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
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
32 #include <linux/module.h>
33 #include <linux/nmi.h>
34 #include <linux/init.h>
35 #include <linux/uaccess.h>
36 #include <linux/highmem.h>
37 #include <linux/smp_lock.h>
38 #include <asm/mmu_context.h>
39 #include <linux/interrupt.h>
40 #include <linux/capability.h>
41 #include <linux/completion.h>
42 #include <linux/kernel_stat.h>
43 #include <linux/debug_locks.h>
44 #include <linux/perf_event.h>
45 #include <linux/security.h>
46 #include <linux/notifier.h>
47 #include <linux/profile.h>
48 #include <linux/freezer.h>
49 #include <linux/vmalloc.h>
50 #include <linux/blkdev.h>
51 #include <linux/delay.h>
52 #include <linux/pid_namespace.h>
53 #include <linux/smp.h>
54 #include <linux/threads.h>
55 #include <linux/timer.h>
56 #include <linux/rcupdate.h>
57 #include <linux/cpu.h>
58 #include <linux/cpuset.h>
59 #include <linux/percpu.h>
60 #include <linux/kthread.h>
61 #include <linux/proc_fs.h>
62 #include <linux/seq_file.h>
63 #include <linux/sysctl.h>
64 #include <linux/syscalls.h>
65 #include <linux/times.h>
66 #include <linux/tsacct_kern.h>
67 #include <linux/kprobes.h>
68 #include <linux/delayacct.h>
69 #include <linux/unistd.h>
70 #include <linux/pagemap.h>
71 #include <linux/hrtimer.h>
72 #include <linux/tick.h>
73 #include <linux/debugfs.h>
74 #include <linux/ctype.h>
75 #include <linux/ftrace.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.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 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 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 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 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_GROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups);
246 /* task group related information */
248 #ifdef CONFIG_CGROUP_SCHED
249 struct cgroup_subsys_state css;
252 #ifdef CONFIG_USER_SCHED
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 /* schedulable entities of this group on each cpu */
258 struct sched_entity **se;
259 /* runqueue "owned" by this group on each cpu */
260 struct cfs_rq **cfs_rq;
261 unsigned long shares;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
268 struct rt_bandwidth rt_bandwidth;
272 struct list_head list;
274 struct task_group *parent;
275 struct list_head siblings;
276 struct list_head children;
279 #ifdef CONFIG_USER_SCHED
281 /* Helper function to pass uid information to create_sched_user() */
282 void set_tg_uid(struct user_struct *user)
284 user->tg->uid = user->uid;
289 * Every UID task group (including init_task_group aka UID-0) will
290 * be a child to this group.
292 struct task_group root_task_group;
294 #ifdef CONFIG_FAIR_GROUP_SCHED
295 /* Default task group's sched entity on each cpu */
296 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
297 /* Default task group's cfs_rq on each cpu */
298 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
299 #endif /* CONFIG_FAIR_GROUP_SCHED */
301 #ifdef CONFIG_RT_GROUP_SCHED
302 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
303 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq);
304 #endif /* CONFIG_RT_GROUP_SCHED */
305 #else /* !CONFIG_USER_SCHED */
306 #define root_task_group init_task_group
307 #endif /* CONFIG_USER_SCHED */
309 /* task_group_lock serializes add/remove of task groups and also changes to
310 * a task group's cpu shares.
312 static DEFINE_SPINLOCK(task_group_lock);
314 #ifdef CONFIG_FAIR_GROUP_SCHED
317 static int root_task_group_empty(void)
319 return list_empty(&root_task_group.children);
323 #ifdef CONFIG_USER_SCHED
324 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
325 #else /* !CONFIG_USER_SCHED */
326 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
327 #endif /* CONFIG_USER_SCHED */
330 * A weight of 0 or 1 can cause arithmetics problems.
331 * A weight of a cfs_rq is the sum of weights of which entities
332 * are queued on this cfs_rq, so a weight of a entity should not be
333 * too large, so as the shares value of a task group.
334 * (The default weight is 1024 - so there's no practical
335 * limitation from this.)
338 #define MAX_SHARES (1UL << 18)
340 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
343 /* Default task group.
344 * Every task in system belong to this group at bootup.
346 struct task_group init_task_group;
348 /* return group to which a task belongs */
349 static inline struct task_group *task_group(struct task_struct *p)
351 struct task_group *tg;
353 #ifdef CONFIG_USER_SCHED
355 tg = __task_cred(p)->user->tg;
357 #elif defined(CONFIG_CGROUP_SCHED)
358 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
359 struct task_group, css);
361 tg = &init_task_group;
366 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
367 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
369 #ifdef CONFIG_FAIR_GROUP_SCHED
370 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
371 p->se.parent = task_group(p)->se[cpu];
374 #ifdef CONFIG_RT_GROUP_SCHED
375 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
376 p->rt.parent = task_group(p)->rt_se[cpu];
382 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
383 static inline struct task_group *task_group(struct task_struct *p)
388 #endif /* CONFIG_GROUP_SCHED */
390 /* CFS-related fields in a runqueue */
392 struct load_weight load;
393 unsigned long nr_running;
398 struct rb_root tasks_timeline;
399 struct rb_node *rb_leftmost;
401 struct list_head tasks;
402 struct list_head *balance_iterator;
405 * 'curr' points to currently running entity on this cfs_rq.
406 * It is set to NULL otherwise (i.e when none are currently running).
408 struct sched_entity *curr, *next, *last;
410 unsigned int nr_spread_over;
412 #ifdef CONFIG_FAIR_GROUP_SCHED
413 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
416 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
417 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
418 * (like users, containers etc.)
420 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
421 * list is used during load balance.
423 struct list_head leaf_cfs_rq_list;
424 struct task_group *tg; /* group that "owns" this runqueue */
428 * the part of load.weight contributed by tasks
430 unsigned long task_weight;
433 * h_load = weight * f(tg)
435 * Where f(tg) is the recursive weight fraction assigned to
438 unsigned long h_load;
441 * this cpu's part of tg->shares
443 unsigned long shares;
446 * load.weight at the time we set shares
448 unsigned long rq_weight;
453 /* Real-Time classes' related field in a runqueue: */
455 struct rt_prio_array active;
456 unsigned long rt_nr_running;
457 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
459 int curr; /* highest queued rt task prio */
461 int next; /* next highest */
466 unsigned long rt_nr_migratory;
467 unsigned long rt_nr_total;
469 struct plist_head pushable_tasks;
474 /* Nests inside the rq lock: */
475 spinlock_t rt_runtime_lock;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 unsigned long rt_nr_boosted;
481 struct list_head leaf_rt_rq_list;
482 struct task_group *tg;
483 struct sched_rt_entity *rt_se;
490 * We add the notion of a root-domain which will be used to define per-domain
491 * variables. Each exclusive cpuset essentially defines an island domain by
492 * fully partitioning the member cpus from any other cpuset. Whenever a new
493 * exclusive cpuset is created, we also create and attach a new root-domain
500 cpumask_var_t online;
503 * The "RT overload" flag: it gets set if a CPU has more than
504 * one runnable RT task.
506 cpumask_var_t rto_mask;
509 struct cpupri cpupri;
514 * By default the system creates a single root-domain with all cpus as
515 * members (mimicking the global state we have today).
517 static struct root_domain def_root_domain;
522 * This is the main, per-CPU runqueue data structure.
524 * Locking rule: those places that want to lock multiple runqueues
525 * (such as the load balancing or the thread migration code), lock
526 * acquire operations must be ordered by ascending &runqueue.
533 * nr_running and cpu_load should be in the same cacheline because
534 * remote CPUs use both these fields when doing load calculation.
536 unsigned long nr_running;
537 #define CPU_LOAD_IDX_MAX 5
538 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
540 unsigned char in_nohz_recently;
542 /* capture load from *all* tasks on this cpu: */
543 struct load_weight load;
544 unsigned long nr_load_updates;
550 #ifdef CONFIG_FAIR_GROUP_SCHED
551 /* list of leaf cfs_rq on this cpu: */
552 struct list_head leaf_cfs_rq_list;
554 #ifdef CONFIG_RT_GROUP_SCHED
555 struct list_head leaf_rt_rq_list;
559 * This is part of a global counter where only the total sum
560 * over all CPUs matters. A task can increase this counter on
561 * one CPU and if it got migrated afterwards it may decrease
562 * it on another CPU. Always updated under the runqueue lock:
564 unsigned long nr_uninterruptible;
566 struct task_struct *curr, *idle;
567 unsigned long next_balance;
568 struct mm_struct *prev_mm;
575 struct root_domain *rd;
576 struct sched_domain *sd;
578 unsigned char idle_at_tick;
579 /* For active balancing */
583 /* cpu of this runqueue: */
587 unsigned long avg_load_per_task;
589 struct task_struct *migration_thread;
590 struct list_head migration_queue;
598 /* calc_load related fields */
599 unsigned long calc_load_update;
600 long calc_load_active;
602 #ifdef CONFIG_SCHED_HRTICK
604 int hrtick_csd_pending;
605 struct call_single_data hrtick_csd;
607 struct hrtimer hrtick_timer;
610 #ifdef CONFIG_SCHEDSTATS
612 struct sched_info rq_sched_info;
613 unsigned long long rq_cpu_time;
614 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
616 /* sys_sched_yield() stats */
617 unsigned int yld_count;
619 /* schedule() stats */
620 unsigned int sched_switch;
621 unsigned int sched_count;
622 unsigned int sched_goidle;
624 /* try_to_wake_up() stats */
625 unsigned int ttwu_count;
626 unsigned int ttwu_local;
629 unsigned int bkl_count;
633 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
636 void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
638 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
641 static inline int cpu_of(struct rq *rq)
651 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
652 * See detach_destroy_domains: synchronize_sched for details.
654 * The domain tree of any CPU may only be accessed from within
655 * preempt-disabled sections.
657 #define for_each_domain(cpu, __sd) \
658 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
660 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
661 #define this_rq() (&__get_cpu_var(runqueues))
662 #define task_rq(p) cpu_rq(task_cpu(p))
663 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
664 #define raw_rq() (&__raw_get_cpu_var(runqueues))
666 inline void update_rq_clock(struct rq *rq)
668 rq->clock = sched_clock_cpu(cpu_of(rq));
672 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
674 #ifdef CONFIG_SCHED_DEBUG
675 # define const_debug __read_mostly
677 # define const_debug static const
682 * @cpu: the processor in question.
684 * Returns true if the current cpu runqueue is locked.
685 * This interface allows printk to be called with the runqueue lock
686 * held and know whether or not it is OK to wake up the klogd.
688 int runqueue_is_locked(int cpu)
690 return spin_is_locked(&cpu_rq(cpu)->lock);
694 * Debugging: various feature bits
697 #define SCHED_FEAT(name, enabled) \
698 __SCHED_FEAT_##name ,
701 #include "sched_features.h"
706 #define SCHED_FEAT(name, enabled) \
707 (1UL << __SCHED_FEAT_##name) * enabled |
709 const_debug unsigned int sysctl_sched_features =
710 #include "sched_features.h"
715 #ifdef CONFIG_SCHED_DEBUG
716 #define SCHED_FEAT(name, enabled) \
719 static __read_mostly char *sched_feat_names[] = {
720 #include "sched_features.h"
726 static int sched_feat_show(struct seq_file *m, void *v)
730 for (i = 0; sched_feat_names[i]; i++) {
731 if (!(sysctl_sched_features & (1UL << i)))
733 seq_printf(m, "%s ", sched_feat_names[i]);
741 sched_feat_write(struct file *filp, const char __user *ubuf,
742 size_t cnt, loff_t *ppos)
752 if (copy_from_user(&buf, ubuf, cnt))
757 if (strncmp(buf, "NO_", 3) == 0) {
762 for (i = 0; sched_feat_names[i]; i++) {
763 int len = strlen(sched_feat_names[i]);
765 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
767 sysctl_sched_features &= ~(1UL << i);
769 sysctl_sched_features |= (1UL << i);
774 if (!sched_feat_names[i])
782 static int sched_feat_open(struct inode *inode, struct file *filp)
784 return single_open(filp, sched_feat_show, NULL);
787 static const struct file_operations sched_feat_fops = {
788 .open = sched_feat_open,
789 .write = sched_feat_write,
792 .release = single_release,
795 static __init int sched_init_debug(void)
797 debugfs_create_file("sched_features", 0644, NULL, NULL,
802 late_initcall(sched_init_debug);
806 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
809 * Number of tasks to iterate in a single balance run.
810 * Limited because this is done with IRQs disabled.
812 const_debug unsigned int sysctl_sched_nr_migrate = 32;
815 * ratelimit for updating the group shares.
818 unsigned int sysctl_sched_shares_ratelimit = 250000;
819 unsigned int normalized_sysctl_sched_shares_ratelimit = 250000;
822 * Inject some fuzzyness into changing the per-cpu group shares
823 * this avoids remote rq-locks at the expense of fairness.
826 unsigned int sysctl_sched_shares_thresh = 4;
829 * period over which we average the RT time consumption, measured
834 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
837 * period over which we measure -rt task cpu usage in us.
840 unsigned int sysctl_sched_rt_period = 1000000;
842 static __read_mostly int scheduler_running;
845 * part of the period that we allow rt tasks to run in us.
848 int sysctl_sched_rt_runtime = 950000;
850 static inline u64 global_rt_period(void)
852 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
855 static inline u64 global_rt_runtime(void)
857 if (sysctl_sched_rt_runtime < 0)
860 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
863 #ifndef prepare_arch_switch
864 # define prepare_arch_switch(next) do { } while (0)
866 #ifndef finish_arch_switch
867 # define finish_arch_switch(prev) do { } while (0)
870 static inline int task_current(struct rq *rq, struct task_struct *p)
872 return rq->curr == p;
875 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
876 static inline int task_running(struct rq *rq, struct task_struct *p)
878 return task_current(rq, p);
881 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
885 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
887 #ifdef CONFIG_DEBUG_SPINLOCK
888 /* this is a valid case when another task releases the spinlock */
889 rq->lock.owner = current;
892 * If we are tracking spinlock dependencies then we have to
893 * fix up the runqueue lock - which gets 'carried over' from
896 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
898 spin_unlock_irq(&rq->lock);
901 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
902 static inline int task_running(struct rq *rq, struct task_struct *p)
907 return task_current(rq, p);
911 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
915 * We can optimise this out completely for !SMP, because the
916 * SMP rebalancing from interrupt is the only thing that cares
921 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
922 spin_unlock_irq(&rq->lock);
924 spin_unlock(&rq->lock);
928 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
932 * After ->oncpu is cleared, the task can be moved to a different CPU.
933 * We must ensure this doesn't happen until the switch is completely
939 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
943 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
946 * __task_rq_lock - lock the runqueue a given task resides on.
947 * Must be called interrupts disabled.
949 static inline struct rq *__task_rq_lock(struct task_struct *p)
953 struct rq *rq = task_rq(p);
954 spin_lock(&rq->lock);
955 if (likely(rq == task_rq(p)))
957 spin_unlock(&rq->lock);
962 * task_rq_lock - lock the runqueue a given task resides on and disable
963 * interrupts. Note the ordering: we can safely lookup the task_rq without
964 * explicitly disabling preemption.
966 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
972 local_irq_save(*flags);
974 spin_lock(&rq->lock);
975 if (likely(rq == task_rq(p)))
977 spin_unlock_irqrestore(&rq->lock, *flags);
981 void task_rq_unlock_wait(struct task_struct *p)
983 struct rq *rq = task_rq(p);
985 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
986 spin_unlock_wait(&rq->lock);
989 static void __task_rq_unlock(struct rq *rq)
992 spin_unlock(&rq->lock);
995 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
998 spin_unlock_irqrestore(&rq->lock, *flags);
1002 * this_rq_lock - lock this runqueue and disable interrupts.
1004 static struct rq *this_rq_lock(void)
1005 __acquires(rq->lock)
1009 local_irq_disable();
1011 spin_lock(&rq->lock);
1016 #ifdef CONFIG_SCHED_HRTICK
1018 * Use HR-timers to deliver accurate preemption points.
1020 * Its all a bit involved since we cannot program an hrt while holding the
1021 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1024 * When we get rescheduled we reprogram the hrtick_timer outside of the
1030 * - enabled by features
1031 * - hrtimer is actually high res
1033 static inline int hrtick_enabled(struct rq *rq)
1035 if (!sched_feat(HRTICK))
1037 if (!cpu_active(cpu_of(rq)))
1039 return hrtimer_is_hres_active(&rq->hrtick_timer);
1042 static void hrtick_clear(struct rq *rq)
1044 if (hrtimer_active(&rq->hrtick_timer))
1045 hrtimer_cancel(&rq->hrtick_timer);
1049 * High-resolution timer tick.
1050 * Runs from hardirq context with interrupts disabled.
1052 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1054 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1056 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1058 spin_lock(&rq->lock);
1059 update_rq_clock(rq);
1060 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1061 spin_unlock(&rq->lock);
1063 return HRTIMER_NORESTART;
1068 * called from hardirq (IPI) context
1070 static void __hrtick_start(void *arg)
1072 struct rq *rq = arg;
1074 spin_lock(&rq->lock);
1075 hrtimer_restart(&rq->hrtick_timer);
1076 rq->hrtick_csd_pending = 0;
1077 spin_unlock(&rq->lock);
1081 * Called to set the hrtick timer state.
1083 * called with rq->lock held and irqs disabled
1085 static void hrtick_start(struct rq *rq, u64 delay)
1087 struct hrtimer *timer = &rq->hrtick_timer;
1088 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1090 hrtimer_set_expires(timer, time);
1092 if (rq == this_rq()) {
1093 hrtimer_restart(timer);
1094 } else if (!rq->hrtick_csd_pending) {
1095 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1096 rq->hrtick_csd_pending = 1;
1101 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1103 int cpu = (int)(long)hcpu;
1106 case CPU_UP_CANCELED:
1107 case CPU_UP_CANCELED_FROZEN:
1108 case CPU_DOWN_PREPARE:
1109 case CPU_DOWN_PREPARE_FROZEN:
1111 case CPU_DEAD_FROZEN:
1112 hrtick_clear(cpu_rq(cpu));
1119 static __init void init_hrtick(void)
1121 hotcpu_notifier(hotplug_hrtick, 0);
1125 * Called to set the hrtick timer state.
1127 * called with rq->lock held and irqs disabled
1129 static void hrtick_start(struct rq *rq, u64 delay)
1131 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1132 HRTIMER_MODE_REL_PINNED, 0);
1135 static inline void init_hrtick(void)
1138 #endif /* CONFIG_SMP */
1140 static void init_rq_hrtick(struct rq *rq)
1143 rq->hrtick_csd_pending = 0;
1145 rq->hrtick_csd.flags = 0;
1146 rq->hrtick_csd.func = __hrtick_start;
1147 rq->hrtick_csd.info = rq;
1150 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1151 rq->hrtick_timer.function = hrtick;
1153 #else /* CONFIG_SCHED_HRTICK */
1154 static inline void hrtick_clear(struct rq *rq)
1158 static inline void init_rq_hrtick(struct rq *rq)
1162 static inline void init_hrtick(void)
1165 #endif /* CONFIG_SCHED_HRTICK */
1168 * resched_task - mark a task 'to be rescheduled now'.
1170 * On UP this means the setting of the need_resched flag, on SMP it
1171 * might also involve a cross-CPU call to trigger the scheduler on
1176 #ifndef tsk_is_polling
1177 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1180 static void resched_task(struct task_struct *p)
1184 assert_spin_locked(&task_rq(p)->lock);
1186 if (test_tsk_need_resched(p))
1189 set_tsk_need_resched(p);
1192 if (cpu == smp_processor_id())
1195 /* NEED_RESCHED must be visible before we test polling */
1197 if (!tsk_is_polling(p))
1198 smp_send_reschedule(cpu);
1201 static void resched_cpu(int cpu)
1203 struct rq *rq = cpu_rq(cpu);
1204 unsigned long flags;
1206 if (!spin_trylock_irqsave(&rq->lock, flags))
1208 resched_task(cpu_curr(cpu));
1209 spin_unlock_irqrestore(&rq->lock, flags);
1214 * When add_timer_on() enqueues a timer into the timer wheel of an
1215 * idle CPU then this timer might expire before the next timer event
1216 * which is scheduled to wake up that CPU. In case of a completely
1217 * idle system the next event might even be infinite time into the
1218 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1219 * leaves the inner idle loop so the newly added timer is taken into
1220 * account when the CPU goes back to idle and evaluates the timer
1221 * wheel for the next timer event.
1223 void wake_up_idle_cpu(int cpu)
1225 struct rq *rq = cpu_rq(cpu);
1227 if (cpu == smp_processor_id())
1231 * This is safe, as this function is called with the timer
1232 * wheel base lock of (cpu) held. When the CPU is on the way
1233 * to idle and has not yet set rq->curr to idle then it will
1234 * be serialized on the timer wheel base lock and take the new
1235 * timer into account automatically.
1237 if (rq->curr != rq->idle)
1241 * We can set TIF_RESCHED on the idle task of the other CPU
1242 * lockless. The worst case is that the other CPU runs the
1243 * idle task through an additional NOOP schedule()
1245 set_tsk_need_resched(rq->idle);
1247 /* NEED_RESCHED must be visible before we test polling */
1249 if (!tsk_is_polling(rq->idle))
1250 smp_send_reschedule(cpu);
1252 #endif /* CONFIG_NO_HZ */
1254 static u64 sched_avg_period(void)
1256 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1259 static void sched_avg_update(struct rq *rq)
1261 s64 period = sched_avg_period();
1263 while ((s64)(rq->clock - rq->age_stamp) > period) {
1264 rq->age_stamp += period;
1269 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1271 rq->rt_avg += rt_delta;
1272 sched_avg_update(rq);
1275 #else /* !CONFIG_SMP */
1276 static void resched_task(struct task_struct *p)
1278 assert_spin_locked(&task_rq(p)->lock);
1279 set_tsk_need_resched(p);
1282 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1285 #endif /* CONFIG_SMP */
1287 #if BITS_PER_LONG == 32
1288 # define WMULT_CONST (~0UL)
1290 # define WMULT_CONST (1UL << 32)
1293 #define WMULT_SHIFT 32
1296 * Shift right and round:
1298 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1301 * delta *= weight / lw
1303 static unsigned long
1304 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1305 struct load_weight *lw)
1309 if (!lw->inv_weight) {
1310 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1313 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1317 tmp = (u64)delta_exec * weight;
1319 * Check whether we'd overflow the 64-bit multiplication:
1321 if (unlikely(tmp > WMULT_CONST))
1322 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1325 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1327 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1330 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1336 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1343 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1344 * of tasks with abnormal "nice" values across CPUs the contribution that
1345 * each task makes to its run queue's load is weighted according to its
1346 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1347 * scaled version of the new time slice allocation that they receive on time
1351 #define WEIGHT_IDLEPRIO 3
1352 #define WMULT_IDLEPRIO 1431655765
1355 * Nice levels are multiplicative, with a gentle 10% change for every
1356 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1357 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1358 * that remained on nice 0.
1360 * The "10% effect" is relative and cumulative: from _any_ nice level,
1361 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1362 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1363 * If a task goes up by ~10% and another task goes down by ~10% then
1364 * the relative distance between them is ~25%.)
1366 static const int prio_to_weight[40] = {
1367 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1368 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1369 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1370 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1371 /* 0 */ 1024, 820, 655, 526, 423,
1372 /* 5 */ 335, 272, 215, 172, 137,
1373 /* 10 */ 110, 87, 70, 56, 45,
1374 /* 15 */ 36, 29, 23, 18, 15,
1378 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1380 * In cases where the weight does not change often, we can use the
1381 * precalculated inverse to speed up arithmetics by turning divisions
1382 * into multiplications:
1384 static const u32 prio_to_wmult[40] = {
1385 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1386 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1387 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1388 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1389 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1390 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1391 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1392 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1395 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1398 * runqueue iterator, to support SMP load-balancing between different
1399 * scheduling classes, without having to expose their internal data
1400 * structures to the load-balancing proper:
1402 struct rq_iterator {
1404 struct task_struct *(*start)(void *);
1405 struct task_struct *(*next)(void *);
1409 static unsigned long
1410 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1411 unsigned long max_load_move, struct sched_domain *sd,
1412 enum cpu_idle_type idle, int *all_pinned,
1413 int *this_best_prio, struct rq_iterator *iterator);
1416 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1417 struct sched_domain *sd, enum cpu_idle_type idle,
1418 struct rq_iterator *iterator);
1421 /* Time spent by the tasks of the cpu accounting group executing in ... */
1422 enum cpuacct_stat_index {
1423 CPUACCT_STAT_USER, /* ... user mode */
1424 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1426 CPUACCT_STAT_NSTATS,
1429 #ifdef CONFIG_CGROUP_CPUACCT
1430 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1431 static void cpuacct_update_stats(struct task_struct *tsk,
1432 enum cpuacct_stat_index idx, cputime_t val);
1434 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1435 static inline void cpuacct_update_stats(struct task_struct *tsk,
1436 enum cpuacct_stat_index idx, cputime_t val) {}
1439 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1441 update_load_add(&rq->load, load);
1444 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1446 update_load_sub(&rq->load, load);
1449 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1450 typedef int (*tg_visitor)(struct task_group *, void *);
1453 * Iterate the full tree, calling @down when first entering a node and @up when
1454 * leaving it for the final time.
1456 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1458 struct task_group *parent, *child;
1462 parent = &root_task_group;
1464 ret = (*down)(parent, data);
1467 list_for_each_entry_rcu(child, &parent->children, siblings) {
1474 ret = (*up)(parent, data);
1479 parent = parent->parent;
1488 static int tg_nop(struct task_group *tg, void *data)
1495 /* Used instead of source_load when we know the type == 0 */
1496 static unsigned long weighted_cpuload(const int cpu)
1498 return cpu_rq(cpu)->load.weight;
1502 * Return a low guess at the load of a migration-source cpu weighted
1503 * according to the scheduling class and "nice" value.
1505 * We want to under-estimate the load of migration sources, to
1506 * balance conservatively.
1508 static unsigned long source_load(int cpu, int type)
1510 struct rq *rq = cpu_rq(cpu);
1511 unsigned long total = weighted_cpuload(cpu);
1513 if (type == 0 || !sched_feat(LB_BIAS))
1516 return min(rq->cpu_load[type-1], total);
1520 * Return a high guess at the load of a migration-target cpu weighted
1521 * according to the scheduling class and "nice" value.
1523 static unsigned long target_load(int cpu, int type)
1525 struct rq *rq = cpu_rq(cpu);
1526 unsigned long total = weighted_cpuload(cpu);
1528 if (type == 0 || !sched_feat(LB_BIAS))
1531 return max(rq->cpu_load[type-1], total);
1534 static struct sched_group *group_of(int cpu)
1536 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1544 static unsigned long power_of(int cpu)
1546 struct sched_group *group = group_of(cpu);
1549 return SCHED_LOAD_SCALE;
1551 return group->cpu_power;
1554 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1556 static unsigned long cpu_avg_load_per_task(int cpu)
1558 struct rq *rq = cpu_rq(cpu);
1559 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1562 rq->avg_load_per_task = rq->load.weight / nr_running;
1564 rq->avg_load_per_task = 0;
1566 return rq->avg_load_per_task;
1569 #ifdef CONFIG_FAIR_GROUP_SCHED
1571 static __read_mostly unsigned long *update_shares_data;
1573 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1576 * Calculate and set the cpu's group shares.
1578 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1579 unsigned long sd_shares,
1580 unsigned long sd_rq_weight,
1581 unsigned long *usd_rq_weight)
1583 unsigned long shares, rq_weight;
1586 rq_weight = usd_rq_weight[cpu];
1589 rq_weight = NICE_0_LOAD;
1593 * \Sum_j shares_j * rq_weight_i
1594 * shares_i = -----------------------------
1595 * \Sum_j rq_weight_j
1597 shares = (sd_shares * rq_weight) / sd_rq_weight;
1598 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1600 if (abs(shares - tg->se[cpu]->load.weight) >
1601 sysctl_sched_shares_thresh) {
1602 struct rq *rq = cpu_rq(cpu);
1603 unsigned long flags;
1605 spin_lock_irqsave(&rq->lock, flags);
1606 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1607 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1608 __set_se_shares(tg->se[cpu], shares);
1609 spin_unlock_irqrestore(&rq->lock, flags);
1614 * Re-compute the task group their per cpu shares over the given domain.
1615 * This needs to be done in a bottom-up fashion because the rq weight of a
1616 * parent group depends on the shares of its child groups.
1618 static int tg_shares_up(struct task_group *tg, void *data)
1620 unsigned long weight, rq_weight = 0, sum_weight = 0, shares = 0;
1621 unsigned long *usd_rq_weight;
1622 struct sched_domain *sd = data;
1623 unsigned long flags;
1629 local_irq_save(flags);
1630 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
1632 for_each_cpu(i, sched_domain_span(sd)) {
1633 weight = tg->cfs_rq[i]->load.weight;
1634 usd_rq_weight[i] = weight;
1636 rq_weight += weight;
1638 * If there are currently no tasks on the cpu pretend there
1639 * is one of average load so that when a new task gets to
1640 * run here it will not get delayed by group starvation.
1643 weight = NICE_0_LOAD;
1645 sum_weight += weight;
1646 shares += tg->cfs_rq[i]->shares;
1650 rq_weight = sum_weight;
1652 if ((!shares && rq_weight) || shares > tg->shares)
1653 shares = tg->shares;
1655 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1656 shares = tg->shares;
1658 for_each_cpu(i, sched_domain_span(sd))
1659 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
1661 local_irq_restore(flags);
1667 * Compute the cpu's hierarchical load factor for each task group.
1668 * This needs to be done in a top-down fashion because the load of a child
1669 * group is a fraction of its parents load.
1671 static int tg_load_down(struct task_group *tg, void *data)
1674 long cpu = (long)data;
1677 load = cpu_rq(cpu)->load.weight;
1679 load = tg->parent->cfs_rq[cpu]->h_load;
1680 load *= tg->cfs_rq[cpu]->shares;
1681 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1684 tg->cfs_rq[cpu]->h_load = load;
1689 static void update_shares(struct sched_domain *sd)
1694 if (root_task_group_empty())
1697 now = cpu_clock(raw_smp_processor_id());
1698 elapsed = now - sd->last_update;
1700 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1701 sd->last_update = now;
1702 walk_tg_tree(tg_nop, tg_shares_up, sd);
1706 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1708 if (root_task_group_empty())
1711 spin_unlock(&rq->lock);
1713 spin_lock(&rq->lock);
1716 static void update_h_load(long cpu)
1718 if (root_task_group_empty())
1721 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1726 static inline void update_shares(struct sched_domain *sd)
1730 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1736 #ifdef CONFIG_PREEMPT
1738 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1741 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1742 * way at the expense of forcing extra atomic operations in all
1743 * invocations. This assures that the double_lock is acquired using the
1744 * same underlying policy as the spinlock_t on this architecture, which
1745 * reduces latency compared to the unfair variant below. However, it
1746 * also adds more overhead and therefore may reduce throughput.
1748 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1749 __releases(this_rq->lock)
1750 __acquires(busiest->lock)
1751 __acquires(this_rq->lock)
1753 spin_unlock(&this_rq->lock);
1754 double_rq_lock(this_rq, busiest);
1761 * Unfair double_lock_balance: Optimizes throughput at the expense of
1762 * latency by eliminating extra atomic operations when the locks are
1763 * already in proper order on entry. This favors lower cpu-ids and will
1764 * grant the double lock to lower cpus over higher ids under contention,
1765 * regardless of entry order into the function.
1767 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1768 __releases(this_rq->lock)
1769 __acquires(busiest->lock)
1770 __acquires(this_rq->lock)
1774 if (unlikely(!spin_trylock(&busiest->lock))) {
1775 if (busiest < this_rq) {
1776 spin_unlock(&this_rq->lock);
1777 spin_lock(&busiest->lock);
1778 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1781 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1786 #endif /* CONFIG_PREEMPT */
1789 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1791 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1793 if (unlikely(!irqs_disabled())) {
1794 /* printk() doesn't work good under rq->lock */
1795 spin_unlock(&this_rq->lock);
1799 return _double_lock_balance(this_rq, busiest);
1802 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1803 __releases(busiest->lock)
1805 spin_unlock(&busiest->lock);
1806 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1810 #ifdef CONFIG_FAIR_GROUP_SCHED
1811 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1814 cfs_rq->shares = shares;
1819 static void calc_load_account_active(struct rq *this_rq);
1820 static void update_sysctl(void);
1821 static int get_update_sysctl_factor(void);
1823 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1825 set_task_rq(p, cpu);
1828 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1829 * successfuly executed on another CPU. We must ensure that updates of
1830 * per-task data have been completed by this moment.
1833 task_thread_info(p)->cpu = cpu;
1837 #include "sched_stats.h"
1838 #include "sched_idletask.c"
1839 #include "sched_fair.c"
1840 #include "sched_rt.c"
1841 #ifdef CONFIG_SCHED_DEBUG
1842 # include "sched_debug.c"
1845 #define sched_class_highest (&rt_sched_class)
1846 #define for_each_class(class) \
1847 for (class = sched_class_highest; class; class = class->next)
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 = prio_to_weight[0] * 2;
1863 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
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 update_avg(u64 *avg, u64 sample)
1882 s64 diff = sample - *avg;
1886 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1889 p->se.start_runtime = p->se.sum_exec_runtime;
1891 sched_info_queued(p);
1892 p->sched_class->enqueue_task(rq, p, wakeup);
1896 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1899 if (p->se.last_wakeup) {
1900 update_avg(&p->se.avg_overlap,
1901 p->se.sum_exec_runtime - p->se.last_wakeup);
1902 p->se.last_wakeup = 0;
1904 update_avg(&p->se.avg_wakeup,
1905 sysctl_sched_wakeup_granularity);
1909 sched_info_dequeued(p);
1910 p->sched_class->dequeue_task(rq, p, sleep);
1915 * __normal_prio - return the priority that is based on the static prio
1917 static inline int __normal_prio(struct task_struct *p)
1919 return p->static_prio;
1923 * Calculate the expected normal priority: i.e. priority
1924 * without taking RT-inheritance into account. Might be
1925 * boosted by interactivity modifiers. Changes upon fork,
1926 * setprio syscalls, and whenever the interactivity
1927 * estimator recalculates.
1929 static inline int normal_prio(struct task_struct *p)
1933 if (task_has_rt_policy(p))
1934 prio = MAX_RT_PRIO-1 - p->rt_priority;
1936 prio = __normal_prio(p);
1941 * Calculate the current priority, i.e. the priority
1942 * taken into account by the scheduler. This value might
1943 * be boosted by RT tasks, or might be boosted by
1944 * interactivity modifiers. Will be RT if the task got
1945 * RT-boosted. If not then it returns p->normal_prio.
1947 static int effective_prio(struct task_struct *p)
1949 p->normal_prio = normal_prio(p);
1951 * If we are RT tasks or we were boosted to RT priority,
1952 * keep the priority unchanged. Otherwise, update priority
1953 * to the normal priority:
1955 if (!rt_prio(p->prio))
1956 return p->normal_prio;
1961 * activate_task - move a task to the runqueue.
1963 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1965 if (task_contributes_to_load(p))
1966 rq->nr_uninterruptible--;
1968 enqueue_task(rq, p, wakeup);
1973 * deactivate_task - remove a task from the runqueue.
1975 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1977 if (task_contributes_to_load(p))
1978 rq->nr_uninterruptible++;
1980 dequeue_task(rq, p, sleep);
1985 * task_curr - is this task currently executing on a CPU?
1986 * @p: the task in question.
1988 inline int task_curr(const struct task_struct *p)
1990 return cpu_curr(task_cpu(p)) == p;
1993 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1994 const struct sched_class *prev_class,
1995 int oldprio, int running)
1997 if (prev_class != p->sched_class) {
1998 if (prev_class->switched_from)
1999 prev_class->switched_from(rq, p, running);
2000 p->sched_class->switched_to(rq, p, running);
2002 p->sched_class->prio_changed(rq, p, oldprio, running);
2006 * kthread_bind - bind a just-created kthread to a cpu.
2007 * @p: thread created by kthread_create().
2008 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2010 * Description: This function is equivalent to set_cpus_allowed(),
2011 * except that @cpu doesn't need to be online, and the thread must be
2012 * stopped (i.e., just returned from kthread_create()).
2014 * Function lives here instead of kthread.c because it messes with
2015 * scheduler internals which require locking.
2017 void kthread_bind(struct task_struct *p, unsigned int cpu)
2019 struct rq *rq = cpu_rq(cpu);
2020 unsigned long flags;
2022 /* Must have done schedule() in kthread() before we set_task_cpu */
2023 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2028 spin_lock_irqsave(&rq->lock, flags);
2029 update_rq_clock(rq);
2030 set_task_cpu(p, cpu);
2031 p->cpus_allowed = cpumask_of_cpu(cpu);
2032 p->rt.nr_cpus_allowed = 1;
2033 p->flags |= PF_THREAD_BOUND;
2034 spin_unlock_irqrestore(&rq->lock, flags);
2036 EXPORT_SYMBOL(kthread_bind);
2040 * Is this task likely cache-hot:
2043 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2048 * Buddy candidates are cache hot:
2050 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2051 (&p->se == cfs_rq_of(&p->se)->next ||
2052 &p->se == cfs_rq_of(&p->se)->last))
2055 if (p->sched_class != &fair_sched_class)
2058 if (sysctl_sched_migration_cost == -1)
2060 if (sysctl_sched_migration_cost == 0)
2063 delta = now - p->se.exec_start;
2065 return delta < (s64)sysctl_sched_migration_cost;
2069 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2071 int old_cpu = task_cpu(p);
2072 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2073 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2075 trace_sched_migrate_task(p, new_cpu);
2077 if (old_cpu != new_cpu) {
2078 p->se.nr_migrations++;
2079 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2082 p->se.vruntime -= old_cfsrq->min_vruntime -
2083 new_cfsrq->min_vruntime;
2085 __set_task_cpu(p, new_cpu);
2088 struct migration_req {
2089 struct list_head list;
2091 struct task_struct *task;
2094 struct completion done;
2098 * The task's runqueue lock must be held.
2099 * Returns true if you have to wait for migration thread.
2102 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2104 struct rq *rq = task_rq(p);
2107 * If the task is not on a runqueue (and not running), then
2108 * it is sufficient to simply update the task's cpu field.
2110 if (!p->se.on_rq && !task_running(rq, p)) {
2111 update_rq_clock(rq);
2112 set_task_cpu(p, dest_cpu);
2116 init_completion(&req->done);
2118 req->dest_cpu = dest_cpu;
2119 list_add(&req->list, &rq->migration_queue);
2125 * wait_task_context_switch - wait for a thread to complete at least one
2128 * @p must not be current.
2130 void wait_task_context_switch(struct task_struct *p)
2132 unsigned long nvcsw, nivcsw, flags;
2140 * The runqueue is assigned before the actual context
2141 * switch. We need to take the runqueue lock.
2143 * We could check initially without the lock but it is
2144 * very likely that we need to take the lock in every
2147 rq = task_rq_lock(p, &flags);
2148 running = task_running(rq, p);
2149 task_rq_unlock(rq, &flags);
2151 if (likely(!running))
2154 * The switch count is incremented before the actual
2155 * context switch. We thus wait for two switches to be
2156 * sure at least one completed.
2158 if ((p->nvcsw - nvcsw) > 1)
2160 if ((p->nivcsw - nivcsw) > 1)
2168 * wait_task_inactive - wait for a thread to unschedule.
2170 * If @match_state is nonzero, it's the @p->state value just checked and
2171 * not expected to change. If it changes, i.e. @p might have woken up,
2172 * then return zero. When we succeed in waiting for @p to be off its CPU,
2173 * we return a positive number (its total switch count). If a second call
2174 * a short while later returns the same number, the caller can be sure that
2175 * @p has remained unscheduled the whole time.
2177 * The caller must ensure that the task *will* unschedule sometime soon,
2178 * else this function might spin for a *long* time. This function can't
2179 * be called with interrupts off, or it may introduce deadlock with
2180 * smp_call_function() if an IPI is sent by the same process we are
2181 * waiting to become inactive.
2183 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2185 unsigned long flags;
2192 * We do the initial early heuristics without holding
2193 * any task-queue locks at all. We'll only try to get
2194 * the runqueue lock when things look like they will
2200 * If the task is actively running on another CPU
2201 * still, just relax and busy-wait without holding
2204 * NOTE! Since we don't hold any locks, it's not
2205 * even sure that "rq" stays as the right runqueue!
2206 * But we don't care, since "task_running()" will
2207 * return false if the runqueue has changed and p
2208 * is actually now running somewhere else!
2210 while (task_running(rq, p)) {
2211 if (match_state && unlikely(p->state != match_state))
2217 * Ok, time to look more closely! We need the rq
2218 * lock now, to be *sure*. If we're wrong, we'll
2219 * just go back and repeat.
2221 rq = task_rq_lock(p, &flags);
2222 trace_sched_wait_task(rq, p);
2223 running = task_running(rq, p);
2224 on_rq = p->se.on_rq;
2226 if (!match_state || p->state == match_state)
2227 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2228 task_rq_unlock(rq, &flags);
2231 * If it changed from the expected state, bail out now.
2233 if (unlikely(!ncsw))
2237 * Was it really running after all now that we
2238 * checked with the proper locks actually held?
2240 * Oops. Go back and try again..
2242 if (unlikely(running)) {
2248 * It's not enough that it's not actively running,
2249 * it must be off the runqueue _entirely_, and not
2252 * So if it was still runnable (but just not actively
2253 * running right now), it's preempted, and we should
2254 * yield - it could be a while.
2256 if (unlikely(on_rq)) {
2257 schedule_timeout_uninterruptible(1);
2262 * Ahh, all good. It wasn't running, and it wasn't
2263 * runnable, which means that it will never become
2264 * running in the future either. We're all done!
2273 * kick_process - kick a running thread to enter/exit the kernel
2274 * @p: the to-be-kicked thread
2276 * Cause a process which is running on another CPU to enter
2277 * kernel-mode, without any delay. (to get signals handled.)
2279 * NOTE: this function doesnt have to take the runqueue lock,
2280 * because all it wants to ensure is that the remote task enters
2281 * the kernel. If the IPI races and the task has been migrated
2282 * to another CPU then no harm is done and the purpose has been
2285 void kick_process(struct task_struct *p)
2291 if ((cpu != smp_processor_id()) && task_curr(p))
2292 smp_send_reschedule(cpu);
2295 EXPORT_SYMBOL_GPL(kick_process);
2296 #endif /* CONFIG_SMP */
2299 * task_oncpu_function_call - call a function on the cpu on which a task runs
2300 * @p: the task to evaluate
2301 * @func: the function to be called
2302 * @info: the function call argument
2304 * Calls the function @func when the task is currently running. This might
2305 * be on the current CPU, which just calls the function directly
2307 void task_oncpu_function_call(struct task_struct *p,
2308 void (*func) (void *info), void *info)
2315 smp_call_function_single(cpu, func, info, 1);
2321 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2323 return p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2328 * try_to_wake_up - wake up a thread
2329 * @p: the to-be-woken-up thread
2330 * @state: the mask of task states that can be woken
2331 * @sync: do a synchronous wakeup?
2333 * Put it on the run-queue if it's not already there. The "current"
2334 * thread is always on the run-queue (except when the actual
2335 * re-schedule is in progress), and as such you're allowed to do
2336 * the simpler "current->state = TASK_RUNNING" to mark yourself
2337 * runnable without the overhead of this.
2339 * returns failure only if the task is already active.
2341 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2344 int cpu, orig_cpu, this_cpu, success = 0;
2345 unsigned long flags;
2346 struct rq *rq, *orig_rq;
2348 if (!sched_feat(SYNC_WAKEUPS))
2349 wake_flags &= ~WF_SYNC;
2351 this_cpu = get_cpu();
2354 rq = orig_rq = task_rq_lock(p, &flags);
2355 update_rq_clock(rq);
2356 if (!(p->state & state))
2366 if (unlikely(task_running(rq, p)))
2370 * In order to handle concurrent wakeups and release the rq->lock
2371 * we put the task in TASK_WAKING state.
2373 * First fix up the nr_uninterruptible count:
2375 if (task_contributes_to_load(p))
2376 rq->nr_uninterruptible--;
2377 p->state = TASK_WAKING;
2378 __task_rq_unlock(rq);
2380 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2381 if (cpu != orig_cpu)
2382 set_task_cpu(p, cpu);
2384 rq = __task_rq_lock(p);
2385 update_rq_clock(rq);
2387 WARN_ON(p->state != TASK_WAKING);
2390 #ifdef CONFIG_SCHEDSTATS
2391 schedstat_inc(rq, ttwu_count);
2392 if (cpu == this_cpu)
2393 schedstat_inc(rq, ttwu_local);
2395 struct sched_domain *sd;
2396 for_each_domain(this_cpu, sd) {
2397 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2398 schedstat_inc(sd, ttwu_wake_remote);
2403 #endif /* CONFIG_SCHEDSTATS */
2406 #endif /* CONFIG_SMP */
2407 schedstat_inc(p, se.nr_wakeups);
2408 if (wake_flags & WF_SYNC)
2409 schedstat_inc(p, se.nr_wakeups_sync);
2410 if (orig_cpu != cpu)
2411 schedstat_inc(p, se.nr_wakeups_migrate);
2412 if (cpu == this_cpu)
2413 schedstat_inc(p, se.nr_wakeups_local);
2415 schedstat_inc(p, se.nr_wakeups_remote);
2416 activate_task(rq, p, 1);
2420 * Only attribute actual wakeups done by this task.
2422 if (!in_interrupt()) {
2423 struct sched_entity *se = ¤t->se;
2424 u64 sample = se->sum_exec_runtime;
2426 if (se->last_wakeup)
2427 sample -= se->last_wakeup;
2429 sample -= se->start_runtime;
2430 update_avg(&se->avg_wakeup, sample);
2432 se->last_wakeup = se->sum_exec_runtime;
2436 trace_sched_wakeup(rq, p, success);
2437 check_preempt_curr(rq, p, wake_flags);
2439 p->state = TASK_RUNNING;
2441 if (p->sched_class->task_wake_up)
2442 p->sched_class->task_wake_up(rq, p);
2444 if (unlikely(rq->idle_stamp)) {
2445 u64 delta = rq->clock - rq->idle_stamp;
2446 u64 max = 2*sysctl_sched_migration_cost;
2451 update_avg(&rq->avg_idle, delta);
2456 task_rq_unlock(rq, &flags);
2463 * wake_up_process - Wake up a specific process
2464 * @p: The process to be woken up.
2466 * Attempt to wake up the nominated process and move it to the set of runnable
2467 * processes. Returns 1 if the process was woken up, 0 if it was already
2470 * It may be assumed that this function implies a write memory barrier before
2471 * changing the task state if and only if any tasks are woken up.
2473 int wake_up_process(struct task_struct *p)
2475 return try_to_wake_up(p, TASK_ALL, 0);
2477 EXPORT_SYMBOL(wake_up_process);
2479 int wake_up_state(struct task_struct *p, unsigned int state)
2481 return try_to_wake_up(p, state, 0);
2485 * Perform scheduler related setup for a newly forked process p.
2486 * p is forked by current.
2488 * __sched_fork() is basic setup used by init_idle() too:
2490 static void __sched_fork(struct task_struct *p)
2492 p->se.exec_start = 0;
2493 p->se.sum_exec_runtime = 0;
2494 p->se.prev_sum_exec_runtime = 0;
2495 p->se.nr_migrations = 0;
2496 p->se.last_wakeup = 0;
2497 p->se.avg_overlap = 0;
2498 p->se.start_runtime = 0;
2499 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2501 #ifdef CONFIG_SCHEDSTATS
2502 p->se.wait_start = 0;
2504 p->se.wait_count = 0;
2507 p->se.sleep_start = 0;
2508 p->se.sleep_max = 0;
2509 p->se.sum_sleep_runtime = 0;
2511 p->se.block_start = 0;
2512 p->se.block_max = 0;
2514 p->se.slice_max = 0;
2516 p->se.nr_migrations_cold = 0;
2517 p->se.nr_failed_migrations_affine = 0;
2518 p->se.nr_failed_migrations_running = 0;
2519 p->se.nr_failed_migrations_hot = 0;
2520 p->se.nr_forced_migrations = 0;
2522 p->se.nr_wakeups = 0;
2523 p->se.nr_wakeups_sync = 0;
2524 p->se.nr_wakeups_migrate = 0;
2525 p->se.nr_wakeups_local = 0;
2526 p->se.nr_wakeups_remote = 0;
2527 p->se.nr_wakeups_affine = 0;
2528 p->se.nr_wakeups_affine_attempts = 0;
2529 p->se.nr_wakeups_passive = 0;
2530 p->se.nr_wakeups_idle = 0;
2534 INIT_LIST_HEAD(&p->rt.run_list);
2536 INIT_LIST_HEAD(&p->se.group_node);
2538 #ifdef CONFIG_PREEMPT_NOTIFIERS
2539 INIT_HLIST_HEAD(&p->preempt_notifiers);
2543 * We mark the process as running here, but have not actually
2544 * inserted it onto the runqueue yet. This guarantees that
2545 * nobody will actually run it, and a signal or other external
2546 * event cannot wake it up and insert it on the runqueue either.
2548 p->state = TASK_RUNNING;
2552 * fork()/clone()-time setup:
2554 void sched_fork(struct task_struct *p, int clone_flags)
2556 int cpu = get_cpu();
2561 * Revert to default priority/policy on fork if requested.
2563 if (unlikely(p->sched_reset_on_fork)) {
2564 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2565 p->policy = SCHED_NORMAL;
2566 p->normal_prio = p->static_prio;
2569 if (PRIO_TO_NICE(p->static_prio) < 0) {
2570 p->static_prio = NICE_TO_PRIO(0);
2571 p->normal_prio = p->static_prio;
2576 * We don't need the reset flag anymore after the fork. It has
2577 * fulfilled its duty:
2579 p->sched_reset_on_fork = 0;
2583 * Make sure we do not leak PI boosting priority to the child.
2585 p->prio = current->normal_prio;
2587 if (!rt_prio(p->prio))
2588 p->sched_class = &fair_sched_class;
2590 if (p->sched_class->task_fork)
2591 p->sched_class->task_fork(p);
2594 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2596 set_task_cpu(p, cpu);
2598 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2599 if (likely(sched_info_on()))
2600 memset(&p->sched_info, 0, sizeof(p->sched_info));
2602 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2605 #ifdef CONFIG_PREEMPT
2606 /* Want to start with kernel preemption disabled. */
2607 task_thread_info(p)->preempt_count = 1;
2609 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2615 * wake_up_new_task - wake up a newly created task for the first time.
2617 * This function will do some initial scheduler statistics housekeeping
2618 * that must be done for every newly created context, then puts the task
2619 * on the runqueue and wakes it.
2621 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2623 unsigned long flags;
2626 rq = task_rq_lock(p, &flags);
2627 BUG_ON(p->state != TASK_RUNNING);
2628 update_rq_clock(rq);
2629 activate_task(rq, p, 0);
2630 trace_sched_wakeup_new(rq, p, 1);
2631 check_preempt_curr(rq, p, WF_FORK);
2633 if (p->sched_class->task_wake_up)
2634 p->sched_class->task_wake_up(rq, p);
2636 task_rq_unlock(rq, &flags);
2639 #ifdef CONFIG_PREEMPT_NOTIFIERS
2642 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2643 * @notifier: notifier struct to register
2645 void preempt_notifier_register(struct preempt_notifier *notifier)
2647 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2649 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2652 * preempt_notifier_unregister - no longer interested in preemption notifications
2653 * @notifier: notifier struct to unregister
2655 * This is safe to call from within a preemption notifier.
2657 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2659 hlist_del(¬ifier->link);
2661 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2663 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2665 struct preempt_notifier *notifier;
2666 struct hlist_node *node;
2668 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2669 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2673 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2674 struct task_struct *next)
2676 struct preempt_notifier *notifier;
2677 struct hlist_node *node;
2679 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2680 notifier->ops->sched_out(notifier, next);
2683 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2685 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2690 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2691 struct task_struct *next)
2695 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2698 * prepare_task_switch - prepare to switch tasks
2699 * @rq: the runqueue preparing to switch
2700 * @prev: the current task that is being switched out
2701 * @next: the task we are going to switch to.
2703 * This is called with the rq lock held and interrupts off. It must
2704 * be paired with a subsequent finish_task_switch after the context
2707 * prepare_task_switch sets up locking and calls architecture specific
2711 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2712 struct task_struct *next)
2714 fire_sched_out_preempt_notifiers(prev, next);
2715 prepare_lock_switch(rq, next);
2716 prepare_arch_switch(next);
2720 * finish_task_switch - clean up after a task-switch
2721 * @rq: runqueue associated with task-switch
2722 * @prev: the thread we just switched away from.
2724 * finish_task_switch must be called after the context switch, paired
2725 * with a prepare_task_switch call before the context switch.
2726 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2727 * and do any other architecture-specific cleanup actions.
2729 * Note that we may have delayed dropping an mm in context_switch(). If
2730 * so, we finish that here outside of the runqueue lock. (Doing it
2731 * with the lock held can cause deadlocks; see schedule() for
2734 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2735 __releases(rq->lock)
2737 struct mm_struct *mm = rq->prev_mm;
2743 * A task struct has one reference for the use as "current".
2744 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2745 * schedule one last time. The schedule call will never return, and
2746 * the scheduled task must drop that reference.
2747 * The test for TASK_DEAD must occur while the runqueue locks are
2748 * still held, otherwise prev could be scheduled on another cpu, die
2749 * there before we look at prev->state, and then the reference would
2751 * Manfred Spraul <manfred@colorfullife.com>
2753 prev_state = prev->state;
2754 finish_arch_switch(prev);
2755 perf_event_task_sched_in(current, cpu_of(rq));
2756 finish_lock_switch(rq, prev);
2758 fire_sched_in_preempt_notifiers(current);
2761 if (unlikely(prev_state == TASK_DEAD)) {
2763 * Remove function-return probe instances associated with this
2764 * task and put them back on the free list.
2766 kprobe_flush_task(prev);
2767 put_task_struct(prev);
2773 /* assumes rq->lock is held */
2774 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2776 if (prev->sched_class->pre_schedule)
2777 prev->sched_class->pre_schedule(rq, prev);
2780 /* rq->lock is NOT held, but preemption is disabled */
2781 static inline void post_schedule(struct rq *rq)
2783 if (rq->post_schedule) {
2784 unsigned long flags;
2786 spin_lock_irqsave(&rq->lock, flags);
2787 if (rq->curr->sched_class->post_schedule)
2788 rq->curr->sched_class->post_schedule(rq);
2789 spin_unlock_irqrestore(&rq->lock, flags);
2791 rq->post_schedule = 0;
2797 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2801 static inline void post_schedule(struct rq *rq)
2808 * schedule_tail - first thing a freshly forked thread must call.
2809 * @prev: the thread we just switched away from.
2811 asmlinkage void schedule_tail(struct task_struct *prev)
2812 __releases(rq->lock)
2814 struct rq *rq = this_rq();
2816 finish_task_switch(rq, prev);
2819 * FIXME: do we need to worry about rq being invalidated by the
2824 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2825 /* In this case, finish_task_switch does not reenable preemption */
2828 if (current->set_child_tid)
2829 put_user(task_pid_vnr(current), current->set_child_tid);
2833 * context_switch - switch to the new MM and the new
2834 * thread's register state.
2837 context_switch(struct rq *rq, struct task_struct *prev,
2838 struct task_struct *next)
2840 struct mm_struct *mm, *oldmm;
2842 prepare_task_switch(rq, prev, next);
2843 trace_sched_switch(rq, prev, next);
2845 oldmm = prev->active_mm;
2847 * For paravirt, this is coupled with an exit in switch_to to
2848 * combine the page table reload and the switch backend into
2851 arch_start_context_switch(prev);
2854 next->active_mm = oldmm;
2855 atomic_inc(&oldmm->mm_count);
2856 enter_lazy_tlb(oldmm, next);
2858 switch_mm(oldmm, mm, next);
2860 if (likely(!prev->mm)) {
2861 prev->active_mm = NULL;
2862 rq->prev_mm = oldmm;
2865 * Since the runqueue lock will be released by the next
2866 * task (which is an invalid locking op but in the case
2867 * of the scheduler it's an obvious special-case), so we
2868 * do an early lockdep release here:
2870 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2871 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2874 /* Here we just switch the register state and the stack. */
2875 switch_to(prev, next, prev);
2879 * this_rq must be evaluated again because prev may have moved
2880 * CPUs since it called schedule(), thus the 'rq' on its stack
2881 * frame will be invalid.
2883 finish_task_switch(this_rq(), prev);
2887 * nr_running, nr_uninterruptible and nr_context_switches:
2889 * externally visible scheduler statistics: current number of runnable
2890 * threads, current number of uninterruptible-sleeping threads, total
2891 * number of context switches performed since bootup.
2893 unsigned long nr_running(void)
2895 unsigned long i, sum = 0;
2897 for_each_online_cpu(i)
2898 sum += cpu_rq(i)->nr_running;
2903 unsigned long nr_uninterruptible(void)
2905 unsigned long i, sum = 0;
2907 for_each_possible_cpu(i)
2908 sum += cpu_rq(i)->nr_uninterruptible;
2911 * Since we read the counters lockless, it might be slightly
2912 * inaccurate. Do not allow it to go below zero though:
2914 if (unlikely((long)sum < 0))
2920 unsigned long long nr_context_switches(void)
2923 unsigned long long sum = 0;
2925 for_each_possible_cpu(i)
2926 sum += cpu_rq(i)->nr_switches;
2931 unsigned long nr_iowait(void)
2933 unsigned long i, sum = 0;
2935 for_each_possible_cpu(i)
2936 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2941 unsigned long nr_iowait_cpu(void)
2943 struct rq *this = this_rq();
2944 return atomic_read(&this->nr_iowait);
2947 unsigned long this_cpu_load(void)
2949 struct rq *this = this_rq();
2950 return this->cpu_load[0];
2954 /* Variables and functions for calc_load */
2955 static atomic_long_t calc_load_tasks;
2956 static unsigned long calc_load_update;
2957 unsigned long avenrun[3];
2958 EXPORT_SYMBOL(avenrun);
2961 * get_avenrun - get the load average array
2962 * @loads: pointer to dest load array
2963 * @offset: offset to add
2964 * @shift: shift count to shift the result left
2966 * These values are estimates at best, so no need for locking.
2968 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2970 loads[0] = (avenrun[0] + offset) << shift;
2971 loads[1] = (avenrun[1] + offset) << shift;
2972 loads[2] = (avenrun[2] + offset) << shift;
2975 static unsigned long
2976 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2979 load += active * (FIXED_1 - exp);
2980 return load >> FSHIFT;
2984 * calc_load - update the avenrun load estimates 10 ticks after the
2985 * CPUs have updated calc_load_tasks.
2987 void calc_global_load(void)
2989 unsigned long upd = calc_load_update + 10;
2992 if (time_before(jiffies, upd))
2995 active = atomic_long_read(&calc_load_tasks);
2996 active = active > 0 ? active * FIXED_1 : 0;
2998 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2999 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3000 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3002 calc_load_update += LOAD_FREQ;
3006 * Either called from update_cpu_load() or from a cpu going idle
3008 static void calc_load_account_active(struct rq *this_rq)
3010 long nr_active, delta;
3012 nr_active = this_rq->nr_running;
3013 nr_active += (long) this_rq->nr_uninterruptible;
3015 if (nr_active != this_rq->calc_load_active) {
3016 delta = nr_active - this_rq->calc_load_active;
3017 this_rq->calc_load_active = nr_active;
3018 atomic_long_add(delta, &calc_load_tasks);
3023 * Update rq->cpu_load[] statistics. This function is usually called every
3024 * scheduler tick (TICK_NSEC).
3026 static void update_cpu_load(struct rq *this_rq)
3028 unsigned long this_load = this_rq->load.weight;
3031 this_rq->nr_load_updates++;
3033 /* Update our load: */
3034 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3035 unsigned long old_load, new_load;
3037 /* scale is effectively 1 << i now, and >> i divides by scale */
3039 old_load = this_rq->cpu_load[i];
3040 new_load = this_load;
3042 * Round up the averaging division if load is increasing. This
3043 * prevents us from getting stuck on 9 if the load is 10, for
3046 if (new_load > old_load)
3047 new_load += scale-1;
3048 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3051 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3052 this_rq->calc_load_update += LOAD_FREQ;
3053 calc_load_account_active(this_rq);
3060 * double_rq_lock - safely lock two runqueues
3062 * Note this does not disable interrupts like task_rq_lock,
3063 * you need to do so manually before calling.
3065 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3066 __acquires(rq1->lock)
3067 __acquires(rq2->lock)
3069 BUG_ON(!irqs_disabled());
3071 spin_lock(&rq1->lock);
3072 __acquire(rq2->lock); /* Fake it out ;) */
3075 spin_lock(&rq1->lock);
3076 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3078 spin_lock(&rq2->lock);
3079 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3082 update_rq_clock(rq1);
3083 update_rq_clock(rq2);
3087 * double_rq_unlock - safely unlock two runqueues
3089 * Note this does not restore interrupts like task_rq_unlock,
3090 * you need to do so manually after calling.
3092 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3093 __releases(rq1->lock)
3094 __releases(rq2->lock)
3096 spin_unlock(&rq1->lock);
3098 spin_unlock(&rq2->lock);
3100 __release(rq2->lock);
3104 * If dest_cpu is allowed for this process, migrate the task to it.
3105 * This is accomplished by forcing the cpu_allowed mask to only
3106 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3107 * the cpu_allowed mask is restored.
3109 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3111 struct migration_req req;
3112 unsigned long flags;
3115 rq = task_rq_lock(p, &flags);
3116 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3117 || unlikely(!cpu_active(dest_cpu)))
3120 /* force the process onto the specified CPU */
3121 if (migrate_task(p, dest_cpu, &req)) {
3122 /* Need to wait for migration thread (might exit: take ref). */
3123 struct task_struct *mt = rq->migration_thread;
3125 get_task_struct(mt);
3126 task_rq_unlock(rq, &flags);
3127 wake_up_process(mt);
3128 put_task_struct(mt);
3129 wait_for_completion(&req.done);
3134 task_rq_unlock(rq, &flags);
3138 * sched_exec - execve() is a valuable balancing opportunity, because at
3139 * this point the task has the smallest effective memory and cache footprint.
3141 void sched_exec(void)
3143 int new_cpu, this_cpu = get_cpu();
3144 new_cpu = select_task_rq(current, SD_BALANCE_EXEC, 0);
3146 if (new_cpu != this_cpu)
3147 sched_migrate_task(current, new_cpu);
3151 * pull_task - move a task from a remote runqueue to the local runqueue.
3152 * Both runqueues must be locked.
3154 static void pull_task(struct rq *src_rq, struct task_struct *p,
3155 struct rq *this_rq, int this_cpu)
3157 deactivate_task(src_rq, p, 0);
3158 set_task_cpu(p, this_cpu);
3159 activate_task(this_rq, p, 0);
3160 check_preempt_curr(this_rq, p, 0);
3164 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3167 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3168 struct sched_domain *sd, enum cpu_idle_type idle,
3171 int tsk_cache_hot = 0;
3173 * We do not migrate tasks that are:
3174 * 1) running (obviously), or
3175 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3176 * 3) are cache-hot on their current CPU.
3178 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3179 schedstat_inc(p, se.nr_failed_migrations_affine);
3184 if (task_running(rq, p)) {
3185 schedstat_inc(p, se.nr_failed_migrations_running);
3190 * Aggressive migration if:
3191 * 1) task is cache cold, or
3192 * 2) too many balance attempts have failed.
3195 tsk_cache_hot = task_hot(p, rq->clock, sd);
3196 if (!tsk_cache_hot ||
3197 sd->nr_balance_failed > sd->cache_nice_tries) {
3198 #ifdef CONFIG_SCHEDSTATS
3199 if (tsk_cache_hot) {
3200 schedstat_inc(sd, lb_hot_gained[idle]);
3201 schedstat_inc(p, se.nr_forced_migrations);
3207 if (tsk_cache_hot) {
3208 schedstat_inc(p, se.nr_failed_migrations_hot);
3214 static unsigned long
3215 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3216 unsigned long max_load_move, struct sched_domain *sd,
3217 enum cpu_idle_type idle, int *all_pinned,
3218 int *this_best_prio, struct rq_iterator *iterator)
3220 int loops = 0, pulled = 0, pinned = 0;
3221 struct task_struct *p;
3222 long rem_load_move = max_load_move;
3224 if (max_load_move == 0)
3230 * Start the load-balancing iterator:
3232 p = iterator->start(iterator->arg);
3234 if (!p || loops++ > sysctl_sched_nr_migrate)
3237 if ((p->se.load.weight >> 1) > rem_load_move ||
3238 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3239 p = iterator->next(iterator->arg);
3243 pull_task(busiest, p, this_rq, this_cpu);
3245 rem_load_move -= p->se.load.weight;
3247 #ifdef CONFIG_PREEMPT
3249 * NEWIDLE balancing is a source of latency, so preemptible kernels
3250 * will stop after the first task is pulled to minimize the critical
3253 if (idle == CPU_NEWLY_IDLE)
3258 * We only want to steal up to the prescribed amount of weighted load.
3260 if (rem_load_move > 0) {
3261 if (p->prio < *this_best_prio)
3262 *this_best_prio = p->prio;
3263 p = iterator->next(iterator->arg);
3268 * Right now, this is one of only two places pull_task() is called,
3269 * so we can safely collect pull_task() stats here rather than
3270 * inside pull_task().
3272 schedstat_add(sd, lb_gained[idle], pulled);
3275 *all_pinned = pinned;
3277 return max_load_move - rem_load_move;
3281 * move_tasks tries to move up to max_load_move weighted load from busiest to
3282 * this_rq, as part of a balancing operation within domain "sd".
3283 * Returns 1 if successful and 0 otherwise.
3285 * Called with both runqueues locked.
3287 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3288 unsigned long max_load_move,
3289 struct sched_domain *sd, enum cpu_idle_type idle,
3292 const struct sched_class *class = sched_class_highest;
3293 unsigned long total_load_moved = 0;
3294 int this_best_prio = this_rq->curr->prio;
3298 class->load_balance(this_rq, this_cpu, busiest,
3299 max_load_move - total_load_moved,
3300 sd, idle, all_pinned, &this_best_prio);
3301 class = class->next;
3303 #ifdef CONFIG_PREEMPT
3305 * NEWIDLE balancing is a source of latency, so preemptible
3306 * kernels will stop after the first task is pulled to minimize
3307 * the critical section.
3309 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3312 } while (class && max_load_move > total_load_moved);
3314 return total_load_moved > 0;
3318 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3319 struct sched_domain *sd, enum cpu_idle_type idle,
3320 struct rq_iterator *iterator)
3322 struct task_struct *p = iterator->start(iterator->arg);
3326 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3327 pull_task(busiest, p, this_rq, this_cpu);
3329 * Right now, this is only the second place pull_task()
3330 * is called, so we can safely collect pull_task()
3331 * stats here rather than inside pull_task().
3333 schedstat_inc(sd, lb_gained[idle]);
3337 p = iterator->next(iterator->arg);
3344 * move_one_task tries to move exactly one task from busiest to this_rq, as
3345 * part of active balancing operations within "domain".
3346 * Returns 1 if successful and 0 otherwise.
3348 * Called with both runqueues locked.
3350 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3351 struct sched_domain *sd, enum cpu_idle_type idle)
3353 const struct sched_class *class;
3355 for_each_class(class) {
3356 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3362 /********** Helpers for find_busiest_group ************************/
3364 * sd_lb_stats - Structure to store the statistics of a sched_domain
3365 * during load balancing.
3367 struct sd_lb_stats {
3368 struct sched_group *busiest; /* Busiest group in this sd */
3369 struct sched_group *this; /* Local group in this sd */
3370 unsigned long total_load; /* Total load of all groups in sd */
3371 unsigned long total_pwr; /* Total power of all groups in sd */
3372 unsigned long avg_load; /* Average load across all groups in sd */
3374 /** Statistics of this group */
3375 unsigned long this_load;
3376 unsigned long this_load_per_task;
3377 unsigned long this_nr_running;
3379 /* Statistics of the busiest group */
3380 unsigned long max_load;
3381 unsigned long busiest_load_per_task;
3382 unsigned long busiest_nr_running;
3384 int group_imb; /* Is there imbalance in this sd */
3385 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3386 int power_savings_balance; /* Is powersave balance needed for this sd */
3387 struct sched_group *group_min; /* Least loaded group in sd */
3388 struct sched_group *group_leader; /* Group which relieves group_min */
3389 unsigned long min_load_per_task; /* load_per_task in group_min */
3390 unsigned long leader_nr_running; /* Nr running of group_leader */
3391 unsigned long min_nr_running; /* Nr running of group_min */
3396 * sg_lb_stats - stats of a sched_group required for load_balancing
3398 struct sg_lb_stats {
3399 unsigned long avg_load; /*Avg load across the CPUs of the group */
3400 unsigned long group_load; /* Total load over the CPUs of the group */
3401 unsigned long sum_nr_running; /* Nr tasks running in the group */
3402 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3403 unsigned long group_capacity;
3404 int group_imb; /* Is there an imbalance in the group ? */
3408 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3409 * @group: The group whose first cpu is to be returned.
3411 static inline unsigned int group_first_cpu(struct sched_group *group)
3413 return cpumask_first(sched_group_cpus(group));
3417 * get_sd_load_idx - Obtain the load index for a given sched domain.
3418 * @sd: The sched_domain whose load_idx is to be obtained.
3419 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3421 static inline int get_sd_load_idx(struct sched_domain *sd,
3422 enum cpu_idle_type idle)
3428 load_idx = sd->busy_idx;
3431 case CPU_NEWLY_IDLE:
3432 load_idx = sd->newidle_idx;
3435 load_idx = sd->idle_idx;
3443 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3445 * init_sd_power_savings_stats - Initialize power savings statistics for
3446 * the given sched_domain, during load balancing.
3448 * @sd: Sched domain whose power-savings statistics are to be initialized.
3449 * @sds: Variable containing the statistics for sd.
3450 * @idle: Idle status of the CPU at which we're performing load-balancing.
3452 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3453 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3456 * Busy processors will not participate in power savings
3459 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3460 sds->power_savings_balance = 0;
3462 sds->power_savings_balance = 1;
3463 sds->min_nr_running = ULONG_MAX;
3464 sds->leader_nr_running = 0;
3469 * update_sd_power_savings_stats - Update the power saving stats for a
3470 * sched_domain while performing load balancing.
3472 * @group: sched_group belonging to the sched_domain under consideration.
3473 * @sds: Variable containing the statistics of the sched_domain
3474 * @local_group: Does group contain the CPU for which we're performing
3476 * @sgs: Variable containing the statistics of the group.
3478 static inline void update_sd_power_savings_stats(struct sched_group *group,
3479 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3482 if (!sds->power_savings_balance)
3486 * If the local group is idle or completely loaded
3487 * no need to do power savings balance at this domain
3489 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3490 !sds->this_nr_running))
3491 sds->power_savings_balance = 0;
3494 * If a group is already running at full capacity or idle,
3495 * don't include that group in power savings calculations
3497 if (!sds->power_savings_balance ||
3498 sgs->sum_nr_running >= sgs->group_capacity ||
3499 !sgs->sum_nr_running)
3503 * Calculate the group which has the least non-idle load.
3504 * This is the group from where we need to pick up the load
3507 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3508 (sgs->sum_nr_running == sds->min_nr_running &&
3509 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3510 sds->group_min = group;
3511 sds->min_nr_running = sgs->sum_nr_running;
3512 sds->min_load_per_task = sgs->sum_weighted_load /
3513 sgs->sum_nr_running;
3517 * Calculate the group which is almost near its
3518 * capacity but still has some space to pick up some load
3519 * from other group and save more power
3521 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3524 if (sgs->sum_nr_running > sds->leader_nr_running ||
3525 (sgs->sum_nr_running == sds->leader_nr_running &&
3526 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3527 sds->group_leader = group;
3528 sds->leader_nr_running = sgs->sum_nr_running;
3533 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3534 * @sds: Variable containing the statistics of the sched_domain
3535 * under consideration.
3536 * @this_cpu: Cpu at which we're currently performing load-balancing.
3537 * @imbalance: Variable to store the imbalance.
3540 * Check if we have potential to perform some power-savings balance.
3541 * If yes, set the busiest group to be the least loaded group in the
3542 * sched_domain, so that it's CPUs can be put to idle.
3544 * Returns 1 if there is potential to perform power-savings balance.
3547 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3548 int this_cpu, unsigned long *imbalance)
3550 if (!sds->power_savings_balance)
3553 if (sds->this != sds->group_leader ||
3554 sds->group_leader == sds->group_min)
3557 *imbalance = sds->min_load_per_task;
3558 sds->busiest = sds->group_min;
3563 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3564 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3565 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3570 static inline void update_sd_power_savings_stats(struct sched_group *group,
3571 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3576 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3577 int this_cpu, unsigned long *imbalance)
3581 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3584 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3586 return SCHED_LOAD_SCALE;
3589 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3591 return default_scale_freq_power(sd, cpu);
3594 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3596 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3597 unsigned long smt_gain = sd->smt_gain;
3604 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3606 return default_scale_smt_power(sd, cpu);
3609 unsigned long scale_rt_power(int cpu)
3611 struct rq *rq = cpu_rq(cpu);
3612 u64 total, available;
3614 sched_avg_update(rq);
3616 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3617 available = total - rq->rt_avg;
3619 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3620 total = SCHED_LOAD_SCALE;
3622 total >>= SCHED_LOAD_SHIFT;
3624 return div_u64(available, total);
3627 static void update_cpu_power(struct sched_domain *sd, int cpu)
3629 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3630 unsigned long power = SCHED_LOAD_SCALE;
3631 struct sched_group *sdg = sd->groups;
3633 if (sched_feat(ARCH_POWER))
3634 power *= arch_scale_freq_power(sd, cpu);
3636 power *= default_scale_freq_power(sd, cpu);
3638 power >>= SCHED_LOAD_SHIFT;
3640 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3641 if (sched_feat(ARCH_POWER))
3642 power *= arch_scale_smt_power(sd, cpu);
3644 power *= default_scale_smt_power(sd, cpu);
3646 power >>= SCHED_LOAD_SHIFT;
3649 power *= scale_rt_power(cpu);
3650 power >>= SCHED_LOAD_SHIFT;
3655 sdg->cpu_power = power;
3658 static void update_group_power(struct sched_domain *sd, int cpu)
3660 struct sched_domain *child = sd->child;
3661 struct sched_group *group, *sdg = sd->groups;
3662 unsigned long power;
3665 update_cpu_power(sd, cpu);
3671 group = child->groups;
3673 power += group->cpu_power;
3674 group = group->next;
3675 } while (group != child->groups);
3677 sdg->cpu_power = power;
3681 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3682 * @sd: The sched_domain whose statistics are to be updated.
3683 * @group: sched_group whose statistics are to be updated.
3684 * @this_cpu: Cpu for which load balance is currently performed.
3685 * @idle: Idle status of this_cpu
3686 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3687 * @sd_idle: Idle status of the sched_domain containing group.
3688 * @local_group: Does group contain this_cpu.
3689 * @cpus: Set of cpus considered for load balancing.
3690 * @balance: Should we balance.
3691 * @sgs: variable to hold the statistics for this group.
3693 static inline void update_sg_lb_stats(struct sched_domain *sd,
3694 struct sched_group *group, int this_cpu,
3695 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3696 int local_group, const struct cpumask *cpus,
3697 int *balance, struct sg_lb_stats *sgs)
3699 unsigned long load, max_cpu_load, min_cpu_load;
3701 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3702 unsigned long sum_avg_load_per_task;
3703 unsigned long avg_load_per_task;
3706 balance_cpu = group_first_cpu(group);
3707 if (balance_cpu == this_cpu)
3708 update_group_power(sd, this_cpu);
3711 /* Tally up the load of all CPUs in the group */
3712 sum_avg_load_per_task = avg_load_per_task = 0;
3714 min_cpu_load = ~0UL;
3716 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3717 struct rq *rq = cpu_rq(i);
3719 if (*sd_idle && rq->nr_running)
3722 /* Bias balancing toward cpus of our domain */
3724 if (idle_cpu(i) && !first_idle_cpu) {
3729 load = target_load(i, load_idx);
3731 load = source_load(i, load_idx);
3732 if (load > max_cpu_load)
3733 max_cpu_load = load;
3734 if (min_cpu_load > load)
3735 min_cpu_load = load;
3738 sgs->group_load += load;
3739 sgs->sum_nr_running += rq->nr_running;
3740 sgs->sum_weighted_load += weighted_cpuload(i);
3742 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3746 * First idle cpu or the first cpu(busiest) in this sched group
3747 * is eligible for doing load balancing at this and above
3748 * domains. In the newly idle case, we will allow all the cpu's
3749 * to do the newly idle load balance.
3751 if (idle != CPU_NEWLY_IDLE && local_group &&
3752 balance_cpu != this_cpu && balance) {
3757 /* Adjust by relative CPU power of the group */
3758 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3762 * Consider the group unbalanced when the imbalance is larger
3763 * than the average weight of two tasks.
3765 * APZ: with cgroup the avg task weight can vary wildly and
3766 * might not be a suitable number - should we keep a
3767 * normalized nr_running number somewhere that negates
3770 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3773 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3776 sgs->group_capacity =
3777 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3781 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3782 * @sd: sched_domain whose statistics are to be updated.
3783 * @this_cpu: Cpu for which load balance is currently performed.
3784 * @idle: Idle status of this_cpu
3785 * @sd_idle: Idle status of the sched_domain containing group.
3786 * @cpus: Set of cpus considered for load balancing.
3787 * @balance: Should we balance.
3788 * @sds: variable to hold the statistics for this sched_domain.
3790 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3791 enum cpu_idle_type idle, int *sd_idle,
3792 const struct cpumask *cpus, int *balance,
3793 struct sd_lb_stats *sds)
3795 struct sched_domain *child = sd->child;
3796 struct sched_group *group = sd->groups;
3797 struct sg_lb_stats sgs;
3798 int load_idx, prefer_sibling = 0;
3800 if (child && child->flags & SD_PREFER_SIBLING)
3803 init_sd_power_savings_stats(sd, sds, idle);
3804 load_idx = get_sd_load_idx(sd, idle);
3809 local_group = cpumask_test_cpu(this_cpu,
3810 sched_group_cpus(group));
3811 memset(&sgs, 0, sizeof(sgs));
3812 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3813 local_group, cpus, balance, &sgs);
3815 if (local_group && balance && !(*balance))
3818 sds->total_load += sgs.group_load;
3819 sds->total_pwr += group->cpu_power;
3822 * In case the child domain prefers tasks go to siblings
3823 * first, lower the group capacity to one so that we'll try
3824 * and move all the excess tasks away.
3827 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3830 sds->this_load = sgs.avg_load;
3832 sds->this_nr_running = sgs.sum_nr_running;
3833 sds->this_load_per_task = sgs.sum_weighted_load;
3834 } else if (sgs.avg_load > sds->max_load &&
3835 (sgs.sum_nr_running > sgs.group_capacity ||
3837 sds->max_load = sgs.avg_load;
3838 sds->busiest = group;
3839 sds->busiest_nr_running = sgs.sum_nr_running;
3840 sds->busiest_load_per_task = sgs.sum_weighted_load;
3841 sds->group_imb = sgs.group_imb;
3844 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3845 group = group->next;
3846 } while (group != sd->groups);
3850 * fix_small_imbalance - Calculate the minor imbalance that exists
3851 * amongst the groups of a sched_domain, during
3853 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3854 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3855 * @imbalance: Variable to store the imbalance.
3857 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3858 int this_cpu, unsigned long *imbalance)
3860 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3861 unsigned int imbn = 2;
3863 if (sds->this_nr_running) {
3864 sds->this_load_per_task /= sds->this_nr_running;
3865 if (sds->busiest_load_per_task >
3866 sds->this_load_per_task)
3869 sds->this_load_per_task =
3870 cpu_avg_load_per_task(this_cpu);
3872 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3873 sds->busiest_load_per_task * imbn) {
3874 *imbalance = sds->busiest_load_per_task;
3879 * OK, we don't have enough imbalance to justify moving tasks,
3880 * however we may be able to increase total CPU power used by
3884 pwr_now += sds->busiest->cpu_power *
3885 min(sds->busiest_load_per_task, sds->max_load);
3886 pwr_now += sds->this->cpu_power *
3887 min(sds->this_load_per_task, sds->this_load);
3888 pwr_now /= SCHED_LOAD_SCALE;
3890 /* Amount of load we'd subtract */
3891 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3892 sds->busiest->cpu_power;
3893 if (sds->max_load > tmp)
3894 pwr_move += sds->busiest->cpu_power *
3895 min(sds->busiest_load_per_task, sds->max_load - tmp);
3897 /* Amount of load we'd add */
3898 if (sds->max_load * sds->busiest->cpu_power <
3899 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3900 tmp = (sds->max_load * sds->busiest->cpu_power) /
3901 sds->this->cpu_power;
3903 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3904 sds->this->cpu_power;
3905 pwr_move += sds->this->cpu_power *
3906 min(sds->this_load_per_task, sds->this_load + tmp);
3907 pwr_move /= SCHED_LOAD_SCALE;
3909 /* Move if we gain throughput */
3910 if (pwr_move > pwr_now)
3911 *imbalance = sds->busiest_load_per_task;
3915 * calculate_imbalance - Calculate the amount of imbalance present within the
3916 * groups of a given sched_domain during load balance.
3917 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3918 * @this_cpu: Cpu for which currently load balance is being performed.
3919 * @imbalance: The variable to store the imbalance.
3921 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3922 unsigned long *imbalance)
3924 unsigned long max_pull;
3926 * In the presence of smp nice balancing, certain scenarios can have
3927 * max load less than avg load(as we skip the groups at or below
3928 * its cpu_power, while calculating max_load..)
3930 if (sds->max_load < sds->avg_load) {
3932 return fix_small_imbalance(sds, this_cpu, imbalance);
3935 /* Don't want to pull so many tasks that a group would go idle */
3936 max_pull = min(sds->max_load - sds->avg_load,
3937 sds->max_load - sds->busiest_load_per_task);
3939 /* How much load to actually move to equalise the imbalance */
3940 *imbalance = min(max_pull * sds->busiest->cpu_power,
3941 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3945 * if *imbalance is less than the average load per runnable task
3946 * there is no gaurantee that any tasks will be moved so we'll have
3947 * a think about bumping its value to force at least one task to be
3950 if (*imbalance < sds->busiest_load_per_task)
3951 return fix_small_imbalance(sds, this_cpu, imbalance);
3954 /******* find_busiest_group() helpers end here *********************/
3957 * find_busiest_group - Returns the busiest group within the sched_domain
3958 * if there is an imbalance. If there isn't an imbalance, and
3959 * the user has opted for power-savings, it returns a group whose
3960 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3961 * such a group exists.
3963 * Also calculates the amount of weighted load which should be moved
3964 * to restore balance.
3966 * @sd: The sched_domain whose busiest group is to be returned.
3967 * @this_cpu: The cpu for which load balancing is currently being performed.
3968 * @imbalance: Variable which stores amount of weighted load which should
3969 * be moved to restore balance/put a group to idle.
3970 * @idle: The idle status of this_cpu.
3971 * @sd_idle: The idleness of sd
3972 * @cpus: The set of CPUs under consideration for load-balancing.
3973 * @balance: Pointer to a variable indicating if this_cpu
3974 * is the appropriate cpu to perform load balancing at this_level.
3976 * Returns: - the busiest group if imbalance exists.
3977 * - If no imbalance and user has opted for power-savings balance,
3978 * return the least loaded group whose CPUs can be
3979 * put to idle by rebalancing its tasks onto our group.
3981 static struct sched_group *
3982 find_busiest_group(struct sched_domain *sd, int this_cpu,
3983 unsigned long *imbalance, enum cpu_idle_type idle,
3984 int *sd_idle, const struct cpumask *cpus, int *balance)
3986 struct sd_lb_stats sds;
3988 memset(&sds, 0, sizeof(sds));
3991 * Compute the various statistics relavent for load balancing at
3994 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3997 /* Cases where imbalance does not exist from POV of this_cpu */
3998 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4000 * 2) There is no busy sibling group to pull from.
4001 * 3) This group is the busiest group.
4002 * 4) This group is more busy than the avg busieness at this
4004 * 5) The imbalance is within the specified limit.
4005 * 6) Any rebalance would lead to ping-pong
4007 if (balance && !(*balance))
4010 if (!sds.busiest || sds.busiest_nr_running == 0)
4013 if (sds.this_load >= sds.max_load)
4016 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4018 if (sds.this_load >= sds.avg_load)
4021 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4024 sds.busiest_load_per_task /= sds.busiest_nr_running;
4026 sds.busiest_load_per_task =
4027 min(sds.busiest_load_per_task, sds.avg_load);
4030 * We're trying to get all the cpus to the average_load, so we don't
4031 * want to push ourselves above the average load, nor do we wish to
4032 * reduce the max loaded cpu below the average load, as either of these
4033 * actions would just result in more rebalancing later, and ping-pong
4034 * tasks around. Thus we look for the minimum possible imbalance.
4035 * Negative imbalances (*we* are more loaded than anyone else) will
4036 * be counted as no imbalance for these purposes -- we can't fix that
4037 * by pulling tasks to us. Be careful of negative numbers as they'll
4038 * appear as very large values with unsigned longs.
4040 if (sds.max_load <= sds.busiest_load_per_task)
4043 /* Looks like there is an imbalance. Compute it */
4044 calculate_imbalance(&sds, this_cpu, imbalance);
4049 * There is no obvious imbalance. But check if we can do some balancing
4052 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4060 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4063 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4064 unsigned long imbalance, const struct cpumask *cpus)
4066 struct rq *busiest = NULL, *rq;
4067 unsigned long max_load = 0;
4070 for_each_cpu(i, sched_group_cpus(group)) {
4071 unsigned long power = power_of(i);
4072 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4075 if (!cpumask_test_cpu(i, cpus))
4079 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4082 if (capacity && rq->nr_running == 1 && wl > imbalance)
4085 if (wl > max_load) {
4095 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4096 * so long as it is large enough.
4098 #define MAX_PINNED_INTERVAL 512
4100 /* Working cpumask for load_balance and load_balance_newidle. */
4101 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4104 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4105 * tasks if there is an imbalance.
4107 static int load_balance(int this_cpu, struct rq *this_rq,
4108 struct sched_domain *sd, enum cpu_idle_type idle,
4111 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4112 struct sched_group *group;
4113 unsigned long imbalance;
4115 unsigned long flags;
4116 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4118 cpumask_copy(cpus, cpu_active_mask);
4121 * When power savings policy is enabled for the parent domain, idle
4122 * sibling can pick up load irrespective of busy siblings. In this case,
4123 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4124 * portraying it as CPU_NOT_IDLE.
4126 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4127 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4130 schedstat_inc(sd, lb_count[idle]);
4134 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4141 schedstat_inc(sd, lb_nobusyg[idle]);
4145 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4147 schedstat_inc(sd, lb_nobusyq[idle]);
4151 BUG_ON(busiest == this_rq);
4153 schedstat_add(sd, lb_imbalance[idle], imbalance);
4156 if (busiest->nr_running > 1) {
4158 * Attempt to move tasks. If find_busiest_group has found
4159 * an imbalance but busiest->nr_running <= 1, the group is
4160 * still unbalanced. ld_moved simply stays zero, so it is
4161 * correctly treated as an imbalance.
4163 local_irq_save(flags);
4164 double_rq_lock(this_rq, busiest);
4165 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4166 imbalance, sd, idle, &all_pinned);
4167 double_rq_unlock(this_rq, busiest);
4168 local_irq_restore(flags);
4171 * some other cpu did the load balance for us.
4173 if (ld_moved && this_cpu != smp_processor_id())
4174 resched_cpu(this_cpu);
4176 /* All tasks on this runqueue were pinned by CPU affinity */
4177 if (unlikely(all_pinned)) {
4178 cpumask_clear_cpu(cpu_of(busiest), cpus);
4179 if (!cpumask_empty(cpus))
4186 schedstat_inc(sd, lb_failed[idle]);
4187 sd->nr_balance_failed++;
4189 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4191 spin_lock_irqsave(&busiest->lock, flags);
4193 /* don't kick the migration_thread, if the curr
4194 * task on busiest cpu can't be moved to this_cpu
4196 if (!cpumask_test_cpu(this_cpu,
4197 &busiest->curr->cpus_allowed)) {
4198 spin_unlock_irqrestore(&busiest->lock, flags);
4200 goto out_one_pinned;
4203 if (!busiest->active_balance) {
4204 busiest->active_balance = 1;
4205 busiest->push_cpu = this_cpu;
4208 spin_unlock_irqrestore(&busiest->lock, flags);
4210 wake_up_process(busiest->migration_thread);
4213 * We've kicked active balancing, reset the failure
4216 sd->nr_balance_failed = sd->cache_nice_tries+1;
4219 sd->nr_balance_failed = 0;
4221 if (likely(!active_balance)) {
4222 /* We were unbalanced, so reset the balancing interval */
4223 sd->balance_interval = sd->min_interval;
4226 * If we've begun active balancing, start to back off. This
4227 * case may not be covered by the all_pinned logic if there
4228 * is only 1 task on the busy runqueue (because we don't call
4231 if (sd->balance_interval < sd->max_interval)
4232 sd->balance_interval *= 2;
4235 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4236 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4242 schedstat_inc(sd, lb_balanced[idle]);
4244 sd->nr_balance_failed = 0;
4247 /* tune up the balancing interval */
4248 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4249 (sd->balance_interval < sd->max_interval))
4250 sd->balance_interval *= 2;
4252 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4253 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4264 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4265 * tasks if there is an imbalance.
4267 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4268 * this_rq is locked.
4271 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4273 struct sched_group *group;
4274 struct rq *busiest = NULL;
4275 unsigned long imbalance;
4279 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4281 cpumask_copy(cpus, cpu_active_mask);
4284 * When power savings policy is enabled for the parent domain, idle
4285 * sibling can pick up load irrespective of busy siblings. In this case,
4286 * let the state of idle sibling percolate up as IDLE, instead of
4287 * portraying it as CPU_NOT_IDLE.
4289 if (sd->flags & SD_SHARE_CPUPOWER &&
4290 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4293 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4295 update_shares_locked(this_rq, sd);
4296 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4297 &sd_idle, cpus, NULL);
4299 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4303 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4305 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4309 BUG_ON(busiest == this_rq);
4311 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4314 if (busiest->nr_running > 1) {
4315 /* Attempt to move tasks */
4316 double_lock_balance(this_rq, busiest);
4317 /* this_rq->clock is already updated */
4318 update_rq_clock(busiest);
4319 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4320 imbalance, sd, CPU_NEWLY_IDLE,
4322 double_unlock_balance(this_rq, busiest);
4324 if (unlikely(all_pinned)) {
4325 cpumask_clear_cpu(cpu_of(busiest), cpus);
4326 if (!cpumask_empty(cpus))
4332 int active_balance = 0;
4334 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4335 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4336 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4339 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4342 if (sd->nr_balance_failed++ < 2)
4346 * The only task running in a non-idle cpu can be moved to this
4347 * cpu in an attempt to completely freeup the other CPU
4348 * package. The same method used to move task in load_balance()
4349 * have been extended for load_balance_newidle() to speedup
4350 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4352 * The package power saving logic comes from
4353 * find_busiest_group(). If there are no imbalance, then
4354 * f_b_g() will return NULL. However when sched_mc={1,2} then
4355 * f_b_g() will select a group from which a running task may be
4356 * pulled to this cpu in order to make the other package idle.
4357 * If there is no opportunity to make a package idle and if
4358 * there are no imbalance, then f_b_g() will return NULL and no
4359 * action will be taken in load_balance_newidle().
4361 * Under normal task pull operation due to imbalance, there
4362 * will be more than one task in the source run queue and
4363 * move_tasks() will succeed. ld_moved will be true and this
4364 * active balance code will not be triggered.
4367 /* Lock busiest in correct order while this_rq is held */
4368 double_lock_balance(this_rq, busiest);
4371 * don't kick the migration_thread, if the curr
4372 * task on busiest cpu can't be moved to this_cpu
4374 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4375 double_unlock_balance(this_rq, busiest);
4380 if (!busiest->active_balance) {
4381 busiest->active_balance = 1;
4382 busiest->push_cpu = this_cpu;
4386 double_unlock_balance(this_rq, busiest);
4388 * Should not call ttwu while holding a rq->lock
4390 spin_unlock(&this_rq->lock);
4392 wake_up_process(busiest->migration_thread);
4393 spin_lock(&this_rq->lock);
4396 sd->nr_balance_failed = 0;
4398 update_shares_locked(this_rq, sd);
4402 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4403 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4404 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4406 sd->nr_balance_failed = 0;
4412 * idle_balance is called by schedule() if this_cpu is about to become
4413 * idle. Attempts to pull tasks from other CPUs.
4415 static void idle_balance(int this_cpu, struct rq *this_rq)
4417 struct sched_domain *sd;
4418 int pulled_task = 0;
4419 unsigned long next_balance = jiffies + HZ;
4421 this_rq->idle_stamp = this_rq->clock;
4423 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4426 for_each_domain(this_cpu, sd) {
4427 unsigned long interval;
4429 if (!(sd->flags & SD_LOAD_BALANCE))
4432 if (sd->flags & SD_BALANCE_NEWIDLE)
4433 /* If we've pulled tasks over stop searching: */
4434 pulled_task = load_balance_newidle(this_cpu, this_rq,
4437 interval = msecs_to_jiffies(sd->balance_interval);
4438 if (time_after(next_balance, sd->last_balance + interval))
4439 next_balance = sd->last_balance + interval;
4441 this_rq->idle_stamp = 0;
4445 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4447 * We are going idle. next_balance may be set based on
4448 * a busy processor. So reset next_balance.
4450 this_rq->next_balance = next_balance;
4455 * active_load_balance is run by migration threads. It pushes running tasks
4456 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4457 * running on each physical CPU where possible, and avoids physical /
4458 * logical imbalances.
4460 * Called with busiest_rq locked.
4462 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4464 int target_cpu = busiest_rq->push_cpu;
4465 struct sched_domain *sd;
4466 struct rq *target_rq;
4468 /* Is there any task to move? */
4469 if (busiest_rq->nr_running <= 1)
4472 target_rq = cpu_rq(target_cpu);
4475 * This condition is "impossible", if it occurs
4476 * we need to fix it. Originally reported by
4477 * Bjorn Helgaas on a 128-cpu setup.
4479 BUG_ON(busiest_rq == target_rq);
4481 /* move a task from busiest_rq to target_rq */
4482 double_lock_balance(busiest_rq, target_rq);
4483 update_rq_clock(busiest_rq);
4484 update_rq_clock(target_rq);
4486 /* Search for an sd spanning us and the target CPU. */
4487 for_each_domain(target_cpu, sd) {
4488 if ((sd->flags & SD_LOAD_BALANCE) &&
4489 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4494 schedstat_inc(sd, alb_count);
4496 if (move_one_task(target_rq, target_cpu, busiest_rq,
4498 schedstat_inc(sd, alb_pushed);
4500 schedstat_inc(sd, alb_failed);
4502 double_unlock_balance(busiest_rq, target_rq);
4507 atomic_t load_balancer;
4508 cpumask_var_t cpu_mask;
4509 cpumask_var_t ilb_grp_nohz_mask;
4510 } nohz ____cacheline_aligned = {
4511 .load_balancer = ATOMIC_INIT(-1),
4514 int get_nohz_load_balancer(void)
4516 return atomic_read(&nohz.load_balancer);
4519 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4521 * lowest_flag_domain - Return lowest sched_domain containing flag.
4522 * @cpu: The cpu whose lowest level of sched domain is to
4524 * @flag: The flag to check for the lowest sched_domain
4525 * for the given cpu.
4527 * Returns the lowest sched_domain of a cpu which contains the given flag.
4529 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4531 struct sched_domain *sd;
4533 for_each_domain(cpu, sd)
4534 if (sd && (sd->flags & flag))
4541 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4542 * @cpu: The cpu whose domains we're iterating over.
4543 * @sd: variable holding the value of the power_savings_sd
4545 * @flag: The flag to filter the sched_domains to be iterated.
4547 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4548 * set, starting from the lowest sched_domain to the highest.
4550 #define for_each_flag_domain(cpu, sd, flag) \
4551 for (sd = lowest_flag_domain(cpu, flag); \
4552 (sd && (sd->flags & flag)); sd = sd->parent)
4555 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4556 * @ilb_group: group to be checked for semi-idleness
4558 * Returns: 1 if the group is semi-idle. 0 otherwise.
4560 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4561 * and atleast one non-idle CPU. This helper function checks if the given
4562 * sched_group is semi-idle or not.
4564 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4566 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4567 sched_group_cpus(ilb_group));
4570 * A sched_group is semi-idle when it has atleast one busy cpu
4571 * and atleast one idle cpu.
4573 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4576 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4582 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4583 * @cpu: The cpu which is nominating a new idle_load_balancer.
4585 * Returns: Returns the id of the idle load balancer if it exists,
4586 * Else, returns >= nr_cpu_ids.
4588 * This algorithm picks the idle load balancer such that it belongs to a
4589 * semi-idle powersavings sched_domain. The idea is to try and avoid
4590 * completely idle packages/cores just for the purpose of idle load balancing
4591 * when there are other idle cpu's which are better suited for that job.
4593 static int find_new_ilb(int cpu)
4595 struct sched_domain *sd;
4596 struct sched_group *ilb_group;
4599 * Have idle load balancer selection from semi-idle packages only
4600 * when power-aware load balancing is enabled
4602 if (!(sched_smt_power_savings || sched_mc_power_savings))
4606 * Optimize for the case when we have no idle CPUs or only one
4607 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4609 if (cpumask_weight(nohz.cpu_mask) < 2)
4612 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4613 ilb_group = sd->groups;
4616 if (is_semi_idle_group(ilb_group))
4617 return cpumask_first(nohz.ilb_grp_nohz_mask);
4619 ilb_group = ilb_group->next;
4621 } while (ilb_group != sd->groups);
4625 return cpumask_first(nohz.cpu_mask);
4627 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4628 static inline int find_new_ilb(int call_cpu)
4630 return cpumask_first(nohz.cpu_mask);
4635 * This routine will try to nominate the ilb (idle load balancing)
4636 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4637 * load balancing on behalf of all those cpus. If all the cpus in the system
4638 * go into this tickless mode, then there will be no ilb owner (as there is
4639 * no need for one) and all the cpus will sleep till the next wakeup event
4642 * For the ilb owner, tick is not stopped. And this tick will be used
4643 * for idle load balancing. ilb owner will still be part of
4646 * While stopping the tick, this cpu will become the ilb owner if there
4647 * is no other owner. And will be the owner till that cpu becomes busy
4648 * or if all cpus in the system stop their ticks at which point
4649 * there is no need for ilb owner.
4651 * When the ilb owner becomes busy, it nominates another owner, during the
4652 * next busy scheduler_tick()
4654 int select_nohz_load_balancer(int stop_tick)
4656 int cpu = smp_processor_id();
4659 cpu_rq(cpu)->in_nohz_recently = 1;
4661 if (!cpu_active(cpu)) {
4662 if (atomic_read(&nohz.load_balancer) != cpu)
4666 * If we are going offline and still the leader,
4669 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4675 cpumask_set_cpu(cpu, nohz.cpu_mask);
4677 /* time for ilb owner also to sleep */
4678 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4679 if (atomic_read(&nohz.load_balancer) == cpu)
4680 atomic_set(&nohz.load_balancer, -1);
4684 if (atomic_read(&nohz.load_balancer) == -1) {
4685 /* make me the ilb owner */
4686 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4688 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4691 if (!(sched_smt_power_savings ||
4692 sched_mc_power_savings))
4695 * Check to see if there is a more power-efficient
4698 new_ilb = find_new_ilb(cpu);
4699 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4700 atomic_set(&nohz.load_balancer, -1);
4701 resched_cpu(new_ilb);
4707 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4710 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4712 if (atomic_read(&nohz.load_balancer) == cpu)
4713 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4720 static DEFINE_SPINLOCK(balancing);
4723 * It checks each scheduling domain to see if it is due to be balanced,
4724 * and initiates a balancing operation if so.
4726 * Balancing parameters are set up in arch_init_sched_domains.
4728 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4731 struct rq *rq = cpu_rq(cpu);
4732 unsigned long interval;
4733 struct sched_domain *sd;
4734 /* Earliest time when we have to do rebalance again */
4735 unsigned long next_balance = jiffies + 60*HZ;
4736 int update_next_balance = 0;
4739 for_each_domain(cpu, sd) {
4740 if (!(sd->flags & SD_LOAD_BALANCE))
4743 interval = sd->balance_interval;
4744 if (idle != CPU_IDLE)
4745 interval *= sd->busy_factor;
4747 /* scale ms to jiffies */
4748 interval = msecs_to_jiffies(interval);
4749 if (unlikely(!interval))
4751 if (interval > HZ*NR_CPUS/10)
4752 interval = HZ*NR_CPUS/10;
4754 need_serialize = sd->flags & SD_SERIALIZE;
4756 if (need_serialize) {
4757 if (!spin_trylock(&balancing))
4761 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4762 if (load_balance(cpu, rq, sd, idle, &balance)) {
4764 * We've pulled tasks over so either we're no
4765 * longer idle, or one of our SMT siblings is
4768 idle = CPU_NOT_IDLE;
4770 sd->last_balance = jiffies;
4773 spin_unlock(&balancing);
4775 if (time_after(next_balance, sd->last_balance + interval)) {
4776 next_balance = sd->last_balance + interval;
4777 update_next_balance = 1;
4781 * Stop the load balance at this level. There is another
4782 * CPU in our sched group which is doing load balancing more
4790 * next_balance will be updated only when there is a need.
4791 * When the cpu is attached to null domain for ex, it will not be
4794 if (likely(update_next_balance))
4795 rq->next_balance = next_balance;
4799 * run_rebalance_domains is triggered when needed from the scheduler tick.
4800 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4801 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4803 static void run_rebalance_domains(struct softirq_action *h)
4805 int this_cpu = smp_processor_id();
4806 struct rq *this_rq = cpu_rq(this_cpu);
4807 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4808 CPU_IDLE : CPU_NOT_IDLE;
4810 rebalance_domains(this_cpu, idle);
4814 * If this cpu is the owner for idle load balancing, then do the
4815 * balancing on behalf of the other idle cpus whose ticks are
4818 if (this_rq->idle_at_tick &&
4819 atomic_read(&nohz.load_balancer) == this_cpu) {
4823 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4824 if (balance_cpu == this_cpu)
4828 * If this cpu gets work to do, stop the load balancing
4829 * work being done for other cpus. Next load
4830 * balancing owner will pick it up.
4835 rebalance_domains(balance_cpu, CPU_IDLE);
4837 rq = cpu_rq(balance_cpu);
4838 if (time_after(this_rq->next_balance, rq->next_balance))
4839 this_rq->next_balance = rq->next_balance;
4845 static inline int on_null_domain(int cpu)
4847 return !rcu_dereference(cpu_rq(cpu)->sd);
4851 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4853 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4854 * idle load balancing owner or decide to stop the periodic load balancing,
4855 * if the whole system is idle.
4857 static inline void trigger_load_balance(struct rq *rq, int cpu)
4861 * If we were in the nohz mode recently and busy at the current
4862 * scheduler tick, then check if we need to nominate new idle
4865 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4866 rq->in_nohz_recently = 0;
4868 if (atomic_read(&nohz.load_balancer) == cpu) {
4869 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4870 atomic_set(&nohz.load_balancer, -1);
4873 if (atomic_read(&nohz.load_balancer) == -1) {
4874 int ilb = find_new_ilb(cpu);
4876 if (ilb < nr_cpu_ids)
4882 * If this cpu is idle and doing idle load balancing for all the
4883 * cpus with ticks stopped, is it time for that to stop?
4885 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4886 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4892 * If this cpu is idle and the idle load balancing is done by
4893 * someone else, then no need raise the SCHED_SOFTIRQ
4895 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4896 cpumask_test_cpu(cpu, nohz.cpu_mask))
4899 /* Don't need to rebalance while attached to NULL domain */
4900 if (time_after_eq(jiffies, rq->next_balance) &&
4901 likely(!on_null_domain(cpu)))
4902 raise_softirq(SCHED_SOFTIRQ);
4905 #else /* CONFIG_SMP */
4908 * on UP we do not need to balance between CPUs:
4910 static inline void idle_balance(int cpu, struct rq *rq)
4916 DEFINE_PER_CPU(struct kernel_stat, kstat);
4918 EXPORT_PER_CPU_SYMBOL(kstat);
4921 * Return any ns on the sched_clock that have not yet been accounted in
4922 * @p in case that task is currently running.
4924 * Called with task_rq_lock() held on @rq.
4926 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4930 if (task_current(rq, p)) {
4931 update_rq_clock(rq);
4932 ns = rq->clock - p->se.exec_start;
4940 unsigned long long task_delta_exec(struct task_struct *p)
4942 unsigned long flags;
4946 rq = task_rq_lock(p, &flags);
4947 ns = do_task_delta_exec(p, rq);
4948 task_rq_unlock(rq, &flags);
4954 * Return accounted runtime for the task.
4955 * In case the task is currently running, return the runtime plus current's
4956 * pending runtime that have not been accounted yet.
4958 unsigned long long task_sched_runtime(struct task_struct *p)
4960 unsigned long flags;
4964 rq = task_rq_lock(p, &flags);
4965 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4966 task_rq_unlock(rq, &flags);
4972 * Return sum_exec_runtime for the thread group.
4973 * In case the task is currently running, return the sum plus current's
4974 * pending runtime that have not been accounted yet.
4976 * Note that the thread group might have other running tasks as well,
4977 * so the return value not includes other pending runtime that other
4978 * running tasks might have.
4980 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4982 struct task_cputime totals;
4983 unsigned long flags;
4987 rq = task_rq_lock(p, &flags);
4988 thread_group_cputime(p, &totals);
4989 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4990 task_rq_unlock(rq, &flags);
4996 * Account user cpu time to a process.
4997 * @p: the process that the cpu time gets accounted to
4998 * @cputime: the cpu time spent in user space since the last update
4999 * @cputime_scaled: cputime scaled by cpu frequency
5001 void account_user_time(struct task_struct *p, cputime_t cputime,
5002 cputime_t cputime_scaled)
5004 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5007 /* Add user time to process. */
5008 p->utime = cputime_add(p->utime, cputime);
5009 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5010 account_group_user_time(p, cputime);
5012 /* Add user time to cpustat. */
5013 tmp = cputime_to_cputime64(cputime);
5014 if (TASK_NICE(p) > 0)
5015 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5017 cpustat->user = cputime64_add(cpustat->user, tmp);
5019 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5020 /* Account for user time used */
5021 acct_update_integrals(p);
5025 * Account guest cpu time to a process.
5026 * @p: the process that the cpu time gets accounted to
5027 * @cputime: the cpu time spent in virtual machine since the last update
5028 * @cputime_scaled: cputime scaled by cpu frequency
5030 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5031 cputime_t cputime_scaled)
5034 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5036 tmp = cputime_to_cputime64(cputime);
5038 /* Add guest time to process. */
5039 p->utime = cputime_add(p->utime, cputime);
5040 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5041 account_group_user_time(p, cputime);
5042 p->gtime = cputime_add(p->gtime, cputime);
5044 /* Add guest time to cpustat. */
5045 if (TASK_NICE(p) > 0) {
5046 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5047 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5049 cpustat->user = cputime64_add(cpustat->user, tmp);
5050 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5055 * Account system cpu time to a process.
5056 * @p: the process that the cpu time gets accounted to
5057 * @hardirq_offset: the offset to subtract from hardirq_count()
5058 * @cputime: the cpu time spent in kernel space since the last update
5059 * @cputime_scaled: cputime scaled by cpu frequency
5061 void account_system_time(struct task_struct *p, int hardirq_offset,
5062 cputime_t cputime, cputime_t cputime_scaled)
5064 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5067 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5068 account_guest_time(p, cputime, cputime_scaled);
5072 /* Add system time to process. */
5073 p->stime = cputime_add(p->stime, cputime);
5074 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5075 account_group_system_time(p, cputime);
5077 /* Add system time to cpustat. */
5078 tmp = cputime_to_cputime64(cputime);
5079 if (hardirq_count() - hardirq_offset)
5080 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5081 else if (softirq_count())
5082 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5084 cpustat->system = cputime64_add(cpustat->system, tmp);
5086 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5088 /* Account for system time used */
5089 acct_update_integrals(p);
5093 * Account for involuntary wait time.
5094 * @steal: the cpu time spent in involuntary wait
5096 void account_steal_time(cputime_t cputime)
5098 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5099 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5101 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5105 * Account for idle time.
5106 * @cputime: the cpu time spent in idle wait
5108 void account_idle_time(cputime_t cputime)
5110 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5111 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5112 struct rq *rq = this_rq();
5114 if (atomic_read(&rq->nr_iowait) > 0)
5115 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5117 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5120 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5123 * Account a single tick of cpu time.
5124 * @p: the process that the cpu time gets accounted to
5125 * @user_tick: indicates if the tick is a user or a system tick
5127 void account_process_tick(struct task_struct *p, int user_tick)
5129 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5130 struct rq *rq = this_rq();
5133 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5134 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5135 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5138 account_idle_time(cputime_one_jiffy);
5142 * Account multiple ticks of steal time.
5143 * @p: the process from which the cpu time has been stolen
5144 * @ticks: number of stolen ticks
5146 void account_steal_ticks(unsigned long ticks)
5148 account_steal_time(jiffies_to_cputime(ticks));
5152 * Account multiple ticks of idle time.
5153 * @ticks: number of stolen ticks
5155 void account_idle_ticks(unsigned long ticks)
5157 account_idle_time(jiffies_to_cputime(ticks));
5163 * Use precise platform statistics if available:
5165 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5166 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5172 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5174 struct task_cputime cputime;
5176 thread_group_cputime(p, &cputime);
5178 *ut = cputime.utime;
5179 *st = cputime.stime;
5183 #ifndef nsecs_to_cputime
5184 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5187 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5189 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5192 * Use CFS's precise accounting:
5194 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5199 temp = (u64)(rtime * utime);
5200 do_div(temp, total);
5201 utime = (cputime_t)temp;
5206 * Compare with previous values, to keep monotonicity:
5208 p->prev_utime = max(p->prev_utime, utime);
5209 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5211 *ut = p->prev_utime;
5212 *st = p->prev_stime;
5216 * Must be called with siglock held.
5218 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5220 struct signal_struct *sig = p->signal;
5221 struct task_cputime cputime;
5222 cputime_t rtime, utime, total;
5224 thread_group_cputime(p, &cputime);
5226 total = cputime_add(cputime.utime, cputime.stime);
5227 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5232 temp = (u64)(rtime * cputime.utime);
5233 do_div(temp, total);
5234 utime = (cputime_t)temp;
5238 sig->prev_utime = max(sig->prev_utime, utime);
5239 sig->prev_stime = max(sig->prev_stime,
5240 cputime_sub(rtime, sig->prev_utime));
5242 *ut = sig->prev_utime;
5243 *st = sig->prev_stime;
5248 * This function gets called by the timer code, with HZ frequency.
5249 * We call it with interrupts disabled.
5251 * It also gets called by the fork code, when changing the parent's
5254 void scheduler_tick(void)
5256 int cpu = smp_processor_id();
5257 struct rq *rq = cpu_rq(cpu);
5258 struct task_struct *curr = rq->curr;
5262 spin_lock(&rq->lock);
5263 update_rq_clock(rq);
5264 update_cpu_load(rq);
5265 curr->sched_class->task_tick(rq, curr, 0);
5266 spin_unlock(&rq->lock);
5268 perf_event_task_tick(curr, cpu);
5271 rq->idle_at_tick = idle_cpu(cpu);
5272 trigger_load_balance(rq, cpu);
5276 notrace unsigned long get_parent_ip(unsigned long addr)
5278 if (in_lock_functions(addr)) {
5279 addr = CALLER_ADDR2;
5280 if (in_lock_functions(addr))
5281 addr = CALLER_ADDR3;
5286 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5287 defined(CONFIG_PREEMPT_TRACER))
5289 void __kprobes add_preempt_count(int val)
5291 #ifdef CONFIG_DEBUG_PREEMPT
5295 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5298 preempt_count() += val;
5299 #ifdef CONFIG_DEBUG_PREEMPT
5301 * Spinlock count overflowing soon?
5303 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5306 if (preempt_count() == val)
5307 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5309 EXPORT_SYMBOL(add_preempt_count);
5311 void __kprobes sub_preempt_count(int val)
5313 #ifdef CONFIG_DEBUG_PREEMPT
5317 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5320 * Is the spinlock portion underflowing?
5322 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5323 !(preempt_count() & PREEMPT_MASK)))
5327 if (preempt_count() == val)
5328 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5329 preempt_count() -= val;
5331 EXPORT_SYMBOL(sub_preempt_count);
5336 * Print scheduling while atomic bug:
5338 static noinline void __schedule_bug(struct task_struct *prev)
5340 struct pt_regs *regs = get_irq_regs();
5342 pr_err("BUG: scheduling while atomic: %s/%d/0x%08x\n",
5343 prev->comm, prev->pid, preempt_count());
5345 debug_show_held_locks(prev);
5347 if (irqs_disabled())
5348 print_irqtrace_events(prev);
5357 * Various schedule()-time debugging checks and statistics:
5359 static inline void schedule_debug(struct task_struct *prev)
5362 * Test if we are atomic. Since do_exit() needs to call into
5363 * schedule() atomically, we ignore that path for now.
5364 * Otherwise, whine if we are scheduling when we should not be.
5366 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5367 __schedule_bug(prev);
5369 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5371 schedstat_inc(this_rq(), sched_count);
5372 #ifdef CONFIG_SCHEDSTATS
5373 if (unlikely(prev->lock_depth >= 0)) {
5374 schedstat_inc(this_rq(), bkl_count);
5375 schedstat_inc(prev, sched_info.bkl_count);
5380 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5382 if (prev->state == TASK_RUNNING) {
5383 u64 runtime = prev->se.sum_exec_runtime;
5385 runtime -= prev->se.prev_sum_exec_runtime;
5386 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5389 * In order to avoid avg_overlap growing stale when we are
5390 * indeed overlapping and hence not getting put to sleep, grow
5391 * the avg_overlap on preemption.
5393 * We use the average preemption runtime because that
5394 * correlates to the amount of cache footprint a task can
5397 update_avg(&prev->se.avg_overlap, runtime);
5399 prev->sched_class->put_prev_task(rq, prev);
5403 * Pick up the highest-prio task:
5405 static inline struct task_struct *
5406 pick_next_task(struct rq *rq)
5408 const struct sched_class *class;
5409 struct task_struct *p;
5412 * Optimization: we know that if all tasks are in
5413 * the fair class we can call that function directly:
5415 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5416 p = fair_sched_class.pick_next_task(rq);
5421 class = sched_class_highest;
5423 p = class->pick_next_task(rq);
5427 * Will never be NULL as the idle class always
5428 * returns a non-NULL p:
5430 class = class->next;
5435 * schedule() is the main scheduler function.
5437 asmlinkage void __sched schedule(void)
5439 struct task_struct *prev, *next;
5440 unsigned long *switch_count;
5446 cpu = smp_processor_id();
5450 switch_count = &prev->nivcsw;
5452 release_kernel_lock(prev);
5453 need_resched_nonpreemptible:
5455 schedule_debug(prev);
5457 if (sched_feat(HRTICK))
5460 spin_lock_irq(&rq->lock);
5461 update_rq_clock(rq);
5462 clear_tsk_need_resched(prev);
5464 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5465 if (unlikely(signal_pending_state(prev->state, prev)))
5466 prev->state = TASK_RUNNING;
5468 deactivate_task(rq, prev, 1);
5469 switch_count = &prev->nvcsw;
5472 pre_schedule(rq, prev);
5474 if (unlikely(!rq->nr_running))
5475 idle_balance(cpu, rq);
5477 put_prev_task(rq, prev);
5478 next = pick_next_task(rq);
5480 if (likely(prev != next)) {
5481 sched_info_switch(prev, next);
5482 perf_event_task_sched_out(prev, next, cpu);
5488 context_switch(rq, prev, next); /* unlocks the rq */
5490 * the context switch might have flipped the stack from under
5491 * us, hence refresh the local variables.
5493 cpu = smp_processor_id();
5496 spin_unlock_irq(&rq->lock);
5500 if (unlikely(reacquire_kernel_lock(current) < 0))
5501 goto need_resched_nonpreemptible;
5503 preempt_enable_no_resched();
5507 EXPORT_SYMBOL(schedule);
5509 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5511 * Look out! "owner" is an entirely speculative pointer
5512 * access and not reliable.
5514 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5519 if (!sched_feat(OWNER_SPIN))
5522 #ifdef CONFIG_DEBUG_PAGEALLOC
5524 * Need to access the cpu field knowing that
5525 * DEBUG_PAGEALLOC could have unmapped it if
5526 * the mutex owner just released it and exited.
5528 if (probe_kernel_address(&owner->cpu, cpu))
5535 * Even if the access succeeded (likely case),
5536 * the cpu field may no longer be valid.
5538 if (cpu >= nr_cpumask_bits)
5542 * We need to validate that we can do a
5543 * get_cpu() and that we have the percpu area.
5545 if (!cpu_online(cpu))
5552 * Owner changed, break to re-assess state.
5554 if (lock->owner != owner)
5558 * Is that owner really running on that cpu?
5560 if (task_thread_info(rq->curr) != owner || need_resched())
5570 #ifdef CONFIG_PREEMPT
5572 * this is the entry point to schedule() from in-kernel preemption
5573 * off of preempt_enable. Kernel preemptions off return from interrupt
5574 * occur there and call schedule directly.
5576 asmlinkage void __sched preempt_schedule(void)
5578 struct thread_info *ti = current_thread_info();
5581 * If there is a non-zero preempt_count or interrupts are disabled,
5582 * we do not want to preempt the current task. Just return..
5584 if (likely(ti->preempt_count || irqs_disabled()))
5588 add_preempt_count(PREEMPT_ACTIVE);
5590 sub_preempt_count(PREEMPT_ACTIVE);
5593 * Check again in case we missed a preemption opportunity
5594 * between schedule and now.
5597 } while (need_resched());
5599 EXPORT_SYMBOL(preempt_schedule);
5602 * this is the entry point to schedule() from kernel preemption
5603 * off of irq context.
5604 * Note, that this is called and return with irqs disabled. This will
5605 * protect us against recursive calling from irq.
5607 asmlinkage void __sched preempt_schedule_irq(void)
5609 struct thread_info *ti = current_thread_info();
5611 /* Catch callers which need to be fixed */
5612 BUG_ON(ti->preempt_count || !irqs_disabled());
5615 add_preempt_count(PREEMPT_ACTIVE);
5618 local_irq_disable();
5619 sub_preempt_count(PREEMPT_ACTIVE);
5622 * Check again in case we missed a preemption opportunity
5623 * between schedule and now.
5626 } while (need_resched());
5629 #endif /* CONFIG_PREEMPT */
5631 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5634 return try_to_wake_up(curr->private, mode, wake_flags);
5636 EXPORT_SYMBOL(default_wake_function);
5639 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5640 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5641 * number) then we wake all the non-exclusive tasks and one exclusive task.
5643 * There are circumstances in which we can try to wake a task which has already
5644 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5645 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5647 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5648 int nr_exclusive, int wake_flags, void *key)
5650 wait_queue_t *curr, *next;
5652 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5653 unsigned flags = curr->flags;
5655 if (curr->func(curr, mode, wake_flags, key) &&
5656 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5662 * __wake_up - wake up threads blocked on a waitqueue.
5664 * @mode: which threads
5665 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5666 * @key: is directly passed to the wakeup function
5668 * It may be assumed that this function implies a write memory barrier before
5669 * changing the task state if and only if any tasks are woken up.
5671 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5672 int nr_exclusive, void *key)
5674 unsigned long flags;
5676 spin_lock_irqsave(&q->lock, flags);
5677 __wake_up_common(q, mode, nr_exclusive, 0, key);
5678 spin_unlock_irqrestore(&q->lock, flags);
5680 EXPORT_SYMBOL(__wake_up);
5683 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5685 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5687 __wake_up_common(q, mode, 1, 0, NULL);
5690 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5692 __wake_up_common(q, mode, 1, 0, key);
5696 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5698 * @mode: which threads
5699 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5700 * @key: opaque value to be passed to wakeup targets
5702 * The sync wakeup differs that the waker knows that it will schedule
5703 * away soon, so while the target thread will be woken up, it will not
5704 * be migrated to another CPU - ie. the two threads are 'synchronized'
5705 * with each other. This can prevent needless bouncing between CPUs.
5707 * On UP it can prevent extra preemption.
5709 * It may be assumed that this function implies a write memory barrier before
5710 * changing the task state if and only if any tasks are woken up.
5712 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5713 int nr_exclusive, void *key)
5715 unsigned long flags;
5716 int wake_flags = WF_SYNC;
5721 if (unlikely(!nr_exclusive))
5724 spin_lock_irqsave(&q->lock, flags);
5725 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5726 spin_unlock_irqrestore(&q->lock, flags);
5728 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5731 * __wake_up_sync - see __wake_up_sync_key()
5733 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5735 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5737 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5740 * complete: - signals a single thread waiting on this completion
5741 * @x: holds the state of this particular completion
5743 * This will wake up a single thread waiting on this completion. Threads will be
5744 * awakened in the same order in which they were queued.
5746 * See also complete_all(), wait_for_completion() and related routines.
5748 * It may be assumed that this function implies a write memory barrier before
5749 * changing the task state if and only if any tasks are woken up.
5751 void complete(struct completion *x)
5753 unsigned long flags;
5755 spin_lock_irqsave(&x->wait.lock, flags);
5757 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5758 spin_unlock_irqrestore(&x->wait.lock, flags);
5760 EXPORT_SYMBOL(complete);
5763 * complete_all: - signals all threads waiting on this completion
5764 * @x: holds the state of this particular completion
5766 * This will wake up all threads waiting on this particular completion event.
5768 * It may be assumed that this function implies a write memory barrier before
5769 * changing the task state if and only if any tasks are woken up.
5771 void complete_all(struct completion *x)
5773 unsigned long flags;
5775 spin_lock_irqsave(&x->wait.lock, flags);
5776 x->done += UINT_MAX/2;
5777 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5778 spin_unlock_irqrestore(&x->wait.lock, flags);
5780 EXPORT_SYMBOL(complete_all);
5782 static inline long __sched
5783 do_wait_for_common(struct completion *x, long timeout, int state)
5786 DECLARE_WAITQUEUE(wait, current);
5788 wait.flags |= WQ_FLAG_EXCLUSIVE;
5789 __add_wait_queue_tail(&x->wait, &wait);
5791 if (signal_pending_state(state, current)) {
5792 timeout = -ERESTARTSYS;
5795 __set_current_state(state);
5796 spin_unlock_irq(&x->wait.lock);
5797 timeout = schedule_timeout(timeout);
5798 spin_lock_irq(&x->wait.lock);
5799 } while (!x->done && timeout);
5800 __remove_wait_queue(&x->wait, &wait);
5805 return timeout ?: 1;
5809 wait_for_common(struct completion *x, long timeout, int state)
5813 spin_lock_irq(&x->wait.lock);
5814 timeout = do_wait_for_common(x, timeout, state);
5815 spin_unlock_irq(&x->wait.lock);
5820 * wait_for_completion: - waits for completion of a task
5821 * @x: holds the state of this particular completion
5823 * This waits to be signaled for completion of a specific task. It is NOT
5824 * interruptible and there is no timeout.
5826 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5827 * and interrupt capability. Also see complete().
5829 void __sched wait_for_completion(struct completion *x)
5831 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5833 EXPORT_SYMBOL(wait_for_completion);
5836 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5837 * @x: holds the state of this particular completion
5838 * @timeout: timeout value in jiffies
5840 * This waits for either a completion of a specific task to be signaled or for a
5841 * specified timeout to expire. The timeout is in jiffies. It is not
5844 unsigned long __sched
5845 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5847 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5849 EXPORT_SYMBOL(wait_for_completion_timeout);
5852 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5853 * @x: holds the state of this particular completion
5855 * This waits for completion of a specific task to be signaled. It is
5858 int __sched wait_for_completion_interruptible(struct completion *x)
5860 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5861 if (t == -ERESTARTSYS)
5865 EXPORT_SYMBOL(wait_for_completion_interruptible);
5868 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5869 * @x: holds the state of this particular completion
5870 * @timeout: timeout value in jiffies
5872 * This waits for either a completion of a specific task to be signaled or for a
5873 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5875 unsigned long __sched
5876 wait_for_completion_interruptible_timeout(struct completion *x,
5877 unsigned long timeout)
5879 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5881 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5884 * wait_for_completion_killable: - waits for completion of a task (killable)
5885 * @x: holds the state of this particular completion
5887 * This waits to be signaled for completion of a specific task. It can be
5888 * interrupted by a kill signal.
5890 int __sched wait_for_completion_killable(struct completion *x)
5892 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5893 if (t == -ERESTARTSYS)
5897 EXPORT_SYMBOL(wait_for_completion_killable);
5900 * try_wait_for_completion - try to decrement a completion without blocking
5901 * @x: completion structure
5903 * Returns: 0 if a decrement cannot be done without blocking
5904 * 1 if a decrement succeeded.
5906 * If a completion is being used as a counting completion,
5907 * attempt to decrement the counter without blocking. This
5908 * enables us to avoid waiting if the resource the completion
5909 * is protecting is not available.
5911 bool try_wait_for_completion(struct completion *x)
5913 unsigned long flags;
5916 spin_lock_irqsave(&x->wait.lock, flags);
5921 spin_unlock_irqrestore(&x->wait.lock, flags);
5924 EXPORT_SYMBOL(try_wait_for_completion);
5927 * completion_done - Test to see if a completion has any waiters
5928 * @x: completion structure
5930 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5931 * 1 if there are no waiters.
5934 bool completion_done(struct completion *x)
5936 unsigned long flags;
5939 spin_lock_irqsave(&x->wait.lock, flags);
5942 spin_unlock_irqrestore(&x->wait.lock, flags);
5945 EXPORT_SYMBOL(completion_done);
5948 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5950 unsigned long flags;
5953 init_waitqueue_entry(&wait, current);
5955 __set_current_state(state);
5957 spin_lock_irqsave(&q->lock, flags);
5958 __add_wait_queue(q, &wait);
5959 spin_unlock(&q->lock);
5960 timeout = schedule_timeout(timeout);
5961 spin_lock_irq(&q->lock);
5962 __remove_wait_queue(q, &wait);
5963 spin_unlock_irqrestore(&q->lock, flags);
5968 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5970 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5972 EXPORT_SYMBOL(interruptible_sleep_on);
5975 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5977 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5979 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5981 void __sched sleep_on(wait_queue_head_t *q)
5983 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5985 EXPORT_SYMBOL(sleep_on);
5987 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5989 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5991 EXPORT_SYMBOL(sleep_on_timeout);
5993 #ifdef CONFIG_RT_MUTEXES
5996 * rt_mutex_setprio - set the current priority of a task
5998 * @prio: prio value (kernel-internal form)
6000 * This function changes the 'effective' priority of a task. It does
6001 * not touch ->normal_prio like __setscheduler().
6003 * Used by the rt_mutex code to implement priority inheritance logic.
6005 void rt_mutex_setprio(struct task_struct *p, int prio)
6007 unsigned long flags;
6008 int oldprio, on_rq, running;
6010 const struct sched_class *prev_class = p->sched_class;
6012 BUG_ON(prio < 0 || prio > MAX_PRIO);
6014 rq = task_rq_lock(p, &flags);
6015 update_rq_clock(rq);
6018 on_rq = p->se.on_rq;
6019 running = task_current(rq, p);
6021 dequeue_task(rq, p, 0);
6023 p->sched_class->put_prev_task(rq, p);
6026 p->sched_class = &rt_sched_class;
6028 p->sched_class = &fair_sched_class;
6033 p->sched_class->set_curr_task(rq);
6035 enqueue_task(rq, p, 0);
6037 check_class_changed(rq, p, prev_class, oldprio, running);
6039 task_rq_unlock(rq, &flags);
6044 void set_user_nice(struct task_struct *p, long nice)
6046 int old_prio, delta, on_rq;
6047 unsigned long flags;
6050 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6053 * We have to be careful, if called from sys_setpriority(),
6054 * the task might be in the middle of scheduling on another CPU.
6056 rq = task_rq_lock(p, &flags);
6057 update_rq_clock(rq);
6059 * The RT priorities are set via sched_setscheduler(), but we still
6060 * allow the 'normal' nice value to be set - but as expected
6061 * it wont have any effect on scheduling until the task is
6062 * SCHED_FIFO/SCHED_RR:
6064 if (task_has_rt_policy(p)) {
6065 p->static_prio = NICE_TO_PRIO(nice);
6068 on_rq = p->se.on_rq;
6070 dequeue_task(rq, p, 0);
6072 p->static_prio = NICE_TO_PRIO(nice);
6075 p->prio = effective_prio(p);
6076 delta = p->prio - old_prio;
6079 enqueue_task(rq, p, 0);
6081 * If the task increased its priority or is running and
6082 * lowered its priority, then reschedule its CPU:
6084 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6085 resched_task(rq->curr);
6088 task_rq_unlock(rq, &flags);
6090 EXPORT_SYMBOL(set_user_nice);
6093 * can_nice - check if a task can reduce its nice value
6097 int can_nice(const struct task_struct *p, const int nice)
6099 /* convert nice value [19,-20] to rlimit style value [1,40] */
6100 int nice_rlim = 20 - nice;
6102 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6103 capable(CAP_SYS_NICE));
6106 #ifdef __ARCH_WANT_SYS_NICE
6109 * sys_nice - change the priority of the current process.
6110 * @increment: priority increment
6112 * sys_setpriority is a more generic, but much slower function that
6113 * does similar things.
6115 SYSCALL_DEFINE1(nice, int, increment)
6120 * Setpriority might change our priority at the same moment.
6121 * We don't have to worry. Conceptually one call occurs first
6122 * and we have a single winner.
6124 if (increment < -40)
6129 nice = TASK_NICE(current) + increment;
6135 if (increment < 0 && !can_nice(current, nice))
6138 retval = security_task_setnice(current, nice);
6142 set_user_nice(current, nice);
6149 * task_prio - return the priority value of a given task.
6150 * @p: the task in question.
6152 * This is the priority value as seen by users in /proc.
6153 * RT tasks are offset by -200. Normal tasks are centered
6154 * around 0, value goes from -16 to +15.
6156 int task_prio(const struct task_struct *p)
6158 return p->prio - MAX_RT_PRIO;
6162 * task_nice - return the nice value of a given task.
6163 * @p: the task in question.
6165 int task_nice(const struct task_struct *p)
6167 return TASK_NICE(p);
6169 EXPORT_SYMBOL(task_nice);
6172 * idle_cpu - is a given cpu idle currently?
6173 * @cpu: the processor in question.
6175 int idle_cpu(int cpu)
6177 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6181 * idle_task - return the idle task for a given cpu.
6182 * @cpu: the processor in question.
6184 struct task_struct *idle_task(int cpu)
6186 return cpu_rq(cpu)->idle;
6190 * find_process_by_pid - find a process with a matching PID value.
6191 * @pid: the pid in question.
6193 static struct task_struct *find_process_by_pid(pid_t pid)
6195 return pid ? find_task_by_vpid(pid) : current;
6198 /* Actually do priority change: must hold rq lock. */
6200 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6202 BUG_ON(p->se.on_rq);
6205 p->rt_priority = prio;
6206 p->normal_prio = normal_prio(p);
6207 /* we are holding p->pi_lock already */
6208 p->prio = rt_mutex_getprio(p);
6209 if (rt_prio(p->prio))
6210 p->sched_class = &rt_sched_class;
6212 p->sched_class = &fair_sched_class;
6217 * check the target process has a UID that matches the current process's
6219 static bool check_same_owner(struct task_struct *p)
6221 const struct cred *cred = current_cred(), *pcred;
6225 pcred = __task_cred(p);
6226 match = (cred->euid == pcred->euid ||
6227 cred->euid == pcred->uid);
6232 static int __sched_setscheduler(struct task_struct *p, int policy,
6233 struct sched_param *param, bool user)
6235 int retval, oldprio, oldpolicy = -1, on_rq, running;
6236 unsigned long flags;
6237 const struct sched_class *prev_class = p->sched_class;
6241 /* may grab non-irq protected spin_locks */
6242 BUG_ON(in_interrupt());
6244 /* double check policy once rq lock held */
6246 reset_on_fork = p->sched_reset_on_fork;
6247 policy = oldpolicy = p->policy;
6249 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6250 policy &= ~SCHED_RESET_ON_FORK;
6252 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6253 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6254 policy != SCHED_IDLE)
6259 * Valid priorities for SCHED_FIFO and SCHED_RR are
6260 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6261 * SCHED_BATCH and SCHED_IDLE is 0.
6263 if (param->sched_priority < 0 ||
6264 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6265 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6267 if (rt_policy(policy) != (param->sched_priority != 0))
6271 * Allow unprivileged RT tasks to decrease priority:
6273 if (user && !capable(CAP_SYS_NICE)) {
6274 if (rt_policy(policy)) {
6275 unsigned long rlim_rtprio;
6277 if (!lock_task_sighand(p, &flags))
6279 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6280 unlock_task_sighand(p, &flags);
6282 /* can't set/change the rt policy */
6283 if (policy != p->policy && !rlim_rtprio)
6286 /* can't increase priority */
6287 if (param->sched_priority > p->rt_priority &&
6288 param->sched_priority > rlim_rtprio)
6292 * Like positive nice levels, dont allow tasks to
6293 * move out of SCHED_IDLE either:
6295 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6298 /* can't change other user's priorities */
6299 if (!check_same_owner(p))
6302 /* Normal users shall not reset the sched_reset_on_fork flag */
6303 if (p->sched_reset_on_fork && !reset_on_fork)
6308 #ifdef CONFIG_RT_GROUP_SCHED
6310 * Do not allow realtime tasks into groups that have no runtime
6313 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6314 task_group(p)->rt_bandwidth.rt_runtime == 0)
6318 retval = security_task_setscheduler(p, policy, param);
6324 * make sure no PI-waiters arrive (or leave) while we are
6325 * changing the priority of the task:
6327 spin_lock_irqsave(&p->pi_lock, flags);
6329 * To be able to change p->policy safely, the apropriate
6330 * runqueue lock must be held.
6332 rq = __task_rq_lock(p);
6333 /* recheck policy now with rq lock held */
6334 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6335 policy = oldpolicy = -1;
6336 __task_rq_unlock(rq);
6337 spin_unlock_irqrestore(&p->pi_lock, flags);
6340 update_rq_clock(rq);
6341 on_rq = p->se.on_rq;
6342 running = task_current(rq, p);
6344 deactivate_task(rq, p, 0);
6346 p->sched_class->put_prev_task(rq, p);
6348 p->sched_reset_on_fork = reset_on_fork;
6351 __setscheduler(rq, p, policy, param->sched_priority);
6354 p->sched_class->set_curr_task(rq);
6356 activate_task(rq, p, 0);
6358 check_class_changed(rq, p, prev_class, oldprio, running);
6360 __task_rq_unlock(rq);
6361 spin_unlock_irqrestore(&p->pi_lock, flags);
6363 rt_mutex_adjust_pi(p);
6369 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6370 * @p: the task in question.
6371 * @policy: new policy.
6372 * @param: structure containing the new RT priority.
6374 * NOTE that the task may be already dead.
6376 int sched_setscheduler(struct task_struct *p, int policy,
6377 struct sched_param *param)
6379 return __sched_setscheduler(p, policy, param, true);
6381 EXPORT_SYMBOL_GPL(sched_setscheduler);
6384 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6385 * @p: the task in question.
6386 * @policy: new policy.
6387 * @param: structure containing the new RT priority.
6389 * Just like sched_setscheduler, only don't bother checking if the
6390 * current context has permission. For example, this is needed in
6391 * stop_machine(): we create temporary high priority worker threads,
6392 * but our caller might not have that capability.
6394 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6395 struct sched_param *param)
6397 return __sched_setscheduler(p, policy, param, false);
6401 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6403 struct sched_param lparam;
6404 struct task_struct *p;
6407 if (!param || pid < 0)
6409 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6414 p = find_process_by_pid(pid);
6416 retval = sched_setscheduler(p, policy, &lparam);
6423 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6424 * @pid: the pid in question.
6425 * @policy: new policy.
6426 * @param: structure containing the new RT priority.
6428 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6429 struct sched_param __user *, param)
6431 /* negative values for policy are not valid */
6435 return do_sched_setscheduler(pid, policy, param);
6439 * sys_sched_setparam - set/change the RT priority of a thread
6440 * @pid: the pid in question.
6441 * @param: structure containing the new RT priority.
6443 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6445 return do_sched_setscheduler(pid, -1, param);
6449 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6450 * @pid: the pid in question.
6452 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6454 struct task_struct *p;
6462 p = find_process_by_pid(pid);
6464 retval = security_task_getscheduler(p);
6467 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6474 * sys_sched_getparam - get the RT priority of a thread
6475 * @pid: the pid in question.
6476 * @param: structure containing the RT priority.
6478 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6480 struct sched_param lp;
6481 struct task_struct *p;
6484 if (!param || pid < 0)
6488 p = find_process_by_pid(pid);
6493 retval = security_task_getscheduler(p);
6497 lp.sched_priority = p->rt_priority;
6501 * This one might sleep, we cannot do it with a spinlock held ...
6503 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6512 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6514 cpumask_var_t cpus_allowed, new_mask;
6515 struct task_struct *p;
6519 read_lock(&tasklist_lock);
6521 p = find_process_by_pid(pid);
6523 read_unlock(&tasklist_lock);
6529 * It is not safe to call set_cpus_allowed with the
6530 * tasklist_lock held. We will bump the task_struct's
6531 * usage count and then drop tasklist_lock.
6534 read_unlock(&tasklist_lock);
6536 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6540 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6542 goto out_free_cpus_allowed;
6545 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6548 retval = security_task_setscheduler(p, 0, NULL);
6552 cpuset_cpus_allowed(p, cpus_allowed);
6553 cpumask_and(new_mask, in_mask, cpus_allowed);
6555 retval = set_cpus_allowed_ptr(p, new_mask);
6558 cpuset_cpus_allowed(p, cpus_allowed);
6559 if (!cpumask_subset(new_mask, cpus_allowed)) {
6561 * We must have raced with a concurrent cpuset
6562 * update. Just reset the cpus_allowed to the
6563 * cpuset's cpus_allowed
6565 cpumask_copy(new_mask, cpus_allowed);
6570 free_cpumask_var(new_mask);
6571 out_free_cpus_allowed:
6572 free_cpumask_var(cpus_allowed);
6579 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6580 struct cpumask *new_mask)
6582 if (len < cpumask_size())
6583 cpumask_clear(new_mask);
6584 else if (len > cpumask_size())
6585 len = cpumask_size();
6587 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6591 * sys_sched_setaffinity - set the cpu affinity of a process
6592 * @pid: pid of the process
6593 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6594 * @user_mask_ptr: user-space pointer to the new cpu mask
6596 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6597 unsigned long __user *, user_mask_ptr)
6599 cpumask_var_t new_mask;
6602 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6605 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6607 retval = sched_setaffinity(pid, new_mask);
6608 free_cpumask_var(new_mask);
6612 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6614 struct task_struct *p;
6615 unsigned long flags;
6620 read_lock(&tasklist_lock);
6623 p = find_process_by_pid(pid);
6627 retval = security_task_getscheduler(p);
6631 rq = task_rq_lock(p, &flags);
6632 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6633 task_rq_unlock(rq, &flags);
6636 read_unlock(&tasklist_lock);
6643 * sys_sched_getaffinity - get the cpu affinity of a process
6644 * @pid: pid of the process
6645 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6646 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6648 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6649 unsigned long __user *, user_mask_ptr)
6654 if (len < cpumask_size())
6657 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6660 ret = sched_getaffinity(pid, mask);
6662 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6665 ret = cpumask_size();
6667 free_cpumask_var(mask);
6673 * sys_sched_yield - yield the current processor to other threads.
6675 * This function yields the current CPU to other tasks. If there are no
6676 * other threads running on this CPU then this function will return.
6678 SYSCALL_DEFINE0(sched_yield)
6680 struct rq *rq = this_rq_lock();
6682 schedstat_inc(rq, yld_count);
6683 current->sched_class->yield_task(rq);
6686 * Since we are going to call schedule() anyway, there's
6687 * no need to preempt or enable interrupts:
6689 __release(rq->lock);
6690 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6691 _raw_spin_unlock(&rq->lock);
6692 preempt_enable_no_resched();
6699 static inline int should_resched(void)
6701 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6704 static void __cond_resched(void)
6706 add_preempt_count(PREEMPT_ACTIVE);
6708 sub_preempt_count(PREEMPT_ACTIVE);
6711 int __sched _cond_resched(void)
6713 if (should_resched()) {
6719 EXPORT_SYMBOL(_cond_resched);
6722 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6723 * call schedule, and on return reacquire the lock.
6725 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6726 * operations here to prevent schedule() from being called twice (once via
6727 * spin_unlock(), once by hand).
6729 int __cond_resched_lock(spinlock_t *lock)
6731 int resched = should_resched();
6734 lockdep_assert_held(lock);
6736 if (spin_needbreak(lock) || resched) {
6747 EXPORT_SYMBOL(__cond_resched_lock);
6749 int __sched __cond_resched_softirq(void)
6751 BUG_ON(!in_softirq());
6753 if (should_resched()) {
6761 EXPORT_SYMBOL(__cond_resched_softirq);
6764 * yield - yield the current processor to other threads.
6766 * This is a shortcut for kernel-space yielding - it marks the
6767 * thread runnable and calls sys_sched_yield().
6769 void __sched yield(void)
6771 set_current_state(TASK_RUNNING);
6774 EXPORT_SYMBOL(yield);
6777 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6778 * that process accounting knows that this is a task in IO wait state.
6780 void __sched io_schedule(void)
6782 struct rq *rq = raw_rq();
6784 delayacct_blkio_start();
6785 atomic_inc(&rq->nr_iowait);
6786 current->in_iowait = 1;
6788 current->in_iowait = 0;
6789 atomic_dec(&rq->nr_iowait);
6790 delayacct_blkio_end();
6792 EXPORT_SYMBOL(io_schedule);
6794 long __sched io_schedule_timeout(long timeout)
6796 struct rq *rq = raw_rq();
6799 delayacct_blkio_start();
6800 atomic_inc(&rq->nr_iowait);
6801 current->in_iowait = 1;
6802 ret = schedule_timeout(timeout);
6803 current->in_iowait = 0;
6804 atomic_dec(&rq->nr_iowait);
6805 delayacct_blkio_end();
6810 * sys_sched_get_priority_max - return maximum RT priority.
6811 * @policy: scheduling class.
6813 * this syscall returns the maximum rt_priority that can be used
6814 * by a given scheduling class.
6816 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6823 ret = MAX_USER_RT_PRIO-1;
6835 * sys_sched_get_priority_min - return minimum RT priority.
6836 * @policy: scheduling class.
6838 * this syscall returns the minimum rt_priority that can be used
6839 * by a given scheduling class.
6841 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6859 * sys_sched_rr_get_interval - return the default timeslice of a process.
6860 * @pid: pid of the process.
6861 * @interval: userspace pointer to the timeslice value.
6863 * this syscall writes the default timeslice value of a given process
6864 * into the user-space timespec buffer. A value of '0' means infinity.
6866 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6867 struct timespec __user *, interval)
6869 struct task_struct *p;
6870 unsigned int time_slice;
6871 unsigned long flags;
6880 read_lock(&tasklist_lock);
6881 p = find_process_by_pid(pid);
6885 retval = security_task_getscheduler(p);
6889 rq = task_rq_lock(p, &flags);
6890 time_slice = p->sched_class->get_rr_interval(rq, p);
6891 task_rq_unlock(rq, &flags);
6893 read_unlock(&tasklist_lock);
6894 jiffies_to_timespec(time_slice, &t);
6895 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6899 read_unlock(&tasklist_lock);
6903 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6905 void sched_show_task(struct task_struct *p)
6907 unsigned long free = 0;
6910 state = p->state ? __ffs(p->state) + 1 : 0;
6911 pr_info("%-13.13s %c", p->comm,
6912 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6913 #if BITS_PER_LONG == 32
6914 if (state == TASK_RUNNING)
6915 pr_cont(" running ");
6917 pr_cont(" %08lx ", thread_saved_pc(p));
6919 if (state == TASK_RUNNING)
6920 pr_cont(" running task ");
6922 pr_cont(" %016lx ", thread_saved_pc(p));
6924 #ifdef CONFIG_DEBUG_STACK_USAGE
6925 free = stack_not_used(p);
6927 pr_cont("%5lu %5d %6d 0x%08lx\n", free,
6928 task_pid_nr(p), task_pid_nr(p->real_parent),
6929 (unsigned long)task_thread_info(p)->flags);
6931 show_stack(p, NULL);
6934 void show_state_filter(unsigned long state_filter)
6936 struct task_struct *g, *p;
6938 #if BITS_PER_LONG == 32
6939 pr_info(" task PC stack pid father\n");
6941 pr_info(" task PC stack pid father\n");
6943 read_lock(&tasklist_lock);
6944 do_each_thread(g, p) {
6946 * reset the NMI-timeout, listing all files on a slow
6947 * console might take alot of time:
6949 touch_nmi_watchdog();
6950 if (!state_filter || (p->state & state_filter))
6952 } while_each_thread(g, p);
6954 touch_all_softlockup_watchdogs();
6956 #ifdef CONFIG_SCHED_DEBUG
6957 sysrq_sched_debug_show();
6959 read_unlock(&tasklist_lock);
6961 * Only show locks if all tasks are dumped:
6964 debug_show_all_locks();
6967 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6969 idle->sched_class = &idle_sched_class;
6973 * init_idle - set up an idle thread for a given CPU
6974 * @idle: task in question
6975 * @cpu: cpu the idle task belongs to
6977 * NOTE: this function does not set the idle thread's NEED_RESCHED
6978 * flag, to make booting more robust.
6980 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6982 struct rq *rq = cpu_rq(cpu);
6983 unsigned long flags;
6985 spin_lock_irqsave(&rq->lock, flags);
6988 idle->se.exec_start = sched_clock();
6990 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6991 __set_task_cpu(idle, cpu);
6993 rq->curr = rq->idle = idle;
6994 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6997 spin_unlock_irqrestore(&rq->lock, flags);
6999 /* Set the preempt count _outside_ the spinlocks! */
7000 #if defined(CONFIG_PREEMPT)
7001 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7003 task_thread_info(idle)->preempt_count = 0;
7006 * The idle tasks have their own, simple scheduling class:
7008 idle->sched_class = &idle_sched_class;
7009 ftrace_graph_init_task(idle);
7013 * In a system that switches off the HZ timer nohz_cpu_mask
7014 * indicates which cpus entered this state. This is used
7015 * in the rcu update to wait only for active cpus. For system
7016 * which do not switch off the HZ timer nohz_cpu_mask should
7017 * always be CPU_BITS_NONE.
7019 cpumask_var_t nohz_cpu_mask;
7022 * Increase the granularity value when there are more CPUs,
7023 * because with more CPUs the 'effective latency' as visible
7024 * to users decreases. But the relationship is not linear,
7025 * so pick a second-best guess by going with the log2 of the
7028 * This idea comes from the SD scheduler of Con Kolivas:
7030 static int get_update_sysctl_factor(void)
7032 unsigned int cpus = min_t(int, num_online_cpus(), 8);
7033 unsigned int factor;
7035 switch (sysctl_sched_tunable_scaling) {
7036 case SCHED_TUNABLESCALING_NONE:
7039 case SCHED_TUNABLESCALING_LINEAR:
7042 case SCHED_TUNABLESCALING_LOG:
7044 factor = 1 + ilog2(cpus);
7051 static void update_sysctl(void)
7053 unsigned int factor = get_update_sysctl_factor();
7055 #define SET_SYSCTL(name) \
7056 (sysctl_##name = (factor) * normalized_sysctl_##name)
7057 SET_SYSCTL(sched_min_granularity);
7058 SET_SYSCTL(sched_latency);
7059 SET_SYSCTL(sched_wakeup_granularity);
7060 SET_SYSCTL(sched_shares_ratelimit);
7064 static inline void sched_init_granularity(void)
7071 * This is how migration works:
7073 * 1) we queue a struct migration_req structure in the source CPU's
7074 * runqueue and wake up that CPU's migration thread.
7075 * 2) we down() the locked semaphore => thread blocks.
7076 * 3) migration thread wakes up (implicitly it forces the migrated
7077 * thread off the CPU)
7078 * 4) it gets the migration request and checks whether the migrated
7079 * task is still in the wrong runqueue.
7080 * 5) if it's in the wrong runqueue then the migration thread removes
7081 * it and puts it into the right queue.
7082 * 6) migration thread up()s the semaphore.
7083 * 7) we wake up and the migration is done.
7087 * Change a given task's CPU affinity. Migrate the thread to a
7088 * proper CPU and schedule it away if the CPU it's executing on
7089 * is removed from the allowed bitmask.
7091 * NOTE: the caller must have a valid reference to the task, the
7092 * task must not exit() & deallocate itself prematurely. The
7093 * call is not atomic; no spinlocks may be held.
7095 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7097 struct migration_req req;
7098 unsigned long flags;
7102 rq = task_rq_lock(p, &flags);
7103 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7108 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7109 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7114 if (p->sched_class->set_cpus_allowed)
7115 p->sched_class->set_cpus_allowed(p, new_mask);
7117 cpumask_copy(&p->cpus_allowed, new_mask);
7118 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7121 /* Can the task run on the task's current CPU? If so, we're done */
7122 if (cpumask_test_cpu(task_cpu(p), new_mask))
7125 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7126 /* Need help from migration thread: drop lock and wait. */
7127 struct task_struct *mt = rq->migration_thread;
7129 get_task_struct(mt);
7130 task_rq_unlock(rq, &flags);
7131 wake_up_process(rq->migration_thread);
7132 put_task_struct(mt);
7133 wait_for_completion(&req.done);
7134 tlb_migrate_finish(p->mm);
7138 task_rq_unlock(rq, &flags);
7142 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7145 * Move (not current) task off this cpu, onto dest cpu. We're doing
7146 * this because either it can't run here any more (set_cpus_allowed()
7147 * away from this CPU, or CPU going down), or because we're
7148 * attempting to rebalance this task on exec (sched_exec).
7150 * So we race with normal scheduler movements, but that's OK, as long
7151 * as the task is no longer on this CPU.
7153 * Returns non-zero if task was successfully migrated.
7155 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7157 struct rq *rq_dest, *rq_src;
7160 if (unlikely(!cpu_active(dest_cpu)))
7163 rq_src = cpu_rq(src_cpu);
7164 rq_dest = cpu_rq(dest_cpu);
7166 double_rq_lock(rq_src, rq_dest);
7167 /* Already moved. */
7168 if (task_cpu(p) != src_cpu)
7170 /* Affinity changed (again). */
7171 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7174 on_rq = p->se.on_rq;
7176 deactivate_task(rq_src, p, 0);
7178 set_task_cpu(p, dest_cpu);
7180 activate_task(rq_dest, p, 0);
7181 check_preempt_curr(rq_dest, p, 0);
7186 double_rq_unlock(rq_src, rq_dest);
7190 #define RCU_MIGRATION_IDLE 0
7191 #define RCU_MIGRATION_NEED_QS 1
7192 #define RCU_MIGRATION_GOT_QS 2
7193 #define RCU_MIGRATION_MUST_SYNC 3
7196 * migration_thread - this is a highprio system thread that performs
7197 * thread migration by bumping thread off CPU then 'pushing' onto
7200 static int migration_thread(void *data)
7203 int cpu = (long)data;
7207 BUG_ON(rq->migration_thread != current);
7209 set_current_state(TASK_INTERRUPTIBLE);
7210 while (!kthread_should_stop()) {
7211 struct migration_req *req;
7212 struct list_head *head;
7214 spin_lock_irq(&rq->lock);
7216 if (cpu_is_offline(cpu)) {
7217 spin_unlock_irq(&rq->lock);
7221 if (rq->active_balance) {
7222 active_load_balance(rq, cpu);
7223 rq->active_balance = 0;
7226 head = &rq->migration_queue;
7228 if (list_empty(head)) {
7229 spin_unlock_irq(&rq->lock);
7231 set_current_state(TASK_INTERRUPTIBLE);
7234 req = list_entry(head->next, struct migration_req, list);
7235 list_del_init(head->next);
7237 if (req->task != NULL) {
7238 spin_unlock(&rq->lock);
7239 __migrate_task(req->task, cpu, req->dest_cpu);
7240 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7241 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7242 spin_unlock(&rq->lock);
7244 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7245 spin_unlock(&rq->lock);
7246 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7250 complete(&req->done);
7252 __set_current_state(TASK_RUNNING);
7257 #ifdef CONFIG_HOTPLUG_CPU
7259 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7263 local_irq_disable();
7264 ret = __migrate_task(p, src_cpu, dest_cpu);
7270 * Figure out where task on dead CPU should go, use force if necessary.
7272 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7275 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7278 /* Look for allowed, online CPU in same node. */
7279 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
7280 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7283 /* Any allowed, online CPU? */
7284 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
7285 if (dest_cpu < nr_cpu_ids)
7288 /* No more Mr. Nice Guy. */
7289 if (dest_cpu >= nr_cpu_ids) {
7290 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7291 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
7294 * Don't tell them about moving exiting tasks or
7295 * kernel threads (both mm NULL), since they never
7298 if (p->mm && printk_ratelimit()) {
7299 pr_info("process %d (%s) no longer affine to cpu%d\n",
7300 task_pid_nr(p), p->comm, dead_cpu);
7305 /* It can have affinity changed while we were choosing. */
7306 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7311 * While a dead CPU has no uninterruptible tasks queued at this point,
7312 * it might still have a nonzero ->nr_uninterruptible counter, because
7313 * for performance reasons the counter is not stricly tracking tasks to
7314 * their home CPUs. So we just add the counter to another CPU's counter,
7315 * to keep the global sum constant after CPU-down:
7317 static void migrate_nr_uninterruptible(struct rq *rq_src)
7319 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7320 unsigned long flags;
7322 local_irq_save(flags);
7323 double_rq_lock(rq_src, rq_dest);
7324 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7325 rq_src->nr_uninterruptible = 0;
7326 double_rq_unlock(rq_src, rq_dest);
7327 local_irq_restore(flags);
7330 /* Run through task list and migrate tasks from the dead cpu. */
7331 static void migrate_live_tasks(int src_cpu)
7333 struct task_struct *p, *t;
7335 read_lock(&tasklist_lock);
7337 do_each_thread(t, p) {
7341 if (task_cpu(p) == src_cpu)
7342 move_task_off_dead_cpu(src_cpu, p);
7343 } while_each_thread(t, p);
7345 read_unlock(&tasklist_lock);
7349 * Schedules idle task to be the next runnable task on current CPU.
7350 * It does so by boosting its priority to highest possible.
7351 * Used by CPU offline code.
7353 void sched_idle_next(void)
7355 int this_cpu = smp_processor_id();
7356 struct rq *rq = cpu_rq(this_cpu);
7357 struct task_struct *p = rq->idle;
7358 unsigned long flags;
7360 /* cpu has to be offline */
7361 BUG_ON(cpu_online(this_cpu));
7364 * Strictly not necessary since rest of the CPUs are stopped by now
7365 * and interrupts disabled on the current cpu.
7367 spin_lock_irqsave(&rq->lock, flags);
7369 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7371 update_rq_clock(rq);
7372 activate_task(rq, p, 0);
7374 spin_unlock_irqrestore(&rq->lock, flags);
7378 * Ensures that the idle task is using init_mm right before its cpu goes
7381 void idle_task_exit(void)
7383 struct mm_struct *mm = current->active_mm;
7385 BUG_ON(cpu_online(smp_processor_id()));
7388 switch_mm(mm, &init_mm, current);
7392 /* called under rq->lock with disabled interrupts */
7393 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7395 struct rq *rq = cpu_rq(dead_cpu);
7397 /* Must be exiting, otherwise would be on tasklist. */
7398 BUG_ON(!p->exit_state);
7400 /* Cannot have done final schedule yet: would have vanished. */
7401 BUG_ON(p->state == TASK_DEAD);
7406 * Drop lock around migration; if someone else moves it,
7407 * that's OK. No task can be added to this CPU, so iteration is
7410 spin_unlock_irq(&rq->lock);
7411 move_task_off_dead_cpu(dead_cpu, p);
7412 spin_lock_irq(&rq->lock);
7417 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7418 static void migrate_dead_tasks(unsigned int dead_cpu)
7420 struct rq *rq = cpu_rq(dead_cpu);
7421 struct task_struct *next;
7424 if (!rq->nr_running)
7426 update_rq_clock(rq);
7427 next = pick_next_task(rq);
7430 next->sched_class->put_prev_task(rq, next);
7431 migrate_dead(dead_cpu, next);
7437 * remove the tasks which were accounted by rq from calc_load_tasks.
7439 static void calc_global_load_remove(struct rq *rq)
7441 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7442 rq->calc_load_active = 0;
7444 #endif /* CONFIG_HOTPLUG_CPU */
7446 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7448 static struct ctl_table sd_ctl_dir[] = {
7450 .procname = "sched_domain",
7456 static struct ctl_table sd_ctl_root[] = {
7458 .procname = "kernel",
7460 .child = sd_ctl_dir,
7465 static struct ctl_table *sd_alloc_ctl_entry(int n)
7467 struct ctl_table *entry =
7468 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7473 static void sd_free_ctl_entry(struct ctl_table **tablep)
7475 struct ctl_table *entry;
7478 * In the intermediate directories, both the child directory and
7479 * procname are dynamically allocated and could fail but the mode
7480 * will always be set. In the lowest directory the names are
7481 * static strings and all have proc handlers.
7483 for (entry = *tablep; entry->mode; entry++) {
7485 sd_free_ctl_entry(&entry->child);
7486 if (entry->proc_handler == NULL)
7487 kfree(entry->procname);
7495 set_table_entry(struct ctl_table *entry,
7496 const char *procname, void *data, int maxlen,
7497 mode_t mode, proc_handler *proc_handler)
7499 entry->procname = procname;
7501 entry->maxlen = maxlen;
7503 entry->proc_handler = proc_handler;
7506 static struct ctl_table *
7507 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7509 struct ctl_table *table = sd_alloc_ctl_entry(13);
7514 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7515 sizeof(long), 0644, proc_doulongvec_minmax);
7516 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7517 sizeof(long), 0644, proc_doulongvec_minmax);
7518 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7519 sizeof(int), 0644, proc_dointvec_minmax);
7520 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7521 sizeof(int), 0644, proc_dointvec_minmax);
7522 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7523 sizeof(int), 0644, proc_dointvec_minmax);
7524 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7525 sizeof(int), 0644, proc_dointvec_minmax);
7526 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7527 sizeof(int), 0644, proc_dointvec_minmax);
7528 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7529 sizeof(int), 0644, proc_dointvec_minmax);
7530 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7531 sizeof(int), 0644, proc_dointvec_minmax);
7532 set_table_entry(&table[9], "cache_nice_tries",
7533 &sd->cache_nice_tries,
7534 sizeof(int), 0644, proc_dointvec_minmax);
7535 set_table_entry(&table[10], "flags", &sd->flags,
7536 sizeof(int), 0644, proc_dointvec_minmax);
7537 set_table_entry(&table[11], "name", sd->name,
7538 CORENAME_MAX_SIZE, 0444, proc_dostring);
7539 /* &table[12] is terminator */
7544 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7546 struct ctl_table *entry, *table;
7547 struct sched_domain *sd;
7548 int domain_num = 0, i;
7551 for_each_domain(cpu, sd)
7553 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7558 for_each_domain(cpu, sd) {
7559 snprintf(buf, 32, "domain%d", i);
7560 entry->procname = kstrdup(buf, GFP_KERNEL);
7562 entry->child = sd_alloc_ctl_domain_table(sd);
7569 static struct ctl_table_header *sd_sysctl_header;
7570 static void register_sched_domain_sysctl(void)
7572 int i, cpu_num = num_possible_cpus();
7573 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7576 WARN_ON(sd_ctl_dir[0].child);
7577 sd_ctl_dir[0].child = entry;
7582 for_each_possible_cpu(i) {
7583 snprintf(buf, 32, "cpu%d", i);
7584 entry->procname = kstrdup(buf, GFP_KERNEL);
7586 entry->child = sd_alloc_ctl_cpu_table(i);
7590 WARN_ON(sd_sysctl_header);
7591 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7594 /* may be called multiple times per register */
7595 static void unregister_sched_domain_sysctl(void)
7597 if (sd_sysctl_header)
7598 unregister_sysctl_table(sd_sysctl_header);
7599 sd_sysctl_header = NULL;
7600 if (sd_ctl_dir[0].child)
7601 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7604 static void register_sched_domain_sysctl(void)
7607 static void unregister_sched_domain_sysctl(void)
7612 static void set_rq_online(struct rq *rq)
7615 const struct sched_class *class;
7617 cpumask_set_cpu(rq->cpu, rq->rd->online);
7620 for_each_class(class) {
7621 if (class->rq_online)
7622 class->rq_online(rq);
7627 static void set_rq_offline(struct rq *rq)
7630 const struct sched_class *class;
7632 for_each_class(class) {
7633 if (class->rq_offline)
7634 class->rq_offline(rq);
7637 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7643 * migration_call - callback that gets triggered when a CPU is added.
7644 * Here we can start up the necessary migration thread for the new CPU.
7646 static int __cpuinit
7647 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7649 struct task_struct *p;
7650 int cpu = (long)hcpu;
7651 unsigned long flags;
7656 case CPU_UP_PREPARE:
7657 case CPU_UP_PREPARE_FROZEN:
7658 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7661 kthread_bind(p, cpu);
7662 /* Must be high prio: stop_machine expects to yield to it. */
7663 rq = task_rq_lock(p, &flags);
7664 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7665 task_rq_unlock(rq, &flags);
7667 cpu_rq(cpu)->migration_thread = p;
7668 rq->calc_load_update = calc_load_update;
7672 case CPU_ONLINE_FROZEN:
7673 /* Strictly unnecessary, as first user will wake it. */
7674 wake_up_process(cpu_rq(cpu)->migration_thread);
7676 /* Update our root-domain */
7678 spin_lock_irqsave(&rq->lock, flags);
7680 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7684 spin_unlock_irqrestore(&rq->lock, flags);
7687 #ifdef CONFIG_HOTPLUG_CPU
7688 case CPU_UP_CANCELED:
7689 case CPU_UP_CANCELED_FROZEN:
7690 if (!cpu_rq(cpu)->migration_thread)
7692 /* Unbind it from offline cpu so it can run. Fall thru. */
7693 kthread_bind(cpu_rq(cpu)->migration_thread,
7694 cpumask_any(cpu_online_mask));
7695 kthread_stop(cpu_rq(cpu)->migration_thread);
7696 put_task_struct(cpu_rq(cpu)->migration_thread);
7697 cpu_rq(cpu)->migration_thread = NULL;
7701 case CPU_DEAD_FROZEN:
7702 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7703 migrate_live_tasks(cpu);
7705 kthread_stop(rq->migration_thread);
7706 put_task_struct(rq->migration_thread);
7707 rq->migration_thread = NULL;
7708 /* Idle task back to normal (off runqueue, low prio) */
7709 spin_lock_irq(&rq->lock);
7710 update_rq_clock(rq);
7711 deactivate_task(rq, rq->idle, 0);
7712 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7713 rq->idle->sched_class = &idle_sched_class;
7714 migrate_dead_tasks(cpu);
7715 spin_unlock_irq(&rq->lock);
7717 migrate_nr_uninterruptible(rq);
7718 BUG_ON(rq->nr_running != 0);
7719 calc_global_load_remove(rq);
7721 * No need to migrate the tasks: it was best-effort if
7722 * they didn't take sched_hotcpu_mutex. Just wake up
7725 spin_lock_irq(&rq->lock);
7726 while (!list_empty(&rq->migration_queue)) {
7727 struct migration_req *req;
7729 req = list_entry(rq->migration_queue.next,
7730 struct migration_req, list);
7731 list_del_init(&req->list);
7732 spin_unlock_irq(&rq->lock);
7733 complete(&req->done);
7734 spin_lock_irq(&rq->lock);
7736 spin_unlock_irq(&rq->lock);
7740 case CPU_DYING_FROZEN:
7741 /* Update our root-domain */
7743 spin_lock_irqsave(&rq->lock, flags);
7745 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7748 spin_unlock_irqrestore(&rq->lock, flags);
7756 * Register at high priority so that task migration (migrate_all_tasks)
7757 * happens before everything else. This has to be lower priority than
7758 * the notifier in the perf_event subsystem, though.
7760 static struct notifier_block __cpuinitdata migration_notifier = {
7761 .notifier_call = migration_call,
7765 static int __init migration_init(void)
7767 void *cpu = (void *)(long)smp_processor_id();
7770 /* Start one for the boot CPU: */
7771 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7772 BUG_ON(err == NOTIFY_BAD);
7773 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7774 register_cpu_notifier(&migration_notifier);
7778 early_initcall(migration_init);
7783 #ifdef CONFIG_SCHED_DEBUG
7785 static __read_mostly int sched_domain_debug_enabled;
7787 static int __init sched_domain_debug_setup(char *str)
7789 sched_domain_debug_enabled = 1;
7793 early_param("sched_debug", sched_domain_debug_setup);
7795 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7796 struct cpumask *groupmask)
7798 struct sched_group *group = sd->groups;
7801 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7802 cpumask_clear(groupmask);
7804 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7806 if (!(sd->flags & SD_LOAD_BALANCE)) {
7807 pr_cont("does not load-balance\n");
7809 pr_err("ERROR: !SD_LOAD_BALANCE domain has parent\n");
7813 pr_cont("span %s level %s\n", str, sd->name);
7815 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7816 pr_err("ERROR: domain->span does not contain CPU%d\n", cpu);
7818 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7819 pr_err("ERROR: domain->groups does not contain CPU%d\n", cpu);
7822 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7826 pr_err("ERROR: group is NULL\n");
7830 if (!group->cpu_power) {
7832 pr_err("ERROR: domain->cpu_power not set\n");
7836 if (!cpumask_weight(sched_group_cpus(group))) {
7838 pr_err("ERROR: empty group\n");
7842 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7844 pr_err("ERROR: repeated CPUs\n");
7848 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7850 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7852 pr_cont(" %s", str);
7853 if (group->cpu_power != SCHED_LOAD_SCALE) {
7854 pr_cont(" (cpu_power = %d)", group->cpu_power);
7857 group = group->next;
7858 } while (group != sd->groups);
7861 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7862 pr_err("ERROR: groups don't span domain->span\n");
7865 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7866 pr_err("ERROR: parent span is not a superset of domain->span\n");
7870 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7872 cpumask_var_t groupmask;
7875 if (!sched_domain_debug_enabled)
7879 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7883 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7885 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7886 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7891 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7898 free_cpumask_var(groupmask);
7900 #else /* !CONFIG_SCHED_DEBUG */
7901 # define sched_domain_debug(sd, cpu) do { } while (0)
7902 #endif /* CONFIG_SCHED_DEBUG */
7904 static int sd_degenerate(struct sched_domain *sd)
7906 if (cpumask_weight(sched_domain_span(sd)) == 1)
7909 /* Following flags need at least 2 groups */
7910 if (sd->flags & (SD_LOAD_BALANCE |
7911 SD_BALANCE_NEWIDLE |
7915 SD_SHARE_PKG_RESOURCES)) {
7916 if (sd->groups != sd->groups->next)
7920 /* Following flags don't use groups */
7921 if (sd->flags & (SD_WAKE_AFFINE))
7928 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7930 unsigned long cflags = sd->flags, pflags = parent->flags;
7932 if (sd_degenerate(parent))
7935 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7938 /* Flags needing groups don't count if only 1 group in parent */
7939 if (parent->groups == parent->groups->next) {
7940 pflags &= ~(SD_LOAD_BALANCE |
7941 SD_BALANCE_NEWIDLE |
7945 SD_SHARE_PKG_RESOURCES);
7946 if (nr_node_ids == 1)
7947 pflags &= ~SD_SERIALIZE;
7949 if (~cflags & pflags)
7955 static void free_rootdomain(struct root_domain *rd)
7957 synchronize_sched();
7959 cpupri_cleanup(&rd->cpupri);
7961 free_cpumask_var(rd->rto_mask);
7962 free_cpumask_var(rd->online);
7963 free_cpumask_var(rd->span);
7967 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7969 struct root_domain *old_rd = NULL;
7970 unsigned long flags;
7972 spin_lock_irqsave(&rq->lock, flags);
7977 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7980 cpumask_clear_cpu(rq->cpu, old_rd->span);
7983 * If we dont want to free the old_rt yet then
7984 * set old_rd to NULL to skip the freeing later
7987 if (!atomic_dec_and_test(&old_rd->refcount))
7991 atomic_inc(&rd->refcount);
7994 cpumask_set_cpu(rq->cpu, rd->span);
7995 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7998 spin_unlock_irqrestore(&rq->lock, flags);
8001 free_rootdomain(old_rd);
8004 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8006 gfp_t gfp = GFP_KERNEL;
8008 memset(rd, 0, sizeof(*rd));
8013 if (!alloc_cpumask_var(&rd->span, gfp))
8015 if (!alloc_cpumask_var(&rd->online, gfp))
8017 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8020 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8025 free_cpumask_var(rd->rto_mask);
8027 free_cpumask_var(rd->online);
8029 free_cpumask_var(rd->span);
8034 static void init_defrootdomain(void)
8036 init_rootdomain(&def_root_domain, true);
8038 atomic_set(&def_root_domain.refcount, 1);
8041 static struct root_domain *alloc_rootdomain(void)
8043 struct root_domain *rd;
8045 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8049 if (init_rootdomain(rd, false) != 0) {
8058 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8059 * hold the hotplug lock.
8062 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8064 struct rq *rq = cpu_rq(cpu);
8065 struct sched_domain *tmp;
8067 /* Remove the sched domains which do not contribute to scheduling. */
8068 for (tmp = sd; tmp; ) {
8069 struct sched_domain *parent = tmp->parent;
8073 if (sd_parent_degenerate(tmp, parent)) {
8074 tmp->parent = parent->parent;
8076 parent->parent->child = tmp;
8081 if (sd && sd_degenerate(sd)) {
8087 sched_domain_debug(sd, cpu);
8089 rq_attach_root(rq, rd);
8090 rcu_assign_pointer(rq->sd, sd);
8093 /* cpus with isolated domains */
8094 static cpumask_var_t cpu_isolated_map;
8096 /* Setup the mask of cpus configured for isolated domains */
8097 static int __init isolated_cpu_setup(char *str)
8099 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8100 cpulist_parse(str, cpu_isolated_map);
8104 __setup("isolcpus=", isolated_cpu_setup);
8107 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8108 * to a function which identifies what group(along with sched group) a CPU
8109 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8110 * (due to the fact that we keep track of groups covered with a struct cpumask).
8112 * init_sched_build_groups will build a circular linked list of the groups
8113 * covered by the given span, and will set each group's ->cpumask correctly,
8114 * and ->cpu_power to 0.
8117 init_sched_build_groups(const struct cpumask *span,
8118 const struct cpumask *cpu_map,
8119 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8120 struct sched_group **sg,
8121 struct cpumask *tmpmask),
8122 struct cpumask *covered, struct cpumask *tmpmask)
8124 struct sched_group *first = NULL, *last = NULL;
8127 cpumask_clear(covered);
8129 for_each_cpu(i, span) {
8130 struct sched_group *sg;
8131 int group = group_fn(i, cpu_map, &sg, tmpmask);
8134 if (cpumask_test_cpu(i, covered))
8137 cpumask_clear(sched_group_cpus(sg));
8140 for_each_cpu(j, span) {
8141 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8144 cpumask_set_cpu(j, covered);
8145 cpumask_set_cpu(j, sched_group_cpus(sg));
8156 #define SD_NODES_PER_DOMAIN 16
8161 * find_next_best_node - find the next node to include in a sched_domain
8162 * @node: node whose sched_domain we're building
8163 * @used_nodes: nodes already in the sched_domain
8165 * Find the next node to include in a given scheduling domain. Simply
8166 * finds the closest node not already in the @used_nodes map.
8168 * Should use nodemask_t.
8170 static int find_next_best_node(int node, nodemask_t *used_nodes)
8172 int i, n, val, min_val, best_node = 0;
8176 for (i = 0; i < nr_node_ids; i++) {
8177 /* Start at @node */
8178 n = (node + i) % nr_node_ids;
8180 if (!nr_cpus_node(n))
8183 /* Skip already used nodes */
8184 if (node_isset(n, *used_nodes))
8187 /* Simple min distance search */
8188 val = node_distance(node, n);
8190 if (val < min_val) {
8196 node_set(best_node, *used_nodes);
8201 * sched_domain_node_span - get a cpumask for a node's sched_domain
8202 * @node: node whose cpumask we're constructing
8203 * @span: resulting cpumask
8205 * Given a node, construct a good cpumask for its sched_domain to span. It
8206 * should be one that prevents unnecessary balancing, but also spreads tasks
8209 static void sched_domain_node_span(int node, struct cpumask *span)
8211 nodemask_t used_nodes;
8214 cpumask_clear(span);
8215 nodes_clear(used_nodes);
8217 cpumask_or(span, span, cpumask_of_node(node));
8218 node_set(node, used_nodes);
8220 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8221 int next_node = find_next_best_node(node, &used_nodes);
8223 cpumask_or(span, span, cpumask_of_node(next_node));
8226 #endif /* CONFIG_NUMA */
8228 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8231 * The cpus mask in sched_group and sched_domain hangs off the end.
8233 * ( See the the comments in include/linux/sched.h:struct sched_group
8234 * and struct sched_domain. )
8236 struct static_sched_group {
8237 struct sched_group sg;
8238 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8241 struct static_sched_domain {
8242 struct sched_domain sd;
8243 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8249 cpumask_var_t domainspan;
8250 cpumask_var_t covered;
8251 cpumask_var_t notcovered;
8253 cpumask_var_t nodemask;
8254 cpumask_var_t this_sibling_map;
8255 cpumask_var_t this_core_map;
8256 cpumask_var_t send_covered;
8257 cpumask_var_t tmpmask;
8258 struct sched_group **sched_group_nodes;
8259 struct root_domain *rd;
8263 sa_sched_groups = 0,
8268 sa_this_sibling_map,
8270 sa_sched_group_nodes,
8280 * SMT sched-domains:
8282 #ifdef CONFIG_SCHED_SMT
8283 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8284 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8287 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8288 struct sched_group **sg, struct cpumask *unused)
8291 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8294 #endif /* CONFIG_SCHED_SMT */
8297 * multi-core sched-domains:
8299 #ifdef CONFIG_SCHED_MC
8300 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8301 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8302 #endif /* CONFIG_SCHED_MC */
8304 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8306 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8307 struct sched_group **sg, struct cpumask *mask)
8311 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8312 group = cpumask_first(mask);
8314 *sg = &per_cpu(sched_group_core, group).sg;
8317 #elif defined(CONFIG_SCHED_MC)
8319 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8320 struct sched_group **sg, struct cpumask *unused)
8323 *sg = &per_cpu(sched_group_core, cpu).sg;
8328 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8329 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8332 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8333 struct sched_group **sg, struct cpumask *mask)
8336 #ifdef CONFIG_SCHED_MC
8337 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8338 group = cpumask_first(mask);
8339 #elif defined(CONFIG_SCHED_SMT)
8340 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8341 group = cpumask_first(mask);
8346 *sg = &per_cpu(sched_group_phys, group).sg;
8352 * The init_sched_build_groups can't handle what we want to do with node
8353 * groups, so roll our own. Now each node has its own list of groups which
8354 * gets dynamically allocated.
8356 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8357 static struct sched_group ***sched_group_nodes_bycpu;
8359 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8360 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8362 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8363 struct sched_group **sg,
8364 struct cpumask *nodemask)
8368 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8369 group = cpumask_first(nodemask);
8372 *sg = &per_cpu(sched_group_allnodes, group).sg;
8376 static void init_numa_sched_groups_power(struct sched_group *group_head)
8378 struct sched_group *sg = group_head;
8384 for_each_cpu(j, sched_group_cpus(sg)) {
8385 struct sched_domain *sd;
8387 sd = &per_cpu(phys_domains, j).sd;
8388 if (j != group_first_cpu(sd->groups)) {
8390 * Only add "power" once for each
8396 sg->cpu_power += sd->groups->cpu_power;
8399 } while (sg != group_head);
8402 static int build_numa_sched_groups(struct s_data *d,
8403 const struct cpumask *cpu_map, int num)
8405 struct sched_domain *sd;
8406 struct sched_group *sg, *prev;
8409 cpumask_clear(d->covered);
8410 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8411 if (cpumask_empty(d->nodemask)) {
8412 d->sched_group_nodes[num] = NULL;
8416 sched_domain_node_span(num, d->domainspan);
8417 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8419 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8422 pr_warning("Can not alloc domain group for node %d\n", num);
8425 d->sched_group_nodes[num] = sg;
8427 for_each_cpu(j, d->nodemask) {
8428 sd = &per_cpu(node_domains, j).sd;
8433 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8435 cpumask_or(d->covered, d->covered, d->nodemask);
8438 for (j = 0; j < nr_node_ids; j++) {
8439 n = (num + j) % nr_node_ids;
8440 cpumask_complement(d->notcovered, d->covered);
8441 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8442 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8443 if (cpumask_empty(d->tmpmask))
8445 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8446 if (cpumask_empty(d->tmpmask))
8448 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8451 pr_warning("Can not alloc domain group for node %d\n",
8456 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8457 sg->next = prev->next;
8458 cpumask_or(d->covered, d->covered, d->tmpmask);
8465 #endif /* CONFIG_NUMA */
8468 /* Free memory allocated for various sched_group structures */
8469 static void free_sched_groups(const struct cpumask *cpu_map,
8470 struct cpumask *nodemask)
8474 for_each_cpu(cpu, cpu_map) {
8475 struct sched_group **sched_group_nodes
8476 = sched_group_nodes_bycpu[cpu];
8478 if (!sched_group_nodes)
8481 for (i = 0; i < nr_node_ids; i++) {
8482 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8484 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8485 if (cpumask_empty(nodemask))
8495 if (oldsg != sched_group_nodes[i])
8498 kfree(sched_group_nodes);
8499 sched_group_nodes_bycpu[cpu] = NULL;
8502 #else /* !CONFIG_NUMA */
8503 static void free_sched_groups(const struct cpumask *cpu_map,
8504 struct cpumask *nodemask)
8507 #endif /* CONFIG_NUMA */
8510 * Initialize sched groups cpu_power.
8512 * cpu_power indicates the capacity of sched group, which is used while
8513 * distributing the load between different sched groups in a sched domain.
8514 * Typically cpu_power for all the groups in a sched domain will be same unless
8515 * there are asymmetries in the topology. If there are asymmetries, group
8516 * having more cpu_power will pickup more load compared to the group having
8519 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8521 struct sched_domain *child;
8522 struct sched_group *group;
8526 WARN_ON(!sd || !sd->groups);
8528 if (cpu != group_first_cpu(sd->groups))
8533 sd->groups->cpu_power = 0;
8536 power = SCHED_LOAD_SCALE;
8537 weight = cpumask_weight(sched_domain_span(sd));
8539 * SMT siblings share the power of a single core.
8540 * Usually multiple threads get a better yield out of
8541 * that one core than a single thread would have,
8542 * reflect that in sd->smt_gain.
8544 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8545 power *= sd->smt_gain;
8547 power >>= SCHED_LOAD_SHIFT;
8549 sd->groups->cpu_power += power;
8554 * Add cpu_power of each child group to this groups cpu_power.
8556 group = child->groups;
8558 sd->groups->cpu_power += group->cpu_power;
8559 group = group->next;
8560 } while (group != child->groups);
8564 * Initializers for schedule domains
8565 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8568 #ifdef CONFIG_SCHED_DEBUG
8569 # define SD_INIT_NAME(sd, type) sd->name = #type
8571 # define SD_INIT_NAME(sd, type) do { } while (0)
8574 #define SD_INIT(sd, type) sd_init_##type(sd)
8576 #define SD_INIT_FUNC(type) \
8577 static noinline void sd_init_##type(struct sched_domain *sd) \
8579 memset(sd, 0, sizeof(*sd)); \
8580 *sd = SD_##type##_INIT; \
8581 sd->level = SD_LV_##type; \
8582 SD_INIT_NAME(sd, type); \
8587 SD_INIT_FUNC(ALLNODES)
8590 #ifdef CONFIG_SCHED_SMT
8591 SD_INIT_FUNC(SIBLING)
8593 #ifdef CONFIG_SCHED_MC
8597 static int default_relax_domain_level = -1;
8599 static int __init setup_relax_domain_level(char *str)
8603 val = simple_strtoul(str, NULL, 0);
8604 if (val < SD_LV_MAX)
8605 default_relax_domain_level = val;
8609 __setup("relax_domain_level=", setup_relax_domain_level);
8611 static void set_domain_attribute(struct sched_domain *sd,
8612 struct sched_domain_attr *attr)
8616 if (!attr || attr->relax_domain_level < 0) {
8617 if (default_relax_domain_level < 0)
8620 request = default_relax_domain_level;
8622 request = attr->relax_domain_level;
8623 if (request < sd->level) {
8624 /* turn off idle balance on this domain */
8625 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8627 /* turn on idle balance on this domain */
8628 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8632 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8633 const struct cpumask *cpu_map)
8636 case sa_sched_groups:
8637 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8638 d->sched_group_nodes = NULL;
8640 free_rootdomain(d->rd); /* fall through */
8642 free_cpumask_var(d->tmpmask); /* fall through */
8643 case sa_send_covered:
8644 free_cpumask_var(d->send_covered); /* fall through */
8645 case sa_this_core_map:
8646 free_cpumask_var(d->this_core_map); /* fall through */
8647 case sa_this_sibling_map:
8648 free_cpumask_var(d->this_sibling_map); /* fall through */
8650 free_cpumask_var(d->nodemask); /* fall through */
8651 case sa_sched_group_nodes:
8653 kfree(d->sched_group_nodes); /* fall through */
8655 free_cpumask_var(d->notcovered); /* fall through */
8657 free_cpumask_var(d->covered); /* fall through */
8659 free_cpumask_var(d->domainspan); /* fall through */
8666 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8667 const struct cpumask *cpu_map)
8670 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8672 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8673 return sa_domainspan;
8674 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8676 /* Allocate the per-node list of sched groups */
8677 d->sched_group_nodes = kcalloc(nr_node_ids,
8678 sizeof(struct sched_group *), GFP_KERNEL);
8679 if (!d->sched_group_nodes) {
8680 pr_warning("Can not alloc sched group node list\n");
8681 return sa_notcovered;
8683 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8685 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8686 return sa_sched_group_nodes;
8687 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8689 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8690 return sa_this_sibling_map;
8691 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8692 return sa_this_core_map;
8693 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8694 return sa_send_covered;
8695 d->rd = alloc_rootdomain();
8697 pr_warning("Cannot alloc root domain\n");
8700 return sa_rootdomain;
8703 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8704 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8706 struct sched_domain *sd = NULL;
8708 struct sched_domain *parent;
8711 if (cpumask_weight(cpu_map) >
8712 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8713 sd = &per_cpu(allnodes_domains, i).sd;
8714 SD_INIT(sd, ALLNODES);
8715 set_domain_attribute(sd, attr);
8716 cpumask_copy(sched_domain_span(sd), cpu_map);
8717 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8722 sd = &per_cpu(node_domains, i).sd;
8724 set_domain_attribute(sd, attr);
8725 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8726 sd->parent = parent;
8729 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8734 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8735 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8736 struct sched_domain *parent, int i)
8738 struct sched_domain *sd;
8739 sd = &per_cpu(phys_domains, i).sd;
8741 set_domain_attribute(sd, attr);
8742 cpumask_copy(sched_domain_span(sd), d->nodemask);
8743 sd->parent = parent;
8746 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8750 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8751 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8752 struct sched_domain *parent, int i)
8754 struct sched_domain *sd = parent;
8755 #ifdef CONFIG_SCHED_MC
8756 sd = &per_cpu(core_domains, i).sd;
8758 set_domain_attribute(sd, attr);
8759 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8760 sd->parent = parent;
8762 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8767 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8768 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8769 struct sched_domain *parent, int i)
8771 struct sched_domain *sd = parent;
8772 #ifdef CONFIG_SCHED_SMT
8773 sd = &per_cpu(cpu_domains, i).sd;
8774 SD_INIT(sd, SIBLING);
8775 set_domain_attribute(sd, attr);
8776 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8777 sd->parent = parent;
8779 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8784 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8785 const struct cpumask *cpu_map, int cpu)
8788 #ifdef CONFIG_SCHED_SMT
8789 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8790 cpumask_and(d->this_sibling_map, cpu_map,
8791 topology_thread_cpumask(cpu));
8792 if (cpu == cpumask_first(d->this_sibling_map))
8793 init_sched_build_groups(d->this_sibling_map, cpu_map,
8795 d->send_covered, d->tmpmask);
8798 #ifdef CONFIG_SCHED_MC
8799 case SD_LV_MC: /* set up multi-core groups */
8800 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8801 if (cpu == cpumask_first(d->this_core_map))
8802 init_sched_build_groups(d->this_core_map, cpu_map,
8804 d->send_covered, d->tmpmask);
8807 case SD_LV_CPU: /* set up physical groups */
8808 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8809 if (!cpumask_empty(d->nodemask))
8810 init_sched_build_groups(d->nodemask, cpu_map,
8812 d->send_covered, d->tmpmask);
8815 case SD_LV_ALLNODES:
8816 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8817 d->send_covered, d->tmpmask);
8826 * Build sched domains for a given set of cpus and attach the sched domains
8827 * to the individual cpus
8829 static int __build_sched_domains(const struct cpumask *cpu_map,
8830 struct sched_domain_attr *attr)
8832 enum s_alloc alloc_state = sa_none;
8834 struct sched_domain *sd;
8840 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8841 if (alloc_state != sa_rootdomain)
8843 alloc_state = sa_sched_groups;
8846 * Set up domains for cpus specified by the cpu_map.
8848 for_each_cpu(i, cpu_map) {
8849 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8852 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8853 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8854 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8855 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8858 for_each_cpu(i, cpu_map) {
8859 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8860 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8863 /* Set up physical groups */
8864 for (i = 0; i < nr_node_ids; i++)
8865 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8868 /* Set up node groups */
8870 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8872 for (i = 0; i < nr_node_ids; i++)
8873 if (build_numa_sched_groups(&d, cpu_map, i))
8877 /* Calculate CPU power for physical packages and nodes */
8878 #ifdef CONFIG_SCHED_SMT
8879 for_each_cpu(i, cpu_map) {
8880 sd = &per_cpu(cpu_domains, i).sd;
8881 init_sched_groups_power(i, sd);
8884 #ifdef CONFIG_SCHED_MC
8885 for_each_cpu(i, cpu_map) {
8886 sd = &per_cpu(core_domains, i).sd;
8887 init_sched_groups_power(i, sd);
8891 for_each_cpu(i, cpu_map) {
8892 sd = &per_cpu(phys_domains, i).sd;
8893 init_sched_groups_power(i, sd);
8897 for (i = 0; i < nr_node_ids; i++)
8898 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8900 if (d.sd_allnodes) {
8901 struct sched_group *sg;
8903 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8905 init_numa_sched_groups_power(sg);
8909 /* Attach the domains */
8910 for_each_cpu(i, cpu_map) {
8911 #ifdef CONFIG_SCHED_SMT
8912 sd = &per_cpu(cpu_domains, i).sd;
8913 #elif defined(CONFIG_SCHED_MC)
8914 sd = &per_cpu(core_domains, i).sd;
8916 sd = &per_cpu(phys_domains, i).sd;
8918 cpu_attach_domain(sd, d.rd, i);
8921 d.sched_group_nodes = NULL; /* don't free this we still need it */
8922 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8926 __free_domain_allocs(&d, alloc_state, cpu_map);
8930 static int build_sched_domains(const struct cpumask *cpu_map)
8932 return __build_sched_domains(cpu_map, NULL);
8935 static cpumask_var_t *doms_cur; /* current sched domains */
8936 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8937 static struct sched_domain_attr *dattr_cur;
8938 /* attribues of custom domains in 'doms_cur' */
8941 * Special case: If a kmalloc of a doms_cur partition (array of
8942 * cpumask) fails, then fallback to a single sched domain,
8943 * as determined by the single cpumask fallback_doms.
8945 static cpumask_var_t fallback_doms;
8948 * arch_update_cpu_topology lets virtualized architectures update the
8949 * cpu core maps. It is supposed to return 1 if the topology changed
8950 * or 0 if it stayed the same.
8952 int __attribute__((weak)) arch_update_cpu_topology(void)
8957 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8960 cpumask_var_t *doms;
8962 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8965 for (i = 0; i < ndoms; i++) {
8966 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
8967 free_sched_domains(doms, i);
8974 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
8977 for (i = 0; i < ndoms; i++)
8978 free_cpumask_var(doms[i]);
8983 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8984 * For now this just excludes isolated cpus, but could be used to
8985 * exclude other special cases in the future.
8987 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8991 arch_update_cpu_topology();
8993 doms_cur = alloc_sched_domains(ndoms_cur);
8995 doms_cur = &fallback_doms;
8996 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
8998 err = build_sched_domains(doms_cur[0]);
8999 register_sched_domain_sysctl();
9004 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9005 struct cpumask *tmpmask)
9007 free_sched_groups(cpu_map, tmpmask);
9011 * Detach sched domains from a group of cpus specified in cpu_map
9012 * These cpus will now be attached to the NULL domain
9014 static void detach_destroy_domains(const struct cpumask *cpu_map)
9016 /* Save because hotplug lock held. */
9017 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9020 for_each_cpu(i, cpu_map)
9021 cpu_attach_domain(NULL, &def_root_domain, i);
9022 synchronize_sched();
9023 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9026 /* handle null as "default" */
9027 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9028 struct sched_domain_attr *new, int idx_new)
9030 struct sched_domain_attr tmp;
9037 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9038 new ? (new + idx_new) : &tmp,
9039 sizeof(struct sched_domain_attr));
9043 * Partition sched domains as specified by the 'ndoms_new'
9044 * cpumasks in the array doms_new[] of cpumasks. This compares
9045 * doms_new[] to the current sched domain partitioning, doms_cur[].
9046 * It destroys each deleted domain and builds each new domain.
9048 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9049 * The masks don't intersect (don't overlap.) We should setup one
9050 * sched domain for each mask. CPUs not in any of the cpumasks will
9051 * not be load balanced. If the same cpumask appears both in the
9052 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9055 * The passed in 'doms_new' should be allocated using
9056 * alloc_sched_domains. This routine takes ownership of it and will
9057 * free_sched_domains it when done with it. If the caller failed the
9058 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9059 * and partition_sched_domains() will fallback to the single partition
9060 * 'fallback_doms', it also forces the domains to be rebuilt.
9062 * If doms_new == NULL it will be replaced with cpu_online_mask.
9063 * ndoms_new == 0 is a special case for destroying existing domains,
9064 * and it will not create the default domain.
9066 * Call with hotplug lock held
9068 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9069 struct sched_domain_attr *dattr_new)
9074 mutex_lock(&sched_domains_mutex);
9076 /* always unregister in case we don't destroy any domains */
9077 unregister_sched_domain_sysctl();
9079 /* Let architecture update cpu core mappings. */
9080 new_topology = arch_update_cpu_topology();
9082 n = doms_new ? ndoms_new : 0;
9084 /* Destroy deleted domains */
9085 for (i = 0; i < ndoms_cur; i++) {
9086 for (j = 0; j < n && !new_topology; j++) {
9087 if (cpumask_equal(doms_cur[i], doms_new[j])
9088 && dattrs_equal(dattr_cur, i, dattr_new, j))
9091 /* no match - a current sched domain not in new doms_new[] */
9092 detach_destroy_domains(doms_cur[i]);
9097 if (doms_new == NULL) {
9099 doms_new = &fallback_doms;
9100 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9101 WARN_ON_ONCE(dattr_new);
9104 /* Build new domains */
9105 for (i = 0; i < ndoms_new; i++) {
9106 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9107 if (cpumask_equal(doms_new[i], doms_cur[j])
9108 && dattrs_equal(dattr_new, i, dattr_cur, j))
9111 /* no match - add a new doms_new */
9112 __build_sched_domains(doms_new[i],
9113 dattr_new ? dattr_new + i : NULL);
9118 /* Remember the new sched domains */
9119 if (doms_cur != &fallback_doms)
9120 free_sched_domains(doms_cur, ndoms_cur);
9121 kfree(dattr_cur); /* kfree(NULL) is safe */
9122 doms_cur = doms_new;
9123 dattr_cur = dattr_new;
9124 ndoms_cur = ndoms_new;
9126 register_sched_domain_sysctl();
9128 mutex_unlock(&sched_domains_mutex);
9131 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9132 static void arch_reinit_sched_domains(void)
9136 /* Destroy domains first to force the rebuild */
9137 partition_sched_domains(0, NULL, NULL);
9139 rebuild_sched_domains();
9143 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9145 unsigned int level = 0;
9147 if (sscanf(buf, "%u", &level) != 1)
9151 * level is always be positive so don't check for
9152 * level < POWERSAVINGS_BALANCE_NONE which is 0
9153 * What happens on 0 or 1 byte write,
9154 * need to check for count as well?
9157 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9161 sched_smt_power_savings = level;
9163 sched_mc_power_savings = level;
9165 arch_reinit_sched_domains();
9170 #ifdef CONFIG_SCHED_MC
9171 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9174 return sprintf(page, "%u\n", sched_mc_power_savings);
9176 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9177 const char *buf, size_t count)
9179 return sched_power_savings_store(buf, count, 0);
9181 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9182 sched_mc_power_savings_show,
9183 sched_mc_power_savings_store);
9186 #ifdef CONFIG_SCHED_SMT
9187 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9190 return sprintf(page, "%u\n", sched_smt_power_savings);
9192 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9193 const char *buf, size_t count)
9195 return sched_power_savings_store(buf, count, 1);
9197 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9198 sched_smt_power_savings_show,
9199 sched_smt_power_savings_store);
9202 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9206 #ifdef CONFIG_SCHED_SMT
9208 err = sysfs_create_file(&cls->kset.kobj,
9209 &attr_sched_smt_power_savings.attr);
9211 #ifdef CONFIG_SCHED_MC
9212 if (!err && mc_capable())
9213 err = sysfs_create_file(&cls->kset.kobj,
9214 &attr_sched_mc_power_savings.attr);
9218 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9220 #ifndef CONFIG_CPUSETS
9222 * Add online and remove offline CPUs from the scheduler domains.
9223 * When cpusets are enabled they take over this function.
9225 static int update_sched_domains(struct notifier_block *nfb,
9226 unsigned long action, void *hcpu)
9230 case CPU_ONLINE_FROZEN:
9231 case CPU_DOWN_PREPARE:
9232 case CPU_DOWN_PREPARE_FROZEN:
9233 case CPU_DOWN_FAILED:
9234 case CPU_DOWN_FAILED_FROZEN:
9235 partition_sched_domains(1, NULL, NULL);
9244 static int update_runtime(struct notifier_block *nfb,
9245 unsigned long action, void *hcpu)
9247 int cpu = (int)(long)hcpu;
9250 case CPU_DOWN_PREPARE:
9251 case CPU_DOWN_PREPARE_FROZEN:
9252 disable_runtime(cpu_rq(cpu));
9255 case CPU_DOWN_FAILED:
9256 case CPU_DOWN_FAILED_FROZEN:
9258 case CPU_ONLINE_FROZEN:
9259 enable_runtime(cpu_rq(cpu));
9267 void __init sched_init_smp(void)
9269 cpumask_var_t non_isolated_cpus;
9271 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9272 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9274 #if defined(CONFIG_NUMA)
9275 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9277 BUG_ON(sched_group_nodes_bycpu == NULL);
9280 mutex_lock(&sched_domains_mutex);
9281 arch_init_sched_domains(cpu_active_mask);
9282 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9283 if (cpumask_empty(non_isolated_cpus))
9284 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9285 mutex_unlock(&sched_domains_mutex);
9288 #ifndef CONFIG_CPUSETS
9289 /* XXX: Theoretical race here - CPU may be hotplugged now */
9290 hotcpu_notifier(update_sched_domains, 0);
9293 /* RT runtime code needs to handle some hotplug events */
9294 hotcpu_notifier(update_runtime, 0);
9298 /* Move init over to a non-isolated CPU */
9299 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9301 sched_init_granularity();
9302 free_cpumask_var(non_isolated_cpus);
9304 init_sched_rt_class();
9307 void __init sched_init_smp(void)
9309 sched_init_granularity();
9311 #endif /* CONFIG_SMP */
9313 const_debug unsigned int sysctl_timer_migration = 1;
9315 int in_sched_functions(unsigned long addr)
9317 return in_lock_functions(addr) ||
9318 (addr >= (unsigned long)__sched_text_start
9319 && addr < (unsigned long)__sched_text_end);
9322 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9324 cfs_rq->tasks_timeline = RB_ROOT;
9325 INIT_LIST_HEAD(&cfs_rq->tasks);
9326 #ifdef CONFIG_FAIR_GROUP_SCHED
9329 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9332 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9334 struct rt_prio_array *array;
9337 array = &rt_rq->active;
9338 for (i = 0; i < MAX_RT_PRIO; i++) {
9339 INIT_LIST_HEAD(array->queue + i);
9340 __clear_bit(i, array->bitmap);
9342 /* delimiter for bitsearch: */
9343 __set_bit(MAX_RT_PRIO, array->bitmap);
9345 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9346 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9348 rt_rq->highest_prio.next = MAX_RT_PRIO;
9352 rt_rq->rt_nr_migratory = 0;
9353 rt_rq->overloaded = 0;
9354 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9358 rt_rq->rt_throttled = 0;
9359 rt_rq->rt_runtime = 0;
9360 spin_lock_init(&rt_rq->rt_runtime_lock);
9362 #ifdef CONFIG_RT_GROUP_SCHED
9363 rt_rq->rt_nr_boosted = 0;
9368 #ifdef CONFIG_FAIR_GROUP_SCHED
9369 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9370 struct sched_entity *se, int cpu, int add,
9371 struct sched_entity *parent)
9373 struct rq *rq = cpu_rq(cpu);
9374 tg->cfs_rq[cpu] = cfs_rq;
9375 init_cfs_rq(cfs_rq, rq);
9378 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9381 /* se could be NULL for init_task_group */
9386 se->cfs_rq = &rq->cfs;
9388 se->cfs_rq = parent->my_q;
9391 se->load.weight = tg->shares;
9392 se->load.inv_weight = 0;
9393 se->parent = parent;
9397 #ifdef CONFIG_RT_GROUP_SCHED
9398 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9399 struct sched_rt_entity *rt_se, int cpu, int add,
9400 struct sched_rt_entity *parent)
9402 struct rq *rq = cpu_rq(cpu);
9404 tg->rt_rq[cpu] = rt_rq;
9405 init_rt_rq(rt_rq, rq);
9407 rt_rq->rt_se = rt_se;
9408 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9410 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9412 tg->rt_se[cpu] = rt_se;
9417 rt_se->rt_rq = &rq->rt;
9419 rt_se->rt_rq = parent->my_q;
9421 rt_se->my_q = rt_rq;
9422 rt_se->parent = parent;
9423 INIT_LIST_HEAD(&rt_se->run_list);
9427 void __init sched_init(void)
9430 unsigned long alloc_size = 0, ptr;
9432 #ifdef CONFIG_FAIR_GROUP_SCHED
9433 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9435 #ifdef CONFIG_RT_GROUP_SCHED
9436 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9438 #ifdef CONFIG_USER_SCHED
9441 #ifdef CONFIG_CPUMASK_OFFSTACK
9442 alloc_size += num_possible_cpus() * cpumask_size();
9445 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9447 #ifdef CONFIG_FAIR_GROUP_SCHED
9448 init_task_group.se = (struct sched_entity **)ptr;
9449 ptr += nr_cpu_ids * sizeof(void **);
9451 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9452 ptr += nr_cpu_ids * sizeof(void **);
9454 #ifdef CONFIG_USER_SCHED
9455 root_task_group.se = (struct sched_entity **)ptr;
9456 ptr += nr_cpu_ids * sizeof(void **);
9458 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9459 ptr += nr_cpu_ids * sizeof(void **);
9460 #endif /* CONFIG_USER_SCHED */
9461 #endif /* CONFIG_FAIR_GROUP_SCHED */
9462 #ifdef CONFIG_RT_GROUP_SCHED
9463 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9464 ptr += nr_cpu_ids * sizeof(void **);
9466 init_task_group.rt_rq = (struct rt_rq **)ptr;
9467 ptr += nr_cpu_ids * sizeof(void **);
9469 #ifdef CONFIG_USER_SCHED
9470 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9471 ptr += nr_cpu_ids * sizeof(void **);
9473 root_task_group.rt_rq = (struct rt_rq **)ptr;
9474 ptr += nr_cpu_ids * sizeof(void **);
9475 #endif /* CONFIG_USER_SCHED */
9476 #endif /* CONFIG_RT_GROUP_SCHED */
9477 #ifdef CONFIG_CPUMASK_OFFSTACK
9478 for_each_possible_cpu(i) {
9479 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9480 ptr += cpumask_size();
9482 #endif /* CONFIG_CPUMASK_OFFSTACK */
9486 init_defrootdomain();
9489 init_rt_bandwidth(&def_rt_bandwidth,
9490 global_rt_period(), global_rt_runtime());
9492 #ifdef CONFIG_RT_GROUP_SCHED
9493 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9494 global_rt_period(), global_rt_runtime());
9495 #ifdef CONFIG_USER_SCHED
9496 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9497 global_rt_period(), RUNTIME_INF);
9498 #endif /* CONFIG_USER_SCHED */
9499 #endif /* CONFIG_RT_GROUP_SCHED */
9501 #ifdef CONFIG_GROUP_SCHED
9502 list_add(&init_task_group.list, &task_groups);
9503 INIT_LIST_HEAD(&init_task_group.children);
9505 #ifdef CONFIG_USER_SCHED
9506 INIT_LIST_HEAD(&root_task_group.children);
9507 init_task_group.parent = &root_task_group;
9508 list_add(&init_task_group.siblings, &root_task_group.children);
9509 #endif /* CONFIG_USER_SCHED */
9510 #endif /* CONFIG_GROUP_SCHED */
9512 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9513 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9514 __alignof__(unsigned long));
9516 for_each_possible_cpu(i) {
9520 spin_lock_init(&rq->lock);
9522 rq->calc_load_active = 0;
9523 rq->calc_load_update = jiffies + LOAD_FREQ;
9524 init_cfs_rq(&rq->cfs, rq);
9525 init_rt_rq(&rq->rt, rq);
9526 #ifdef CONFIG_FAIR_GROUP_SCHED
9527 init_task_group.shares = init_task_group_load;
9528 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9529 #ifdef CONFIG_CGROUP_SCHED
9531 * How much cpu bandwidth does init_task_group get?
9533 * In case of task-groups formed thr' the cgroup filesystem, it
9534 * gets 100% of the cpu resources in the system. This overall
9535 * system cpu resource is divided among the tasks of
9536 * init_task_group and its child task-groups in a fair manner,
9537 * based on each entity's (task or task-group's) weight
9538 * (se->load.weight).
9540 * In other words, if init_task_group has 10 tasks of weight
9541 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9542 * then A0's share of the cpu resource is:
9544 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9546 * We achieve this by letting init_task_group's tasks sit
9547 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9549 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9550 #elif defined CONFIG_USER_SCHED
9551 root_task_group.shares = NICE_0_LOAD;
9552 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9554 * In case of task-groups formed thr' the user id of tasks,
9555 * init_task_group represents tasks belonging to root user.
9556 * Hence it forms a sibling of all subsequent groups formed.
9557 * In this case, init_task_group gets only a fraction of overall
9558 * system cpu resource, based on the weight assigned to root
9559 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9560 * by letting tasks of init_task_group sit in a separate cfs_rq
9561 * (init_tg_cfs_rq) and having one entity represent this group of
9562 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9564 init_tg_cfs_entry(&init_task_group,
9565 &per_cpu(init_tg_cfs_rq, i),
9566 &per_cpu(init_sched_entity, i), i, 1,
9567 root_task_group.se[i]);
9570 #endif /* CONFIG_FAIR_GROUP_SCHED */
9572 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9573 #ifdef CONFIG_RT_GROUP_SCHED
9574 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9575 #ifdef CONFIG_CGROUP_SCHED
9576 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9577 #elif defined CONFIG_USER_SCHED
9578 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9579 init_tg_rt_entry(&init_task_group,
9580 &per_cpu(init_rt_rq, i),
9581 &per_cpu(init_sched_rt_entity, i), i, 1,
9582 root_task_group.rt_se[i]);
9586 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9587 rq->cpu_load[j] = 0;
9591 rq->post_schedule = 0;
9592 rq->active_balance = 0;
9593 rq->next_balance = jiffies;
9597 rq->migration_thread = NULL;
9599 rq->avg_idle = 2*sysctl_sched_migration_cost;
9600 INIT_LIST_HEAD(&rq->migration_queue);
9601 rq_attach_root(rq, &def_root_domain);
9604 atomic_set(&rq->nr_iowait, 0);
9607 set_load_weight(&init_task);
9609 #ifdef CONFIG_PREEMPT_NOTIFIERS
9610 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9614 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9617 #ifdef CONFIG_RT_MUTEXES
9618 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9622 * The boot idle thread does lazy MMU switching as well:
9624 atomic_inc(&init_mm.mm_count);
9625 enter_lazy_tlb(&init_mm, current);
9628 * Make us the idle thread. Technically, schedule() should not be
9629 * called from this thread, however somewhere below it might be,
9630 * but because we are the idle thread, we just pick up running again
9631 * when this runqueue becomes "idle".
9633 init_idle(current, smp_processor_id());
9635 calc_load_update = jiffies + LOAD_FREQ;
9638 * During early bootup we pretend to be a normal task:
9640 current->sched_class = &fair_sched_class;
9642 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9643 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9646 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9647 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9649 /* May be allocated at isolcpus cmdline parse time */
9650 if (cpu_isolated_map == NULL)
9651 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9656 scheduler_running = 1;
9659 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9660 static inline int preempt_count_equals(int preempt_offset)
9662 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9664 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9667 void __might_sleep(char *file, int line, int preempt_offset)
9670 static unsigned long prev_jiffy; /* ratelimiting */
9672 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9673 system_state != SYSTEM_RUNNING || oops_in_progress)
9675 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9677 prev_jiffy = jiffies;
9679 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9681 pr_err("in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9682 in_atomic(), irqs_disabled(),
9683 current->pid, current->comm);
9685 debug_show_held_locks(current);
9686 if (irqs_disabled())
9687 print_irqtrace_events(current);
9691 EXPORT_SYMBOL(__might_sleep);
9694 #ifdef CONFIG_MAGIC_SYSRQ
9695 static void normalize_task(struct rq *rq, struct task_struct *p)
9699 update_rq_clock(rq);
9700 on_rq = p->se.on_rq;
9702 deactivate_task(rq, p, 0);
9703 __setscheduler(rq, p, SCHED_NORMAL, 0);
9705 activate_task(rq, p, 0);
9706 resched_task(rq->curr);
9710 void normalize_rt_tasks(void)
9712 struct task_struct *g, *p;
9713 unsigned long flags;
9716 read_lock_irqsave(&tasklist_lock, flags);
9717 do_each_thread(g, p) {
9719 * Only normalize user tasks:
9724 p->se.exec_start = 0;
9725 #ifdef CONFIG_SCHEDSTATS
9726 p->se.wait_start = 0;
9727 p->se.sleep_start = 0;
9728 p->se.block_start = 0;
9733 * Renice negative nice level userspace
9736 if (TASK_NICE(p) < 0 && p->mm)
9737 set_user_nice(p, 0);
9741 spin_lock(&p->pi_lock);
9742 rq = __task_rq_lock(p);
9744 normalize_task(rq, p);
9746 __task_rq_unlock(rq);
9747 spin_unlock(&p->pi_lock);
9748 } while_each_thread(g, p);
9750 read_unlock_irqrestore(&tasklist_lock, flags);
9753 #endif /* CONFIG_MAGIC_SYSRQ */
9757 * These functions are only useful for the IA64 MCA handling.
9759 * They can only be called when the whole system has been
9760 * stopped - every CPU needs to be quiescent, and no scheduling
9761 * activity can take place. Using them for anything else would
9762 * be a serious bug, and as a result, they aren't even visible
9763 * under any other configuration.
9767 * curr_task - return the current task for a given cpu.
9768 * @cpu: the processor in question.
9770 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9772 struct task_struct *curr_task(int cpu)
9774 return cpu_curr(cpu);
9778 * set_curr_task - set the current task for a given cpu.
9779 * @cpu: the processor in question.
9780 * @p: the task pointer to set.
9782 * Description: This function must only be used when non-maskable interrupts
9783 * are serviced on a separate stack. It allows the architecture to switch the
9784 * notion of the current task on a cpu in a non-blocking manner. This function
9785 * must be called with all CPU's synchronized, and interrupts disabled, the
9786 * and caller must save the original value of the current task (see
9787 * curr_task() above) and restore that value before reenabling interrupts and
9788 * re-starting the system.
9790 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9792 void set_curr_task(int cpu, struct task_struct *p)
9799 #ifdef CONFIG_FAIR_GROUP_SCHED
9800 static void free_fair_sched_group(struct task_group *tg)
9804 for_each_possible_cpu(i) {
9806 kfree(tg->cfs_rq[i]);
9816 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9818 struct cfs_rq *cfs_rq;
9819 struct sched_entity *se;
9823 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9826 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9830 tg->shares = NICE_0_LOAD;
9832 for_each_possible_cpu(i) {
9835 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9836 GFP_KERNEL, cpu_to_node(i));
9840 se = kzalloc_node(sizeof(struct sched_entity),
9841 GFP_KERNEL, cpu_to_node(i));
9845 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9856 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9858 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9859 &cpu_rq(cpu)->leaf_cfs_rq_list);
9862 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9864 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9866 #else /* !CONFG_FAIR_GROUP_SCHED */
9867 static inline void free_fair_sched_group(struct task_group *tg)
9872 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9877 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9881 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9884 #endif /* CONFIG_FAIR_GROUP_SCHED */
9886 #ifdef CONFIG_RT_GROUP_SCHED
9887 static void free_rt_sched_group(struct task_group *tg)
9891 destroy_rt_bandwidth(&tg->rt_bandwidth);
9893 for_each_possible_cpu(i) {
9895 kfree(tg->rt_rq[i]);
9897 kfree(tg->rt_se[i]);
9905 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9907 struct rt_rq *rt_rq;
9908 struct sched_rt_entity *rt_se;
9912 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9915 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9919 init_rt_bandwidth(&tg->rt_bandwidth,
9920 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9922 for_each_possible_cpu(i) {
9925 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9926 GFP_KERNEL, cpu_to_node(i));
9930 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9931 GFP_KERNEL, cpu_to_node(i));
9935 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9946 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9948 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9949 &cpu_rq(cpu)->leaf_rt_rq_list);
9952 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9954 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9956 #else /* !CONFIG_RT_GROUP_SCHED */
9957 static inline void free_rt_sched_group(struct task_group *tg)
9962 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9967 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9971 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9974 #endif /* CONFIG_RT_GROUP_SCHED */
9976 #ifdef CONFIG_GROUP_SCHED
9977 static void free_sched_group(struct task_group *tg)
9979 free_fair_sched_group(tg);
9980 free_rt_sched_group(tg);
9984 /* allocate runqueue etc for a new task group */
9985 struct task_group *sched_create_group(struct task_group *parent)
9987 struct task_group *tg;
9988 unsigned long flags;
9991 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9993 return ERR_PTR(-ENOMEM);
9995 if (!alloc_fair_sched_group(tg, parent))
9998 if (!alloc_rt_sched_group(tg, parent))
10001 spin_lock_irqsave(&task_group_lock, flags);
10002 for_each_possible_cpu(i) {
10003 register_fair_sched_group(tg, i);
10004 register_rt_sched_group(tg, i);
10006 list_add_rcu(&tg->list, &task_groups);
10008 WARN_ON(!parent); /* root should already exist */
10010 tg->parent = parent;
10011 INIT_LIST_HEAD(&tg->children);
10012 list_add_rcu(&tg->siblings, &parent->children);
10013 spin_unlock_irqrestore(&task_group_lock, flags);
10018 free_sched_group(tg);
10019 return ERR_PTR(-ENOMEM);
10022 /* rcu callback to free various structures associated with a task group */
10023 static void free_sched_group_rcu(struct rcu_head *rhp)
10025 /* now it should be safe to free those cfs_rqs */
10026 free_sched_group(container_of(rhp, struct task_group, rcu));
10029 /* Destroy runqueue etc associated with a task group */
10030 void sched_destroy_group(struct task_group *tg)
10032 unsigned long flags;
10035 spin_lock_irqsave(&task_group_lock, flags);
10036 for_each_possible_cpu(i) {
10037 unregister_fair_sched_group(tg, i);
10038 unregister_rt_sched_group(tg, i);
10040 list_del_rcu(&tg->list);
10041 list_del_rcu(&tg->siblings);
10042 spin_unlock_irqrestore(&task_group_lock, flags);
10044 /* wait for possible concurrent references to cfs_rqs complete */
10045 call_rcu(&tg->rcu, free_sched_group_rcu);
10048 /* change task's runqueue when it moves between groups.
10049 * The caller of this function should have put the task in its new group
10050 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10051 * reflect its new group.
10053 void sched_move_task(struct task_struct *tsk)
10055 int on_rq, running;
10056 unsigned long flags;
10059 rq = task_rq_lock(tsk, &flags);
10061 update_rq_clock(rq);
10063 running = task_current(rq, tsk);
10064 on_rq = tsk->se.on_rq;
10067 dequeue_task(rq, tsk, 0);
10068 if (unlikely(running))
10069 tsk->sched_class->put_prev_task(rq, tsk);
10071 set_task_rq(tsk, task_cpu(tsk));
10073 #ifdef CONFIG_FAIR_GROUP_SCHED
10074 if (tsk->sched_class->moved_group)
10075 tsk->sched_class->moved_group(tsk);
10078 if (unlikely(running))
10079 tsk->sched_class->set_curr_task(rq);
10081 enqueue_task(rq, tsk, 0);
10083 task_rq_unlock(rq, &flags);
10085 #endif /* CONFIG_GROUP_SCHED */
10087 #ifdef CONFIG_FAIR_GROUP_SCHED
10088 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10090 struct cfs_rq *cfs_rq = se->cfs_rq;
10095 dequeue_entity(cfs_rq, se, 0);
10097 se->load.weight = shares;
10098 se->load.inv_weight = 0;
10101 enqueue_entity(cfs_rq, se, 0);
10104 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10106 struct cfs_rq *cfs_rq = se->cfs_rq;
10107 struct rq *rq = cfs_rq->rq;
10108 unsigned long flags;
10110 spin_lock_irqsave(&rq->lock, flags);
10111 __set_se_shares(se, shares);
10112 spin_unlock_irqrestore(&rq->lock, flags);
10115 static DEFINE_MUTEX(shares_mutex);
10117 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10120 unsigned long flags;
10123 * We can't change the weight of the root cgroup.
10128 if (shares < MIN_SHARES)
10129 shares = MIN_SHARES;
10130 else if (shares > MAX_SHARES)
10131 shares = MAX_SHARES;
10133 mutex_lock(&shares_mutex);
10134 if (tg->shares == shares)
10137 spin_lock_irqsave(&task_group_lock, flags);
10138 for_each_possible_cpu(i)
10139 unregister_fair_sched_group(tg, i);
10140 list_del_rcu(&tg->siblings);
10141 spin_unlock_irqrestore(&task_group_lock, flags);
10143 /* wait for any ongoing reference to this group to finish */
10144 synchronize_sched();
10147 * Now we are free to modify the group's share on each cpu
10148 * w/o tripping rebalance_share or load_balance_fair.
10150 tg->shares = shares;
10151 for_each_possible_cpu(i) {
10153 * force a rebalance
10155 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10156 set_se_shares(tg->se[i], shares);
10160 * Enable load balance activity on this group, by inserting it back on
10161 * each cpu's rq->leaf_cfs_rq_list.
10163 spin_lock_irqsave(&task_group_lock, flags);
10164 for_each_possible_cpu(i)
10165 register_fair_sched_group(tg, i);
10166 list_add_rcu(&tg->siblings, &tg->parent->children);
10167 spin_unlock_irqrestore(&task_group_lock, flags);
10169 mutex_unlock(&shares_mutex);
10173 unsigned long sched_group_shares(struct task_group *tg)
10179 #ifdef CONFIG_RT_GROUP_SCHED
10181 * Ensure that the real time constraints are schedulable.
10183 static DEFINE_MUTEX(rt_constraints_mutex);
10185 static unsigned long to_ratio(u64 period, u64 runtime)
10187 if (runtime == RUNTIME_INF)
10190 return div64_u64(runtime << 20, period);
10193 /* Must be called with tasklist_lock held */
10194 static inline int tg_has_rt_tasks(struct task_group *tg)
10196 struct task_struct *g, *p;
10198 do_each_thread(g, p) {
10199 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10201 } while_each_thread(g, p);
10206 struct rt_schedulable_data {
10207 struct task_group *tg;
10212 static int tg_schedulable(struct task_group *tg, void *data)
10214 struct rt_schedulable_data *d = data;
10215 struct task_group *child;
10216 unsigned long total, sum = 0;
10217 u64 period, runtime;
10219 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10220 runtime = tg->rt_bandwidth.rt_runtime;
10223 period = d->rt_period;
10224 runtime = d->rt_runtime;
10227 #ifdef CONFIG_USER_SCHED
10228 if (tg == &root_task_group) {
10229 period = global_rt_period();
10230 runtime = global_rt_runtime();
10235 * Cannot have more runtime than the period.
10237 if (runtime > period && runtime != RUNTIME_INF)
10241 * Ensure we don't starve existing RT tasks.
10243 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10246 total = to_ratio(period, runtime);
10249 * Nobody can have more than the global setting allows.
10251 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10255 * The sum of our children's runtime should not exceed our own.
10257 list_for_each_entry_rcu(child, &tg->children, siblings) {
10258 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10259 runtime = child->rt_bandwidth.rt_runtime;
10261 if (child == d->tg) {
10262 period = d->rt_period;
10263 runtime = d->rt_runtime;
10266 sum += to_ratio(period, runtime);
10275 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10277 struct rt_schedulable_data data = {
10279 .rt_period = period,
10280 .rt_runtime = runtime,
10283 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10286 static int tg_set_bandwidth(struct task_group *tg,
10287 u64 rt_period, u64 rt_runtime)
10291 mutex_lock(&rt_constraints_mutex);
10292 read_lock(&tasklist_lock);
10293 err = __rt_schedulable(tg, rt_period, rt_runtime);
10297 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10298 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10299 tg->rt_bandwidth.rt_runtime = rt_runtime;
10301 for_each_possible_cpu(i) {
10302 struct rt_rq *rt_rq = tg->rt_rq[i];
10304 spin_lock(&rt_rq->rt_runtime_lock);
10305 rt_rq->rt_runtime = rt_runtime;
10306 spin_unlock(&rt_rq->rt_runtime_lock);
10308 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10310 read_unlock(&tasklist_lock);
10311 mutex_unlock(&rt_constraints_mutex);
10316 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10318 u64 rt_runtime, rt_period;
10320 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10321 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10322 if (rt_runtime_us < 0)
10323 rt_runtime = RUNTIME_INF;
10325 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10328 long sched_group_rt_runtime(struct task_group *tg)
10332 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10335 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10336 do_div(rt_runtime_us, NSEC_PER_USEC);
10337 return rt_runtime_us;
10340 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10342 u64 rt_runtime, rt_period;
10344 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10345 rt_runtime = tg->rt_bandwidth.rt_runtime;
10347 if (rt_period == 0)
10350 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10353 long sched_group_rt_period(struct task_group *tg)
10357 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10358 do_div(rt_period_us, NSEC_PER_USEC);
10359 return rt_period_us;
10362 static int sched_rt_global_constraints(void)
10364 u64 runtime, period;
10367 if (sysctl_sched_rt_period <= 0)
10370 runtime = global_rt_runtime();
10371 period = global_rt_period();
10374 * Sanity check on the sysctl variables.
10376 if (runtime > period && runtime != RUNTIME_INF)
10379 mutex_lock(&rt_constraints_mutex);
10380 read_lock(&tasklist_lock);
10381 ret = __rt_schedulable(NULL, 0, 0);
10382 read_unlock(&tasklist_lock);
10383 mutex_unlock(&rt_constraints_mutex);
10388 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10390 /* Don't accept realtime tasks when there is no way for them to run */
10391 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10397 #else /* !CONFIG_RT_GROUP_SCHED */
10398 static int sched_rt_global_constraints(void)
10400 unsigned long flags;
10403 if (sysctl_sched_rt_period <= 0)
10407 * There's always some RT tasks in the root group
10408 * -- migration, kstopmachine etc..
10410 if (sysctl_sched_rt_runtime == 0)
10413 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10414 for_each_possible_cpu(i) {
10415 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10417 spin_lock(&rt_rq->rt_runtime_lock);
10418 rt_rq->rt_runtime = global_rt_runtime();
10419 spin_unlock(&rt_rq->rt_runtime_lock);
10421 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10425 #endif /* CONFIG_RT_GROUP_SCHED */
10427 int sched_rt_handler(struct ctl_table *table, int write,
10428 void __user *buffer, size_t *lenp,
10432 int old_period, old_runtime;
10433 static DEFINE_MUTEX(mutex);
10435 mutex_lock(&mutex);
10436 old_period = sysctl_sched_rt_period;
10437 old_runtime = sysctl_sched_rt_runtime;
10439 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10441 if (!ret && write) {
10442 ret = sched_rt_global_constraints();
10444 sysctl_sched_rt_period = old_period;
10445 sysctl_sched_rt_runtime = old_runtime;
10447 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10448 def_rt_bandwidth.rt_period =
10449 ns_to_ktime(global_rt_period());
10452 mutex_unlock(&mutex);
10457 #ifdef CONFIG_CGROUP_SCHED
10459 /* return corresponding task_group object of a cgroup */
10460 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10462 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10463 struct task_group, css);
10466 static struct cgroup_subsys_state *
10467 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10469 struct task_group *tg, *parent;
10471 if (!cgrp->parent) {
10472 /* This is early initialization for the top cgroup */
10473 return &init_task_group.css;
10476 parent = cgroup_tg(cgrp->parent);
10477 tg = sched_create_group(parent);
10479 return ERR_PTR(-ENOMEM);
10485 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10487 struct task_group *tg = cgroup_tg(cgrp);
10489 sched_destroy_group(tg);
10493 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10495 #ifdef CONFIG_RT_GROUP_SCHED
10496 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10499 /* We don't support RT-tasks being in separate groups */
10500 if (tsk->sched_class != &fair_sched_class)
10507 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10508 struct task_struct *tsk, bool threadgroup)
10510 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10514 struct task_struct *c;
10516 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10517 retval = cpu_cgroup_can_attach_task(cgrp, c);
10529 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10530 struct cgroup *old_cont, struct task_struct *tsk,
10533 sched_move_task(tsk);
10535 struct task_struct *c;
10537 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10538 sched_move_task(c);
10544 #ifdef CONFIG_FAIR_GROUP_SCHED
10545 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10548 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10551 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10553 struct task_group *tg = cgroup_tg(cgrp);
10555 return (u64) tg->shares;
10557 #endif /* CONFIG_FAIR_GROUP_SCHED */
10559 #ifdef CONFIG_RT_GROUP_SCHED
10560 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10563 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10566 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10568 return sched_group_rt_runtime(cgroup_tg(cgrp));
10571 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10574 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10577 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10579 return sched_group_rt_period(cgroup_tg(cgrp));
10581 #endif /* CONFIG_RT_GROUP_SCHED */
10583 static struct cftype cpu_files[] = {
10584 #ifdef CONFIG_FAIR_GROUP_SCHED
10587 .read_u64 = cpu_shares_read_u64,
10588 .write_u64 = cpu_shares_write_u64,
10591 #ifdef CONFIG_RT_GROUP_SCHED
10593 .name = "rt_runtime_us",
10594 .read_s64 = cpu_rt_runtime_read,
10595 .write_s64 = cpu_rt_runtime_write,
10598 .name = "rt_period_us",
10599 .read_u64 = cpu_rt_period_read_uint,
10600 .write_u64 = cpu_rt_period_write_uint,
10605 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10607 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10610 struct cgroup_subsys cpu_cgroup_subsys = {
10612 .create = cpu_cgroup_create,
10613 .destroy = cpu_cgroup_destroy,
10614 .can_attach = cpu_cgroup_can_attach,
10615 .attach = cpu_cgroup_attach,
10616 .populate = cpu_cgroup_populate,
10617 .subsys_id = cpu_cgroup_subsys_id,
10621 #endif /* CONFIG_CGROUP_SCHED */
10623 #ifdef CONFIG_CGROUP_CPUACCT
10626 * CPU accounting code for task groups.
10628 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10629 * (balbir@in.ibm.com).
10632 /* track cpu usage of a group of tasks and its child groups */
10634 struct cgroup_subsys_state css;
10635 /* cpuusage holds pointer to a u64-type object on every cpu */
10637 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10638 struct cpuacct *parent;
10641 struct cgroup_subsys cpuacct_subsys;
10643 /* return cpu accounting group corresponding to this container */
10644 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10646 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10647 struct cpuacct, css);
10650 /* return cpu accounting group to which this task belongs */
10651 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10653 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10654 struct cpuacct, css);
10657 /* create a new cpu accounting group */
10658 static struct cgroup_subsys_state *cpuacct_create(
10659 struct cgroup_subsys *ss, struct cgroup *cgrp)
10661 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10667 ca->cpuusage = alloc_percpu(u64);
10671 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10672 if (percpu_counter_init(&ca->cpustat[i], 0))
10673 goto out_free_counters;
10676 ca->parent = cgroup_ca(cgrp->parent);
10682 percpu_counter_destroy(&ca->cpustat[i]);
10683 free_percpu(ca->cpuusage);
10687 return ERR_PTR(-ENOMEM);
10690 /* destroy an existing cpu accounting group */
10692 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10694 struct cpuacct *ca = cgroup_ca(cgrp);
10697 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10698 percpu_counter_destroy(&ca->cpustat[i]);
10699 free_percpu(ca->cpuusage);
10703 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10705 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10708 #ifndef CONFIG_64BIT
10710 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10712 spin_lock_irq(&cpu_rq(cpu)->lock);
10714 spin_unlock_irq(&cpu_rq(cpu)->lock);
10722 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10724 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10726 #ifndef CONFIG_64BIT
10728 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10730 spin_lock_irq(&cpu_rq(cpu)->lock);
10732 spin_unlock_irq(&cpu_rq(cpu)->lock);
10738 /* return total cpu usage (in nanoseconds) of a group */
10739 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10741 struct cpuacct *ca = cgroup_ca(cgrp);
10742 u64 totalcpuusage = 0;
10745 for_each_present_cpu(i)
10746 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10748 return totalcpuusage;
10751 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10754 struct cpuacct *ca = cgroup_ca(cgrp);
10763 for_each_present_cpu(i)
10764 cpuacct_cpuusage_write(ca, i, 0);
10770 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10771 struct seq_file *m)
10773 struct cpuacct *ca = cgroup_ca(cgroup);
10777 for_each_present_cpu(i) {
10778 percpu = cpuacct_cpuusage_read(ca, i);
10779 seq_printf(m, "%llu ", (unsigned long long) percpu);
10781 seq_printf(m, "\n");
10785 static const char *cpuacct_stat_desc[] = {
10786 [CPUACCT_STAT_USER] = "user",
10787 [CPUACCT_STAT_SYSTEM] = "system",
10790 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10791 struct cgroup_map_cb *cb)
10793 struct cpuacct *ca = cgroup_ca(cgrp);
10796 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10797 s64 val = percpu_counter_read(&ca->cpustat[i]);
10798 val = cputime64_to_clock_t(val);
10799 cb->fill(cb, cpuacct_stat_desc[i], val);
10804 static struct cftype files[] = {
10807 .read_u64 = cpuusage_read,
10808 .write_u64 = cpuusage_write,
10811 .name = "usage_percpu",
10812 .read_seq_string = cpuacct_percpu_seq_read,
10816 .read_map = cpuacct_stats_show,
10820 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10822 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10826 * charge this task's execution time to its accounting group.
10828 * called with rq->lock held.
10830 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10832 struct cpuacct *ca;
10835 if (unlikely(!cpuacct_subsys.active))
10838 cpu = task_cpu(tsk);
10844 for (; ca; ca = ca->parent) {
10845 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10846 *cpuusage += cputime;
10853 * Charge the system/user time to the task's accounting group.
10855 static void cpuacct_update_stats(struct task_struct *tsk,
10856 enum cpuacct_stat_index idx, cputime_t val)
10858 struct cpuacct *ca;
10860 if (unlikely(!cpuacct_subsys.active))
10867 percpu_counter_add(&ca->cpustat[idx], val);
10873 struct cgroup_subsys cpuacct_subsys = {
10875 .create = cpuacct_create,
10876 .destroy = cpuacct_destroy,
10877 .populate = cpuacct_populate,
10878 .subsys_id = cpuacct_subsys_id,
10880 #endif /* CONFIG_CGROUP_CPUACCT */
10884 int rcu_expedited_torture_stats(char *page)
10888 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10890 void synchronize_sched_expedited(void)
10893 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10895 #else /* #ifndef CONFIG_SMP */
10897 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10898 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10900 #define RCU_EXPEDITED_STATE_POST -2
10901 #define RCU_EXPEDITED_STATE_IDLE -1
10903 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10905 int rcu_expedited_torture_stats(char *page)
10910 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10911 for_each_online_cpu(cpu) {
10912 cnt += sprintf(&page[cnt], " %d:%d",
10913 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10915 cnt += sprintf(&page[cnt], "\n");
10918 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10920 static long synchronize_sched_expedited_count;
10923 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10924 * approach to force grace period to end quickly. This consumes
10925 * significant time on all CPUs, and is thus not recommended for
10926 * any sort of common-case code.
10928 * Note that it is illegal to call this function while holding any
10929 * lock that is acquired by a CPU-hotplug notifier. Failing to
10930 * observe this restriction will result in deadlock.
10932 void synchronize_sched_expedited(void)
10935 unsigned long flags;
10936 bool need_full_sync = 0;
10938 struct migration_req *req;
10942 smp_mb(); /* ensure prior mod happens before capturing snap. */
10943 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10945 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10947 if (trycount++ < 10)
10948 udelay(trycount * num_online_cpus());
10950 synchronize_sched();
10953 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10954 smp_mb(); /* ensure test happens before caller kfree */
10959 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10960 for_each_online_cpu(cpu) {
10962 req = &per_cpu(rcu_migration_req, cpu);
10963 init_completion(&req->done);
10965 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10966 spin_lock_irqsave(&rq->lock, flags);
10967 list_add(&req->list, &rq->migration_queue);
10968 spin_unlock_irqrestore(&rq->lock, flags);
10969 wake_up_process(rq->migration_thread);
10971 for_each_online_cpu(cpu) {
10972 rcu_expedited_state = cpu;
10973 req = &per_cpu(rcu_migration_req, cpu);
10975 wait_for_completion(&req->done);
10976 spin_lock_irqsave(&rq->lock, flags);
10977 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10978 need_full_sync = 1;
10979 req->dest_cpu = RCU_MIGRATION_IDLE;
10980 spin_unlock_irqrestore(&rq->lock, flags);
10982 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10983 synchronize_sched_expedited_count++;
10984 mutex_unlock(&rcu_sched_expedited_mutex);
10986 if (need_full_sync)
10987 synchronize_sched();
10989 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10991 #endif /* #else #ifndef CONFIG_SMP */