done for every newly created context, then puts the task
2569 * on the runqueue and wakes it.
2571 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2573 unsigned long flags;
2575 int cpu __maybe_unused = get_cpu();
2578 rq = task_rq_lock(p, &flags);
2579 p->state = TASK_WAKING;
2582 * Fork balancing, do it here and not earlier because:
2583 * - cpus_allowed can change in the fork path
2584 * - any previously selected cpu might disappear through hotplug
2586 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2587 * without people poking at ->cpus_allowed.
2589 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2590 set_task_cpu(p, cpu);
2592 p->state = TASK_RUNNING;
2593 task_rq_unlock(rq, &flags);
2596 rq = task_rq_lock(p, &flags);
2597 activate_task(rq, p, 0);
2598 trace_sched_wakeup_new(p, 1);
2599 check_preempt_curr(rq, p, WF_FORK);
2601 if (p->sched_class->task_woken)
2602 p->sched_class->task_woken(rq, p);
2604 task_rq_unlock(rq, &flags);
2608 #ifdef CONFIG_PREEMPT_NOTIFIERS
2611 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2612 * @notifier: notifier struct to register
2614 void preempt_notifier_register(struct preempt_notifier *notifier)
2616 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2618 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2621 * preempt_notifier_unregister - no longer interested in preemption notifications
2622 * @notifier: notifier struct to unregister
2624 * This is safe to call from within a preemption notifier.
2626 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2628 hlist_del(¬ifier->link);
2630 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2632 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2634 struct preempt_notifier *notifier;
2635 struct hlist_node *node;
2637 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2638 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2642 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2643 struct task_struct *next)
2645 struct preempt_notifier *notifier;
2646 struct hlist_node *node;
2648 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2649 notifier->ops->sched_out(notifier, next);
2652 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2654 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2659 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2660 struct task_struct *next)
2664 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2667 * prepare_task_switch - prepare to switch tasks
2668 * @rq: the runqueue preparing to switch
2669 * @prev: the current task that is being switched out
2670 * @next: the task we are going to switch to.
2672 * This is called with the rq lock held and interrupts off. It must
2673 * be paired with a subsequent finish_task_switch after the context
2676 * prepare_task_switch sets up locking and calls architecture specific
2680 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2681 struct task_struct *next)
2683 fire_sched_out_preempt_notifiers(prev, next);
2684 prepare_lock_switch(rq, next);
2685 prepare_arch_switch(next);
2689 * finish_task_switch - clean up after a task-switch
2690 * @rq: runqueue associated with task-switch
2691 * @prev: the thread we just switched away from.
2693 * finish_task_switch must be called after the context switch, paired
2694 * with a prepare_task_switch call before the context switch.
2695 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2696 * and do any other architecture-specific cleanup actions.
2698 * Note that we may have delayed dropping an mm in context_switch(). If
2699 * so, we finish that here outside of the runqueue lock. (Doing it
2700 * with the lock held can cause deadlocks; see schedule() for
2703 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2704 __releases(rq->lock)
2706 struct mm_struct *mm = rq->prev_mm;
2712 * A task struct has one reference for the use as "current".
2713 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2714 * schedule one last time. The schedule call will never return, and
2715 * the scheduled task must drop that reference.
2716 * The test for TASK_DEAD must occur while the runqueue locks are
2717 * still held, otherwise prev could be scheduled on another cpu, die
2718 * there before we look at prev->state, and then the reference would
2720 * Manfred Spraul <manfred@colorfullife.com>
2722 prev_state = prev->state;
2723 finish_arch_switch(prev);
2724 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2725 local_irq_disable();
2726 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2727 perf_event_task_sched_in(current);
2728 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2730 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2731 finish_lock_switch(rq, prev);
2733 fire_sched_in_preempt_notifiers(current);
2736 if (unlikely(prev_state == TASK_DEAD)) {
2738 * Remove function-return probe instances associated with this
2739 * task and put them back on the free list.
2741 kprobe_flush_task(prev);
2742 put_task_struct(prev);
2748 /* assumes rq->lock is held */
2749 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2751 if (prev->sched_class->pre_schedule)
2752 prev->sched_class->pre_schedule(rq, prev);
2755 /* rq->lock is NOT held, but preemption is disabled */
2756 static inline void post_schedule(struct rq *rq)
2758 if (rq->post_schedule) {
2759 unsigned long flags;
2761 raw_spin_lock_irqsave(&rq->lock, flags);
2762 if (rq->curr->sched_class->post_schedule)
2763 rq->curr->sched_class->post_schedule(rq);
2764 raw_spin_unlock_irqrestore(&rq->lock, flags);
2766 rq->post_schedule = 0;
2772 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2776 static inline void post_schedule(struct rq *rq)
2783 * schedule_tail - first thing a freshly forked thread must call.
2784 * @prev: the thread we just switched away from.
2786 asmlinkage void schedule_tail(struct task_struct *prev)
2787 __releases(rq->lock)
2789 struct rq *rq = this_rq();
2791 finish_task_switch(rq, prev);
2794 * FIXME: do we need to worry about rq being invalidated by the
2799 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2800 /* In this case, finish_task_switch does not reenable preemption */
2803 if (current->set_child_tid)
2804 put_user(task_pid_vnr(current), current->set_child_tid);
2808 * context_switch - switch to the new MM and the new
2809 * thread's register state.
2812 context_switch(struct rq *rq, struct task_struct *prev,
2813 struct task_struct *next)
2815 struct mm_struct *mm, *oldmm;
2817 prepare_task_switch(rq, prev, next);
2818 trace_sched_switch(prev, next);
2820 oldmm = prev->active_mm;
2822 * For paravirt, this is coupled with an exit in switch_to to
2823 * combine the page table reload and the switch backend into
2826 arch_start_context_switch(prev);
2829 next->active_mm = oldmm;
2830 atomic_inc(&oldmm->mm_count);
2831 enter_lazy_tlb(oldmm, next);
2833 switch_mm(oldmm, mm, next);
2835 if (likely(!prev->mm)) {
2836 prev->active_mm = NULL;
2837 rq->prev_mm = oldmm;
2840 * Since the runqueue lock will be released by the next
2841 * task (which is an invalid locking op but in the case
2842 * of the scheduler it's an obvious special-case), so we
2843 * do an early lockdep release here:
2845 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2846 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2849 /* Here we just switch the register state and the stack. */
2850 switch_to(prev, next, prev);
2854 * this_rq must be evaluated again because prev may have moved
2855 * CPUs since it called schedule(), thus the 'rq' on its stack
2856 * frame will be invalid.
2858 finish_task_switch(this_rq(), prev);
2862 * nr_running, nr_uninterruptible and nr_context_switches:
2864 * externally visible scheduler statistics: current number of runnable
2865 * threads, current number of uninterruptible-sleeping threads, total
2866 * number of context switches performed since bootup.
2868 unsigned long nr_running(void)
2870 unsigned long i, sum = 0;
2872 for_each_online_cpu(i)
2873 sum += cpu_rq(i)->nr_running;
2878 unsigned long nr_uninterruptible(void)
2880 unsigned long i, sum = 0;
2882 for_each_possible_cpu(i)
2883 sum += cpu_rq(i)->nr_uninterruptible;
2886 * Since we read the counters lockless, it might be slightly
2887 * inaccurate. Do not allow it to go below zero though:
2889 if (unlikely((long)sum < 0))
2895 unsigned long long nr_context_switches(void)
2898 unsigned long long sum = 0;
2900 for_each_possible_cpu(i)
2901 sum += cpu_rq(i)->nr_switches;
2906 unsigned long nr_iowait(void)
2908 unsigned long i, sum = 0;
2910 for_each_possible_cpu(i)
2911 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2916 unsigned long nr_iowait_cpu(void)
2918 struct rq *this = this_rq();
2919 return atomic_read(&this->nr_iowait);
2922 unsigned long this_cpu_load(void)
2924 struct rq *this = this_rq();
2925 return this->cpu_load[0];
2929 /* Variables and functions for calc_load */
2930 static atomic_long_t calc_load_tasks;
2931 static unsigned long calc_load_update;
2932 unsigned long avenrun[3];
2933 EXPORT_SYMBOL(avenrun);
2935 static long calc_load_fold_active(struct rq *this_rq)
2937 long nr_active, delta = 0;
2939 nr_active = this_rq->nr_running;
2940 nr_active += (long) this_rq->nr_uninterruptible;
2942 if (nr_active != this_rq->calc_load_active) {
2943 delta = nr_active - this_rq->calc_load_active;
2944 this_rq->calc_load_active = nr_active;
2952 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2954 * When making the ILB scale, we should try to pull this in as well.
2956 static atomic_long_t calc_load_tasks_idle;
2958 static void calc_load_account_idle(struct rq *this_rq)
2962 delta = calc_load_fold_active(this_rq);
2964 atomic_long_add(delta, &calc_load_tasks_idle);
2967 static long calc_load_fold_idle(void)
2972 * Its got a race, we don't care...
2974 if (atomic_long_read(&calc_load_tasks_idle))
2975 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
2980 static void calc_load_account_idle(struct rq *this_rq)
2984 static inline long calc_load_fold_idle(void)
2991 * get_avenrun - get the load average array
2992 * @loads: pointer to dest load array
2993 * @offset: offset to add
2994 * @shift: shift count to shift the result left
2996 * These values are estimates at best, so no need for locking.
2998 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3000 loads[0] = (avenrun[0] + offset) << shift;
3001 loads[1] = (avenrun[1] + offset) << shift;
3002 loads[2] = (avenrun[2] + offset) << shift;
3005 static unsigned long
3006 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3009 load += active * (FIXED_1 - exp);
3010 return load >> FSHIFT;
3014 * calc_load - update the avenrun load estimates 10 ticks after the
3015 * CPUs have updated calc_load_tasks.
3017 void calc_global_load(void)
3019 unsigned long upd = calc_load_update + 10;
3022 if (time_before(jiffies, upd))
3025 active = atomic_long_read(&calc_load_tasks);
3026 active = active > 0 ? active * FIXED_1 : 0;
3028 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3029 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3030 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3032 calc_load_update += LOAD_FREQ;
3036 * Called from update_cpu_load() to periodically update this CPU's
3039 static void calc_load_account_active(struct rq *this_rq)
3043 if (time_before(jiffies, this_rq->calc_load_update))
3046 delta = calc_load_fold_active(this_rq);
3047 delta += calc_load_fold_idle();
3049 atomic_long_add(delta, &calc_load_tasks);
3051 this_rq->calc_load_update += LOAD_FREQ;
3055 * The exact cpuload at various idx values, calculated at every tick would be
3056 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3058 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3059 * on nth tick when cpu may be busy, then we have:
3060 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3061 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3063 * decay_load_missed() below does efficient calculation of
3064 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3065 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3067 * The calculation is approximated on a 128 point scale.
3068 * degrade_zero_ticks is the number of ticks after which load at any
3069 * particular idx is approximated to be zero.
3070 * degrade_factor is a precomputed table, a row for each load idx.
3071 * Each column corresponds to degradation factor for a power of two ticks,
3072 * based on 128 point scale.
3074 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3075 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3077 * With this power of 2 load factors, we can degrade the load n times
3078 * by looking at 1 bits in n and doing as many mult/shift instead of
3079 * n mult/shifts needed by the exact degradation.
3081 #define DEGRADE_SHIFT 7
3082 static const unsigned char
3083 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3084 static const unsigned char
3085 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3086 {0, 0, 0, 0, 0, 0, 0, 0},
3087 {64, 32, 8, 0, 0, 0, 0, 0},
3088 {96, 72, 40, 12, 1, 0, 0},
3089 {112, 98, 75, 43, 15, 1, 0},
3090 {120, 112, 98, 76, 45, 16, 2} };
3093 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3094 * would be when CPU is idle and so we just decay the old load without
3095 * adding any new load.
3097 static unsigned long
3098 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3102 if (!missed_updates)
3105 if (missed_updates >= degrade_zero_ticks[idx])
3109 return load >> missed_updates;
3111 while (missed_updates) {
3112 if (missed_updates % 2)
3113 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3115 missed_updates >>= 1;
3122 * Update rq->cpu_load[] statistics. This function is usually called every
3123 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3124 * every tick. We fix it up based on jiffies.
3126 static void update_cpu_load(struct rq *this_rq)
3128 unsigned long this_load = this_rq->load.weight;
3129 unsigned long curr_jiffies = jiffies;
3130 unsigned long pending_updates;
3133 this_rq->nr_load_updates++;
3135 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3136 if (curr_jiffies == this_rq->last_load_update_tick)
3139 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3140 this_rq->last_load_update_tick = curr_jiffies;
3142 /* Update our load: */
3143 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3144 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3145 unsigned long old_load, new_load;
3147 /* scale is effectively 1 << i now, and >> i divides by scale */
3149 old_load = this_rq->cpu_load[i];
3150 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3151 new_load = this_load;
3153 * Round up the averaging division if load is increasing. This
3154 * prevents us from getting stuck on 9 if the load is 10, for
3157 if (new_load > old_load)
3158 new_load += scale - 1;
3160 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3164 static void update_cpu_load_active(struct rq *this_rq)
3166 update_cpu_load(this_rq);
3168 calc_load_account_active(this_rq);
3174 * sched_exec - execve() is a valuable balancing opportunity, because at
3175 * this point the task has the smallest effective memory and cache footprint.
3177 void sched_exec(void)
3179 struct task_struct *p = current;
3180 unsigned long flags;
3184 rq = task_rq_lock(p, &flags);
3185 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3186 if (dest_cpu == smp_processor_id())
3190 * select_task_rq() can race against ->cpus_allowed
3192 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3193 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3194 struct migration_arg arg = { p, dest_cpu };
3196 task_rq_unlock(rq, &flags);
3197 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3201 task_rq_unlock(rq, &flags);
3206 DEFINE_PER_CPU(struct kernel_stat, kstat);
3208 EXPORT_PER_CPU_SYMBOL(kstat);
3211 * Return any ns on the sched_clock that have not yet been accounted in
3212 * @p in case that task is currently running.
3214 * Called with task_rq_lock() held on @rq.
3216 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3220 if (task_current(rq, p)) {
3221 update_rq_clock(rq);
3222 ns = rq->clock - p->se.exec_start;
3230 unsigned long long task_delta_exec(struct task_struct *p)
3232 unsigned long flags;
3236 rq = task_rq_lock(p, &flags);
3237 ns = do_task_delta_exec(p, rq);
3238 task_rq_unlock(rq, &flags);
3244 * Return accounted runtime for the task.
3245 * In case the task is currently running, return the runtime plus current's
3246 * pending runtime that have not been accounted yet.
3248 unsigned long long task_sched_runtime(struct task_struct *p)
3250 unsigned long flags;
3254 rq = task_rq_lock(p, &flags);
3255 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3256 task_rq_unlock(rq, &flags);
3262 * Return sum_exec_runtime for the thread group.
3263 * In case the task is currently running, return the sum plus current's
3264 * pending runtime that have not been accounted yet.
3266 * Note that the thread group might have other running tasks as well,
3267 * so the return value not includes other pending runtime that other
3268 * running tasks might have.
3270 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3272 struct task_cputime totals;
3273 unsigned long flags;
3277 rq = task_rq_lock(p, &flags);
3278 thread_group_cputime(p, &totals);
3279 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3280 task_rq_unlock(rq, &flags);
3286 * Account user cpu time to a process.
3287 * @p: the process that the cpu time gets accounted to
3288 * @cputime: the cpu time spent in user space since the last update
3289 * @cputime_scaled: cputime scaled by cpu frequency
3291 void account_user_time(struct task_struct *p, cputime_t cputime,
3292 cputime_t cputime_scaled)
3294 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3297 /* Add user time to process. */
3298 p->utime = cputime_add(p->utime, cputime);
3299 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3300 account_group_user_time(p, cputime);
3302 /* Add user time to cpustat. */
3303 tmp = cputime_to_cputime64(cputime);
3304 if (TASK_NICE(p) > 0)
3305 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3307 cpustat->user = cputime64_add(cpustat->user, tmp);
3309 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3310 /* Account for user time used */
3311 acct_update_integrals(p);
3315 * Account guest cpu time to a process.
3316 * @p: the process that the cpu time gets accounted to
3317 * @cputime: the cpu time spent in virtual machine since the last update
3318 * @cputime_scaled: cputime scaled by cpu frequency
3320 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3321 cputime_t cputime_scaled)
3324 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3326 tmp = cputime_to_cputime64(cputime);
3328 /* Add guest time to process. */
3329 p->utime = cputime_add(p->utime, cputime);
3330 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3331 account_group_user_time(p, cputime);
3332 p->gtime = cputime_add(p->gtime, cputime);
3334 /* Add guest time to cpustat. */
3335 if (TASK_NICE(p) > 0) {
3336 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3337 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3339 cpustat->user = cputime64_add(cpustat->user, tmp);
3340 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3345 * Account system cpu time to a process.
3346 * @p: the process that the cpu time gets accounted to
3347 * @hardirq_offset: the offset to subtract from hardirq_count()
3348 * @cputime: the cpu time spent in kernel space since the last update
3349 * @cputime_scaled: cputime scaled by cpu frequency
3351 void account_system_time(struct task_struct *p, int hardirq_offset,
3352 cputime_t cputime, cputime_t cputime_scaled)
3354 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3357 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3358 account_guest_time(p, cputime, cputime_scaled);
3362 /* Add system time to process. */
3363 p->stime = cputime_add(p->stime, cputime);
3364 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3365 account_group_system_time(p, cputime);
3367 /* Add system time to cpustat. */
3368 tmp = cputime_to_cputime64(cputime);
3369 if (hardirq_count() - hardirq_offset)
3370 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3371 else if (softirq_count())
3372 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3374 cpustat->system = cputime64_add(cpustat->system, tmp);
3376 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3378 /* Account for system time used */
3379 acct_update_integrals(p);
3383 * Account for involuntary wait time.
3384 * @steal: the cpu time spent in involuntary wait
3386 void account_steal_time(cputime_t cputime)
3388 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3389 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3391 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3395 * Account for idle time.
3396 * @cputime: the cpu time spent in idle wait
3398 void account_idle_time(cputime_t cputime)
3400 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3401 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3402 struct rq *rq = this_rq();
3404 if (atomic_read(&rq->nr_iowait) > 0)
3405 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3407 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3410 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3413 * Account a single tick of cpu time.
3414 * @p: the process that the cpu time gets accounted to
3415 * @user_tick: indicates if the tick is a user or a system tick
3417 void account_process_tick(struct task_struct *p, int user_tick)
3419 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3420 struct rq *rq = this_rq();
3423 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3424 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3425 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3428 account_idle_time(cputime_one_jiffy);
3432 * Account multiple ticks of steal time.
3433 * @p: the process from which the cpu time has been stolen
3434 * @ticks: number of stolen ticks
3436 void account_steal_ticks(unsigned long ticks)
3438 account_steal_time(jiffies_to_cputime(ticks));
3442 * Account multiple ticks of idle time.
3443 * @ticks: number of stolen ticks
3445 void account_idle_ticks(unsigned long ticks)
3447 account_idle_time(jiffies_to_cputime(ticks));
3453 * Use precise platform statistics if available:
3455 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3456 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3462 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3464 struct task_cputime cputime;
3466 thread_group_cputime(p, &cputime);
3468 *ut = cputime.utime;
3469 *st = cputime.stime;
3473 #ifndef nsecs_to_cputime
3474 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3477 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3479 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3482 * Use CFS's precise accounting:
3484 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3489 temp = (u64)(rtime * utime);
3490 do_div(temp, total);
3491 utime = (cputime_t)temp;
3496 * Compare with previous values, to keep monotonicity:
3498 p->prev_utime = max(p->prev_utime, utime);
3499 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3501 *ut = p->prev_utime;
3502 *st = p->prev_stime;
3506 * Must be called with siglock held.
3508 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3510 struct signal_struct *sig = p->signal;
3511 struct task_cputime cputime;
3512 cputime_t rtime, utime, total;
3514 thread_group_cputime(p, &cputime);
3516 total = cputime_add(cputime.utime, cputime.stime);
3517 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3522 temp = (u64)(rtime * cputime.utime);
3523 do_div(temp, total);
3524 utime = (cputime_t)temp;
3528 sig->prev_utime = max(sig->prev_utime, utime);
3529 sig->prev_stime = max(sig->prev_stime,
3530 cputime_sub(rtime, sig->prev_utime));
3532 *ut = sig->prev_utime;
3533 *st = sig->prev_stime;
3538 * This function gets called by the timer code, with HZ frequency.
3539 * We call it with interrupts disabled.
3541 * It also gets called by the fork code, when changing the parent's
3544 void scheduler_tick(void)
3546 int cpu = smp_processor_id();
3547 struct rq *rq = cpu_rq(cpu);
3548 struct task_struct *curr = rq->curr;
3552 raw_spin_lock(&rq->lock);
3553 update_rq_clock(rq);
3554 update_cpu_load_active(rq);
3555 curr->sched_class->task_tick(rq, curr, 0);
3556 raw_spin_unlock(&rq->lock);
3558 perf_event_task_tick(curr);
3561 rq->idle_at_tick = idle_cpu(cpu);
3562 trigger_load_balance(rq, cpu);
3566 notrace unsigned long get_parent_ip(unsigned long addr)
3568 if (in_lock_functions(addr)) {
3569 addr = CALLER_ADDR2;
3570 if (in_lock_functions(addr))
3571 addr = CALLER_ADDR3;
3576 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3577 defined(CONFIG_PREEMPT_TRACER))
3579 void __kprobes add_preempt_count(int val)
3581 #ifdef CONFIG_DEBUG_PREEMPT
3585 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3588 preempt_count() += val;
3589 #ifdef CONFIG_DEBUG_PREEMPT
3591 * Spinlock count overflowing soon?
3593 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3596 if (preempt_count() == val)
3597 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3599 EXPORT_SYMBOL(add_preempt_count);
3601 void __kprobes sub_preempt_count(int val)
3603 #ifdef CONFIG_DEBUG_PREEMPT
3607 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3610 * Is the spinlock portion underflowing?
3612 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3613 !(preempt_count() & PREEMPT_MASK)))
3617 if (preempt_count() == val)
3618 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3619 preempt_count() -= val;
3621 EXPORT_SYMBOL(sub_preempt_count);
3626 * Print scheduling while atomic bug:
3628 static noinline void __schedule_bug(struct task_struct *prev)
3630 struct pt_regs *regs = get_irq_regs();
3632 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3633 prev->comm, prev->pid, preempt_count());
3635 debug_show_held_locks(prev);
3637 if (irqs_disabled())
3638 print_irqtrace_events(prev);
3647 * Various schedule()-time debugging checks and statistics:
3649 static inline void schedule_debug(struct task_struct *prev)
3652 * Test if we are atomic. Since do_exit() needs to call into
3653 * schedule() atomically, we ignore that path for now.
3654 * Otherwise, whine if we are scheduling when we should not be.
3656 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3657 __schedule_bug(prev);
3659 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3661 schedstat_inc(this_rq(), sched_count);
3662 #ifdef CONFIG_SCHEDSTATS
3663 if (unlikely(prev->lock_depth >= 0)) {
3664 schedstat_inc(this_rq(), bkl_count);
3665 schedstat_inc(prev, sched_info.bkl_count);
3670 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3673 update_rq_clock(rq);
3674 rq->skip_clock_update = 0;
3675 prev->sched_class->put_prev_task(rq, prev);
3679 * Pick up the highest-prio task:
3681 static inline struct task_struct *
3682 pick_next_task(struct rq *rq)
3684 const struct sched_class *class;
3685 struct task_struct *p;
3688 * Optimization: we know that if all tasks are in
3689 * the fair class we can call that function directly:
3691 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3692 p = fair_sched_class.pick_next_task(rq);
3697 class = sched_class_highest;
3699 p = class->pick_next_task(rq);
3703 * Will never be NULL as the idle class always
3704 * returns a non-NULL p:
3706 class = class->next;
3711 * schedule() is the main scheduler function.
3713 asmlinkage void __sched schedule(void)
3715 struct task_struct *prev, *next;
3716 unsigned long *switch_count;
3722 cpu = smp_processor_id();
3724 rcu_note_context_switch(cpu);
3727 release_kernel_lock(prev);
3728 need_resched_nonpreemptible:
3730 schedule_debug(prev);
3732 if (sched_feat(HRTICK))
3735 raw_spin_lock_irq(&rq->lock);
3736 clear_tsk_need_resched(prev);
3738 switch_count = &prev->nivcsw;
3739 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3740 if (unlikely(signal_pending_state(prev->state, prev))) {
3741 prev->state = TASK_RUNNING;
3744 * If a worker is going to sleep, notify and
3745 * ask workqueue whether it wants to wake up a
3746 * task to maintain concurrency. If so, wake
3749 if (prev->flags & PF_WQ_WORKER) {
3750 struct task_struct *to_wakeup;
3752 to_wakeup = wq_worker_sleeping(prev, cpu);
3754 try_to_wake_up_local(to_wakeup);
3756 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3758 switch_count = &prev->nvcsw;
3761 pre_schedule(rq, prev);
3763 if (unlikely(!rq->nr_running))
3764 idle_balance(cpu, rq);
3766 put_prev_task(rq, prev);
3767 next = pick_next_task(rq);
3769 if (likely(prev != next)) {
3770 sched_info_switch(prev, next);
3771 perf_event_task_sched_out(prev, next);
3777 context_switch(rq, prev, next); /* unlocks the rq */
3779 * The context switch have flipped the stack from under us
3780 * and restored the local variables which were saved when
3781 * this task called schedule() in the past. prev == current
3782 * is still correct, but it can be moved to another cpu/rq.
3784 cpu = smp_processor_id();
3787 raw_spin_unlock_irq(&rq->lock);
3791 if (unlikely(reacquire_kernel_lock(prev)))
3792 goto need_resched_nonpreemptible;
3794 preempt_enable_no_resched();
3798 EXPORT_SYMBOL(schedule);
3800 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3802 * Look out! "owner" is an entirely speculative pointer
3803 * access and not reliable.
3805 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3810 if (!sched_feat(OWNER_SPIN))
3813 #ifdef CONFIG_DEBUG_PAGEALLOC
3815 * Need to access the cpu field knowing that
3816 * DEBUG_PAGEALLOC could have unmapped it if
3817 * the mutex owner just released it and exited.
3819 if (probe_kernel_address(&owner->cpu, cpu))
3826 * Even if the access succeeded (likely case),
3827 * the cpu field may no longer be valid.
3829 if (cpu >= nr_cpumask_bits)
3833 * We need to validate that we can do a
3834 * get_cpu() and that we have the percpu area.
3836 if (!cpu_online(cpu))
3843 * Owner changed, break to re-assess state.
3845 if (lock->owner != owner)
3849 * Is that owner really running on that cpu?
3851 if (task_thread_info(rq->curr) != owner || need_resched())
3861 #ifdef CONFIG_PREEMPT
3863 * this is the entry point to schedule() from in-kernel preemption
3864 * off of preempt_enable. Kernel preemptions off return from interrupt
3865 * occur there and call schedule directly.
3867 asmlinkage void __sched preempt_schedule(void)
3869 struct thread_info *ti = current_thread_info();
3872 * If there is a non-zero preempt_count or interrupts are disabled,
3873 * we do not want to preempt the current task. Just return..
3875 if (likely(ti->preempt_count || irqs_disabled()))
3879 add_preempt_count(PREEMPT_ACTIVE);
3881 sub_preempt_count(PREEMPT_ACTIVE);
3884 * Check again in case we missed a preemption opportunity
3885 * between schedule and now.
3888 } while (need_resched());
3890 EXPORT_SYMBOL(preempt_schedule);
3893 * this is the entry point to schedule() from kernel preemption
3894 * off of irq context.
3895 * Note, that this is called and return with irqs disabled. This will
3896 * protect us against recursive calling from irq.
3898 asmlinkage void __sched preempt_schedule_irq(void)
3900 struct thread_info *ti = current_thread_info();
3902 /* Catch callers which need to be fixed */
3903 BUG_ON(ti->preempt_count || !irqs_disabled());
3906 add_preempt_count(PREEMPT_ACTIVE);
3909 local_irq_disable();
3910 sub_preempt_count(PREEMPT_ACTIVE);
3913 * Check again in case we missed a preemption opportunity
3914 * between schedule and now.
3917 } while (need_resched());
3920 #endif /* CONFIG_PREEMPT */
3922 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3925 return try_to_wake_up(curr->private, mode, wake_flags);
3927 EXPORT_SYMBOL(default_wake_function);
3930 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3931 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3932 * number) then we wake all the non-exclusive tasks and one exclusive task.
3934 * There are circumstances in which we can try to wake a task which has already
3935 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3936 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3938 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3939 int nr_exclusive, int wake_flags, void *key)
3941 wait_queue_t *curr, *next;
3943 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3944 unsigned flags = curr->flags;
3946 if (curr->func(curr, mode, wake_flags, key) &&
3947 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3953 * __wake_up - wake up threads blocked on a waitqueue.
3955 * @mode: which threads
3956 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3957 * @key: is directly passed to the wakeup function
3959 * It may be assumed that this function implies a write memory barrier before
3960 * changing the task state if and only if any tasks are woken up.
3962 void __wake_up(wait_queue_head_t *q, unsigned int mode,
3963 int nr_exclusive, void *key)
3965 unsigned long flags;
3967 spin_lock_irqsave(&q->lock, flags);
3968 __wake_up_common(q, mode, nr_exclusive, 0, key);
3969 spin_unlock_irqrestore(&q->lock, flags);
3971 EXPORT_SYMBOL(__wake_up);
3974 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3976 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
3978 __wake_up_common(q, mode, 1, 0, NULL);
3980 EXPORT_SYMBOL_GPL(__wake_up_locked);
3982 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
3984 __wake_up_common(q, mode, 1, 0, key);
3988 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3990 * @mode: which threads
3991 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3992 * @key: opaque value to be passed to wakeup targets
3994 * The sync wakeup differs that the waker knows that it will schedule
3995 * away soon, so while the target thread will be woken up, it will not
3996 * be migrated to another CPU - ie. the two threads are 'synchronized'
3997 * with each other. This can prevent needless bouncing between CPUs.
3999 * On UP it can prevent extra preemption.
4001 * It may be assumed that this function implies a write memory barrier before
4002 * changing the task state if and only if any tasks are woken up.
4004 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4005 int nr_exclusive, void *key)
4007 unsigned long flags;
4008 int wake_flags = WF_SYNC;
4013 if (unlikely(!nr_exclusive))
4016 spin_lock_irqsave(&q->lock, flags);
4017 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4018 spin_unlock_irqrestore(&q->lock, flags);
4020 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4023 * __wake_up_sync - see __wake_up_sync_key()
4025 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4027 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4029 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4032 * complete: - signals a single thread waiting on this completion
4033 * @x: holds the state of this particular completion
4035 * This will wake up a single thread waiting on this completion. Threads will be
4036 * awakened in the same order in which they were queued.
4038 * See also complete_all(), wait_for_completion() and related routines.
4040 * It may be assumed that this function implies a write memory barrier before
4041 * changing the task state if and only if any tasks are woken up.
4043 void complete(struct completion *x)
4045 unsigned long flags;
4047 spin_lock_irqsave(&x->wait.lock, flags);
4049 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4050 spin_unlock_irqrestore(&x->wait.lock, flags);
4052 EXPORT_SYMBOL(complete);
4055 * complete_all: - signals all threads waiting on this completion
4056 * @x: holds the state of this particular completion
4058 * This will wake up all threads waiting on this particular completion event.
4060 * It may be assumed that this function implies a write memory barrier before
4061 * changing the task state if and only if any tasks are woken up.
4063 void complete_all(struct completion *x)
4065 unsigned long flags;
4067 spin_lock_irqsave(&x->wait.lock, flags);
4068 x->done += UINT_MAX/2;
4069 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4070 spin_unlock_irqrestore(&x->wait.lock, flags);
4072 EXPORT_SYMBOL(complete_all);
4074 static inline long __sched
4075 do_wait_for_common(struct completion *x, long timeout, int state)
4078 DECLARE_WAITQUEUE(wait, current);
4080 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4082 if (signal_pending_state(state, current)) {
4083 timeout = -ERESTARTSYS;
4086 __set_current_state(state);
4087 spin_unlock_irq(&x->wait.lock);
4088 timeout = schedule_timeout(timeout);
4089 spin_lock_irq(&x->wait.lock);
4090 } while (!x->done && timeout);
4091 __remove_wait_queue(&x->wait, &wait);
4096 return timeout ?: 1;
4100 wait_for_common(struct completion *x, long timeout, int state)
4104 spin_lock_irq(&x->wait.lock);
4105 timeout = do_wait_for_common(x, timeout, state);
4106 spin_unlock_irq(&x->wait.lock);
4111 * wait_for_completion: - waits for completion of a task
4112 * @x: holds the state of this particular completion
4114 * This waits to be signaled for completion of a specific task. It is NOT
4115 * interruptible and there is no timeout.
4117 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4118 * and interrupt capability. Also see complete().
4120 void __sched wait_for_completion(struct completion *x)
4122 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4124 EXPORT_SYMBOL(wait_for_completion);
4127 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4128 * @x: holds the state of this particular completion
4129 * @timeout: timeout value in jiffies
4131 * This waits for either a completion of a specific task to be signaled or for a
4132 * specified timeout to expire. The timeout is in jiffies. It is not
4135 unsigned long __sched
4136 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4138 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4140 EXPORT_SYMBOL(wait_for_completion_timeout);
4143 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4144 * @x: holds the state of this particular completion
4146 * This waits for completion of a specific task to be signaled. It is
4149 int __sched wait_for_completion_interruptible(struct completion *x)
4151 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4152 if (t == -ERESTARTSYS)
4156 EXPORT_SYMBOL(wait_for_completion_interruptible);
4159 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4160 * @x: holds the state of this particular completion
4161 * @timeout: timeout value in jiffies
4163 * This waits for either a completion of a specific task to be signaled or for a
4164 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4166 unsigned long __sched
4167 wait_for_completion_interruptible_timeout(struct completion *x,
4168 unsigned long timeout)
4170 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4172 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4175 * wait_for_completion_killable: - waits for completion of a task (killable)
4176 * @x: holds the state of this particular completion
4178 * This waits to be signaled for completion of a specific task. It can be
4179 * interrupted by a kill signal.
4181 int __sched wait_for_completion_killable(struct completion *x)
4183 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4184 if (t == -ERESTARTSYS)
4188 EXPORT_SYMBOL(wait_for_completion_killable);
4191 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4192 * @x: holds the state of this particular completion
4193 * @timeout: timeout value in jiffies
4195 * This waits for either a completion of a specific task to be
4196 * signaled or for a specified timeout to expire. It can be
4197 * interrupted by a kill signal. The timeout is in jiffies.
4199 unsigned long __sched
4200 wait_for_completion_killable_timeout(struct completion *x,
4201 unsigned long timeout)
4203 return wait_for_common(x, timeout, TASK_KILLABLE);
4205 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4208 * try_wait_for_completion - try to decrement a completion without blocking
4209 * @x: completion structure
4211 * Returns: 0 if a decrement cannot be done without blocking
4212 * 1 if a decrement succeeded.
4214 * If a completion is being used as a counting completion,
4215 * attempt to decrement the counter without blocking. This
4216 * enables us to avoid waiting if the resource the completion
4217 * is protecting is not available.
4219 bool try_wait_for_completion(struct completion *x)
4221 unsigned long flags;
4224 spin_lock_irqsave(&x->wait.lock, flags);
4229 spin_unlock_irqrestore(&x->wait.lock, flags);
4232 EXPORT_SYMBOL(try_wait_for_completion);
4235 * completion_done - Test to see if a completion has any waiters
4236 * @x: completion structure
4238 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4239 * 1 if there are no waiters.
4242 bool completion_done(struct completion *x)
4244 unsigned long flags;
4247 spin_lock_irqsave(&x->wait.lock, flags);
4250 spin_unlock_irqrestore(&x->wait.lock, flags);
4253 EXPORT_SYMBOL(completion_done);
4256 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4258 unsigned long flags;
4261 init_waitqueue_entry(&wait, current);
4263 __set_current_state(state);
4265 spin_lock_irqsave(&q->lock, flags);
4266 __add_wait_queue(q, &wait);
4267 spin_unlock(&q->lock);
4268 timeout = schedule_timeout(timeout);
4269 spin_lock_irq(&q->lock);
4270 __remove_wait_queue(q, &wait);
4271 spin_unlock_irqrestore(&q->lock, flags);
4276 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4278 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4280 EXPORT_SYMBOL(interruptible_sleep_on);
4283 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4285 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4287 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4289 void __sched sleep_on(wait_queue_head_t *q)
4291 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4293 EXPORT_SYMBOL(sleep_on);
4295 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4297 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4299 EXPORT_SYMBOL(sleep_on_timeout);
4301 #ifdef CONFIG_RT_MUTEXES
4304 * rt_mutex_setprio - set the current priority of a task
4306 * @prio: prio value (kernel-internal form)
4308 * This function changes the 'effective' priority of a task. It does
4309 * not touch ->normal_prio like __setscheduler().
4311 * Used by the rt_mutex code to implement priority inheritance logic.
4313 void rt_mutex_setprio(struct task_struct *p, int prio)
4315 unsigned long flags;
4316 int oldprio, on_rq, running;
4318 const struct sched_class *prev_class;
4320 BUG_ON(prio < 0 || prio > MAX_PRIO);
4322 rq = task_rq_lock(p, &flags);
4325 prev_class = p->sched_class;
4326 on_rq = p->se.on_rq;
4327 running = task_current(rq, p);
4329 dequeue_task(rq, p, 0);
4331 p->sched_class->put_prev_task(rq, p);
4334 p->sched_class = &rt_sched_class;
4336 p->sched_class = &fair_sched_class;
4341 p->sched_class->set_curr_task(rq);
4343 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4345 check_class_changed(rq, p, prev_class, oldprio, running);
4347 task_rq_unlock(rq, &flags);
4352 void set_user_nice(struct task_struct *p, long nice)
4354 int old_prio, delta, on_rq;
4355 unsigned long flags;
4358 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4361 * We have to be careful, if called from sys_setpriority(),
4362 * the task might be in the middle of scheduling on another CPU.
4364 rq = task_rq_lock(p, &flags);
4366 * The RT priorities are set via sched_setscheduler(), but we still
4367 * allow the 'normal' nice value to be set - but as expected
4368 * it wont have any effect on scheduling until the task is
4369 * SCHED_FIFO/SCHED_RR:
4371 if (task_has_rt_policy(p)) {
4372 p->static_prio = NICE_TO_PRIO(nice);
4375 on_rq = p->se.on_rq;
4377 dequeue_task(rq, p, 0);
4379 p->static_prio = NICE_TO_PRIO(nice);
4382 p->prio = effective_prio(p);
4383 delta = p->prio - old_prio;
4386 enqueue_task(rq, p, 0);
4388 * If the task increased its priority or is running and
4389 * lowered its priority, then reschedule its CPU:
4391 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4392 resched_task(rq->curr);
4395 task_rq_unlock(rq, &flags);
4397 EXPORT_SYMBOL(set_user_nice);
4400 * can_nice - check if a task can reduce its nice value
4404 int can_nice(const struct task_struct *p, const int nice)
4406 /* convert nice value [19,-20] to rlimit style value [1,40] */
4407 int nice_rlim = 20 - nice;
4409 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4410 capable(CAP_SYS_NICE));
4413 #ifdef __ARCH_WANT_SYS_NICE
4416 * sys_nice - change the priority of the current process.
4417 * @increment: priority increment
4419 * sys_setpriority is a more generic, but much slower function that
4420 * does similar things.
4422 SYSCALL_DEFINE1(nice, int, increment)
4427 * Setpriority might change our priority at the same moment.
4428 * We don't have to worry. Conceptually one call occurs first
4429 * and we have a single winner.
4431 if (increment < -40)
4436 nice = TASK_NICE(current) + increment;
4442 if (increment < 0 && !can_nice(current, nice))
4445 retval = security_task_setnice(current, nice);
4449 set_user_nice(current, nice);
4456 * task_prio - return the priority value of a given task.
4457 * @p: the task in question.
4459 * This is the priority value as seen by users in /proc.
4460 * RT tasks are offset by -200. Normal tasks are centered
4461 * around 0, value goes from -16 to +15.
4463 int task_prio(const struct task_struct *p)
4465 return p->prio - MAX_RT_PRIO;
4469 * task_nice - return the nice value of a given task.
4470 * @p: the task in question.
4472 int task_nice(const struct task_struct *p)
4474 return TASK_NICE(p);
4476 EXPORT_SYMBOL(task_nice);
4479 * idle_cpu - is a given cpu idle currently?
4480 * @cpu: the processor in question.
4482 int idle_cpu(int cpu)
4484 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4488 * idle_task - return the idle task for a given cpu.
4489 * @cpu: the processor in question.
4491 struct task_struct *idle_task(int cpu)
4493 return cpu_rq(cpu)->idle;
4497 * find_process_by_pid - find a process with a matching PID value.
4498 * @pid: the pid in question.
4500 static struct task_struct *find_process_by_pid(pid_t pid)
4502 return pid ? find_task_by_vpid(pid) : current;
4505 /* Actually do priority change: must hold rq lock. */
4507 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4509 BUG_ON(p->se.on_rq);
4512 p->rt_priority = prio;
4513 p->normal_prio = normal_prio(p);
4514 /* we are holding p->pi_lock already */
4515 p->prio = rt_mutex_getprio(p);
4516 if (rt_prio(p->prio))
4517 p->sched_class = &rt_sched_class;
4519 p->sched_class = &fair_sched_class;
4524 * check the target process has a UID that matches the current process's
4526 static bool check_same_owner(struct task_struct *p)
4528 const struct cred *cred = current_cred(), *pcred;
4532 pcred = __task_cred(p);
4533 match = (cred->euid == pcred->euid ||
4534 cred->euid == pcred->uid);
4539 static int __sched_setscheduler(struct task_struct *p, int policy,
4540 struct sched_param *param, bool user)
4542 int retval, oldprio, oldpolicy = -1, on_rq, running;
4543 unsigned long flags;
4544 const struct sched_class *prev_class;
4548 /* may grab non-irq protected spin_locks */
4549 BUG_ON(in_interrupt());
4551 /* double check policy once rq lock held */
4553 reset_on_fork = p->sched_reset_on_fork;
4554 policy = oldpolicy = p->policy;
4556 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4557 policy &= ~SCHED_RESET_ON_FORK;
4559 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4560 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4561 policy != SCHED_IDLE)
4566 * Valid priorities for SCHED_FIFO and SCHED_RR are
4567 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4568 * SCHED_BATCH and SCHED_IDLE is 0.
4570 if (param->sched_priority < 0 ||
4571 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4572 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4574 if (rt_policy(policy) != (param->sched_priority != 0))
4578 * Allow unprivileged RT tasks to decrease priority:
4580 if (user && !capable(CAP_SYS_NICE)) {
4581 if (rt_policy(policy)) {
4582 unsigned long rlim_rtprio;
4584 if (!lock_task_sighand(p, &flags))
4586 rlim_rtprio = task_rlimit(p, RLIMIT_RTPRIO);
4587 unlock_task_sighand(p, &flags);
4589 /* can't set/change the rt policy */
4590 if (policy != p->policy && !rlim_rtprio)
4593 /* can't increase priority */
4594 if (param->sched_priority > p->rt_priority &&
4595 param->sched_priority > rlim_rtprio)
4599 * Like positive nice levels, dont allow tasks to
4600 * move out of SCHED_IDLE either:
4602 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4605 /* can't change other user's priorities */
4606 if (!check_same_owner(p))
4609 /* Normal users shall not reset the sched_reset_on_fork flag */
4610 if (p->sched_reset_on_fork && !reset_on_fork)
4615 retval = security_task_setscheduler(p, policy, param);
4621 * make sure no PI-waiters arrive (or leave) while we are
4622 * changing the priority of the task:
4624 raw_spin_lock_irqsave(&p->pi_lock, flags);
4626 * To be able to change p->policy safely, the apropriate
4627 * runqueue lock must be held.
4629 rq = __task_rq_lock(p);
4631 #ifdef CONFIG_RT_GROUP_SCHED
4634 * Do not allow realtime tasks into groups that have no runtime
4637 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4638 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4639 __task_rq_unlock(rq);
4640 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4646 /* recheck policy now with rq lock held */
4647 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4648 policy = oldpolicy = -1;
4649 __task_rq_unlock(rq);
4650 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4653 on_rq = p->se.on_rq;
4654 running = task_current(rq, p);
4656 deactivate_task(rq, p, 0);
4658 p->sched_class->put_prev_task(rq, p);
4660 p->sched_reset_on_fork = reset_on_fork;
4663 prev_class = p->sched_class;
4664 __setscheduler(rq, p, policy, param->sched_priority);
4667 p->sched_class->set_curr_task(rq);
4669 activate_task(rq, p, 0);
4671 check_class_changed(rq, p, prev_class, oldprio, running);
4673 __task_rq_unlock(rq);
4674 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4676 rt_mutex_adjust_pi(p);
4682 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4683 * @p: the task in question.
4684 * @policy: new policy.
4685 * @param: structure containing the new RT priority.
4687 * NOTE that the task may be already dead.
4689 int sched_setscheduler(struct task_struct *p, int policy,
4690 struct sched_param *param)
4692 return __sched_setscheduler(p, policy, param, true);
4694 EXPORT_SYMBOL_GPL(sched_setscheduler);
4697 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4698 * @p: the task in question.
4699 * @policy: new policy.
4700 * @param: structure containing the new RT priority.
4702 * Just like sched_setscheduler, only don't bother checking if the
4703 * current context has permission. For example, this is needed in
4704 * stop_machine(): we create temporary high priority worker threads,
4705 * but our caller might not have that capability.
4707 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4708 struct sched_param *param)
4710 return __sched_setscheduler(p, policy, param, false);
4714 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4716 struct sched_param lparam;
4717 struct task_struct *p;
4720 if (!param || pid < 0)
4722 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4727 p = find_process_by_pid(pid);
4729 retval = sched_setscheduler(p, policy, &lparam);
4736 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4737 * @pid: the pid in question.
4738 * @policy: new policy.
4739 * @param: structure containing the new RT priority.
4741 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4742 struct sched_param __user *, param)
4744 /* negative values for policy are not valid */
4748 return do_sched_setscheduler(pid, policy, param);
4752 * sys_sched_setparam - set/change the RT priority of a thread
4753 * @pid: the pid in question.
4754 * @param: structure containing the new RT priority.
4756 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4758 return do_sched_setscheduler(pid, -1, param);
4762 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4763 * @pid: the pid in question.
4765 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4767 struct task_struct *p;
4775 p = find_process_by_pid(pid);
4777 retval = security_task_getscheduler(p);
4780 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4787 * sys_sched_getparam - get the RT priority of a thread
4788 * @pid: the pid in question.
4789 * @param: structure containing the RT priority.
4791 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4793 struct sched_param lp;
4794 struct task_struct *p;
4797 if (!param || pid < 0)
4801 p = find_process_by_pid(pid);
4806 retval = security_task_getscheduler(p);
4810 lp.sched_priority = p->rt_priority;
4814 * This one might sleep, we cannot do it with a spinlock held ...
4816 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4825 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4827 cpumask_var_t cpus_allowed, new_mask;
4828 struct task_struct *p;
4834 p = find_process_by_pid(pid);
4841 /* Prevent p going away */
4845 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4849 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4851 goto out_free_cpus_allowed;
4854 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4857 retval = security_task_setscheduler(p, 0, NULL);
4861 cpuset_cpus_allowed(p, cpus_allowed);
4862 cpumask_and(new_mask, in_mask, cpus_allowed);
4864 retval = set_cpus_allowed_ptr(p, new_mask);
4867 cpuset_cpus_allowed(p, cpus_allowed);
4868 if (!cpumask_subset(new_mask, cpus_allowed)) {
4870 * We must have raced with a concurrent cpuset
4871 * update. Just reset the cpus_allowed to the
4872 * cpuset's cpus_allowed
4874 cpumask_copy(new_mask, cpus_allowed);
4879 free_cpumask_var(new_mask);
4880 out_free_cpus_allowed:
4881 free_cpumask_var(cpus_allowed);
4888 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4889 struct cpumask *new_mask)
4891 if (len < cpumask_size())
4892 cpumask_clear(new_mask);
4893 else if (len > cpumask_size())
4894 len = cpumask_size();
4896 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4900 * sys_sched_setaffinity - set the cpu affinity of a process
4901 * @pid: pid of the process
4902 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4903 * @user_mask_ptr: user-space pointer to the new cpu mask
4905 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4906 unsigned long __user *, user_mask_ptr)
4908 cpumask_var_t new_mask;
4911 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4914 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4916 retval = sched_setaffinity(pid, new_mask);
4917 free_cpumask_var(new_mask);
4921 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4923 struct task_struct *p;
4924 unsigned long flags;
4932 p = find_process_by_pid(pid);
4936 retval = security_task_getscheduler(p);
4940 rq = task_rq_lock(p, &flags);
4941 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4942 task_rq_unlock(rq, &flags);
4952 * sys_sched_getaffinity - get the cpu affinity of a process
4953 * @pid: pid of the process
4954 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4955 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4957 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4958 unsigned long __user *, user_mask_ptr)
4963 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4965 if (len & (sizeof(unsigned long)-1))
4968 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
4971 ret = sched_getaffinity(pid, mask);
4973 size_t retlen = min_t(size_t, len, cpumask_size());
4975 if (copy_to_user(user_mask_ptr, mask, retlen))
4980 free_cpumask_var(mask);
4986 * sys_sched_yield - yield the current processor to other threads.
4988 * This function yields the current CPU to other tasks. If there are no
4989 * other threads running on this CPU then this function will return.
4991 SYSCALL_DEFINE0(sched_yield)
4993 struct rq *rq = this_rq_lock();
4995 schedstat_inc(rq, yld_count);
4996 current->sched_class->yield_task(rq);
4999 * Since we are going to call schedule() anyway, there's
5000 * no need to preempt or enable interrupts:
5002 __release(rq->lock);
5003 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5004 do_raw_spin_unlock(&rq->lock);
5005 preempt_enable_no_resched();
5012 static inline int should_resched(void)
5014 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5017 static void __cond_resched(void)
5019 add_preempt_count(PREEMPT_ACTIVE);
5021 sub_preempt_count(PREEMPT_ACTIVE);
5024 int __sched _cond_resched(void)
5026 if (should_resched()) {
5032 EXPORT_SYMBOL(_cond_resched);
5035 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5036 * call schedule, and on return reacquire the lock.
5038 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5039 * operations here to prevent schedule() from being called twice (once via
5040 * spin_unlock(), once by hand).
5042 int __cond_resched_lock(spinlock_t *lock)
5044 int resched = should_resched();
5047 lockdep_assert_held(lock);
5049 if (spin_needbreak(lock) || resched) {
5060 EXPORT_SYMBOL(__cond_resched_lock);
5062 int __sched __cond_resched_softirq(void)
5064 BUG_ON(!in_softirq());
5066 if (should_resched()) {
5074 EXPORT_SYMBOL(__cond_resched_softirq);
5077 * yield - yield the current processor to other threads.
5079 * This is a shortcut for kernel-space yielding - it marks the
5080 * thread runnable and calls sys_sched_yield().
5082 void __sched yield(void)
5084 set_current_state(TASK_RUNNING);
5087 EXPORT_SYMBOL(yield);
5090 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5091 * that process accounting knows that this is a task in IO wait state.
5093 void __sched io_schedule(void)
5095 struct rq *rq = raw_rq();
5097 delayacct_blkio_start();
5098 atomic_inc(&rq->nr_iowait);
5099 current->in_iowait = 1;
5101 current->in_iowait = 0;
5102 atomic_dec(&rq->nr_iowait);
5103 delayacct_blkio_end();
5105 EXPORT_SYMBOL(io_schedule);
5107 long __sched io_schedule_timeout(long timeout)
5109 struct rq *rq = raw_rq();
5112 delayacct_blkio_start();
5113 atomic_inc(&rq->nr_iowait);
5114 current->in_iowait = 1;
5115 ret = schedule_timeout(timeout);
5116 current->in_iowait = 0;
5117 atomic_dec(&rq->nr_iowait);
5118 delayacct_blkio_end();
5123 * sys_sched_get_priority_max - return maximum RT priority.
5124 * @policy: scheduling class.
5126 * this syscall returns the maximum rt_priority that can be used
5127 * by a given scheduling class.
5129 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5136 ret = MAX_USER_RT_PRIO-1;
5148 * sys_sched_get_priority_min - return minimum RT priority.
5149 * @policy: scheduling class.
5151 * this syscall returns the minimum rt_priority that can be used
5152 * by a given scheduling class.
5154 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5172 * sys_sched_rr_get_interval - return the default timeslice of a process.
5173 * @pid: pid of the process.
5174 * @interval: userspace pointer to the timeslice value.
5176 * this syscall writes the default timeslice value of a given process
5177 * into the user-space timespec buffer. A value of '0' means infinity.
5179 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5180 struct timespec __user *, interval)
5182 struct task_struct *p;
5183 unsigned int time_slice;
5184 unsigned long flags;
5194 p = find_process_by_pid(pid);
5198 retval = security_task_getscheduler(p);
5202 rq = task_rq_lock(p, &flags);
5203 time_slice = p->sched_class->get_rr_interval(rq, p);
5204 task_rq_unlock(rq, &flags);
5207 jiffies_to_timespec(time_slice, &t);
5208 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5216 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5218 void sched_show_task(struct task_struct *p)
5220 unsigned long free = 0;
5223 state = p->state ? __ffs(p->state) + 1 : 0;
5224 printk(KERN_INFO "%-13.13s %c", p->comm,
5225 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5226 #if BITS_PER_LONG == 32
5227 if (state == TASK_RUNNING)
5228 printk(KERN_CONT " running ");
5230 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5232 if (state == TASK_RUNNING)
5233 printk(KERN_CONT " running task ");
5235 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5237 #ifdef CONFIG_DEBUG_STACK_USAGE
5238 free = stack_not_used(p);
5240 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5241 task_pid_nr(p), task_pid_nr(p->real_parent),
5242 (unsigned long)task_thread_info(p)->flags);
5244 show_stack(p, NULL);
5247 void show_state_filter(unsigned long state_filter)
5249 struct task_struct *g, *p;
5251 #if BITS_PER_LONG == 32
5253 " task PC stack pid father\n");
5256 " task PC stack pid father\n");
5258 read_lock(&tasklist_lock);
5259 do_each_thread(g, p) {
5261 * reset the NMI-timeout, listing all files on a slow
5262 * console might take alot of time:
5264 touch_nmi_watchdog();
5265 if (!state_filter || (p->state & state_filter))
5267 } while_each_thread(g, p);
5269 touch_all_softlockup_watchdogs();
5271 #ifdef CONFIG_SCHED_DEBUG
5272 sysrq_sched_debug_show();
5274 read_unlock(&tasklist_lock);
5276 * Only show locks if all tasks are dumped:
5279 debug_show_all_locks();
5282 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5284 idle->sched_class = &idle_sched_class;
5288 * init_idle - set up an idle thread for a given CPU
5289 * @idle: task in question
5290 * @cpu: cpu the idle task belongs to
5292 * NOTE: this function does not set the idle thread's NEED_RESCHED
5293 * flag, to make booting more robust.
5295 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5297 struct rq *rq = cpu_rq(cpu);
5298 unsigned long flags;
5300 raw_spin_lock_irqsave(&rq->lock, flags);
5303 idle->state = TASK_RUNNING;
5304 idle->se.exec_start = sched_clock();
5306 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5307 __set_task_cpu(idle, cpu);
5309 rq->curr = rq->idle = idle;
5310 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5313 raw_spin_unlock_irqrestore(&rq->lock, flags);
5315 /* Set the preempt count _outside_ the spinlocks! */
5316 #if defined(CONFIG_PREEMPT)
5317 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5319 task_thread_info(idle)->preempt_count = 0;
5322 * The idle tasks have their own, simple scheduling class:
5324 idle->sched_class = &idle_sched_class;
5325 ftrace_graph_init_task(idle);
5329 * In a system that switches off the HZ timer nohz_cpu_mask
5330 * indicates which cpus entered this state. This is used
5331 * in the rcu update to wait only for active cpus. For system
5332 * which do not switch off the HZ timer nohz_cpu_mask should
5333 * always be CPU_BITS_NONE.
5335 cpumask_var_t nohz_cpu_mask;
5338 * Increase the granularity value when there are more CPUs,
5339 * because with more CPUs the 'effective latency' as visible
5340 * to users decreases. But the relationship is not linear,
5341 * so pick a second-best guess by going with the log2 of the
5344 * This idea comes from the SD scheduler of Con Kolivas:
5346 static int get_update_sysctl_factor(void)
5348 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5349 unsigned int factor;
5351 switch (sysctl_sched_tunable_scaling) {
5352 case SCHED_TUNABLESCALING_NONE:
5355 case SCHED_TUNABLESCALING_LINEAR:
5358 case SCHED_TUNABLESCALING_LOG:
5360 factor = 1 + ilog2(cpus);
5367 static void update_sysctl(void)
5369 unsigned int factor = get_update_sysctl_factor();
5371 #define SET_SYSCTL(name) \
5372 (sysctl_##name = (factor) * normalized_sysctl_##name)
5373 SET_SYSCTL(sched_min_granularity);
5374 SET_SYSCTL(sched_latency);
5375 SET_SYSCTL(sched_wakeup_granularity);
5376 SET_SYSCTL(sched_shares_ratelimit);
5380 static inline void sched_init_granularity(void)
5387 * This is how migration works:
5389 * 1) we invoke migration_cpu_stop() on the target CPU using
5391 * 2) stopper starts to run (implicitly forcing the migrated thread
5393 * 3) it checks whether the migrated task is still in the wrong runqueue.
5394 * 4) if it's in the wrong runqueue then the migration thread removes
5395 * it and puts it into the right queue.
5396 * 5) stopper completes and stop_one_cpu() returns and the migration
5401 * Change a given task's CPU affinity. Migrate the thread to a
5402 * proper CPU and schedule it away if the CPU it's executing on
5403 * is removed from the allowed bitmask.
5405 * NOTE: the caller must have a valid reference to the task, the
5406 * task must not exit() & deallocate itself prematurely. The
5407 * call is not atomic; no spinlocks may be held.
5409 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5411 unsigned long flags;
5413 unsigned int dest_cpu;
5417 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5418 * drop the rq->lock and still rely on ->cpus_allowed.
5421 while (task_is_waking(p))
5423 rq = task_rq_lock(p, &flags);
5424 if (task_is_waking(p)) {
5425 task_rq_unlock(rq, &flags);
5429 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5434 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5435 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5440 if (p->sched_class->set_cpus_allowed)
5441 p->sched_class->set_cpus_allowed(p, new_mask);
5443 cpumask_copy(&p->cpus_allowed, new_mask);
5444 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5447 /* Can the task run on the task's current CPU? If so, we're done */
5448 if (cpumask_test_cpu(task_cpu(p), new_mask))
5451 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5452 if (migrate_task(p, dest_cpu)) {
5453 struct migration_arg arg = { p, dest_cpu };
5454 /* Need help from migration thread: drop lock and wait. */
5455 task_rq_unlock(rq, &flags);
5456 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5457 tlb_migrate_finish(p->mm);
5461 task_rq_unlock(rq, &flags);
5465 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5468 * Move (not current) task off this cpu, onto dest cpu. We're doing
5469 * this because either it can't run here any more (set_cpus_allowed()
5470 * away from this CPU, or CPU going down), or because we're
5471 * attempting to rebalance this task on exec (sched_exec).
5473 * So we race with normal scheduler movements, but that's OK, as long
5474 * as the task is no longer on this CPU.
5476 * Returns non-zero if task was successfully migrated.
5478 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5480 struct rq *rq_dest, *rq_src;
5483 if (unlikely(!cpu_active(dest_cpu)))
5486 rq_src = cpu_rq(src_cpu);
5487 rq_dest = cpu_rq(dest_cpu);
5489 double_rq_lock(rq_src, rq_dest);
5490 /* Already moved. */
5491 if (task_cpu(p) != src_cpu)
5493 /* Affinity changed (again). */
5494 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5498 * If we're not on a rq, the next wake-up will ensure we're
5502 deactivate_task(rq_src, p, 0);
5503 set_task_cpu(p, dest_cpu);
5504 activate_task(rq_dest, p, 0);
5505 check_preempt_curr(rq_dest, p, 0);
5510 double_rq_unlock(rq_src, rq_dest);
5515 * migration_cpu_stop - this will be executed by a highprio stopper thread
5516 * and performs thread migration by bumping thread off CPU then
5517 * 'pushing' onto another runqueue.
5519 static int migration_cpu_stop(void *data)
5521 struct migration_arg *arg = data;
5524 * The original target cpu might have gone down and we might
5525 * be on another cpu but it doesn't matter.
5527 local_irq_disable();
5528 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5533 #ifdef CONFIG_HOTPLUG_CPU
5535 * Figure out where task on dead CPU should go, use force if necessary.
5537 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5539 struct rq *rq = cpu_rq(dead_cpu);
5540 int needs_cpu, uninitialized_var(dest_cpu);
5541 unsigned long flags;
5543 local_irq_save(flags);
5545 raw_spin_lock(&rq->lock);
5546 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5548 dest_cpu = select_fallback_rq(dead_cpu, p);
5549 raw_spin_unlock(&rq->lock);
5551 * It can only fail if we race with set_cpus_allowed(),
5552 * in the racer should migrate the task anyway.
5555 __migrate_task(p, dead_cpu, dest_cpu);
5556 local_irq_restore(flags);
5560 * While a dead CPU has no uninterruptible tasks queued at this point,
5561 * it might still have a nonzero ->nr_uninterruptible counter, because
5562 * for performance reasons the counter is not stricly tracking tasks to
5563 * their home CPUs. So we just add the counter to another CPU's counter,
5564 * to keep the global sum constant after CPU-down:
5566 static void migrate_nr_uninterruptible(struct rq *rq_src)
5568 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5569 unsigned long flags;
5571 local_irq_save(flags);
5572 double_rq_lock(rq_src, rq_dest);
5573 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5574 rq_src->nr_uninterruptible = 0;
5575 double_rq_unlock(rq_src, rq_dest);
5576 local_irq_restore(flags);
5579 /* Run through task list and migrate tasks from the dead cpu. */
5580 static void migrate_live_tasks(int src_cpu)
5582 struct task_struct *p, *t;
5584 read_lock(&tasklist_lock);
5586 do_each_thread(t, p) {
5590 if (task_cpu(p) == src_cpu)
5591 move_task_off_dead_cpu(src_cpu, p);
5592 } while_each_thread(t, p);
5594 read_unlock(&tasklist_lock);
5598 * Schedules idle task to be the next runnable task on current CPU.
5599 * It does so by boosting its priority to highest possible.
5600 * Used by CPU offline code.
5602 void sched_idle_next(void)
5604 int this_cpu = smp_processor_id();
5605 struct rq *rq = cpu_rq(this_cpu);
5606 struct task_struct *p = rq->idle;
5607 unsigned long flags;
5609 /* cpu has to be offline */
5610 BUG_ON(cpu_online(this_cpu));
5613 * Strictly not necessary since rest of the CPUs are stopped by now
5614 * and interrupts disabled on the current cpu.
5616 raw_spin_lock_irqsave(&rq->lock, flags);
5618 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5620 activate_task(rq, p, 0);
5622 raw_spin_unlock_irqrestore(&rq->lock, flags);
5626 * Ensures that the idle task is using init_mm right before its cpu goes
5629 void idle_task_exit(void)
5631 struct mm_struct *mm = current->active_mm;
5633 BUG_ON(cpu_online(smp_processor_id()));
5636 switch_mm(mm, &init_mm, current);
5640 /* called under rq->lock with disabled interrupts */
5641 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5643 struct rq *rq = cpu_rq(dead_cpu);
5645 /* Must be exiting, otherwise would be on tasklist. */
5646 BUG_ON(!p->exit_state);
5648 /* Cannot have done final schedule yet: would have vanished. */
5649 BUG_ON(p->state == TASK_DEAD);
5654 * Drop lock around migration; if someone else moves it,
5655 * that's OK. No task can be added to this CPU, so iteration is
5658 raw_spin_unlock_irq(&rq->lock);
5659 move_task_off_dead_cpu(dead_cpu, p);
5660 raw_spin_lock_irq(&rq->lock);
5665 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5666 static void migrate_dead_tasks(unsigned int dead_cpu)
5668 struct rq *rq = cpu_rq(dead_cpu);
5669 struct task_struct *next;
5672 if (!rq->nr_running)
5674 next = pick_next_task(rq);
5677 next->sched_class->put_prev_task(rq, next);
5678 migrate_dead(dead_cpu, next);
5684 * remove the tasks which were accounted by rq from calc_load_tasks.
5686 static void calc_global_load_remove(struct rq *rq)
5688 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5689 rq->calc_load_active = 0;
5691 #endif /* CONFIG_HOTPLUG_CPU */
5693 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5695 static struct ctl_table sd_ctl_dir[] = {
5697 .procname = "sched_domain",
5703 static struct ctl_table sd_ctl_root[] = {
5705 .procname = "kernel",
5707 .child = sd_ctl_dir,
5712 static struct ctl_table *sd_alloc_ctl_entry(int n)
5714 struct ctl_table *entry =
5715 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5720 static void sd_free_ctl_entry(struct ctl_table **tablep)
5722 struct ctl_table *entry;
5725 * In the intermediate directories, both the child directory and
5726 * procname are dynamically allocated and could fail but the mode
5727 * will always be set. In the lowest directory the names are
5728 * static strings and all have proc handlers.
5730 for (entry = *tablep; entry->mode; entry++) {
5732 sd_free_ctl_entry(&entry->child);
5733 if (entry->proc_handler == NULL)
5734 kfree(entry->procname);
5742 set_table_entry(struct ctl_table *entry,
5743 const char *procname, void *data, int maxlen,
5744 mode_t mode, proc_handler *proc_handler)
5746 entry->procname = procname;
5748 entry->maxlen = maxlen;
5750 entry->proc_handler = proc_handler;
5753 static struct ctl_table *
5754 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5756 struct ctl_table *table = sd_alloc_ctl_entry(13);
5761 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5762 sizeof(long), 0644, proc_doulongvec_minmax);
5763 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5764 sizeof(long), 0644, proc_doulongvec_minmax);
5765 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5766 sizeof(int), 0644, proc_dointvec_minmax);
5767 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5768 sizeof(int), 0644, proc_dointvec_minmax);
5769 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5770 sizeof(int), 0644, proc_dointvec_minmax);
5771 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5772 sizeof(int), 0644, proc_dointvec_minmax);
5773 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5774 sizeof(int), 0644, proc_dointvec_minmax);
5775 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5776 sizeof(int), 0644, proc_dointvec_minmax);
5777 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5778 sizeof(int), 0644, proc_dointvec_minmax);
5779 set_table_entry(&table[9], "cache_nice_tries",
5780 &sd->cache_nice_tries,
5781 sizeof(int), 0644, proc_dointvec_minmax);
5782 set_table_entry(&table[10], "flags", &sd->flags,
5783 sizeof(int), 0644, proc_dointvec_minmax);
5784 set_table_entry(&table[11], "name", sd->name,
5785 CORENAME_MAX_SIZE, 0444, proc_dostring);
5786 /* &table[12] is terminator */
5791 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5793 struct ctl_table *entry, *table;
5794 struct sched_domain *sd;
5795 int domain_num = 0, i;
5798 for_each_domain(cpu, sd)
5800 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5805 for_each_domain(cpu, sd) {
5806 snprintf(buf, 32, "domain%d", i);
5807 entry->procname = kstrdup(buf, GFP_KERNEL);
5809 entry->child = sd_alloc_ctl_domain_table(sd);
5816 static struct ctl_table_header *sd_sysctl_header;
5817 static void register_sched_domain_sysctl(void)
5819 int i, cpu_num = num_possible_cpus();
5820 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5823 WARN_ON(sd_ctl_dir[0].child);
5824 sd_ctl_dir[0].child = entry;
5829 for_each_possible_cpu(i) {
5830 snprintf(buf, 32, "cpu%d", i);
5831 entry->procname = kstrdup(buf, GFP_KERNEL);
5833 entry->child = sd_alloc_ctl_cpu_table(i);
5837 WARN_ON(sd_sysctl_header);
5838 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5841 /* may be called multiple times per register */
5842 static void unregister_sched_domain_sysctl(void)
5844 if (sd_sysctl_header)
5845 unregister_sysctl_table(sd_sysctl_header);
5846 sd_sysctl_header = NULL;
5847 if (sd_ctl_dir[0].child)
5848 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5851 static void register_sched_domain_sysctl(void)
5854 static void unregister_sched_domain_sysctl(void)
5859 static void set_rq_online(struct rq *rq)
5862 const struct sched_class *class;
5864 cpumask_set_cpu(rq->cpu, rq->rd->online);
5867 for_each_class(class) {
5868 if (class->rq_online)
5869 class->rq_online(rq);
5874 static void set_rq_offline(struct rq *rq)
5877 const struct sched_class *class;
5879 for_each_class(class) {
5880 if (class->rq_offline)
5881 class->rq_offline(rq);
5884 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5890 * migration_call - callback that gets triggered when a CPU is added.
5891 * Here we can start up the necessary migration thread for the new CPU.
5893 static int __cpuinit
5894 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5896 int cpu = (long)hcpu;
5897 unsigned long flags;
5898 struct rq *rq = cpu_rq(cpu);
5902 case CPU_UP_PREPARE:
5903 case CPU_UP_PREPARE_FROZEN:
5904 rq->calc_load_update = calc_load_update;
5908 case CPU_ONLINE_FROZEN:
5909 /* Update our root-domain */
5910 raw_spin_lock_irqsave(&rq->lock, flags);
5912 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5916 raw_spin_unlock_irqrestore(&rq->lock, flags);
5919 #ifdef CONFIG_HOTPLUG_CPU
5921 case CPU_DEAD_FROZEN:
5922 migrate_live_tasks(cpu);
5923 /* Idle task back to normal (off runqueue, low prio) */
5924 raw_spin_lock_irq(&rq->lock);
5925 deactivate_task(rq, rq->idle, 0);
5926 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5927 rq->idle->sched_class = &idle_sched_class;
5928 migrate_dead_tasks(cpu);
5929 raw_spin_unlock_irq(&rq->lock);
5930 migrate_nr_uninterruptible(rq);
5931 BUG_ON(rq->nr_running != 0);
5932 calc_global_load_remove(rq);
5936 case CPU_DYING_FROZEN:
5937 /* Update our root-domain */
5938 raw_spin_lock_irqsave(&rq->lock, flags);
5940 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5943 raw_spin_unlock_irqrestore(&rq->lock, flags);
5951 * Register at high priority so that task migration (migrate_all_tasks)
5952 * happens before everything else. This has to be lower priority than
5953 * the notifier in the perf_event subsystem, though.
5955 static struct notifier_block __cpuinitdata migration_notifier = {
5956 .notifier_call = migration_call,
5957 .priority = CPU_PRI_MIGRATION,
5960 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
5961 unsigned long action, void *hcpu)
5963 switch (action & ~CPU_TASKS_FROZEN) {
5965 case CPU_DOWN_FAILED:
5966 set_cpu_active((long)hcpu, true);
5973 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
5974 unsigned long action, void *hcpu)
5976 switch (action & ~CPU_TASKS_FROZEN) {
5977 case CPU_DOWN_PREPARE:
5978 set_cpu_active((long)hcpu, false);
5985 static int __init migration_init(void)
5987 void *cpu = (void *)(long)smp_processor_id();
5990 /* Initialize migration for the boot CPU */
5991 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
5992 BUG_ON(err == NOTIFY_BAD);
5993 migration_call(&migration_notifier, CPU_ONLINE, cpu);
5994 register_cpu_notifier(&migration_notifier);
5996 /* Register cpu active notifiers */
5997 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
5998 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6002 early_initcall(migration_init);
6007 #ifdef CONFIG_SCHED_DEBUG
6009 static __read_mostly int sched_domain_debug_enabled;
6011 static int __init sched_domain_debug_setup(char *str)
6013 sched_domain_debug_enabled = 1;
6017 early_param("sched_debug", sched_domain_debug_setup);
6019 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6020 struct cpumask *groupmask)
6022 struct sched_group *group = sd->groups;
6025 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6026 cpumask_clear(groupmask);
6028 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6030 if (!(sd->flags & SD_LOAD_BALANCE)) {
6031 printk("does not load-balance\n");
6033 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6038 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6040 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6041 printk(KERN_ERR "ERROR: domain->span does not contain "
6044 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6045 printk(KERN_ERR "ERROR: domain->groups does not contain"
6049 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6053 printk(KERN_ERR "ERROR: group is NULL\n");
6057 if (!group->cpu_power) {
6058 printk(KERN_CONT "\n");
6059 printk(KERN_ERR "ERROR: domain->cpu_power not "
6064 if (!cpumask_weight(sched_group_cpus(group))) {
6065 printk(KERN_CONT "\n");
6066 printk(KERN_ERR "ERROR: empty group\n");
6070 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6071 printk(KERN_CONT "\n");
6072 printk(KERN_ERR "ERROR: repeated CPUs\n");
6076 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6078 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6080 printk(KERN_CONT " %s", str);
6081 if (group->cpu_power != SCHED_LOAD_SCALE) {
6082 printk(KERN_CONT " (cpu_power = %d)",
6086 group = group->next;
6087 } while (group != sd->groups);
6088 printk(KERN_CONT "\n");
6090 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6091 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6094 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6095 printk(KERN_ERR "ERROR: parent span is not a superset "
6096 "of domain->span\n");
6100 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6102 cpumask_var_t groupmask;
6105 if (!sched_domain_debug_enabled)
6109 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6113 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6115 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6116 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6121 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6128 free_cpumask_var(groupmask);
6130 #else /* !CONFIG_SCHED_DEBUG */
6131 # define sched_domain_debug(sd, cpu) do { } while (0)
6132 #endif /* CONFIG_SCHED_DEBUG */
6134 static int sd_degenerate(struct sched_domain *sd)
6136 if (cpumask_weight(sched_domain_span(sd)) == 1)
6139 /* Following flags need at least 2 groups */
6140 if (sd->flags & (SD_LOAD_BALANCE |
6141 SD_BALANCE_NEWIDLE |
6145 SD_SHARE_PKG_RESOURCES)) {
6146 if (sd->groups != sd->groups->next)
6150 /* Following flags don't use groups */
6151 if (sd->flags & (SD_WAKE_AFFINE))
6158 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6160 unsigned long cflags = sd->flags, pflags = parent->flags;
6162 if (sd_degenerate(parent))
6165 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6168 /* Flags needing groups don't count if only 1 group in parent */
6169 if (parent->groups == parent->groups->next) {
6170 pflags &= ~(SD_LOAD_BALANCE |
6171 SD_BALANCE_NEWIDLE |
6175 SD_SHARE_PKG_RESOURCES);
6176 if (nr_node_ids == 1)
6177 pflags &= ~SD_SERIALIZE;
6179 if (~cflags & pflags)
6185 static void free_rootdomain(struct root_domain *rd)
6187 synchronize_sched();
6189 cpupri_cleanup(&rd->cpupri);
6191 free_cpumask_var(rd->rto_mask);
6192 free_cpumask_var(rd->online);
6193 free_cpumask_var(rd->span);
6197 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6199 struct root_domain *old_rd = NULL;
6200 unsigned long flags;
6202 raw_spin_lock_irqsave(&rq->lock, flags);
6207 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6210 cpumask_clear_cpu(rq->cpu, old_rd->span);
6213 * If we dont want to free the old_rt yet then
6214 * set old_rd to NULL to skip the freeing later
6217 if (!atomic_dec_and_test(&old_rd->refcount))
6221 atomic_inc(&rd->refcount);
6224 cpumask_set_cpu(rq->cpu, rd->span);
6225 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6228 raw_spin_unlock_irqrestore(&rq->lock, flags);
6231 free_rootdomain(old_rd);
6234 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6236 gfp_t gfp = GFP_KERNEL;
6238 memset(rd, 0, sizeof(*rd));
6243 if (!alloc_cpumask_var(&rd->span, gfp))
6245 if (!alloc_cpumask_var(&rd->online, gfp))
6247 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
6250 if (cpupri_init(&rd->cpupri, bootmem) != 0)
6255 free_cpumask_var(rd->rto_mask);
6257 free_cpumask_var(rd->online);
6259 free_cpumask_var(rd->span);
6264 static void init_defrootdomain(void)
6266 init_rootdomain(&def_root_domain, true);
6268 atomic_set(&def_root_domain.refcount, 1);
6271 static struct root_domain *alloc_rootdomain(void)
6273 struct root_domain *rd;
6275 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6279 if (init_rootdomain(rd, false) != 0) {
6288 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6289 * hold the hotplug lock.
6292 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6294 struct rq *rq = cpu_rq(cpu);
6295 struct sched_domain *tmp;
6297 for (tmp = sd; tmp; tmp = tmp->parent)
6298 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6300 /* Remove the sched domains which do not contribute to scheduling. */
6301 for (tmp = sd; tmp; ) {
6302 struct sched_domain *parent = tmp->parent;
6306 if (sd_parent_degenerate(tmp, parent)) {
6307 tmp->parent = parent->parent;
6309 parent->parent->child = tmp;
6314 if (sd && sd_degenerate(sd)) {
6320 sched_domain_debug(sd, cpu);
6322 rq_attach_root(rq, rd);
6323 rcu_assign_pointer(rq->sd, sd);
6326 /* cpus with isolated domains */
6327 static cpumask_var_t cpu_isolated_map;
6329 /* Setup the mask of cpus configured for isolated domains */
6330 static int __init isolated_cpu_setup(char *str)
6332 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6333 cpulist_parse(str, cpu_isolated_map);
6337 __setup("isolcpus=", isolated_cpu_setup);
6340 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6341 * to a function which identifies what group(along with sched group) a CPU
6342 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6343 * (due to the fact that we keep track of groups covered with a struct cpumask).
6345 * init_sched_build_groups will build a circular linked list of the groups
6346 * covered by the given span, and will set each group's ->cpumask correctly,
6347 * and ->cpu_power to 0.
6350 init_sched_build_groups(const struct cpumask *span,
6351 const struct cpumask *cpu_map,
6352 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6353 struct sched_group **sg,
6354 struct cpumask *tmpmask),
6355 struct cpumask *covered, struct cpumask *tmpmask)
6357 struct sched_group *first = NULL, *last = NULL;
6360 cpumask_clear(covered);
6362 for_each_cpu(i, span) {
6363 struct sched_group *sg;
6364 int group = group_fn(i, cpu_map, &sg, tmpmask);
6367 if (cpumask_test_cpu(i, covered))
6370 cpumask_clear(sched_group_cpus(sg));
6373 for_each_cpu(j, span) {
6374 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6377 cpumask_set_cpu(j, covered);
6378 cpumask_set_cpu(j, sched_group_cpus(sg));
6389 #define SD_NODES_PER_DOMAIN 16
6394 * find_next_best_node - find the next node to include in a sched_domain
6395 * @node: node whose sched_domain we're building
6396 * @used_nodes: nodes already in the sched_domain
6398 * Find the next node to include in a given scheduling domain. Simply
6399 * finds the closest node not already in the @used_nodes map.
6401 * Should use nodemask_t.
6403 static int find_next_best_node(int node, nodemask_t *used_nodes)
6405 int i, n, val, min_val, best_node = 0;
6409 for (i = 0; i < nr_node_ids; i++) {
6410 /* Start at @node */
6411 n = (node + i) % nr_node_ids;
6413 if (!nr_cpus_node(n))
6416 /* Skip already used nodes */
6417 if (node_isset(n, *used_nodes))
6420 /* Simple min distance search */
6421 val = node_distance(node, n);
6423 if (val < min_val) {
6429 node_set(best_node, *used_nodes);
6434 * sched_domain_node_span - get a cpumask for a node's sched_domain
6435 * @node: node whose cpumask we're constructing
6436 * @span: resulting cpumask
6438 * Given a node, construct a good cpumask for its sched_domain to span. It
6439 * should be one that prevents unnecessary balancing, but also spreads tasks
6442 static void sched_domain_node_span(int node, struct cpumask *span)
6444 nodemask_t used_nodes;
6447 cpumask_clear(span);
6448 nodes_clear(used_nodes);
6450 cpumask_or(span, span, cpumask_of_node(node));
6451 node_set(node, used_nodes);
6453 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6454 int next_node = find_next_best_node(node, &used_nodes);
6456 cpumask_or(span, span, cpumask_of_node(next_node));
6459 #endif /* CONFIG_NUMA */
6461 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6464 * The cpus mask in sched_group and sched_domain hangs off the end.
6466 * ( See the the comments in include/linux/sched.h:struct sched_group
6467 * and struct sched_domain. )
6469 struct static_sched_group {
6470 struct sched_group sg;
6471 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6474 struct static_sched_domain {
6475 struct sched_domain sd;
6476 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6482 cpumask_var_t domainspan;
6483 cpumask_var_t covered;
6484 cpumask_var_t notcovered;
6486 cpumask_var_t nodemask;
6487 cpumask_var_t this_sibling_map;
6488 cpumask_var_t this_core_map;
6489 cpumask_var_t send_covered;
6490 cpumask_var_t tmpmask;
6491 struct sched_group **sched_group_nodes;
6492 struct root_domain *rd;
6496 sa_sched_groups = 0,
6501 sa_this_sibling_map,
6503 sa_sched_group_nodes,
6513 * SMT sched-domains:
6515 #ifdef CONFIG_SCHED_SMT
6516 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6517 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6520 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6521 struct sched_group **sg, struct cpumask *unused)
6524 *sg = &per_cpu(sched_groups, cpu).sg;
6527 #endif /* CONFIG_SCHED_SMT */
6530 * multi-core sched-domains:
6532 #ifdef CONFIG_SCHED_MC
6533 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6534 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6535 #endif /* CONFIG_SCHED_MC */
6537 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6539 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6540 struct sched_group **sg, struct cpumask *mask)
6544 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6545 group = cpumask_first(mask);
6547 *sg = &per_cpu(sched_group_core, group).sg;
6550 #elif defined(CONFIG_SCHED_MC)
6552 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6553 struct sched_group **sg, struct cpumask *unused)
6556 *sg = &per_cpu(sched_group_core, cpu).sg;
6561 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6562 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6565 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6566 struct sched_group **sg, struct cpumask *mask)
6569 #ifdef CONFIG_SCHED_MC
6570 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6571 group = cpumask_first(mask);
6572 #elif defined(CONFIG_SCHED_SMT)
6573 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6574 group = cpumask_first(mask);
6579 *sg = &per_cpu(sched_group_phys, group).sg;
6585 * The init_sched_build_groups can't handle what we want to do with node
6586 * groups, so roll our own. Now each node has its own list of groups which
6587 * gets dynamically allocated.
6589 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6590 static struct sched_group ***sched_group_nodes_bycpu;
6592 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6593 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6595 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6596 struct sched_group **sg,
6597 struct cpumask *nodemask)
6601 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6602 group = cpumask_first(nodemask);
6605 *sg = &per_cpu(sched_group_allnodes, group).sg;
6609 static void init_numa_sched_groups_power(struct sched_group *group_head)
6611 struct sched_group *sg = group_head;
6617 for_each_cpu(j, sched_group_cpus(sg)) {
6618 struct sched_domain *sd;
6620 sd = &per_cpu(phys_domains, j).sd;
6621 if (j != group_first_cpu(sd->groups)) {
6623 * Only add "power" once for each
6629 sg->cpu_power += sd->groups->cpu_power;
6632 } while (sg != group_head);
6635 static int build_numa_sched_groups(struct s_data *d,
6636 const struct cpumask *cpu_map, int num)
6638 struct sched_domain *sd;
6639 struct sched_group *sg, *prev;
6642 cpumask_clear(d->covered);
6643 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6644 if (cpumask_empty(d->nodemask)) {
6645 d->sched_group_nodes[num] = NULL;
6649 sched_domain_node_span(num, d->domainspan);
6650 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6652 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6655 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6659 d->sched_group_nodes[num] = sg;
6661 for_each_cpu(j, d->nodemask) {
6662 sd = &per_cpu(node_domains, j).sd;
6667 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6669 cpumask_or(d->covered, d->covered, d->nodemask);
6672 for (j = 0; j < nr_node_ids; j++) {
6673 n = (num + j) % nr_node_ids;
6674 cpumask_complement(d->notcovered, d->covered);
6675 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6676 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6677 if (cpumask_empty(d->tmpmask))
6679 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6680 if (cpumask_empty(d->tmpmask))
6682 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6686 "Can not alloc domain group for node %d\n", j);
6690 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6691 sg->next = prev->next;
6692 cpumask_or(d->covered, d->covered, d->tmpmask);
6699 #endif /* CONFIG_NUMA */
6702 /* Free memory allocated for various sched_group structures */
6703 static void free_sched_groups(const struct cpumask *cpu_map,
6704 struct cpumask *nodemask)
6708 for_each_cpu(cpu, cpu_map) {
6709 struct sched_group **sched_group_nodes
6710 = sched_group_nodes_bycpu[cpu];
6712 if (!sched_group_nodes)
6715 for (i = 0; i < nr_node_ids; i++) {
6716 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6718 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6719 if (cpumask_empty(nodemask))
6729 if (oldsg != sched_group_nodes[i])
6732 kfree(sched_group_nodes);
6733 sched_group_nodes_bycpu[cpu] = NULL;
6736 #else /* !CONFIG_NUMA */
6737 static void free_sched_groups(const struct cpumask *cpu_map,
6738 struct cpumask *nodemask)
6741 #endif /* CONFIG_NUMA */
6744 * Initialize sched groups cpu_power.
6746 * cpu_power indicates the capacity of sched group, which is used while
6747 * distributing the load between different sched groups in a sched domain.
6748 * Typically cpu_power for all the groups in a sched domain will be same unless
6749 * there are asymmetries in the topology. If there are asymmetries, group
6750 * having more cpu_power will pickup more load compared to the group having
6753 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6755 struct sched_domain *child;
6756 struct sched_group *group;
6760 WARN_ON(!sd || !sd->groups);
6762 if (cpu != group_first_cpu(sd->groups))
6767 sd->groups->cpu_power = 0;
6770 power = SCHED_LOAD_SCALE;
6771 weight = cpumask_weight(sched_domain_span(sd));
6773 * SMT siblings share the power of a single core.
6774 * Usually multiple threads get a better yield out of
6775 * that one core than a single thread would have,
6776 * reflect that in sd->smt_gain.
6778 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6779 power *= sd->smt_gain;
6781 power >>= SCHED_LOAD_SHIFT;
6783 sd->groups->cpu_power += power;
6788 * Add cpu_power of each child group to this groups cpu_power.
6790 group = child->groups;
6792 sd->groups->cpu_power += group->cpu_power;
6793 group = group->next;
6794 } while (group != child->groups);
6798 * Initializers for schedule domains
6799 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6802 #ifdef CONFIG_SCHED_DEBUG
6803 # define SD_INIT_NAME(sd, type) sd->name = #type
6805 # define SD_INIT_NAME(sd, type) do { } while (0)
6808 #define SD_INIT(sd, type) sd_init_##type(sd)
6810 #define SD_INIT_FUNC(type) \
6811 static noinline void sd_init_##type(struct sched_domain *sd) \
6813 memset(sd, 0, sizeof(*sd)); \
6814 *sd = SD_##type##_INIT; \
6815 sd->level = SD_LV_##type; \
6816 SD_INIT_NAME(sd, type); \
6821 SD_INIT_FUNC(ALLNODES)
6824 #ifdef CONFIG_SCHED_SMT
6825 SD_INIT_FUNC(SIBLING)
6827 #ifdef CONFIG_SCHED_MC
6831 static int default_relax_domain_level = -1;
6833 static int __init setup_relax_domain_level(char *str)
6837 val = simple_strtoul(str, NULL, 0);
6838 if (val < SD_LV_MAX)
6839 default_relax_domain_level = val;
6843 __setup("relax_domain_level=", setup_relax_domain_level);
6845 static void set_domain_attribute(struct sched_domain *sd,
6846 struct sched_domain_attr *attr)
6850 if (!attr || attr->relax_domain_level < 0) {
6851 if (default_relax_domain_level < 0)
6854 request = default_relax_domain_level;
6856 request = attr->relax_domain_level;
6857 if (request < sd->level) {
6858 /* turn off idle balance on this domain */
6859 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6861 /* turn on idle balance on this domain */
6862 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6866 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6867 const struct cpumask *cpu_map)
6870 case sa_sched_groups:
6871 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6872 d->sched_group_nodes = NULL;
6874 free_rootdomain(d->rd); /* fall through */
6876 free_cpumask_var(d->tmpmask); /* fall through */
6877 case sa_send_covered:
6878 free_cpumask_var(d->send_covered); /* fall through */
6879 case sa_this_core_map:
6880 free_cpumask_var(d->this_core_map); /* fall through */
6881 case sa_this_sibling_map:
6882 free_cpumask_var(d->this_sibling_map); /* fall through */
6884 free_cpumask_var(d->nodemask); /* fall through */
6885 case sa_sched_group_nodes:
6887 kfree(d->sched_group_nodes); /* fall through */
6889 free_cpumask_var(d->notcovered); /* fall through */
6891 free_cpumask_var(d->covered); /* fall through */
6893 free_cpumask_var(d->domainspan); /* fall through */
6900 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6901 const struct cpumask *cpu_map)
6904 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6906 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6907 return sa_domainspan;
6908 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6910 /* Allocate the per-node list of sched groups */
6911 d->sched_group_nodes = kcalloc(nr_node_ids,
6912 sizeof(struct sched_group *), GFP_KERNEL);
6913 if (!d->sched_group_nodes) {
6914 printk(KERN_WARNING "Can not alloc sched group node list\n");
6915 return sa_notcovered;
6917 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6919 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6920 return sa_sched_group_nodes;
6921 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6923 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6924 return sa_this_sibling_map;
6925 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6926 return sa_this_core_map;
6927 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6928 return sa_send_covered;
6929 d->rd = alloc_rootdomain();
6931 printk(KERN_WARNING "Cannot alloc root domain\n");
6934 return sa_rootdomain;
6937 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6938 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6940 struct sched_domain *sd = NULL;
6942 struct sched_domain *parent;
6945 if (cpumask_weight(cpu_map) >
6946 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6947 sd = &per_cpu(allnodes_domains, i).sd;
6948 SD_INIT(sd, ALLNODES);
6949 set_domain_attribute(sd, attr);
6950 cpumask_copy(sched_domain_span(sd), cpu_map);
6951 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6956 sd = &per_cpu(node_domains, i).sd;
6958 set_domain_attribute(sd, attr);
6959 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
6960 sd->parent = parent;
6963 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
6968 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
6969 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6970 struct sched_domain *parent, int i)
6972 struct sched_domain *sd;
6973 sd = &per_cpu(phys_domains, i).sd;
6975 set_domain_attribute(sd, attr);
6976 cpumask_copy(sched_domain_span(sd), d->nodemask);
6977 sd->parent = parent;
6980 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
6984 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
6985 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
6986 struct sched_domain *parent, int i)
6988 struct sched_domain *sd = parent;
6989 #ifdef CONFIG_SCHED_MC
6990 sd = &per_cpu(core_domains, i).sd;
6992 set_domain_attribute(sd, attr);
6993 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
6994 sd->parent = parent;
6996 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7001 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7002 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7003 struct sched_domain *parent, int i)
7005 struct sched_domain *sd = parent;
7006 #ifdef CONFIG_SCHED_SMT
7007 sd = &per_cpu(cpu_domains, i).sd;
7008 SD_INIT(sd, SIBLING);
7009 set_domain_attribute(sd, attr);
7010 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7011 sd->parent = parent;
7013 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7018 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7019 const struct cpumask *cpu_map, int cpu)
7022 #ifdef CONFIG_SCHED_SMT
7023 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7024 cpumask_and(d->this_sibling_map, cpu_map,
7025 topology_thread_cpumask(cpu));
7026 if (cpu == cpumask_first(d->this_sibling_map))
7027 init_sched_build_groups(d->this_sibling_map, cpu_map,
7029 d->send_covered, d->tmpmask);
7032 #ifdef CONFIG_SCHED_MC
7033 case SD_LV_MC: /* set up multi-core groups */
7034 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7035 if (cpu == cpumask_first(d->this_core_map))
7036 init_sched_build_groups(d->this_core_map, cpu_map,
7038 d->send_covered, d->tmpmask);
7041 case SD_LV_CPU: /* set up physical groups */
7042 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7043 if (!cpumask_empty(d->nodemask))
7044 init_sched_build_groups(d->nodemask, cpu_map,
7046 d->send_covered, d->tmpmask);
7049 case SD_LV_ALLNODES:
7050 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7051 d->send_covered, d->tmpmask);
7060 * Build sched domains for a given set of cpus and attach the sched domains
7061 * to the individual cpus
7063 static int __build_sched_domains(const struct cpumask *cpu_map,
7064 struct sched_domain_attr *attr)
7066 enum s_alloc alloc_state = sa_none;
7068 struct sched_domain *sd;
7074 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7075 if (alloc_state != sa_rootdomain)
7077 alloc_state = sa_sched_groups;
7080 * Set up domains for cpus specified by the cpu_map.
7082 for_each_cpu(i, cpu_map) {
7083 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7086 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7087 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7088 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7089 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7092 for_each_cpu(i, cpu_map) {
7093 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7094 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7097 /* Set up physical groups */
7098 for (i = 0; i < nr_node_ids; i++)
7099 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7102 /* Set up node groups */
7104 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7106 for (i = 0; i < nr_node_ids; i++)
7107 if (build_numa_sched_groups(&d, cpu_map, i))
7111 /* Calculate CPU power for physical packages and nodes */
7112 #ifdef CONFIG_SCHED_SMT
7113 for_each_cpu(i, cpu_map) {
7114 sd = &per_cpu(cpu_domains, i).sd;
7115 init_sched_groups_power(i, sd);
7118 #ifdef CONFIG_SCHED_MC
7119 for_each_cpu(i, cpu_map) {
7120 sd = &per_cpu(core_domains, i).sd;
7121 init_sched_groups_power(i, sd);
7125 for_each_cpu(i, cpu_map) {
7126 sd = &per_cpu(phys_domains, i).sd;
7127 init_sched_groups_power(i, sd);
7131 for (i = 0; i < nr_node_ids; i++)
7132 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7134 if (d.sd_allnodes) {
7135 struct sched_group *sg;
7137 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7139 init_numa_sched_groups_power(sg);
7143 /* Attach the domains */
7144 for_each_cpu(i, cpu_map) {
7145 #ifdef CONFIG_SCHED_SMT
7146 sd = &per_cpu(cpu_domains, i).sd;
7147 #elif defined(CONFIG_SCHED_MC)
7148 sd = &per_cpu(core_domains, i).sd;
7150 sd = &per_cpu(phys_domains, i).sd;
7152 cpu_attach_domain(sd, d.rd, i);
7155 d.sched_group_nodes = NULL; /* don't free this we still need it */
7156 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7160 __free_domain_allocs(&d, alloc_state, cpu_map);
7164 static int build_sched_domains(const struct cpumask *cpu_map)
7166 return __build_sched_domains(cpu_map, NULL);
7169 static cpumask_var_t *doms_cur; /* current sched domains */
7170 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7171 static struct sched_domain_attr *dattr_cur;
7172 /* attribues of custom domains in 'doms_cur' */
7175 * Special case: If a kmalloc of a doms_cur partition (array of
7176 * cpumask) fails, then fallback to a single sched domain,
7177 * as determined by the single cpumask fallback_doms.
7179 static cpumask_var_t fallback_doms;
7182 * arch_update_cpu_topology lets virtualized architectures update the
7183 * cpu core maps. It is supposed to return 1 if the topology changed
7184 * or 0 if it stayed the same.
7186 int __attribute__((weak)) arch_update_cpu_topology(void)
7191 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7194 cpumask_var_t *doms;
7196 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7199 for (i = 0; i < ndoms; i++) {
7200 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7201 free_sched_domains(doms, i);
7208 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7211 for (i = 0; i < ndoms; i++)
7212 free_cpumask_var(doms[i]);
7217 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7218 * For now this just excludes isolated cpus, but could be used to
7219 * exclude other special cases in the future.
7221 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7225 arch_update_cpu_topology();
7227 doms_cur = alloc_sched_domains(ndoms_cur);
7229 doms_cur = &fallback_doms;
7230 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7232 err = build_sched_domains(doms_cur[0]);
7233 register_sched_domain_sysctl();
7238 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7239 struct cpumask *tmpmask)
7241 free_sched_groups(cpu_map, tmpmask);
7245 * Detach sched domains from a group of cpus specified in cpu_map
7246 * These cpus will now be attached to the NULL domain
7248 static void detach_destroy_domains(const struct cpumask *cpu_map)
7250 /* Save because hotplug lock held. */
7251 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7254 for_each_cpu(i, cpu_map)
7255 cpu_attach_domain(NULL, &def_root_domain, i);
7256 synchronize_sched();
7257 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7260 /* handle null as "default" */
7261 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7262 struct sched_domain_attr *new, int idx_new)
7264 struct sched_domain_attr tmp;
7271 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7272 new ? (new + idx_new) : &tmp,
7273 sizeof(struct sched_domain_attr));
7277 * Partition sched domains as specified by the 'ndoms_new'
7278 * cpumasks in the array doms_new[] of cpumasks. This compares
7279 * doms_new[] to the current sched domain partitioning, doms_cur[].
7280 * It destroys each deleted domain and builds each new domain.
7282 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7283 * The masks don't intersect (don't overlap.) We should setup one
7284 * sched domain for each mask. CPUs not in any of the cpumasks will
7285 * not be load balanced. If the same cpumask appears both in the
7286 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7289 * The passed in 'doms_new' should be allocated using
7290 * alloc_sched_domains. This routine takes ownership of it and will
7291 * free_sched_domains it when done with it. If the caller failed the
7292 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7293 * and partition_sched_domains() will fallback to the single partition
7294 * 'fallback_doms', it also forces the domains to be rebuilt.
7296 * If doms_new == NULL it will be replaced with cpu_online_mask.
7297 * ndoms_new == 0 is a special case for destroying existing domains,
7298 * and it will not create the default domain.
7300 * Call with hotplug lock held
7302 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7303 struct sched_domain_attr *dattr_new)
7308 mutex_lock(&sched_domains_mutex);
7310 /* always unregister in case we don't destroy any domains */
7311 unregister_sched_domain_sysctl();
7313 /* Let architecture update cpu core mappings. */
7314 new_topology = arch_update_cpu_topology();
7316 n = doms_new ? ndoms_new : 0;
7318 /* Destroy deleted domains */
7319 for (i = 0; i < ndoms_cur; i++) {
7320 for (j = 0; j < n && !new_topology; j++) {
7321 if (cpumask_equal(doms_cur[i], doms_new[j])
7322 && dattrs_equal(dattr_cur, i, dattr_new, j))
7325 /* no match - a current sched domain not in new doms_new[] */
7326 detach_destroy_domains(doms_cur[i]);
7331 if (doms_new == NULL) {
7333 doms_new = &fallback_doms;
7334 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7335 WARN_ON_ONCE(dattr_new);
7338 /* Build new domains */
7339 for (i = 0; i < ndoms_new; i++) {
7340 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7341 if (cpumask_equal(doms_new[i], doms_cur[j])
7342 && dattrs_equal(dattr_new, i, dattr_cur, j))
7345 /* no match - add a new doms_new */
7346 __build_sched_domains(doms_new[i],
7347 dattr_new ? dattr_new + i : NULL);
7352 /* Remember the new sched domains */
7353 if (doms_cur != &fallback_doms)
7354 free_sched_domains(doms_cur, ndoms_cur);
7355 kfree(dattr_cur); /* kfree(NULL) is safe */
7356 doms_cur = doms_new;
7357 dattr_cur = dattr_new;
7358 ndoms_cur = ndoms_new;
7360 register_sched_domain_sysctl();
7362 mutex_unlock(&sched_domains_mutex);
7365 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7366 static void arch_reinit_sched_domains(void)
7370 /* Destroy domains first to force the rebuild */
7371 partition_sched_domains(0, NULL, NULL);
7373 rebuild_sched_domains();
7377 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7379 unsigned int level = 0;
7381 if (sscanf(buf, "%u", &level) != 1)
7385 * level is always be positive so don't check for
7386 * level < POWERSAVINGS_BALANCE_NONE which is 0
7387 * What happens on 0 or 1 byte write,
7388 * need to check for count as well?
7391 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7395 sched_smt_power_savings = level;
7397 sched_mc_power_savings = level;
7399 arch_reinit_sched_domains();
7404 #ifdef CONFIG_SCHED_MC
7405 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7406 struct sysdev_class_attribute *attr,
7409 return sprintf(page, "%u\n", sched_mc_power_savings);
7411 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7412 struct sysdev_class_attribute *attr,
7413 const char *buf, size_t count)
7415 return sched_power_savings_store(buf, count, 0);
7417 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7418 sched_mc_power_savings_show,
7419 sched_mc_power_savings_store);
7422 #ifdef CONFIG_SCHED_SMT
7423 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7424 struct sysdev_class_attribute *attr,
7427 return sprintf(page, "%u\n", sched_smt_power_savings);
7429 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7430 struct sysdev_class_attribute *attr,
7431 const char *buf, size_t count)
7433 return sched_power_savings_store(buf, count, 1);
7435 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7436 sched_smt_power_savings_show,
7437 sched_smt_power_savings_store);
7440 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7444 #ifdef CONFIG_SCHED_SMT
7446 err = sysfs_create_file(&cls->kset.kobj,
7447 &attr_sched_smt_power_savings.attr);
7449 #ifdef CONFIG_SCHED_MC
7450 if (!err && mc_capable())
7451 err = sysfs_create_file(&cls->kset.kobj,
7452 &attr_sched_mc_power_savings.attr);
7456 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7459 * Update cpusets according to cpu_active mask. If cpusets are
7460 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7461 * around partition_sched_domains().
7463 static int __cpuexit cpuset_cpu_active(struct notifier_block *nfb,
7464 unsigned long action, void *hcpu)
7466 switch (action & ~CPU_TASKS_FROZEN) {
7468 case CPU_DOWN_FAILED:
7469 cpuset_update_active_cpus();
7476 static int __cpuexit cpuset_cpu_inactive(struct notifier_block *nfb,
7477 unsigned long action, void *hcpu)
7479 switch (action & ~CPU_TASKS_FROZEN) {
7480 case CPU_DOWN_PREPARE:
7481 cpuset_update_active_cpus();
7488 static int update_runtime(struct notifier_block *nfb,
7489 unsigned long action, void *hcpu)
7491 int cpu = (int)(long)hcpu;
7494 case CPU_DOWN_PREPARE:
7495 case CPU_DOWN_PREPARE_FROZEN:
7496 disable_runtime(cpu_rq(cpu));
7499 case CPU_DOWN_FAILED:
7500 case CPU_DOWN_FAILED_FROZEN:
7502 case CPU_ONLINE_FROZEN:
7503 enable_runtime(cpu_rq(cpu));
7511 void __init sched_init_smp(void)
7513 cpumask_var_t non_isolated_cpus;
7515 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7516 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7518 #if defined(CONFIG_NUMA)
7519 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7521 BUG_ON(sched_group_nodes_bycpu == NULL);
7524 mutex_lock(&sched_domains_mutex);
7525 arch_init_sched_domains(cpu_active_mask);
7526 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7527 if (cpumask_empty(non_isolated_cpus))
7528 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7529 mutex_unlock(&sched_domains_mutex);
7532 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7533 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7535 /* RT runtime code needs to handle some hotplug events */
7536 hotcpu_notifier(update_runtime, 0);
7540 /* Move init over to a non-isolated CPU */
7541 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7543 sched_init_granularity();
7544 free_cpumask_var(non_isolated_cpus);
7546 init_sched_rt_class();
7549 void __init sched_init_smp(void)
7551 sched_init_granularity();
7553 #endif /* CONFIG_SMP */
7555 const_debug unsigned int sysctl_timer_migration = 1;
7557 int in_sched_functions(unsigned long addr)
7559 return in_lock_functions(addr) ||
7560 (addr >= (unsigned long)__sched_text_start
7561 && addr < (unsigned long)__sched_text_end);
7564 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7566 cfs_rq->tasks_timeline = RB_ROOT;
7567 INIT_LIST_HEAD(&cfs_rq->tasks);
7568 #ifdef CONFIG_FAIR_GROUP_SCHED
7571 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7574 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7576 struct rt_prio_array *array;
7579 array = &rt_rq->active;
7580 for (i = 0; i < MAX_RT_PRIO; i++) {
7581 INIT_LIST_HEAD(array->queue + i);
7582 __clear_bit(i, array->bitmap);
7584 /* delimiter for bitsearch: */
7585 __set_bit(MAX_RT_PRIO, array->bitmap);
7587 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7588 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7590 rt_rq->highest_prio.next = MAX_RT_PRIO;
7594 rt_rq->rt_nr_migratory = 0;
7595 rt_rq->overloaded = 0;
7596 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7600 rt_rq->rt_throttled = 0;
7601 rt_rq->rt_runtime = 0;
7602 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7604 #ifdef CONFIG_RT_GROUP_SCHED
7605 rt_rq->rt_nr_boosted = 0;
7610 #ifdef CONFIG_FAIR_GROUP_SCHED
7611 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7612 struct sched_entity *se, int cpu, int add,
7613 struct sched_entity *parent)
7615 struct rq *rq = cpu_rq(cpu);
7616 tg->cfs_rq[cpu] = cfs_rq;
7617 init_cfs_rq(cfs_rq, rq);
7620 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7623 /* se could be NULL for init_task_group */
7628 se->cfs_rq = &rq->cfs;
7630 se->cfs_rq = parent->my_q;
7633 se->load.weight = tg->shares;
7634 se->load.inv_weight = 0;
7635 se->parent = parent;
7639 #ifdef CONFIG_RT_GROUP_SCHED
7640 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7641 struct sched_rt_entity *rt_se, int cpu, int add,
7642 struct sched_rt_entity *parent)
7644 struct rq *rq = cpu_rq(cpu);
7646 tg->rt_rq[cpu] = rt_rq;
7647 init_rt_rq(rt_rq, rq);
7649 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7651 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7653 tg->rt_se[cpu] = rt_se;
7658 rt_se->rt_rq = &rq->rt;
7660 rt_se->rt_rq = parent->my_q;
7662 rt_se->my_q = rt_rq;
7663 rt_se->parent = parent;
7664 INIT_LIST_HEAD(&rt_se->run_list);
7668 void __init sched_init(void)
7671 unsigned long alloc_size = 0, ptr;
7673 #ifdef CONFIG_FAIR_GROUP_SCHED
7674 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7676 #ifdef CONFIG_RT_GROUP_SCHED
7677 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7679 #ifdef CONFIG_CPUMASK_OFFSTACK
7680 alloc_size += num_possible_cpus() * cpumask_size();
7683 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7685 #ifdef CONFIG_FAIR_GROUP_SCHED
7686 init_task_group.se = (struct sched_entity **)ptr;
7687 ptr += nr_cpu_ids * sizeof(void **);
7689 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7690 ptr += nr_cpu_ids * sizeof(void **);
7692 #endif /* CONFIG_FAIR_GROUP_SCHED */
7693 #ifdef CONFIG_RT_GROUP_SCHED
7694 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7695 ptr += nr_cpu_ids * sizeof(void **);
7697 init_task_group.rt_rq = (struct rt_rq **)ptr;
7698 ptr += nr_cpu_ids * sizeof(void **);
7700 #endif /* CONFIG_RT_GROUP_SCHED */
7701 #ifdef CONFIG_CPUMASK_OFFSTACK
7702 for_each_possible_cpu(i) {
7703 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7704 ptr += cpumask_size();
7706 #endif /* CONFIG_CPUMASK_OFFSTACK */
7710 init_defrootdomain();
7713 init_rt_bandwidth(&def_rt_bandwidth,
7714 global_rt_period(), global_rt_runtime());
7716 #ifdef CONFIG_RT_GROUP_SCHED
7717 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7718 global_rt_period(), global_rt_runtime());
7719 #endif /* CONFIG_RT_GROUP_SCHED */
7721 #ifdef CONFIG_CGROUP_SCHED
7722 list_add(&init_task_group.list, &task_groups);
7723 INIT_LIST_HEAD(&init_task_group.children);
7725 #endif /* CONFIG_CGROUP_SCHED */
7727 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7728 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7729 __alignof__(unsigned long));
7731 for_each_possible_cpu(i) {
7735 raw_spin_lock_init(&rq->lock);
7737 rq->calc_load_active = 0;
7738 rq->calc_load_update = jiffies + LOAD_FREQ;
7739 init_cfs_rq(&rq->cfs, rq);
7740 init_rt_rq(&rq->rt, rq);
7741 #ifdef CONFIG_FAIR_GROUP_SCHED
7742 init_task_group.shares = init_task_group_load;
7743 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7744 #ifdef CONFIG_CGROUP_SCHED
7746 * How much cpu bandwidth does init_task_group get?
7748 * In case of task-groups formed thr' the cgroup filesystem, it
7749 * gets 100% of the cpu resources in the system. This overall
7750 * system cpu resource is divided among the tasks of
7751 * init_task_group and its child task-groups in a fair manner,
7752 * based on each entity's (task or task-group's) weight
7753 * (se->load.weight).
7755 * In other words, if init_task_group has 10 tasks of weight
7756 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7757 * then A0's share of the cpu resource is:
7759 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7761 * We achieve this by letting init_task_group's tasks sit
7762 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7764 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7766 #endif /* CONFIG_FAIR_GROUP_SCHED */
7768 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7769 #ifdef CONFIG_RT_GROUP_SCHED
7770 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7771 #ifdef CONFIG_CGROUP_SCHED
7772 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7776 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7777 rq->cpu_load[j] = 0;
7779 rq->last_load_update_tick = jiffies;
7784 rq->cpu_power = SCHED_LOAD_SCALE;
7785 rq->post_schedule = 0;
7786 rq->active_balance = 0;
7787 rq->next_balance = jiffies;
7792 rq->avg_idle = 2*sysctl_sched_migration_cost;
7793 rq_attach_root(rq, &def_root_domain);
7796 atomic_set(&rq->nr_iowait, 0);
7799 set_load_weight(&init_task);
7801 #ifdef CONFIG_PREEMPT_NOTIFIERS
7802 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7806 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7809 #ifdef CONFIG_RT_MUTEXES
7810 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7814 * The boot idle thread does lazy MMU switching as well:
7816 atomic_inc(&init_mm.mm_count);
7817 enter_lazy_tlb(&init_mm, current);
7820 * Make us the idle thread. Technically, schedule() should not be
7821 * called from this thread, however somewhere below it might be,
7822 * but because we are the idle thread, we just pick up running again
7823 * when this runqueue becomes "idle".
7825 init_idle(current, smp_processor_id());
7827 calc_load_update = jiffies + LOAD_FREQ;
7830 * During early bootup we pretend to be a normal task:
7832 current->sched_class = &fair_sched_class;
7834 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7835 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7838 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
7839 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
7841 /* May be allocated at isolcpus cmdline parse time */
7842 if (cpu_isolated_map == NULL)
7843 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7848 scheduler_running = 1;
7851 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7852 static inline int preempt_count_equals(int preempt_offset)
7854 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7856 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7859 void __might_sleep(const char *file, int line, int preempt_offset)
7862 static unsigned long prev_jiffy; /* ratelimiting */
7864 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7865 system_state != SYSTEM_RUNNING || oops_in_progress)
7867 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7869 prev_jiffy = jiffies;
7872 "BUG: sleeping function called from invalid context at %s:%d\n",
7875 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7876 in_atomic(), irqs_disabled(),
7877 current->pid, current->comm);
7879 debug_show_held_locks(current);
7880 if (irqs_disabled())
7881 print_irqtrace_events(current);
7885 EXPORT_SYMBOL(__might_sleep);
7888 #ifdef CONFIG_MAGIC_SYSRQ
7889 static void normalize_task(struct rq *rq, struct task_struct *p)
7893 on_rq = p->se.on_rq;
7895 deactivate_task(rq, p, 0);
7896 __setscheduler(rq, p, SCHED_NORMAL, 0);
7898 activate_task(rq, p, 0);
7899 resched_task(rq->curr);
7903 void normalize_rt_tasks(void)
7905 struct task_struct *g, *p;
7906 unsigned long flags;
7909 read_lock_irqsave(&tasklist_lock, flags);
7910 do_each_thread(g, p) {
7912 * Only normalize user tasks:
7917 p->se.exec_start = 0;
7918 #ifdef CONFIG_SCHEDSTATS
7919 p->se.statistics.wait_start = 0;
7920 p->se.statistics.sleep_start = 0;
7921 p->se.statistics.block_start = 0;
7926 * Renice negative nice level userspace
7929 if (TASK_NICE(p) < 0 && p->mm)
7930 set_user_nice(p, 0);
7934 raw_spin_lock(&p->pi_lock);
7935 rq = __task_rq_lock(p);
7937 normalize_task(rq, p);
7939 __task_rq_unlock(rq);
7940 raw_spin_unlock(&p->pi_lock);
7941 } while_each_thread(g, p);
7943 read_unlock_irqrestore(&tasklist_lock, flags);
7946 #endif /* CONFIG_MAGIC_SYSRQ */
7948 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7950 * These functions are only useful for the IA64 MCA handling, or kdb.
7952 * They can only be called when the whole system has been
7953 * stopped - every CPU needs to be quiescent, and no scheduling
7954 * activity can take place. Using them for anything else would
7955 * be a serious bug, and as a result, they aren't even visible
7956 * under any other configuration.
7960 * curr_task - return the current task for a given cpu.
7961 * @cpu: the processor in question.
7963 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7965 struct task_struct *curr_task(int cpu)
7967 return cpu_curr(cpu);
7970 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7974 * set_curr_task - set the current task for a given cpu.
7975 * @cpu: the processor in question.
7976 * @p: the task pointer to set.
7978 * Description: This function must only be used when non-maskable interrupts
7979 * are serviced on a separate stack. It allows the architecture to switch the
7980 * notion of the current task on a cpu in a non-blocking manner. This function
7981 * must be called with all CPU's synchronized, and interrupts disabled, the
7982 * and caller must save the original value of the current task (see
7983 * curr_task() above) and restore that value before reenabling interrupts and
7984 * re-starting the system.
7986 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7988 void set_curr_task(int cpu, struct task_struct *p)
7995 #ifdef CONFIG_FAIR_GROUP_SCHED
7996 static void free_fair_sched_group(struct task_group *tg)
8000 for_each_possible_cpu(i) {
8002 kfree(tg->cfs_rq[i]);
8012 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8014 struct cfs_rq *cfs_rq;
8015 struct sched_entity *se;
8019 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8022 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8026 tg->shares = NICE_0_LOAD;
8028 for_each_possible_cpu(i) {
8031 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8032 GFP_KERNEL, cpu_to_node(i));
8036 se = kzalloc_node(sizeof(struct sched_entity),
8037 GFP_KERNEL, cpu_to_node(i));
8041 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8052 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8054 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8055 &cpu_rq(cpu)->leaf_cfs_rq_list);
8058 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8060 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8062 #else /* !CONFG_FAIR_GROUP_SCHED */
8063 static inline void free_fair_sched_group(struct task_group *tg)
8068 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8073 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8077 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8080 #endif /* CONFIG_FAIR_GROUP_SCHED */
8082 #ifdef CONFIG_RT_GROUP_SCHED
8083 static void free_rt_sched_group(struct task_group *tg)
8087 destroy_rt_bandwidth(&tg->rt_bandwidth);
8089 for_each_possible_cpu(i) {
8091 kfree(tg->rt_rq[i]);
8093 kfree(tg->rt_se[i]);
8101 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8103 struct rt_rq *rt_rq;
8104 struct sched_rt_entity *rt_se;
8108 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8111 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8115 init_rt_bandwidth(&tg->rt_bandwidth,
8116 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8118 for_each_possible_cpu(i) {
8121 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8122 GFP_KERNEL, cpu_to_node(i));
8126 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8127 GFP_KERNEL, cpu_to_node(i));
8131 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8142 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8144 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8145 &cpu_rq(cpu)->leaf_rt_rq_list);
8148 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8150 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8152 #else /* !CONFIG_RT_GROUP_SCHED */
8153 static inline void free_rt_sched_group(struct task_group *tg)
8158 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8163 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8167 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8170 #endif /* CONFIG_RT_GROUP_SCHED */
8172 #ifdef CONFIG_CGROUP_SCHED
8173 static void free_sched_group(struct task_group *tg)
8175 free_fair_sched_group(tg);
8176 free_rt_sched_group(tg);
8180 /* allocate runqueue etc for a new task group */
8181 struct task_group *sched_create_group(struct task_group *parent)
8183 struct task_group *tg;
8184 unsigned long flags;
8187 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8189 return ERR_PTR(-ENOMEM);
8191 if (!alloc_fair_sched_group(tg, parent))
8194 if (!alloc_rt_sched_group(tg, parent))
8197 spin_lock_irqsave(&task_group_lock, flags);
8198 for_each_possible_cpu(i) {
8199 register_fair_sched_group(tg, i);
8200 register_rt_sched_group(tg, i);
8202 list_add_rcu(&tg->list, &task_groups);
8204 WARN_ON(!parent); /* root should already exist */
8206 tg->parent = parent;
8207 INIT_LIST_HEAD(&tg->children);
8208 list_add_rcu(&tg->siblings, &parent->children);
8209 spin_unlock_irqrestore(&task_group_lock, flags);
8214 free_sched_group(tg);
8215 return ERR_PTR(-ENOMEM);
8218 /* rcu callback to free various structures associated with a task group */
8219 static void free_sched_group_rcu(struct rcu_head *rhp)
8221 /* now it should be safe to free those cfs_rqs */
8222 free_sched_group(container_of(rhp, struct task_group, rcu));
8225 /* Destroy runqueue etc associated with a task group */
8226 void sched_destroy_group(struct task_group *tg)
8228 unsigned long flags;
8231 spin_lock_irqsave(&task_group_lock, flags);
8232 for_each_possible_cpu(i) {
8233 unregister_fair_sched_group(tg, i);
8234 unregister_rt_sched_group(tg, i);
8236 list_del_rcu(&tg->list);
8237 list_del_rcu(&tg->siblings);
8238 spin_unlock_irqrestore(&task_group_lock, flags);
8240 /* wait for possible concurrent references to cfs_rqs complete */
8241 call_rcu(&tg->rcu, free_sched_group_rcu);
8244 /* change task's runqueue when it moves between groups.
8245 * The caller of this function should have put the task in its new group
8246 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8247 * reflect its new group.
8249 void sched_move_task(struct task_struct *tsk)
8252 unsigned long flags;
8255 rq = task_rq_lock(tsk, &flags);
8257 running = task_current(rq, tsk);
8258 on_rq = tsk->se.on_rq;
8261 dequeue_task(rq, tsk, 0);
8262 if (unlikely(running))
8263 tsk->sched_class->put_prev_task(rq, tsk);
8265 set_task_rq(tsk, task_cpu(tsk));
8267 #ifdef CONFIG_FAIR_GROUP_SCHED
8268 if (tsk->sched_class->moved_group)
8269 tsk->sched_class->moved_group(tsk, on_rq);
8272 if (unlikely(running))
8273 tsk->sched_class->set_curr_task(rq);
8275 enqueue_task(rq, tsk, 0);
8277 task_rq_unlock(rq, &flags);
8279 #endif /* CONFIG_CGROUP_SCHED */
8281 #ifdef CONFIG_FAIR_GROUP_SCHED
8282 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8284 struct cfs_rq *cfs_rq = se->cfs_rq;
8289 dequeue_entity(cfs_rq, se, 0);
8291 se->load.weight = shares;
8292 se->load.inv_weight = 0;
8295 enqueue_entity(cfs_rq, se, 0);
8298 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8300 struct cfs_rq *cfs_rq = se->cfs_rq;
8301 struct rq *rq = cfs_rq->rq;
8302 unsigned long flags;
8304 raw_spin_lock_irqsave(&rq->lock, flags);
8305 __set_se_shares(se, shares);
8306 raw_spin_unlock_irqrestore(&rq->lock, flags);
8309 static DEFINE_MUTEX(shares_mutex);
8311 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8314 unsigned long flags;
8317 * We can't change the weight of the root cgroup.
8322 if (shares < MIN_SHARES)
8323 shares = MIN_SHARES;
8324 else if (shares > MAX_SHARES)
8325 shares = MAX_SHARES;
8327 mutex_lock(&shares_mutex);
8328 if (tg->shares == shares)
8331 spin_lock_irqsave(&task_group_lock, flags);
8332 for_each_possible_cpu(i)
8333 unregister_fair_sched_group(tg, i);
8334 list_del_rcu(&tg->siblings);
8335 spin_unlock_irqrestore(&task_group_lock, flags);
8337 /* wait for any ongoing reference to this group to finish */
8338 synchronize_sched();
8341 * Now we are free to modify the group's share on each cpu
8342 * w/o tripping rebalance_share or load_balance_fair.
8344 tg->shares = shares;
8345 for_each_possible_cpu(i) {
8349 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8350 set_se_shares(tg->se[i], shares);
8354 * Enable load balance activity on this group, by inserting it back on
8355 * each cpu's rq->leaf_cfs_rq_list.
8357 spin_lock_irqsave(&task_group_lock, flags);
8358 for_each_possible_cpu(i)
8359 register_fair_sched_group(tg, i);
8360 list_add_rcu(&tg->siblings, &tg->parent->children);
8361 spin_unlock_irqrestore(&task_group_lock, flags);
8363 mutex_unlock(&shares_mutex);
8367 unsigned long sched_group_shares(struct task_group *tg)
8373 #ifdef CONFIG_RT_GROUP_SCHED
8375 * Ensure that the real time constraints are schedulable.
8377 static DEFINE_MUTEX(rt_constraints_mutex);
8379 static unsigned long to_ratio(u64 period, u64 runtime)
8381 if (runtime == RUNTIME_INF)
8384 return div64_u64(runtime << 20, period);
8387 /* Must be called with tasklist_lock held */
8388 static inline int tg_has_rt_tasks(struct task_group *tg)
8390 struct task_struct *g, *p;
8392 do_each_thread(g, p) {
8393 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8395 } while_each_thread(g, p);
8400 struct rt_schedulable_data {
8401 struct task_group *tg;
8406 static int tg_schedulable(struct task_group *tg, void *data)
8408 struct rt_schedulable_data *d = data;
8409 struct task_group *child;
8410 unsigned long total, sum = 0;
8411 u64 period, runtime;
8413 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8414 runtime = tg->rt_bandwidth.rt_runtime;
8417 period = d->rt_period;
8418 runtime = d->rt_runtime;
8422 * Cannot have more runtime than the period.
8424 if (runtime > period && runtime != RUNTIME_INF)
8428 * Ensure we don't starve existing RT tasks.
8430 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8433 total = to_ratio(period, runtime);
8436 * Nobody can have more than the global setting allows.
8438 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8442 * The sum of our children's runtime should not exceed our own.
8444 list_for_each_entry_rcu(child, &tg->children, siblings) {
8445 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8446 runtime = child->rt_bandwidth.rt_runtime;
8448 if (child == d->tg) {
8449 period = d->rt_period;
8450 runtime = d->rt_runtime;
8453 sum += to_ratio(period, runtime);
8462 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8464 struct rt_schedulable_data data = {
8466 .rt_period = period,
8467 .rt_runtime = runtime,
8470 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8473 static int tg_set_bandwidth(struct task_group *tg,
8474 u64 rt_period, u64 rt_runtime)
8478 mutex_lock(&rt_constraints_mutex);
8479 read_lock(&tasklist_lock);
8480 err = __rt_schedulable(tg, rt_period, rt_runtime);
8484 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8485 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8486 tg->rt_bandwidth.rt_runtime = rt_runtime;
8488 for_each_possible_cpu(i) {
8489 struct rt_rq *rt_rq = tg->rt_rq[i];
8491 raw_spin_lock(&rt_rq->rt_runtime_lock);
8492 rt_rq->rt_runtime = rt_runtime;
8493 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8495 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8497 read_unlock(&tasklist_lock);
8498 mutex_unlock(&rt_constraints_mutex);
8503 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8505 u64 rt_runtime, rt_period;
8507 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8508 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8509 if (rt_runtime_us < 0)
8510 rt_runtime = RUNTIME_INF;
8512 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8515 long sched_group_rt_runtime(struct task_group *tg)
8519 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8522 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8523 do_div(rt_runtime_us, NSEC_PER_USEC);
8524 return rt_runtime_us;
8527 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8529 u64 rt_runtime, rt_period;
8531 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8532 rt_runtime = tg->rt_bandwidth.rt_runtime;
8537 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8540 long sched_group_rt_period(struct task_group *tg)
8544 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8545 do_div(rt_period_us, NSEC_PER_USEC);
8546 return rt_period_us;
8549 static int sched_rt_global_constraints(void)
8551 u64 runtime, period;
8554 if (sysctl_sched_rt_period <= 0)
8557 runtime = global_rt_runtime();
8558 period = global_rt_period();
8561 * Sanity check on the sysctl variables.
8563 if (runtime > period && runtime != RUNTIME_INF)
8566 mutex_lock(&rt_constraints_mutex);
8567 read_lock(&tasklist_lock);
8568 ret = __rt_schedulable(NULL, 0, 0);
8569 read_unlock(&tasklist_lock);
8570 mutex_unlock(&rt_constraints_mutex);
8575 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8577 /* Don't accept realtime tasks when there is no way for them to run */
8578 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8584 #else /* !CONFIG_RT_GROUP_SCHED */
8585 static int sched_rt_global_constraints(void)
8587 unsigned long flags;
8590 if (sysctl_sched_rt_period <= 0)
8594 * There's always some RT tasks in the root group
8595 * -- migration, kstopmachine etc..
8597 if (sysctl_sched_rt_runtime == 0)
8600 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8601 for_each_possible_cpu(i) {
8602 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8604 raw_spin_lock(&rt_rq->rt_runtime_lock);
8605 rt_rq->rt_runtime = global_rt_runtime();
8606 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8608 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8612 #endif /* CONFIG_RT_GROUP_SCHED */
8614 int sched_rt_handler(struct ctl_table *table, int write,
8615 void __user *buffer, size_t *lenp,
8619 int old_period, old_runtime;
8620 static DEFINE_MUTEX(mutex);
8623 old_period = sysctl_sched_rt_period;
8624 old_runtime = sysctl_sched_rt_runtime;
8626 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8628 if (!ret && write) {
8629 ret = sched_rt_global_constraints();
8631 sysctl_sched_rt_period = old_period;
8632 sysctl_sched_rt_runtime = old_runtime;
8634 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8635 def_rt_bandwidth.rt_period =
8636 ns_to_ktime(global_rt_period());
8639 mutex_unlock(&mutex);
8644 #ifdef CONFIG_CGROUP_SCHED
8646 /* return corresponding task_group object of a cgroup */
8647 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8649 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8650 struct task_group, css);
8653 static struct cgroup_subsys_state *
8654 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8656 struct task_group *tg, *parent;
8658 if (!cgrp->parent) {
8659 /* This is early initialization for the top cgroup */
8660 return &init_task_group.css;
8663 parent = cgroup_tg(cgrp->parent);
8664 tg = sched_create_group(parent);
8666 return ERR_PTR(-ENOMEM);
8672 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8674 struct task_group *tg = cgroup_tg(cgrp);
8676 sched_destroy_group(tg);
8680 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8682 #ifdef CONFIG_RT_GROUP_SCHED
8683 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8686 /* We don't support RT-tasks being in separate groups */
8687 if (tsk->sched_class != &fair_sched_class)
8694 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8695 struct task_struct *tsk, bool threadgroup)
8697 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8701 struct task_struct *c;
8703 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8704 retval = cpu_cgroup_can_attach_task(cgrp, c);
8716 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8717 struct cgroup *old_cont, struct task_struct *tsk,
8720 sched_move_task(tsk);
8722 struct task_struct *c;
8724 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8731 #ifdef CONFIG_FAIR_GROUP_SCHED
8732 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8735 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8738 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8740 struct task_group *tg = cgroup_tg(cgrp);
8742 return (u64) tg->shares;
8744 #endif /* CONFIG_FAIR_GROUP_SCHED */
8746 #ifdef CONFIG_RT_GROUP_SCHED
8747 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8750 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8753 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8755 return sched_group_rt_runtime(cgroup_tg(cgrp));
8758 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8761 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8764 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8766 return sched_group_rt_period(cgroup_tg(cgrp));
8768 #endif /* CONFIG_RT_GROUP_SCHED */
8770 static struct cftype cpu_files[] = {
8771 #ifdef CONFIG_FAIR_GROUP_SCHED
8774 .read_u64 = cpu_shares_read_u64,
8775 .write_u64 = cpu_shares_write_u64,
8778 #ifdef CONFIG_RT_GROUP_SCHED
8780 .name = "rt_runtime_us",
8781 .read_s64 = cpu_rt_runtime_read,
8782 .write_s64 = cpu_rt_runtime_write,
8785 .name = "rt_period_us",
8786 .read_u64 = cpu_rt_period_read_uint,
8787 .write_u64 = cpu_rt_period_write_uint,
8792 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8794 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8797 struct cgroup_subsys cpu_cgroup_subsys = {
8799 .create = cpu_cgroup_create,
8800 .destroy = cpu_cgroup_destroy,
8801 .can_attach = cpu_cgroup_can_attach,
8802 .attach = cpu_cgroup_attach,
8803 .populate = cpu_cgroup_populate,
8804 .subsys_id = cpu_cgroup_subsys_id,
8808 #endif /* CONFIG_CGROUP_SCHED */
8810 #ifdef CONFIG_CGROUP_CPUACCT
8813 * CPU accounting code for task groups.
8815 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8816 * (balbir@in.ibm.com).
8819 /* track cpu usage of a group of tasks and its child groups */
8821 struct cgroup_subsys_state css;
8822 /* cpuusage holds pointer to a u64-type object on every cpu */
8823 u64 __percpu *cpuusage;
8824 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8825 struct cpuacct *parent;
8828 struct cgroup_subsys cpuacct_subsys;
8830 /* return cpu accounting group corresponding to this container */
8831 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8833 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8834 struct cpuacct, css);
8837 /* return cpu accounting group to which this task belongs */
8838 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8840 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8841 struct cpuacct, css);
8844 /* create a new cpu accounting group */
8845 static struct cgroup_subsys_state *cpuacct_create(
8846 struct cgroup_subsys *ss, struct cgroup *cgrp)
8848 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8854 ca->cpuusage = alloc_percpu(u64);
8858 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8859 if (percpu_counter_init(&ca->cpustat[i], 0))
8860 goto out_free_counters;
8863 ca->parent = cgroup_ca(cgrp->parent);
8869 percpu_counter_destroy(&ca->cpustat[i]);
8870 free_percpu(ca->cpuusage);
8874 return ERR_PTR(-ENOMEM);
8877 /* destroy an existing cpu accounting group */
8879 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8881 struct cpuacct *ca = cgroup_ca(cgrp);
8884 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8885 percpu_counter_destroy(&ca->cpustat[i]);
8886 free_percpu(ca->cpuusage);
8890 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8892 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8895 #ifndef CONFIG_64BIT
8897 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8899 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8901 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8909 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8911 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8913 #ifndef CONFIG_64BIT
8915 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8917 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8919 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8925 /* return total cpu usage (in nanoseconds) of a group */
8926 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8928 struct cpuacct *ca = cgroup_ca(cgrp);
8929 u64 totalcpuusage = 0;
8932 for_each_present_cpu(i)
8933 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8935 return totalcpuusage;
8938 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8941 struct cpuacct *ca = cgroup_ca(cgrp);
8950 for_each_present_cpu(i)
8951 cpuacct_cpuusage_write(ca, i, 0);
8957 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
8960 struct cpuacct *ca = cgroup_ca(cgroup);
8964 for_each_present_cpu(i) {
8965 percpu = cpuacct_cpuusage_read(ca, i);
8966 seq_printf(m, "%llu ", (unsigned long long) percpu);
8968 seq_printf(m, "\n");
8972 static const char *cpuacct_stat_desc[] = {
8973 [CPUACCT_STAT_USER] = "user",
8974 [CPUACCT_STAT_SYSTEM] = "system",
8977 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
8978 struct cgroup_map_cb *cb)
8980 struct cpuacct *ca = cgroup_ca(cgrp);
8983 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
8984 s64 val = percpu_counter_read(&ca->cpustat[i]);
8985 val = cputime64_to_clock_t(val);
8986 cb->fill(cb, cpuacct_stat_desc[i], val);
8991 static struct cftype files[] = {
8994 .read_u64 = cpuusage_read,
8995 .write_u64 = cpuusage_write,
8998 .name = "usage_percpu",
8999 .read_seq_string = cpuacct_percpu_seq_read,
9003 .read_map = cpuacct_stats_show,
9007 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9009 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9013 * charge this task's execution time to its accounting group.
9015 * called with rq->lock held.
9017 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9022 if (unlikely(!cpuacct_subsys.active))
9025 cpu = task_cpu(tsk);
9031 for (; ca; ca = ca->parent) {
9032 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9033 *cpuusage += cputime;
9040 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9041 * in cputime_t units. As a result, cpuacct_update_stats calls
9042 * percpu_counter_add with values large enough to always overflow the
9043 * per cpu batch limit causing bad SMP scalability.
9045 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9046 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9047 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9050 #define CPUACCT_BATCH \
9051 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9053 #define CPUACCT_BATCH 0
9057 * Charge the system/user time to the task's accounting group.
9059 static void cpuacct_update_stats(struct task_struct *tsk,
9060 enum cpuacct_stat_index idx, cputime_t val)
9063 int batch = CPUACCT_BATCH;
9065 if (unlikely(!cpuacct_subsys.active))
9072 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9078 struct cgroup_subsys cpuacct_subsys = {
9080 .create = cpuacct_create,
9081 .destroy = cpuacct_destroy,
9082 .populate = cpuacct_populate,
9083 .subsys_id = cpuacct_subsys_id,
9085 #endif /* CONFIG_CGROUP_CPUACCT */
9089 void synchronize_sched_expedited(void)
9093 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9095 #else /* #ifndef CONFIG_SMP */
9097 static atomic_t synchronize_sched_expedited_count = ATOMIC_INIT(0);
9099 static int synchronize_sched_expedited_cpu_stop(void *data)
9102 * There must be a full memory barrier on each affected CPU
9103 * between the time that try_stop_cpus() is called and the
9104 * time that it returns.
9106 * In the current initial implementation of cpu_stop, the
9107 * above condition is already met when the control reaches
9108 * this point and the following smp_mb() is not strictly
9109 * necessary. Do smp_mb() anyway for documentation and
9110 * robustness against future implementation changes.
9112 smp_mb(); /* See above comment block. */
9117 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9118 * approach to force grace period to end quickly. This consumes
9119 * significant time on all CPUs, and is thus not recommended for
9120 * any sort of common-case code.
9122 * Note that it is illegal to call this function while holding any
9123 * lock that is acquired by a CPU-hotplug notifier. Failing to
9124 * observe this restriction will result in deadlock.
9126 void synchronize_sched_expedited(void)
9128 int snap, trycount = 0;
9130 smp_mb(); /* ensure prior mod happens before capturing snap. */
9131 snap = atomic_read(&synchronize_sched_expedited_count) + 1;
9133 while (try_stop_cpus(cpu_online_mask,
9134 synchronize_sched_expedited_cpu_stop,
9137 if (trycount++ < 10)
9138 udelay(trycount * num_online_cpus());
9140 synchronize_sched();
9143 if (atomic_read(&synchronize_sched_expedited_count) - snap > 0) {
9144 smp_mb(); /* ensure test happens before caller kfree */
9149 atomic_inc(&synchronize_sched_expedited_count);
9150 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9153 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
9155 #endif /* #else #ifndef CONFIG_SMP */