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 nohz_balance_kick;
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 * In the semi idle case, use the nearest busy cpu for migrating timers
1199 * from an idle cpu. This is good for power-savings.
1201 * We don't do similar optimization for completely idle system, as
1202 * selecting an idle cpu will add more delays to the timers than intended
1203 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1205 int get_nohz_timer_target(void)
1207 int cpu = smp_processor_id();
1209 struct sched_domain *sd;
1211 for_each_domain(cpu, sd) {
1212 for_each_cpu(i, sched_domain_span(sd))
1219 * When add_timer_on() enqueues a timer into the timer wheel of an
1220 * idle CPU then this timer might expire before the next timer event
1221 * which is scheduled to wake up that CPU. In case of a completely
1222 * idle system the next event might even be infinite time into the
1223 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1224 * leaves the inner idle loop so the newly added timer is taken into
1225 * account when the CPU goes back to idle and evaluates the timer
1226 * wheel for the next timer event.
1228 void wake_up_idle_cpu(int cpu)
1230 struct rq *rq = cpu_rq(cpu);
1232 if (cpu == smp_processor_id())
1236 * This is safe, as this function is called with the timer
1237 * wheel base lock of (cpu) held. When the CPU is on the way
1238 * to idle and has not yet set rq->curr to idle then it will
1239 * be serialized on the timer wheel base lock and take the new
1240 * timer into account automatically.
1242 if (rq->curr != rq->idle)
1246 * We can set TIF_RESCHED on the idle task of the other CPU
1247 * lockless. The worst case is that the other CPU runs the
1248 * idle task through an additional NOOP schedule()
1250 set_tsk_need_resched(rq->idle);
1252 /* NEED_RESCHED must be visible before we test polling */
1254 if (!tsk_is_polling(rq->idle))
1255 smp_send_reschedule(cpu);
1258 #endif /* CONFIG_NO_HZ */
1260 static u64 sched_avg_period(void)
1262 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1265 static void sched_avg_update(struct rq *rq)
1267 s64 period = sched_avg_period();
1269 while ((s64)(rq->clock - rq->age_stamp) > period) {
1271 * Inline assembly required to prevent the compiler
1272 * optimising this loop into a divmod call.
1273 * See __iter_div_u64_rem() for another example of this.
1275 asm("" : "+rm" (rq->age_stamp));
1276 rq->age_stamp += period;
1281 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1283 rq->rt_avg += rt_delta;
1284 sched_avg_update(rq);
1287 #else /* !CONFIG_SMP */
1288 static void resched_task(struct task_struct *p)
1290 assert_raw_spin_locked(&task_rq(p)->lock);
1291 set_tsk_need_resched(p);
1294 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1298 static void sched_avg_update(struct rq *rq)
1301 #endif /* CONFIG_SMP */
1303 #if BITS_PER_LONG == 32
1304 # define WMULT_CONST (~0UL)
1306 # define WMULT_CONST (1UL << 32)
1309 #define WMULT_SHIFT 32
1312 * Shift right and round:
1314 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1317 * delta *= weight / lw
1319 static unsigned long
1320 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1321 struct load_weight *lw)
1325 if (!lw->inv_weight) {
1326 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1329 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1333 tmp = (u64)delta_exec * weight;
1335 * Check whether we'd overflow the 64-bit multiplication:
1337 if (unlikely(tmp > WMULT_CONST))
1338 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1341 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1343 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1346 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1352 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1359 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1360 * of tasks with abnormal "nice" values across CPUs the contribution that
1361 * each task makes to its run queue's load is weighted according to its
1362 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1363 * scaled version of the new time slice allocation that they receive on time
1367 #define WEIGHT_IDLEPRIO 3
1368 #define WMULT_IDLEPRIO 1431655765
1371 * Nice levels are multiplicative, with a gentle 10% change for every
1372 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1373 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1374 * that remained on nice 0.
1376 * The "10% effect" is relative and cumulative: from _any_ nice level,
1377 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1378 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1379 * If a task goes up by ~10% and another task goes down by ~10% then
1380 * the relative distance between them is ~25%.)
1382 static const int prio_to_weight[40] = {
1383 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1384 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1385 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1386 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1387 /* 0 */ 1024, 820, 655, 526, 423,
1388 /* 5 */ 335, 272, 215, 172, 137,
1389 /* 10 */ 110, 87, 70, 56, 45,
1390 /* 15 */ 36, 29, 23, 18, 15,
1394 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1396 * In cases where the weight does not change often, we can use the
1397 * precalculated inverse to speed up arithmetics by turning divisions
1398 * into multiplications:
1400 static const u32 prio_to_wmult[40] = {
1401 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1402 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1403 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1404 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1405 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1406 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1407 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1408 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1411 /* Time spent by the tasks of the cpu accounting group executing in ... */
1412 enum cpuacct_stat_index {
1413 CPUACCT_STAT_USER, /* ... user mode */
1414 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1416 CPUACCT_STAT_NSTATS,
1419 #ifdef CONFIG_CGROUP_CPUACCT
1420 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1421 static void cpuacct_update_stats(struct task_struct *tsk,
1422 enum cpuacct_stat_index idx, cputime_t val);
1424 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1425 static inline void cpuacct_update_stats(struct task_struct *tsk,
1426 enum cpuacct_stat_index idx, cputime_t val) {}
1429 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1431 update_load_add(&rq->load, load);
1434 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1436 update_load_sub(&rq->load, load);
1439 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1440 typedef int (*tg_visitor)(struct task_group *, void *);
1443 * Iterate the full tree, calling @down when first entering a node and @up when
1444 * leaving it for the final time.
1446 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1448 struct task_group *parent, *child;
1452 parent = &root_task_group;
1454 ret = (*down)(parent, data);
1457 list_for_each_entry_rcu(child, &parent->children, siblings) {
1464 ret = (*up)(parent, data);
1469 parent = parent->parent;
1478 static int tg_nop(struct task_group *tg, void *data)
1485 /* Used instead of source_load when we know the type == 0 */
1486 static unsigned long weighted_cpuload(const int cpu)
1488 return cpu_rq(cpu)->load.weight;
1492 * Return a low guess at the load of a migration-source cpu weighted
1493 * according to the scheduling class and "nice" value.
1495 * We want to under-estimate the load of migration sources, to
1496 * balance conservatively.
1498 static unsigned long source_load(int cpu, int type)
1500 struct rq *rq = cpu_rq(cpu);
1501 unsigned long total = weighted_cpuload(cpu);
1503 if (type == 0 || !sched_feat(LB_BIAS))
1506 return min(rq->cpu_load[type-1], total);
1510 * Return a high guess at the load of a migration-target cpu weighted
1511 * according to the scheduling class and "nice" value.
1513 static unsigned long target_load(int cpu, int type)
1515 struct rq *rq = cpu_rq(cpu);
1516 unsigned long total = weighted_cpuload(cpu);
1518 if (type == 0 || !sched_feat(LB_BIAS))
1521 return max(rq->cpu_load[type-1], total);
1524 static unsigned long power_of(int cpu)
1526 return cpu_rq(cpu)->cpu_power;
1529 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1531 static unsigned long cpu_avg_load_per_task(int cpu)
1533 struct rq *rq = cpu_rq(cpu);
1534 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1537 rq->avg_load_per_task = rq->load.weight / nr_running;
1539 rq->avg_load_per_task = 0;
1541 return rq->avg_load_per_task;
1544 #ifdef CONFIG_FAIR_GROUP_SCHED
1546 static __read_mostly unsigned long __percpu *update_shares_data;
1548 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1551 * Calculate and set the cpu's group shares.
1553 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1554 unsigned long sd_shares,
1555 unsigned long sd_rq_weight,
1556 unsigned long *usd_rq_weight)
1558 unsigned long shares, rq_weight;
1561 rq_weight = usd_rq_weight[cpu];
1564 rq_weight = NICE_0_LOAD;
1568 * \Sum_j shares_j * rq_weight_i
1569 * shares_i = -----------------------------
1570 * \Sum_j rq_weight_j
1572 shares = (sd_shares * rq_weight) / sd_rq_weight;
1573 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1575 if (abs(shares - tg->se[cpu]->load.weight) >
1576 sysctl_sched_shares_thresh) {
1577 struct rq *rq = cpu_rq(cpu);
1578 unsigned long flags;
1580 raw_spin_lock_irqsave(&rq->lock, flags);
1581 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1582 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1583 __set_se_shares(tg->se[cpu], shares);
1584 raw_spin_unlock_irqrestore(&rq->lock, flags);
1589 * Re-compute the task group their per cpu shares over the given domain.
1590 * This needs to be done in a bottom-up fashion because the rq weight of a
1591 * parent group depends on the shares of its child groups.
1593 static int tg_shares_up(struct task_group *tg, void *data)
1595 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1596 unsigned long *usd_rq_weight;
1597 struct sched_domain *sd = data;
1598 unsigned long flags;
1604 local_irq_save(flags);
1605 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1607 for_each_cpu(i, sched_domain_span(sd)) {
1608 weight = tg->cfs_rq[i]->load.weight;
1609 usd_rq_weight[i] = weight;
1611 rq_weight += weight;
1613 * If there are currently no tasks on the cpu pretend there
1614 * is one of average load so that when a new task gets to
1615 * run here it will not get delayed by group starvation.
1618 weight = NICE_0_LOAD;
1620 sum_weight += weight;
1621 shares += tg->cfs_rq[i]->shares;
1625 rq_weight = sum_weight;
1627 if ((!shares && rq_weight) || shares > tg->shares)
1628 shares = tg->shares;
1630 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1631 shares = tg->shares;
1633 for_each_cpu(i, sched_domain_span(sd))
1634 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1636 local_irq_restore(flags);
1642 * Compute the cpu's hierarchical load factor for each task group.
1643 * This needs to be done in a top-down fashion because the load of a child
1644 * group is a fraction of its parents load.
1646 static int tg_load_down(struct task_group *tg, void *data)
1649 long cpu = (long)data;
1652 load = cpu_rq(cpu)->load.weight;
1654 load = tg->parent->cfs_rq[cpu]->h_load;
1655 load *= tg->cfs_rq[cpu]->shares;
1656 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1659 tg->cfs_rq[cpu]->h_load = load;
1664 static void update_shares(struct sched_domain *sd)
1669 if (root_task_group_empty())
1672 now = local_clock();
1673 elapsed = now - sd->last_update;
1675 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1676 sd->last_update = now;
1677 walk_tg_tree(tg_nop, tg_shares_up, sd);
1681 static void update_h_load(long cpu)
1683 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1688 static inline void update_shares(struct sched_domain *sd)
1694 #ifdef CONFIG_PREEMPT
1696 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1699 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1700 * way at the expense of forcing extra atomic operations in all
1701 * invocations. This assures that the double_lock is acquired using the
1702 * same underlying policy as the spinlock_t on this architecture, which
1703 * reduces latency compared to the unfair variant below. However, it
1704 * also adds more overhead and therefore may reduce throughput.
1706 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1707 __releases(this_rq->lock)
1708 __acquires(busiest->lock)
1709 __acquires(this_rq->lock)
1711 raw_spin_unlock(&this_rq->lock);
1712 double_rq_lock(this_rq, busiest);
1719 * Unfair double_lock_balance: Optimizes throughput at the expense of
1720 * latency by eliminating extra atomic operations when the locks are
1721 * already in proper order on entry. This favors lower cpu-ids and will
1722 * grant the double lock to lower cpus over higher ids under contention,
1723 * regardless of entry order into the function.
1725 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1726 __releases(this_rq->lock)
1727 __acquires(busiest->lock)
1728 __acquires(this_rq->lock)
1732 if (unlikely(!raw_spin_trylock(&busiest->lock))) {
1733 if (busiest < this_rq) {
1734 raw_spin_unlock(&this_rq->lock);
1735 raw_spin_lock(&busiest->lock);
1736 raw_spin_lock_nested(&this_rq->lock,
1737 SINGLE_DEPTH_NESTING);
1740 raw_spin_lock_nested(&busiest->lock,
1741 SINGLE_DEPTH_NESTING);
1746 #endif /* CONFIG_PREEMPT */
1749 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1751 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1753 if (unlikely(!irqs_disabled())) {
1754 /* printk() doesn't work good under rq->lock */
1755 raw_spin_unlock(&this_rq->lock);
1759 return _double_lock_balance(this_rq, busiest);
1762 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1763 __releases(busiest->lock)
1765 raw_spin_unlock(&busiest->lock);
1766 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1770 * double_rq_lock - safely lock two runqueues
1772 * Note this does not disable interrupts like task_rq_lock,
1773 * you need to do so manually before calling.
1775 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
1776 __acquires(rq1->lock)
1777 __acquires(rq2->lock)
1779 BUG_ON(!irqs_disabled());
1781 raw_spin_lock(&rq1->lock);
1782 __acquire(rq2->lock); /* Fake it out ;) */
1785 raw_spin_lock(&rq1->lock);
1786 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
1788 raw_spin_lock(&rq2->lock);
1789 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
1795 * double_rq_unlock - safely unlock two runqueues
1797 * Note this does not restore interrupts like task_rq_unlock,
1798 * you need to do so manually after calling.
1800 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
1801 __releases(rq1->lock)
1802 __releases(rq2->lock)
1804 raw_spin_unlock(&rq1->lock);
1806 raw_spin_unlock(&rq2->lock);
1808 __release(rq2->lock);
1813 #ifdef CONFIG_FAIR_GROUP_SCHED
1814 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1817 cfs_rq->shares = shares;
1822 static void calc_load_account_idle(struct rq *this_rq);
1823 static void update_sysctl(void);
1824 static int get_update_sysctl_factor(void);
1825 static void update_cpu_load(struct rq *this_rq);
1827 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1829 set_task_rq(p, cpu);
1832 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1833 * successfuly executed on another CPU. We must ensure that updates of
1834 * per-task data have been completed by this moment.
1837 task_thread_info(p)->cpu = cpu;
1841 static const struct sched_class rt_sched_class;
1843 #define sched_class_highest (&rt_sched_class)
1844 #define for_each_class(class) \
1845 for (class = sched_class_highest; class; class = class->next)
1847 #include "sched_stats.h"
1849 static void inc_nr_running(struct rq *rq)
1854 static void dec_nr_running(struct rq *rq)
1859 static void set_load_weight(struct task_struct *p)
1861 if (task_has_rt_policy(p)) {
1862 p->se.load.weight = 0;
1863 p->se.load.inv_weight = WMULT_CONST;
1868 * SCHED_IDLE tasks get minimal weight:
1870 if (p->policy == SCHED_IDLE) {
1871 p->se.load.weight = WEIGHT_IDLEPRIO;
1872 p->se.load.inv_weight = WMULT_IDLEPRIO;
1876 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1877 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1880 static void enqueue_task(struct rq *rq, struct task_struct *p, int flags)
1882 update_rq_clock(rq);
1883 sched_info_queued(p);
1884 p->sched_class->enqueue_task(rq, p, flags);
1888 static void dequeue_task(struct rq *rq, struct task_struct *p, int flags)
1890 update_rq_clock(rq);
1891 sched_info_dequeued(p);
1892 p->sched_class->dequeue_task(rq, p, flags);
1897 * activate_task - move a task to the runqueue.
1899 static void activate_task(struct rq *rq, struct task_struct *p, int flags)
1901 if (task_contributes_to_load(p))
1902 rq->nr_uninterruptible--;
1904 enqueue_task(rq, p, flags);
1909 * deactivate_task - remove a task from the runqueue.
1911 static void deactivate_task(struct rq *rq, struct task_struct *p, int flags)
1913 if (task_contributes_to_load(p))
1914 rq->nr_uninterruptible++;
1916 dequeue_task(rq, p, flags);
1920 #include "sched_idletask.c"
1921 #include "sched_fair.c"
1922 #include "sched_rt.c"
1923 #ifdef CONFIG_SCHED_DEBUG
1924 # include "sched_debug.c"
1928 * __normal_prio - return the priority that is based on the static prio
1930 static inline int __normal_prio(struct task_struct *p)
1932 return p->static_prio;
1936 * Calculate the expected normal priority: i.e. priority
1937 * without taking RT-inheritance into account. Might be
1938 * boosted by interactivity modifiers. Changes upon fork,
1939 * setprio syscalls, and whenever the interactivity
1940 * estimator recalculates.
1942 static inline int normal_prio(struct task_struct *p)
1946 if (task_has_rt_policy(p))
1947 prio = MAX_RT_PRIO-1 - p->rt_priority;
1949 prio = __normal_prio(p);
1954 * Calculate the current priority, i.e. the priority
1955 * taken into account by the scheduler. This value might
1956 * be boosted by RT tasks, or might be boosted by
1957 * interactivity modifiers. Will be RT if the task got
1958 * RT-boosted. If not then it returns p->normal_prio.
1960 static int effective_prio(struct task_struct *p)
1962 p->normal_prio = normal_prio(p);
1964 * If we are RT tasks or we were boosted to RT priority,
1965 * keep the priority unchanged. Otherwise, update priority
1966 * to the normal priority:
1968 if (!rt_prio(p->prio))
1969 return p->normal_prio;
1974 * task_curr - is this task currently executing on a CPU?
1975 * @p: the task in question.
1977 inline int task_curr(const struct task_struct *p)
1979 return cpu_curr(task_cpu(p)) == p;
1982 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1983 const struct sched_class *prev_class,
1984 int oldprio, int running)
1986 if (prev_class != p->sched_class) {
1987 if (prev_class->switched_from)
1988 prev_class->switched_from(rq, p, running);
1989 p->sched_class->switched_to(rq, p, running);
1991 p->sched_class->prio_changed(rq, p, oldprio, running);
1996 * Is this task likely cache-hot:
1999 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2003 if (p->sched_class != &fair_sched_class)
2007 * Buddy candidates are cache hot:
2009 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2010 (&p->se == cfs_rq_of(&p->se)->next ||
2011 &p->se == cfs_rq_of(&p->se)->last))
2014 if (sysctl_sched_migration_cost == -1)
2016 if (sysctl_sched_migration_cost == 0)
2019 delta = now - p->se.exec_start;
2021 return delta < (s64)sysctl_sched_migration_cost;
2024 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2026 #ifdef CONFIG_SCHED_DEBUG
2028 * We should never call set_task_cpu() on a blocked task,
2029 * ttwu() will sort out the placement.
2031 WARN_ON_ONCE(p->state != TASK_RUNNING && p->state != TASK_WAKING &&
2032 !(task_thread_info(p)->preempt_count & PREEMPT_ACTIVE));
2035 trace_sched_migrate_task(p, new_cpu);
2037 if (task_cpu(p) != new_cpu) {
2038 p->se.nr_migrations++;
2039 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS, 1, 1, NULL, 0);
2042 __set_task_cpu(p, new_cpu);
2045 struct migration_arg {
2046 struct task_struct *task;
2050 static int migration_cpu_stop(void *data);
2053 * The task's runqueue lock must be held.
2054 * Returns true if you have to wait for migration thread.
2056 static bool migrate_task(struct task_struct *p, int dest_cpu)
2058 struct rq *rq = task_rq(p);
2061 * If the task is not on a runqueue (and not running), then
2062 * the next wake-up will properly place the task.
2064 return p->se.on_rq || task_running(rq, p);
2068 * wait_task_inactive - wait for a thread to unschedule.
2070 * If @match_state is nonzero, it's the @p->state value just checked and
2071 * not expected to change. If it changes, i.e. @p might have woken up,
2072 * then return zero. When we succeed in waiting for @p to be off its CPU,
2073 * we return a positive number (its total switch count). If a second call
2074 * a short while later returns the same number, the caller can be sure that
2075 * @p has remained unscheduled the whole time.
2077 * The caller must ensure that the task *will* unschedule sometime soon,
2078 * else this function might spin for a *long* time. This function can't
2079 * be called with interrupts off, or it may introduce deadlock with
2080 * smp_call_function() if an IPI is sent by the same process we are
2081 * waiting to become inactive.
2083 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2085 unsigned long flags;
2092 * We do the initial early heuristics without holding
2093 * any task-queue locks at all. We'll only try to get
2094 * the runqueue lock when things look like they will
2100 * If the task is actively running on another CPU
2101 * still, just relax and busy-wait without holding
2104 * NOTE! Since we don't hold any locks, it's not
2105 * even sure that "rq" stays as the right runqueue!
2106 * But we don't care, since "task_running()" will
2107 * return false if the runqueue has changed and p
2108 * is actually now running somewhere else!
2110 while (task_running(rq, p)) {
2111 if (match_state && unlikely(p->state != match_state))
2117 * Ok, time to look more closely! We need the rq
2118 * lock now, to be *sure*. If we're wrong, we'll
2119 * just go back and repeat.
2121 rq = task_rq_lock(p, &flags);
2122 trace_sched_wait_task(p);
2123 running = task_running(rq, p);
2124 on_rq = p->se.on_rq;
2126 if (!match_state || p->state == match_state)
2127 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2128 task_rq_unlock(rq, &flags);
2131 * If it changed from the expected state, bail out now.
2133 if (unlikely(!ncsw))
2137 * Was it really running after all now that we
2138 * checked with the proper locks actually held?
2140 * Oops. Go back and try again..
2142 if (unlikely(running)) {
2148 * It's not enough that it's not actively running,
2149 * it must be off the runqueue _entirely_, and not
2152 * So if it was still runnable (but just not actively
2153 * running right now), it's preempted, and we should
2154 * yield - it could be a while.
2156 if (unlikely(on_rq)) {
2157 schedule_timeout_uninterruptible(1);
2162 * Ahh, all good. It wasn't running, and it wasn't
2163 * runnable, which means that it will never become
2164 * running in the future either. We're all done!
2173 * kick_process - kick a running thread to enter/exit the kernel
2174 * @p: the to-be-kicked thread
2176 * Cause a process which is running on another CPU to enter
2177 * kernel-mode, without any delay. (to get signals handled.)
2179 * NOTE: this function doesnt have to take the runqueue lock,
2180 * because all it wants to ensure is that the remote task enters
2181 * the kernel. If the IPI races and the task has been migrated
2182 * to another CPU then no harm is done and the purpose has been
2185 void kick_process(struct task_struct *p)
2191 if ((cpu != smp_processor_id()) && task_curr(p))
2192 smp_send_reschedule(cpu);
2195 EXPORT_SYMBOL_GPL(kick_process);
2196 #endif /* CONFIG_SMP */
2199 * task_oncpu_function_call - call a function on the cpu on which a task runs
2200 * @p: the task to evaluate
2201 * @func: the function to be called
2202 * @info: the function call argument
2204 * Calls the function @func when the task is currently running. This might
2205 * be on the current CPU, which just calls the function directly
2207 void task_oncpu_function_call(struct task_struct *p,
2208 void (*func) (void *info), void *info)
2215 smp_call_function_single(cpu, func, info, 1);
2221 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2223 static int select_fallback_rq(int cpu, struct task_struct *p)
2226 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2228 /* Look for allowed, online CPU in same node. */
2229 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2230 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2233 /* Any allowed, online CPU? */
2234 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2235 if (dest_cpu < nr_cpu_ids)
2238 /* No more Mr. Nice Guy. */
2239 if (unlikely(dest_cpu >= nr_cpu_ids)) {
2240 dest_cpu = cpuset_cpus_allowed_fallback(p);
2242 * Don't tell them about moving exiting tasks or
2243 * kernel threads (both mm NULL), since they never
2246 if (p->mm && printk_ratelimit()) {
2247 printk(KERN_INFO "process %d (%s) no "
2248 "longer affine to cpu%d\n",
2249 task_pid_nr(p), p->comm, cpu);
2257 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2260 int select_task_rq(struct rq *rq, struct task_struct *p, int sd_flags, int wake_flags)
2262 int cpu = p->sched_class->select_task_rq(rq, p, sd_flags, wake_flags);
2265 * In order not to call set_task_cpu() on a blocking task we need
2266 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2269 * Since this is common to all placement strategies, this lives here.
2271 * [ this allows ->select_task() to simply return task_cpu(p) and
2272 * not worry about this generic constraint ]
2274 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2276 cpu = select_fallback_rq(task_cpu(p), p);
2281 static void update_avg(u64 *avg, u64 sample)
2283 s64 diff = sample - *avg;
2288 static inline void ttwu_activate(struct task_struct *p, struct rq *rq,
2289 bool is_sync, bool is_migrate, bool is_local,
2290 unsigned long en_flags)
2292 schedstat_inc(p, se.statistics.nr_wakeups);
2294 schedstat_inc(p, se.statistics.nr_wakeups_sync);
2296 schedstat_inc(p, se.statistics.nr_wakeups_migrate);
2298 schedstat_inc(p, se.statistics.nr_wakeups_local);
2300 schedstat_inc(p, se.statistics.nr_wakeups_remote);
2302 activate_task(rq, p, en_flags);
2305 static inline void ttwu_post_activation(struct task_struct *p, struct rq *rq,
2306 int wake_flags, bool success)
2308 trace_sched_wakeup(p, success);
2309 check_preempt_curr(rq, p, wake_flags);
2311 p->state = TASK_RUNNING;
2313 if (p->sched_class->task_woken)
2314 p->sched_class->task_woken(rq, p);
2316 if (unlikely(rq->idle_stamp)) {
2317 u64 delta = rq->clock - rq->idle_stamp;
2318 u64 max = 2*sysctl_sched_migration_cost;
2323 update_avg(&rq->avg_idle, delta);
2327 /* if a worker is waking up, notify workqueue */
2328 if ((p->flags & PF_WQ_WORKER) && success)
2329 wq_worker_waking_up(p, cpu_of(rq));
2333 * try_to_wake_up - wake up a thread
2334 * @p: the thread to be awakened
2335 * @state: the mask of task states that can be woken
2336 * @wake_flags: wake modifier flags (WF_*)
2338 * Put it on the run-queue if it's not already there. The "current"
2339 * thread is always on the run-queue (except when the actual
2340 * re-schedule is in progress), and as such you're allowed to do
2341 * the simpler "current->state = TASK_RUNNING" to mark yourself
2342 * runnable without the overhead of this.
2344 * Returns %true if @p was woken up, %false if it was already running
2345 * or @state didn't match @p's state.
2347 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2350 int cpu, orig_cpu, this_cpu, success = 0;
2351 unsigned long flags;
2352 unsigned long en_flags = ENQUEUE_WAKEUP;
2355 this_cpu = get_cpu();
2358 rq = task_rq_lock(p, &flags);
2359 if (!(p->state & state))
2369 if (unlikely(task_running(rq, p)))
2373 * In order to handle concurrent wakeups and release the rq->lock
2374 * we put the task in TASK_WAKING state.
2376 * First fix up the nr_uninterruptible count:
2378 if (task_contributes_to_load(p)) {
2379 if (likely(cpu_online(orig_cpu)))
2380 rq->nr_uninterruptible--;
2382 this_rq()->nr_uninterruptible--;
2384 p->state = TASK_WAKING;
2386 if (p->sched_class->task_waking) {
2387 p->sched_class->task_waking(rq, p);
2388 en_flags |= ENQUEUE_WAKING;
2391 cpu = select_task_rq(rq, p, SD_BALANCE_WAKE, wake_flags);
2392 if (cpu != orig_cpu)
2393 set_task_cpu(p, cpu);
2394 __task_rq_unlock(rq);
2397 raw_spin_lock(&rq->lock);
2400 * We migrated the task without holding either rq->lock, however
2401 * since the task is not on the task list itself, nobody else
2402 * will try and migrate the task, hence the rq should match the
2403 * cpu we just moved it to.
2405 WARN_ON(task_cpu(p) != cpu);
2406 WARN_ON(p->state != TASK_WAKING);
2408 #ifdef CONFIG_SCHEDSTATS
2409 schedstat_inc(rq, ttwu_count);
2410 if (cpu == this_cpu)
2411 schedstat_inc(rq, ttwu_local);
2413 struct sched_domain *sd;
2414 for_each_domain(this_cpu, sd) {
2415 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2416 schedstat_inc(sd, ttwu_wake_remote);
2421 #endif /* CONFIG_SCHEDSTATS */
2424 #endif /* CONFIG_SMP */
2425 ttwu_activate(p, rq, wake_flags & WF_SYNC, orig_cpu != cpu,
2426 cpu == this_cpu, en_flags);
2429 ttwu_post_activation(p, rq, wake_flags, success);
2431 task_rq_unlock(rq, &flags);
2438 * try_to_wake_up_local - try to wake up a local task with rq lock held
2439 * @p: the thread to be awakened
2441 * Put @p on the run-queue if it's not alredy there. The caller must
2442 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2443 * the current task. this_rq() stays locked over invocation.
2445 static void try_to_wake_up_local(struct task_struct *p)
2447 struct rq *rq = task_rq(p);
2448 bool success = false;
2450 BUG_ON(rq != this_rq());
2451 BUG_ON(p == current);
2452 lockdep_assert_held(&rq->lock);
2454 if (!(p->state & TASK_NORMAL))
2458 if (likely(!task_running(rq, p))) {
2459 schedstat_inc(rq, ttwu_count);
2460 schedstat_inc(rq, ttwu_local);
2462 ttwu_activate(p, rq, false, false, true, ENQUEUE_WAKEUP);
2465 ttwu_post_activation(p, rq, 0, success);
2469 * wake_up_process - Wake up a specific process
2470 * @p: The process to be woken up.
2472 * Attempt to wake up the nominated process and move it to the set of runnable
2473 * processes. Returns 1 if the process was woken up, 0 if it was already
2476 * It may be assumed that this function implies a write memory barrier before
2477 * changing the task state if and only if any tasks are woken up.
2479 int wake_up_process(struct task_struct *p)
2481 return try_to_wake_up(p, TASK_ALL, 0);
2483 EXPORT_SYMBOL(wake_up_process);
2485 int wake_up_state(struct task_struct *p, unsigned int state)
2487 return try_to_wake_up(p, state, 0);
2491 * Perform scheduler related setup for a newly forked process p.
2492 * p is forked by current.
2494 * __sched_fork() is basic setup used by init_idle() too:
2496 static void __sched_fork(struct task_struct *p)
2498 p->se.exec_start = 0;
2499 p->se.sum_exec_runtime = 0;
2500 p->se.prev_sum_exec_runtime = 0;
2501 p->se.nr_migrations = 0;
2503 #ifdef CONFIG_SCHEDSTATS
2504 memset(&p->se.statistics, 0, sizeof(p->se.statistics));
2507 INIT_LIST_HEAD(&p->rt.run_list);
2509 INIT_LIST_HEAD(&p->se.group_node);
2511 #ifdef CONFIG_PREEMPT_NOTIFIERS
2512 INIT_HLIST_HEAD(&p->preempt_notifiers);
2517 * fork()/clone()-time setup:
2519 void sched_fork(struct task_struct *p, int clone_flags)
2521 int cpu = get_cpu();
2525 * We mark the process as running here. This guarantees that
2526 * nobody will actually run it, and a signal or other external
2527 * event cannot wake it up and insert it on the runqueue either.
2529 p->state = TASK_RUNNING;
2532 * Revert to default priority/policy on fork if requested.
2534 if (unlikely(p->sched_reset_on_fork)) {
2535 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2536 p->policy = SCHED_NORMAL;
2537 p->normal_prio = p->static_prio;
2540 if (PRIO_TO_NICE(p->static_prio) < 0) {
2541 p->static_prio = NICE_TO_PRIO(0);
2542 p->normal_prio = p->static_prio;
2547 * We don't need the reset flag anymore after the fork. It has
2548 * fulfilled its duty:
2550 p->sched_reset_on_fork = 0;
2554 * Make sure we do not leak PI boosting priority to the child.
2556 p->prio = current->normal_prio;
2558 if (!rt_prio(p->prio))
2559 p->sched_class = &fair_sched_class;
2561 if (p->sched_class->task_fork)
2562 p->sched_class->task_fork(p);
2565 * The child is not yet in the pid-hash so no cgroup attach races,
2566 * and the cgroup is pinned to this child due to cgroup_fork()
2567 * is ran before sched_fork().
2569 * Silence PROVE_RCU.
2572 set_task_cpu(p, cpu);
2575 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2576 if (likely(sched_info_on()))
2577 memset(&p->sched_info, 0, sizeof(p->sched_info));
2579 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2582 #ifdef CONFIG_PREEMPT
2583 /* Want to start with kernel preemption disabled. */
2584 task_thread_info(p)->preempt_count = 1;
2586 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2592 * wake_up_new_task - wake up a newly created task for the first time.
2594 * This function will do some initial scheduler statistics housekeeping
2595 * that must be done for every newly created context, then puts the task
2596 * on the runqueue and wakes it.
2598 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2600 unsigned long flags;
2602 int cpu __maybe_unused = get_cpu();
2605 rq = task_rq_lock(p, &flags);
2606 p->state = TASK_WAKING;
2609 * Fork balancing, do it here and not earlier because:
2610 * - cpus_allowed can change in the fork path
2611 * - any previously selected cpu might disappear through hotplug
2613 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2614 * without people poking at ->cpus_allowed.
2616 cpu = select_task_rq(rq, p, SD_BALANCE_FORK, 0);
2617 set_task_cpu(p, cpu);
2619 p->state = TASK_RUNNING;
2620 task_rq_unlock(rq, &flags);
2623 rq = task_rq_lock(p, &flags);
2624 activate_task(rq, p, 0);
2625 trace_sched_wakeup_new(p, 1);
2626 check_preempt_curr(rq, p, WF_FORK);
2628 if (p->sched_class->task_woken)
2629 p->sched_class->task_woken(rq, p);
2631 task_rq_unlock(rq, &flags);
2635 #ifdef CONFIG_PREEMPT_NOTIFIERS
2638 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2639 * @notifier: notifier struct to register
2641 void preempt_notifier_register(struct preempt_notifier *notifier)
2643 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2645 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2648 * preempt_notifier_unregister - no longer interested in preemption notifications
2649 * @notifier: notifier struct to unregister
2651 * This is safe to call from within a preemption notifier.
2653 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2655 hlist_del(¬ifier->link);
2657 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2659 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2661 struct preempt_notifier *notifier;
2662 struct hlist_node *node;
2664 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2665 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2669 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2670 struct task_struct *next)
2672 struct preempt_notifier *notifier;
2673 struct hlist_node *node;
2675 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2676 notifier->ops->sched_out(notifier, next);
2679 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2681 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2686 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2687 struct task_struct *next)
2691 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2694 * prepare_task_switch - prepare to switch tasks
2695 * @rq: the runqueue preparing to switch
2696 * @prev: the current task that is being switched out
2697 * @next: the task we are going to switch to.
2699 * This is called with the rq lock held and interrupts off. It must
2700 * be paired with a subsequent finish_task_switch after the context
2703 * prepare_task_switch sets up locking and calls architecture specific
2707 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2708 struct task_struct *next)
2710 fire_sched_out_preempt_notifiers(prev, next);
2711 prepare_lock_switch(rq, next);
2712 prepare_arch_switch(next);
2716 * finish_task_switch - clean up after a task-switch
2717 * @rq: runqueue associated with task-switch
2718 * @prev: the thread we just switched away from.
2720 * finish_task_switch must be called after the context switch, paired
2721 * with a prepare_task_switch call before the context switch.
2722 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2723 * and do any other architecture-specific cleanup actions.
2725 * Note that we may have delayed dropping an mm in context_switch(). If
2726 * so, we finish that here outside of the runqueue lock. (Doing it
2727 * with the lock held can cause deadlocks; see schedule() for
2730 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2731 __releases(rq->lock)
2733 struct mm_struct *mm = rq->prev_mm;
2739 * A task struct has one reference for the use as "current".
2740 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2741 * schedule one last time. The schedule call will never return, and
2742 * the scheduled task must drop that reference.
2743 * The test for TASK_DEAD must occur while the runqueue locks are
2744 * still held, otherwise prev could be scheduled on another cpu, die
2745 * there before we look at prev->state, and then the reference would
2747 * Manfred Spraul <manfred@colorfullife.com>
2749 prev_state = prev->state;
2750 finish_arch_switch(prev);
2751 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2752 local_irq_disable();
2753 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2754 perf_event_task_sched_in(current);
2755 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2757 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2758 finish_lock_switch(rq, prev);
2760 fire_sched_in_preempt_notifiers(current);
2763 if (unlikely(prev_state == TASK_DEAD)) {
2765 * Remove function-return probe instances associated with this
2766 * task and put them back on the free list.
2768 kprobe_flush_task(prev);
2769 put_task_struct(prev);
2775 /* assumes rq->lock is held */
2776 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2778 if (prev->sched_class->pre_schedule)
2779 prev->sched_class->pre_schedule(rq, prev);
2782 /* rq->lock is NOT held, but preemption is disabled */
2783 static inline void post_schedule(struct rq *rq)
2785 if (rq->post_schedule) {
2786 unsigned long flags;
2788 raw_spin_lock_irqsave(&rq->lock, flags);
2789 if (rq->curr->sched_class->post_schedule)
2790 rq->curr->sched_class->post_schedule(rq);
2791 raw_spin_unlock_irqrestore(&rq->lock, flags);
2793 rq->post_schedule = 0;
2799 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2803 static inline void post_schedule(struct rq *rq)
2810 * schedule_tail - first thing a freshly forked thread must call.
2811 * @prev: the thread we just switched away from.
2813 asmlinkage void schedule_tail(struct task_struct *prev)
2814 __releases(rq->lock)
2816 struct rq *rq = this_rq();
2818 finish_task_switch(rq, prev);
2821 * FIXME: do we need to worry about rq being invalidated by the
2826 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2827 /* In this case, finish_task_switch does not reenable preemption */
2830 if (current->set_child_tid)
2831 put_user(task_pid_vnr(current), current->set_child_tid);
2835 * context_switch - switch to the new MM and the new
2836 * thread's register state.
2839 context_switch(struct rq *rq, struct task_struct *prev,
2840 struct task_struct *next)
2842 struct mm_struct *mm, *oldmm;
2844 prepare_task_switch(rq, prev, next);
2845 trace_sched_switch(prev, next);
2847 oldmm = prev->active_mm;
2849 * For paravirt, this is coupled with an exit in switch_to to
2850 * combine the page table reload and the switch backend into
2853 arch_start_context_switch(prev);
2856 next->active_mm = oldmm;
2857 atomic_inc(&oldmm->mm_count);
2858 enter_lazy_tlb(oldmm, next);
2860 switch_mm(oldmm, mm, next);
2862 if (likely(!prev->mm)) {
2863 prev->active_mm = NULL;
2864 rq->prev_mm = oldmm;
2867 * Since the runqueue lock will be released by the next
2868 * task (which is an invalid locking op but in the case
2869 * of the scheduler it's an obvious special-case), so we
2870 * do an early lockdep release here:
2872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2873 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2876 /* Here we just switch the register state and the stack. */
2877 switch_to(prev, next, prev);
2881 * this_rq must be evaluated again because prev may have moved
2882 * CPUs since it called schedule(), thus the 'rq' on its stack
2883 * frame will be invalid.
2885 finish_task_switch(this_rq(), prev);
2889 * nr_running, nr_uninterruptible and nr_context_switches:
2891 * externally visible scheduler statistics: current number of runnable
2892 * threads, current number of uninterruptible-sleeping threads, total
2893 * number of context switches performed since bootup.
2895 unsigned long nr_running(void)
2897 unsigned long i, sum = 0;
2899 for_each_online_cpu(i)
2900 sum += cpu_rq(i)->nr_running;
2905 unsigned long nr_uninterruptible(void)
2907 unsigned long i, sum = 0;
2909 for_each_possible_cpu(i)
2910 sum += cpu_rq(i)->nr_uninterruptible;
2913 * Since we read the counters lockless, it might be slightly
2914 * inaccurate. Do not allow it to go below zero though:
2916 if (unlikely((long)sum < 0))
2922 unsigned long long nr_context_switches(void)
2925 unsigned long long sum = 0;
2927 for_each_possible_cpu(i)
2928 sum += cpu_rq(i)->nr_switches;
2933 unsigned long nr_iowait(void)
2935 unsigned long i, sum = 0;
2937 for_each_possible_cpu(i)
2938 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2943 unsigned long nr_iowait_cpu(int cpu)
2945 struct rq *this = cpu_rq(cpu);
2946 return atomic_read(&this->nr_iowait);
2949 unsigned long this_cpu_load(void)
2951 struct rq *this = this_rq();
2952 return this->cpu_load[0];
2956 /* Variables and functions for calc_load */
2957 static atomic_long_t calc_load_tasks;
2958 static unsigned long calc_load_update;
2959 unsigned long avenrun[3];
2960 EXPORT_SYMBOL(avenrun);
2962 static long calc_load_fold_active(struct rq *this_rq)
2964 long nr_active, delta = 0;
2966 nr_active = this_rq->nr_running;
2967 nr_active += (long) this_rq->nr_uninterruptible;
2969 if (nr_active != this_rq->calc_load_active) {
2970 delta = nr_active - this_rq->calc_load_active;
2971 this_rq->calc_load_active = nr_active;
2979 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2981 * When making the ILB scale, we should try to pull this in as well.
2983 static atomic_long_t calc_load_tasks_idle;
2985 static void calc_load_account_idle(struct rq *this_rq)
2989 delta = calc_load_fold_active(this_rq);
2991 atomic_long_add(delta, &calc_load_tasks_idle);
2994 static long calc_load_fold_idle(void)
2999 * Its got a race, we don't care...
3001 if (atomic_long_read(&calc_load_tasks_idle))
3002 delta = atomic_long_xchg(&calc_load_tasks_idle, 0);
3007 static void calc_load_account_idle(struct rq *this_rq)
3011 static inline long calc_load_fold_idle(void)
3018 * get_avenrun - get the load average array
3019 * @loads: pointer to dest load array
3020 * @offset: offset to add
3021 * @shift: shift count to shift the result left
3023 * These values are estimates at best, so no need for locking.
3025 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3027 loads[0] = (avenrun[0] + offset) << shift;
3028 loads[1] = (avenrun[1] + offset) << shift;
3029 loads[2] = (avenrun[2] + offset) << shift;
3032 static unsigned long
3033 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3036 load += active * (FIXED_1 - exp);
3037 return load >> FSHIFT;
3041 * calc_load - update the avenrun load estimates 10 ticks after the
3042 * CPUs have updated calc_load_tasks.
3044 void calc_global_load(void)
3046 unsigned long upd = calc_load_update + 10;
3049 if (time_before(jiffies, upd))
3052 active = atomic_long_read(&calc_load_tasks);
3053 active = active > 0 ? active * FIXED_1 : 0;
3055 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3056 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3057 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3059 calc_load_update += LOAD_FREQ;
3063 * Called from update_cpu_load() to periodically update this CPU's
3066 static void calc_load_account_active(struct rq *this_rq)
3070 if (time_before(jiffies, this_rq->calc_load_update))
3073 delta = calc_load_fold_active(this_rq);
3074 delta += calc_load_fold_idle();
3076 atomic_long_add(delta, &calc_load_tasks);
3078 this_rq->calc_load_update += LOAD_FREQ;
3082 * The exact cpuload at various idx values, calculated at every tick would be
3083 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3085 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3086 * on nth tick when cpu may be busy, then we have:
3087 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3088 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3090 * decay_load_missed() below does efficient calculation of
3091 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3092 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3094 * The calculation is approximated on a 128 point scale.
3095 * degrade_zero_ticks is the number of ticks after which load at any
3096 * particular idx is approximated to be zero.
3097 * degrade_factor is a precomputed table, a row for each load idx.
3098 * Each column corresponds to degradation factor for a power of two ticks,
3099 * based on 128 point scale.
3101 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3102 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3104 * With this power of 2 load factors, we can degrade the load n times
3105 * by looking at 1 bits in n and doing as many mult/shift instead of
3106 * n mult/shifts needed by the exact degradation.
3108 #define DEGRADE_SHIFT 7
3109 static const unsigned char
3110 degrade_zero_ticks[CPU_LOAD_IDX_MAX] = {0, 8, 32, 64, 128};
3111 static const unsigned char
3112 degrade_factor[CPU_LOAD_IDX_MAX][DEGRADE_SHIFT + 1] = {
3113 {0, 0, 0, 0, 0, 0, 0, 0},
3114 {64, 32, 8, 0, 0, 0, 0, 0},
3115 {96, 72, 40, 12, 1, 0, 0},
3116 {112, 98, 75, 43, 15, 1, 0},
3117 {120, 112, 98, 76, 45, 16, 2} };
3120 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3121 * would be when CPU is idle and so we just decay the old load without
3122 * adding any new load.
3124 static unsigned long
3125 decay_load_missed(unsigned long load, unsigned long missed_updates, int idx)
3129 if (!missed_updates)
3132 if (missed_updates >= degrade_zero_ticks[idx])
3136 return load >> missed_updates;
3138 while (missed_updates) {
3139 if (missed_updates % 2)
3140 load = (load * degrade_factor[idx][j]) >> DEGRADE_SHIFT;
3142 missed_updates >>= 1;
3149 * Update rq->cpu_load[] statistics. This function is usually called every
3150 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3151 * every tick. We fix it up based on jiffies.
3153 static void update_cpu_load(struct rq *this_rq)
3155 unsigned long this_load = this_rq->load.weight;
3156 unsigned long curr_jiffies = jiffies;
3157 unsigned long pending_updates;
3160 this_rq->nr_load_updates++;
3162 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3163 if (curr_jiffies == this_rq->last_load_update_tick)
3166 pending_updates = curr_jiffies - this_rq->last_load_update_tick;
3167 this_rq->last_load_update_tick = curr_jiffies;
3169 /* Update our load: */
3170 this_rq->cpu_load[0] = this_load; /* Fasttrack for idx 0 */
3171 for (i = 1, scale = 2; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3172 unsigned long old_load, new_load;
3174 /* scale is effectively 1 << i now, and >> i divides by scale */
3176 old_load = this_rq->cpu_load[i];
3177 old_load = decay_load_missed(old_load, pending_updates - 1, i);
3178 new_load = this_load;
3180 * Round up the averaging division if load is increasing. This
3181 * prevents us from getting stuck on 9 if the load is 10, for
3184 if (new_load > old_load)
3185 new_load += scale - 1;
3187 this_rq->cpu_load[i] = (old_load * (scale - 1) + new_load) >> i;
3190 sched_avg_update(this_rq);
3193 static void update_cpu_load_active(struct rq *this_rq)
3195 update_cpu_load(this_rq);
3197 calc_load_account_active(this_rq);
3203 * sched_exec - execve() is a valuable balancing opportunity, because at
3204 * this point the task has the smallest effective memory and cache footprint.
3206 void sched_exec(void)
3208 struct task_struct *p = current;
3209 unsigned long flags;
3213 rq = task_rq_lock(p, &flags);
3214 dest_cpu = p->sched_class->select_task_rq(rq, p, SD_BALANCE_EXEC, 0);
3215 if (dest_cpu == smp_processor_id())
3219 * select_task_rq() can race against ->cpus_allowed
3221 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed) &&
3222 likely(cpu_active(dest_cpu)) && migrate_task(p, dest_cpu)) {
3223 struct migration_arg arg = { p, dest_cpu };
3225 task_rq_unlock(rq, &flags);
3226 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
3230 task_rq_unlock(rq, &flags);
3235 DEFINE_PER_CPU(struct kernel_stat, kstat);
3237 EXPORT_PER_CPU_SYMBOL(kstat);
3240 * Return any ns on the sched_clock that have not yet been accounted in
3241 * @p in case that task is currently running.
3243 * Called with task_rq_lock() held on @rq.
3245 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
3249 if (task_current(rq, p)) {
3250 update_rq_clock(rq);
3251 ns = rq->clock - p->se.exec_start;
3259 unsigned long long task_delta_exec(struct task_struct *p)
3261 unsigned long flags;
3265 rq = task_rq_lock(p, &flags);
3266 ns = do_task_delta_exec(p, rq);
3267 task_rq_unlock(rq, &flags);
3273 * Return accounted runtime for the task.
3274 * In case the task is currently running, return the runtime plus current's
3275 * pending runtime that have not been accounted yet.
3277 unsigned long long task_sched_runtime(struct task_struct *p)
3279 unsigned long flags;
3283 rq = task_rq_lock(p, &flags);
3284 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
3285 task_rq_unlock(rq, &flags);
3291 * Return sum_exec_runtime for the thread group.
3292 * In case the task is currently running, return the sum plus current's
3293 * pending runtime that have not been accounted yet.
3295 * Note that the thread group might have other running tasks as well,
3296 * so the return value not includes other pending runtime that other
3297 * running tasks might have.
3299 unsigned long long thread_group_sched_runtime(struct task_struct *p)
3301 struct task_cputime totals;
3302 unsigned long flags;
3306 rq = task_rq_lock(p, &flags);
3307 thread_group_cputime(p, &totals);
3308 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
3309 task_rq_unlock(rq, &flags);
3315 * Account user cpu time to a process.
3316 * @p: the process that the cpu time gets accounted to
3317 * @cputime: the cpu time spent in user space since the last update
3318 * @cputime_scaled: cputime scaled by cpu frequency
3320 void account_user_time(struct task_struct *p, cputime_t cputime,
3321 cputime_t cputime_scaled)
3323 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3326 /* Add user time to process. */
3327 p->utime = cputime_add(p->utime, cputime);
3328 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3329 account_group_user_time(p, cputime);
3331 /* Add user time to cpustat. */
3332 tmp = cputime_to_cputime64(cputime);
3333 if (TASK_NICE(p) > 0)
3334 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3336 cpustat->user = cputime64_add(cpustat->user, tmp);
3338 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
3339 /* Account for user time used */
3340 acct_update_integrals(p);
3344 * Account guest cpu time to a process.
3345 * @p: the process that the cpu time gets accounted to
3346 * @cputime: the cpu time spent in virtual machine since the last update
3347 * @cputime_scaled: cputime scaled by cpu frequency
3349 static void account_guest_time(struct task_struct *p, cputime_t cputime,
3350 cputime_t cputime_scaled)
3353 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3355 tmp = cputime_to_cputime64(cputime);
3357 /* Add guest time to process. */
3358 p->utime = cputime_add(p->utime, cputime);
3359 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
3360 account_group_user_time(p, cputime);
3361 p->gtime = cputime_add(p->gtime, cputime);
3363 /* Add guest time to cpustat. */
3364 if (TASK_NICE(p) > 0) {
3365 cpustat->nice = cputime64_add(cpustat->nice, tmp);
3366 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
3368 cpustat->user = cputime64_add(cpustat->user, tmp);
3369 cpustat->guest = cputime64_add(cpustat->guest, tmp);
3374 * Account system cpu time to a process.
3375 * @p: the process that the cpu time gets accounted to
3376 * @hardirq_offset: the offset to subtract from hardirq_count()
3377 * @cputime: the cpu time spent in kernel space since the last update
3378 * @cputime_scaled: cputime scaled by cpu frequency
3380 void account_system_time(struct task_struct *p, int hardirq_offset,
3381 cputime_t cputime, cputime_t cputime_scaled)
3383 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3386 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
3387 account_guest_time(p, cputime, cputime_scaled);
3391 /* Add system time to process. */
3392 p->stime = cputime_add(p->stime, cputime);
3393 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
3394 account_group_system_time(p, cputime);
3396 /* Add system time to cpustat. */
3397 tmp = cputime_to_cputime64(cputime);
3398 if (hardirq_count() - hardirq_offset)
3399 cpustat->irq = cputime64_add(cpustat->irq, tmp);
3400 else if (softirq_count())
3401 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
3403 cpustat->system = cputime64_add(cpustat->system, tmp);
3405 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
3407 /* Account for system time used */
3408 acct_update_integrals(p);
3412 * Account for involuntary wait time.
3413 * @steal: the cpu time spent in involuntary wait
3415 void account_steal_time(cputime_t cputime)
3417 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3418 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3420 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
3424 * Account for idle time.
3425 * @cputime: the cpu time spent in idle wait
3427 void account_idle_time(cputime_t cputime)
3429 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
3430 cputime64_t cputime64 = cputime_to_cputime64(cputime);
3431 struct rq *rq = this_rq();
3433 if (atomic_read(&rq->nr_iowait) > 0)
3434 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
3436 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
3439 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3442 * Account a single tick of cpu time.
3443 * @p: the process that the cpu time gets accounted to
3444 * @user_tick: indicates if the tick is a user or a system tick
3446 void account_process_tick(struct task_struct *p, int user_tick)
3448 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
3449 struct rq *rq = this_rq();
3452 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
3453 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
3454 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
3457 account_idle_time(cputime_one_jiffy);
3461 * Account multiple ticks of steal time.
3462 * @p: the process from which the cpu time has been stolen
3463 * @ticks: number of stolen ticks
3465 void account_steal_ticks(unsigned long ticks)
3467 account_steal_time(jiffies_to_cputime(ticks));
3471 * Account multiple ticks of idle time.
3472 * @ticks: number of stolen ticks
3474 void account_idle_ticks(unsigned long ticks)
3476 account_idle_time(jiffies_to_cputime(ticks));
3482 * Use precise platform statistics if available:
3484 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3485 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3491 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3493 struct task_cputime cputime;
3495 thread_group_cputime(p, &cputime);
3497 *ut = cputime.utime;
3498 *st = cputime.stime;
3502 #ifndef nsecs_to_cputime
3503 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3506 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3508 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
3511 * Use CFS's precise accounting:
3513 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
3519 do_div(temp, total);
3520 utime = (cputime_t)temp;
3525 * Compare with previous values, to keep monotonicity:
3527 p->prev_utime = max(p->prev_utime, utime);
3528 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
3530 *ut = p->prev_utime;
3531 *st = p->prev_stime;
3535 * Must be called with siglock held.
3537 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
3539 struct signal_struct *sig = p->signal;
3540 struct task_cputime cputime;
3541 cputime_t rtime, utime, total;
3543 thread_group_cputime(p, &cputime);
3545 total = cputime_add(cputime.utime, cputime.stime);
3546 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
3551 temp *= cputime.utime;
3552 do_div(temp, total);
3553 utime = (cputime_t)temp;
3557 sig->prev_utime = max(sig->prev_utime, utime);
3558 sig->prev_stime = max(sig->prev_stime,
3559 cputime_sub(rtime, sig->prev_utime));
3561 *ut = sig->prev_utime;
3562 *st = sig->prev_stime;
3567 * This function gets called by the timer code, with HZ frequency.
3568 * We call it with interrupts disabled.
3570 * It also gets called by the fork code, when changing the parent's
3573 void scheduler_tick(void)
3575 int cpu = smp_processor_id();
3576 struct rq *rq = cpu_rq(cpu);
3577 struct task_struct *curr = rq->curr;
3581 raw_spin_lock(&rq->lock);
3582 update_rq_clock(rq);
3583 update_cpu_load_active(rq);
3584 curr->sched_class->task_tick(rq, curr, 0);
3585 raw_spin_unlock(&rq->lock);
3587 perf_event_task_tick(curr);
3590 rq->idle_at_tick = idle_cpu(cpu);
3591 trigger_load_balance(rq, cpu);
3595 notrace unsigned long get_parent_ip(unsigned long addr)
3597 if (in_lock_functions(addr)) {
3598 addr = CALLER_ADDR2;
3599 if (in_lock_functions(addr))
3600 addr = CALLER_ADDR3;
3605 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3606 defined(CONFIG_PREEMPT_TRACER))
3608 void __kprobes add_preempt_count(int val)
3610 #ifdef CONFIG_DEBUG_PREEMPT
3614 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3617 preempt_count() += val;
3618 #ifdef CONFIG_DEBUG_PREEMPT
3620 * Spinlock count overflowing soon?
3622 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
3625 if (preempt_count() == val)
3626 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3628 EXPORT_SYMBOL(add_preempt_count);
3630 void __kprobes sub_preempt_count(int val)
3632 #ifdef CONFIG_DEBUG_PREEMPT
3636 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
3639 * Is the spinlock portion underflowing?
3641 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
3642 !(preempt_count() & PREEMPT_MASK)))
3646 if (preempt_count() == val)
3647 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
3648 preempt_count() -= val;
3650 EXPORT_SYMBOL(sub_preempt_count);
3655 * Print scheduling while atomic bug:
3657 static noinline void __schedule_bug(struct task_struct *prev)
3659 struct pt_regs *regs = get_irq_regs();
3661 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
3662 prev->comm, prev->pid, preempt_count());
3664 debug_show_held_locks(prev);
3666 if (irqs_disabled())
3667 print_irqtrace_events(prev);
3676 * Various schedule()-time debugging checks and statistics:
3678 static inline void schedule_debug(struct task_struct *prev)
3681 * Test if we are atomic. Since do_exit() needs to call into
3682 * schedule() atomically, we ignore that path for now.
3683 * Otherwise, whine if we are scheduling when we should not be.
3685 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
3686 __schedule_bug(prev);
3688 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
3690 schedstat_inc(this_rq(), sched_count);
3691 #ifdef CONFIG_SCHEDSTATS
3692 if (unlikely(prev->lock_depth >= 0)) {
3693 schedstat_inc(this_rq(), bkl_count);
3694 schedstat_inc(prev, sched_info.bkl_count);
3699 static void put_prev_task(struct rq *rq, struct task_struct *prev)
3702 update_rq_clock(rq);
3703 rq->skip_clock_update = 0;
3704 prev->sched_class->put_prev_task(rq, prev);
3708 * Pick up the highest-prio task:
3710 static inline struct task_struct *
3711 pick_next_task(struct rq *rq)
3713 const struct sched_class *class;
3714 struct task_struct *p;
3717 * Optimization: we know that if all tasks are in
3718 * the fair class we can call that function directly:
3720 if (likely(rq->nr_running == rq->cfs.nr_running)) {
3721 p = fair_sched_class.pick_next_task(rq);
3726 class = sched_class_highest;
3728 p = class->pick_next_task(rq);
3732 * Will never be NULL as the idle class always
3733 * returns a non-NULL p:
3735 class = class->next;
3740 * schedule() is the main scheduler function.
3742 asmlinkage void __sched schedule(void)
3744 struct task_struct *prev, *next;
3745 unsigned long *switch_count;
3751 cpu = smp_processor_id();
3753 rcu_note_context_switch(cpu);
3756 release_kernel_lock(prev);
3757 need_resched_nonpreemptible:
3759 schedule_debug(prev);
3761 if (sched_feat(HRTICK))
3764 raw_spin_lock_irq(&rq->lock);
3765 clear_tsk_need_resched(prev);
3767 switch_count = &prev->nivcsw;
3768 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
3769 if (unlikely(signal_pending_state(prev->state, prev))) {
3770 prev->state = TASK_RUNNING;
3773 * If a worker is going to sleep, notify and
3774 * ask workqueue whether it wants to wake up a
3775 * task to maintain concurrency. If so, wake
3778 if (prev->flags & PF_WQ_WORKER) {
3779 struct task_struct *to_wakeup;
3781 to_wakeup = wq_worker_sleeping(prev, cpu);
3783 try_to_wake_up_local(to_wakeup);
3785 deactivate_task(rq, prev, DEQUEUE_SLEEP);
3787 switch_count = &prev->nvcsw;
3790 pre_schedule(rq, prev);
3792 if (unlikely(!rq->nr_running))
3793 idle_balance(cpu, rq);
3795 put_prev_task(rq, prev);
3796 next = pick_next_task(rq);
3798 if (likely(prev != next)) {
3799 sched_info_switch(prev, next);
3800 perf_event_task_sched_out(prev, next);
3806 context_switch(rq, prev, next); /* unlocks the rq */
3808 * The context switch have flipped the stack from under us
3809 * and restored the local variables which were saved when
3810 * this task called schedule() in the past. prev == current
3811 * is still correct, but it can be moved to another cpu/rq.
3813 cpu = smp_processor_id();
3816 raw_spin_unlock_irq(&rq->lock);
3820 if (unlikely(reacquire_kernel_lock(prev)))
3821 goto need_resched_nonpreemptible;
3823 preempt_enable_no_resched();
3827 EXPORT_SYMBOL(schedule);
3829 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3831 * Look out! "owner" is an entirely speculative pointer
3832 * access and not reliable.
3834 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
3839 if (!sched_feat(OWNER_SPIN))
3842 #ifdef CONFIG_DEBUG_PAGEALLOC
3844 * Need to access the cpu field knowing that
3845 * DEBUG_PAGEALLOC could have unmapped it if
3846 * the mutex owner just released it and exited.
3848 if (probe_kernel_address(&owner->cpu, cpu))
3855 * Even if the access succeeded (likely case),
3856 * the cpu field may no longer be valid.
3858 if (cpu >= nr_cpumask_bits)
3862 * We need to validate that we can do a
3863 * get_cpu() and that we have the percpu area.
3865 if (!cpu_online(cpu))
3872 * Owner changed, break to re-assess state.
3874 if (lock->owner != owner) {
3876 * If the lock has switched to a different owner,
3877 * we likely have heavy contention. Return 0 to quit
3878 * optimistic spinning and not contend further:
3886 * Is that owner really running on that cpu?
3888 if (task_thread_info(rq->curr) != owner || need_resched())
3898 #ifdef CONFIG_PREEMPT
3900 * this is the entry point to schedule() from in-kernel preemption
3901 * off of preempt_enable. Kernel preemptions off return from interrupt
3902 * occur there and call schedule directly.
3904 asmlinkage void __sched notrace preempt_schedule(void)
3906 struct thread_info *ti = current_thread_info();
3909 * If there is a non-zero preempt_count or interrupts are disabled,
3910 * we do not want to preempt the current task. Just return..
3912 if (likely(ti->preempt_count || irqs_disabled()))
3916 add_preempt_count_notrace(PREEMPT_ACTIVE);
3918 sub_preempt_count_notrace(PREEMPT_ACTIVE);
3921 * Check again in case we missed a preemption opportunity
3922 * between schedule and now.
3925 } while (need_resched());
3927 EXPORT_SYMBOL(preempt_schedule);
3930 * this is the entry point to schedule() from kernel preemption
3931 * off of irq context.
3932 * Note, that this is called and return with irqs disabled. This will
3933 * protect us against recursive calling from irq.
3935 asmlinkage void __sched preempt_schedule_irq(void)
3937 struct thread_info *ti = current_thread_info();
3939 /* Catch callers which need to be fixed */
3940 BUG_ON(ti->preempt_count || !irqs_disabled());
3943 add_preempt_count(PREEMPT_ACTIVE);
3946 local_irq_disable();
3947 sub_preempt_count(PREEMPT_ACTIVE);
3950 * Check again in case we missed a preemption opportunity
3951 * between schedule and now.
3954 } while (need_resched());
3957 #endif /* CONFIG_PREEMPT */
3959 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
3962 return try_to_wake_up(curr->private, mode, wake_flags);
3964 EXPORT_SYMBOL(default_wake_function);
3967 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3968 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3969 * number) then we wake all the non-exclusive tasks and one exclusive task.
3971 * There are circumstances in which we can try to wake a task which has already
3972 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3973 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3975 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
3976 int nr_exclusive, int wake_flags, void *key)
3978 wait_queue_t *curr, *next;
3980 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
3981 unsigned flags = curr->flags;
3983 if (curr->func(curr, mode, wake_flags, key) &&
3984 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
3990 * __wake_up - wake up threads blocked on a waitqueue.
3992 * @mode: which threads
3993 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3994 * @key: is directly passed to the wakeup function
3996 * It may be assumed that this function implies a write memory barrier before
3997 * changing the task state if and only if any tasks are woken up.
3999 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4000 int nr_exclusive, void *key)
4002 unsigned long flags;
4004 spin_lock_irqsave(&q->lock, flags);
4005 __wake_up_common(q, mode, nr_exclusive, 0, key);
4006 spin_unlock_irqrestore(&q->lock, flags);
4008 EXPORT_SYMBOL(__wake_up);
4011 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4013 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4015 __wake_up_common(q, mode, 1, 0, NULL);
4017 EXPORT_SYMBOL_GPL(__wake_up_locked);
4019 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
4021 __wake_up_common(q, mode, 1, 0, key);
4025 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4027 * @mode: which threads
4028 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4029 * @key: opaque value to be passed to wakeup targets
4031 * The sync wakeup differs that the waker knows that it will schedule
4032 * away soon, so while the target thread will be woken up, it will not
4033 * be migrated to another CPU - ie. the two threads are 'synchronized'
4034 * with each other. This can prevent needless bouncing between CPUs.
4036 * On UP it can prevent extra preemption.
4038 * It may be assumed that this function implies a write memory barrier before
4039 * changing the task state if and only if any tasks are woken up.
4041 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
4042 int nr_exclusive, void *key)
4044 unsigned long flags;
4045 int wake_flags = WF_SYNC;
4050 if (unlikely(!nr_exclusive))
4053 spin_lock_irqsave(&q->lock, flags);
4054 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
4055 spin_unlock_irqrestore(&q->lock, flags);
4057 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
4060 * __wake_up_sync - see __wake_up_sync_key()
4062 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4064 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
4066 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4069 * complete: - signals a single thread waiting on this completion
4070 * @x: holds the state of this particular completion
4072 * This will wake up a single thread waiting on this completion. Threads will be
4073 * awakened in the same order in which they were queued.
4075 * See also complete_all(), wait_for_completion() and related routines.
4077 * It may be assumed that this function implies a write memory barrier before
4078 * changing the task state if and only if any tasks are woken up.
4080 void complete(struct completion *x)
4082 unsigned long flags;
4084 spin_lock_irqsave(&x->wait.lock, flags);
4086 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4087 spin_unlock_irqrestore(&x->wait.lock, flags);
4089 EXPORT_SYMBOL(complete);
4092 * complete_all: - signals all threads waiting on this completion
4093 * @x: holds the state of this particular completion
4095 * This will wake up all threads waiting on this particular completion event.
4097 * It may be assumed that this function implies a write memory barrier before
4098 * changing the task state if and only if any tasks are woken up.
4100 void complete_all(struct completion *x)
4102 unsigned long flags;
4104 spin_lock_irqsave(&x->wait.lock, flags);
4105 x->done += UINT_MAX/2;
4106 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4107 spin_unlock_irqrestore(&x->wait.lock, flags);
4109 EXPORT_SYMBOL(complete_all);
4111 static inline long __sched
4112 do_wait_for_common(struct completion *x, long timeout, int state)
4115 DECLARE_WAITQUEUE(wait, current);
4117 __add_wait_queue_tail_exclusive(&x->wait, &wait);
4119 if (signal_pending_state(state, current)) {
4120 timeout = -ERESTARTSYS;
4123 __set_current_state(state);
4124 spin_unlock_irq(&x->wait.lock);
4125 timeout = schedule_timeout(timeout);
4126 spin_lock_irq(&x->wait.lock);
4127 } while (!x->done && timeout);
4128 __remove_wait_queue(&x->wait, &wait);
4133 return timeout ?: 1;
4137 wait_for_common(struct completion *x, long timeout, int state)
4141 spin_lock_irq(&x->wait.lock);
4142 timeout = do_wait_for_common(x, timeout, state);
4143 spin_unlock_irq(&x->wait.lock);
4148 * wait_for_completion: - waits for completion of a task
4149 * @x: holds the state of this particular completion
4151 * This waits to be signaled for completion of a specific task. It is NOT
4152 * interruptible and there is no timeout.
4154 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4155 * and interrupt capability. Also see complete().
4157 void __sched wait_for_completion(struct completion *x)
4159 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4161 EXPORT_SYMBOL(wait_for_completion);
4164 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4165 * @x: holds the state of this particular completion
4166 * @timeout: timeout value in jiffies
4168 * This waits for either a completion of a specific task to be signaled or for a
4169 * specified timeout to expire. The timeout is in jiffies. It is not
4172 unsigned long __sched
4173 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4175 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4177 EXPORT_SYMBOL(wait_for_completion_timeout);
4180 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4181 * @x: holds the state of this particular completion
4183 * This waits for completion of a specific task to be signaled. It is
4186 int __sched wait_for_completion_interruptible(struct completion *x)
4188 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4189 if (t == -ERESTARTSYS)
4193 EXPORT_SYMBOL(wait_for_completion_interruptible);
4196 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4197 * @x: holds the state of this particular completion
4198 * @timeout: timeout value in jiffies
4200 * This waits for either a completion of a specific task to be signaled or for a
4201 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4203 unsigned long __sched
4204 wait_for_completion_interruptible_timeout(struct completion *x,
4205 unsigned long timeout)
4207 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4209 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4212 * wait_for_completion_killable: - waits for completion of a task (killable)
4213 * @x: holds the state of this particular completion
4215 * This waits to be signaled for completion of a specific task. It can be
4216 * interrupted by a kill signal.
4218 int __sched wait_for_completion_killable(struct completion *x)
4220 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4221 if (t == -ERESTARTSYS)
4225 EXPORT_SYMBOL(wait_for_completion_killable);
4228 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4229 * @x: holds the state of this particular completion
4230 * @timeout: timeout value in jiffies
4232 * This waits for either a completion of a specific task to be
4233 * signaled or for a specified timeout to expire. It can be
4234 * interrupted by a kill signal. The timeout is in jiffies.
4236 unsigned long __sched
4237 wait_for_completion_killable_timeout(struct completion *x,
4238 unsigned long timeout)
4240 return wait_for_common(x, timeout, TASK_KILLABLE);
4242 EXPORT_SYMBOL(wait_for_completion_killable_timeout);
4245 * try_wait_for_completion - try to decrement a completion without blocking
4246 * @x: completion structure
4248 * Returns: 0 if a decrement cannot be done without blocking
4249 * 1 if a decrement succeeded.
4251 * If a completion is being used as a counting completion,
4252 * attempt to decrement the counter without blocking. This
4253 * enables us to avoid waiting if the resource the completion
4254 * is protecting is not available.
4256 bool try_wait_for_completion(struct completion *x)
4258 unsigned long flags;
4261 spin_lock_irqsave(&x->wait.lock, flags);
4266 spin_unlock_irqrestore(&x->wait.lock, flags);
4269 EXPORT_SYMBOL(try_wait_for_completion);
4272 * completion_done - Test to see if a completion has any waiters
4273 * @x: completion structure
4275 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4276 * 1 if there are no waiters.
4279 bool completion_done(struct completion *x)
4281 unsigned long flags;
4284 spin_lock_irqsave(&x->wait.lock, flags);
4287 spin_unlock_irqrestore(&x->wait.lock, flags);
4290 EXPORT_SYMBOL(completion_done);
4293 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4295 unsigned long flags;
4298 init_waitqueue_entry(&wait, current);
4300 __set_current_state(state);
4302 spin_lock_irqsave(&q->lock, flags);
4303 __add_wait_queue(q, &wait);
4304 spin_unlock(&q->lock);
4305 timeout = schedule_timeout(timeout);
4306 spin_lock_irq(&q->lock);
4307 __remove_wait_queue(q, &wait);
4308 spin_unlock_irqrestore(&q->lock, flags);
4313 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4315 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4317 EXPORT_SYMBOL(interruptible_sleep_on);
4320 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4322 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4324 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4326 void __sched sleep_on(wait_queue_head_t *q)
4328 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4330 EXPORT_SYMBOL(sleep_on);
4332 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4334 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4336 EXPORT_SYMBOL(sleep_on_timeout);
4338 #ifdef CONFIG_RT_MUTEXES
4341 * rt_mutex_setprio - set the current priority of a task
4343 * @prio: prio value (kernel-internal form)
4345 * This function changes the 'effective' priority of a task. It does
4346 * not touch ->normal_prio like __setscheduler().
4348 * Used by the rt_mutex code to implement priority inheritance logic.
4350 void rt_mutex_setprio(struct task_struct *p, int prio)
4352 unsigned long flags;
4353 int oldprio, on_rq, running;
4355 const struct sched_class *prev_class;
4357 BUG_ON(prio < 0 || prio > MAX_PRIO);
4359 rq = task_rq_lock(p, &flags);
4362 prev_class = p->sched_class;
4363 on_rq = p->se.on_rq;
4364 running = task_current(rq, p);
4366 dequeue_task(rq, p, 0);
4368 p->sched_class->put_prev_task(rq, p);
4371 p->sched_class = &rt_sched_class;
4373 p->sched_class = &fair_sched_class;
4378 p->sched_class->set_curr_task(rq);
4380 enqueue_task(rq, p, oldprio < prio ? ENQUEUE_HEAD : 0);
4382 check_class_changed(rq, p, prev_class, oldprio, running);
4384 task_rq_unlock(rq, &flags);
4389 void set_user_nice(struct task_struct *p, long nice)
4391 int old_prio, delta, on_rq;
4392 unsigned long flags;
4395 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4398 * We have to be careful, if called from sys_setpriority(),
4399 * the task might be in the middle of scheduling on another CPU.
4401 rq = task_rq_lock(p, &flags);
4403 * The RT priorities are set via sched_setscheduler(), but we still
4404 * allow the 'normal' nice value to be set - but as expected
4405 * it wont have any effect on scheduling until the task is
4406 * SCHED_FIFO/SCHED_RR:
4408 if (task_has_rt_policy(p)) {
4409 p->static_prio = NICE_TO_PRIO(nice);
4412 on_rq = p->se.on_rq;
4414 dequeue_task(rq, p, 0);
4416 p->static_prio = NICE_TO_PRIO(nice);
4419 p->prio = effective_prio(p);
4420 delta = p->prio - old_prio;
4423 enqueue_task(rq, p, 0);
4425 * If the task increased its priority or is running and
4426 * lowered its priority, then reschedule its CPU:
4428 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4429 resched_task(rq->curr);
4432 task_rq_unlock(rq, &flags);
4434 EXPORT_SYMBOL(set_user_nice);
4437 * can_nice - check if a task can reduce its nice value
4441 int can_nice(const struct task_struct *p, const int nice)
4443 /* convert nice value [19,-20] to rlimit style value [1,40] */
4444 int nice_rlim = 20 - nice;
4446 return (nice_rlim <= task_rlimit(p, RLIMIT_NICE) ||
4447 capable(CAP_SYS_NICE));
4450 #ifdef __ARCH_WANT_SYS_NICE
4453 * sys_nice - change the priority of the current process.
4454 * @increment: priority increment
4456 * sys_setpriority is a more generic, but much slower function that
4457 * does similar things.
4459 SYSCALL_DEFINE1(nice, int, increment)
4464 * Setpriority might change our priority at the same moment.
4465 * We don't have to worry. Conceptually one call occurs first
4466 * and we have a single winner.
4468 if (increment < -40)
4473 nice = TASK_NICE(current) + increment;
4479 if (increment < 0 && !can_nice(current, nice))
4482 retval = security_task_setnice(current, nice);
4486 set_user_nice(current, nice);
4493 * task_prio - return the priority value of a given task.
4494 * @p: the task in question.
4496 * This is the priority value as seen by users in /proc.
4497 * RT tasks are offset by -200. Normal tasks are centered
4498 * around 0, value goes from -16 to +15.
4500 int task_prio(const struct task_struct *p)
4502 return p->prio - MAX_RT_PRIO;
4506 * task_nice - return the nice value of a given task.
4507 * @p: the task in question.
4509 int task_nice(const struct task_struct *p)
4511 return TASK_NICE(p);
4513 EXPORT_SYMBOL(task_nice);
4516 * idle_cpu - is a given cpu idle currently?
4517 * @cpu: the processor in question.
4519 int idle_cpu(int cpu)
4521 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
4525 * idle_task - return the idle task for a given cpu.
4526 * @cpu: the processor in question.
4528 struct task_struct *idle_task(int cpu)
4530 return cpu_rq(cpu)->idle;
4534 * find_process_by_pid - find a process with a matching PID value.
4535 * @pid: the pid in question.
4537 static struct task_struct *find_process_by_pid(pid_t pid)
4539 return pid ? find_task_by_vpid(pid) : current;
4542 /* Actually do priority change: must hold rq lock. */
4544 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
4546 BUG_ON(p->se.on_rq);
4549 p->rt_priority = prio;
4550 p->normal_prio = normal_prio(p);
4551 /* we are holding p->pi_lock already */
4552 p->prio = rt_mutex_getprio(p);
4553 if (rt_prio(p->prio))
4554 p->sched_class = &rt_sched_class;
4556 p->sched_class = &fair_sched_class;
4561 * check the target process has a UID that matches the current process's
4563 static bool check_same_owner(struct task_struct *p)
4565 const struct cred *cred = current_cred(), *pcred;
4569 pcred = __task_cred(p);
4570 match = (cred->euid == pcred->euid ||
4571 cred->euid == pcred->uid);
4576 static int __sched_setscheduler(struct task_struct *p, int policy,
4577 struct sched_param *param, bool user)
4579 int retval, oldprio, oldpolicy = -1, on_rq, running;
4580 unsigned long flags;
4581 const struct sched_class *prev_class;
4585 /* may grab non-irq protected spin_locks */
4586 BUG_ON(in_interrupt());
4588 /* double check policy once rq lock held */
4590 reset_on_fork = p->sched_reset_on_fork;
4591 policy = oldpolicy = p->policy;
4593 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
4594 policy &= ~SCHED_RESET_ON_FORK;
4596 if (policy != SCHED_FIFO && policy != SCHED_RR &&
4597 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
4598 policy != SCHED_IDLE)
4603 * Valid priorities for SCHED_FIFO and SCHED_RR are
4604 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4605 * SCHED_BATCH and SCHED_IDLE is 0.
4607 if (param->sched_priority < 0 ||
4608 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
4609 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
4611 if (rt_policy(policy) != (param->sched_priority != 0))
4615 * Allow unprivileged RT tasks to decrease priority:
4617 if (user && !capable(CAP_SYS_NICE)) {
4618 if (rt_policy(policy)) {
4619 unsigned long rlim_rtprio =
4620 task_rlimit(p, RLIMIT_RTPRIO);
4622 /* can't set/change the rt policy */
4623 if (policy != p->policy && !rlim_rtprio)
4626 /* can't increase priority */
4627 if (param->sched_priority > p->rt_priority &&
4628 param->sched_priority > rlim_rtprio)
4632 * Like positive nice levels, dont allow tasks to
4633 * move out of SCHED_IDLE either:
4635 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
4638 /* can't change other user's priorities */
4639 if (!check_same_owner(p))
4642 /* Normal users shall not reset the sched_reset_on_fork flag */
4643 if (p->sched_reset_on_fork && !reset_on_fork)
4648 retval = security_task_setscheduler(p, policy, param);
4654 * make sure no PI-waiters arrive (or leave) while we are
4655 * changing the priority of the task:
4657 raw_spin_lock_irqsave(&p->pi_lock, flags);
4659 * To be able to change p->policy safely, the apropriate
4660 * runqueue lock must be held.
4662 rq = __task_rq_lock(p);
4664 #ifdef CONFIG_RT_GROUP_SCHED
4667 * Do not allow realtime tasks into groups that have no runtime
4670 if (rt_bandwidth_enabled() && rt_policy(policy) &&
4671 task_group(p)->rt_bandwidth.rt_runtime == 0) {
4672 __task_rq_unlock(rq);
4673 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4679 /* recheck policy now with rq lock held */
4680 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
4681 policy = oldpolicy = -1;
4682 __task_rq_unlock(rq);
4683 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4686 on_rq = p->se.on_rq;
4687 running = task_current(rq, p);
4689 deactivate_task(rq, p, 0);
4691 p->sched_class->put_prev_task(rq, p);
4693 p->sched_reset_on_fork = reset_on_fork;
4696 prev_class = p->sched_class;
4697 __setscheduler(rq, p, policy, param->sched_priority);
4700 p->sched_class->set_curr_task(rq);
4702 activate_task(rq, p, 0);
4704 check_class_changed(rq, p, prev_class, oldprio, running);
4706 __task_rq_unlock(rq);
4707 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
4709 rt_mutex_adjust_pi(p);
4715 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4716 * @p: the task in question.
4717 * @policy: new policy.
4718 * @param: structure containing the new RT priority.
4720 * NOTE that the task may be already dead.
4722 int sched_setscheduler(struct task_struct *p, int policy,
4723 struct sched_param *param)
4725 return __sched_setscheduler(p, policy, param, true);
4727 EXPORT_SYMBOL_GPL(sched_setscheduler);
4730 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4731 * @p: the task in question.
4732 * @policy: new policy.
4733 * @param: structure containing the new RT priority.
4735 * Just like sched_setscheduler, only don't bother checking if the
4736 * current context has permission. For example, this is needed in
4737 * stop_machine(): we create temporary high priority worker threads,
4738 * but our caller might not have that capability.
4740 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
4741 struct sched_param *param)
4743 return __sched_setscheduler(p, policy, param, false);
4747 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
4749 struct sched_param lparam;
4750 struct task_struct *p;
4753 if (!param || pid < 0)
4755 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
4760 p = find_process_by_pid(pid);
4762 retval = sched_setscheduler(p, policy, &lparam);
4769 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4770 * @pid: the pid in question.
4771 * @policy: new policy.
4772 * @param: structure containing the new RT priority.
4774 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
4775 struct sched_param __user *, param)
4777 /* negative values for policy are not valid */
4781 return do_sched_setscheduler(pid, policy, param);
4785 * sys_sched_setparam - set/change the RT priority of a thread
4786 * @pid: the pid in question.
4787 * @param: structure containing the new RT priority.
4789 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
4791 return do_sched_setscheduler(pid, -1, param);
4795 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4796 * @pid: the pid in question.
4798 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
4800 struct task_struct *p;
4808 p = find_process_by_pid(pid);
4810 retval = security_task_getscheduler(p);
4813 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
4820 * sys_sched_getparam - get the RT priority of a thread
4821 * @pid: the pid in question.
4822 * @param: structure containing the RT priority.
4824 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
4826 struct sched_param lp;
4827 struct task_struct *p;
4830 if (!param || pid < 0)
4834 p = find_process_by_pid(pid);
4839 retval = security_task_getscheduler(p);
4843 lp.sched_priority = p->rt_priority;
4847 * This one might sleep, we cannot do it with a spinlock held ...
4849 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
4858 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
4860 cpumask_var_t cpus_allowed, new_mask;
4861 struct task_struct *p;
4867 p = find_process_by_pid(pid);
4874 /* Prevent p going away */
4878 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
4882 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
4884 goto out_free_cpus_allowed;
4887 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
4890 retval = security_task_setscheduler(p, 0, NULL);
4894 cpuset_cpus_allowed(p, cpus_allowed);
4895 cpumask_and(new_mask, in_mask, cpus_allowed);
4897 retval = set_cpus_allowed_ptr(p, new_mask);
4900 cpuset_cpus_allowed(p, cpus_allowed);
4901 if (!cpumask_subset(new_mask, cpus_allowed)) {
4903 * We must have raced with a concurrent cpuset
4904 * update. Just reset the cpus_allowed to the
4905 * cpuset's cpus_allowed
4907 cpumask_copy(new_mask, cpus_allowed);
4912 free_cpumask_var(new_mask);
4913 out_free_cpus_allowed:
4914 free_cpumask_var(cpus_allowed);
4921 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
4922 struct cpumask *new_mask)
4924 if (len < cpumask_size())
4925 cpumask_clear(new_mask);
4926 else if (len > cpumask_size())
4927 len = cpumask_size();
4929 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
4933 * sys_sched_setaffinity - set the cpu affinity of a process
4934 * @pid: pid of the process
4935 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4936 * @user_mask_ptr: user-space pointer to the new cpu mask
4938 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
4939 unsigned long __user *, user_mask_ptr)
4941 cpumask_var_t new_mask;
4944 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
4947 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
4949 retval = sched_setaffinity(pid, new_mask);
4950 free_cpumask_var(new_mask);
4954 long sched_getaffinity(pid_t pid, struct cpumask *mask)
4956 struct task_struct *p;
4957 unsigned long flags;
4965 p = find_process_by_pid(pid);
4969 retval = security_task_getscheduler(p);
4973 rq = task_rq_lock(p, &flags);
4974 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
4975 task_rq_unlock(rq, &flags);
4985 * sys_sched_getaffinity - get the cpu affinity of a process
4986 * @pid: pid of the process
4987 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4988 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4990 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
4991 unsigned long __user *, user_mask_ptr)
4996 if ((len * BITS_PER_BYTE) < nr_cpu_ids)
4998 if (len & (sizeof(unsigned long)-1))
5001 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5004 ret = sched_getaffinity(pid, mask);
5006 size_t retlen = min_t(size_t, len, cpumask_size());
5008 if (copy_to_user(user_mask_ptr, mask, retlen))
5013 free_cpumask_var(mask);
5019 * sys_sched_yield - yield the current processor to other threads.
5021 * This function yields the current CPU to other tasks. If there are no
5022 * other threads running on this CPU then this function will return.
5024 SYSCALL_DEFINE0(sched_yield)
5026 struct rq *rq = this_rq_lock();
5028 schedstat_inc(rq, yld_count);
5029 current->sched_class->yield_task(rq);
5032 * Since we are going to call schedule() anyway, there's
5033 * no need to preempt or enable interrupts:
5035 __release(rq->lock);
5036 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5037 do_raw_spin_unlock(&rq->lock);
5038 preempt_enable_no_resched();
5045 static inline int should_resched(void)
5047 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
5050 static void __cond_resched(void)
5052 add_preempt_count(PREEMPT_ACTIVE);
5054 sub_preempt_count(PREEMPT_ACTIVE);
5057 int __sched _cond_resched(void)
5059 if (should_resched()) {
5065 EXPORT_SYMBOL(_cond_resched);
5068 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5069 * call schedule, and on return reacquire the lock.
5071 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5072 * operations here to prevent schedule() from being called twice (once via
5073 * spin_unlock(), once by hand).
5075 int __cond_resched_lock(spinlock_t *lock)
5077 int resched = should_resched();
5080 lockdep_assert_held(lock);
5082 if (spin_needbreak(lock) || resched) {
5093 EXPORT_SYMBOL(__cond_resched_lock);
5095 int __sched __cond_resched_softirq(void)
5097 BUG_ON(!in_softirq());
5099 if (should_resched()) {
5107 EXPORT_SYMBOL(__cond_resched_softirq);
5110 * yield - yield the current processor to other threads.
5112 * This is a shortcut for kernel-space yielding - it marks the
5113 * thread runnable and calls sys_sched_yield().
5115 void __sched yield(void)
5117 set_current_state(TASK_RUNNING);
5120 EXPORT_SYMBOL(yield);
5123 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5124 * that process accounting knows that this is a task in IO wait state.
5126 void __sched io_schedule(void)
5128 struct rq *rq = raw_rq();
5130 delayacct_blkio_start();
5131 atomic_inc(&rq->nr_iowait);
5132 current->in_iowait = 1;
5134 current->in_iowait = 0;
5135 atomic_dec(&rq->nr_iowait);
5136 delayacct_blkio_end();
5138 EXPORT_SYMBOL(io_schedule);
5140 long __sched io_schedule_timeout(long timeout)
5142 struct rq *rq = raw_rq();
5145 delayacct_blkio_start();
5146 atomic_inc(&rq->nr_iowait);
5147 current->in_iowait = 1;
5148 ret = schedule_timeout(timeout);
5149 current->in_iowait = 0;
5150 atomic_dec(&rq->nr_iowait);
5151 delayacct_blkio_end();
5156 * sys_sched_get_priority_max - return maximum RT priority.
5157 * @policy: scheduling class.
5159 * this syscall returns the maximum rt_priority that can be used
5160 * by a given scheduling class.
5162 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5169 ret = MAX_USER_RT_PRIO-1;
5181 * sys_sched_get_priority_min - return minimum RT priority.
5182 * @policy: scheduling class.
5184 * this syscall returns the minimum rt_priority that can be used
5185 * by a given scheduling class.
5187 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5205 * sys_sched_rr_get_interval - return the default timeslice of a process.
5206 * @pid: pid of the process.
5207 * @interval: userspace pointer to the timeslice value.
5209 * this syscall writes the default timeslice value of a given process
5210 * into the user-space timespec buffer. A value of '0' means infinity.
5212 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5213 struct timespec __user *, interval)
5215 struct task_struct *p;
5216 unsigned int time_slice;
5217 unsigned long flags;
5227 p = find_process_by_pid(pid);
5231 retval = security_task_getscheduler(p);
5235 rq = task_rq_lock(p, &flags);
5236 time_slice = p->sched_class->get_rr_interval(rq, p);
5237 task_rq_unlock(rq, &flags);
5240 jiffies_to_timespec(time_slice, &t);
5241 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5249 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5251 void sched_show_task(struct task_struct *p)
5253 unsigned long free = 0;
5256 state = p->state ? __ffs(p->state) + 1 : 0;
5257 printk(KERN_INFO "%-13.13s %c", p->comm,
5258 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5259 #if BITS_PER_LONG == 32
5260 if (state == TASK_RUNNING)
5261 printk(KERN_CONT " running ");
5263 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5265 if (state == TASK_RUNNING)
5266 printk(KERN_CONT " running task ");
5268 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5270 #ifdef CONFIG_DEBUG_STACK_USAGE
5271 free = stack_not_used(p);
5273 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
5274 task_pid_nr(p), task_pid_nr(p->real_parent),
5275 (unsigned long)task_thread_info(p)->flags);
5277 show_stack(p, NULL);
5280 void show_state_filter(unsigned long state_filter)
5282 struct task_struct *g, *p;
5284 #if BITS_PER_LONG == 32
5286 " task PC stack pid father\n");
5289 " task PC stack pid father\n");
5291 read_lock(&tasklist_lock);
5292 do_each_thread(g, p) {
5294 * reset the NMI-timeout, listing all files on a slow
5295 * console might take alot of time:
5297 touch_nmi_watchdog();
5298 if (!state_filter || (p->state & state_filter))
5300 } while_each_thread(g, p);
5302 touch_all_softlockup_watchdogs();
5304 #ifdef CONFIG_SCHED_DEBUG
5305 sysrq_sched_debug_show();
5307 read_unlock(&tasklist_lock);
5309 * Only show locks if all tasks are dumped:
5312 debug_show_all_locks();
5315 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5317 idle->sched_class = &idle_sched_class;
5321 * init_idle - set up an idle thread for a given CPU
5322 * @idle: task in question
5323 * @cpu: cpu the idle task belongs to
5325 * NOTE: this function does not set the idle thread's NEED_RESCHED
5326 * flag, to make booting more robust.
5328 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5330 struct rq *rq = cpu_rq(cpu);
5331 unsigned long flags;
5333 raw_spin_lock_irqsave(&rq->lock, flags);
5336 idle->state = TASK_RUNNING;
5337 idle->se.exec_start = sched_clock();
5339 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5341 * We're having a chicken and egg problem, even though we are
5342 * holding rq->lock, the cpu isn't yet set to this cpu so the
5343 * lockdep check in task_group() will fail.
5345 * Similar case to sched_fork(). / Alternatively we could
5346 * use task_rq_lock() here and obtain the other rq->lock.
5351 __set_task_cpu(idle, cpu);
5354 rq->curr = rq->idle = idle;
5355 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5358 raw_spin_unlock_irqrestore(&rq->lock, flags);
5360 /* Set the preempt count _outside_ the spinlocks! */
5361 #if defined(CONFIG_PREEMPT)
5362 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5364 task_thread_info(idle)->preempt_count = 0;
5367 * The idle tasks have their own, simple scheduling class:
5369 idle->sched_class = &idle_sched_class;
5370 ftrace_graph_init_task(idle);
5374 * In a system that switches off the HZ timer nohz_cpu_mask
5375 * indicates which cpus entered this state. This is used
5376 * in the rcu update to wait only for active cpus. For system
5377 * which do not switch off the HZ timer nohz_cpu_mask should
5378 * always be CPU_BITS_NONE.
5380 cpumask_var_t nohz_cpu_mask;
5383 * Increase the granularity value when there are more CPUs,
5384 * because with more CPUs the 'effective latency' as visible
5385 * to users decreases. But the relationship is not linear,
5386 * so pick a second-best guess by going with the log2 of the
5389 * This idea comes from the SD scheduler of Con Kolivas:
5391 static int get_update_sysctl_factor(void)
5393 unsigned int cpus = min_t(int, num_online_cpus(), 8);
5394 unsigned int factor;
5396 switch (sysctl_sched_tunable_scaling) {
5397 case SCHED_TUNABLESCALING_NONE:
5400 case SCHED_TUNABLESCALING_LINEAR:
5403 case SCHED_TUNABLESCALING_LOG:
5405 factor = 1 + ilog2(cpus);
5412 static void update_sysctl(void)
5414 unsigned int factor = get_update_sysctl_factor();
5416 #define SET_SYSCTL(name) \
5417 (sysctl_##name = (factor) * normalized_sysctl_##name)
5418 SET_SYSCTL(sched_min_granularity);
5419 SET_SYSCTL(sched_latency);
5420 SET_SYSCTL(sched_wakeup_granularity);
5421 SET_SYSCTL(sched_shares_ratelimit);
5425 static inline void sched_init_granularity(void)
5432 * This is how migration works:
5434 * 1) we invoke migration_cpu_stop() on the target CPU using
5436 * 2) stopper starts to run (implicitly forcing the migrated thread
5438 * 3) it checks whether the migrated task is still in the wrong runqueue.
5439 * 4) if it's in the wrong runqueue then the migration thread removes
5440 * it and puts it into the right queue.
5441 * 5) stopper completes and stop_one_cpu() returns and the migration
5446 * Change a given task's CPU affinity. Migrate the thread to a
5447 * proper CPU and schedule it away if the CPU it's executing on
5448 * is removed from the allowed bitmask.
5450 * NOTE: the caller must have a valid reference to the task, the
5451 * task must not exit() & deallocate itself prematurely. The
5452 * call is not atomic; no spinlocks may be held.
5454 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
5456 unsigned long flags;
5458 unsigned int dest_cpu;
5462 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5463 * drop the rq->lock and still rely on ->cpus_allowed.
5466 while (task_is_waking(p))
5468 rq = task_rq_lock(p, &flags);
5469 if (task_is_waking(p)) {
5470 task_rq_unlock(rq, &flags);
5474 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
5479 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5480 !cpumask_equal(&p->cpus_allowed, new_mask))) {
5485 if (p->sched_class->set_cpus_allowed)
5486 p->sched_class->set_cpus_allowed(p, new_mask);
5488 cpumask_copy(&p->cpus_allowed, new_mask);
5489 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
5492 /* Can the task run on the task's current CPU? If so, we're done */
5493 if (cpumask_test_cpu(task_cpu(p), new_mask))
5496 dest_cpu = cpumask_any_and(cpu_active_mask, new_mask);
5497 if (migrate_task(p, dest_cpu)) {
5498 struct migration_arg arg = { p, dest_cpu };
5499 /* Need help from migration thread: drop lock and wait. */
5500 task_rq_unlock(rq, &flags);
5501 stop_one_cpu(cpu_of(rq), migration_cpu_stop, &arg);
5502 tlb_migrate_finish(p->mm);
5506 task_rq_unlock(rq, &flags);
5510 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5513 * Move (not current) task off this cpu, onto dest cpu. We're doing
5514 * this because either it can't run here any more (set_cpus_allowed()
5515 * away from this CPU, or CPU going down), or because we're
5516 * attempting to rebalance this task on exec (sched_exec).
5518 * So we race with normal scheduler movements, but that's OK, as long
5519 * as the task is no longer on this CPU.
5521 * Returns non-zero if task was successfully migrated.
5523 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5525 struct rq *rq_dest, *rq_src;
5528 if (unlikely(!cpu_active(dest_cpu)))
5531 rq_src = cpu_rq(src_cpu);
5532 rq_dest = cpu_rq(dest_cpu);
5534 double_rq_lock(rq_src, rq_dest);
5535 /* Already moved. */
5536 if (task_cpu(p) != src_cpu)
5538 /* Affinity changed (again). */
5539 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
5543 * If we're not on a rq, the next wake-up will ensure we're
5547 deactivate_task(rq_src, p, 0);
5548 set_task_cpu(p, dest_cpu);
5549 activate_task(rq_dest, p, 0);
5550 check_preempt_curr(rq_dest, p, 0);
5555 double_rq_unlock(rq_src, rq_dest);
5560 * migration_cpu_stop - this will be executed by a highprio stopper thread
5561 * and performs thread migration by bumping thread off CPU then
5562 * 'pushing' onto another runqueue.
5564 static int migration_cpu_stop(void *data)
5566 struct migration_arg *arg = data;
5569 * The original target cpu might have gone down and we might
5570 * be on another cpu but it doesn't matter.
5572 local_irq_disable();
5573 __migrate_task(arg->task, raw_smp_processor_id(), arg->dest_cpu);
5578 #ifdef CONFIG_HOTPLUG_CPU
5580 * Figure out where task on dead CPU should go, use force if necessary.
5582 void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
5584 struct rq *rq = cpu_rq(dead_cpu);
5585 int needs_cpu, uninitialized_var(dest_cpu);
5586 unsigned long flags;
5588 local_irq_save(flags);
5590 raw_spin_lock(&rq->lock);
5591 needs_cpu = (task_cpu(p) == dead_cpu) && (p->state != TASK_WAKING);
5593 dest_cpu = select_fallback_rq(dead_cpu, p);
5594 raw_spin_unlock(&rq->lock);
5596 * It can only fail if we race with set_cpus_allowed(),
5597 * in the racer should migrate the task anyway.
5600 __migrate_task(p, dead_cpu, dest_cpu);
5601 local_irq_restore(flags);
5605 * While a dead CPU has no uninterruptible tasks queued at this point,
5606 * it might still have a nonzero ->nr_uninterruptible counter, because
5607 * for performance reasons the counter is not stricly tracking tasks to
5608 * their home CPUs. So we just add the counter to another CPU's counter,
5609 * to keep the global sum constant after CPU-down:
5611 static void migrate_nr_uninterruptible(struct rq *rq_src)
5613 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
5614 unsigned long flags;
5616 local_irq_save(flags);
5617 double_rq_lock(rq_src, rq_dest);
5618 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
5619 rq_src->nr_uninterruptible = 0;
5620 double_rq_unlock(rq_src, rq_dest);
5621 local_irq_restore(flags);
5624 /* Run through task list and migrate tasks from the dead cpu. */
5625 static void migrate_live_tasks(int src_cpu)
5627 struct task_struct *p, *t;
5629 read_lock(&tasklist_lock);
5631 do_each_thread(t, p) {
5635 if (task_cpu(p) == src_cpu)
5636 move_task_off_dead_cpu(src_cpu, p);
5637 } while_each_thread(t, p);
5639 read_unlock(&tasklist_lock);
5643 * Schedules idle task to be the next runnable task on current CPU.
5644 * It does so by boosting its priority to highest possible.
5645 * Used by CPU offline code.
5647 void sched_idle_next(void)
5649 int this_cpu = smp_processor_id();
5650 struct rq *rq = cpu_rq(this_cpu);
5651 struct task_struct *p = rq->idle;
5652 unsigned long flags;
5654 /* cpu has to be offline */
5655 BUG_ON(cpu_online(this_cpu));
5658 * Strictly not necessary since rest of the CPUs are stopped by now
5659 * and interrupts disabled on the current cpu.
5661 raw_spin_lock_irqsave(&rq->lock, flags);
5663 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
5665 activate_task(rq, p, 0);
5667 raw_spin_unlock_irqrestore(&rq->lock, flags);
5671 * Ensures that the idle task is using init_mm right before its cpu goes
5674 void idle_task_exit(void)
5676 struct mm_struct *mm = current->active_mm;
5678 BUG_ON(cpu_online(smp_processor_id()));
5681 switch_mm(mm, &init_mm, current);
5685 /* called under rq->lock with disabled interrupts */
5686 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
5688 struct rq *rq = cpu_rq(dead_cpu);
5690 /* Must be exiting, otherwise would be on tasklist. */
5691 BUG_ON(!p->exit_state);
5693 /* Cannot have done final schedule yet: would have vanished. */
5694 BUG_ON(p->state == TASK_DEAD);
5699 * Drop lock around migration; if someone else moves it,
5700 * that's OK. No task can be added to this CPU, so iteration is
5703 raw_spin_unlock_irq(&rq->lock);
5704 move_task_off_dead_cpu(dead_cpu, p);
5705 raw_spin_lock_irq(&rq->lock);
5710 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5711 static void migrate_dead_tasks(unsigned int dead_cpu)
5713 struct rq *rq = cpu_rq(dead_cpu);
5714 struct task_struct *next;
5717 if (!rq->nr_running)
5719 next = pick_next_task(rq);
5722 next->sched_class->put_prev_task(rq, next);
5723 migrate_dead(dead_cpu, next);
5729 * remove the tasks which were accounted by rq from calc_load_tasks.
5731 static void calc_global_load_remove(struct rq *rq)
5733 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
5734 rq->calc_load_active = 0;
5736 #endif /* CONFIG_HOTPLUG_CPU */
5738 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5740 static struct ctl_table sd_ctl_dir[] = {
5742 .procname = "sched_domain",
5748 static struct ctl_table sd_ctl_root[] = {
5750 .procname = "kernel",
5752 .child = sd_ctl_dir,
5757 static struct ctl_table *sd_alloc_ctl_entry(int n)
5759 struct ctl_table *entry =
5760 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
5765 static void sd_free_ctl_entry(struct ctl_table **tablep)
5767 struct ctl_table *entry;
5770 * In the intermediate directories, both the child directory and
5771 * procname are dynamically allocated and could fail but the mode
5772 * will always be set. In the lowest directory the names are
5773 * static strings and all have proc handlers.
5775 for (entry = *tablep; entry->mode; entry++) {
5777 sd_free_ctl_entry(&entry->child);
5778 if (entry->proc_handler == NULL)
5779 kfree(entry->procname);
5787 set_table_entry(struct ctl_table *entry,
5788 const char *procname, void *data, int maxlen,
5789 mode_t mode, proc_handler *proc_handler)
5791 entry->procname = procname;
5793 entry->maxlen = maxlen;
5795 entry->proc_handler = proc_handler;
5798 static struct ctl_table *
5799 sd_alloc_ctl_domain_table(struct sched_domain *sd)
5801 struct ctl_table *table = sd_alloc_ctl_entry(13);
5806 set_table_entry(&table[0], "min_interval", &sd->min_interval,
5807 sizeof(long), 0644, proc_doulongvec_minmax);
5808 set_table_entry(&table[1], "max_interval", &sd->max_interval,
5809 sizeof(long), 0644, proc_doulongvec_minmax);
5810 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
5811 sizeof(int), 0644, proc_dointvec_minmax);
5812 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
5813 sizeof(int), 0644, proc_dointvec_minmax);
5814 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
5815 sizeof(int), 0644, proc_dointvec_minmax);
5816 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
5817 sizeof(int), 0644, proc_dointvec_minmax);
5818 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
5819 sizeof(int), 0644, proc_dointvec_minmax);
5820 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
5821 sizeof(int), 0644, proc_dointvec_minmax);
5822 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
5823 sizeof(int), 0644, proc_dointvec_minmax);
5824 set_table_entry(&table[9], "cache_nice_tries",
5825 &sd->cache_nice_tries,
5826 sizeof(int), 0644, proc_dointvec_minmax);
5827 set_table_entry(&table[10], "flags", &sd->flags,
5828 sizeof(int), 0644, proc_dointvec_minmax);
5829 set_table_entry(&table[11], "name", sd->name,
5830 CORENAME_MAX_SIZE, 0444, proc_dostring);
5831 /* &table[12] is terminator */
5836 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
5838 struct ctl_table *entry, *table;
5839 struct sched_domain *sd;
5840 int domain_num = 0, i;
5843 for_each_domain(cpu, sd)
5845 entry = table = sd_alloc_ctl_entry(domain_num + 1);
5850 for_each_domain(cpu, sd) {
5851 snprintf(buf, 32, "domain%d", i);
5852 entry->procname = kstrdup(buf, GFP_KERNEL);
5854 entry->child = sd_alloc_ctl_domain_table(sd);
5861 static struct ctl_table_header *sd_sysctl_header;
5862 static void register_sched_domain_sysctl(void)
5864 int i, cpu_num = num_possible_cpus();
5865 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
5868 WARN_ON(sd_ctl_dir[0].child);
5869 sd_ctl_dir[0].child = entry;
5874 for_each_possible_cpu(i) {
5875 snprintf(buf, 32, "cpu%d", i);
5876 entry->procname = kstrdup(buf, GFP_KERNEL);
5878 entry->child = sd_alloc_ctl_cpu_table(i);
5882 WARN_ON(sd_sysctl_header);
5883 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
5886 /* may be called multiple times per register */
5887 static void unregister_sched_domain_sysctl(void)
5889 if (sd_sysctl_header)
5890 unregister_sysctl_table(sd_sysctl_header);
5891 sd_sysctl_header = NULL;
5892 if (sd_ctl_dir[0].child)
5893 sd_free_ctl_entry(&sd_ctl_dir[0].child);
5896 static void register_sched_domain_sysctl(void)
5899 static void unregister_sched_domain_sysctl(void)
5904 static void set_rq_online(struct rq *rq)
5907 const struct sched_class *class;
5909 cpumask_set_cpu(rq->cpu, rq->rd->online);
5912 for_each_class(class) {
5913 if (class->rq_online)
5914 class->rq_online(rq);
5919 static void set_rq_offline(struct rq *rq)
5922 const struct sched_class *class;
5924 for_each_class(class) {
5925 if (class->rq_offline)
5926 class->rq_offline(rq);
5929 cpumask_clear_cpu(rq->cpu, rq->rd->online);
5935 * migration_call - callback that gets triggered when a CPU is added.
5936 * Here we can start up the necessary migration thread for the new CPU.
5938 static int __cpuinit
5939 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
5941 int cpu = (long)hcpu;
5942 unsigned long flags;
5943 struct rq *rq = cpu_rq(cpu);
5947 case CPU_UP_PREPARE:
5948 case CPU_UP_PREPARE_FROZEN:
5949 rq->calc_load_update = calc_load_update;
5953 case CPU_ONLINE_FROZEN:
5954 /* Update our root-domain */
5955 raw_spin_lock_irqsave(&rq->lock, flags);
5957 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5961 raw_spin_unlock_irqrestore(&rq->lock, flags);
5964 #ifdef CONFIG_HOTPLUG_CPU
5966 case CPU_DEAD_FROZEN:
5967 migrate_live_tasks(cpu);
5968 /* Idle task back to normal (off runqueue, low prio) */
5969 raw_spin_lock_irq(&rq->lock);
5970 deactivate_task(rq, rq->idle, 0);
5971 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
5972 rq->idle->sched_class = &idle_sched_class;
5973 migrate_dead_tasks(cpu);
5974 raw_spin_unlock_irq(&rq->lock);
5975 migrate_nr_uninterruptible(rq);
5976 BUG_ON(rq->nr_running != 0);
5977 calc_global_load_remove(rq);
5981 case CPU_DYING_FROZEN:
5982 /* Update our root-domain */
5983 raw_spin_lock_irqsave(&rq->lock, flags);
5985 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
5988 raw_spin_unlock_irqrestore(&rq->lock, flags);
5996 * Register at high priority so that task migration (migrate_all_tasks)
5997 * happens before everything else. This has to be lower priority than
5998 * the notifier in the perf_event subsystem, though.
6000 static struct notifier_block __cpuinitdata migration_notifier = {
6001 .notifier_call = migration_call,
6002 .priority = CPU_PRI_MIGRATION,
6005 static int __cpuinit sched_cpu_active(struct notifier_block *nfb,
6006 unsigned long action, void *hcpu)
6008 switch (action & ~CPU_TASKS_FROZEN) {
6010 case CPU_DOWN_FAILED:
6011 set_cpu_active((long)hcpu, true);
6018 static int __cpuinit sched_cpu_inactive(struct notifier_block *nfb,
6019 unsigned long action, void *hcpu)
6021 switch (action & ~CPU_TASKS_FROZEN) {
6022 case CPU_DOWN_PREPARE:
6023 set_cpu_active((long)hcpu, false);
6030 static int __init migration_init(void)
6032 void *cpu = (void *)(long)smp_processor_id();
6035 /* Initialize migration for the boot CPU */
6036 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6037 BUG_ON(err == NOTIFY_BAD);
6038 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6039 register_cpu_notifier(&migration_notifier);
6041 /* Register cpu active notifiers */
6042 cpu_notifier(sched_cpu_active, CPU_PRI_SCHED_ACTIVE);
6043 cpu_notifier(sched_cpu_inactive, CPU_PRI_SCHED_INACTIVE);
6047 early_initcall(migration_init);
6052 #ifdef CONFIG_SCHED_DEBUG
6054 static __read_mostly int sched_domain_debug_enabled;
6056 static int __init sched_domain_debug_setup(char *str)
6058 sched_domain_debug_enabled = 1;
6062 early_param("sched_debug", sched_domain_debug_setup);
6064 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6065 struct cpumask *groupmask)
6067 struct sched_group *group = sd->groups;
6070 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6071 cpumask_clear(groupmask);
6073 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6075 if (!(sd->flags & SD_LOAD_BALANCE)) {
6076 printk("does not load-balance\n");
6078 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6083 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6085 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6086 printk(KERN_ERR "ERROR: domain->span does not contain "
6089 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6090 printk(KERN_ERR "ERROR: domain->groups does not contain"
6094 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6098 printk(KERN_ERR "ERROR: group is NULL\n");
6102 if (!group->cpu_power) {
6103 printk(KERN_CONT "\n");
6104 printk(KERN_ERR "ERROR: domain->cpu_power not "
6109 if (!cpumask_weight(sched_group_cpus(group))) {
6110 printk(KERN_CONT "\n");
6111 printk(KERN_ERR "ERROR: empty group\n");
6115 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6116 printk(KERN_CONT "\n");
6117 printk(KERN_ERR "ERROR: repeated CPUs\n");
6121 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6123 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6125 printk(KERN_CONT " %s", str);
6126 if (group->cpu_power != SCHED_LOAD_SCALE) {
6127 printk(KERN_CONT " (cpu_power = %d)",
6131 group = group->next;
6132 } while (group != sd->groups);
6133 printk(KERN_CONT "\n");
6135 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6136 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6139 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6140 printk(KERN_ERR "ERROR: parent span is not a superset "
6141 "of domain->span\n");
6145 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6147 cpumask_var_t groupmask;
6150 if (!sched_domain_debug_enabled)
6154 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6158 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6160 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6161 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6166 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6173 free_cpumask_var(groupmask);
6175 #else /* !CONFIG_SCHED_DEBUG */
6176 # define sched_domain_debug(sd, cpu) do { } while (0)
6177 #endif /* CONFIG_SCHED_DEBUG */
6179 static int sd_degenerate(struct sched_domain *sd)
6181 if (cpumask_weight(sched_domain_span(sd)) == 1)
6184 /* Following flags need at least 2 groups */
6185 if (sd->flags & (SD_LOAD_BALANCE |
6186 SD_BALANCE_NEWIDLE |
6190 SD_SHARE_PKG_RESOURCES)) {
6191 if (sd->groups != sd->groups->next)
6195 /* Following flags don't use groups */
6196 if (sd->flags & (SD_WAKE_AFFINE))
6203 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6205 unsigned long cflags = sd->flags, pflags = parent->flags;
6207 if (sd_degenerate(parent))
6210 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6213 /* Flags needing groups don't count if only 1 group in parent */
6214 if (parent->groups == parent->groups->next) {
6215 pflags &= ~(SD_LOAD_BALANCE |
6216 SD_BALANCE_NEWIDLE |
6220 SD_SHARE_PKG_RESOURCES);
6221 if (nr_node_ids == 1)
6222 pflags &= ~SD_SERIALIZE;
6224 if (~cflags & pflags)
6230 static void free_rootdomain(struct root_domain *rd)
6232 synchronize_sched();
6234 cpupri_cleanup(&rd->cpupri);
6236 free_cpumask_var(rd->rto_mask);
6237 free_cpumask_var(rd->online);
6238 free_cpumask_var(rd->span);
6242 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6244 struct root_domain *old_rd = NULL;
6245 unsigned long flags;
6247 raw_spin_lock_irqsave(&rq->lock, flags);
6252 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6255 cpumask_clear_cpu(rq->cpu, old_rd->span);
6258 * If we dont want to free the old_rt yet then
6259 * set old_rd to NULL to skip the freeing later
6262 if (!atomic_dec_and_test(&old_rd->refcount))
6266 atomic_inc(&rd->refcount);
6269 cpumask_set_cpu(rq->cpu, rd->span);
6270 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
6273 raw_spin_unlock_irqrestore(&rq->lock, flags);
6276 free_rootdomain(old_rd);
6279 static int init_rootdomain(struct root_domain *rd)
6281 memset(rd, 0, sizeof(*rd));
6283 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6285 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6287 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6290 if (cpupri_init(&rd->cpupri) != 0)
6295 free_cpumask_var(rd->rto_mask);
6297 free_cpumask_var(rd->online);
6299 free_cpumask_var(rd->span);
6304 static void init_defrootdomain(void)
6306 init_rootdomain(&def_root_domain);
6308 atomic_set(&def_root_domain.refcount, 1);
6311 static struct root_domain *alloc_rootdomain(void)
6313 struct root_domain *rd;
6315 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6319 if (init_rootdomain(rd) != 0) {
6328 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6329 * hold the hotplug lock.
6332 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6334 struct rq *rq = cpu_rq(cpu);
6335 struct sched_domain *tmp;
6337 for (tmp = sd; tmp; tmp = tmp->parent)
6338 tmp->span_weight = cpumask_weight(sched_domain_span(tmp));
6340 /* Remove the sched domains which do not contribute to scheduling. */
6341 for (tmp = sd; tmp; ) {
6342 struct sched_domain *parent = tmp->parent;
6346 if (sd_parent_degenerate(tmp, parent)) {
6347 tmp->parent = parent->parent;
6349 parent->parent->child = tmp;
6354 if (sd && sd_degenerate(sd)) {
6360 sched_domain_debug(sd, cpu);
6362 rq_attach_root(rq, rd);
6363 rcu_assign_pointer(rq->sd, sd);
6366 /* cpus with isolated domains */
6367 static cpumask_var_t cpu_isolated_map;
6369 /* Setup the mask of cpus configured for isolated domains */
6370 static int __init isolated_cpu_setup(char *str)
6372 alloc_bootmem_cpumask_var(&cpu_isolated_map);
6373 cpulist_parse(str, cpu_isolated_map);
6377 __setup("isolcpus=", isolated_cpu_setup);
6380 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6381 * to a function which identifies what group(along with sched group) a CPU
6382 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6383 * (due to the fact that we keep track of groups covered with a struct cpumask).
6385 * init_sched_build_groups will build a circular linked list of the groups
6386 * covered by the given span, and will set each group's ->cpumask correctly,
6387 * and ->cpu_power to 0.
6390 init_sched_build_groups(const struct cpumask *span,
6391 const struct cpumask *cpu_map,
6392 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
6393 struct sched_group **sg,
6394 struct cpumask *tmpmask),
6395 struct cpumask *covered, struct cpumask *tmpmask)
6397 struct sched_group *first = NULL, *last = NULL;
6400 cpumask_clear(covered);
6402 for_each_cpu(i, span) {
6403 struct sched_group *sg;
6404 int group = group_fn(i, cpu_map, &sg, tmpmask);
6407 if (cpumask_test_cpu(i, covered))
6410 cpumask_clear(sched_group_cpus(sg));
6413 for_each_cpu(j, span) {
6414 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6417 cpumask_set_cpu(j, covered);
6418 cpumask_set_cpu(j, sched_group_cpus(sg));
6429 #define SD_NODES_PER_DOMAIN 16
6434 * find_next_best_node - find the next node to include in a sched_domain
6435 * @node: node whose sched_domain we're building
6436 * @used_nodes: nodes already in the sched_domain
6438 * Find the next node to include in a given scheduling domain. Simply
6439 * finds the closest node not already in the @used_nodes map.
6441 * Should use nodemask_t.
6443 static int find_next_best_node(int node, nodemask_t *used_nodes)
6445 int i, n, val, min_val, best_node = 0;
6449 for (i = 0; i < nr_node_ids; i++) {
6450 /* Start at @node */
6451 n = (node + i) % nr_node_ids;
6453 if (!nr_cpus_node(n))
6456 /* Skip already used nodes */
6457 if (node_isset(n, *used_nodes))
6460 /* Simple min distance search */
6461 val = node_distance(node, n);
6463 if (val < min_val) {
6469 node_set(best_node, *used_nodes);
6474 * sched_domain_node_span - get a cpumask for a node's sched_domain
6475 * @node: node whose cpumask we're constructing
6476 * @span: resulting cpumask
6478 * Given a node, construct a good cpumask for its sched_domain to span. It
6479 * should be one that prevents unnecessary balancing, but also spreads tasks
6482 static void sched_domain_node_span(int node, struct cpumask *span)
6484 nodemask_t used_nodes;
6487 cpumask_clear(span);
6488 nodes_clear(used_nodes);
6490 cpumask_or(span, span, cpumask_of_node(node));
6491 node_set(node, used_nodes);
6493 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6494 int next_node = find_next_best_node(node, &used_nodes);
6496 cpumask_or(span, span, cpumask_of_node(next_node));
6499 #endif /* CONFIG_NUMA */
6501 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6504 * The cpus mask in sched_group and sched_domain hangs off the end.
6506 * ( See the the comments in include/linux/sched.h:struct sched_group
6507 * and struct sched_domain. )
6509 struct static_sched_group {
6510 struct sched_group sg;
6511 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
6514 struct static_sched_domain {
6515 struct sched_domain sd;
6516 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
6522 cpumask_var_t domainspan;
6523 cpumask_var_t covered;
6524 cpumask_var_t notcovered;
6526 cpumask_var_t nodemask;
6527 cpumask_var_t this_sibling_map;
6528 cpumask_var_t this_core_map;
6529 cpumask_var_t send_covered;
6530 cpumask_var_t tmpmask;
6531 struct sched_group **sched_group_nodes;
6532 struct root_domain *rd;
6536 sa_sched_groups = 0,
6541 sa_this_sibling_map,
6543 sa_sched_group_nodes,
6553 * SMT sched-domains:
6555 #ifdef CONFIG_SCHED_SMT
6556 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
6557 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
6560 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
6561 struct sched_group **sg, struct cpumask *unused)
6564 *sg = &per_cpu(sched_groups, cpu).sg;
6567 #endif /* CONFIG_SCHED_SMT */
6570 * multi-core sched-domains:
6572 #ifdef CONFIG_SCHED_MC
6573 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
6574 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
6575 #endif /* CONFIG_SCHED_MC */
6577 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6579 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6580 struct sched_group **sg, struct cpumask *mask)
6584 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6585 group = cpumask_first(mask);
6587 *sg = &per_cpu(sched_group_core, group).sg;
6590 #elif defined(CONFIG_SCHED_MC)
6592 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
6593 struct sched_group **sg, struct cpumask *unused)
6596 *sg = &per_cpu(sched_group_core, cpu).sg;
6601 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
6602 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
6605 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
6606 struct sched_group **sg, struct cpumask *mask)
6609 #ifdef CONFIG_SCHED_MC
6610 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
6611 group = cpumask_first(mask);
6612 #elif defined(CONFIG_SCHED_SMT)
6613 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
6614 group = cpumask_first(mask);
6619 *sg = &per_cpu(sched_group_phys, group).sg;
6625 * The init_sched_build_groups can't handle what we want to do with node
6626 * groups, so roll our own. Now each node has its own list of groups which
6627 * gets dynamically allocated.
6629 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
6630 static struct sched_group ***sched_group_nodes_bycpu;
6632 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
6633 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
6635 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
6636 struct sched_group **sg,
6637 struct cpumask *nodemask)
6641 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
6642 group = cpumask_first(nodemask);
6645 *sg = &per_cpu(sched_group_allnodes, group).sg;
6649 static void init_numa_sched_groups_power(struct sched_group *group_head)
6651 struct sched_group *sg = group_head;
6657 for_each_cpu(j, sched_group_cpus(sg)) {
6658 struct sched_domain *sd;
6660 sd = &per_cpu(phys_domains, j).sd;
6661 if (j != group_first_cpu(sd->groups)) {
6663 * Only add "power" once for each
6669 sg->cpu_power += sd->groups->cpu_power;
6672 } while (sg != group_head);
6675 static int build_numa_sched_groups(struct s_data *d,
6676 const struct cpumask *cpu_map, int num)
6678 struct sched_domain *sd;
6679 struct sched_group *sg, *prev;
6682 cpumask_clear(d->covered);
6683 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
6684 if (cpumask_empty(d->nodemask)) {
6685 d->sched_group_nodes[num] = NULL;
6689 sched_domain_node_span(num, d->domainspan);
6690 cpumask_and(d->domainspan, d->domainspan, cpu_map);
6692 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6695 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
6699 d->sched_group_nodes[num] = sg;
6701 for_each_cpu(j, d->nodemask) {
6702 sd = &per_cpu(node_domains, j).sd;
6707 cpumask_copy(sched_group_cpus(sg), d->nodemask);
6709 cpumask_or(d->covered, d->covered, d->nodemask);
6712 for (j = 0; j < nr_node_ids; j++) {
6713 n = (num + j) % nr_node_ids;
6714 cpumask_complement(d->notcovered, d->covered);
6715 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
6716 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
6717 if (cpumask_empty(d->tmpmask))
6719 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
6720 if (cpumask_empty(d->tmpmask))
6722 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
6726 "Can not alloc domain group for node %d\n", j);
6730 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
6731 sg->next = prev->next;
6732 cpumask_or(d->covered, d->covered, d->tmpmask);
6739 #endif /* CONFIG_NUMA */
6742 /* Free memory allocated for various sched_group structures */
6743 static void free_sched_groups(const struct cpumask *cpu_map,
6744 struct cpumask *nodemask)
6748 for_each_cpu(cpu, cpu_map) {
6749 struct sched_group **sched_group_nodes
6750 = sched_group_nodes_bycpu[cpu];
6752 if (!sched_group_nodes)
6755 for (i = 0; i < nr_node_ids; i++) {
6756 struct sched_group *oldsg, *sg = sched_group_nodes[i];
6758 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
6759 if (cpumask_empty(nodemask))
6769 if (oldsg != sched_group_nodes[i])
6772 kfree(sched_group_nodes);
6773 sched_group_nodes_bycpu[cpu] = NULL;
6776 #else /* !CONFIG_NUMA */
6777 static void free_sched_groups(const struct cpumask *cpu_map,
6778 struct cpumask *nodemask)
6781 #endif /* CONFIG_NUMA */
6784 * Initialize sched groups cpu_power.
6786 * cpu_power indicates the capacity of sched group, which is used while
6787 * distributing the load between different sched groups in a sched domain.
6788 * Typically cpu_power for all the groups in a sched domain will be same unless
6789 * there are asymmetries in the topology. If there are asymmetries, group
6790 * having more cpu_power will pickup more load compared to the group having
6793 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
6795 struct sched_domain *child;
6796 struct sched_group *group;
6800 WARN_ON(!sd || !sd->groups);
6802 if (cpu != group_first_cpu(sd->groups))
6807 sd->groups->cpu_power = 0;
6810 power = SCHED_LOAD_SCALE;
6811 weight = cpumask_weight(sched_domain_span(sd));
6813 * SMT siblings share the power of a single core.
6814 * Usually multiple threads get a better yield out of
6815 * that one core than a single thread would have,
6816 * reflect that in sd->smt_gain.
6818 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
6819 power *= sd->smt_gain;
6821 power >>= SCHED_LOAD_SHIFT;
6823 sd->groups->cpu_power += power;
6828 * Add cpu_power of each child group to this groups cpu_power.
6830 group = child->groups;
6832 sd->groups->cpu_power += group->cpu_power;
6833 group = group->next;
6834 } while (group != child->groups);
6838 * Initializers for schedule domains
6839 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6842 #ifdef CONFIG_SCHED_DEBUG
6843 # define SD_INIT_NAME(sd, type) sd->name = #type
6845 # define SD_INIT_NAME(sd, type) do { } while (0)
6848 #define SD_INIT(sd, type) sd_init_##type(sd)
6850 #define SD_INIT_FUNC(type) \
6851 static noinline void sd_init_##type(struct sched_domain *sd) \
6853 memset(sd, 0, sizeof(*sd)); \
6854 *sd = SD_##type##_INIT; \
6855 sd->level = SD_LV_##type; \
6856 SD_INIT_NAME(sd, type); \
6861 SD_INIT_FUNC(ALLNODES)
6864 #ifdef CONFIG_SCHED_SMT
6865 SD_INIT_FUNC(SIBLING)
6867 #ifdef CONFIG_SCHED_MC
6871 static int default_relax_domain_level = -1;
6873 static int __init setup_relax_domain_level(char *str)
6877 val = simple_strtoul(str, NULL, 0);
6878 if (val < SD_LV_MAX)
6879 default_relax_domain_level = val;
6883 __setup("relax_domain_level=", setup_relax_domain_level);
6885 static void set_domain_attribute(struct sched_domain *sd,
6886 struct sched_domain_attr *attr)
6890 if (!attr || attr->relax_domain_level < 0) {
6891 if (default_relax_domain_level < 0)
6894 request = default_relax_domain_level;
6896 request = attr->relax_domain_level;
6897 if (request < sd->level) {
6898 /* turn off idle balance on this domain */
6899 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6901 /* turn on idle balance on this domain */
6902 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
6906 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
6907 const struct cpumask *cpu_map)
6910 case sa_sched_groups:
6911 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
6912 d->sched_group_nodes = NULL;
6914 free_rootdomain(d->rd); /* fall through */
6916 free_cpumask_var(d->tmpmask); /* fall through */
6917 case sa_send_covered:
6918 free_cpumask_var(d->send_covered); /* fall through */
6919 case sa_this_core_map:
6920 free_cpumask_var(d->this_core_map); /* fall through */
6921 case sa_this_sibling_map:
6922 free_cpumask_var(d->this_sibling_map); /* fall through */
6924 free_cpumask_var(d->nodemask); /* fall through */
6925 case sa_sched_group_nodes:
6927 kfree(d->sched_group_nodes); /* fall through */
6929 free_cpumask_var(d->notcovered); /* fall through */
6931 free_cpumask_var(d->covered); /* fall through */
6933 free_cpumask_var(d->domainspan); /* fall through */
6940 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
6941 const struct cpumask *cpu_map)
6944 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
6946 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
6947 return sa_domainspan;
6948 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
6950 /* Allocate the per-node list of sched groups */
6951 d->sched_group_nodes = kcalloc(nr_node_ids,
6952 sizeof(struct sched_group *), GFP_KERNEL);
6953 if (!d->sched_group_nodes) {
6954 printk(KERN_WARNING "Can not alloc sched group node list\n");
6955 return sa_notcovered;
6957 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
6959 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
6960 return sa_sched_group_nodes;
6961 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
6963 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
6964 return sa_this_sibling_map;
6965 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
6966 return sa_this_core_map;
6967 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
6968 return sa_send_covered;
6969 d->rd = alloc_rootdomain();
6971 printk(KERN_WARNING "Cannot alloc root domain\n");
6974 return sa_rootdomain;
6977 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
6978 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
6980 struct sched_domain *sd = NULL;
6982 struct sched_domain *parent;
6985 if (cpumask_weight(cpu_map) >
6986 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
6987 sd = &per_cpu(allnodes_domains, i).sd;
6988 SD_INIT(sd, ALLNODES);
6989 set_domain_attribute(sd, attr);
6990 cpumask_copy(sched_domain_span(sd), cpu_map);
6991 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
6996 sd = &per_cpu(node_domains, i).sd;
6998 set_domain_attribute(sd, attr);
6999 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7000 sd->parent = parent;
7003 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
7008 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
7009 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7010 struct sched_domain *parent, int i)
7012 struct sched_domain *sd;
7013 sd = &per_cpu(phys_domains, i).sd;
7015 set_domain_attribute(sd, attr);
7016 cpumask_copy(sched_domain_span(sd), d->nodemask);
7017 sd->parent = parent;
7020 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
7024 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
7025 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7026 struct sched_domain *parent, int i)
7028 struct sched_domain *sd = parent;
7029 #ifdef CONFIG_SCHED_MC
7030 sd = &per_cpu(core_domains, i).sd;
7032 set_domain_attribute(sd, attr);
7033 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
7034 sd->parent = parent;
7036 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
7041 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
7042 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
7043 struct sched_domain *parent, int i)
7045 struct sched_domain *sd = parent;
7046 #ifdef CONFIG_SCHED_SMT
7047 sd = &per_cpu(cpu_domains, i).sd;
7048 SD_INIT(sd, SIBLING);
7049 set_domain_attribute(sd, attr);
7050 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
7051 sd->parent = parent;
7053 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
7058 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
7059 const struct cpumask *cpu_map, int cpu)
7062 #ifdef CONFIG_SCHED_SMT
7063 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
7064 cpumask_and(d->this_sibling_map, cpu_map,
7065 topology_thread_cpumask(cpu));
7066 if (cpu == cpumask_first(d->this_sibling_map))
7067 init_sched_build_groups(d->this_sibling_map, cpu_map,
7069 d->send_covered, d->tmpmask);
7072 #ifdef CONFIG_SCHED_MC
7073 case SD_LV_MC: /* set up multi-core groups */
7074 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
7075 if (cpu == cpumask_first(d->this_core_map))
7076 init_sched_build_groups(d->this_core_map, cpu_map,
7078 d->send_covered, d->tmpmask);
7081 case SD_LV_CPU: /* set up physical groups */
7082 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
7083 if (!cpumask_empty(d->nodemask))
7084 init_sched_build_groups(d->nodemask, cpu_map,
7086 d->send_covered, d->tmpmask);
7089 case SD_LV_ALLNODES:
7090 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
7091 d->send_covered, d->tmpmask);
7100 * Build sched domains for a given set of cpus and attach the sched domains
7101 * to the individual cpus
7103 static int __build_sched_domains(const struct cpumask *cpu_map,
7104 struct sched_domain_attr *attr)
7106 enum s_alloc alloc_state = sa_none;
7108 struct sched_domain *sd;
7114 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
7115 if (alloc_state != sa_rootdomain)
7117 alloc_state = sa_sched_groups;
7120 * Set up domains for cpus specified by the cpu_map.
7122 for_each_cpu(i, cpu_map) {
7123 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
7126 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
7127 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
7128 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
7129 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
7132 for_each_cpu(i, cpu_map) {
7133 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
7134 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
7137 /* Set up physical groups */
7138 for (i = 0; i < nr_node_ids; i++)
7139 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
7142 /* Set up node groups */
7144 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
7146 for (i = 0; i < nr_node_ids; i++)
7147 if (build_numa_sched_groups(&d, cpu_map, i))
7151 /* Calculate CPU power for physical packages and nodes */
7152 #ifdef CONFIG_SCHED_SMT
7153 for_each_cpu(i, cpu_map) {
7154 sd = &per_cpu(cpu_domains, i).sd;
7155 init_sched_groups_power(i, sd);
7158 #ifdef CONFIG_SCHED_MC
7159 for_each_cpu(i, cpu_map) {
7160 sd = &per_cpu(core_domains, i).sd;
7161 init_sched_groups_power(i, sd);
7165 for_each_cpu(i, cpu_map) {
7166 sd = &per_cpu(phys_domains, i).sd;
7167 init_sched_groups_power(i, sd);
7171 for (i = 0; i < nr_node_ids; i++)
7172 init_numa_sched_groups_power(d.sched_group_nodes[i]);
7174 if (d.sd_allnodes) {
7175 struct sched_group *sg;
7177 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7179 init_numa_sched_groups_power(sg);
7183 /* Attach the domains */
7184 for_each_cpu(i, cpu_map) {
7185 #ifdef CONFIG_SCHED_SMT
7186 sd = &per_cpu(cpu_domains, i).sd;
7187 #elif defined(CONFIG_SCHED_MC)
7188 sd = &per_cpu(core_domains, i).sd;
7190 sd = &per_cpu(phys_domains, i).sd;
7192 cpu_attach_domain(sd, d.rd, i);
7195 d.sched_group_nodes = NULL; /* don't free this we still need it */
7196 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
7200 __free_domain_allocs(&d, alloc_state, cpu_map);
7204 static int build_sched_domains(const struct cpumask *cpu_map)
7206 return __build_sched_domains(cpu_map, NULL);
7209 static cpumask_var_t *doms_cur; /* current sched domains */
7210 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7211 static struct sched_domain_attr *dattr_cur;
7212 /* attribues of custom domains in 'doms_cur' */
7215 * Special case: If a kmalloc of a doms_cur partition (array of
7216 * cpumask) fails, then fallback to a single sched domain,
7217 * as determined by the single cpumask fallback_doms.
7219 static cpumask_var_t fallback_doms;
7222 * arch_update_cpu_topology lets virtualized architectures update the
7223 * cpu core maps. It is supposed to return 1 if the topology changed
7224 * or 0 if it stayed the same.
7226 int __attribute__((weak)) arch_update_cpu_topology(void)
7231 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
7234 cpumask_var_t *doms;
7236 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
7239 for (i = 0; i < ndoms; i++) {
7240 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
7241 free_sched_domains(doms, i);
7248 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
7251 for (i = 0; i < ndoms; i++)
7252 free_cpumask_var(doms[i]);
7257 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7258 * For now this just excludes isolated cpus, but could be used to
7259 * exclude other special cases in the future.
7261 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7265 arch_update_cpu_topology();
7267 doms_cur = alloc_sched_domains(ndoms_cur);
7269 doms_cur = &fallback_doms;
7270 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
7272 err = build_sched_domains(doms_cur[0]);
7273 register_sched_domain_sysctl();
7278 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7279 struct cpumask *tmpmask)
7281 free_sched_groups(cpu_map, tmpmask);
7285 * Detach sched domains from a group of cpus specified in cpu_map
7286 * These cpus will now be attached to the NULL domain
7288 static void detach_destroy_domains(const struct cpumask *cpu_map)
7290 /* Save because hotplug lock held. */
7291 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7294 for_each_cpu(i, cpu_map)
7295 cpu_attach_domain(NULL, &def_root_domain, i);
7296 synchronize_sched();
7297 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7300 /* handle null as "default" */
7301 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7302 struct sched_domain_attr *new, int idx_new)
7304 struct sched_domain_attr tmp;
7311 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7312 new ? (new + idx_new) : &tmp,
7313 sizeof(struct sched_domain_attr));
7317 * Partition sched domains as specified by the 'ndoms_new'
7318 * cpumasks in the array doms_new[] of cpumasks. This compares
7319 * doms_new[] to the current sched domain partitioning, doms_cur[].
7320 * It destroys each deleted domain and builds each new domain.
7322 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7323 * The masks don't intersect (don't overlap.) We should setup one
7324 * sched domain for each mask. CPUs not in any of the cpumasks will
7325 * not be load balanced. If the same cpumask appears both in the
7326 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7329 * The passed in 'doms_new' should be allocated using
7330 * alloc_sched_domains. This routine takes ownership of it and will
7331 * free_sched_domains it when done with it. If the caller failed the
7332 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7333 * and partition_sched_domains() will fallback to the single partition
7334 * 'fallback_doms', it also forces the domains to be rebuilt.
7336 * If doms_new == NULL it will be replaced with cpu_online_mask.
7337 * ndoms_new == 0 is a special case for destroying existing domains,
7338 * and it will not create the default domain.
7340 * Call with hotplug lock held
7342 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
7343 struct sched_domain_attr *dattr_new)
7348 mutex_lock(&sched_domains_mutex);
7350 /* always unregister in case we don't destroy any domains */
7351 unregister_sched_domain_sysctl();
7353 /* Let architecture update cpu core mappings. */
7354 new_topology = arch_update_cpu_topology();
7356 n = doms_new ? ndoms_new : 0;
7358 /* Destroy deleted domains */
7359 for (i = 0; i < ndoms_cur; i++) {
7360 for (j = 0; j < n && !new_topology; j++) {
7361 if (cpumask_equal(doms_cur[i], doms_new[j])
7362 && dattrs_equal(dattr_cur, i, dattr_new, j))
7365 /* no match - a current sched domain not in new doms_new[] */
7366 detach_destroy_domains(doms_cur[i]);
7371 if (doms_new == NULL) {
7373 doms_new = &fallback_doms;
7374 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
7375 WARN_ON_ONCE(dattr_new);
7378 /* Build new domains */
7379 for (i = 0; i < ndoms_new; i++) {
7380 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7381 if (cpumask_equal(doms_new[i], doms_cur[j])
7382 && dattrs_equal(dattr_new, i, dattr_cur, j))
7385 /* no match - add a new doms_new */
7386 __build_sched_domains(doms_new[i],
7387 dattr_new ? dattr_new + i : NULL);
7392 /* Remember the new sched domains */
7393 if (doms_cur != &fallback_doms)
7394 free_sched_domains(doms_cur, ndoms_cur);
7395 kfree(dattr_cur); /* kfree(NULL) is safe */
7396 doms_cur = doms_new;
7397 dattr_cur = dattr_new;
7398 ndoms_cur = ndoms_new;
7400 register_sched_domain_sysctl();
7402 mutex_unlock(&sched_domains_mutex);
7405 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7406 static void arch_reinit_sched_domains(void)
7410 /* Destroy domains first to force the rebuild */
7411 partition_sched_domains(0, NULL, NULL);
7413 rebuild_sched_domains();
7417 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7419 unsigned int level = 0;
7421 if (sscanf(buf, "%u", &level) != 1)
7425 * level is always be positive so don't check for
7426 * level < POWERSAVINGS_BALANCE_NONE which is 0
7427 * What happens on 0 or 1 byte write,
7428 * need to check for count as well?
7431 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
7435 sched_smt_power_savings = level;
7437 sched_mc_power_savings = level;
7439 arch_reinit_sched_domains();
7444 #ifdef CONFIG_SCHED_MC
7445 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7446 struct sysdev_class_attribute *attr,
7449 return sprintf(page, "%u\n", sched_mc_power_savings);
7451 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7452 struct sysdev_class_attribute *attr,
7453 const char *buf, size_t count)
7455 return sched_power_savings_store(buf, count, 0);
7457 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7458 sched_mc_power_savings_show,
7459 sched_mc_power_savings_store);
7462 #ifdef CONFIG_SCHED_SMT
7463 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7464 struct sysdev_class_attribute *attr,
7467 return sprintf(page, "%u\n", sched_smt_power_savings);
7469 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7470 struct sysdev_class_attribute *attr,
7471 const char *buf, size_t count)
7473 return sched_power_savings_store(buf, count, 1);
7475 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7476 sched_smt_power_savings_show,
7477 sched_smt_power_savings_store);
7480 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7484 #ifdef CONFIG_SCHED_SMT
7486 err = sysfs_create_file(&cls->kset.kobj,
7487 &attr_sched_smt_power_savings.attr);
7489 #ifdef CONFIG_SCHED_MC
7490 if (!err && mc_capable())
7491 err = sysfs_create_file(&cls->kset.kobj,
7492 &attr_sched_mc_power_savings.attr);
7496 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7499 * Update cpusets according to cpu_active mask. If cpusets are
7500 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7501 * around partition_sched_domains().
7503 static int cpuset_cpu_active(struct notifier_block *nfb, unsigned long action,
7506 switch (action & ~CPU_TASKS_FROZEN) {
7508 case CPU_DOWN_FAILED:
7509 cpuset_update_active_cpus();
7516 static int cpuset_cpu_inactive(struct notifier_block *nfb, unsigned long action,
7519 switch (action & ~CPU_TASKS_FROZEN) {
7520 case CPU_DOWN_PREPARE:
7521 cpuset_update_active_cpus();
7528 static int update_runtime(struct notifier_block *nfb,
7529 unsigned long action, void *hcpu)
7531 int cpu = (int)(long)hcpu;
7534 case CPU_DOWN_PREPARE:
7535 case CPU_DOWN_PREPARE_FROZEN:
7536 disable_runtime(cpu_rq(cpu));
7539 case CPU_DOWN_FAILED:
7540 case CPU_DOWN_FAILED_FROZEN:
7542 case CPU_ONLINE_FROZEN:
7543 enable_runtime(cpu_rq(cpu));
7551 void __init sched_init_smp(void)
7553 cpumask_var_t non_isolated_cpus;
7555 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
7556 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
7558 #if defined(CONFIG_NUMA)
7559 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7561 BUG_ON(sched_group_nodes_bycpu == NULL);
7564 mutex_lock(&sched_domains_mutex);
7565 arch_init_sched_domains(cpu_active_mask);
7566 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
7567 if (cpumask_empty(non_isolated_cpus))
7568 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
7569 mutex_unlock(&sched_domains_mutex);
7572 hotcpu_notifier(cpuset_cpu_active, CPU_PRI_CPUSET_ACTIVE);
7573 hotcpu_notifier(cpuset_cpu_inactive, CPU_PRI_CPUSET_INACTIVE);
7575 /* RT runtime code needs to handle some hotplug events */
7576 hotcpu_notifier(update_runtime, 0);
7580 /* Move init over to a non-isolated CPU */
7581 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
7583 sched_init_granularity();
7584 free_cpumask_var(non_isolated_cpus);
7586 init_sched_rt_class();
7589 void __init sched_init_smp(void)
7591 sched_init_granularity();
7593 #endif /* CONFIG_SMP */
7595 const_debug unsigned int sysctl_timer_migration = 1;
7597 int in_sched_functions(unsigned long addr)
7599 return in_lock_functions(addr) ||
7600 (addr >= (unsigned long)__sched_text_start
7601 && addr < (unsigned long)__sched_text_end);
7604 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7606 cfs_rq->tasks_timeline = RB_ROOT;
7607 INIT_LIST_HEAD(&cfs_rq->tasks);
7608 #ifdef CONFIG_FAIR_GROUP_SCHED
7611 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7614 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7616 struct rt_prio_array *array;
7619 array = &rt_rq->active;
7620 for (i = 0; i < MAX_RT_PRIO; i++) {
7621 INIT_LIST_HEAD(array->queue + i);
7622 __clear_bit(i, array->bitmap);
7624 /* delimiter for bitsearch: */
7625 __set_bit(MAX_RT_PRIO, array->bitmap);
7627 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7628 rt_rq->highest_prio.curr = MAX_RT_PRIO;
7630 rt_rq->highest_prio.next = MAX_RT_PRIO;
7634 rt_rq->rt_nr_migratory = 0;
7635 rt_rq->overloaded = 0;
7636 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
7640 rt_rq->rt_throttled = 0;
7641 rt_rq->rt_runtime = 0;
7642 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
7644 #ifdef CONFIG_RT_GROUP_SCHED
7645 rt_rq->rt_nr_boosted = 0;
7650 #ifdef CONFIG_FAIR_GROUP_SCHED
7651 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7652 struct sched_entity *se, int cpu, int add,
7653 struct sched_entity *parent)
7655 struct rq *rq = cpu_rq(cpu);
7656 tg->cfs_rq[cpu] = cfs_rq;
7657 init_cfs_rq(cfs_rq, rq);
7660 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7663 /* se could be NULL for init_task_group */
7668 se->cfs_rq = &rq->cfs;
7670 se->cfs_rq = parent->my_q;
7673 se->load.weight = tg->shares;
7674 se->load.inv_weight = 0;
7675 se->parent = parent;
7679 #ifdef CONFIG_RT_GROUP_SCHED
7680 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
7681 struct sched_rt_entity *rt_se, int cpu, int add,
7682 struct sched_rt_entity *parent)
7684 struct rq *rq = cpu_rq(cpu);
7686 tg->rt_rq[cpu] = rt_rq;
7687 init_rt_rq(rt_rq, rq);
7689 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
7691 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
7693 tg->rt_se[cpu] = rt_se;
7698 rt_se->rt_rq = &rq->rt;
7700 rt_se->rt_rq = parent->my_q;
7702 rt_se->my_q = rt_rq;
7703 rt_se->parent = parent;
7704 INIT_LIST_HEAD(&rt_se->run_list);
7708 void __init sched_init(void)
7711 unsigned long alloc_size = 0, ptr;
7713 #ifdef CONFIG_FAIR_GROUP_SCHED
7714 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7716 #ifdef CONFIG_RT_GROUP_SCHED
7717 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
7719 #ifdef CONFIG_CPUMASK_OFFSTACK
7720 alloc_size += num_possible_cpus() * cpumask_size();
7723 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
7725 #ifdef CONFIG_FAIR_GROUP_SCHED
7726 init_task_group.se = (struct sched_entity **)ptr;
7727 ptr += nr_cpu_ids * sizeof(void **);
7729 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
7730 ptr += nr_cpu_ids * sizeof(void **);
7732 #endif /* CONFIG_FAIR_GROUP_SCHED */
7733 #ifdef CONFIG_RT_GROUP_SCHED
7734 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
7735 ptr += nr_cpu_ids * sizeof(void **);
7737 init_task_group.rt_rq = (struct rt_rq **)ptr;
7738 ptr += nr_cpu_ids * sizeof(void **);
7740 #endif /* CONFIG_RT_GROUP_SCHED */
7741 #ifdef CONFIG_CPUMASK_OFFSTACK
7742 for_each_possible_cpu(i) {
7743 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
7744 ptr += cpumask_size();
7746 #endif /* CONFIG_CPUMASK_OFFSTACK */
7750 init_defrootdomain();
7753 init_rt_bandwidth(&def_rt_bandwidth,
7754 global_rt_period(), global_rt_runtime());
7756 #ifdef CONFIG_RT_GROUP_SCHED
7757 init_rt_bandwidth(&init_task_group.rt_bandwidth,
7758 global_rt_period(), global_rt_runtime());
7759 #endif /* CONFIG_RT_GROUP_SCHED */
7761 #ifdef CONFIG_CGROUP_SCHED
7762 list_add(&init_task_group.list, &task_groups);
7763 INIT_LIST_HEAD(&init_task_group.children);
7765 #endif /* CONFIG_CGROUP_SCHED */
7767 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7768 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
7769 __alignof__(unsigned long));
7771 for_each_possible_cpu(i) {
7775 raw_spin_lock_init(&rq->lock);
7777 rq->calc_load_active = 0;
7778 rq->calc_load_update = jiffies + LOAD_FREQ;
7779 init_cfs_rq(&rq->cfs, rq);
7780 init_rt_rq(&rq->rt, rq);
7781 #ifdef CONFIG_FAIR_GROUP_SCHED
7782 init_task_group.shares = init_task_group_load;
7783 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
7784 #ifdef CONFIG_CGROUP_SCHED
7786 * How much cpu bandwidth does init_task_group get?
7788 * In case of task-groups formed thr' the cgroup filesystem, it
7789 * gets 100% of the cpu resources in the system. This overall
7790 * system cpu resource is divided among the tasks of
7791 * init_task_group and its child task-groups in a fair manner,
7792 * based on each entity's (task or task-group's) weight
7793 * (se->load.weight).
7795 * In other words, if init_task_group has 10 tasks of weight
7796 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7797 * then A0's share of the cpu resource is:
7799 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7801 * We achieve this by letting init_task_group's tasks sit
7802 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7804 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
7806 #endif /* CONFIG_FAIR_GROUP_SCHED */
7808 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
7809 #ifdef CONFIG_RT_GROUP_SCHED
7810 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
7811 #ifdef CONFIG_CGROUP_SCHED
7812 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
7816 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
7817 rq->cpu_load[j] = 0;
7819 rq->last_load_update_tick = jiffies;
7824 rq->cpu_power = SCHED_LOAD_SCALE;
7825 rq->post_schedule = 0;
7826 rq->active_balance = 0;
7827 rq->next_balance = jiffies;
7832 rq->avg_idle = 2*sysctl_sched_migration_cost;
7833 rq_attach_root(rq, &def_root_domain);
7835 rq->nohz_balance_kick = 0;
7836 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb, i));
7840 atomic_set(&rq->nr_iowait, 0);
7843 set_load_weight(&init_task);
7845 #ifdef CONFIG_PREEMPT_NOTIFIERS
7846 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
7850 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
7853 #ifdef CONFIG_RT_MUTEXES
7854 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
7858 * The boot idle thread does lazy MMU switching as well:
7860 atomic_inc(&init_mm.mm_count);
7861 enter_lazy_tlb(&init_mm, current);
7864 * Make us the idle thread. Technically, schedule() should not be
7865 * called from this thread, however somewhere below it might be,
7866 * but because we are the idle thread, we just pick up running again
7867 * when this runqueue becomes "idle".
7869 init_idle(current, smp_processor_id());
7871 calc_load_update = jiffies + LOAD_FREQ;
7874 * During early bootup we pretend to be a normal task:
7876 current->sched_class = &fair_sched_class;
7878 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7879 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
7882 zalloc_cpumask_var(&nohz.idle_cpus_mask, GFP_NOWAIT);
7883 alloc_cpumask_var(&nohz.grp_idle_mask, GFP_NOWAIT);
7884 atomic_set(&nohz.load_balancer, nr_cpu_ids);
7885 atomic_set(&nohz.first_pick_cpu, nr_cpu_ids);
7886 atomic_set(&nohz.second_pick_cpu, nr_cpu_ids);
7888 /* May be allocated at isolcpus cmdline parse time */
7889 if (cpu_isolated_map == NULL)
7890 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
7895 scheduler_running = 1;
7898 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7899 static inline int preempt_count_equals(int preempt_offset)
7901 int nested = (preempt_count() & ~PREEMPT_ACTIVE) + rcu_preempt_depth();
7903 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
7906 void __might_sleep(const char *file, int line, int preempt_offset)
7909 static unsigned long prev_jiffy; /* ratelimiting */
7911 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
7912 system_state != SYSTEM_RUNNING || oops_in_progress)
7914 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
7916 prev_jiffy = jiffies;
7919 "BUG: sleeping function called from invalid context at %s:%d\n",
7922 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7923 in_atomic(), irqs_disabled(),
7924 current->pid, current->comm);
7926 debug_show_held_locks(current);
7927 if (irqs_disabled())
7928 print_irqtrace_events(current);
7932 EXPORT_SYMBOL(__might_sleep);
7935 #ifdef CONFIG_MAGIC_SYSRQ
7936 static void normalize_task(struct rq *rq, struct task_struct *p)
7940 on_rq = p->se.on_rq;
7942 deactivate_task(rq, p, 0);
7943 __setscheduler(rq, p, SCHED_NORMAL, 0);
7945 activate_task(rq, p, 0);
7946 resched_task(rq->curr);
7950 void normalize_rt_tasks(void)
7952 struct task_struct *g, *p;
7953 unsigned long flags;
7956 read_lock_irqsave(&tasklist_lock, flags);
7957 do_each_thread(g, p) {
7959 * Only normalize user tasks:
7964 p->se.exec_start = 0;
7965 #ifdef CONFIG_SCHEDSTATS
7966 p->se.statistics.wait_start = 0;
7967 p->se.statistics.sleep_start = 0;
7968 p->se.statistics.block_start = 0;
7973 * Renice negative nice level userspace
7976 if (TASK_NICE(p) < 0 && p->mm)
7977 set_user_nice(p, 0);
7981 raw_spin_lock(&p->pi_lock);
7982 rq = __task_rq_lock(p);
7984 normalize_task(rq, p);
7986 __task_rq_unlock(rq);
7987 raw_spin_unlock(&p->pi_lock);
7988 } while_each_thread(g, p);
7990 read_unlock_irqrestore(&tasklist_lock, flags);
7993 #endif /* CONFIG_MAGIC_SYSRQ */
7995 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7997 * These functions are only useful for the IA64 MCA handling, or kdb.
7999 * They can only be called when the whole system has been
8000 * stopped - every CPU needs to be quiescent, and no scheduling
8001 * activity can take place. Using them for anything else would
8002 * be a serious bug, and as a result, they aren't even visible
8003 * under any other configuration.
8007 * curr_task - return the current task for a given cpu.
8008 * @cpu: the processor in question.
8010 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8012 struct task_struct *curr_task(int cpu)
8014 return cpu_curr(cpu);
8017 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8021 * set_curr_task - set the current task for a given cpu.
8022 * @cpu: the processor in question.
8023 * @p: the task pointer to set.
8025 * Description: This function must only be used when non-maskable interrupts
8026 * are serviced on a separate stack. It allows the architecture to switch the
8027 * notion of the current task on a cpu in a non-blocking manner. This function
8028 * must be called with all CPU's synchronized, and interrupts disabled, the
8029 * and caller must save the original value of the current task (see
8030 * curr_task() above) and restore that value before reenabling interrupts and
8031 * re-starting the system.
8033 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8035 void set_curr_task(int cpu, struct task_struct *p)
8042 #ifdef CONFIG_FAIR_GROUP_SCHED
8043 static void free_fair_sched_group(struct task_group *tg)
8047 for_each_possible_cpu(i) {
8049 kfree(tg->cfs_rq[i]);
8059 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8061 struct cfs_rq *cfs_rq;
8062 struct sched_entity *se;
8066 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8069 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8073 tg->shares = NICE_0_LOAD;
8075 for_each_possible_cpu(i) {
8078 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8079 GFP_KERNEL, cpu_to_node(i));
8083 se = kzalloc_node(sizeof(struct sched_entity),
8084 GFP_KERNEL, cpu_to_node(i));
8088 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8099 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8101 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8102 &cpu_rq(cpu)->leaf_cfs_rq_list);
8105 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8107 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8109 #else /* !CONFG_FAIR_GROUP_SCHED */
8110 static inline void free_fair_sched_group(struct task_group *tg)
8115 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8120 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8124 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8127 #endif /* CONFIG_FAIR_GROUP_SCHED */
8129 #ifdef CONFIG_RT_GROUP_SCHED
8130 static void free_rt_sched_group(struct task_group *tg)
8134 destroy_rt_bandwidth(&tg->rt_bandwidth);
8136 for_each_possible_cpu(i) {
8138 kfree(tg->rt_rq[i]);
8140 kfree(tg->rt_se[i]);
8148 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8150 struct rt_rq *rt_rq;
8151 struct sched_rt_entity *rt_se;
8155 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8158 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8162 init_rt_bandwidth(&tg->rt_bandwidth,
8163 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8165 for_each_possible_cpu(i) {
8168 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8169 GFP_KERNEL, cpu_to_node(i));
8173 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8174 GFP_KERNEL, cpu_to_node(i));
8178 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8189 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8191 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8192 &cpu_rq(cpu)->leaf_rt_rq_list);
8195 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8197 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8199 #else /* !CONFIG_RT_GROUP_SCHED */
8200 static inline void free_rt_sched_group(struct task_group *tg)
8205 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8210 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8214 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8217 #endif /* CONFIG_RT_GROUP_SCHED */
8219 #ifdef CONFIG_CGROUP_SCHED
8220 static void free_sched_group(struct task_group *tg)
8222 free_fair_sched_group(tg);
8223 free_rt_sched_group(tg);
8227 /* allocate runqueue etc for a new task group */
8228 struct task_group *sched_create_group(struct task_group *parent)
8230 struct task_group *tg;
8231 unsigned long flags;
8234 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8236 return ERR_PTR(-ENOMEM);
8238 if (!alloc_fair_sched_group(tg, parent))
8241 if (!alloc_rt_sched_group(tg, parent))
8244 spin_lock_irqsave(&task_group_lock, flags);
8245 for_each_possible_cpu(i) {
8246 register_fair_sched_group(tg, i);
8247 register_rt_sched_group(tg, i);
8249 list_add_rcu(&tg->list, &task_groups);
8251 WARN_ON(!parent); /* root should already exist */
8253 tg->parent = parent;
8254 INIT_LIST_HEAD(&tg->children);
8255 list_add_rcu(&tg->siblings, &parent->children);
8256 spin_unlock_irqrestore(&task_group_lock, flags);
8261 free_sched_group(tg);
8262 return ERR_PTR(-ENOMEM);
8265 /* rcu callback to free various structures associated with a task group */
8266 static void free_sched_group_rcu(struct rcu_head *rhp)
8268 /* now it should be safe to free those cfs_rqs */
8269 free_sched_group(container_of(rhp, struct task_group, rcu));
8272 /* Destroy runqueue etc associated with a task group */
8273 void sched_destroy_group(struct task_group *tg)
8275 unsigned long flags;
8278 spin_lock_irqsave(&task_group_lock, flags);
8279 for_each_possible_cpu(i) {
8280 unregister_fair_sched_group(tg, i);
8281 unregister_rt_sched_group(tg, i);
8283 list_del_rcu(&tg->list);
8284 list_del_rcu(&tg->siblings);
8285 spin_unlock_irqrestore(&task_group_lock, flags);
8287 /* wait for possible concurrent references to cfs_rqs complete */
8288 call_rcu(&tg->rcu, free_sched_group_rcu);
8291 /* change task's runqueue when it moves between groups.
8292 * The caller of this function should have put the task in its new group
8293 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8294 * reflect its new group.
8296 void sched_move_task(struct task_struct *tsk)
8299 unsigned long flags;
8302 rq = task_rq_lock(tsk, &flags);
8304 running = task_current(rq, tsk);
8305 on_rq = tsk->se.on_rq;
8308 dequeue_task(rq, tsk, 0);
8309 if (unlikely(running))
8310 tsk->sched_class->put_prev_task(rq, tsk);
8312 set_task_rq(tsk, task_cpu(tsk));
8314 #ifdef CONFIG_FAIR_GROUP_SCHED
8315 if (tsk->sched_class->moved_group)
8316 tsk->sched_class->moved_group(tsk, on_rq);
8319 if (unlikely(running))
8320 tsk->sched_class->set_curr_task(rq);
8322 enqueue_task(rq, tsk, 0);
8324 task_rq_unlock(rq, &flags);
8326 #endif /* CONFIG_CGROUP_SCHED */
8328 #ifdef CONFIG_FAIR_GROUP_SCHED
8329 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8331 struct cfs_rq *cfs_rq = se->cfs_rq;
8336 dequeue_entity(cfs_rq, se, 0);
8338 se->load.weight = shares;
8339 se->load.inv_weight = 0;
8342 enqueue_entity(cfs_rq, se, 0);
8345 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8347 struct cfs_rq *cfs_rq = se->cfs_rq;
8348 struct rq *rq = cfs_rq->rq;
8349 unsigned long flags;
8351 raw_spin_lock_irqsave(&rq->lock, flags);
8352 __set_se_shares(se, shares);
8353 raw_spin_unlock_irqrestore(&rq->lock, flags);
8356 static DEFINE_MUTEX(shares_mutex);
8358 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8361 unsigned long flags;
8364 * We can't change the weight of the root cgroup.
8369 if (shares < MIN_SHARES)
8370 shares = MIN_SHARES;
8371 else if (shares > MAX_SHARES)
8372 shares = MAX_SHARES;
8374 mutex_lock(&shares_mutex);
8375 if (tg->shares == shares)
8378 spin_lock_irqsave(&task_group_lock, flags);
8379 for_each_possible_cpu(i)
8380 unregister_fair_sched_group(tg, i);
8381 list_del_rcu(&tg->siblings);
8382 spin_unlock_irqrestore(&task_group_lock, flags);
8384 /* wait for any ongoing reference to this group to finish */
8385 synchronize_sched();
8388 * Now we are free to modify the group's share on each cpu
8389 * w/o tripping rebalance_share or load_balance_fair.
8391 tg->shares = shares;
8392 for_each_possible_cpu(i) {
8396 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8397 set_se_shares(tg->se[i], shares);
8401 * Enable load balance activity on this group, by inserting it back on
8402 * each cpu's rq->leaf_cfs_rq_list.
8404 spin_lock_irqsave(&task_group_lock, flags);
8405 for_each_possible_cpu(i)
8406 register_fair_sched_group(tg, i);
8407 list_add_rcu(&tg->siblings, &tg->parent->children);
8408 spin_unlock_irqrestore(&task_group_lock, flags);
8410 mutex_unlock(&shares_mutex);
8414 unsigned long sched_group_shares(struct task_group *tg)
8420 #ifdef CONFIG_RT_GROUP_SCHED
8422 * Ensure that the real time constraints are schedulable.
8424 static DEFINE_MUTEX(rt_constraints_mutex);
8426 static unsigned long to_ratio(u64 period, u64 runtime)
8428 if (runtime == RUNTIME_INF)
8431 return div64_u64(runtime << 20, period);
8434 /* Must be called with tasklist_lock held */
8435 static inline int tg_has_rt_tasks(struct task_group *tg)
8437 struct task_struct *g, *p;
8439 do_each_thread(g, p) {
8440 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8442 } while_each_thread(g, p);
8447 struct rt_schedulable_data {
8448 struct task_group *tg;
8453 static int tg_schedulable(struct task_group *tg, void *data)
8455 struct rt_schedulable_data *d = data;
8456 struct task_group *child;
8457 unsigned long total, sum = 0;
8458 u64 period, runtime;
8460 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8461 runtime = tg->rt_bandwidth.rt_runtime;
8464 period = d->rt_period;
8465 runtime = d->rt_runtime;
8469 * Cannot have more runtime than the period.
8471 if (runtime > period && runtime != RUNTIME_INF)
8475 * Ensure we don't starve existing RT tasks.
8477 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8480 total = to_ratio(period, runtime);
8483 * Nobody can have more than the global setting allows.
8485 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8489 * The sum of our children's runtime should not exceed our own.
8491 list_for_each_entry_rcu(child, &tg->children, siblings) {
8492 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8493 runtime = child->rt_bandwidth.rt_runtime;
8495 if (child == d->tg) {
8496 period = d->rt_period;
8497 runtime = d->rt_runtime;
8500 sum += to_ratio(period, runtime);
8509 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8511 struct rt_schedulable_data data = {
8513 .rt_period = period,
8514 .rt_runtime = runtime,
8517 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8520 static int tg_set_bandwidth(struct task_group *tg,
8521 u64 rt_period, u64 rt_runtime)
8525 mutex_lock(&rt_constraints_mutex);
8526 read_lock(&tasklist_lock);
8527 err = __rt_schedulable(tg, rt_period, rt_runtime);
8531 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8532 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8533 tg->rt_bandwidth.rt_runtime = rt_runtime;
8535 for_each_possible_cpu(i) {
8536 struct rt_rq *rt_rq = tg->rt_rq[i];
8538 raw_spin_lock(&rt_rq->rt_runtime_lock);
8539 rt_rq->rt_runtime = rt_runtime;
8540 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8542 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8544 read_unlock(&tasklist_lock);
8545 mutex_unlock(&rt_constraints_mutex);
8550 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8552 u64 rt_runtime, rt_period;
8554 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8555 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8556 if (rt_runtime_us < 0)
8557 rt_runtime = RUNTIME_INF;
8559 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8562 long sched_group_rt_runtime(struct task_group *tg)
8566 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8569 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8570 do_div(rt_runtime_us, NSEC_PER_USEC);
8571 return rt_runtime_us;
8574 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8576 u64 rt_runtime, rt_period;
8578 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8579 rt_runtime = tg->rt_bandwidth.rt_runtime;
8584 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8587 long sched_group_rt_period(struct task_group *tg)
8591 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8592 do_div(rt_period_us, NSEC_PER_USEC);
8593 return rt_period_us;
8596 static int sched_rt_global_constraints(void)
8598 u64 runtime, period;
8601 if (sysctl_sched_rt_period <= 0)
8604 runtime = global_rt_runtime();
8605 period = global_rt_period();
8608 * Sanity check on the sysctl variables.
8610 if (runtime > period && runtime != RUNTIME_INF)
8613 mutex_lock(&rt_constraints_mutex);
8614 read_lock(&tasklist_lock);
8615 ret = __rt_schedulable(NULL, 0, 0);
8616 read_unlock(&tasklist_lock);
8617 mutex_unlock(&rt_constraints_mutex);
8622 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
8624 /* Don't accept realtime tasks when there is no way for them to run */
8625 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
8631 #else /* !CONFIG_RT_GROUP_SCHED */
8632 static int sched_rt_global_constraints(void)
8634 unsigned long flags;
8637 if (sysctl_sched_rt_period <= 0)
8641 * There's always some RT tasks in the root group
8642 * -- migration, kstopmachine etc..
8644 if (sysctl_sched_rt_runtime == 0)
8647 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8648 for_each_possible_cpu(i) {
8649 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8651 raw_spin_lock(&rt_rq->rt_runtime_lock);
8652 rt_rq->rt_runtime = global_rt_runtime();
8653 raw_spin_unlock(&rt_rq->rt_runtime_lock);
8655 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8659 #endif /* CONFIG_RT_GROUP_SCHED */
8661 int sched_rt_handler(struct ctl_table *table, int write,
8662 void __user *buffer, size_t *lenp,
8666 int old_period, old_runtime;
8667 static DEFINE_MUTEX(mutex);
8670 old_period = sysctl_sched_rt_period;
8671 old_runtime = sysctl_sched_rt_runtime;
8673 ret = proc_dointvec(table, write, buffer, lenp, ppos);
8675 if (!ret && write) {
8676 ret = sched_rt_global_constraints();
8678 sysctl_sched_rt_period = old_period;
8679 sysctl_sched_rt_runtime = old_runtime;
8681 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8682 def_rt_bandwidth.rt_period =
8683 ns_to_ktime(global_rt_period());
8686 mutex_unlock(&mutex);
8691 #ifdef CONFIG_CGROUP_SCHED
8693 /* return corresponding task_group object of a cgroup */
8694 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8696 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8697 struct task_group, css);
8700 static struct cgroup_subsys_state *
8701 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8703 struct task_group *tg, *parent;
8705 if (!cgrp->parent) {
8706 /* This is early initialization for the top cgroup */
8707 return &init_task_group.css;
8710 parent = cgroup_tg(cgrp->parent);
8711 tg = sched_create_group(parent);
8713 return ERR_PTR(-ENOMEM);
8719 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8721 struct task_group *tg = cgroup_tg(cgrp);
8723 sched_destroy_group(tg);
8727 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
8729 #ifdef CONFIG_RT_GROUP_SCHED
8730 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
8733 /* We don't support RT-tasks being in separate groups */
8734 if (tsk->sched_class != &fair_sched_class)
8741 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8742 struct task_struct *tsk, bool threadgroup)
8744 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
8748 struct task_struct *c;
8750 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8751 retval = cpu_cgroup_can_attach_task(cgrp, c);
8763 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
8764 struct cgroup *old_cont, struct task_struct *tsk,
8767 sched_move_task(tsk);
8769 struct task_struct *c;
8771 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
8778 #ifdef CONFIG_FAIR_GROUP_SCHED
8779 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
8782 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
8785 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
8787 struct task_group *tg = cgroup_tg(cgrp);
8789 return (u64) tg->shares;
8791 #endif /* CONFIG_FAIR_GROUP_SCHED */
8793 #ifdef CONFIG_RT_GROUP_SCHED
8794 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
8797 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
8800 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
8802 return sched_group_rt_runtime(cgroup_tg(cgrp));
8805 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
8808 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
8811 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
8813 return sched_group_rt_period(cgroup_tg(cgrp));
8815 #endif /* CONFIG_RT_GROUP_SCHED */
8817 static struct cftype cpu_files[] = {
8818 #ifdef CONFIG_FAIR_GROUP_SCHED
8821 .read_u64 = cpu_shares_read_u64,
8822 .write_u64 = cpu_shares_write_u64,
8825 #ifdef CONFIG_RT_GROUP_SCHED
8827 .name = "rt_runtime_us",
8828 .read_s64 = cpu_rt_runtime_read,
8829 .write_s64 = cpu_rt_runtime_write,
8832 .name = "rt_period_us",
8833 .read_u64 = cpu_rt_period_read_uint,
8834 .write_u64 = cpu_rt_period_write_uint,
8839 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
8841 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
8844 struct cgroup_subsys cpu_cgroup_subsys = {
8846 .create = cpu_cgroup_create,
8847 .destroy = cpu_cgroup_destroy,
8848 .can_attach = cpu_cgroup_can_attach,
8849 .attach = cpu_cgroup_attach,
8850 .populate = cpu_cgroup_populate,
8851 .subsys_id = cpu_cgroup_subsys_id,
8855 #endif /* CONFIG_CGROUP_SCHED */
8857 #ifdef CONFIG_CGROUP_CPUACCT
8860 * CPU accounting code for task groups.
8862 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8863 * (balbir@in.ibm.com).
8866 /* track cpu usage of a group of tasks and its child groups */
8868 struct cgroup_subsys_state css;
8869 /* cpuusage holds pointer to a u64-type object on every cpu */
8870 u64 __percpu *cpuusage;
8871 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
8872 struct cpuacct *parent;
8875 struct cgroup_subsys cpuacct_subsys;
8877 /* return cpu accounting group corresponding to this container */
8878 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
8880 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
8881 struct cpuacct, css);
8884 /* return cpu accounting group to which this task belongs */
8885 static inline struct cpuacct *task_ca(struct task_struct *tsk)
8887 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
8888 struct cpuacct, css);
8891 /* create a new cpu accounting group */
8892 static struct cgroup_subsys_state *cpuacct_create(
8893 struct cgroup_subsys *ss, struct cgroup *cgrp)
8895 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
8901 ca->cpuusage = alloc_percpu(u64);
8905 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8906 if (percpu_counter_init(&ca->cpustat[i], 0))
8907 goto out_free_counters;
8910 ca->parent = cgroup_ca(cgrp->parent);
8916 percpu_counter_destroy(&ca->cpustat[i]);
8917 free_percpu(ca->cpuusage);
8921 return ERR_PTR(-ENOMEM);
8924 /* destroy an existing cpu accounting group */
8926 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8928 struct cpuacct *ca = cgroup_ca(cgrp);
8931 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
8932 percpu_counter_destroy(&ca->cpustat[i]);
8933 free_percpu(ca->cpuusage);
8937 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
8939 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8942 #ifndef CONFIG_64BIT
8944 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8946 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8948 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8956 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
8958 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
8960 #ifndef CONFIG_64BIT
8962 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8964 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
8966 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
8972 /* return total cpu usage (in nanoseconds) of a group */
8973 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
8975 struct cpuacct *ca = cgroup_ca(cgrp);
8976 u64 totalcpuusage = 0;
8979 for_each_present_cpu(i)
8980 totalcpuusage += cpuacct_cpuusage_read(ca, i);
8982 return totalcpuusage;
8985 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
8988 struct cpuacct *ca = cgroup_ca(cgrp);
8997 for_each_present_cpu(i)
8998 cpuacct_cpuusage_write(ca, i, 0);
9004 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9007 struct cpuacct *ca = cgroup_ca(cgroup);
9011 for_each_present_cpu(i) {
9012 percpu = cpuacct_cpuusage_read(ca, i);
9013 seq_printf(m, "%llu ", (unsigned long long) percpu);
9015 seq_printf(m, "\n");
9019 static const char *cpuacct_stat_desc[] = {
9020 [CPUACCT_STAT_USER] = "user",
9021 [CPUACCT_STAT_SYSTEM] = "system",
9024 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9025 struct cgroup_map_cb *cb)
9027 struct cpuacct *ca = cgroup_ca(cgrp);
9030 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9031 s64 val = percpu_counter_read(&ca->cpustat[i]);
9032 val = cputime64_to_clock_t(val);
9033 cb->fill(cb, cpuacct_stat_desc[i], val);
9038 static struct cftype files[] = {
9041 .read_u64 = cpuusage_read,
9042 .write_u64 = cpuusage_write,
9045 .name = "usage_percpu",
9046 .read_seq_string = cpuacct_percpu_seq_read,
9050 .read_map = cpuacct_stats_show,
9054 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9056 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9060 * charge this task's execution time to its accounting group.
9062 * called with rq->lock held.
9064 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9069 if (unlikely(!cpuacct_subsys.active))
9072 cpu = task_cpu(tsk);
9078 for (; ca; ca = ca->parent) {
9079 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
9080 *cpuusage += cputime;
9087 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9088 * in cputime_t units. As a result, cpuacct_update_stats calls
9089 * percpu_counter_add with values large enough to always overflow the
9090 * per cpu batch limit causing bad SMP scalability.
9092 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9093 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9094 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9097 #define CPUACCT_BATCH \
9098 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9100 #define CPUACCT_BATCH 0
9104 * Charge the system/user time to the task's accounting group.
9106 static void cpuacct_update_stats(struct task_struct *tsk,
9107 enum cpuacct_stat_index idx, cputime_t val)
9110 int batch = CPUACCT_BATCH;
9112 if (unlikely(!cpuacct_subsys.active))
9119 __percpu_counter_add(&ca->cpustat[idx], val, batch);
9125 struct cgroup_subsys cpuacct_subsys = {
9127 .create = cpuacct_create,
9128 .destroy = cpuacct_destroy,
9129 .populate = cpuacct_populate,
9130 .subsys_id = cpuacct_subsys_id,
9132 #endif /* CONFIG_CGROUP_CPUACCT */