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 raw_spinlock_t rt_runtime_lock;
149 struct hrtimer rt_period_timer;
152 static struct rt_bandwidth def_rt_bandwidth;
154 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
156 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
158 struct rt_bandwidth *rt_b =
159 container_of(timer, struct rt_bandwidth, rt_period_timer);
165 now = hrtimer_cb_get_time(timer);
166 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
171 idle = do_sched_rt_period_timer(rt_b, overrun);
174 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
180 rt_b->rt_period = ns_to_ktime(period);
181 rt_b->rt_runtime = runtime;
183 raw_spin_lock_init(&rt_b->rt_runtime_lock);
185 hrtimer_init(&rt_b->rt_period_timer,
186 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
187 rt_b->rt_period_timer.function = sched_rt_period_timer;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime >= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
199 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
202 if (hrtimer_active(&rt_b->rt_period_timer))
205 raw_spin_lock(&rt_b->rt_runtime_lock);
210 if (hrtimer_active(&rt_b->rt_period_timer))
213 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
214 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
216 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
217 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
218 delta = ktime_to_ns(ktime_sub(hard, soft));
219 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
220 HRTIMER_MODE_ABS_PINNED, 0);
222 raw_spin_unlock(&rt_b->rt_runtime_lock);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
228 hrtimer_cancel(&rt_b->rt_period_timer);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex);
238 #ifdef CONFIG_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_var);
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 raw_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 raw_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 raw_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 raw_spin_unlock_irq(&rq->lock);
924 raw_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 raw_spin_lock(&rq->lock);
955 if (likely(rq == task_rq(p)))
957 raw_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 raw_spin_lock(&rq->lock);
975 if (likely(rq == task_rq(p)))
977 raw_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 raw_spin_unlock_wait(&rq->lock);
989 static void __task_rq_unlock(struct rq *rq)
992 raw_spin_unlock(&rq->lock);
995 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
998 raw_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 raw_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 raw_spin_lock(&rq->lock);
1059 update_rq_clock(rq);
1060 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1061 raw_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 raw_spin_lock(&rq->lock);
1075 hrtimer_restart(&rq->hrtick_timer);
1076 rq->hrtick_csd_pending = 0;
1077 raw_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_raw_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 (!raw_spin_trylock_irqsave(&rq->lock, flags))
1208 resched_task(cpu_curr(cpu));
1209 raw_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_raw_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 raw_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 raw_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 raw_spin_unlock(&rq->lock);
1713 raw_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 raw_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(!raw_spin_trylock(&busiest->lock))) {
1775 if (busiest < this_rq) {
1776 raw_spin_unlock(&this_rq->lock);
1777 raw_spin_lock(&busiest->lock);
1778 raw_spin_lock_nested(&this_rq->lock,
1779 SINGLE_DEPTH_NESTING);
1782 raw_spin_lock_nested(&busiest->lock,
1783 SINGLE_DEPTH_NESTING);
1788 #endif /* CONFIG_PREEMPT */
1791 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1793 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1795 if (unlikely(!irqs_disabled())) {
1796 /* printk() doesn't work good under rq->lock */
1797 raw_spin_unlock(&this_rq->lock);
1801 return _double_lock_balance(this_rq, busiest);
1804 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1805 __releases(busiest->lock)
1807 raw_spin_unlock(&busiest->lock);
1808 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1812 #ifdef CONFIG_FAIR_GROUP_SCHED
1813 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1816 cfs_rq->shares = shares;
1821 static void calc_load_account_active(struct rq *this_rq);
1822 static void update_sysctl(void);
1823 static int get_update_sysctl_factor(void);
1825 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1827 set_task_rq(p, cpu);
1830 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1831 * successfuly executed on another CPU. We must ensure that updates of
1832 * per-task data have been completed by this moment.
1835 task_thread_info(p)->cpu = cpu;
1839 #include "sched_stats.h"
1840 #include "sched_idletask.c"
1841 #include "sched_fair.c"
1842 #include "sched_rt.c"
1843 #ifdef CONFIG_SCHED_DEBUG
1844 # include "sched_debug.c"
1847 #define sched_class_highest (&rt_sched_class)
1848 #define for_each_class(class) \
1849 for (class = sched_class_highest; class; class = class->next)
1851 static void inc_nr_running(struct rq *rq)
1856 static void dec_nr_running(struct rq *rq)
1861 static void set_load_weight(struct task_struct *p)
1863 if (task_has_rt_policy(p)) {
1864 p->se.load.weight = prio_to_weight[0] * 2;
1865 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1870 * SCHED_IDLE tasks get minimal weight:
1872 if (p->policy == SCHED_IDLE) {
1873 p->se.load.weight = WEIGHT_IDLEPRIO;
1874 p->se.load.inv_weight = WMULT_IDLEPRIO;
1878 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1879 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1882 static void update_avg(u64 *avg, u64 sample)
1884 s64 diff = sample - *avg;
1888 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1891 p->se.start_runtime = p->se.sum_exec_runtime;
1893 sched_info_queued(p);
1894 p->sched_class->enqueue_task(rq, p, wakeup);
1898 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1901 if (p->se.last_wakeup) {
1902 update_avg(&p->se.avg_overlap,
1903 p->se.sum_exec_runtime - p->se.last_wakeup);
1904 p->se.last_wakeup = 0;
1906 update_avg(&p->se.avg_wakeup,
1907 sysctl_sched_wakeup_granularity);
1911 sched_info_dequeued(p);
1912 p->sched_class->dequeue_task(rq, p, sleep);
1917 * __normal_prio - return the priority that is based on the static prio
1919 static inline int __normal_prio(struct task_struct *p)
1921 return p->static_prio;
1925 * Calculate the expected normal priority: i.e. priority
1926 * without taking RT-inheritance into account. Might be
1927 * boosted by interactivity modifiers. Changes upon fork,
1928 * setprio syscalls, and whenever the interactivity
1929 * estimator recalculates.
1931 static inline int normal_prio(struct task_struct *p)
1935 if (task_has_rt_policy(p))
1936 prio = MAX_RT_PRIO-1 - p->rt_priority;
1938 prio = __normal_prio(p);
1943 * Calculate the current priority, i.e. the priority
1944 * taken into account by the scheduler. This value might
1945 * be boosted by RT tasks, or might be boosted by
1946 * interactivity modifiers. Will be RT if the task got
1947 * RT-boosted. If not then it returns p->normal_prio.
1949 static int effective_prio(struct task_struct *p)
1951 p->normal_prio = normal_prio(p);
1953 * If we are RT tasks or we were boosted to RT priority,
1954 * keep the priority unchanged. Otherwise, update priority
1955 * to the normal priority:
1957 if (!rt_prio(p->prio))
1958 return p->normal_prio;
1963 * activate_task - move a task to the runqueue.
1965 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1967 if (task_contributes_to_load(p))
1968 rq->nr_uninterruptible--;
1970 enqueue_task(rq, p, wakeup);
1975 * deactivate_task - remove a task from the runqueue.
1977 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1979 if (task_contributes_to_load(p))
1980 rq->nr_uninterruptible++;
1982 dequeue_task(rq, p, sleep);
1987 * task_curr - is this task currently executing on a CPU?
1988 * @p: the task in question.
1990 inline int task_curr(const struct task_struct *p)
1992 return cpu_curr(task_cpu(p)) == p;
1995 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1996 const struct sched_class *prev_class,
1997 int oldprio, int running)
1999 if (prev_class != p->sched_class) {
2000 if (prev_class->switched_from)
2001 prev_class->switched_from(rq, p, running);
2002 p->sched_class->switched_to(rq, p, running);
2004 p->sched_class->prio_changed(rq, p, oldprio, running);
2009 * Is this task likely cache-hot:
2012 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2016 if (p->sched_class != &fair_sched_class)
2020 * Buddy candidates are cache hot:
2022 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2023 (&p->se == cfs_rq_of(&p->se)->next ||
2024 &p->se == cfs_rq_of(&p->se)->last))
2027 if (sysctl_sched_migration_cost == -1)
2029 if (sysctl_sched_migration_cost == 0)
2032 delta = now - p->se.exec_start;
2034 return delta < (s64)sysctl_sched_migration_cost;
2038 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2040 int old_cpu = task_cpu(p);
2041 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2042 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2044 #ifdef CONFIG_SCHED_DEBUG
2046 * We should never call set_task_cpu() on a blocked task,
2047 * ttwu() will sort out the placement.
2049 WARN_ON(p->state != TASK_RUNNING && p->state != TASK_WAKING);
2052 trace_sched_migrate_task(p, new_cpu);
2054 if (old_cpu != new_cpu) {
2055 p->se.nr_migrations++;
2056 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2059 p->se.vruntime -= old_cfsrq->min_vruntime -
2060 new_cfsrq->min_vruntime;
2062 __set_task_cpu(p, new_cpu);
2065 struct migration_req {
2066 struct list_head list;
2068 struct task_struct *task;
2071 struct completion done;
2075 * The task's runqueue lock must be held.
2076 * Returns true if you have to wait for migration thread.
2079 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2081 struct rq *rq = task_rq(p);
2084 * If the task is not on a runqueue (and not running), then
2085 * the next wake-up will properly place the task.
2087 if (!p->se.on_rq && !task_running(rq, p))
2090 init_completion(&req->done);
2092 req->dest_cpu = dest_cpu;
2093 list_add(&req->list, &rq->migration_queue);
2099 * wait_task_context_switch - wait for a thread to complete at least one
2102 * @p must not be current.
2104 void wait_task_context_switch(struct task_struct *p)
2106 unsigned long nvcsw, nivcsw, flags;
2114 * The runqueue is assigned before the actual context
2115 * switch. We need to take the runqueue lock.
2117 * We could check initially without the lock but it is
2118 * very likely that we need to take the lock in every
2121 rq = task_rq_lock(p, &flags);
2122 running = task_running(rq, p);
2123 task_rq_unlock(rq, &flags);
2125 if (likely(!running))
2128 * The switch count is incremented before the actual
2129 * context switch. We thus wait for two switches to be
2130 * sure at least one completed.
2132 if ((p->nvcsw - nvcsw) > 1)
2134 if ((p->nivcsw - nivcsw) > 1)
2142 * wait_task_inactive - wait for a thread to unschedule.
2144 * If @match_state is nonzero, it's the @p->state value just checked and
2145 * not expected to change. If it changes, i.e. @p might have woken up,
2146 * then return zero. When we succeed in waiting for @p to be off its CPU,
2147 * we return a positive number (its total switch count). If a second call
2148 * a short while later returns the same number, the caller can be sure that
2149 * @p has remained unscheduled the whole time.
2151 * The caller must ensure that the task *will* unschedule sometime soon,
2152 * else this function might spin for a *long* time. This function can't
2153 * be called with interrupts off, or it may introduce deadlock with
2154 * smp_call_function() if an IPI is sent by the same process we are
2155 * waiting to become inactive.
2157 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2159 unsigned long flags;
2166 * We do the initial early heuristics without holding
2167 * any task-queue locks at all. We'll only try to get
2168 * the runqueue lock when things look like they will
2174 * If the task is actively running on another CPU
2175 * still, just relax and busy-wait without holding
2178 * NOTE! Since we don't hold any locks, it's not
2179 * even sure that "rq" stays as the right runqueue!
2180 * But we don't care, since "task_running()" will
2181 * return false if the runqueue has changed and p
2182 * is actually now running somewhere else!
2184 while (task_running(rq, p)) {
2185 if (match_state && unlikely(p->state != match_state))
2191 * Ok, time to look more closely! We need the rq
2192 * lock now, to be *sure*. If we're wrong, we'll
2193 * just go back and repeat.
2195 rq = task_rq_lock(p, &flags);
2196 trace_sched_wait_task(rq, p);
2197 running = task_running(rq, p);
2198 on_rq = p->se.on_rq;
2200 if (!match_state || p->state == match_state)
2201 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2202 task_rq_unlock(rq, &flags);
2205 * If it changed from the expected state, bail out now.
2207 if (unlikely(!ncsw))
2211 * Was it really running after all now that we
2212 * checked with the proper locks actually held?
2214 * Oops. Go back and try again..
2216 if (unlikely(running)) {
2222 * It's not enough that it's not actively running,
2223 * it must be off the runqueue _entirely_, and not
2226 * So if it was still runnable (but just not actively
2227 * running right now), it's preempted, and we should
2228 * yield - it could be a while.
2230 if (unlikely(on_rq)) {
2231 schedule_timeout_uninterruptible(1);
2236 * Ahh, all good. It wasn't running, and it wasn't
2237 * runnable, which means that it will never become
2238 * running in the future either. We're all done!
2247 * kick_process - kick a running thread to enter/exit the kernel
2248 * @p: the to-be-kicked thread
2250 * Cause a process which is running on another CPU to enter
2251 * kernel-mode, without any delay. (to get signals handled.)
2253 * NOTE: this function doesnt have to take the runqueue lock,
2254 * because all it wants to ensure is that the remote task enters
2255 * the kernel. If the IPI races and the task has been migrated
2256 * to another CPU then no harm is done and the purpose has been
2259 void kick_process(struct task_struct *p)
2265 if ((cpu != smp_processor_id()) && task_curr(p))
2266 smp_send_reschedule(cpu);
2269 EXPORT_SYMBOL_GPL(kick_process);
2270 #endif /* CONFIG_SMP */
2273 * task_oncpu_function_call - call a function on the cpu on which a task runs
2274 * @p: the task to evaluate
2275 * @func: the function to be called
2276 * @info: the function call argument
2278 * Calls the function @func when the task is currently running. This might
2279 * be on the current CPU, which just calls the function directly
2281 void task_oncpu_function_call(struct task_struct *p,
2282 void (*func) (void *info), void *info)
2289 smp_call_function_single(cpu, func, info, 1);
2294 static int select_fallback_rq(int cpu, struct task_struct *p)
2297 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(cpu));
2299 /* Look for allowed, online CPU in same node. */
2300 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
2301 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
2304 /* Any allowed, online CPU? */
2305 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
2306 if (dest_cpu < nr_cpu_ids)
2309 /* No more Mr. Nice Guy. */
2310 if (dest_cpu >= nr_cpu_ids) {
2312 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
2314 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
2317 * Don't tell them about moving exiting tasks or
2318 * kernel threads (both mm NULL), since they never
2321 if (p->mm && printk_ratelimit()) {
2322 printk(KERN_INFO "process %d (%s) no "
2323 "longer affine to cpu%d\n",
2324 task_pid_nr(p), p->comm, cpu);
2334 * - fork, @p is stable because it isn't on the tasklist yet
2336 * - exec, @p is unstable, retry loop
2338 * - wake-up, we serialize ->cpus_allowed against TASK_WAKING so
2339 * we should be good.
2342 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2344 int cpu = p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2347 * In order not to call set_task_cpu() on a blocking task we need
2348 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2351 * Since this is common to all placement strategies, this lives here.
2353 * [ this allows ->select_task() to simply return task_cpu(p) and
2354 * not worry about this generic constraint ]
2356 if (unlikely(!cpumask_test_cpu(cpu, &p->cpus_allowed) ||
2358 cpu = select_fallback_rq(task_cpu(p), p);
2365 * try_to_wake_up - wake up a thread
2366 * @p: the to-be-woken-up thread
2367 * @state: the mask of task states that can be woken
2368 * @sync: do a synchronous wakeup?
2370 * Put it on the run-queue if it's not already there. The "current"
2371 * thread is always on the run-queue (except when the actual
2372 * re-schedule is in progress), and as such you're allowed to do
2373 * the simpler "current->state = TASK_RUNNING" to mark yourself
2374 * runnable without the overhead of this.
2376 * returns failure only if the task is already active.
2378 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2381 int cpu, orig_cpu, this_cpu, success = 0;
2382 unsigned long flags;
2383 struct rq *rq, *orig_rq;
2385 if (!sched_feat(SYNC_WAKEUPS))
2386 wake_flags &= ~WF_SYNC;
2388 this_cpu = get_cpu();
2391 rq = orig_rq = task_rq_lock(p, &flags);
2392 update_rq_clock(rq);
2393 if (!(p->state & state))
2403 if (unlikely(task_running(rq, p)))
2407 * In order to handle concurrent wakeups and release the rq->lock
2408 * we put the task in TASK_WAKING state.
2410 * First fix up the nr_uninterruptible count:
2412 if (task_contributes_to_load(p))
2413 rq->nr_uninterruptible--;
2414 p->state = TASK_WAKING;
2415 __task_rq_unlock(rq);
2417 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2418 if (cpu != orig_cpu)
2419 set_task_cpu(p, cpu);
2421 rq = __task_rq_lock(p);
2422 update_rq_clock(rq);
2424 WARN_ON(p->state != TASK_WAKING);
2427 #ifdef CONFIG_SCHEDSTATS
2428 schedstat_inc(rq, ttwu_count);
2429 if (cpu == this_cpu)
2430 schedstat_inc(rq, ttwu_local);
2432 struct sched_domain *sd;
2433 for_each_domain(this_cpu, sd) {
2434 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2435 schedstat_inc(sd, ttwu_wake_remote);
2440 #endif /* CONFIG_SCHEDSTATS */
2443 #endif /* CONFIG_SMP */
2444 schedstat_inc(p, se.nr_wakeups);
2445 if (wake_flags & WF_SYNC)
2446 schedstat_inc(p, se.nr_wakeups_sync);
2447 if (orig_cpu != cpu)
2448 schedstat_inc(p, se.nr_wakeups_migrate);
2449 if (cpu == this_cpu)
2450 schedstat_inc(p, se.nr_wakeups_local);
2452 schedstat_inc(p, se.nr_wakeups_remote);
2453 activate_task(rq, p, 1);
2457 * Only attribute actual wakeups done by this task.
2459 if (!in_interrupt()) {
2460 struct sched_entity *se = ¤t->se;
2461 u64 sample = se->sum_exec_runtime;
2463 if (se->last_wakeup)
2464 sample -= se->last_wakeup;
2466 sample -= se->start_runtime;
2467 update_avg(&se->avg_wakeup, sample);
2469 se->last_wakeup = se->sum_exec_runtime;
2473 trace_sched_wakeup(rq, p, success);
2474 check_preempt_curr(rq, p, wake_flags);
2476 p->state = TASK_RUNNING;
2478 if (p->sched_class->task_wake_up)
2479 p->sched_class->task_wake_up(rq, p);
2481 if (unlikely(rq->idle_stamp)) {
2482 u64 delta = rq->clock - rq->idle_stamp;
2483 u64 max = 2*sysctl_sched_migration_cost;
2488 update_avg(&rq->avg_idle, delta);
2493 task_rq_unlock(rq, &flags);
2500 * wake_up_process - Wake up a specific process
2501 * @p: The process to be woken up.
2503 * Attempt to wake up the nominated process and move it to the set of runnable
2504 * processes. Returns 1 if the process was woken up, 0 if it was already
2507 * It may be assumed that this function implies a write memory barrier before
2508 * changing the task state if and only if any tasks are woken up.
2510 int wake_up_process(struct task_struct *p)
2512 return try_to_wake_up(p, TASK_ALL, 0);
2514 EXPORT_SYMBOL(wake_up_process);
2516 int wake_up_state(struct task_struct *p, unsigned int state)
2518 return try_to_wake_up(p, state, 0);
2522 * Perform scheduler related setup for a newly forked process p.
2523 * p is forked by current.
2525 * __sched_fork() is basic setup used by init_idle() too:
2527 static void __sched_fork(struct task_struct *p)
2529 p->se.exec_start = 0;
2530 p->se.sum_exec_runtime = 0;
2531 p->se.prev_sum_exec_runtime = 0;
2532 p->se.nr_migrations = 0;
2533 p->se.last_wakeup = 0;
2534 p->se.avg_overlap = 0;
2535 p->se.start_runtime = 0;
2536 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2538 #ifdef CONFIG_SCHEDSTATS
2539 p->se.wait_start = 0;
2541 p->se.wait_count = 0;
2544 p->se.sleep_start = 0;
2545 p->se.sleep_max = 0;
2546 p->se.sum_sleep_runtime = 0;
2548 p->se.block_start = 0;
2549 p->se.block_max = 0;
2551 p->se.slice_max = 0;
2553 p->se.nr_migrations_cold = 0;
2554 p->se.nr_failed_migrations_affine = 0;
2555 p->se.nr_failed_migrations_running = 0;
2556 p->se.nr_failed_migrations_hot = 0;
2557 p->se.nr_forced_migrations = 0;
2559 p->se.nr_wakeups = 0;
2560 p->se.nr_wakeups_sync = 0;
2561 p->se.nr_wakeups_migrate = 0;
2562 p->se.nr_wakeups_local = 0;
2563 p->se.nr_wakeups_remote = 0;
2564 p->se.nr_wakeups_affine = 0;
2565 p->se.nr_wakeups_affine_attempts = 0;
2566 p->se.nr_wakeups_passive = 0;
2567 p->se.nr_wakeups_idle = 0;
2571 INIT_LIST_HEAD(&p->rt.run_list);
2573 INIT_LIST_HEAD(&p->se.group_node);
2575 #ifdef CONFIG_PREEMPT_NOTIFIERS
2576 INIT_HLIST_HEAD(&p->preempt_notifiers);
2581 * fork()/clone()-time setup:
2583 void sched_fork(struct task_struct *p, int clone_flags)
2585 int cpu = get_cpu();
2589 * We mark the process as waking here. This guarantees that
2590 * nobody will actually run it, and a signal or other external
2591 * event cannot wake it up and insert it on the runqueue either.
2593 p->state = TASK_WAKING;
2596 * Revert to default priority/policy on fork if requested.
2598 if (unlikely(p->sched_reset_on_fork)) {
2599 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2600 p->policy = SCHED_NORMAL;
2601 p->normal_prio = p->static_prio;
2604 if (PRIO_TO_NICE(p->static_prio) < 0) {
2605 p->static_prio = NICE_TO_PRIO(0);
2606 p->normal_prio = p->static_prio;
2611 * We don't need the reset flag anymore after the fork. It has
2612 * fulfilled its duty:
2614 p->sched_reset_on_fork = 0;
2618 * Make sure we do not leak PI boosting priority to the child.
2620 p->prio = current->normal_prio;
2622 if (!rt_prio(p->prio))
2623 p->sched_class = &fair_sched_class;
2625 if (p->sched_class->task_fork)
2626 p->sched_class->task_fork(p);
2629 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2631 set_task_cpu(p, cpu);
2633 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2634 if (likely(sched_info_on()))
2635 memset(&p->sched_info, 0, sizeof(p->sched_info));
2637 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2640 #ifdef CONFIG_PREEMPT
2641 /* Want to start with kernel preemption disabled. */
2642 task_thread_info(p)->preempt_count = 1;
2644 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2650 * wake_up_new_task - wake up a newly created task for the first time.
2652 * This function will do some initial scheduler statistics housekeeping
2653 * that must be done for every newly created context, then puts the task
2654 * on the runqueue and wakes it.
2656 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2658 unsigned long flags;
2661 rq = task_rq_lock(p, &flags);
2662 BUG_ON(p->state != TASK_WAKING);
2663 p->state = TASK_RUNNING;
2664 update_rq_clock(rq);
2665 activate_task(rq, p, 0);
2666 trace_sched_wakeup_new(rq, p, 1);
2667 check_preempt_curr(rq, p, WF_FORK);
2669 if (p->sched_class->task_wake_up)
2670 p->sched_class->task_wake_up(rq, p);
2672 task_rq_unlock(rq, &flags);
2675 #ifdef CONFIG_PREEMPT_NOTIFIERS
2678 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2679 * @notifier: notifier struct to register
2681 void preempt_notifier_register(struct preempt_notifier *notifier)
2683 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2685 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2688 * preempt_notifier_unregister - no longer interested in preemption notifications
2689 * @notifier: notifier struct to unregister
2691 * This is safe to call from within a preemption notifier.
2693 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2695 hlist_del(¬ifier->link);
2697 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2699 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2701 struct preempt_notifier *notifier;
2702 struct hlist_node *node;
2704 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2705 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2709 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2710 struct task_struct *next)
2712 struct preempt_notifier *notifier;
2713 struct hlist_node *node;
2715 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2716 notifier->ops->sched_out(notifier, next);
2719 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2721 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2726 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2727 struct task_struct *next)
2731 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2734 * prepare_task_switch - prepare to switch tasks
2735 * @rq: the runqueue preparing to switch
2736 * @prev: the current task that is being switched out
2737 * @next: the task we are going to switch to.
2739 * This is called with the rq lock held and interrupts off. It must
2740 * be paired with a subsequent finish_task_switch after the context
2743 * prepare_task_switch sets up locking and calls architecture specific
2747 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2748 struct task_struct *next)
2750 fire_sched_out_preempt_notifiers(prev, next);
2751 prepare_lock_switch(rq, next);
2752 prepare_arch_switch(next);
2756 * finish_task_switch - clean up after a task-switch
2757 * @rq: runqueue associated with task-switch
2758 * @prev: the thread we just switched away from.
2760 * finish_task_switch must be called after the context switch, paired
2761 * with a prepare_task_switch call before the context switch.
2762 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2763 * and do any other architecture-specific cleanup actions.
2765 * Note that we may have delayed dropping an mm in context_switch(). If
2766 * so, we finish that here outside of the runqueue lock. (Doing it
2767 * with the lock held can cause deadlocks; see schedule() for
2770 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2771 __releases(rq->lock)
2773 struct mm_struct *mm = rq->prev_mm;
2779 * A task struct has one reference for the use as "current".
2780 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2781 * schedule one last time. The schedule call will never return, and
2782 * the scheduled task must drop that reference.
2783 * The test for TASK_DEAD must occur while the runqueue locks are
2784 * still held, otherwise prev could be scheduled on another cpu, die
2785 * there before we look at prev->state, and then the reference would
2787 * Manfred Spraul <manfred@colorfullife.com>
2789 prev_state = prev->state;
2790 finish_arch_switch(prev);
2791 perf_event_task_sched_in(current, cpu_of(rq));
2792 finish_lock_switch(rq, prev);
2794 fire_sched_in_preempt_notifiers(current);
2797 if (unlikely(prev_state == TASK_DEAD)) {
2799 * Remove function-return probe instances associated with this
2800 * task and put them back on the free list.
2802 kprobe_flush_task(prev);
2803 put_task_struct(prev);
2809 /* assumes rq->lock is held */
2810 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2812 if (prev->sched_class->pre_schedule)
2813 prev->sched_class->pre_schedule(rq, prev);
2816 /* rq->lock is NOT held, but preemption is disabled */
2817 static inline void post_schedule(struct rq *rq)
2819 if (rq->post_schedule) {
2820 unsigned long flags;
2822 raw_spin_lock_irqsave(&rq->lock, flags);
2823 if (rq->curr->sched_class->post_schedule)
2824 rq->curr->sched_class->post_schedule(rq);
2825 raw_spin_unlock_irqrestore(&rq->lock, flags);
2827 rq->post_schedule = 0;
2833 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2837 static inline void post_schedule(struct rq *rq)
2844 * schedule_tail - first thing a freshly forked thread must call.
2845 * @prev: the thread we just switched away from.
2847 asmlinkage void schedule_tail(struct task_struct *prev)
2848 __releases(rq->lock)
2850 struct rq *rq = this_rq();
2852 finish_task_switch(rq, prev);
2855 * FIXME: do we need to worry about rq being invalidated by the
2860 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2861 /* In this case, finish_task_switch does not reenable preemption */
2864 if (current->set_child_tid)
2865 put_user(task_pid_vnr(current), current->set_child_tid);
2869 * context_switch - switch to the new MM and the new
2870 * thread's register state.
2873 context_switch(struct rq *rq, struct task_struct *prev,
2874 struct task_struct *next)
2876 struct mm_struct *mm, *oldmm;
2878 prepare_task_switch(rq, prev, next);
2879 trace_sched_switch(rq, prev, next);
2881 oldmm = prev->active_mm;
2883 * For paravirt, this is coupled with an exit in switch_to to
2884 * combine the page table reload and the switch backend into
2887 arch_start_context_switch(prev);
2890 next->active_mm = oldmm;
2891 atomic_inc(&oldmm->mm_count);
2892 enter_lazy_tlb(oldmm, next);
2894 switch_mm(oldmm, mm, next);
2896 if (likely(!prev->mm)) {
2897 prev->active_mm = NULL;
2898 rq->prev_mm = oldmm;
2901 * Since the runqueue lock will be released by the next
2902 * task (which is an invalid locking op but in the case
2903 * of the scheduler it's an obvious special-case), so we
2904 * do an early lockdep release here:
2906 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2907 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2910 /* Here we just switch the register state and the stack. */
2911 switch_to(prev, next, prev);
2915 * this_rq must be evaluated again because prev may have moved
2916 * CPUs since it called schedule(), thus the 'rq' on its stack
2917 * frame will be invalid.
2919 finish_task_switch(this_rq(), prev);
2923 * nr_running, nr_uninterruptible and nr_context_switches:
2925 * externally visible scheduler statistics: current number of runnable
2926 * threads, current number of uninterruptible-sleeping threads, total
2927 * number of context switches performed since bootup.
2929 unsigned long nr_running(void)
2931 unsigned long i, sum = 0;
2933 for_each_online_cpu(i)
2934 sum += cpu_rq(i)->nr_running;
2939 unsigned long nr_uninterruptible(void)
2941 unsigned long i, sum = 0;
2943 for_each_possible_cpu(i)
2944 sum += cpu_rq(i)->nr_uninterruptible;
2947 * Since we read the counters lockless, it might be slightly
2948 * inaccurate. Do not allow it to go below zero though:
2950 if (unlikely((long)sum < 0))
2956 unsigned long long nr_context_switches(void)
2959 unsigned long long sum = 0;
2961 for_each_possible_cpu(i)
2962 sum += cpu_rq(i)->nr_switches;
2967 unsigned long nr_iowait(void)
2969 unsigned long i, sum = 0;
2971 for_each_possible_cpu(i)
2972 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2977 unsigned long nr_iowait_cpu(void)
2979 struct rq *this = this_rq();
2980 return atomic_read(&this->nr_iowait);
2983 unsigned long this_cpu_load(void)
2985 struct rq *this = this_rq();
2986 return this->cpu_load[0];
2990 /* Variables and functions for calc_load */
2991 static atomic_long_t calc_load_tasks;
2992 static unsigned long calc_load_update;
2993 unsigned long avenrun[3];
2994 EXPORT_SYMBOL(avenrun);
2997 * get_avenrun - get the load average array
2998 * @loads: pointer to dest load array
2999 * @offset: offset to add
3000 * @shift: shift count to shift the result left
3002 * These values are estimates at best, so no need for locking.
3004 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3006 loads[0] = (avenrun[0] + offset) << shift;
3007 loads[1] = (avenrun[1] + offset) << shift;
3008 loads[2] = (avenrun[2] + offset) << shift;
3011 static unsigned long
3012 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3015 load += active * (FIXED_1 - exp);
3016 return load >> FSHIFT;
3020 * calc_load - update the avenrun load estimates 10 ticks after the
3021 * CPUs have updated calc_load_tasks.
3023 void calc_global_load(void)
3025 unsigned long upd = calc_load_update + 10;
3028 if (time_before(jiffies, upd))
3031 active = atomic_long_read(&calc_load_tasks);
3032 active = active > 0 ? active * FIXED_1 : 0;
3034 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3035 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3036 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3038 calc_load_update += LOAD_FREQ;
3042 * Either called from update_cpu_load() or from a cpu going idle
3044 static void calc_load_account_active(struct rq *this_rq)
3046 long nr_active, delta;
3048 nr_active = this_rq->nr_running;
3049 nr_active += (long) this_rq->nr_uninterruptible;
3051 if (nr_active != this_rq->calc_load_active) {
3052 delta = nr_active - this_rq->calc_load_active;
3053 this_rq->calc_load_active = nr_active;
3054 atomic_long_add(delta, &calc_load_tasks);
3059 * Update rq->cpu_load[] statistics. This function is usually called every
3060 * scheduler tick (TICK_NSEC).
3062 static void update_cpu_load(struct rq *this_rq)
3064 unsigned long this_load = this_rq->load.weight;
3067 this_rq->nr_load_updates++;
3069 /* Update our load: */
3070 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3071 unsigned long old_load, new_load;
3073 /* scale is effectively 1 << i now, and >> i divides by scale */
3075 old_load = this_rq->cpu_load[i];
3076 new_load = this_load;
3078 * Round up the averaging division if load is increasing. This
3079 * prevents us from getting stuck on 9 if the load is 10, for
3082 if (new_load > old_load)
3083 new_load += scale-1;
3084 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3087 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3088 this_rq->calc_load_update += LOAD_FREQ;
3089 calc_load_account_active(this_rq);
3096 * double_rq_lock - safely lock two runqueues
3098 * Note this does not disable interrupts like task_rq_lock,
3099 * you need to do so manually before calling.
3101 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3102 __acquires(rq1->lock)
3103 __acquires(rq2->lock)
3105 BUG_ON(!irqs_disabled());
3107 raw_spin_lock(&rq1->lock);
3108 __acquire(rq2->lock); /* Fake it out ;) */
3111 raw_spin_lock(&rq1->lock);
3112 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3114 raw_spin_lock(&rq2->lock);
3115 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3118 update_rq_clock(rq1);
3119 update_rq_clock(rq2);
3123 * double_rq_unlock - safely unlock two runqueues
3125 * Note this does not restore interrupts like task_rq_unlock,
3126 * you need to do so manually after calling.
3128 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3129 __releases(rq1->lock)
3130 __releases(rq2->lock)
3132 raw_spin_unlock(&rq1->lock);
3134 raw_spin_unlock(&rq2->lock);
3136 __release(rq2->lock);
3140 * sched_exec - execve() is a valuable balancing opportunity, because at
3141 * this point the task has the smallest effective memory and cache footprint.
3143 void sched_exec(void)
3145 struct task_struct *p = current;
3146 struct migration_req req;
3147 int dest_cpu, this_cpu;
3148 unsigned long flags;
3152 this_cpu = get_cpu();
3153 dest_cpu = select_task_rq(p, SD_BALANCE_EXEC, 0);
3154 if (dest_cpu == this_cpu) {
3159 rq = task_rq_lock(p, &flags);
3163 * select_task_rq() can race against ->cpus_allowed
3165 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3166 || unlikely(!cpu_active(dest_cpu))) {
3167 task_rq_unlock(rq, &flags);
3171 /* force the process onto the specified CPU */
3172 if (migrate_task(p, dest_cpu, &req)) {
3173 /* Need to wait for migration thread (might exit: take ref). */
3174 struct task_struct *mt = rq->migration_thread;
3176 get_task_struct(mt);
3177 task_rq_unlock(rq, &flags);
3178 wake_up_process(mt);
3179 put_task_struct(mt);
3180 wait_for_completion(&req.done);
3184 task_rq_unlock(rq, &flags);
3188 * pull_task - move a task from a remote runqueue to the local runqueue.
3189 * Both runqueues must be locked.
3191 static void pull_task(struct rq *src_rq, struct task_struct *p,
3192 struct rq *this_rq, int this_cpu)
3194 deactivate_task(src_rq, p, 0);
3195 set_task_cpu(p, this_cpu);
3196 activate_task(this_rq, p, 0);
3197 check_preempt_curr(this_rq, p, 0);
3201 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3204 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3205 struct sched_domain *sd, enum cpu_idle_type idle,
3208 int tsk_cache_hot = 0;
3210 * We do not migrate tasks that are:
3211 * 1) running (obviously), or
3212 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3213 * 3) are cache-hot on their current CPU.
3215 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3216 schedstat_inc(p, se.nr_failed_migrations_affine);
3221 if (task_running(rq, p)) {
3222 schedstat_inc(p, se.nr_failed_migrations_running);
3227 * Aggressive migration if:
3228 * 1) task is cache cold, or
3229 * 2) too many balance attempts have failed.
3232 tsk_cache_hot = task_hot(p, rq->clock, sd);
3233 if (!tsk_cache_hot ||
3234 sd->nr_balance_failed > sd->cache_nice_tries) {
3235 #ifdef CONFIG_SCHEDSTATS
3236 if (tsk_cache_hot) {
3237 schedstat_inc(sd, lb_hot_gained[idle]);
3238 schedstat_inc(p, se.nr_forced_migrations);
3244 if (tsk_cache_hot) {
3245 schedstat_inc(p, se.nr_failed_migrations_hot);
3251 static unsigned long
3252 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3253 unsigned long max_load_move, struct sched_domain *sd,
3254 enum cpu_idle_type idle, int *all_pinned,
3255 int *this_best_prio, struct rq_iterator *iterator)
3257 int loops = 0, pulled = 0, pinned = 0;
3258 struct task_struct *p;
3259 long rem_load_move = max_load_move;
3261 if (max_load_move == 0)
3267 * Start the load-balancing iterator:
3269 p = iterator->start(iterator->arg);
3271 if (!p || loops++ > sysctl_sched_nr_migrate)
3274 if ((p->se.load.weight >> 1) > rem_load_move ||
3275 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3276 p = iterator->next(iterator->arg);
3280 pull_task(busiest, p, this_rq, this_cpu);
3282 rem_load_move -= p->se.load.weight;
3284 #ifdef CONFIG_PREEMPT
3286 * NEWIDLE balancing is a source of latency, so preemptible kernels
3287 * will stop after the first task is pulled to minimize the critical
3290 if (idle == CPU_NEWLY_IDLE)
3295 * We only want to steal up to the prescribed amount of weighted load.
3297 if (rem_load_move > 0) {
3298 if (p->prio < *this_best_prio)
3299 *this_best_prio = p->prio;
3300 p = iterator->next(iterator->arg);
3305 * Right now, this is one of only two places pull_task() is called,
3306 * so we can safely collect pull_task() stats here rather than
3307 * inside pull_task().
3309 schedstat_add(sd, lb_gained[idle], pulled);
3312 *all_pinned = pinned;
3314 return max_load_move - rem_load_move;
3318 * move_tasks tries to move up to max_load_move weighted load from busiest to
3319 * this_rq, as part of a balancing operation within domain "sd".
3320 * Returns 1 if successful and 0 otherwise.
3322 * Called with both runqueues locked.
3324 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3325 unsigned long max_load_move,
3326 struct sched_domain *sd, enum cpu_idle_type idle,
3329 const struct sched_class *class = sched_class_highest;
3330 unsigned long total_load_moved = 0;
3331 int this_best_prio = this_rq->curr->prio;
3335 class->load_balance(this_rq, this_cpu, busiest,
3336 max_load_move - total_load_moved,
3337 sd, idle, all_pinned, &this_best_prio);
3338 class = class->next;
3340 #ifdef CONFIG_PREEMPT
3342 * NEWIDLE balancing is a source of latency, so preemptible
3343 * kernels will stop after the first task is pulled to minimize
3344 * the critical section.
3346 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3349 } while (class && max_load_move > total_load_moved);
3351 return total_load_moved > 0;
3355 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3356 struct sched_domain *sd, enum cpu_idle_type idle,
3357 struct rq_iterator *iterator)
3359 struct task_struct *p = iterator->start(iterator->arg);
3363 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3364 pull_task(busiest, p, this_rq, this_cpu);
3366 * Right now, this is only the second place pull_task()
3367 * is called, so we can safely collect pull_task()
3368 * stats here rather than inside pull_task().
3370 schedstat_inc(sd, lb_gained[idle]);
3374 p = iterator->next(iterator->arg);
3381 * move_one_task tries to move exactly one task from busiest to this_rq, as
3382 * part of active balancing operations within "domain".
3383 * Returns 1 if successful and 0 otherwise.
3385 * Called with both runqueues locked.
3387 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3388 struct sched_domain *sd, enum cpu_idle_type idle)
3390 const struct sched_class *class;
3392 for_each_class(class) {
3393 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3399 /********** Helpers for find_busiest_group ************************/
3401 * sd_lb_stats - Structure to store the statistics of a sched_domain
3402 * during load balancing.
3404 struct sd_lb_stats {
3405 struct sched_group *busiest; /* Busiest group in this sd */
3406 struct sched_group *this; /* Local group in this sd */
3407 unsigned long total_load; /* Total load of all groups in sd */
3408 unsigned long total_pwr; /* Total power of all groups in sd */
3409 unsigned long avg_load; /* Average load across all groups in sd */
3411 /** Statistics of this group */
3412 unsigned long this_load;
3413 unsigned long this_load_per_task;
3414 unsigned long this_nr_running;
3416 /* Statistics of the busiest group */
3417 unsigned long max_load;
3418 unsigned long busiest_load_per_task;
3419 unsigned long busiest_nr_running;
3421 int group_imb; /* Is there imbalance in this sd */
3422 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3423 int power_savings_balance; /* Is powersave balance needed for this sd */
3424 struct sched_group *group_min; /* Least loaded group in sd */
3425 struct sched_group *group_leader; /* Group which relieves group_min */
3426 unsigned long min_load_per_task; /* load_per_task in group_min */
3427 unsigned long leader_nr_running; /* Nr running of group_leader */
3428 unsigned long min_nr_running; /* Nr running of group_min */
3433 * sg_lb_stats - stats of a sched_group required for load_balancing
3435 struct sg_lb_stats {
3436 unsigned long avg_load; /*Avg load across the CPUs of the group */
3437 unsigned long group_load; /* Total load over the CPUs of the group */
3438 unsigned long sum_nr_running; /* Nr tasks running in the group */
3439 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3440 unsigned long group_capacity;
3441 int group_imb; /* Is there an imbalance in the group ? */
3445 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3446 * @group: The group whose first cpu is to be returned.
3448 static inline unsigned int group_first_cpu(struct sched_group *group)
3450 return cpumask_first(sched_group_cpus(group));
3454 * get_sd_load_idx - Obtain the load index for a given sched domain.
3455 * @sd: The sched_domain whose load_idx is to be obtained.
3456 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3458 static inline int get_sd_load_idx(struct sched_domain *sd,
3459 enum cpu_idle_type idle)
3465 load_idx = sd->busy_idx;
3468 case CPU_NEWLY_IDLE:
3469 load_idx = sd->newidle_idx;
3472 load_idx = sd->idle_idx;
3480 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3482 * init_sd_power_savings_stats - Initialize power savings statistics for
3483 * the given sched_domain, during load balancing.
3485 * @sd: Sched domain whose power-savings statistics are to be initialized.
3486 * @sds: Variable containing the statistics for sd.
3487 * @idle: Idle status of the CPU at which we're performing load-balancing.
3489 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3490 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3493 * Busy processors will not participate in power savings
3496 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3497 sds->power_savings_balance = 0;
3499 sds->power_savings_balance = 1;
3500 sds->min_nr_running = ULONG_MAX;
3501 sds->leader_nr_running = 0;
3506 * update_sd_power_savings_stats - Update the power saving stats for a
3507 * sched_domain while performing load balancing.
3509 * @group: sched_group belonging to the sched_domain under consideration.
3510 * @sds: Variable containing the statistics of the sched_domain
3511 * @local_group: Does group contain the CPU for which we're performing
3513 * @sgs: Variable containing the statistics of the group.
3515 static inline void update_sd_power_savings_stats(struct sched_group *group,
3516 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3519 if (!sds->power_savings_balance)
3523 * If the local group is idle or completely loaded
3524 * no need to do power savings balance at this domain
3526 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3527 !sds->this_nr_running))
3528 sds->power_savings_balance = 0;
3531 * If a group is already running at full capacity or idle,
3532 * don't include that group in power savings calculations
3534 if (!sds->power_savings_balance ||
3535 sgs->sum_nr_running >= sgs->group_capacity ||
3536 !sgs->sum_nr_running)
3540 * Calculate the group which has the least non-idle load.
3541 * This is the group from where we need to pick up the load
3544 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3545 (sgs->sum_nr_running == sds->min_nr_running &&
3546 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3547 sds->group_min = group;
3548 sds->min_nr_running = sgs->sum_nr_running;
3549 sds->min_load_per_task = sgs->sum_weighted_load /
3550 sgs->sum_nr_running;
3554 * Calculate the group which is almost near its
3555 * capacity but still has some space to pick up some load
3556 * from other group and save more power
3558 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3561 if (sgs->sum_nr_running > sds->leader_nr_running ||
3562 (sgs->sum_nr_running == sds->leader_nr_running &&
3563 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3564 sds->group_leader = group;
3565 sds->leader_nr_running = sgs->sum_nr_running;
3570 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3571 * @sds: Variable containing the statistics of the sched_domain
3572 * under consideration.
3573 * @this_cpu: Cpu at which we're currently performing load-balancing.
3574 * @imbalance: Variable to store the imbalance.
3577 * Check if we have potential to perform some power-savings balance.
3578 * If yes, set the busiest group to be the least loaded group in the
3579 * sched_domain, so that it's CPUs can be put to idle.
3581 * Returns 1 if there is potential to perform power-savings balance.
3584 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3585 int this_cpu, unsigned long *imbalance)
3587 if (!sds->power_savings_balance)
3590 if (sds->this != sds->group_leader ||
3591 sds->group_leader == sds->group_min)
3594 *imbalance = sds->min_load_per_task;
3595 sds->busiest = sds->group_min;
3600 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3601 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3602 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3607 static inline void update_sd_power_savings_stats(struct sched_group *group,
3608 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3613 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3614 int this_cpu, unsigned long *imbalance)
3618 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3621 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3623 return SCHED_LOAD_SCALE;
3626 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3628 return default_scale_freq_power(sd, cpu);
3631 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3633 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3634 unsigned long smt_gain = sd->smt_gain;
3641 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3643 return default_scale_smt_power(sd, cpu);
3646 unsigned long scale_rt_power(int cpu)
3648 struct rq *rq = cpu_rq(cpu);
3649 u64 total, available;
3651 sched_avg_update(rq);
3653 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3654 available = total - rq->rt_avg;
3656 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3657 total = SCHED_LOAD_SCALE;
3659 total >>= SCHED_LOAD_SHIFT;
3661 return div_u64(available, total);
3664 static void update_cpu_power(struct sched_domain *sd, int cpu)
3666 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3667 unsigned long power = SCHED_LOAD_SCALE;
3668 struct sched_group *sdg = sd->groups;
3670 if (sched_feat(ARCH_POWER))
3671 power *= arch_scale_freq_power(sd, cpu);
3673 power *= default_scale_freq_power(sd, cpu);
3675 power >>= SCHED_LOAD_SHIFT;
3677 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3678 if (sched_feat(ARCH_POWER))
3679 power *= arch_scale_smt_power(sd, cpu);
3681 power *= default_scale_smt_power(sd, cpu);
3683 power >>= SCHED_LOAD_SHIFT;
3686 power *= scale_rt_power(cpu);
3687 power >>= SCHED_LOAD_SHIFT;
3692 sdg->cpu_power = power;
3695 static void update_group_power(struct sched_domain *sd, int cpu)
3697 struct sched_domain *child = sd->child;
3698 struct sched_group *group, *sdg = sd->groups;
3699 unsigned long power;
3702 update_cpu_power(sd, cpu);
3708 group = child->groups;
3710 power += group->cpu_power;
3711 group = group->next;
3712 } while (group != child->groups);
3714 sdg->cpu_power = power;
3718 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3719 * @sd: The sched_domain whose statistics are to be updated.
3720 * @group: sched_group whose statistics are to be updated.
3721 * @this_cpu: Cpu for which load balance is currently performed.
3722 * @idle: Idle status of this_cpu
3723 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3724 * @sd_idle: Idle status of the sched_domain containing group.
3725 * @local_group: Does group contain this_cpu.
3726 * @cpus: Set of cpus considered for load balancing.
3727 * @balance: Should we balance.
3728 * @sgs: variable to hold the statistics for this group.
3730 static inline void update_sg_lb_stats(struct sched_domain *sd,
3731 struct sched_group *group, int this_cpu,
3732 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3733 int local_group, const struct cpumask *cpus,
3734 int *balance, struct sg_lb_stats *sgs)
3736 unsigned long load, max_cpu_load, min_cpu_load;
3738 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3739 unsigned long sum_avg_load_per_task;
3740 unsigned long avg_load_per_task;
3743 balance_cpu = group_first_cpu(group);
3744 if (balance_cpu == this_cpu)
3745 update_group_power(sd, this_cpu);
3748 /* Tally up the load of all CPUs in the group */
3749 sum_avg_load_per_task = avg_load_per_task = 0;
3751 min_cpu_load = ~0UL;
3753 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3754 struct rq *rq = cpu_rq(i);
3756 if (*sd_idle && rq->nr_running)
3759 /* Bias balancing toward cpus of our domain */
3761 if (idle_cpu(i) && !first_idle_cpu) {
3766 load = target_load(i, load_idx);
3768 load = source_load(i, load_idx);
3769 if (load > max_cpu_load)
3770 max_cpu_load = load;
3771 if (min_cpu_load > load)
3772 min_cpu_load = load;
3775 sgs->group_load += load;
3776 sgs->sum_nr_running += rq->nr_running;
3777 sgs->sum_weighted_load += weighted_cpuload(i);
3779 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3783 * First idle cpu or the first cpu(busiest) in this sched group
3784 * is eligible for doing load balancing at this and above
3785 * domains. In the newly idle case, we will allow all the cpu's
3786 * to do the newly idle load balance.
3788 if (idle != CPU_NEWLY_IDLE && local_group &&
3789 balance_cpu != this_cpu && balance) {
3794 /* Adjust by relative CPU power of the group */
3795 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3799 * Consider the group unbalanced when the imbalance is larger
3800 * than the average weight of two tasks.
3802 * APZ: with cgroup the avg task weight can vary wildly and
3803 * might not be a suitable number - should we keep a
3804 * normalized nr_running number somewhere that negates
3807 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3810 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3813 sgs->group_capacity =
3814 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3818 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3819 * @sd: sched_domain whose statistics are to be updated.
3820 * @this_cpu: Cpu for which load balance is currently performed.
3821 * @idle: Idle status of this_cpu
3822 * @sd_idle: Idle status of the sched_domain containing group.
3823 * @cpus: Set of cpus considered for load balancing.
3824 * @balance: Should we balance.
3825 * @sds: variable to hold the statistics for this sched_domain.
3827 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3828 enum cpu_idle_type idle, int *sd_idle,
3829 const struct cpumask *cpus, int *balance,
3830 struct sd_lb_stats *sds)
3832 struct sched_domain *child = sd->child;
3833 struct sched_group *group = sd->groups;
3834 struct sg_lb_stats sgs;
3835 int load_idx, prefer_sibling = 0;
3837 if (child && child->flags & SD_PREFER_SIBLING)
3840 init_sd_power_savings_stats(sd, sds, idle);
3841 load_idx = get_sd_load_idx(sd, idle);
3846 local_group = cpumask_test_cpu(this_cpu,
3847 sched_group_cpus(group));
3848 memset(&sgs, 0, sizeof(sgs));
3849 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3850 local_group, cpus, balance, &sgs);
3852 if (local_group && balance && !(*balance))
3855 sds->total_load += sgs.group_load;
3856 sds->total_pwr += group->cpu_power;
3859 * In case the child domain prefers tasks go to siblings
3860 * first, lower the group capacity to one so that we'll try
3861 * and move all the excess tasks away.
3864 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3867 sds->this_load = sgs.avg_load;
3869 sds->this_nr_running = sgs.sum_nr_running;
3870 sds->this_load_per_task = sgs.sum_weighted_load;
3871 } else if (sgs.avg_load > sds->max_load &&
3872 (sgs.sum_nr_running > sgs.group_capacity ||
3874 sds->max_load = sgs.avg_load;
3875 sds->busiest = group;
3876 sds->busiest_nr_running = sgs.sum_nr_running;
3877 sds->busiest_load_per_task = sgs.sum_weighted_load;
3878 sds->group_imb = sgs.group_imb;
3881 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3882 group = group->next;
3883 } while (group != sd->groups);
3887 * fix_small_imbalance - Calculate the minor imbalance that exists
3888 * amongst the groups of a sched_domain, during
3890 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3891 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3892 * @imbalance: Variable to store the imbalance.
3894 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3895 int this_cpu, unsigned long *imbalance)
3897 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3898 unsigned int imbn = 2;
3900 if (sds->this_nr_running) {
3901 sds->this_load_per_task /= sds->this_nr_running;
3902 if (sds->busiest_load_per_task >
3903 sds->this_load_per_task)
3906 sds->this_load_per_task =
3907 cpu_avg_load_per_task(this_cpu);
3909 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3910 sds->busiest_load_per_task * imbn) {
3911 *imbalance = sds->busiest_load_per_task;
3916 * OK, we don't have enough imbalance to justify moving tasks,
3917 * however we may be able to increase total CPU power used by
3921 pwr_now += sds->busiest->cpu_power *
3922 min(sds->busiest_load_per_task, sds->max_load);
3923 pwr_now += sds->this->cpu_power *
3924 min(sds->this_load_per_task, sds->this_load);
3925 pwr_now /= SCHED_LOAD_SCALE;
3927 /* Amount of load we'd subtract */
3928 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3929 sds->busiest->cpu_power;
3930 if (sds->max_load > tmp)
3931 pwr_move += sds->busiest->cpu_power *
3932 min(sds->busiest_load_per_task, sds->max_load - tmp);
3934 /* Amount of load we'd add */
3935 if (sds->max_load * sds->busiest->cpu_power <
3936 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3937 tmp = (sds->max_load * sds->busiest->cpu_power) /
3938 sds->this->cpu_power;
3940 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3941 sds->this->cpu_power;
3942 pwr_move += sds->this->cpu_power *
3943 min(sds->this_load_per_task, sds->this_load + tmp);
3944 pwr_move /= SCHED_LOAD_SCALE;
3946 /* Move if we gain throughput */
3947 if (pwr_move > pwr_now)
3948 *imbalance = sds->busiest_load_per_task;
3952 * calculate_imbalance - Calculate the amount of imbalance present within the
3953 * groups of a given sched_domain during load balance.
3954 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3955 * @this_cpu: Cpu for which currently load balance is being performed.
3956 * @imbalance: The variable to store the imbalance.
3958 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3959 unsigned long *imbalance)
3961 unsigned long max_pull;
3963 * In the presence of smp nice balancing, certain scenarios can have
3964 * max load less than avg load(as we skip the groups at or below
3965 * its cpu_power, while calculating max_load..)
3967 if (sds->max_load < sds->avg_load) {
3969 return fix_small_imbalance(sds, this_cpu, imbalance);
3972 /* Don't want to pull so many tasks that a group would go idle */
3973 max_pull = min(sds->max_load - sds->avg_load,
3974 sds->max_load - sds->busiest_load_per_task);
3976 /* How much load to actually move to equalise the imbalance */
3977 *imbalance = min(max_pull * sds->busiest->cpu_power,
3978 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3982 * if *imbalance is less than the average load per runnable task
3983 * there is no gaurantee that any tasks will be moved so we'll have
3984 * a think about bumping its value to force at least one task to be
3987 if (*imbalance < sds->busiest_load_per_task)
3988 return fix_small_imbalance(sds, this_cpu, imbalance);
3991 /******* find_busiest_group() helpers end here *********************/
3994 * find_busiest_group - Returns the busiest group within the sched_domain
3995 * if there is an imbalance. If there isn't an imbalance, and
3996 * the user has opted for power-savings, it returns a group whose
3997 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3998 * such a group exists.
4000 * Also calculates the amount of weighted load which should be moved
4001 * to restore balance.
4003 * @sd: The sched_domain whose busiest group is to be returned.
4004 * @this_cpu: The cpu for which load balancing is currently being performed.
4005 * @imbalance: Variable which stores amount of weighted load which should
4006 * be moved to restore balance/put a group to idle.
4007 * @idle: The idle status of this_cpu.
4008 * @sd_idle: The idleness of sd
4009 * @cpus: The set of CPUs under consideration for load-balancing.
4010 * @balance: Pointer to a variable indicating if this_cpu
4011 * is the appropriate cpu to perform load balancing at this_level.
4013 * Returns: - the busiest group if imbalance exists.
4014 * - If no imbalance and user has opted for power-savings balance,
4015 * return the least loaded group whose CPUs can be
4016 * put to idle by rebalancing its tasks onto our group.
4018 static struct sched_group *
4019 find_busiest_group(struct sched_domain *sd, int this_cpu,
4020 unsigned long *imbalance, enum cpu_idle_type idle,
4021 int *sd_idle, const struct cpumask *cpus, int *balance)
4023 struct sd_lb_stats sds;
4025 memset(&sds, 0, sizeof(sds));
4028 * Compute the various statistics relavent for load balancing at
4031 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4034 /* Cases where imbalance does not exist from POV of this_cpu */
4035 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4037 * 2) There is no busy sibling group to pull from.
4038 * 3) This group is the busiest group.
4039 * 4) This group is more busy than the avg busieness at this
4041 * 5) The imbalance is within the specified limit.
4042 * 6) Any rebalance would lead to ping-pong
4044 if (balance && !(*balance))
4047 if (!sds.busiest || sds.busiest_nr_running == 0)
4050 if (sds.this_load >= sds.max_load)
4053 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4055 if (sds.this_load >= sds.avg_load)
4058 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4061 sds.busiest_load_per_task /= sds.busiest_nr_running;
4063 sds.busiest_load_per_task =
4064 min(sds.busiest_load_per_task, sds.avg_load);
4067 * We're trying to get all the cpus to the average_load, so we don't
4068 * want to push ourselves above the average load, nor do we wish to
4069 * reduce the max loaded cpu below the average load, as either of these
4070 * actions would just result in more rebalancing later, and ping-pong
4071 * tasks around. Thus we look for the minimum possible imbalance.
4072 * Negative imbalances (*we* are more loaded than anyone else) will
4073 * be counted as no imbalance for these purposes -- we can't fix that
4074 * by pulling tasks to us. Be careful of negative numbers as they'll
4075 * appear as very large values with unsigned longs.
4077 if (sds.max_load <= sds.busiest_load_per_task)
4080 /* Looks like there is an imbalance. Compute it */
4081 calculate_imbalance(&sds, this_cpu, imbalance);
4086 * There is no obvious imbalance. But check if we can do some balancing
4089 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4097 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4100 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4101 unsigned long imbalance, const struct cpumask *cpus)
4103 struct rq *busiest = NULL, *rq;
4104 unsigned long max_load = 0;
4107 for_each_cpu(i, sched_group_cpus(group)) {
4108 unsigned long power = power_of(i);
4109 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4112 if (!cpumask_test_cpu(i, cpus))
4116 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4119 if (capacity && rq->nr_running == 1 && wl > imbalance)
4122 if (wl > max_load) {
4132 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4133 * so long as it is large enough.
4135 #define MAX_PINNED_INTERVAL 512
4137 /* Working cpumask for load_balance and load_balance_newidle. */
4138 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4141 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4142 * tasks if there is an imbalance.
4144 static int load_balance(int this_cpu, struct rq *this_rq,
4145 struct sched_domain *sd, enum cpu_idle_type idle,
4148 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4149 struct sched_group *group;
4150 unsigned long imbalance;
4152 unsigned long flags;
4153 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4155 cpumask_copy(cpus, cpu_active_mask);
4158 * When power savings policy is enabled for the parent domain, idle
4159 * sibling can pick up load irrespective of busy siblings. In this case,
4160 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4161 * portraying it as CPU_NOT_IDLE.
4163 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4164 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4167 schedstat_inc(sd, lb_count[idle]);
4171 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4178 schedstat_inc(sd, lb_nobusyg[idle]);
4182 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4184 schedstat_inc(sd, lb_nobusyq[idle]);
4188 BUG_ON(busiest == this_rq);
4190 schedstat_add(sd, lb_imbalance[idle], imbalance);
4193 if (busiest->nr_running > 1) {
4195 * Attempt to move tasks. If find_busiest_group has found
4196 * an imbalance but busiest->nr_running <= 1, the group is
4197 * still unbalanced. ld_moved simply stays zero, so it is
4198 * correctly treated as an imbalance.
4200 local_irq_save(flags);
4201 double_rq_lock(this_rq, busiest);
4202 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4203 imbalance, sd, idle, &all_pinned);
4204 double_rq_unlock(this_rq, busiest);
4205 local_irq_restore(flags);
4208 * some other cpu did the load balance for us.
4210 if (ld_moved && this_cpu != smp_processor_id())
4211 resched_cpu(this_cpu);
4213 /* All tasks on this runqueue were pinned by CPU affinity */
4214 if (unlikely(all_pinned)) {
4215 cpumask_clear_cpu(cpu_of(busiest), cpus);
4216 if (!cpumask_empty(cpus))
4223 schedstat_inc(sd, lb_failed[idle]);
4224 sd->nr_balance_failed++;
4226 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4228 raw_spin_lock_irqsave(&busiest->lock, flags);
4230 /* don't kick the migration_thread, if the curr
4231 * task on busiest cpu can't be moved to this_cpu
4233 if (!cpumask_test_cpu(this_cpu,
4234 &busiest->curr->cpus_allowed)) {
4235 raw_spin_unlock_irqrestore(&busiest->lock,
4238 goto out_one_pinned;
4241 if (!busiest->active_balance) {
4242 busiest->active_balance = 1;
4243 busiest->push_cpu = this_cpu;
4246 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4248 wake_up_process(busiest->migration_thread);
4251 * We've kicked active balancing, reset the failure
4254 sd->nr_balance_failed = sd->cache_nice_tries+1;
4257 sd->nr_balance_failed = 0;
4259 if (likely(!active_balance)) {
4260 /* We were unbalanced, so reset the balancing interval */
4261 sd->balance_interval = sd->min_interval;
4264 * If we've begun active balancing, start to back off. This
4265 * case may not be covered by the all_pinned logic if there
4266 * is only 1 task on the busy runqueue (because we don't call
4269 if (sd->balance_interval < sd->max_interval)
4270 sd->balance_interval *= 2;
4273 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4274 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4280 schedstat_inc(sd, lb_balanced[idle]);
4282 sd->nr_balance_failed = 0;
4285 /* tune up the balancing interval */
4286 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4287 (sd->balance_interval < sd->max_interval))
4288 sd->balance_interval *= 2;
4290 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4291 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4302 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4303 * tasks if there is an imbalance.
4305 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4306 * this_rq is locked.
4309 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4311 struct sched_group *group;
4312 struct rq *busiest = NULL;
4313 unsigned long imbalance;
4317 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4319 cpumask_copy(cpus, cpu_active_mask);
4322 * When power savings policy is enabled for the parent domain, idle
4323 * sibling can pick up load irrespective of busy siblings. In this case,
4324 * let the state of idle sibling percolate up as IDLE, instead of
4325 * portraying it as CPU_NOT_IDLE.
4327 if (sd->flags & SD_SHARE_CPUPOWER &&
4328 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4331 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4333 update_shares_locked(this_rq, sd);
4334 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4335 &sd_idle, cpus, NULL);
4337 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4341 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4343 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4347 BUG_ON(busiest == this_rq);
4349 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4352 if (busiest->nr_running > 1) {
4353 /* Attempt to move tasks */
4354 double_lock_balance(this_rq, busiest);
4355 /* this_rq->clock is already updated */
4356 update_rq_clock(busiest);
4357 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4358 imbalance, sd, CPU_NEWLY_IDLE,
4360 double_unlock_balance(this_rq, busiest);
4362 if (unlikely(all_pinned)) {
4363 cpumask_clear_cpu(cpu_of(busiest), cpus);
4364 if (!cpumask_empty(cpus))
4370 int active_balance = 0;
4372 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4373 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4374 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4377 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4380 if (sd->nr_balance_failed++ < 2)
4384 * The only task running in a non-idle cpu can be moved to this
4385 * cpu in an attempt to completely freeup the other CPU
4386 * package. The same method used to move task in load_balance()
4387 * have been extended for load_balance_newidle() to speedup
4388 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4390 * The package power saving logic comes from
4391 * find_busiest_group(). If there are no imbalance, then
4392 * f_b_g() will return NULL. However when sched_mc={1,2} then
4393 * f_b_g() will select a group from which a running task may be
4394 * pulled to this cpu in order to make the other package idle.
4395 * If there is no opportunity to make a package idle and if
4396 * there are no imbalance, then f_b_g() will return NULL and no
4397 * action will be taken in load_balance_newidle().
4399 * Under normal task pull operation due to imbalance, there
4400 * will be more than one task in the source run queue and
4401 * move_tasks() will succeed. ld_moved will be true and this
4402 * active balance code will not be triggered.
4405 /* Lock busiest in correct order while this_rq is held */
4406 double_lock_balance(this_rq, busiest);
4409 * don't kick the migration_thread, if the curr
4410 * task on busiest cpu can't be moved to this_cpu
4412 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4413 double_unlock_balance(this_rq, busiest);
4418 if (!busiest->active_balance) {
4419 busiest->active_balance = 1;
4420 busiest->push_cpu = this_cpu;
4424 double_unlock_balance(this_rq, busiest);
4426 * Should not call ttwu while holding a rq->lock
4428 raw_spin_unlock(&this_rq->lock);
4430 wake_up_process(busiest->migration_thread);
4431 raw_spin_lock(&this_rq->lock);
4434 sd->nr_balance_failed = 0;
4436 update_shares_locked(this_rq, sd);
4440 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4441 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4442 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4444 sd->nr_balance_failed = 0;
4450 * idle_balance is called by schedule() if this_cpu is about to become
4451 * idle. Attempts to pull tasks from other CPUs.
4453 static void idle_balance(int this_cpu, struct rq *this_rq)
4455 struct sched_domain *sd;
4456 int pulled_task = 0;
4457 unsigned long next_balance = jiffies + HZ;
4459 this_rq->idle_stamp = this_rq->clock;
4461 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4464 for_each_domain(this_cpu, sd) {
4465 unsigned long interval;
4467 if (!(sd->flags & SD_LOAD_BALANCE))
4470 if (sd->flags & SD_BALANCE_NEWIDLE)
4471 /* If we've pulled tasks over stop searching: */
4472 pulled_task = load_balance_newidle(this_cpu, this_rq,
4475 interval = msecs_to_jiffies(sd->balance_interval);
4476 if (time_after(next_balance, sd->last_balance + interval))
4477 next_balance = sd->last_balance + interval;
4479 this_rq->idle_stamp = 0;
4483 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4485 * We are going idle. next_balance may be set based on
4486 * a busy processor. So reset next_balance.
4488 this_rq->next_balance = next_balance;
4493 * active_load_balance is run by migration threads. It pushes running tasks
4494 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4495 * running on each physical CPU where possible, and avoids physical /
4496 * logical imbalances.
4498 * Called with busiest_rq locked.
4500 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4502 int target_cpu = busiest_rq->push_cpu;
4503 struct sched_domain *sd;
4504 struct rq *target_rq;
4506 /* Is there any task to move? */
4507 if (busiest_rq->nr_running <= 1)
4510 target_rq = cpu_rq(target_cpu);
4513 * This condition is "impossible", if it occurs
4514 * we need to fix it. Originally reported by
4515 * Bjorn Helgaas on a 128-cpu setup.
4517 BUG_ON(busiest_rq == target_rq);
4519 /* move a task from busiest_rq to target_rq */
4520 double_lock_balance(busiest_rq, target_rq);
4521 update_rq_clock(busiest_rq);
4522 update_rq_clock(target_rq);
4524 /* Search for an sd spanning us and the target CPU. */
4525 for_each_domain(target_cpu, sd) {
4526 if ((sd->flags & SD_LOAD_BALANCE) &&
4527 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4532 schedstat_inc(sd, alb_count);
4534 if (move_one_task(target_rq, target_cpu, busiest_rq,
4536 schedstat_inc(sd, alb_pushed);
4538 schedstat_inc(sd, alb_failed);
4540 double_unlock_balance(busiest_rq, target_rq);
4545 atomic_t load_balancer;
4546 cpumask_var_t cpu_mask;
4547 cpumask_var_t ilb_grp_nohz_mask;
4548 } nohz ____cacheline_aligned = {
4549 .load_balancer = ATOMIC_INIT(-1),
4552 int get_nohz_load_balancer(void)
4554 return atomic_read(&nohz.load_balancer);
4557 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4559 * lowest_flag_domain - Return lowest sched_domain containing flag.
4560 * @cpu: The cpu whose lowest level of sched domain is to
4562 * @flag: The flag to check for the lowest sched_domain
4563 * for the given cpu.
4565 * Returns the lowest sched_domain of a cpu which contains the given flag.
4567 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4569 struct sched_domain *sd;
4571 for_each_domain(cpu, sd)
4572 if (sd && (sd->flags & flag))
4579 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4580 * @cpu: The cpu whose domains we're iterating over.
4581 * @sd: variable holding the value of the power_savings_sd
4583 * @flag: The flag to filter the sched_domains to be iterated.
4585 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4586 * set, starting from the lowest sched_domain to the highest.
4588 #define for_each_flag_domain(cpu, sd, flag) \
4589 for (sd = lowest_flag_domain(cpu, flag); \
4590 (sd && (sd->flags & flag)); sd = sd->parent)
4593 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4594 * @ilb_group: group to be checked for semi-idleness
4596 * Returns: 1 if the group is semi-idle. 0 otherwise.
4598 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4599 * and atleast one non-idle CPU. This helper function checks if the given
4600 * sched_group is semi-idle or not.
4602 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4604 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4605 sched_group_cpus(ilb_group));
4608 * A sched_group is semi-idle when it has atleast one busy cpu
4609 * and atleast one idle cpu.
4611 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4614 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4620 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4621 * @cpu: The cpu which is nominating a new idle_load_balancer.
4623 * Returns: Returns the id of the idle load balancer if it exists,
4624 * Else, returns >= nr_cpu_ids.
4626 * This algorithm picks the idle load balancer such that it belongs to a
4627 * semi-idle powersavings sched_domain. The idea is to try and avoid
4628 * completely idle packages/cores just for the purpose of idle load balancing
4629 * when there are other idle cpu's which are better suited for that job.
4631 static int find_new_ilb(int cpu)
4633 struct sched_domain *sd;
4634 struct sched_group *ilb_group;
4637 * Have idle load balancer selection from semi-idle packages only
4638 * when power-aware load balancing is enabled
4640 if (!(sched_smt_power_savings || sched_mc_power_savings))
4644 * Optimize for the case when we have no idle CPUs or only one
4645 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4647 if (cpumask_weight(nohz.cpu_mask) < 2)
4650 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4651 ilb_group = sd->groups;
4654 if (is_semi_idle_group(ilb_group))
4655 return cpumask_first(nohz.ilb_grp_nohz_mask);
4657 ilb_group = ilb_group->next;
4659 } while (ilb_group != sd->groups);
4663 return cpumask_first(nohz.cpu_mask);
4665 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4666 static inline int find_new_ilb(int call_cpu)
4668 return cpumask_first(nohz.cpu_mask);
4673 * This routine will try to nominate the ilb (idle load balancing)
4674 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4675 * load balancing on behalf of all those cpus. If all the cpus in the system
4676 * go into this tickless mode, then there will be no ilb owner (as there is
4677 * no need for one) and all the cpus will sleep till the next wakeup event
4680 * For the ilb owner, tick is not stopped. And this tick will be used
4681 * for idle load balancing. ilb owner will still be part of
4684 * While stopping the tick, this cpu will become the ilb owner if there
4685 * is no other owner. And will be the owner till that cpu becomes busy
4686 * or if all cpus in the system stop their ticks at which point
4687 * there is no need for ilb owner.
4689 * When the ilb owner becomes busy, it nominates another owner, during the
4690 * next busy scheduler_tick()
4692 int select_nohz_load_balancer(int stop_tick)
4694 int cpu = smp_processor_id();
4697 cpu_rq(cpu)->in_nohz_recently = 1;
4699 if (!cpu_active(cpu)) {
4700 if (atomic_read(&nohz.load_balancer) != cpu)
4704 * If we are going offline and still the leader,
4707 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4713 cpumask_set_cpu(cpu, nohz.cpu_mask);
4715 /* time for ilb owner also to sleep */
4716 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4717 if (atomic_read(&nohz.load_balancer) == cpu)
4718 atomic_set(&nohz.load_balancer, -1);
4722 if (atomic_read(&nohz.load_balancer) == -1) {
4723 /* make me the ilb owner */
4724 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4726 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4729 if (!(sched_smt_power_savings ||
4730 sched_mc_power_savings))
4733 * Check to see if there is a more power-efficient
4736 new_ilb = find_new_ilb(cpu);
4737 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4738 atomic_set(&nohz.load_balancer, -1);
4739 resched_cpu(new_ilb);
4745 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4748 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4750 if (atomic_read(&nohz.load_balancer) == cpu)
4751 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4758 static DEFINE_SPINLOCK(balancing);
4761 * It checks each scheduling domain to see if it is due to be balanced,
4762 * and initiates a balancing operation if so.
4764 * Balancing parameters are set up in arch_init_sched_domains.
4766 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4769 struct rq *rq = cpu_rq(cpu);
4770 unsigned long interval;
4771 struct sched_domain *sd;
4772 /* Earliest time when we have to do rebalance again */
4773 unsigned long next_balance = jiffies + 60*HZ;
4774 int update_next_balance = 0;
4777 for_each_domain(cpu, sd) {
4778 if (!(sd->flags & SD_LOAD_BALANCE))
4781 interval = sd->balance_interval;
4782 if (idle != CPU_IDLE)
4783 interval *= sd->busy_factor;
4785 /* scale ms to jiffies */
4786 interval = msecs_to_jiffies(interval);
4787 if (unlikely(!interval))
4789 if (interval > HZ*NR_CPUS/10)
4790 interval = HZ*NR_CPUS/10;
4792 need_serialize = sd->flags & SD_SERIALIZE;
4794 if (need_serialize) {
4795 if (!spin_trylock(&balancing))
4799 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4800 if (load_balance(cpu, rq, sd, idle, &balance)) {
4802 * We've pulled tasks over so either we're no
4803 * longer idle, or one of our SMT siblings is
4806 idle = CPU_NOT_IDLE;
4808 sd->last_balance = jiffies;
4811 spin_unlock(&balancing);
4813 if (time_after(next_balance, sd->last_balance + interval)) {
4814 next_balance = sd->last_balance + interval;
4815 update_next_balance = 1;
4819 * Stop the load balance at this level. There is another
4820 * CPU in our sched group which is doing load balancing more
4828 * next_balance will be updated only when there is a need.
4829 * When the cpu is attached to null domain for ex, it will not be
4832 if (likely(update_next_balance))
4833 rq->next_balance = next_balance;
4837 * run_rebalance_domains is triggered when needed from the scheduler tick.
4838 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4839 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4841 static void run_rebalance_domains(struct softirq_action *h)
4843 int this_cpu = smp_processor_id();
4844 struct rq *this_rq = cpu_rq(this_cpu);
4845 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4846 CPU_IDLE : CPU_NOT_IDLE;
4848 rebalance_domains(this_cpu, idle);
4852 * If this cpu is the owner for idle load balancing, then do the
4853 * balancing on behalf of the other idle cpus whose ticks are
4856 if (this_rq->idle_at_tick &&
4857 atomic_read(&nohz.load_balancer) == this_cpu) {
4861 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4862 if (balance_cpu == this_cpu)
4866 * If this cpu gets work to do, stop the load balancing
4867 * work being done for other cpus. Next load
4868 * balancing owner will pick it up.
4873 rebalance_domains(balance_cpu, CPU_IDLE);
4875 rq = cpu_rq(balance_cpu);
4876 if (time_after(this_rq->next_balance, rq->next_balance))
4877 this_rq->next_balance = rq->next_balance;
4883 static inline int on_null_domain(int cpu)
4885 return !rcu_dereference(cpu_rq(cpu)->sd);
4889 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4891 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4892 * idle load balancing owner or decide to stop the periodic load balancing,
4893 * if the whole system is idle.
4895 static inline void trigger_load_balance(struct rq *rq, int cpu)
4899 * If we were in the nohz mode recently and busy at the current
4900 * scheduler tick, then check if we need to nominate new idle
4903 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4904 rq->in_nohz_recently = 0;
4906 if (atomic_read(&nohz.load_balancer) == cpu) {
4907 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4908 atomic_set(&nohz.load_balancer, -1);
4911 if (atomic_read(&nohz.load_balancer) == -1) {
4912 int ilb = find_new_ilb(cpu);
4914 if (ilb < nr_cpu_ids)
4920 * If this cpu is idle and doing idle load balancing for all the
4921 * cpus with ticks stopped, is it time for that to stop?
4923 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4924 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4930 * If this cpu is idle and the idle load balancing is done by
4931 * someone else, then no need raise the SCHED_SOFTIRQ
4933 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4934 cpumask_test_cpu(cpu, nohz.cpu_mask))
4937 /* Don't need to rebalance while attached to NULL domain */
4938 if (time_after_eq(jiffies, rq->next_balance) &&
4939 likely(!on_null_domain(cpu)))
4940 raise_softirq(SCHED_SOFTIRQ);
4943 #else /* CONFIG_SMP */
4946 * on UP we do not need to balance between CPUs:
4948 static inline void idle_balance(int cpu, struct rq *rq)
4954 DEFINE_PER_CPU(struct kernel_stat, kstat);
4956 EXPORT_PER_CPU_SYMBOL(kstat);
4959 * Return any ns on the sched_clock that have not yet been accounted in
4960 * @p in case that task is currently running.
4962 * Called with task_rq_lock() held on @rq.
4964 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4968 if (task_current(rq, p)) {
4969 update_rq_clock(rq);
4970 ns = rq->clock - p->se.exec_start;
4978 unsigned long long task_delta_exec(struct task_struct *p)
4980 unsigned long flags;
4984 rq = task_rq_lock(p, &flags);
4985 ns = do_task_delta_exec(p, rq);
4986 task_rq_unlock(rq, &flags);
4992 * Return accounted runtime for the task.
4993 * In case the task is currently running, return the runtime plus current's
4994 * pending runtime that have not been accounted yet.
4996 unsigned long long task_sched_runtime(struct task_struct *p)
4998 unsigned long flags;
5002 rq = task_rq_lock(p, &flags);
5003 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5004 task_rq_unlock(rq, &flags);
5010 * Return sum_exec_runtime for the thread group.
5011 * In case the task is currently running, return the sum plus current's
5012 * pending runtime that have not been accounted yet.
5014 * Note that the thread group might have other running tasks as well,
5015 * so the return value not includes other pending runtime that other
5016 * running tasks might have.
5018 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5020 struct task_cputime totals;
5021 unsigned long flags;
5025 rq = task_rq_lock(p, &flags);
5026 thread_group_cputime(p, &totals);
5027 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5028 task_rq_unlock(rq, &flags);
5034 * Account user cpu time to a process.
5035 * @p: the process that the cpu time gets accounted to
5036 * @cputime: the cpu time spent in user space since the last update
5037 * @cputime_scaled: cputime scaled by cpu frequency
5039 void account_user_time(struct task_struct *p, cputime_t cputime,
5040 cputime_t cputime_scaled)
5042 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5045 /* Add user time to process. */
5046 p->utime = cputime_add(p->utime, cputime);
5047 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5048 account_group_user_time(p, cputime);
5050 /* Add user time to cpustat. */
5051 tmp = cputime_to_cputime64(cputime);
5052 if (TASK_NICE(p) > 0)
5053 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5055 cpustat->user = cputime64_add(cpustat->user, tmp);
5057 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5058 /* Account for user time used */
5059 acct_update_integrals(p);
5063 * Account guest cpu time to a process.
5064 * @p: the process that the cpu time gets accounted to
5065 * @cputime: the cpu time spent in virtual machine since the last update
5066 * @cputime_scaled: cputime scaled by cpu frequency
5068 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5069 cputime_t cputime_scaled)
5072 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5074 tmp = cputime_to_cputime64(cputime);
5076 /* Add guest time to process. */
5077 p->utime = cputime_add(p->utime, cputime);
5078 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5079 account_group_user_time(p, cputime);
5080 p->gtime = cputime_add(p->gtime, cputime);
5082 /* Add guest time to cpustat. */
5083 if (TASK_NICE(p) > 0) {
5084 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5085 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5087 cpustat->user = cputime64_add(cpustat->user, tmp);
5088 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5093 * Account system cpu time to a process.
5094 * @p: the process that the cpu time gets accounted to
5095 * @hardirq_offset: the offset to subtract from hardirq_count()
5096 * @cputime: the cpu time spent in kernel space since the last update
5097 * @cputime_scaled: cputime scaled by cpu frequency
5099 void account_system_time(struct task_struct *p, int hardirq_offset,
5100 cputime_t cputime, cputime_t cputime_scaled)
5102 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5105 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5106 account_guest_time(p, cputime, cputime_scaled);
5110 /* Add system time to process. */
5111 p->stime = cputime_add(p->stime, cputime);
5112 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5113 account_group_system_time(p, cputime);
5115 /* Add system time to cpustat. */
5116 tmp = cputime_to_cputime64(cputime);
5117 if (hardirq_count() - hardirq_offset)
5118 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5119 else if (softirq_count())
5120 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5122 cpustat->system = cputime64_add(cpustat->system, tmp);
5124 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5126 /* Account for system time used */
5127 acct_update_integrals(p);
5131 * Account for involuntary wait time.
5132 * @steal: the cpu time spent in involuntary wait
5134 void account_steal_time(cputime_t cputime)
5136 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5137 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5139 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5143 * Account for idle time.
5144 * @cputime: the cpu time spent in idle wait
5146 void account_idle_time(cputime_t cputime)
5148 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5149 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5150 struct rq *rq = this_rq();
5152 if (atomic_read(&rq->nr_iowait) > 0)
5153 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5155 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5158 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5161 * Account a single tick of cpu time.
5162 * @p: the process that the cpu time gets accounted to
5163 * @user_tick: indicates if the tick is a user or a system tick
5165 void account_process_tick(struct task_struct *p, int user_tick)
5167 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5168 struct rq *rq = this_rq();
5171 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5172 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5173 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5176 account_idle_time(cputime_one_jiffy);
5180 * Account multiple ticks of steal time.
5181 * @p: the process from which the cpu time has been stolen
5182 * @ticks: number of stolen ticks
5184 void account_steal_ticks(unsigned long ticks)
5186 account_steal_time(jiffies_to_cputime(ticks));
5190 * Account multiple ticks of idle time.
5191 * @ticks: number of stolen ticks
5193 void account_idle_ticks(unsigned long ticks)
5195 account_idle_time(jiffies_to_cputime(ticks));
5201 * Use precise platform statistics if available:
5203 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5204 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5210 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5212 struct task_cputime cputime;
5214 thread_group_cputime(p, &cputime);
5216 *ut = cputime.utime;
5217 *st = cputime.stime;
5221 #ifndef nsecs_to_cputime
5222 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5225 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5227 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5230 * Use CFS's precise accounting:
5232 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5237 temp = (u64)(rtime * utime);
5238 do_div(temp, total);
5239 utime = (cputime_t)temp;
5244 * Compare with previous values, to keep monotonicity:
5246 p->prev_utime = max(p->prev_utime, utime);
5247 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5249 *ut = p->prev_utime;
5250 *st = p->prev_stime;
5254 * Must be called with siglock held.
5256 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5258 struct signal_struct *sig = p->signal;
5259 struct task_cputime cputime;
5260 cputime_t rtime, utime, total;
5262 thread_group_cputime(p, &cputime);
5264 total = cputime_add(cputime.utime, cputime.stime);
5265 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5270 temp = (u64)(rtime * cputime.utime);
5271 do_div(temp, total);
5272 utime = (cputime_t)temp;
5276 sig->prev_utime = max(sig->prev_utime, utime);
5277 sig->prev_stime = max(sig->prev_stime,
5278 cputime_sub(rtime, sig->prev_utime));
5280 *ut = sig->prev_utime;
5281 *st = sig->prev_stime;
5286 * This function gets called by the timer code, with HZ frequency.
5287 * We call it with interrupts disabled.
5289 * It also gets called by the fork code, when changing the parent's
5292 void scheduler_tick(void)
5294 int cpu = smp_processor_id();
5295 struct rq *rq = cpu_rq(cpu);
5296 struct task_struct *curr = rq->curr;
5300 raw_spin_lock(&rq->lock);
5301 update_rq_clock(rq);
5302 update_cpu_load(rq);
5303 curr->sched_class->task_tick(rq, curr, 0);
5304 raw_spin_unlock(&rq->lock);
5306 perf_event_task_tick(curr, cpu);
5309 rq->idle_at_tick = idle_cpu(cpu);
5310 trigger_load_balance(rq, cpu);
5314 notrace unsigned long get_parent_ip(unsigned long addr)
5316 if (in_lock_functions(addr)) {
5317 addr = CALLER_ADDR2;
5318 if (in_lock_functions(addr))
5319 addr = CALLER_ADDR3;
5324 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5325 defined(CONFIG_PREEMPT_TRACER))
5327 void __kprobes add_preempt_count(int val)
5329 #ifdef CONFIG_DEBUG_PREEMPT
5333 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5336 preempt_count() += val;
5337 #ifdef CONFIG_DEBUG_PREEMPT
5339 * Spinlock count overflowing soon?
5341 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5344 if (preempt_count() == val)
5345 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5347 EXPORT_SYMBOL(add_preempt_count);
5349 void __kprobes sub_preempt_count(int val)
5351 #ifdef CONFIG_DEBUG_PREEMPT
5355 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5358 * Is the spinlock portion underflowing?
5360 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5361 !(preempt_count() & PREEMPT_MASK)))
5365 if (preempt_count() == val)
5366 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5367 preempt_count() -= val;
5369 EXPORT_SYMBOL(sub_preempt_count);
5374 * Print scheduling while atomic bug:
5376 static noinline void __schedule_bug(struct task_struct *prev)
5378 struct pt_regs *regs = get_irq_regs();
5380 pr_err("BUG: scheduling while atomic: %s/%d/0x%08x\n",
5381 prev->comm, prev->pid, preempt_count());
5383 debug_show_held_locks(prev);
5385 if (irqs_disabled())
5386 print_irqtrace_events(prev);
5395 * Various schedule()-time debugging checks and statistics:
5397 static inline void schedule_debug(struct task_struct *prev)
5400 * Test if we are atomic. Since do_exit() needs to call into
5401 * schedule() atomically, we ignore that path for now.
5402 * Otherwise, whine if we are scheduling when we should not be.
5404 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5405 __schedule_bug(prev);
5407 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5409 schedstat_inc(this_rq(), sched_count);
5410 #ifdef CONFIG_SCHEDSTATS
5411 if (unlikely(prev->lock_depth >= 0)) {
5412 schedstat_inc(this_rq(), bkl_count);
5413 schedstat_inc(prev, sched_info.bkl_count);
5418 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5420 if (prev->state == TASK_RUNNING) {
5421 u64 runtime = prev->se.sum_exec_runtime;
5423 runtime -= prev->se.prev_sum_exec_runtime;
5424 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5427 * In order to avoid avg_overlap growing stale when we are
5428 * indeed overlapping and hence not getting put to sleep, grow
5429 * the avg_overlap on preemption.
5431 * We use the average preemption runtime because that
5432 * correlates to the amount of cache footprint a task can
5435 update_avg(&prev->se.avg_overlap, runtime);
5437 prev->sched_class->put_prev_task(rq, prev);
5441 * Pick up the highest-prio task:
5443 static inline struct task_struct *
5444 pick_next_task(struct rq *rq)
5446 const struct sched_class *class;
5447 struct task_struct *p;
5450 * Optimization: we know that if all tasks are in
5451 * the fair class we can call that function directly:
5453 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5454 p = fair_sched_class.pick_next_task(rq);
5459 class = sched_class_highest;
5461 p = class->pick_next_task(rq);
5465 * Will never be NULL as the idle class always
5466 * returns a non-NULL p:
5468 class = class->next;
5473 * schedule() is the main scheduler function.
5475 asmlinkage void __sched schedule(void)
5477 struct task_struct *prev, *next;
5478 unsigned long *switch_count;
5484 cpu = smp_processor_id();
5488 switch_count = &prev->nivcsw;
5490 release_kernel_lock(prev);
5491 need_resched_nonpreemptible:
5493 schedule_debug(prev);
5495 if (sched_feat(HRTICK))
5498 raw_spin_lock_irq(&rq->lock);
5499 update_rq_clock(rq);
5500 clear_tsk_need_resched(prev);
5502 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5503 if (unlikely(signal_pending_state(prev->state, prev)))
5504 prev->state = TASK_RUNNING;
5506 deactivate_task(rq, prev, 1);
5507 switch_count = &prev->nvcsw;
5510 pre_schedule(rq, prev);
5512 if (unlikely(!rq->nr_running))
5513 idle_balance(cpu, rq);
5515 put_prev_task(rq, prev);
5516 next = pick_next_task(rq);
5518 if (likely(prev != next)) {
5519 sched_info_switch(prev, next);
5520 perf_event_task_sched_out(prev, next, cpu);
5526 context_switch(rq, prev, next); /* unlocks the rq */
5528 * the context switch might have flipped the stack from under
5529 * us, hence refresh the local variables.
5531 cpu = smp_processor_id();
5534 raw_spin_unlock_irq(&rq->lock);
5538 if (unlikely(reacquire_kernel_lock(current) < 0))
5539 goto need_resched_nonpreemptible;
5541 preempt_enable_no_resched();
5545 EXPORT_SYMBOL(schedule);
5547 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5549 * Look out! "owner" is an entirely speculative pointer
5550 * access and not reliable.
5552 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5557 if (!sched_feat(OWNER_SPIN))
5560 #ifdef CONFIG_DEBUG_PAGEALLOC
5562 * Need to access the cpu field knowing that
5563 * DEBUG_PAGEALLOC could have unmapped it if
5564 * the mutex owner just released it and exited.
5566 if (probe_kernel_address(&owner->cpu, cpu))
5573 * Even if the access succeeded (likely case),
5574 * the cpu field may no longer be valid.
5576 if (cpu >= nr_cpumask_bits)
5580 * We need to validate that we can do a
5581 * get_cpu() and that we have the percpu area.
5583 if (!cpu_online(cpu))
5590 * Owner changed, break to re-assess state.
5592 if (lock->owner != owner)
5596 * Is that owner really running on that cpu?
5598 if (task_thread_info(rq->curr) != owner || need_resched())
5608 #ifdef CONFIG_PREEMPT
5610 * this is the entry point to schedule() from in-kernel preemption
5611 * off of preempt_enable. Kernel preemptions off return from interrupt
5612 * occur there and call schedule directly.
5614 asmlinkage void __sched preempt_schedule(void)
5616 struct thread_info *ti = current_thread_info();
5619 * If there is a non-zero preempt_count or interrupts are disabled,
5620 * we do not want to preempt the current task. Just return..
5622 if (likely(ti->preempt_count || irqs_disabled()))
5626 add_preempt_count(PREEMPT_ACTIVE);
5628 sub_preempt_count(PREEMPT_ACTIVE);
5631 * Check again in case we missed a preemption opportunity
5632 * between schedule and now.
5635 } while (need_resched());
5637 EXPORT_SYMBOL(preempt_schedule);
5640 * this is the entry point to schedule() from kernel preemption
5641 * off of irq context.
5642 * Note, that this is called and return with irqs disabled. This will
5643 * protect us against recursive calling from irq.
5645 asmlinkage void __sched preempt_schedule_irq(void)
5647 struct thread_info *ti = current_thread_info();
5649 /* Catch callers which need to be fixed */
5650 BUG_ON(ti->preempt_count || !irqs_disabled());
5653 add_preempt_count(PREEMPT_ACTIVE);
5656 local_irq_disable();
5657 sub_preempt_count(PREEMPT_ACTIVE);
5660 * Check again in case we missed a preemption opportunity
5661 * between schedule and now.
5664 } while (need_resched());
5667 #endif /* CONFIG_PREEMPT */
5669 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5672 return try_to_wake_up(curr->private, mode, wake_flags);
5674 EXPORT_SYMBOL(default_wake_function);
5677 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5678 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5679 * number) then we wake all the non-exclusive tasks and one exclusive task.
5681 * There are circumstances in which we can try to wake a task which has already
5682 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5683 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5685 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5686 int nr_exclusive, int wake_flags, void *key)
5688 wait_queue_t *curr, *next;
5690 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5691 unsigned flags = curr->flags;
5693 if (curr->func(curr, mode, wake_flags, key) &&
5694 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5700 * __wake_up - wake up threads blocked on a waitqueue.
5702 * @mode: which threads
5703 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5704 * @key: is directly passed to the wakeup function
5706 * It may be assumed that this function implies a write memory barrier before
5707 * changing the task state if and only if any tasks are woken up.
5709 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5710 int nr_exclusive, void *key)
5712 unsigned long flags;
5714 spin_lock_irqsave(&q->lock, flags);
5715 __wake_up_common(q, mode, nr_exclusive, 0, key);
5716 spin_unlock_irqrestore(&q->lock, flags);
5718 EXPORT_SYMBOL(__wake_up);
5721 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5723 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5725 __wake_up_common(q, mode, 1, 0, NULL);
5728 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5730 __wake_up_common(q, mode, 1, 0, key);
5734 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5736 * @mode: which threads
5737 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5738 * @key: opaque value to be passed to wakeup targets
5740 * The sync wakeup differs that the waker knows that it will schedule
5741 * away soon, so while the target thread will be woken up, it will not
5742 * be migrated to another CPU - ie. the two threads are 'synchronized'
5743 * with each other. This can prevent needless bouncing between CPUs.
5745 * On UP it can prevent extra preemption.
5747 * It may be assumed that this function implies a write memory barrier before
5748 * changing the task state if and only if any tasks are woken up.
5750 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5751 int nr_exclusive, void *key)
5753 unsigned long flags;
5754 int wake_flags = WF_SYNC;
5759 if (unlikely(!nr_exclusive))
5762 spin_lock_irqsave(&q->lock, flags);
5763 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5764 spin_unlock_irqrestore(&q->lock, flags);
5766 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5769 * __wake_up_sync - see __wake_up_sync_key()
5771 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5773 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5775 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5778 * complete: - signals a single thread waiting on this completion
5779 * @x: holds the state of this particular completion
5781 * This will wake up a single thread waiting on this completion. Threads will be
5782 * awakened in the same order in which they were queued.
5784 * See also complete_all(), wait_for_completion() and related routines.
5786 * It may be assumed that this function implies a write memory barrier before
5787 * changing the task state if and only if any tasks are woken up.
5789 void complete(struct completion *x)
5791 unsigned long flags;
5793 spin_lock_irqsave(&x->wait.lock, flags);
5795 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5796 spin_unlock_irqrestore(&x->wait.lock, flags);
5798 EXPORT_SYMBOL(complete);
5801 * complete_all: - signals all threads waiting on this completion
5802 * @x: holds the state of this particular completion
5804 * This will wake up all threads waiting on this particular completion event.
5806 * It may be assumed that this function implies a write memory barrier before
5807 * changing the task state if and only if any tasks are woken up.
5809 void complete_all(struct completion *x)
5811 unsigned long flags;
5813 spin_lock_irqsave(&x->wait.lock, flags);
5814 x->done += UINT_MAX/2;
5815 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5816 spin_unlock_irqrestore(&x->wait.lock, flags);
5818 EXPORT_SYMBOL(complete_all);
5820 static inline long __sched
5821 do_wait_for_common(struct completion *x, long timeout, int state)
5824 DECLARE_WAITQUEUE(wait, current);
5826 wait.flags |= WQ_FLAG_EXCLUSIVE;
5827 __add_wait_queue_tail(&x->wait, &wait);
5829 if (signal_pending_state(state, current)) {
5830 timeout = -ERESTARTSYS;
5833 __set_current_state(state);
5834 spin_unlock_irq(&x->wait.lock);
5835 timeout = schedule_timeout(timeout);
5836 spin_lock_irq(&x->wait.lock);
5837 } while (!x->done && timeout);
5838 __remove_wait_queue(&x->wait, &wait);
5843 return timeout ?: 1;
5847 wait_for_common(struct completion *x, long timeout, int state)
5851 spin_lock_irq(&x->wait.lock);
5852 timeout = do_wait_for_common(x, timeout, state);
5853 spin_unlock_irq(&x->wait.lock);
5858 * wait_for_completion: - waits for completion of a task
5859 * @x: holds the state of this particular completion
5861 * This waits to be signaled for completion of a specific task. It is NOT
5862 * interruptible and there is no timeout.
5864 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5865 * and interrupt capability. Also see complete().
5867 void __sched wait_for_completion(struct completion *x)
5869 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5871 EXPORT_SYMBOL(wait_for_completion);
5874 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5875 * @x: holds the state of this particular completion
5876 * @timeout: timeout value in jiffies
5878 * This waits for either a completion of a specific task to be signaled or for a
5879 * specified timeout to expire. The timeout is in jiffies. It is not
5882 unsigned long __sched
5883 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5885 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5887 EXPORT_SYMBOL(wait_for_completion_timeout);
5890 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5891 * @x: holds the state of this particular completion
5893 * This waits for completion of a specific task to be signaled. It is
5896 int __sched wait_for_completion_interruptible(struct completion *x)
5898 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5899 if (t == -ERESTARTSYS)
5903 EXPORT_SYMBOL(wait_for_completion_interruptible);
5906 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5907 * @x: holds the state of this particular completion
5908 * @timeout: timeout value in jiffies
5910 * This waits for either a completion of a specific task to be signaled or for a
5911 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5913 unsigned long __sched
5914 wait_for_completion_interruptible_timeout(struct completion *x,
5915 unsigned long timeout)
5917 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5919 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5922 * wait_for_completion_killable: - waits for completion of a task (killable)
5923 * @x: holds the state of this particular completion
5925 * This waits to be signaled for completion of a specific task. It can be
5926 * interrupted by a kill signal.
5928 int __sched wait_for_completion_killable(struct completion *x)
5930 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5931 if (t == -ERESTARTSYS)
5935 EXPORT_SYMBOL(wait_for_completion_killable);
5938 * try_wait_for_completion - try to decrement a completion without blocking
5939 * @x: completion structure
5941 * Returns: 0 if a decrement cannot be done without blocking
5942 * 1 if a decrement succeeded.
5944 * If a completion is being used as a counting completion,
5945 * attempt to decrement the counter without blocking. This
5946 * enables us to avoid waiting if the resource the completion
5947 * is protecting is not available.
5949 bool try_wait_for_completion(struct completion *x)
5951 unsigned long flags;
5954 spin_lock_irqsave(&x->wait.lock, flags);
5959 spin_unlock_irqrestore(&x->wait.lock, flags);
5962 EXPORT_SYMBOL(try_wait_for_completion);
5965 * completion_done - Test to see if a completion has any waiters
5966 * @x: completion structure
5968 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5969 * 1 if there are no waiters.
5972 bool completion_done(struct completion *x)
5974 unsigned long flags;
5977 spin_lock_irqsave(&x->wait.lock, flags);
5980 spin_unlock_irqrestore(&x->wait.lock, flags);
5983 EXPORT_SYMBOL(completion_done);
5986 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5988 unsigned long flags;
5991 init_waitqueue_entry(&wait, current);
5993 __set_current_state(state);
5995 spin_lock_irqsave(&q->lock, flags);
5996 __add_wait_queue(q, &wait);
5997 spin_unlock(&q->lock);
5998 timeout = schedule_timeout(timeout);
5999 spin_lock_irq(&q->lock);
6000 __remove_wait_queue(q, &wait);
6001 spin_unlock_irqrestore(&q->lock, flags);
6006 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6008 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6010 EXPORT_SYMBOL(interruptible_sleep_on);
6013 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6015 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6017 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6019 void __sched sleep_on(wait_queue_head_t *q)
6021 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6023 EXPORT_SYMBOL(sleep_on);
6025 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6027 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6029 EXPORT_SYMBOL(sleep_on_timeout);
6031 #ifdef CONFIG_RT_MUTEXES
6034 * rt_mutex_setprio - set the current priority of a task
6036 * @prio: prio value (kernel-internal form)
6038 * This function changes the 'effective' priority of a task. It does
6039 * not touch ->normal_prio like __setscheduler().
6041 * Used by the rt_mutex code to implement priority inheritance logic.
6043 void rt_mutex_setprio(struct task_struct *p, int prio)
6045 unsigned long flags;
6046 int oldprio, on_rq, running;
6048 const struct sched_class *prev_class = p->sched_class;
6050 BUG_ON(prio < 0 || prio > MAX_PRIO);
6052 rq = task_rq_lock(p, &flags);
6053 update_rq_clock(rq);
6056 on_rq = p->se.on_rq;
6057 running = task_current(rq, p);
6059 dequeue_task(rq, p, 0);
6061 p->sched_class->put_prev_task(rq, p);
6064 p->sched_class = &rt_sched_class;
6066 p->sched_class = &fair_sched_class;
6071 p->sched_class->set_curr_task(rq);
6073 enqueue_task(rq, p, 0);
6075 check_class_changed(rq, p, prev_class, oldprio, running);
6077 task_rq_unlock(rq, &flags);
6082 void set_user_nice(struct task_struct *p, long nice)
6084 int old_prio, delta, on_rq;
6085 unsigned long flags;
6088 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6091 * We have to be careful, if called from sys_setpriority(),
6092 * the task might be in the middle of scheduling on another CPU.
6094 rq = task_rq_lock(p, &flags);
6095 update_rq_clock(rq);
6097 * The RT priorities are set via sched_setscheduler(), but we still
6098 * allow the 'normal' nice value to be set - but as expected
6099 * it wont have any effect on scheduling until the task is
6100 * SCHED_FIFO/SCHED_RR:
6102 if (task_has_rt_policy(p)) {
6103 p->static_prio = NICE_TO_PRIO(nice);
6106 on_rq = p->se.on_rq;
6108 dequeue_task(rq, p, 0);
6110 p->static_prio = NICE_TO_PRIO(nice);
6113 p->prio = effective_prio(p);
6114 delta = p->prio - old_prio;
6117 enqueue_task(rq, p, 0);
6119 * If the task increased its priority or is running and
6120 * lowered its priority, then reschedule its CPU:
6122 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6123 resched_task(rq->curr);
6126 task_rq_unlock(rq, &flags);
6128 EXPORT_SYMBOL(set_user_nice);
6131 * can_nice - check if a task can reduce its nice value
6135 int can_nice(const struct task_struct *p, const int nice)
6137 /* convert nice value [19,-20] to rlimit style value [1,40] */
6138 int nice_rlim = 20 - nice;
6140 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6141 capable(CAP_SYS_NICE));
6144 #ifdef __ARCH_WANT_SYS_NICE
6147 * sys_nice - change the priority of the current process.
6148 * @increment: priority increment
6150 * sys_setpriority is a more generic, but much slower function that
6151 * does similar things.
6153 SYSCALL_DEFINE1(nice, int, increment)
6158 * Setpriority might change our priority at the same moment.
6159 * We don't have to worry. Conceptually one call occurs first
6160 * and we have a single winner.
6162 if (increment < -40)
6167 nice = TASK_NICE(current) + increment;
6173 if (increment < 0 && !can_nice(current, nice))
6176 retval = security_task_setnice(current, nice);
6180 set_user_nice(current, nice);
6187 * task_prio - return the priority value of a given task.
6188 * @p: the task in question.
6190 * This is the priority value as seen by users in /proc.
6191 * RT tasks are offset by -200. Normal tasks are centered
6192 * around 0, value goes from -16 to +15.
6194 int task_prio(const struct task_struct *p)
6196 return p->prio - MAX_RT_PRIO;
6200 * task_nice - return the nice value of a given task.
6201 * @p: the task in question.
6203 int task_nice(const struct task_struct *p)
6205 return TASK_NICE(p);
6207 EXPORT_SYMBOL(task_nice);
6210 * idle_cpu - is a given cpu idle currently?
6211 * @cpu: the processor in question.
6213 int idle_cpu(int cpu)
6215 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6219 * idle_task - return the idle task for a given cpu.
6220 * @cpu: the processor in question.
6222 struct task_struct *idle_task(int cpu)
6224 return cpu_rq(cpu)->idle;
6228 * find_process_by_pid - find a process with a matching PID value.
6229 * @pid: the pid in question.
6231 static struct task_struct *find_process_by_pid(pid_t pid)
6233 return pid ? find_task_by_vpid(pid) : current;
6236 /* Actually do priority change: must hold rq lock. */
6238 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6240 BUG_ON(p->se.on_rq);
6243 p->rt_priority = prio;
6244 p->normal_prio = normal_prio(p);
6245 /* we are holding p->pi_lock already */
6246 p->prio = rt_mutex_getprio(p);
6247 if (rt_prio(p->prio))
6248 p->sched_class = &rt_sched_class;
6250 p->sched_class = &fair_sched_class;
6255 * check the target process has a UID that matches the current process's
6257 static bool check_same_owner(struct task_struct *p)
6259 const struct cred *cred = current_cred(), *pcred;
6263 pcred = __task_cred(p);
6264 match = (cred->euid == pcred->euid ||
6265 cred->euid == pcred->uid);
6270 static int __sched_setscheduler(struct task_struct *p, int policy,
6271 struct sched_param *param, bool user)
6273 int retval, oldprio, oldpolicy = -1, on_rq, running;
6274 unsigned long flags;
6275 const struct sched_class *prev_class = p->sched_class;
6279 /* may grab non-irq protected spin_locks */
6280 BUG_ON(in_interrupt());
6282 /* double check policy once rq lock held */
6284 reset_on_fork = p->sched_reset_on_fork;
6285 policy = oldpolicy = p->policy;
6287 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6288 policy &= ~SCHED_RESET_ON_FORK;
6290 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6291 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6292 policy != SCHED_IDLE)
6297 * Valid priorities for SCHED_FIFO and SCHED_RR are
6298 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6299 * SCHED_BATCH and SCHED_IDLE is 0.
6301 if (param->sched_priority < 0 ||
6302 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6303 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6305 if (rt_policy(policy) != (param->sched_priority != 0))
6309 * Allow unprivileged RT tasks to decrease priority:
6311 if (user && !capable(CAP_SYS_NICE)) {
6312 if (rt_policy(policy)) {
6313 unsigned long rlim_rtprio;
6315 if (!lock_task_sighand(p, &flags))
6317 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6318 unlock_task_sighand(p, &flags);
6320 /* can't set/change the rt policy */
6321 if (policy != p->policy && !rlim_rtprio)
6324 /* can't increase priority */
6325 if (param->sched_priority > p->rt_priority &&
6326 param->sched_priority > rlim_rtprio)
6330 * Like positive nice levels, dont allow tasks to
6331 * move out of SCHED_IDLE either:
6333 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6336 /* can't change other user's priorities */
6337 if (!check_same_owner(p))
6340 /* Normal users shall not reset the sched_reset_on_fork flag */
6341 if (p->sched_reset_on_fork && !reset_on_fork)
6346 #ifdef CONFIG_RT_GROUP_SCHED
6348 * Do not allow realtime tasks into groups that have no runtime
6351 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6352 task_group(p)->rt_bandwidth.rt_runtime == 0)
6356 retval = security_task_setscheduler(p, policy, param);
6362 * make sure no PI-waiters arrive (or leave) while we are
6363 * changing the priority of the task:
6365 raw_spin_lock_irqsave(&p->pi_lock, flags);
6367 * To be able to change p->policy safely, the apropriate
6368 * runqueue lock must be held.
6370 rq = __task_rq_lock(p);
6371 /* recheck policy now with rq lock held */
6372 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6373 policy = oldpolicy = -1;
6374 __task_rq_unlock(rq);
6375 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6378 update_rq_clock(rq);
6379 on_rq = p->se.on_rq;
6380 running = task_current(rq, p);
6382 deactivate_task(rq, p, 0);
6384 p->sched_class->put_prev_task(rq, p);
6386 p->sched_reset_on_fork = reset_on_fork;
6389 __setscheduler(rq, p, policy, param->sched_priority);
6392 p->sched_class->set_curr_task(rq);
6394 activate_task(rq, p, 0);
6396 check_class_changed(rq, p, prev_class, oldprio, running);
6398 __task_rq_unlock(rq);
6399 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6401 rt_mutex_adjust_pi(p);
6407 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6408 * @p: the task in question.
6409 * @policy: new policy.
6410 * @param: structure containing the new RT priority.
6412 * NOTE that the task may be already dead.
6414 int sched_setscheduler(struct task_struct *p, int policy,
6415 struct sched_param *param)
6417 return __sched_setscheduler(p, policy, param, true);
6419 EXPORT_SYMBOL_GPL(sched_setscheduler);
6422 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6423 * @p: the task in question.
6424 * @policy: new policy.
6425 * @param: structure containing the new RT priority.
6427 * Just like sched_setscheduler, only don't bother checking if the
6428 * current context has permission. For example, this is needed in
6429 * stop_machine(): we create temporary high priority worker threads,
6430 * but our caller might not have that capability.
6432 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6433 struct sched_param *param)
6435 return __sched_setscheduler(p, policy, param, false);
6439 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6441 struct sched_param lparam;
6442 struct task_struct *p;
6445 if (!param || pid < 0)
6447 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6452 p = find_process_by_pid(pid);
6454 retval = sched_setscheduler(p, policy, &lparam);
6461 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6462 * @pid: the pid in question.
6463 * @policy: new policy.
6464 * @param: structure containing the new RT priority.
6466 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6467 struct sched_param __user *, param)
6469 /* negative values for policy are not valid */
6473 return do_sched_setscheduler(pid, policy, param);
6477 * sys_sched_setparam - set/change the RT priority of a thread
6478 * @pid: the pid in question.
6479 * @param: structure containing the new RT priority.
6481 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6483 return do_sched_setscheduler(pid, -1, param);
6487 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6488 * @pid: the pid in question.
6490 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6492 struct task_struct *p;
6500 p = find_process_by_pid(pid);
6502 retval = security_task_getscheduler(p);
6505 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6512 * sys_sched_getparam - get the RT priority of a thread
6513 * @pid: the pid in question.
6514 * @param: structure containing the RT priority.
6516 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6518 struct sched_param lp;
6519 struct task_struct *p;
6522 if (!param || pid < 0)
6526 p = find_process_by_pid(pid);
6531 retval = security_task_getscheduler(p);
6535 lp.sched_priority = p->rt_priority;
6539 * This one might sleep, we cannot do it with a spinlock held ...
6541 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6550 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6552 cpumask_var_t cpus_allowed, new_mask;
6553 struct task_struct *p;
6559 p = find_process_by_pid(pid);
6566 /* Prevent p going away */
6570 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6574 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6576 goto out_free_cpus_allowed;
6579 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6582 retval = security_task_setscheduler(p, 0, NULL);
6586 cpuset_cpus_allowed(p, cpus_allowed);
6587 cpumask_and(new_mask, in_mask, cpus_allowed);
6589 retval = set_cpus_allowed_ptr(p, new_mask);
6592 cpuset_cpus_allowed(p, cpus_allowed);
6593 if (!cpumask_subset(new_mask, cpus_allowed)) {
6595 * We must have raced with a concurrent cpuset
6596 * update. Just reset the cpus_allowed to the
6597 * cpuset's cpus_allowed
6599 cpumask_copy(new_mask, cpus_allowed);
6604 free_cpumask_var(new_mask);
6605 out_free_cpus_allowed:
6606 free_cpumask_var(cpus_allowed);
6613 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6614 struct cpumask *new_mask)
6616 if (len < cpumask_size())
6617 cpumask_clear(new_mask);
6618 else if (len > cpumask_size())
6619 len = cpumask_size();
6621 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6625 * sys_sched_setaffinity - set the cpu affinity of a process
6626 * @pid: pid of the process
6627 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6628 * @user_mask_ptr: user-space pointer to the new cpu mask
6630 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6631 unsigned long __user *, user_mask_ptr)
6633 cpumask_var_t new_mask;
6636 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6639 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6641 retval = sched_setaffinity(pid, new_mask);
6642 free_cpumask_var(new_mask);
6646 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6648 struct task_struct *p;
6649 unsigned long flags;
6657 p = find_process_by_pid(pid);
6661 retval = security_task_getscheduler(p);
6665 rq = task_rq_lock(p, &flags);
6666 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6667 task_rq_unlock(rq, &flags);
6677 * sys_sched_getaffinity - get the cpu affinity of a process
6678 * @pid: pid of the process
6679 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6680 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6682 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6683 unsigned long __user *, user_mask_ptr)
6688 if (len < cpumask_size())
6691 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6694 ret = sched_getaffinity(pid, mask);
6696 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6699 ret = cpumask_size();
6701 free_cpumask_var(mask);
6707 * sys_sched_yield - yield the current processor to other threads.
6709 * This function yields the current CPU to other tasks. If there are no
6710 * other threads running on this CPU then this function will return.
6712 SYSCALL_DEFINE0(sched_yield)
6714 struct rq *rq = this_rq_lock();
6716 schedstat_inc(rq, yld_count);
6717 current->sched_class->yield_task(rq);
6720 * Since we are going to call schedule() anyway, there's
6721 * no need to preempt or enable interrupts:
6723 __release(rq->lock);
6724 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6725 do_raw_spin_unlock(&rq->lock);
6726 preempt_enable_no_resched();
6733 static inline int should_resched(void)
6735 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6738 static void __cond_resched(void)
6740 add_preempt_count(PREEMPT_ACTIVE);
6742 sub_preempt_count(PREEMPT_ACTIVE);
6745 int __sched _cond_resched(void)
6747 if (should_resched()) {
6753 EXPORT_SYMBOL(_cond_resched);
6756 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6757 * call schedule, and on return reacquire the lock.
6759 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6760 * operations here to prevent schedule() from being called twice (once via
6761 * spin_unlock(), once by hand).
6763 int __cond_resched_lock(spinlock_t *lock)
6765 int resched = should_resched();
6768 lockdep_assert_held(lock);
6770 if (spin_needbreak(lock) || resched) {
6781 EXPORT_SYMBOL(__cond_resched_lock);
6783 int __sched __cond_resched_softirq(void)
6785 BUG_ON(!in_softirq());
6787 if (should_resched()) {
6795 EXPORT_SYMBOL(__cond_resched_softirq);
6798 * yield - yield the current processor to other threads.
6800 * This is a shortcut for kernel-space yielding - it marks the
6801 * thread runnable and calls sys_sched_yield().
6803 void __sched yield(void)
6805 set_current_state(TASK_RUNNING);
6808 EXPORT_SYMBOL(yield);
6811 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6812 * that process accounting knows that this is a task in IO wait state.
6814 void __sched io_schedule(void)
6816 struct rq *rq = raw_rq();
6818 delayacct_blkio_start();
6819 atomic_inc(&rq->nr_iowait);
6820 current->in_iowait = 1;
6822 current->in_iowait = 0;
6823 atomic_dec(&rq->nr_iowait);
6824 delayacct_blkio_end();
6826 EXPORT_SYMBOL(io_schedule);
6828 long __sched io_schedule_timeout(long timeout)
6830 struct rq *rq = raw_rq();
6833 delayacct_blkio_start();
6834 atomic_inc(&rq->nr_iowait);
6835 current->in_iowait = 1;
6836 ret = schedule_timeout(timeout);
6837 current->in_iowait = 0;
6838 atomic_dec(&rq->nr_iowait);
6839 delayacct_blkio_end();
6844 * sys_sched_get_priority_max - return maximum RT priority.
6845 * @policy: scheduling class.
6847 * this syscall returns the maximum rt_priority that can be used
6848 * by a given scheduling class.
6850 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6857 ret = MAX_USER_RT_PRIO-1;
6869 * sys_sched_get_priority_min - return minimum RT priority.
6870 * @policy: scheduling class.
6872 * this syscall returns the minimum rt_priority that can be used
6873 * by a given scheduling class.
6875 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6893 * sys_sched_rr_get_interval - return the default timeslice of a process.
6894 * @pid: pid of the process.
6895 * @interval: userspace pointer to the timeslice value.
6897 * this syscall writes the default timeslice value of a given process
6898 * into the user-space timespec buffer. A value of '0' means infinity.
6900 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6901 struct timespec __user *, interval)
6903 struct task_struct *p;
6904 unsigned int time_slice;
6905 unsigned long flags;
6915 p = find_process_by_pid(pid);
6919 retval = security_task_getscheduler(p);
6923 rq = task_rq_lock(p, &flags);
6924 time_slice = p->sched_class->get_rr_interval(rq, p);
6925 task_rq_unlock(rq, &flags);
6928 jiffies_to_timespec(time_slice, &t);
6929 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6937 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6939 void sched_show_task(struct task_struct *p)
6941 unsigned long free = 0;
6944 state = p->state ? __ffs(p->state) + 1 : 0;
6945 pr_info("%-13.13s %c", p->comm,
6946 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6947 #if BITS_PER_LONG == 32
6948 if (state == TASK_RUNNING)
6949 pr_cont(" running ");
6951 pr_cont(" %08lx ", thread_saved_pc(p));
6953 if (state == TASK_RUNNING)
6954 pr_cont(" running task ");
6956 pr_cont(" %016lx ", thread_saved_pc(p));
6958 #ifdef CONFIG_DEBUG_STACK_USAGE
6959 free = stack_not_used(p);
6961 pr_cont("%5lu %5d %6d 0x%08lx\n", free,
6962 task_pid_nr(p), task_pid_nr(p->real_parent),
6963 (unsigned long)task_thread_info(p)->flags);
6965 show_stack(p, NULL);
6968 void show_state_filter(unsigned long state_filter)
6970 struct task_struct *g, *p;
6972 #if BITS_PER_LONG == 32
6973 pr_info(" task PC stack pid father\n");
6975 pr_info(" task PC stack pid father\n");
6977 read_lock(&tasklist_lock);
6978 do_each_thread(g, p) {
6980 * reset the NMI-timeout, listing all files on a slow
6981 * console might take alot of time:
6983 touch_nmi_watchdog();
6984 if (!state_filter || (p->state & state_filter))
6986 } while_each_thread(g, p);
6988 touch_all_softlockup_watchdogs();
6990 #ifdef CONFIG_SCHED_DEBUG
6991 sysrq_sched_debug_show();
6993 read_unlock(&tasklist_lock);
6995 * Only show locks if all tasks are dumped:
6998 debug_show_all_locks();
7001 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7003 idle->sched_class = &idle_sched_class;
7007 * init_idle - set up an idle thread for a given CPU
7008 * @idle: task in question
7009 * @cpu: cpu the idle task belongs to
7011 * NOTE: this function does not set the idle thread's NEED_RESCHED
7012 * flag, to make booting more robust.
7014 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7016 struct rq *rq = cpu_rq(cpu);
7017 unsigned long flags;
7019 raw_spin_lock_irqsave(&rq->lock, flags);
7022 idle->state = TASK_RUNNING;
7023 idle->se.exec_start = sched_clock();
7025 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7026 __set_task_cpu(idle, cpu);
7028 rq->curr = rq->idle = idle;
7029 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7032 raw_spin_unlock_irqrestore(&rq->lock, flags);
7034 /* Set the preempt count _outside_ the spinlocks! */
7035 #if defined(CONFIG_PREEMPT)
7036 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7038 task_thread_info(idle)->preempt_count = 0;
7041 * The idle tasks have their own, simple scheduling class:
7043 idle->sched_class = &idle_sched_class;
7044 ftrace_graph_init_task(idle);
7048 * In a system that switches off the HZ timer nohz_cpu_mask
7049 * indicates which cpus entered this state. This is used
7050 * in the rcu update to wait only for active cpus. For system
7051 * which do not switch off the HZ timer nohz_cpu_mask should
7052 * always be CPU_BITS_NONE.
7054 cpumask_var_t nohz_cpu_mask;
7057 * Increase the granularity value when there are more CPUs,
7058 * because with more CPUs the 'effective latency' as visible
7059 * to users decreases. But the relationship is not linear,
7060 * so pick a second-best guess by going with the log2 of the
7063 * This idea comes from the SD scheduler of Con Kolivas:
7065 static int get_update_sysctl_factor(void)
7067 unsigned int cpus = min_t(int, num_online_cpus(), 8);
7068 unsigned int factor;
7070 switch (sysctl_sched_tunable_scaling) {
7071 case SCHED_TUNABLESCALING_NONE:
7074 case SCHED_TUNABLESCALING_LINEAR:
7077 case SCHED_TUNABLESCALING_LOG:
7079 factor = 1 + ilog2(cpus);
7086 static void update_sysctl(void)
7088 unsigned int factor = get_update_sysctl_factor();
7090 #define SET_SYSCTL(name) \
7091 (sysctl_##name = (factor) * normalized_sysctl_##name)
7092 SET_SYSCTL(sched_min_granularity);
7093 SET_SYSCTL(sched_latency);
7094 SET_SYSCTL(sched_wakeup_granularity);
7095 SET_SYSCTL(sched_shares_ratelimit);
7099 static inline void sched_init_granularity(void)
7106 * This is how migration works:
7108 * 1) we queue a struct migration_req structure in the source CPU's
7109 * runqueue and wake up that CPU's migration thread.
7110 * 2) we down() the locked semaphore => thread blocks.
7111 * 3) migration thread wakes up (implicitly it forces the migrated
7112 * thread off the CPU)
7113 * 4) it gets the migration request and checks whether the migrated
7114 * task is still in the wrong runqueue.
7115 * 5) if it's in the wrong runqueue then the migration thread removes
7116 * it and puts it into the right queue.
7117 * 6) migration thread up()s the semaphore.
7118 * 7) we wake up and the migration is done.
7122 * Change a given task's CPU affinity. Migrate the thread to a
7123 * proper CPU and schedule it away if the CPU it's executing on
7124 * is removed from the allowed bitmask.
7126 * NOTE: the caller must have a valid reference to the task, the
7127 * task must not exit() & deallocate itself prematurely. The
7128 * call is not atomic; no spinlocks may be held.
7130 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7132 struct migration_req req;
7133 unsigned long flags;
7138 * Since we rely on wake-ups to migrate sleeping tasks, don't change
7139 * the ->cpus_allowed mask from under waking tasks, which would be
7140 * possible when we change rq->lock in ttwu(), so synchronize against
7141 * TASK_WAKING to avoid that.
7144 while (p->state == TASK_WAKING)
7147 rq = task_rq_lock(p, &flags);
7149 if (p->state == TASK_WAKING) {
7150 task_rq_unlock(rq, &flags);
7154 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7159 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7160 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7165 if (p->sched_class->set_cpus_allowed)
7166 p->sched_class->set_cpus_allowed(p, new_mask);
7168 cpumask_copy(&p->cpus_allowed, new_mask);
7169 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7172 /* Can the task run on the task's current CPU? If so, we're done */
7173 if (cpumask_test_cpu(task_cpu(p), new_mask))
7176 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7177 /* Need help from migration thread: drop lock and wait. */
7178 struct task_struct *mt = rq->migration_thread;
7180 get_task_struct(mt);
7181 task_rq_unlock(rq, &flags);
7182 wake_up_process(rq->migration_thread);
7183 put_task_struct(mt);
7184 wait_for_completion(&req.done);
7185 tlb_migrate_finish(p->mm);
7189 task_rq_unlock(rq, &flags);
7193 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7196 * Move (not current) task off this cpu, onto dest cpu. We're doing
7197 * this because either it can't run here any more (set_cpus_allowed()
7198 * away from this CPU, or CPU going down), or because we're
7199 * attempting to rebalance this task on exec (sched_exec).
7201 * So we race with normal scheduler movements, but that's OK, as long
7202 * as the task is no longer on this CPU.
7204 * Returns non-zero if task was successfully migrated.
7206 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7208 struct rq *rq_dest, *rq_src;
7211 if (unlikely(!cpu_active(dest_cpu)))
7214 rq_src = cpu_rq(src_cpu);
7215 rq_dest = cpu_rq(dest_cpu);
7217 double_rq_lock(rq_src, rq_dest);
7218 /* Already moved. */
7219 if (task_cpu(p) != src_cpu)
7221 /* Affinity changed (again). */
7222 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7226 * If we're not on a rq, the next wake-up will ensure we're
7230 deactivate_task(rq_src, p, 0);
7231 set_task_cpu(p, dest_cpu);
7232 activate_task(rq_dest, p, 0);
7233 check_preempt_curr(rq_dest, p, 0);
7238 double_rq_unlock(rq_src, rq_dest);
7242 #define RCU_MIGRATION_IDLE 0
7243 #define RCU_MIGRATION_NEED_QS 1
7244 #define RCU_MIGRATION_GOT_QS 2
7245 #define RCU_MIGRATION_MUST_SYNC 3
7248 * migration_thread - this is a highprio system thread that performs
7249 * thread migration by bumping thread off CPU then 'pushing' onto
7252 static int migration_thread(void *data)
7255 int cpu = (long)data;
7259 BUG_ON(rq->migration_thread != current);
7261 set_current_state(TASK_INTERRUPTIBLE);
7262 while (!kthread_should_stop()) {
7263 struct migration_req *req;
7264 struct list_head *head;
7266 raw_spin_lock_irq(&rq->lock);
7268 if (cpu_is_offline(cpu)) {
7269 raw_spin_unlock_irq(&rq->lock);
7273 if (rq->active_balance) {
7274 active_load_balance(rq, cpu);
7275 rq->active_balance = 0;
7278 head = &rq->migration_queue;
7280 if (list_empty(head)) {
7281 raw_spin_unlock_irq(&rq->lock);
7283 set_current_state(TASK_INTERRUPTIBLE);
7286 req = list_entry(head->next, struct migration_req, list);
7287 list_del_init(head->next);
7289 if (req->task != NULL) {
7290 raw_spin_unlock(&rq->lock);
7291 __migrate_task(req->task, cpu, req->dest_cpu);
7292 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7293 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7294 raw_spin_unlock(&rq->lock);
7296 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7297 raw_spin_unlock(&rq->lock);
7298 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7302 complete(&req->done);
7304 __set_current_state(TASK_RUNNING);
7309 #ifdef CONFIG_HOTPLUG_CPU
7311 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7315 local_irq_disable();
7316 ret = __migrate_task(p, src_cpu, dest_cpu);
7322 * Figure out where task on dead CPU should go, use force if necessary.
7324 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7329 dest_cpu = select_fallback_rq(dead_cpu, p);
7331 /* It can have affinity changed while we were choosing. */
7332 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7337 * While a dead CPU has no uninterruptible tasks queued at this point,
7338 * it might still have a nonzero ->nr_uninterruptible counter, because
7339 * for performance reasons the counter is not stricly tracking tasks to
7340 * their home CPUs. So we just add the counter to another CPU's counter,
7341 * to keep the global sum constant after CPU-down:
7343 static void migrate_nr_uninterruptible(struct rq *rq_src)
7345 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7346 unsigned long flags;
7348 local_irq_save(flags);
7349 double_rq_lock(rq_src, rq_dest);
7350 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7351 rq_src->nr_uninterruptible = 0;
7352 double_rq_unlock(rq_src, rq_dest);
7353 local_irq_restore(flags);
7356 /* Run through task list and migrate tasks from the dead cpu. */
7357 static void migrate_live_tasks(int src_cpu)
7359 struct task_struct *p, *t;
7361 read_lock(&tasklist_lock);
7363 do_each_thread(t, p) {
7367 if (task_cpu(p) == src_cpu)
7368 move_task_off_dead_cpu(src_cpu, p);
7369 } while_each_thread(t, p);
7371 read_unlock(&tasklist_lock);
7375 * Schedules idle task to be the next runnable task on current CPU.
7376 * It does so by boosting its priority to highest possible.
7377 * Used by CPU offline code.
7379 void sched_idle_next(void)
7381 int this_cpu = smp_processor_id();
7382 struct rq *rq = cpu_rq(this_cpu);
7383 struct task_struct *p = rq->idle;
7384 unsigned long flags;
7386 /* cpu has to be offline */
7387 BUG_ON(cpu_online(this_cpu));
7390 * Strictly not necessary since rest of the CPUs are stopped by now
7391 * and interrupts disabled on the current cpu.
7393 raw_spin_lock_irqsave(&rq->lock, flags);
7395 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7397 update_rq_clock(rq);
7398 activate_task(rq, p, 0);
7400 raw_spin_unlock_irqrestore(&rq->lock, flags);
7404 * Ensures that the idle task is using init_mm right before its cpu goes
7407 void idle_task_exit(void)
7409 struct mm_struct *mm = current->active_mm;
7411 BUG_ON(cpu_online(smp_processor_id()));
7414 switch_mm(mm, &init_mm, current);
7418 /* called under rq->lock with disabled interrupts */
7419 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7421 struct rq *rq = cpu_rq(dead_cpu);
7423 /* Must be exiting, otherwise would be on tasklist. */
7424 BUG_ON(!p->exit_state);
7426 /* Cannot have done final schedule yet: would have vanished. */
7427 BUG_ON(p->state == TASK_DEAD);
7432 * Drop lock around migration; if someone else moves it,
7433 * that's OK. No task can be added to this CPU, so iteration is
7436 raw_spin_unlock_irq(&rq->lock);
7437 move_task_off_dead_cpu(dead_cpu, p);
7438 raw_spin_lock_irq(&rq->lock);
7443 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7444 static void migrate_dead_tasks(unsigned int dead_cpu)
7446 struct rq *rq = cpu_rq(dead_cpu);
7447 struct task_struct *next;
7450 if (!rq->nr_running)
7452 update_rq_clock(rq);
7453 next = pick_next_task(rq);
7456 next->sched_class->put_prev_task(rq, next);
7457 migrate_dead(dead_cpu, next);
7463 * remove the tasks which were accounted by rq from calc_load_tasks.
7465 static void calc_global_load_remove(struct rq *rq)
7467 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7468 rq->calc_load_active = 0;
7470 #endif /* CONFIG_HOTPLUG_CPU */
7472 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7474 static struct ctl_table sd_ctl_dir[] = {
7476 .procname = "sched_domain",
7482 static struct ctl_table sd_ctl_root[] = {
7484 .procname = "kernel",
7486 .child = sd_ctl_dir,
7491 static struct ctl_table *sd_alloc_ctl_entry(int n)
7493 struct ctl_table *entry =
7494 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7499 static void sd_free_ctl_entry(struct ctl_table **tablep)
7501 struct ctl_table *entry;
7504 * In the intermediate directories, both the child directory and
7505 * procname are dynamically allocated and could fail but the mode
7506 * will always be set. In the lowest directory the names are
7507 * static strings and all have proc handlers.
7509 for (entry = *tablep; entry->mode; entry++) {
7511 sd_free_ctl_entry(&entry->child);
7512 if (entry->proc_handler == NULL)
7513 kfree(entry->procname);
7521 set_table_entry(struct ctl_table *entry,
7522 const char *procname, void *data, int maxlen,
7523 mode_t mode, proc_handler *proc_handler)
7525 entry->procname = procname;
7527 entry->maxlen = maxlen;
7529 entry->proc_handler = proc_handler;
7532 static struct ctl_table *
7533 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7535 struct ctl_table *table = sd_alloc_ctl_entry(13);
7540 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7541 sizeof(long), 0644, proc_doulongvec_minmax);
7542 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7543 sizeof(long), 0644, proc_doulongvec_minmax);
7544 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7545 sizeof(int), 0644, proc_dointvec_minmax);
7546 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7547 sizeof(int), 0644, proc_dointvec_minmax);
7548 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7549 sizeof(int), 0644, proc_dointvec_minmax);
7550 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7551 sizeof(int), 0644, proc_dointvec_minmax);
7552 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7553 sizeof(int), 0644, proc_dointvec_minmax);
7554 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7555 sizeof(int), 0644, proc_dointvec_minmax);
7556 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7557 sizeof(int), 0644, proc_dointvec_minmax);
7558 set_table_entry(&table[9], "cache_nice_tries",
7559 &sd->cache_nice_tries,
7560 sizeof(int), 0644, proc_dointvec_minmax);
7561 set_table_entry(&table[10], "flags", &sd->flags,
7562 sizeof(int), 0644, proc_dointvec_minmax);
7563 set_table_entry(&table[11], "name", sd->name,
7564 CORENAME_MAX_SIZE, 0444, proc_dostring);
7565 /* &table[12] is terminator */
7570 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7572 struct ctl_table *entry, *table;
7573 struct sched_domain *sd;
7574 int domain_num = 0, i;
7577 for_each_domain(cpu, sd)
7579 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7584 for_each_domain(cpu, sd) {
7585 snprintf(buf, 32, "domain%d", i);
7586 entry->procname = kstrdup(buf, GFP_KERNEL);
7588 entry->child = sd_alloc_ctl_domain_table(sd);
7595 static struct ctl_table_header *sd_sysctl_header;
7596 static void register_sched_domain_sysctl(void)
7598 int i, cpu_num = num_possible_cpus();
7599 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7602 WARN_ON(sd_ctl_dir[0].child);
7603 sd_ctl_dir[0].child = entry;
7608 for_each_possible_cpu(i) {
7609 snprintf(buf, 32, "cpu%d", i);
7610 entry->procname = kstrdup(buf, GFP_KERNEL);
7612 entry->child = sd_alloc_ctl_cpu_table(i);
7616 WARN_ON(sd_sysctl_header);
7617 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7620 /* may be called multiple times per register */
7621 static void unregister_sched_domain_sysctl(void)
7623 if (sd_sysctl_header)
7624 unregister_sysctl_table(sd_sysctl_header);
7625 sd_sysctl_header = NULL;
7626 if (sd_ctl_dir[0].child)
7627 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7630 static void register_sched_domain_sysctl(void)
7633 static void unregister_sched_domain_sysctl(void)
7638 static void set_rq_online(struct rq *rq)
7641 const struct sched_class *class;
7643 cpumask_set_cpu(rq->cpu, rq->rd->online);
7646 for_each_class(class) {
7647 if (class->rq_online)
7648 class->rq_online(rq);
7653 static void set_rq_offline(struct rq *rq)
7656 const struct sched_class *class;
7658 for_each_class(class) {
7659 if (class->rq_offline)
7660 class->rq_offline(rq);
7663 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7669 * migration_call - callback that gets triggered when a CPU is added.
7670 * Here we can start up the necessary migration thread for the new CPU.
7672 static int __cpuinit
7673 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7675 struct task_struct *p;
7676 int cpu = (long)hcpu;
7677 unsigned long flags;
7682 case CPU_UP_PREPARE:
7683 case CPU_UP_PREPARE_FROZEN:
7684 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7687 kthread_bind(p, cpu);
7688 /* Must be high prio: stop_machine expects to yield to it. */
7689 rq = task_rq_lock(p, &flags);
7690 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7691 task_rq_unlock(rq, &flags);
7693 cpu_rq(cpu)->migration_thread = p;
7694 rq->calc_load_update = calc_load_update;
7698 case CPU_ONLINE_FROZEN:
7699 /* Strictly unnecessary, as first user will wake it. */
7700 wake_up_process(cpu_rq(cpu)->migration_thread);
7702 /* Update our root-domain */
7704 raw_spin_lock_irqsave(&rq->lock, flags);
7706 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7710 raw_spin_unlock_irqrestore(&rq->lock, flags);
7713 #ifdef CONFIG_HOTPLUG_CPU
7714 case CPU_UP_CANCELED:
7715 case CPU_UP_CANCELED_FROZEN:
7716 if (!cpu_rq(cpu)->migration_thread)
7718 /* Unbind it from offline cpu so it can run. Fall thru. */
7719 kthread_bind(cpu_rq(cpu)->migration_thread,
7720 cpumask_any(cpu_online_mask));
7721 kthread_stop(cpu_rq(cpu)->migration_thread);
7722 put_task_struct(cpu_rq(cpu)->migration_thread);
7723 cpu_rq(cpu)->migration_thread = NULL;
7727 case CPU_DEAD_FROZEN:
7728 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7729 migrate_live_tasks(cpu);
7731 kthread_stop(rq->migration_thread);
7732 put_task_struct(rq->migration_thread);
7733 rq->migration_thread = NULL;
7734 /* Idle task back to normal (off runqueue, low prio) */
7735 raw_spin_lock_irq(&rq->lock);
7736 update_rq_clock(rq);
7737 deactivate_task(rq, rq->idle, 0);
7738 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7739 rq->idle->sched_class = &idle_sched_class;
7740 migrate_dead_tasks(cpu);
7741 raw_spin_unlock_irq(&rq->lock);
7743 migrate_nr_uninterruptible(rq);
7744 BUG_ON(rq->nr_running != 0);
7745 calc_global_load_remove(rq);
7747 * No need to migrate the tasks: it was best-effort if
7748 * they didn't take sched_hotcpu_mutex. Just wake up
7751 raw_spin_lock_irq(&rq->lock);
7752 while (!list_empty(&rq->migration_queue)) {
7753 struct migration_req *req;
7755 req = list_entry(rq->migration_queue.next,
7756 struct migration_req, list);
7757 list_del_init(&req->list);
7758 raw_spin_unlock_irq(&rq->lock);
7759 complete(&req->done);
7760 raw_spin_lock_irq(&rq->lock);
7762 raw_spin_unlock_irq(&rq->lock);
7766 case CPU_DYING_FROZEN:
7767 /* Update our root-domain */
7769 raw_spin_lock_irqsave(&rq->lock, flags);
7771 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7774 raw_spin_unlock_irqrestore(&rq->lock, flags);
7782 * Register at high priority so that task migration (migrate_all_tasks)
7783 * happens before everything else. This has to be lower priority than
7784 * the notifier in the perf_event subsystem, though.
7786 static struct notifier_block __cpuinitdata migration_notifier = {
7787 .notifier_call = migration_call,
7791 static int __init migration_init(void)
7793 void *cpu = (void *)(long)smp_processor_id();
7796 /* Start one for the boot CPU: */
7797 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7798 BUG_ON(err == NOTIFY_BAD);
7799 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7800 register_cpu_notifier(&migration_notifier);
7804 early_initcall(migration_init);
7809 #ifdef CONFIG_SCHED_DEBUG
7811 static __read_mostly int sched_domain_debug_enabled;
7813 static int __init sched_domain_debug_setup(char *str)
7815 sched_domain_debug_enabled = 1;
7819 early_param("sched_debug", sched_domain_debug_setup);
7821 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7822 struct cpumask *groupmask)
7824 struct sched_group *group = sd->groups;
7827 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7828 cpumask_clear(groupmask);
7830 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7832 if (!(sd->flags & SD_LOAD_BALANCE)) {
7833 pr_cont("does not load-balance\n");
7835 pr_err("ERROR: !SD_LOAD_BALANCE domain has parent\n");
7839 pr_cont("span %s level %s\n", str, sd->name);
7841 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7842 pr_err("ERROR: domain->span does not contain CPU%d\n", cpu);
7844 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7845 pr_err("ERROR: domain->groups does not contain CPU%d\n", cpu);
7848 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7852 pr_err("ERROR: group is NULL\n");
7856 if (!group->cpu_power) {
7858 pr_err("ERROR: domain->cpu_power not set\n");
7862 if (!cpumask_weight(sched_group_cpus(group))) {
7864 pr_err("ERROR: empty group\n");
7868 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7870 pr_err("ERROR: repeated CPUs\n");
7874 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7876 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7878 pr_cont(" %s", str);
7879 if (group->cpu_power != SCHED_LOAD_SCALE) {
7880 pr_cont(" (cpu_power = %d)", group->cpu_power);
7883 group = group->next;
7884 } while (group != sd->groups);
7887 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7888 pr_err("ERROR: groups don't span domain->span\n");
7891 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7892 pr_err("ERROR: parent span is not a superset of domain->span\n");
7896 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7898 cpumask_var_t groupmask;
7901 if (!sched_domain_debug_enabled)
7905 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7909 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7911 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7912 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7917 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7924 free_cpumask_var(groupmask);
7926 #else /* !CONFIG_SCHED_DEBUG */
7927 # define sched_domain_debug(sd, cpu) do { } while (0)
7928 #endif /* CONFIG_SCHED_DEBUG */
7930 static int sd_degenerate(struct sched_domain *sd)
7932 if (cpumask_weight(sched_domain_span(sd)) == 1)
7935 /* Following flags need at least 2 groups */
7936 if (sd->flags & (SD_LOAD_BALANCE |
7937 SD_BALANCE_NEWIDLE |
7941 SD_SHARE_PKG_RESOURCES)) {
7942 if (sd->groups != sd->groups->next)
7946 /* Following flags don't use groups */
7947 if (sd->flags & (SD_WAKE_AFFINE))
7954 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7956 unsigned long cflags = sd->flags, pflags = parent->flags;
7958 if (sd_degenerate(parent))
7961 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7964 /* Flags needing groups don't count if only 1 group in parent */
7965 if (parent->groups == parent->groups->next) {
7966 pflags &= ~(SD_LOAD_BALANCE |
7967 SD_BALANCE_NEWIDLE |
7971 SD_SHARE_PKG_RESOURCES);
7972 if (nr_node_ids == 1)
7973 pflags &= ~SD_SERIALIZE;
7975 if (~cflags & pflags)
7981 static void free_rootdomain(struct root_domain *rd)
7983 synchronize_sched();
7985 cpupri_cleanup(&rd->cpupri);
7987 free_cpumask_var(rd->rto_mask);
7988 free_cpumask_var(rd->online);
7989 free_cpumask_var(rd->span);
7993 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7995 struct root_domain *old_rd = NULL;
7996 unsigned long flags;
7998 raw_spin_lock_irqsave(&rq->lock, flags);
8003 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8006 cpumask_clear_cpu(rq->cpu, old_rd->span);
8009 * If we dont want to free the old_rt yet then
8010 * set old_rd to NULL to skip the freeing later
8013 if (!atomic_dec_and_test(&old_rd->refcount))
8017 atomic_inc(&rd->refcount);
8020 cpumask_set_cpu(rq->cpu, rd->span);
8021 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8024 raw_spin_unlock_irqrestore(&rq->lock, flags);
8027 free_rootdomain(old_rd);
8030 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8032 gfp_t gfp = GFP_KERNEL;
8034 memset(rd, 0, sizeof(*rd));
8039 if (!alloc_cpumask_var(&rd->span, gfp))
8041 if (!alloc_cpumask_var(&rd->online, gfp))
8043 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8046 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8051 free_cpumask_var(rd->rto_mask);
8053 free_cpumask_var(rd->online);
8055 free_cpumask_var(rd->span);
8060 static void init_defrootdomain(void)
8062 init_rootdomain(&def_root_domain, true);
8064 atomic_set(&def_root_domain.refcount, 1);
8067 static struct root_domain *alloc_rootdomain(void)
8069 struct root_domain *rd;
8071 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8075 if (init_rootdomain(rd, false) != 0) {
8084 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8085 * hold the hotplug lock.
8088 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8090 struct rq *rq = cpu_rq(cpu);
8091 struct sched_domain *tmp;
8093 /* Remove the sched domains which do not contribute to scheduling. */
8094 for (tmp = sd; tmp; ) {
8095 struct sched_domain *parent = tmp->parent;
8099 if (sd_parent_degenerate(tmp, parent)) {
8100 tmp->parent = parent->parent;
8102 parent->parent->child = tmp;
8107 if (sd && sd_degenerate(sd)) {
8113 sched_domain_debug(sd, cpu);
8115 rq_attach_root(rq, rd);
8116 rcu_assign_pointer(rq->sd, sd);
8119 /* cpus with isolated domains */
8120 static cpumask_var_t cpu_isolated_map;
8122 /* Setup the mask of cpus configured for isolated domains */
8123 static int __init isolated_cpu_setup(char *str)
8125 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8126 cpulist_parse(str, cpu_isolated_map);
8130 __setup("isolcpus=", isolated_cpu_setup);
8133 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8134 * to a function which identifies what group(along with sched group) a CPU
8135 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8136 * (due to the fact that we keep track of groups covered with a struct cpumask).
8138 * init_sched_build_groups will build a circular linked list of the groups
8139 * covered by the given span, and will set each group's ->cpumask correctly,
8140 * and ->cpu_power to 0.
8143 init_sched_build_groups(const struct cpumask *span,
8144 const struct cpumask *cpu_map,
8145 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8146 struct sched_group **sg,
8147 struct cpumask *tmpmask),
8148 struct cpumask *covered, struct cpumask *tmpmask)
8150 struct sched_group *first = NULL, *last = NULL;
8153 cpumask_clear(covered);
8155 for_each_cpu(i, span) {
8156 struct sched_group *sg;
8157 int group = group_fn(i, cpu_map, &sg, tmpmask);
8160 if (cpumask_test_cpu(i, covered))
8163 cpumask_clear(sched_group_cpus(sg));
8166 for_each_cpu(j, span) {
8167 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8170 cpumask_set_cpu(j, covered);
8171 cpumask_set_cpu(j, sched_group_cpus(sg));
8182 #define SD_NODES_PER_DOMAIN 16
8187 * find_next_best_node - find the next node to include in a sched_domain
8188 * @node: node whose sched_domain we're building
8189 * @used_nodes: nodes already in the sched_domain
8191 * Find the next node to include in a given scheduling domain. Simply
8192 * finds the closest node not already in the @used_nodes map.
8194 * Should use nodemask_t.
8196 static int find_next_best_node(int node, nodemask_t *used_nodes)
8198 int i, n, val, min_val, best_node = 0;
8202 for (i = 0; i < nr_node_ids; i++) {
8203 /* Start at @node */
8204 n = (node + i) % nr_node_ids;
8206 if (!nr_cpus_node(n))
8209 /* Skip already used nodes */
8210 if (node_isset(n, *used_nodes))
8213 /* Simple min distance search */
8214 val = node_distance(node, n);
8216 if (val < min_val) {
8222 node_set(best_node, *used_nodes);
8227 * sched_domain_node_span - get a cpumask for a node's sched_domain
8228 * @node: node whose cpumask we're constructing
8229 * @span: resulting cpumask
8231 * Given a node, construct a good cpumask for its sched_domain to span. It
8232 * should be one that prevents unnecessary balancing, but also spreads tasks
8235 static void sched_domain_node_span(int node, struct cpumask *span)
8237 nodemask_t used_nodes;
8240 cpumask_clear(span);
8241 nodes_clear(used_nodes);
8243 cpumask_or(span, span, cpumask_of_node(node));
8244 node_set(node, used_nodes);
8246 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8247 int next_node = find_next_best_node(node, &used_nodes);
8249 cpumask_or(span, span, cpumask_of_node(next_node));
8252 #endif /* CONFIG_NUMA */
8254 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8257 * The cpus mask in sched_group and sched_domain hangs off the end.
8259 * ( See the the comments in include/linux/sched.h:struct sched_group
8260 * and struct sched_domain. )
8262 struct static_sched_group {
8263 struct sched_group sg;
8264 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8267 struct static_sched_domain {
8268 struct sched_domain sd;
8269 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8275 cpumask_var_t domainspan;
8276 cpumask_var_t covered;
8277 cpumask_var_t notcovered;
8279 cpumask_var_t nodemask;
8280 cpumask_var_t this_sibling_map;
8281 cpumask_var_t this_core_map;
8282 cpumask_var_t send_covered;
8283 cpumask_var_t tmpmask;
8284 struct sched_group **sched_group_nodes;
8285 struct root_domain *rd;
8289 sa_sched_groups = 0,
8294 sa_this_sibling_map,
8296 sa_sched_group_nodes,
8306 * SMT sched-domains:
8308 #ifdef CONFIG_SCHED_SMT
8309 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8310 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
8313 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8314 struct sched_group **sg, struct cpumask *unused)
8317 *sg = &per_cpu(sched_groups, cpu).sg;
8320 #endif /* CONFIG_SCHED_SMT */
8323 * multi-core sched-domains:
8325 #ifdef CONFIG_SCHED_MC
8326 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8327 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8328 #endif /* CONFIG_SCHED_MC */
8330 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8332 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8333 struct sched_group **sg, struct cpumask *mask)
8337 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8338 group = cpumask_first(mask);
8340 *sg = &per_cpu(sched_group_core, group).sg;
8343 #elif defined(CONFIG_SCHED_MC)
8345 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8346 struct sched_group **sg, struct cpumask *unused)
8349 *sg = &per_cpu(sched_group_core, cpu).sg;
8354 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8355 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8358 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8359 struct sched_group **sg, struct cpumask *mask)
8362 #ifdef CONFIG_SCHED_MC
8363 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8364 group = cpumask_first(mask);
8365 #elif defined(CONFIG_SCHED_SMT)
8366 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8367 group = cpumask_first(mask);
8372 *sg = &per_cpu(sched_group_phys, group).sg;
8378 * The init_sched_build_groups can't handle what we want to do with node
8379 * groups, so roll our own. Now each node has its own list of groups which
8380 * gets dynamically allocated.
8382 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8383 static struct sched_group ***sched_group_nodes_bycpu;
8385 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8386 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8388 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8389 struct sched_group **sg,
8390 struct cpumask *nodemask)
8394 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8395 group = cpumask_first(nodemask);
8398 *sg = &per_cpu(sched_group_allnodes, group).sg;
8402 static void init_numa_sched_groups_power(struct sched_group *group_head)
8404 struct sched_group *sg = group_head;
8410 for_each_cpu(j, sched_group_cpus(sg)) {
8411 struct sched_domain *sd;
8413 sd = &per_cpu(phys_domains, j).sd;
8414 if (j != group_first_cpu(sd->groups)) {
8416 * Only add "power" once for each
8422 sg->cpu_power += sd->groups->cpu_power;
8425 } while (sg != group_head);
8428 static int build_numa_sched_groups(struct s_data *d,
8429 const struct cpumask *cpu_map, int num)
8431 struct sched_domain *sd;
8432 struct sched_group *sg, *prev;
8435 cpumask_clear(d->covered);
8436 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8437 if (cpumask_empty(d->nodemask)) {
8438 d->sched_group_nodes[num] = NULL;
8442 sched_domain_node_span(num, d->domainspan);
8443 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8445 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8448 pr_warning("Can not alloc domain group for node %d\n", num);
8451 d->sched_group_nodes[num] = sg;
8453 for_each_cpu(j, d->nodemask) {
8454 sd = &per_cpu(node_domains, j).sd;
8459 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8461 cpumask_or(d->covered, d->covered, d->nodemask);
8464 for (j = 0; j < nr_node_ids; j++) {
8465 n = (num + j) % nr_node_ids;
8466 cpumask_complement(d->notcovered, d->covered);
8467 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8468 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8469 if (cpumask_empty(d->tmpmask))
8471 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8472 if (cpumask_empty(d->tmpmask))
8474 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8477 pr_warning("Can not alloc domain group for node %d\n",
8482 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8483 sg->next = prev->next;
8484 cpumask_or(d->covered, d->covered, d->tmpmask);
8491 #endif /* CONFIG_NUMA */
8494 /* Free memory allocated for various sched_group structures */
8495 static void free_sched_groups(const struct cpumask *cpu_map,
8496 struct cpumask *nodemask)
8500 for_each_cpu(cpu, cpu_map) {
8501 struct sched_group **sched_group_nodes
8502 = sched_group_nodes_bycpu[cpu];
8504 if (!sched_group_nodes)
8507 for (i = 0; i < nr_node_ids; i++) {
8508 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8510 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8511 if (cpumask_empty(nodemask))
8521 if (oldsg != sched_group_nodes[i])
8524 kfree(sched_group_nodes);
8525 sched_group_nodes_bycpu[cpu] = NULL;
8528 #else /* !CONFIG_NUMA */
8529 static void free_sched_groups(const struct cpumask *cpu_map,
8530 struct cpumask *nodemask)
8533 #endif /* CONFIG_NUMA */
8536 * Initialize sched groups cpu_power.
8538 * cpu_power indicates the capacity of sched group, which is used while
8539 * distributing the load between different sched groups in a sched domain.
8540 * Typically cpu_power for all the groups in a sched domain will be same unless
8541 * there are asymmetries in the topology. If there are asymmetries, group
8542 * having more cpu_power will pickup more load compared to the group having
8545 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8547 struct sched_domain *child;
8548 struct sched_group *group;
8552 WARN_ON(!sd || !sd->groups);
8554 if (cpu != group_first_cpu(sd->groups))
8559 sd->groups->cpu_power = 0;
8562 power = SCHED_LOAD_SCALE;
8563 weight = cpumask_weight(sched_domain_span(sd));
8565 * SMT siblings share the power of a single core.
8566 * Usually multiple threads get a better yield out of
8567 * that one core than a single thread would have,
8568 * reflect that in sd->smt_gain.
8570 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8571 power *= sd->smt_gain;
8573 power >>= SCHED_LOAD_SHIFT;
8575 sd->groups->cpu_power += power;
8580 * Add cpu_power of each child group to this groups cpu_power.
8582 group = child->groups;
8584 sd->groups->cpu_power += group->cpu_power;
8585 group = group->next;
8586 } while (group != child->groups);
8590 * Initializers for schedule domains
8591 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8594 #ifdef CONFIG_SCHED_DEBUG
8595 # define SD_INIT_NAME(sd, type) sd->name = #type
8597 # define SD_INIT_NAME(sd, type) do { } while (0)
8600 #define SD_INIT(sd, type) sd_init_##type(sd)
8602 #define SD_INIT_FUNC(type) \
8603 static noinline void sd_init_##type(struct sched_domain *sd) \
8605 memset(sd, 0, sizeof(*sd)); \
8606 *sd = SD_##type##_INIT; \
8607 sd->level = SD_LV_##type; \
8608 SD_INIT_NAME(sd, type); \
8613 SD_INIT_FUNC(ALLNODES)
8616 #ifdef CONFIG_SCHED_SMT
8617 SD_INIT_FUNC(SIBLING)
8619 #ifdef CONFIG_SCHED_MC
8623 static int default_relax_domain_level = -1;
8625 static int __init setup_relax_domain_level(char *str)
8629 val = simple_strtoul(str, NULL, 0);
8630 if (val < SD_LV_MAX)
8631 default_relax_domain_level = val;
8635 __setup("relax_domain_level=", setup_relax_domain_level);
8637 static void set_domain_attribute(struct sched_domain *sd,
8638 struct sched_domain_attr *attr)
8642 if (!attr || attr->relax_domain_level < 0) {
8643 if (default_relax_domain_level < 0)
8646 request = default_relax_domain_level;
8648 request = attr->relax_domain_level;
8649 if (request < sd->level) {
8650 /* turn off idle balance on this domain */
8651 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8653 /* turn on idle balance on this domain */
8654 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8658 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8659 const struct cpumask *cpu_map)
8662 case sa_sched_groups:
8663 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8664 d->sched_group_nodes = NULL;
8666 free_rootdomain(d->rd); /* fall through */
8668 free_cpumask_var(d->tmpmask); /* fall through */
8669 case sa_send_covered:
8670 free_cpumask_var(d->send_covered); /* fall through */
8671 case sa_this_core_map:
8672 free_cpumask_var(d->this_core_map); /* fall through */
8673 case sa_this_sibling_map:
8674 free_cpumask_var(d->this_sibling_map); /* fall through */
8676 free_cpumask_var(d->nodemask); /* fall through */
8677 case sa_sched_group_nodes:
8679 kfree(d->sched_group_nodes); /* fall through */
8681 free_cpumask_var(d->notcovered); /* fall through */
8683 free_cpumask_var(d->covered); /* fall through */
8685 free_cpumask_var(d->domainspan); /* fall through */
8692 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8693 const struct cpumask *cpu_map)
8696 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8698 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8699 return sa_domainspan;
8700 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8702 /* Allocate the per-node list of sched groups */
8703 d->sched_group_nodes = kcalloc(nr_node_ids,
8704 sizeof(struct sched_group *), GFP_KERNEL);
8705 if (!d->sched_group_nodes) {
8706 pr_warning("Can not alloc sched group node list\n");
8707 return sa_notcovered;
8709 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8711 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8712 return sa_sched_group_nodes;
8713 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8715 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8716 return sa_this_sibling_map;
8717 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8718 return sa_this_core_map;
8719 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8720 return sa_send_covered;
8721 d->rd = alloc_rootdomain();
8723 pr_warning("Cannot alloc root domain\n");
8726 return sa_rootdomain;
8729 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8730 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8732 struct sched_domain *sd = NULL;
8734 struct sched_domain *parent;
8737 if (cpumask_weight(cpu_map) >
8738 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8739 sd = &per_cpu(allnodes_domains, i).sd;
8740 SD_INIT(sd, ALLNODES);
8741 set_domain_attribute(sd, attr);
8742 cpumask_copy(sched_domain_span(sd), cpu_map);
8743 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8748 sd = &per_cpu(node_domains, i).sd;
8750 set_domain_attribute(sd, attr);
8751 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8752 sd->parent = parent;
8755 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8760 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8761 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8762 struct sched_domain *parent, int i)
8764 struct sched_domain *sd;
8765 sd = &per_cpu(phys_domains, i).sd;
8767 set_domain_attribute(sd, attr);
8768 cpumask_copy(sched_domain_span(sd), d->nodemask);
8769 sd->parent = parent;
8772 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8776 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8777 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8778 struct sched_domain *parent, int i)
8780 struct sched_domain *sd = parent;
8781 #ifdef CONFIG_SCHED_MC
8782 sd = &per_cpu(core_domains, i).sd;
8784 set_domain_attribute(sd, attr);
8785 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8786 sd->parent = parent;
8788 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8793 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8794 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8795 struct sched_domain *parent, int i)
8797 struct sched_domain *sd = parent;
8798 #ifdef CONFIG_SCHED_SMT
8799 sd = &per_cpu(cpu_domains, i).sd;
8800 SD_INIT(sd, SIBLING);
8801 set_domain_attribute(sd, attr);
8802 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8803 sd->parent = parent;
8805 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8810 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8811 const struct cpumask *cpu_map, int cpu)
8814 #ifdef CONFIG_SCHED_SMT
8815 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8816 cpumask_and(d->this_sibling_map, cpu_map,
8817 topology_thread_cpumask(cpu));
8818 if (cpu == cpumask_first(d->this_sibling_map))
8819 init_sched_build_groups(d->this_sibling_map, cpu_map,
8821 d->send_covered, d->tmpmask);
8824 #ifdef CONFIG_SCHED_MC
8825 case SD_LV_MC: /* set up multi-core groups */
8826 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8827 if (cpu == cpumask_first(d->this_core_map))
8828 init_sched_build_groups(d->this_core_map, cpu_map,
8830 d->send_covered, d->tmpmask);
8833 case SD_LV_CPU: /* set up physical groups */
8834 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8835 if (!cpumask_empty(d->nodemask))
8836 init_sched_build_groups(d->nodemask, cpu_map,
8838 d->send_covered, d->tmpmask);
8841 case SD_LV_ALLNODES:
8842 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8843 d->send_covered, d->tmpmask);
8852 * Build sched domains for a given set of cpus and attach the sched domains
8853 * to the individual cpus
8855 static int __build_sched_domains(const struct cpumask *cpu_map,
8856 struct sched_domain_attr *attr)
8858 enum s_alloc alloc_state = sa_none;
8860 struct sched_domain *sd;
8866 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8867 if (alloc_state != sa_rootdomain)
8869 alloc_state = sa_sched_groups;
8872 * Set up domains for cpus specified by the cpu_map.
8874 for_each_cpu(i, cpu_map) {
8875 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8878 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8879 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8880 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8881 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8884 for_each_cpu(i, cpu_map) {
8885 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8886 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8889 /* Set up physical groups */
8890 for (i = 0; i < nr_node_ids; i++)
8891 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8894 /* Set up node groups */
8896 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8898 for (i = 0; i < nr_node_ids; i++)
8899 if (build_numa_sched_groups(&d, cpu_map, i))
8903 /* Calculate CPU power for physical packages and nodes */
8904 #ifdef CONFIG_SCHED_SMT
8905 for_each_cpu(i, cpu_map) {
8906 sd = &per_cpu(cpu_domains, i).sd;
8907 init_sched_groups_power(i, sd);
8910 #ifdef CONFIG_SCHED_MC
8911 for_each_cpu(i, cpu_map) {
8912 sd = &per_cpu(core_domains, i).sd;
8913 init_sched_groups_power(i, sd);
8917 for_each_cpu(i, cpu_map) {
8918 sd = &per_cpu(phys_domains, i).sd;
8919 init_sched_groups_power(i, sd);
8923 for (i = 0; i < nr_node_ids; i++)
8924 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8926 if (d.sd_allnodes) {
8927 struct sched_group *sg;
8929 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8931 init_numa_sched_groups_power(sg);
8935 /* Attach the domains */
8936 for_each_cpu(i, cpu_map) {
8937 #ifdef CONFIG_SCHED_SMT
8938 sd = &per_cpu(cpu_domains, i).sd;
8939 #elif defined(CONFIG_SCHED_MC)
8940 sd = &per_cpu(core_domains, i).sd;
8942 sd = &per_cpu(phys_domains, i).sd;
8944 cpu_attach_domain(sd, d.rd, i);
8947 d.sched_group_nodes = NULL; /* don't free this we still need it */
8948 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8952 __free_domain_allocs(&d, alloc_state, cpu_map);
8956 static int build_sched_domains(const struct cpumask *cpu_map)
8958 return __build_sched_domains(cpu_map, NULL);
8961 static cpumask_var_t *doms_cur; /* current sched domains */
8962 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8963 static struct sched_domain_attr *dattr_cur;
8964 /* attribues of custom domains in 'doms_cur' */
8967 * Special case: If a kmalloc of a doms_cur partition (array of
8968 * cpumask) fails, then fallback to a single sched domain,
8969 * as determined by the single cpumask fallback_doms.
8971 static cpumask_var_t fallback_doms;
8974 * arch_update_cpu_topology lets virtualized architectures update the
8975 * cpu core maps. It is supposed to return 1 if the topology changed
8976 * or 0 if it stayed the same.
8978 int __attribute__((weak)) arch_update_cpu_topology(void)
8983 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8986 cpumask_var_t *doms;
8988 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8991 for (i = 0; i < ndoms; i++) {
8992 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
8993 free_sched_domains(doms, i);
9000 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
9003 for (i = 0; i < ndoms; i++)
9004 free_cpumask_var(doms[i]);
9009 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9010 * For now this just excludes isolated cpus, but could be used to
9011 * exclude other special cases in the future.
9013 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9017 arch_update_cpu_topology();
9019 doms_cur = alloc_sched_domains(ndoms_cur);
9021 doms_cur = &fallback_doms;
9022 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
9024 err = build_sched_domains(doms_cur[0]);
9025 register_sched_domain_sysctl();
9030 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9031 struct cpumask *tmpmask)
9033 free_sched_groups(cpu_map, tmpmask);
9037 * Detach sched domains from a group of cpus specified in cpu_map
9038 * These cpus will now be attached to the NULL domain
9040 static void detach_destroy_domains(const struct cpumask *cpu_map)
9042 /* Save because hotplug lock held. */
9043 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9046 for_each_cpu(i, cpu_map)
9047 cpu_attach_domain(NULL, &def_root_domain, i);
9048 synchronize_sched();
9049 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9052 /* handle null as "default" */
9053 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9054 struct sched_domain_attr *new, int idx_new)
9056 struct sched_domain_attr tmp;
9063 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9064 new ? (new + idx_new) : &tmp,
9065 sizeof(struct sched_domain_attr));
9069 * Partition sched domains as specified by the 'ndoms_new'
9070 * cpumasks in the array doms_new[] of cpumasks. This compares
9071 * doms_new[] to the current sched domain partitioning, doms_cur[].
9072 * It destroys each deleted domain and builds each new domain.
9074 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9075 * The masks don't intersect (don't overlap.) We should setup one
9076 * sched domain for each mask. CPUs not in any of the cpumasks will
9077 * not be load balanced. If the same cpumask appears both in the
9078 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9081 * The passed in 'doms_new' should be allocated using
9082 * alloc_sched_domains. This routine takes ownership of it and will
9083 * free_sched_domains it when done with it. If the caller failed the
9084 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9085 * and partition_sched_domains() will fallback to the single partition
9086 * 'fallback_doms', it also forces the domains to be rebuilt.
9088 * If doms_new == NULL it will be replaced with cpu_online_mask.
9089 * ndoms_new == 0 is a special case for destroying existing domains,
9090 * and it will not create the default domain.
9092 * Call with hotplug lock held
9094 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9095 struct sched_domain_attr *dattr_new)
9100 mutex_lock(&sched_domains_mutex);
9102 /* always unregister in case we don't destroy any domains */
9103 unregister_sched_domain_sysctl();
9105 /* Let architecture update cpu core mappings. */
9106 new_topology = arch_update_cpu_topology();
9108 n = doms_new ? ndoms_new : 0;
9110 /* Destroy deleted domains */
9111 for (i = 0; i < ndoms_cur; i++) {
9112 for (j = 0; j < n && !new_topology; j++) {
9113 if (cpumask_equal(doms_cur[i], doms_new[j])
9114 && dattrs_equal(dattr_cur, i, dattr_new, j))
9117 /* no match - a current sched domain not in new doms_new[] */
9118 detach_destroy_domains(doms_cur[i]);
9123 if (doms_new == NULL) {
9125 doms_new = &fallback_doms;
9126 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9127 WARN_ON_ONCE(dattr_new);
9130 /* Build new domains */
9131 for (i = 0; i < ndoms_new; i++) {
9132 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9133 if (cpumask_equal(doms_new[i], doms_cur[j])
9134 && dattrs_equal(dattr_new, i, dattr_cur, j))
9137 /* no match - add a new doms_new */
9138 __build_sched_domains(doms_new[i],
9139 dattr_new ? dattr_new + i : NULL);
9144 /* Remember the new sched domains */
9145 if (doms_cur != &fallback_doms)
9146 free_sched_domains(doms_cur, ndoms_cur);
9147 kfree(dattr_cur); /* kfree(NULL) is safe */
9148 doms_cur = doms_new;
9149 dattr_cur = dattr_new;
9150 ndoms_cur = ndoms_new;
9152 register_sched_domain_sysctl();
9154 mutex_unlock(&sched_domains_mutex);
9157 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9158 static void arch_reinit_sched_domains(void)
9162 /* Destroy domains first to force the rebuild */
9163 partition_sched_domains(0, NULL, NULL);
9165 rebuild_sched_domains();
9169 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9171 unsigned int level = 0;
9173 if (sscanf(buf, "%u", &level) != 1)
9177 * level is always be positive so don't check for
9178 * level < POWERSAVINGS_BALANCE_NONE which is 0
9179 * What happens on 0 or 1 byte write,
9180 * need to check for count as well?
9183 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9187 sched_smt_power_savings = level;
9189 sched_mc_power_savings = level;
9191 arch_reinit_sched_domains();
9196 #ifdef CONFIG_SCHED_MC
9197 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9200 return sprintf(page, "%u\n", sched_mc_power_savings);
9202 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9203 const char *buf, size_t count)
9205 return sched_power_savings_store(buf, count, 0);
9207 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9208 sched_mc_power_savings_show,
9209 sched_mc_power_savings_store);
9212 #ifdef CONFIG_SCHED_SMT
9213 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9216 return sprintf(page, "%u\n", sched_smt_power_savings);
9218 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9219 const char *buf, size_t count)
9221 return sched_power_savings_store(buf, count, 1);
9223 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9224 sched_smt_power_savings_show,
9225 sched_smt_power_savings_store);
9228 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9232 #ifdef CONFIG_SCHED_SMT
9234 err = sysfs_create_file(&cls->kset.kobj,
9235 &attr_sched_smt_power_savings.attr);
9237 #ifdef CONFIG_SCHED_MC
9238 if (!err && mc_capable())
9239 err = sysfs_create_file(&cls->kset.kobj,
9240 &attr_sched_mc_power_savings.attr);
9244 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9246 #ifndef CONFIG_CPUSETS
9248 * Add online and remove offline CPUs from the scheduler domains.
9249 * When cpusets are enabled they take over this function.
9251 static int update_sched_domains(struct notifier_block *nfb,
9252 unsigned long action, void *hcpu)
9256 case CPU_ONLINE_FROZEN:
9257 case CPU_DOWN_PREPARE:
9258 case CPU_DOWN_PREPARE_FROZEN:
9259 case CPU_DOWN_FAILED:
9260 case CPU_DOWN_FAILED_FROZEN:
9261 partition_sched_domains(1, NULL, NULL);
9270 static int update_runtime(struct notifier_block *nfb,
9271 unsigned long action, void *hcpu)
9273 int cpu = (int)(long)hcpu;
9276 case CPU_DOWN_PREPARE:
9277 case CPU_DOWN_PREPARE_FROZEN:
9278 disable_runtime(cpu_rq(cpu));
9281 case CPU_DOWN_FAILED:
9282 case CPU_DOWN_FAILED_FROZEN:
9284 case CPU_ONLINE_FROZEN:
9285 enable_runtime(cpu_rq(cpu));
9293 void __init sched_init_smp(void)
9295 cpumask_var_t non_isolated_cpus;
9297 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9298 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9300 #if defined(CONFIG_NUMA)
9301 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9303 BUG_ON(sched_group_nodes_bycpu == NULL);
9306 mutex_lock(&sched_domains_mutex);
9307 arch_init_sched_domains(cpu_active_mask);
9308 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9309 if (cpumask_empty(non_isolated_cpus))
9310 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9311 mutex_unlock(&sched_domains_mutex);
9314 #ifndef CONFIG_CPUSETS
9315 /* XXX: Theoretical race here - CPU may be hotplugged now */
9316 hotcpu_notifier(update_sched_domains, 0);
9319 /* RT runtime code needs to handle some hotplug events */
9320 hotcpu_notifier(update_runtime, 0);
9324 /* Move init over to a non-isolated CPU */
9325 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9327 sched_init_granularity();
9328 free_cpumask_var(non_isolated_cpus);
9330 init_sched_rt_class();
9333 void __init sched_init_smp(void)
9335 sched_init_granularity();
9337 #endif /* CONFIG_SMP */
9339 const_debug unsigned int sysctl_timer_migration = 1;
9341 int in_sched_functions(unsigned long addr)
9343 return in_lock_functions(addr) ||
9344 (addr >= (unsigned long)__sched_text_start
9345 && addr < (unsigned long)__sched_text_end);
9348 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9350 cfs_rq->tasks_timeline = RB_ROOT;
9351 INIT_LIST_HEAD(&cfs_rq->tasks);
9352 #ifdef CONFIG_FAIR_GROUP_SCHED
9355 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9358 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9360 struct rt_prio_array *array;
9363 array = &rt_rq->active;
9364 for (i = 0; i < MAX_RT_PRIO; i++) {
9365 INIT_LIST_HEAD(array->queue + i);
9366 __clear_bit(i, array->bitmap);
9368 /* delimiter for bitsearch: */
9369 __set_bit(MAX_RT_PRIO, array->bitmap);
9371 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9372 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9374 rt_rq->highest_prio.next = MAX_RT_PRIO;
9378 rt_rq->rt_nr_migratory = 0;
9379 rt_rq->overloaded = 0;
9380 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
9384 rt_rq->rt_throttled = 0;
9385 rt_rq->rt_runtime = 0;
9386 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
9388 #ifdef CONFIG_RT_GROUP_SCHED
9389 rt_rq->rt_nr_boosted = 0;
9394 #ifdef CONFIG_FAIR_GROUP_SCHED
9395 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9396 struct sched_entity *se, int cpu, int add,
9397 struct sched_entity *parent)
9399 struct rq *rq = cpu_rq(cpu);
9400 tg->cfs_rq[cpu] = cfs_rq;
9401 init_cfs_rq(cfs_rq, rq);
9404 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9407 /* se could be NULL for init_task_group */
9412 se->cfs_rq = &rq->cfs;
9414 se->cfs_rq = parent->my_q;
9417 se->load.weight = tg->shares;
9418 se->load.inv_weight = 0;
9419 se->parent = parent;
9423 #ifdef CONFIG_RT_GROUP_SCHED
9424 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9425 struct sched_rt_entity *rt_se, int cpu, int add,
9426 struct sched_rt_entity *parent)
9428 struct rq *rq = cpu_rq(cpu);
9430 tg->rt_rq[cpu] = rt_rq;
9431 init_rt_rq(rt_rq, rq);
9433 rt_rq->rt_se = rt_se;
9434 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9436 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9438 tg->rt_se[cpu] = rt_se;
9443 rt_se->rt_rq = &rq->rt;
9445 rt_se->rt_rq = parent->my_q;
9447 rt_se->my_q = rt_rq;
9448 rt_se->parent = parent;
9449 INIT_LIST_HEAD(&rt_se->run_list);
9453 void __init sched_init(void)
9456 unsigned long alloc_size = 0, ptr;
9458 #ifdef CONFIG_FAIR_GROUP_SCHED
9459 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9461 #ifdef CONFIG_RT_GROUP_SCHED
9462 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9464 #ifdef CONFIG_USER_SCHED
9467 #ifdef CONFIG_CPUMASK_OFFSTACK
9468 alloc_size += num_possible_cpus() * cpumask_size();
9471 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9473 #ifdef CONFIG_FAIR_GROUP_SCHED
9474 init_task_group.se = (struct sched_entity **)ptr;
9475 ptr += nr_cpu_ids * sizeof(void **);
9477 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9478 ptr += nr_cpu_ids * sizeof(void **);
9480 #ifdef CONFIG_USER_SCHED
9481 root_task_group.se = (struct sched_entity **)ptr;
9482 ptr += nr_cpu_ids * sizeof(void **);
9484 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9485 ptr += nr_cpu_ids * sizeof(void **);
9486 #endif /* CONFIG_USER_SCHED */
9487 #endif /* CONFIG_FAIR_GROUP_SCHED */
9488 #ifdef CONFIG_RT_GROUP_SCHED
9489 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9490 ptr += nr_cpu_ids * sizeof(void **);
9492 init_task_group.rt_rq = (struct rt_rq **)ptr;
9493 ptr += nr_cpu_ids * sizeof(void **);
9495 #ifdef CONFIG_USER_SCHED
9496 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9497 ptr += nr_cpu_ids * sizeof(void **);
9499 root_task_group.rt_rq = (struct rt_rq **)ptr;
9500 ptr += nr_cpu_ids * sizeof(void **);
9501 #endif /* CONFIG_USER_SCHED */
9502 #endif /* CONFIG_RT_GROUP_SCHED */
9503 #ifdef CONFIG_CPUMASK_OFFSTACK
9504 for_each_possible_cpu(i) {
9505 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9506 ptr += cpumask_size();
9508 #endif /* CONFIG_CPUMASK_OFFSTACK */
9512 init_defrootdomain();
9515 init_rt_bandwidth(&def_rt_bandwidth,
9516 global_rt_period(), global_rt_runtime());
9518 #ifdef CONFIG_RT_GROUP_SCHED
9519 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9520 global_rt_period(), global_rt_runtime());
9521 #ifdef CONFIG_USER_SCHED
9522 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9523 global_rt_period(), RUNTIME_INF);
9524 #endif /* CONFIG_USER_SCHED */
9525 #endif /* CONFIG_RT_GROUP_SCHED */
9527 #ifdef CONFIG_GROUP_SCHED
9528 list_add(&init_task_group.list, &task_groups);
9529 INIT_LIST_HEAD(&init_task_group.children);
9531 #ifdef CONFIG_USER_SCHED
9532 INIT_LIST_HEAD(&root_task_group.children);
9533 init_task_group.parent = &root_task_group;
9534 list_add(&init_task_group.siblings, &root_task_group.children);
9535 #endif /* CONFIG_USER_SCHED */
9536 #endif /* CONFIG_GROUP_SCHED */
9538 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9539 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9540 __alignof__(unsigned long));
9542 for_each_possible_cpu(i) {
9546 raw_spin_lock_init(&rq->lock);
9548 rq->calc_load_active = 0;
9549 rq->calc_load_update = jiffies + LOAD_FREQ;
9550 init_cfs_rq(&rq->cfs, rq);
9551 init_rt_rq(&rq->rt, rq);
9552 #ifdef CONFIG_FAIR_GROUP_SCHED
9553 init_task_group.shares = init_task_group_load;
9554 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9555 #ifdef CONFIG_CGROUP_SCHED
9557 * How much cpu bandwidth does init_task_group get?
9559 * In case of task-groups formed thr' the cgroup filesystem, it
9560 * gets 100% of the cpu resources in the system. This overall
9561 * system cpu resource is divided among the tasks of
9562 * init_task_group and its child task-groups in a fair manner,
9563 * based on each entity's (task or task-group's) weight
9564 * (se->load.weight).
9566 * In other words, if init_task_group has 10 tasks of weight
9567 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9568 * then A0's share of the cpu resource is:
9570 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9572 * We achieve this by letting init_task_group's tasks sit
9573 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9575 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9576 #elif defined CONFIG_USER_SCHED
9577 root_task_group.shares = NICE_0_LOAD;
9578 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9580 * In case of task-groups formed thr' the user id of tasks,
9581 * init_task_group represents tasks belonging to root user.
9582 * Hence it forms a sibling of all subsequent groups formed.
9583 * In this case, init_task_group gets only a fraction of overall
9584 * system cpu resource, based on the weight assigned to root
9585 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9586 * by letting tasks of init_task_group sit in a separate cfs_rq
9587 * (init_tg_cfs_rq) and having one entity represent this group of
9588 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9590 init_tg_cfs_entry(&init_task_group,
9591 &per_cpu(init_tg_cfs_rq, i),
9592 &per_cpu(init_sched_entity, i), i, 1,
9593 root_task_group.se[i]);
9596 #endif /* CONFIG_FAIR_GROUP_SCHED */
9598 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9599 #ifdef CONFIG_RT_GROUP_SCHED
9600 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9601 #ifdef CONFIG_CGROUP_SCHED
9602 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9603 #elif defined CONFIG_USER_SCHED
9604 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9605 init_tg_rt_entry(&init_task_group,
9606 &per_cpu(init_rt_rq_var, i),
9607 &per_cpu(init_sched_rt_entity, i), i, 1,
9608 root_task_group.rt_se[i]);
9612 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9613 rq->cpu_load[j] = 0;
9617 rq->post_schedule = 0;
9618 rq->active_balance = 0;
9619 rq->next_balance = jiffies;
9623 rq->migration_thread = NULL;
9625 rq->avg_idle = 2*sysctl_sched_migration_cost;
9626 INIT_LIST_HEAD(&rq->migration_queue);
9627 rq_attach_root(rq, &def_root_domain);
9630 atomic_set(&rq->nr_iowait, 0);
9633 set_load_weight(&init_task);
9635 #ifdef CONFIG_PREEMPT_NOTIFIERS
9636 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9640 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9643 #ifdef CONFIG_RT_MUTEXES
9644 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
9648 * The boot idle thread does lazy MMU switching as well:
9650 atomic_inc(&init_mm.mm_count);
9651 enter_lazy_tlb(&init_mm, current);
9654 * Make us the idle thread. Technically, schedule() should not be
9655 * called from this thread, however somewhere below it might be,
9656 * but because we are the idle thread, we just pick up running again
9657 * when this runqueue becomes "idle".
9659 init_idle(current, smp_processor_id());
9661 calc_load_update = jiffies + LOAD_FREQ;
9664 * During early bootup we pretend to be a normal task:
9666 current->sched_class = &fair_sched_class;
9668 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9669 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9672 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9673 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9675 /* May be allocated at isolcpus cmdline parse time */
9676 if (cpu_isolated_map == NULL)
9677 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9682 scheduler_running = 1;
9685 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9686 static inline int preempt_count_equals(int preempt_offset)
9688 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9690 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9693 void __might_sleep(char *file, int line, int preempt_offset)
9696 static unsigned long prev_jiffy; /* ratelimiting */
9698 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9699 system_state != SYSTEM_RUNNING || oops_in_progress)
9701 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9703 prev_jiffy = jiffies;
9705 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9707 pr_err("in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9708 in_atomic(), irqs_disabled(),
9709 current->pid, current->comm);
9711 debug_show_held_locks(current);
9712 if (irqs_disabled())
9713 print_irqtrace_events(current);
9717 EXPORT_SYMBOL(__might_sleep);
9720 #ifdef CONFIG_MAGIC_SYSRQ
9721 static void normalize_task(struct rq *rq, struct task_struct *p)
9725 update_rq_clock(rq);
9726 on_rq = p->se.on_rq;
9728 deactivate_task(rq, p, 0);
9729 __setscheduler(rq, p, SCHED_NORMAL, 0);
9731 activate_task(rq, p, 0);
9732 resched_task(rq->curr);
9736 void normalize_rt_tasks(void)
9738 struct task_struct *g, *p;
9739 unsigned long flags;
9742 read_lock_irqsave(&tasklist_lock, flags);
9743 do_each_thread(g, p) {
9745 * Only normalize user tasks:
9750 p->se.exec_start = 0;
9751 #ifdef CONFIG_SCHEDSTATS
9752 p->se.wait_start = 0;
9753 p->se.sleep_start = 0;
9754 p->se.block_start = 0;
9759 * Renice negative nice level userspace
9762 if (TASK_NICE(p) < 0 && p->mm)
9763 set_user_nice(p, 0);
9767 raw_spin_lock(&p->pi_lock);
9768 rq = __task_rq_lock(p);
9770 normalize_task(rq, p);
9772 __task_rq_unlock(rq);
9773 raw_spin_unlock(&p->pi_lock);
9774 } while_each_thread(g, p);
9776 read_unlock_irqrestore(&tasklist_lock, flags);
9779 #endif /* CONFIG_MAGIC_SYSRQ */
9783 * These functions are only useful for the IA64 MCA handling.
9785 * They can only be called when the whole system has been
9786 * stopped - every CPU needs to be quiescent, and no scheduling
9787 * activity can take place. Using them for anything else would
9788 * be a serious bug, and as a result, they aren't even visible
9789 * under any other configuration.
9793 * curr_task - return the current task for a given cpu.
9794 * @cpu: the processor in question.
9796 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9798 struct task_struct *curr_task(int cpu)
9800 return cpu_curr(cpu);
9804 * set_curr_task - set the current task for a given cpu.
9805 * @cpu: the processor in question.
9806 * @p: the task pointer to set.
9808 * Description: This function must only be used when non-maskable interrupts
9809 * are serviced on a separate stack. It allows the architecture to switch the
9810 * notion of the current task on a cpu in a non-blocking manner. This function
9811 * must be called with all CPU's synchronized, and interrupts disabled, the
9812 * and caller must save the original value of the current task (see
9813 * curr_task() above) and restore that value before reenabling interrupts and
9814 * re-starting the system.
9816 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9818 void set_curr_task(int cpu, struct task_struct *p)
9825 #ifdef CONFIG_FAIR_GROUP_SCHED
9826 static void free_fair_sched_group(struct task_group *tg)
9830 for_each_possible_cpu(i) {
9832 kfree(tg->cfs_rq[i]);
9842 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9844 struct cfs_rq *cfs_rq;
9845 struct sched_entity *se;
9849 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9852 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9856 tg->shares = NICE_0_LOAD;
9858 for_each_possible_cpu(i) {
9861 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9862 GFP_KERNEL, cpu_to_node(i));
9866 se = kzalloc_node(sizeof(struct sched_entity),
9867 GFP_KERNEL, cpu_to_node(i));
9871 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9882 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9884 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9885 &cpu_rq(cpu)->leaf_cfs_rq_list);
9888 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9890 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9892 #else /* !CONFG_FAIR_GROUP_SCHED */
9893 static inline void free_fair_sched_group(struct task_group *tg)
9898 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9903 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9907 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9910 #endif /* CONFIG_FAIR_GROUP_SCHED */
9912 #ifdef CONFIG_RT_GROUP_SCHED
9913 static void free_rt_sched_group(struct task_group *tg)
9917 destroy_rt_bandwidth(&tg->rt_bandwidth);
9919 for_each_possible_cpu(i) {
9921 kfree(tg->rt_rq[i]);
9923 kfree(tg->rt_se[i]);
9931 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9933 struct rt_rq *rt_rq;
9934 struct sched_rt_entity *rt_se;
9938 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9941 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9945 init_rt_bandwidth(&tg->rt_bandwidth,
9946 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9948 for_each_possible_cpu(i) {
9951 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9952 GFP_KERNEL, cpu_to_node(i));
9956 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9957 GFP_KERNEL, cpu_to_node(i));
9961 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9972 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9974 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9975 &cpu_rq(cpu)->leaf_rt_rq_list);
9978 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9980 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9982 #else /* !CONFIG_RT_GROUP_SCHED */
9983 static inline void free_rt_sched_group(struct task_group *tg)
9988 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9993 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9997 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10000 #endif /* CONFIG_RT_GROUP_SCHED */
10002 #ifdef CONFIG_GROUP_SCHED
10003 static void free_sched_group(struct task_group *tg)
10005 free_fair_sched_group(tg);
10006 free_rt_sched_group(tg);
10010 /* allocate runqueue etc for a new task group */
10011 struct task_group *sched_create_group(struct task_group *parent)
10013 struct task_group *tg;
10014 unsigned long flags;
10017 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10019 return ERR_PTR(-ENOMEM);
10021 if (!alloc_fair_sched_group(tg, parent))
10024 if (!alloc_rt_sched_group(tg, parent))
10027 spin_lock_irqsave(&task_group_lock, flags);
10028 for_each_possible_cpu(i) {
10029 register_fair_sched_group(tg, i);
10030 register_rt_sched_group(tg, i);
10032 list_add_rcu(&tg->list, &task_groups);
10034 WARN_ON(!parent); /* root should already exist */
10036 tg->parent = parent;
10037 INIT_LIST_HEAD(&tg->children);
10038 list_add_rcu(&tg->siblings, &parent->children);
10039 spin_unlock_irqrestore(&task_group_lock, flags);
10044 free_sched_group(tg);
10045 return ERR_PTR(-ENOMEM);
10048 /* rcu callback to free various structures associated with a task group */
10049 static void free_sched_group_rcu(struct rcu_head *rhp)
10051 /* now it should be safe to free those cfs_rqs */
10052 free_sched_group(container_of(rhp, struct task_group, rcu));
10055 /* Destroy runqueue etc associated with a task group */
10056 void sched_destroy_group(struct task_group *tg)
10058 unsigned long flags;
10061 spin_lock_irqsave(&task_group_lock, flags);
10062 for_each_possible_cpu(i) {
10063 unregister_fair_sched_group(tg, i);
10064 unregister_rt_sched_group(tg, i);
10066 list_del_rcu(&tg->list);
10067 list_del_rcu(&tg->siblings);
10068 spin_unlock_irqrestore(&task_group_lock, flags);
10070 /* wait for possible concurrent references to cfs_rqs complete */
10071 call_rcu(&tg->rcu, free_sched_group_rcu);
10074 /* change task's runqueue when it moves between groups.
10075 * The caller of this function should have put the task in its new group
10076 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10077 * reflect its new group.
10079 void sched_move_task(struct task_struct *tsk)
10081 int on_rq, running;
10082 unsigned long flags;
10085 rq = task_rq_lock(tsk, &flags);
10087 update_rq_clock(rq);
10089 running = task_current(rq, tsk);
10090 on_rq = tsk->se.on_rq;
10093 dequeue_task(rq, tsk, 0);
10094 if (unlikely(running))
10095 tsk->sched_class->put_prev_task(rq, tsk);
10097 set_task_rq(tsk, task_cpu(tsk));
10099 #ifdef CONFIG_FAIR_GROUP_SCHED
10100 if (tsk->sched_class->moved_group)
10101 tsk->sched_class->moved_group(tsk);
10104 if (unlikely(running))
10105 tsk->sched_class->set_curr_task(rq);
10107 enqueue_task(rq, tsk, 0);
10109 task_rq_unlock(rq, &flags);
10111 #endif /* CONFIG_GROUP_SCHED */
10113 #ifdef CONFIG_FAIR_GROUP_SCHED
10114 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10116 struct cfs_rq *cfs_rq = se->cfs_rq;
10121 dequeue_entity(cfs_rq, se, 0);
10123 se->load.weight = shares;
10124 se->load.inv_weight = 0;
10127 enqueue_entity(cfs_rq, se, 0);
10130 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10132 struct cfs_rq *cfs_rq = se->cfs_rq;
10133 struct rq *rq = cfs_rq->rq;
10134 unsigned long flags;
10136 raw_spin_lock_irqsave(&rq->lock, flags);
10137 __set_se_shares(se, shares);
10138 raw_spin_unlock_irqrestore(&rq->lock, flags);
10141 static DEFINE_MUTEX(shares_mutex);
10143 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10146 unsigned long flags;
10149 * We can't change the weight of the root cgroup.
10154 if (shares < MIN_SHARES)
10155 shares = MIN_SHARES;
10156 else if (shares > MAX_SHARES)
10157 shares = MAX_SHARES;
10159 mutex_lock(&shares_mutex);
10160 if (tg->shares == shares)
10163 spin_lock_irqsave(&task_group_lock, flags);
10164 for_each_possible_cpu(i)
10165 unregister_fair_sched_group(tg, i);
10166 list_del_rcu(&tg->siblings);
10167 spin_unlock_irqrestore(&task_group_lock, flags);
10169 /* wait for any ongoing reference to this group to finish */
10170 synchronize_sched();
10173 * Now we are free to modify the group's share on each cpu
10174 * w/o tripping rebalance_share or load_balance_fair.
10176 tg->shares = shares;
10177 for_each_possible_cpu(i) {
10179 * force a rebalance
10181 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10182 set_se_shares(tg->se[i], shares);
10186 * Enable load balance activity on this group, by inserting it back on
10187 * each cpu's rq->leaf_cfs_rq_list.
10189 spin_lock_irqsave(&task_group_lock, flags);
10190 for_each_possible_cpu(i)
10191 register_fair_sched_group(tg, i);
10192 list_add_rcu(&tg->siblings, &tg->parent->children);
10193 spin_unlock_irqrestore(&task_group_lock, flags);
10195 mutex_unlock(&shares_mutex);
10199 unsigned long sched_group_shares(struct task_group *tg)
10205 #ifdef CONFIG_RT_GROUP_SCHED
10207 * Ensure that the real time constraints are schedulable.
10209 static DEFINE_MUTEX(rt_constraints_mutex);
10211 static unsigned long to_ratio(u64 period, u64 runtime)
10213 if (runtime == RUNTIME_INF)
10216 return div64_u64(runtime << 20, period);
10219 /* Must be called with tasklist_lock held */
10220 static inline int tg_has_rt_tasks(struct task_group *tg)
10222 struct task_struct *g, *p;
10224 do_each_thread(g, p) {
10225 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10227 } while_each_thread(g, p);
10232 struct rt_schedulable_data {
10233 struct task_group *tg;
10238 static int tg_schedulable(struct task_group *tg, void *data)
10240 struct rt_schedulable_data *d = data;
10241 struct task_group *child;
10242 unsigned long total, sum = 0;
10243 u64 period, runtime;
10245 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10246 runtime = tg->rt_bandwidth.rt_runtime;
10249 period = d->rt_period;
10250 runtime = d->rt_runtime;
10253 #ifdef CONFIG_USER_SCHED
10254 if (tg == &root_task_group) {
10255 period = global_rt_period();
10256 runtime = global_rt_runtime();
10261 * Cannot have more runtime than the period.
10263 if (runtime > period && runtime != RUNTIME_INF)
10267 * Ensure we don't starve existing RT tasks.
10269 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10272 total = to_ratio(period, runtime);
10275 * Nobody can have more than the global setting allows.
10277 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10281 * The sum of our children's runtime should not exceed our own.
10283 list_for_each_entry_rcu(child, &tg->children, siblings) {
10284 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10285 runtime = child->rt_bandwidth.rt_runtime;
10287 if (child == d->tg) {
10288 period = d->rt_period;
10289 runtime = d->rt_runtime;
10292 sum += to_ratio(period, runtime);
10301 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10303 struct rt_schedulable_data data = {
10305 .rt_period = period,
10306 .rt_runtime = runtime,
10309 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10312 static int tg_set_bandwidth(struct task_group *tg,
10313 u64 rt_period, u64 rt_runtime)
10317 mutex_lock(&rt_constraints_mutex);
10318 read_lock(&tasklist_lock);
10319 err = __rt_schedulable(tg, rt_period, rt_runtime);
10323 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10324 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10325 tg->rt_bandwidth.rt_runtime = rt_runtime;
10327 for_each_possible_cpu(i) {
10328 struct rt_rq *rt_rq = tg->rt_rq[i];
10330 raw_spin_lock(&rt_rq->rt_runtime_lock);
10331 rt_rq->rt_runtime = rt_runtime;
10332 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10334 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10336 read_unlock(&tasklist_lock);
10337 mutex_unlock(&rt_constraints_mutex);
10342 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10344 u64 rt_runtime, rt_period;
10346 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10347 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10348 if (rt_runtime_us < 0)
10349 rt_runtime = RUNTIME_INF;
10351 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10354 long sched_group_rt_runtime(struct task_group *tg)
10358 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10361 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10362 do_div(rt_runtime_us, NSEC_PER_USEC);
10363 return rt_runtime_us;
10366 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10368 u64 rt_runtime, rt_period;
10370 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10371 rt_runtime = tg->rt_bandwidth.rt_runtime;
10373 if (rt_period == 0)
10376 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10379 long sched_group_rt_period(struct task_group *tg)
10383 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10384 do_div(rt_period_us, NSEC_PER_USEC);
10385 return rt_period_us;
10388 static int sched_rt_global_constraints(void)
10390 u64 runtime, period;
10393 if (sysctl_sched_rt_period <= 0)
10396 runtime = global_rt_runtime();
10397 period = global_rt_period();
10400 * Sanity check on the sysctl variables.
10402 if (runtime > period && runtime != RUNTIME_INF)
10405 mutex_lock(&rt_constraints_mutex);
10406 read_lock(&tasklist_lock);
10407 ret = __rt_schedulable(NULL, 0, 0);
10408 read_unlock(&tasklist_lock);
10409 mutex_unlock(&rt_constraints_mutex);
10414 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10416 /* Don't accept realtime tasks when there is no way for them to run */
10417 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10423 #else /* !CONFIG_RT_GROUP_SCHED */
10424 static int sched_rt_global_constraints(void)
10426 unsigned long flags;
10429 if (sysctl_sched_rt_period <= 0)
10433 * There's always some RT tasks in the root group
10434 * -- migration, kstopmachine etc..
10436 if (sysctl_sched_rt_runtime == 0)
10439 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10440 for_each_possible_cpu(i) {
10441 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10443 raw_spin_lock(&rt_rq->rt_runtime_lock);
10444 rt_rq->rt_runtime = global_rt_runtime();
10445 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10447 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10451 #endif /* CONFIG_RT_GROUP_SCHED */
10453 int sched_rt_handler(struct ctl_table *table, int write,
10454 void __user *buffer, size_t *lenp,
10458 int old_period, old_runtime;
10459 static DEFINE_MUTEX(mutex);
10461 mutex_lock(&mutex);
10462 old_period = sysctl_sched_rt_period;
10463 old_runtime = sysctl_sched_rt_runtime;
10465 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10467 if (!ret && write) {
10468 ret = sched_rt_global_constraints();
10470 sysctl_sched_rt_period = old_period;
10471 sysctl_sched_rt_runtime = old_runtime;
10473 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10474 def_rt_bandwidth.rt_period =
10475 ns_to_ktime(global_rt_period());
10478 mutex_unlock(&mutex);
10483 #ifdef CONFIG_CGROUP_SCHED
10485 /* return corresponding task_group object of a cgroup */
10486 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10488 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10489 struct task_group, css);
10492 static struct cgroup_subsys_state *
10493 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10495 struct task_group *tg, *parent;
10497 if (!cgrp->parent) {
10498 /* This is early initialization for the top cgroup */
10499 return &init_task_group.css;
10502 parent = cgroup_tg(cgrp->parent);
10503 tg = sched_create_group(parent);
10505 return ERR_PTR(-ENOMEM);
10511 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10513 struct task_group *tg = cgroup_tg(cgrp);
10515 sched_destroy_group(tg);
10519 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10521 #ifdef CONFIG_RT_GROUP_SCHED
10522 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10525 /* We don't support RT-tasks being in separate groups */
10526 if (tsk->sched_class != &fair_sched_class)
10533 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10534 struct task_struct *tsk, bool threadgroup)
10536 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10540 struct task_struct *c;
10542 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10543 retval = cpu_cgroup_can_attach_task(cgrp, c);
10555 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10556 struct cgroup *old_cont, struct task_struct *tsk,
10559 sched_move_task(tsk);
10561 struct task_struct *c;
10563 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10564 sched_move_task(c);
10570 #ifdef CONFIG_FAIR_GROUP_SCHED
10571 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10574 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10577 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10579 struct task_group *tg = cgroup_tg(cgrp);
10581 return (u64) tg->shares;
10583 #endif /* CONFIG_FAIR_GROUP_SCHED */
10585 #ifdef CONFIG_RT_GROUP_SCHED
10586 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10589 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10592 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10594 return sched_group_rt_runtime(cgroup_tg(cgrp));
10597 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10600 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10603 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10605 return sched_group_rt_period(cgroup_tg(cgrp));
10607 #endif /* CONFIG_RT_GROUP_SCHED */
10609 static struct cftype cpu_files[] = {
10610 #ifdef CONFIG_FAIR_GROUP_SCHED
10613 .read_u64 = cpu_shares_read_u64,
10614 .write_u64 = cpu_shares_write_u64,
10617 #ifdef CONFIG_RT_GROUP_SCHED
10619 .name = "rt_runtime_us",
10620 .read_s64 = cpu_rt_runtime_read,
10621 .write_s64 = cpu_rt_runtime_write,
10624 .name = "rt_period_us",
10625 .read_u64 = cpu_rt_period_read_uint,
10626 .write_u64 = cpu_rt_period_write_uint,
10631 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10633 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10636 struct cgroup_subsys cpu_cgroup_subsys = {
10638 .create = cpu_cgroup_create,
10639 .destroy = cpu_cgroup_destroy,
10640 .can_attach = cpu_cgroup_can_attach,
10641 .attach = cpu_cgroup_attach,
10642 .populate = cpu_cgroup_populate,
10643 .subsys_id = cpu_cgroup_subsys_id,
10647 #endif /* CONFIG_CGROUP_SCHED */
10649 #ifdef CONFIG_CGROUP_CPUACCT
10652 * CPU accounting code for task groups.
10654 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10655 * (balbir@in.ibm.com).
10658 /* track cpu usage of a group of tasks and its child groups */
10660 struct cgroup_subsys_state css;
10661 /* cpuusage holds pointer to a u64-type object on every cpu */
10663 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10664 struct cpuacct *parent;
10667 struct cgroup_subsys cpuacct_subsys;
10669 /* return cpu accounting group corresponding to this container */
10670 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10672 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10673 struct cpuacct, css);
10676 /* return cpu accounting group to which this task belongs */
10677 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10679 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10680 struct cpuacct, css);
10683 /* create a new cpu accounting group */
10684 static struct cgroup_subsys_state *cpuacct_create(
10685 struct cgroup_subsys *ss, struct cgroup *cgrp)
10687 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10693 ca->cpuusage = alloc_percpu(u64);
10697 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10698 if (percpu_counter_init(&ca->cpustat[i], 0))
10699 goto out_free_counters;
10702 ca->parent = cgroup_ca(cgrp->parent);
10708 percpu_counter_destroy(&ca->cpustat[i]);
10709 free_percpu(ca->cpuusage);
10713 return ERR_PTR(-ENOMEM);
10716 /* destroy an existing cpu accounting group */
10718 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10720 struct cpuacct *ca = cgroup_ca(cgrp);
10723 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10724 percpu_counter_destroy(&ca->cpustat[i]);
10725 free_percpu(ca->cpuusage);
10729 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10731 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10734 #ifndef CONFIG_64BIT
10736 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10738 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10740 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10748 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10750 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10752 #ifndef CONFIG_64BIT
10754 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10756 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10758 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10764 /* return total cpu usage (in nanoseconds) of a group */
10765 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10767 struct cpuacct *ca = cgroup_ca(cgrp);
10768 u64 totalcpuusage = 0;
10771 for_each_present_cpu(i)
10772 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10774 return totalcpuusage;
10777 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10780 struct cpuacct *ca = cgroup_ca(cgrp);
10789 for_each_present_cpu(i)
10790 cpuacct_cpuusage_write(ca, i, 0);
10796 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10797 struct seq_file *m)
10799 struct cpuacct *ca = cgroup_ca(cgroup);
10803 for_each_present_cpu(i) {
10804 percpu = cpuacct_cpuusage_read(ca, i);
10805 seq_printf(m, "%llu ", (unsigned long long) percpu);
10807 seq_printf(m, "\n");
10811 static const char *cpuacct_stat_desc[] = {
10812 [CPUACCT_STAT_USER] = "user",
10813 [CPUACCT_STAT_SYSTEM] = "system",
10816 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10817 struct cgroup_map_cb *cb)
10819 struct cpuacct *ca = cgroup_ca(cgrp);
10822 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10823 s64 val = percpu_counter_read(&ca->cpustat[i]);
10824 val = cputime64_to_clock_t(val);
10825 cb->fill(cb, cpuacct_stat_desc[i], val);
10830 static struct cftype files[] = {
10833 .read_u64 = cpuusage_read,
10834 .write_u64 = cpuusage_write,
10837 .name = "usage_percpu",
10838 .read_seq_string = cpuacct_percpu_seq_read,
10842 .read_map = cpuacct_stats_show,
10846 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10848 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10852 * charge this task's execution time to its accounting group.
10854 * called with rq->lock held.
10856 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10858 struct cpuacct *ca;
10861 if (unlikely(!cpuacct_subsys.active))
10864 cpu = task_cpu(tsk);
10870 for (; ca; ca = ca->parent) {
10871 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10872 *cpuusage += cputime;
10879 * Charge the system/user time to the task's accounting group.
10881 static void cpuacct_update_stats(struct task_struct *tsk,
10882 enum cpuacct_stat_index idx, cputime_t val)
10884 struct cpuacct *ca;
10886 if (unlikely(!cpuacct_subsys.active))
10893 percpu_counter_add(&ca->cpustat[idx], val);
10899 struct cgroup_subsys cpuacct_subsys = {
10901 .create = cpuacct_create,
10902 .destroy = cpuacct_destroy,
10903 .populate = cpuacct_populate,
10904 .subsys_id = cpuacct_subsys_id,
10906 #endif /* CONFIG_CGROUP_CPUACCT */
10910 int rcu_expedited_torture_stats(char *page)
10914 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10916 void synchronize_sched_expedited(void)
10919 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10921 #else /* #ifndef CONFIG_SMP */
10923 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10924 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10926 #define RCU_EXPEDITED_STATE_POST -2
10927 #define RCU_EXPEDITED_STATE_IDLE -1
10929 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10931 int rcu_expedited_torture_stats(char *page)
10936 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10937 for_each_online_cpu(cpu) {
10938 cnt += sprintf(&page[cnt], " %d:%d",
10939 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10941 cnt += sprintf(&page[cnt], "\n");
10944 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10946 static long synchronize_sched_expedited_count;
10949 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10950 * approach to force grace period to end quickly. This consumes
10951 * significant time on all CPUs, and is thus not recommended for
10952 * any sort of common-case code.
10954 * Note that it is illegal to call this function while holding any
10955 * lock that is acquired by a CPU-hotplug notifier. Failing to
10956 * observe this restriction will result in deadlock.
10958 void synchronize_sched_expedited(void)
10961 unsigned long flags;
10962 bool need_full_sync = 0;
10964 struct migration_req *req;
10968 smp_mb(); /* ensure prior mod happens before capturing snap. */
10969 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10971 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10973 if (trycount++ < 10)
10974 udelay(trycount * num_online_cpus());
10976 synchronize_sched();
10979 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10980 smp_mb(); /* ensure test happens before caller kfree */
10985 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10986 for_each_online_cpu(cpu) {
10988 req = &per_cpu(rcu_migration_req, cpu);
10989 init_completion(&req->done);
10991 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10992 raw_spin_lock_irqsave(&rq->lock, flags);
10993 list_add(&req->list, &rq->migration_queue);
10994 raw_spin_unlock_irqrestore(&rq->lock, flags);
10995 wake_up_process(rq->migration_thread);
10997 for_each_online_cpu(cpu) {
10998 rcu_expedited_state = cpu;
10999 req = &per_cpu(rcu_migration_req, cpu);
11001 wait_for_completion(&req->done);
11002 raw_spin_lock_irqsave(&rq->lock, flags);
11003 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
11004 need_full_sync = 1;
11005 req->dest_cpu = RCU_MIGRATION_IDLE;
11006 raw_spin_unlock_irqrestore(&rq->lock, flags);
11008 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
11009 synchronize_sched_expedited_count++;
11010 mutex_unlock(&rcu_sched_expedited_mutex);
11012 if (need_full_sync)
11013 synchronize_sched();
11015 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
11017 #endif /* #else #ifndef CONFIG_SMP */