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);
2008 * kthread_bind - bind a just-created kthread to a cpu.
2009 * @p: thread created by kthread_create().
2010 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2012 * Description: This function is equivalent to set_cpus_allowed(),
2013 * except that @cpu doesn't need to be online, and the thread must be
2014 * stopped (i.e., just returned from kthread_create()).
2016 * Function lives here instead of kthread.c because it messes with
2017 * scheduler internals which require locking.
2019 void kthread_bind(struct task_struct *p, unsigned int cpu)
2021 struct rq *rq = cpu_rq(cpu);
2022 unsigned long flags;
2024 /* Must have done schedule() in kthread() before we set_task_cpu */
2025 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2030 raw_spin_lock_irqsave(&rq->lock, flags);
2031 update_rq_clock(rq);
2032 set_task_cpu(p, cpu);
2033 p->cpus_allowed = cpumask_of_cpu(cpu);
2034 p->rt.nr_cpus_allowed = 1;
2035 p->flags |= PF_THREAD_BOUND;
2036 raw_spin_unlock_irqrestore(&rq->lock, flags);
2038 EXPORT_SYMBOL(kthread_bind);
2042 * Is this task likely cache-hot:
2045 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2050 * Buddy candidates are cache hot:
2052 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
2053 (&p->se == cfs_rq_of(&p->se)->next ||
2054 &p->se == cfs_rq_of(&p->se)->last))
2057 if (p->sched_class != &fair_sched_class)
2060 if (sysctl_sched_migration_cost == -1)
2062 if (sysctl_sched_migration_cost == 0)
2065 delta = now - p->se.exec_start;
2067 return delta < (s64)sysctl_sched_migration_cost;
2071 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2073 int old_cpu = task_cpu(p);
2074 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2075 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2077 trace_sched_migrate_task(p, new_cpu);
2079 if (old_cpu != new_cpu) {
2080 p->se.nr_migrations++;
2081 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2084 p->se.vruntime -= old_cfsrq->min_vruntime -
2085 new_cfsrq->min_vruntime;
2087 __set_task_cpu(p, new_cpu);
2090 struct migration_req {
2091 struct list_head list;
2093 struct task_struct *task;
2096 struct completion done;
2100 * The task's runqueue lock must be held.
2101 * Returns true if you have to wait for migration thread.
2104 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2106 struct rq *rq = task_rq(p);
2109 * If the task is not on a runqueue (and not running), then
2110 * it is sufficient to simply update the task's cpu field.
2112 if (!p->se.on_rq && !task_running(rq, p)) {
2113 update_rq_clock(rq);
2114 set_task_cpu(p, dest_cpu);
2118 init_completion(&req->done);
2120 req->dest_cpu = dest_cpu;
2121 list_add(&req->list, &rq->migration_queue);
2127 * wait_task_context_switch - wait for a thread to complete at least one
2130 * @p must not be current.
2132 void wait_task_context_switch(struct task_struct *p)
2134 unsigned long nvcsw, nivcsw, flags;
2142 * The runqueue is assigned before the actual context
2143 * switch. We need to take the runqueue lock.
2145 * We could check initially without the lock but it is
2146 * very likely that we need to take the lock in every
2149 rq = task_rq_lock(p, &flags);
2150 running = task_running(rq, p);
2151 task_rq_unlock(rq, &flags);
2153 if (likely(!running))
2156 * The switch count is incremented before the actual
2157 * context switch. We thus wait for two switches to be
2158 * sure at least one completed.
2160 if ((p->nvcsw - nvcsw) > 1)
2162 if ((p->nivcsw - nivcsw) > 1)
2170 * wait_task_inactive - wait for a thread to unschedule.
2172 * If @match_state is nonzero, it's the @p->state value just checked and
2173 * not expected to change. If it changes, i.e. @p might have woken up,
2174 * then return zero. When we succeed in waiting for @p to be off its CPU,
2175 * we return a positive number (its total switch count). If a second call
2176 * a short while later returns the same number, the caller can be sure that
2177 * @p has remained unscheduled the whole time.
2179 * The caller must ensure that the task *will* unschedule sometime soon,
2180 * else this function might spin for a *long* time. This function can't
2181 * be called with interrupts off, or it may introduce deadlock with
2182 * smp_call_function() if an IPI is sent by the same process we are
2183 * waiting to become inactive.
2185 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2187 unsigned long flags;
2194 * We do the initial early heuristics without holding
2195 * any task-queue locks at all. We'll only try to get
2196 * the runqueue lock when things look like they will
2202 * If the task is actively running on another CPU
2203 * still, just relax and busy-wait without holding
2206 * NOTE! Since we don't hold any locks, it's not
2207 * even sure that "rq" stays as the right runqueue!
2208 * But we don't care, since "task_running()" will
2209 * return false if the runqueue has changed and p
2210 * is actually now running somewhere else!
2212 while (task_running(rq, p)) {
2213 if (match_state && unlikely(p->state != match_state))
2219 * Ok, time to look more closely! We need the rq
2220 * lock now, to be *sure*. If we're wrong, we'll
2221 * just go back and repeat.
2223 rq = task_rq_lock(p, &flags);
2224 trace_sched_wait_task(rq, p);
2225 running = task_running(rq, p);
2226 on_rq = p->se.on_rq;
2228 if (!match_state || p->state == match_state)
2229 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2230 task_rq_unlock(rq, &flags);
2233 * If it changed from the expected state, bail out now.
2235 if (unlikely(!ncsw))
2239 * Was it really running after all now that we
2240 * checked with the proper locks actually held?
2242 * Oops. Go back and try again..
2244 if (unlikely(running)) {
2250 * It's not enough that it's not actively running,
2251 * it must be off the runqueue _entirely_, and not
2254 * So if it was still runnable (but just not actively
2255 * running right now), it's preempted, and we should
2256 * yield - it could be a while.
2258 if (unlikely(on_rq)) {
2259 schedule_timeout_uninterruptible(1);
2264 * Ahh, all good. It wasn't running, and it wasn't
2265 * runnable, which means that it will never become
2266 * running in the future either. We're all done!
2275 * kick_process - kick a running thread to enter/exit the kernel
2276 * @p: the to-be-kicked thread
2278 * Cause a process which is running on another CPU to enter
2279 * kernel-mode, without any delay. (to get signals handled.)
2281 * NOTE: this function doesnt have to take the runqueue lock,
2282 * because all it wants to ensure is that the remote task enters
2283 * the kernel. If the IPI races and the task has been migrated
2284 * to another CPU then no harm is done and the purpose has been
2287 void kick_process(struct task_struct *p)
2293 if ((cpu != smp_processor_id()) && task_curr(p))
2294 smp_send_reschedule(cpu);
2297 EXPORT_SYMBOL_GPL(kick_process);
2298 #endif /* CONFIG_SMP */
2301 * task_oncpu_function_call - call a function on the cpu on which a task runs
2302 * @p: the task to evaluate
2303 * @func: the function to be called
2304 * @info: the function call argument
2306 * Calls the function @func when the task is currently running. This might
2307 * be on the current CPU, which just calls the function directly
2309 void task_oncpu_function_call(struct task_struct *p,
2310 void (*func) (void *info), void *info)
2317 smp_call_function_single(cpu, func, info, 1);
2323 int select_task_rq(struct task_struct *p, int sd_flags, int wake_flags)
2325 return p->sched_class->select_task_rq(p, sd_flags, wake_flags);
2330 * try_to_wake_up - wake up a thread
2331 * @p: the to-be-woken-up thread
2332 * @state: the mask of task states that can be woken
2333 * @sync: do a synchronous wakeup?
2335 * Put it on the run-queue if it's not already there. The "current"
2336 * thread is always on the run-queue (except when the actual
2337 * re-schedule is in progress), and as such you're allowed to do
2338 * the simpler "current->state = TASK_RUNNING" to mark yourself
2339 * runnable without the overhead of this.
2341 * returns failure only if the task is already active.
2343 static int try_to_wake_up(struct task_struct *p, unsigned int state,
2346 int cpu, orig_cpu, this_cpu, success = 0;
2347 unsigned long flags;
2348 struct rq *rq, *orig_rq;
2350 if (!sched_feat(SYNC_WAKEUPS))
2351 wake_flags &= ~WF_SYNC;
2353 this_cpu = get_cpu();
2356 rq = orig_rq = task_rq_lock(p, &flags);
2357 update_rq_clock(rq);
2358 if (!(p->state & state))
2368 if (unlikely(task_running(rq, p)))
2372 * In order to handle concurrent wakeups and release the rq->lock
2373 * we put the task in TASK_WAKING state.
2375 * First fix up the nr_uninterruptible count:
2377 if (task_contributes_to_load(p))
2378 rq->nr_uninterruptible--;
2379 p->state = TASK_WAKING;
2380 __task_rq_unlock(rq);
2382 cpu = select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
2383 if (cpu != orig_cpu)
2384 set_task_cpu(p, cpu);
2386 rq = __task_rq_lock(p);
2387 update_rq_clock(rq);
2389 WARN_ON(p->state != TASK_WAKING);
2392 #ifdef CONFIG_SCHEDSTATS
2393 schedstat_inc(rq, ttwu_count);
2394 if (cpu == this_cpu)
2395 schedstat_inc(rq, ttwu_local);
2397 struct sched_domain *sd;
2398 for_each_domain(this_cpu, sd) {
2399 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2400 schedstat_inc(sd, ttwu_wake_remote);
2405 #endif /* CONFIG_SCHEDSTATS */
2408 #endif /* CONFIG_SMP */
2409 schedstat_inc(p, se.nr_wakeups);
2410 if (wake_flags & WF_SYNC)
2411 schedstat_inc(p, se.nr_wakeups_sync);
2412 if (orig_cpu != cpu)
2413 schedstat_inc(p, se.nr_wakeups_migrate);
2414 if (cpu == this_cpu)
2415 schedstat_inc(p, se.nr_wakeups_local);
2417 schedstat_inc(p, se.nr_wakeups_remote);
2418 activate_task(rq, p, 1);
2422 * Only attribute actual wakeups done by this task.
2424 if (!in_interrupt()) {
2425 struct sched_entity *se = ¤t->se;
2426 u64 sample = se->sum_exec_runtime;
2428 if (se->last_wakeup)
2429 sample -= se->last_wakeup;
2431 sample -= se->start_runtime;
2432 update_avg(&se->avg_wakeup, sample);
2434 se->last_wakeup = se->sum_exec_runtime;
2438 trace_sched_wakeup(rq, p, success);
2439 check_preempt_curr(rq, p, wake_flags);
2441 p->state = TASK_RUNNING;
2443 if (p->sched_class->task_wake_up)
2444 p->sched_class->task_wake_up(rq, p);
2446 if (unlikely(rq->idle_stamp)) {
2447 u64 delta = rq->clock - rq->idle_stamp;
2448 u64 max = 2*sysctl_sched_migration_cost;
2453 update_avg(&rq->avg_idle, delta);
2458 task_rq_unlock(rq, &flags);
2465 * wake_up_process - Wake up a specific process
2466 * @p: The process to be woken up.
2468 * Attempt to wake up the nominated process and move it to the set of runnable
2469 * processes. Returns 1 if the process was woken up, 0 if it was already
2472 * It may be assumed that this function implies a write memory barrier before
2473 * changing the task state if and only if any tasks are woken up.
2475 int wake_up_process(struct task_struct *p)
2477 return try_to_wake_up(p, TASK_ALL, 0);
2479 EXPORT_SYMBOL(wake_up_process);
2481 int wake_up_state(struct task_struct *p, unsigned int state)
2483 return try_to_wake_up(p, state, 0);
2487 * Perform scheduler related setup for a newly forked process p.
2488 * p is forked by current.
2490 * __sched_fork() is basic setup used by init_idle() too:
2492 static void __sched_fork(struct task_struct *p)
2494 p->se.exec_start = 0;
2495 p->se.sum_exec_runtime = 0;
2496 p->se.prev_sum_exec_runtime = 0;
2497 p->se.nr_migrations = 0;
2498 p->se.last_wakeup = 0;
2499 p->se.avg_overlap = 0;
2500 p->se.start_runtime = 0;
2501 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2503 #ifdef CONFIG_SCHEDSTATS
2504 p->se.wait_start = 0;
2506 p->se.wait_count = 0;
2509 p->se.sleep_start = 0;
2510 p->se.sleep_max = 0;
2511 p->se.sum_sleep_runtime = 0;
2513 p->se.block_start = 0;
2514 p->se.block_max = 0;
2516 p->se.slice_max = 0;
2518 p->se.nr_migrations_cold = 0;
2519 p->se.nr_failed_migrations_affine = 0;
2520 p->se.nr_failed_migrations_running = 0;
2521 p->se.nr_failed_migrations_hot = 0;
2522 p->se.nr_forced_migrations = 0;
2524 p->se.nr_wakeups = 0;
2525 p->se.nr_wakeups_sync = 0;
2526 p->se.nr_wakeups_migrate = 0;
2527 p->se.nr_wakeups_local = 0;
2528 p->se.nr_wakeups_remote = 0;
2529 p->se.nr_wakeups_affine = 0;
2530 p->se.nr_wakeups_affine_attempts = 0;
2531 p->se.nr_wakeups_passive = 0;
2532 p->se.nr_wakeups_idle = 0;
2536 INIT_LIST_HEAD(&p->rt.run_list);
2538 INIT_LIST_HEAD(&p->se.group_node);
2540 #ifdef CONFIG_PREEMPT_NOTIFIERS
2541 INIT_HLIST_HEAD(&p->preempt_notifiers);
2545 * We mark the process as running here, but have not actually
2546 * inserted it onto the runqueue yet. This guarantees that
2547 * nobody will actually run it, and a signal or other external
2548 * event cannot wake it up and insert it on the runqueue either.
2550 p->state = TASK_RUNNING;
2554 * fork()/clone()-time setup:
2556 void sched_fork(struct task_struct *p, int clone_flags)
2558 int cpu = get_cpu();
2563 * Revert to default priority/policy on fork if requested.
2565 if (unlikely(p->sched_reset_on_fork)) {
2566 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
2567 p->policy = SCHED_NORMAL;
2568 p->normal_prio = p->static_prio;
2571 if (PRIO_TO_NICE(p->static_prio) < 0) {
2572 p->static_prio = NICE_TO_PRIO(0);
2573 p->normal_prio = p->static_prio;
2578 * We don't need the reset flag anymore after the fork. It has
2579 * fulfilled its duty:
2581 p->sched_reset_on_fork = 0;
2585 * Make sure we do not leak PI boosting priority to the child.
2587 p->prio = current->normal_prio;
2589 if (!rt_prio(p->prio))
2590 p->sched_class = &fair_sched_class;
2592 if (p->sched_class->task_fork)
2593 p->sched_class->task_fork(p);
2596 cpu = select_task_rq(p, SD_BALANCE_FORK, 0);
2598 set_task_cpu(p, cpu);
2600 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2601 if (likely(sched_info_on()))
2602 memset(&p->sched_info, 0, sizeof(p->sched_info));
2604 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2607 #ifdef CONFIG_PREEMPT
2608 /* Want to start with kernel preemption disabled. */
2609 task_thread_info(p)->preempt_count = 1;
2611 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2617 * wake_up_new_task - wake up a newly created task for the first time.
2619 * This function will do some initial scheduler statistics housekeeping
2620 * that must be done for every newly created context, then puts the task
2621 * on the runqueue and wakes it.
2623 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2625 unsigned long flags;
2628 rq = task_rq_lock(p, &flags);
2629 BUG_ON(p->state != TASK_RUNNING);
2630 update_rq_clock(rq);
2631 activate_task(rq, p, 0);
2632 trace_sched_wakeup_new(rq, p, 1);
2633 check_preempt_curr(rq, p, WF_FORK);
2635 if (p->sched_class->task_wake_up)
2636 p->sched_class->task_wake_up(rq, p);
2638 task_rq_unlock(rq, &flags);
2641 #ifdef CONFIG_PREEMPT_NOTIFIERS
2644 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2645 * @notifier: notifier struct to register
2647 void preempt_notifier_register(struct preempt_notifier *notifier)
2649 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2651 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2654 * preempt_notifier_unregister - no longer interested in preemption notifications
2655 * @notifier: notifier struct to unregister
2657 * This is safe to call from within a preemption notifier.
2659 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2661 hlist_del(¬ifier->link);
2663 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2665 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2667 struct preempt_notifier *notifier;
2668 struct hlist_node *node;
2670 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2671 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2675 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2676 struct task_struct *next)
2678 struct preempt_notifier *notifier;
2679 struct hlist_node *node;
2681 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2682 notifier->ops->sched_out(notifier, next);
2685 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2687 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2692 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2693 struct task_struct *next)
2697 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2700 * prepare_task_switch - prepare to switch tasks
2701 * @rq: the runqueue preparing to switch
2702 * @prev: the current task that is being switched out
2703 * @next: the task we are going to switch to.
2705 * This is called with the rq lock held and interrupts off. It must
2706 * be paired with a subsequent finish_task_switch after the context
2709 * prepare_task_switch sets up locking and calls architecture specific
2713 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2714 struct task_struct *next)
2716 fire_sched_out_preempt_notifiers(prev, next);
2717 prepare_lock_switch(rq, next);
2718 prepare_arch_switch(next);
2722 * finish_task_switch - clean up after a task-switch
2723 * @rq: runqueue associated with task-switch
2724 * @prev: the thread we just switched away from.
2726 * finish_task_switch must be called after the context switch, paired
2727 * with a prepare_task_switch call before the context switch.
2728 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2729 * and do any other architecture-specific cleanup actions.
2731 * Note that we may have delayed dropping an mm in context_switch(). If
2732 * so, we finish that here outside of the runqueue lock. (Doing it
2733 * with the lock held can cause deadlocks; see schedule() for
2736 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2737 __releases(rq->lock)
2739 struct mm_struct *mm = rq->prev_mm;
2745 * A task struct has one reference for the use as "current".
2746 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2747 * schedule one last time. The schedule call will never return, and
2748 * the scheduled task must drop that reference.
2749 * The test for TASK_DEAD must occur while the runqueue locks are
2750 * still held, otherwise prev could be scheduled on another cpu, die
2751 * there before we look at prev->state, and then the reference would
2753 * Manfred Spraul <manfred@colorfullife.com>
2755 prev_state = prev->state;
2756 finish_arch_switch(prev);
2757 perf_event_task_sched_in(current, cpu_of(rq));
2758 finish_lock_switch(rq, prev);
2760 fire_sched_in_preempt_notifiers(current);
2763 if (unlikely(prev_state == TASK_DEAD)) {
2765 * Remove function-return probe instances associated with this
2766 * task and put them back on the free list.
2768 kprobe_flush_task(prev);
2769 put_task_struct(prev);
2775 /* assumes rq->lock is held */
2776 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2778 if (prev->sched_class->pre_schedule)
2779 prev->sched_class->pre_schedule(rq, prev);
2782 /* rq->lock is NOT held, but preemption is disabled */
2783 static inline void post_schedule(struct rq *rq)
2785 if (rq->post_schedule) {
2786 unsigned long flags;
2788 raw_spin_lock_irqsave(&rq->lock, flags);
2789 if (rq->curr->sched_class->post_schedule)
2790 rq->curr->sched_class->post_schedule(rq);
2791 raw_spin_unlock_irqrestore(&rq->lock, flags);
2793 rq->post_schedule = 0;
2799 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2803 static inline void post_schedule(struct rq *rq)
2810 * schedule_tail - first thing a freshly forked thread must call.
2811 * @prev: the thread we just switched away from.
2813 asmlinkage void schedule_tail(struct task_struct *prev)
2814 __releases(rq->lock)
2816 struct rq *rq = this_rq();
2818 finish_task_switch(rq, prev);
2821 * FIXME: do we need to worry about rq being invalidated by the
2826 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2827 /* In this case, finish_task_switch does not reenable preemption */
2830 if (current->set_child_tid)
2831 put_user(task_pid_vnr(current), current->set_child_tid);
2835 * context_switch - switch to the new MM and the new
2836 * thread's register state.
2839 context_switch(struct rq *rq, struct task_struct *prev,
2840 struct task_struct *next)
2842 struct mm_struct *mm, *oldmm;
2844 prepare_task_switch(rq, prev, next);
2845 trace_sched_switch(rq, prev, next);
2847 oldmm = prev->active_mm;
2849 * For paravirt, this is coupled with an exit in switch_to to
2850 * combine the page table reload and the switch backend into
2853 arch_start_context_switch(prev);
2856 next->active_mm = oldmm;
2857 atomic_inc(&oldmm->mm_count);
2858 enter_lazy_tlb(oldmm, next);
2860 switch_mm(oldmm, mm, next);
2862 if (likely(!prev->mm)) {
2863 prev->active_mm = NULL;
2864 rq->prev_mm = oldmm;
2867 * Since the runqueue lock will be released by the next
2868 * task (which is an invalid locking op but in the case
2869 * of the scheduler it's an obvious special-case), so we
2870 * do an early lockdep release here:
2872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2873 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2876 /* Here we just switch the register state and the stack. */
2877 switch_to(prev, next, prev);
2881 * this_rq must be evaluated again because prev may have moved
2882 * CPUs since it called schedule(), thus the 'rq' on its stack
2883 * frame will be invalid.
2885 finish_task_switch(this_rq(), prev);
2889 * nr_running, nr_uninterruptible and nr_context_switches:
2891 * externally visible scheduler statistics: current number of runnable
2892 * threads, current number of uninterruptible-sleeping threads, total
2893 * number of context switches performed since bootup.
2895 unsigned long nr_running(void)
2897 unsigned long i, sum = 0;
2899 for_each_online_cpu(i)
2900 sum += cpu_rq(i)->nr_running;
2905 unsigned long nr_uninterruptible(void)
2907 unsigned long i, sum = 0;
2909 for_each_possible_cpu(i)
2910 sum += cpu_rq(i)->nr_uninterruptible;
2913 * Since we read the counters lockless, it might be slightly
2914 * inaccurate. Do not allow it to go below zero though:
2916 if (unlikely((long)sum < 0))
2922 unsigned long long nr_context_switches(void)
2925 unsigned long long sum = 0;
2927 for_each_possible_cpu(i)
2928 sum += cpu_rq(i)->nr_switches;
2933 unsigned long nr_iowait(void)
2935 unsigned long i, sum = 0;
2937 for_each_possible_cpu(i)
2938 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2943 unsigned long nr_iowait_cpu(void)
2945 struct rq *this = this_rq();
2946 return atomic_read(&this->nr_iowait);
2949 unsigned long this_cpu_load(void)
2951 struct rq *this = this_rq();
2952 return this->cpu_load[0];
2956 /* Variables and functions for calc_load */
2957 static atomic_long_t calc_load_tasks;
2958 static unsigned long calc_load_update;
2959 unsigned long avenrun[3];
2960 EXPORT_SYMBOL(avenrun);
2963 * get_avenrun - get the load average array
2964 * @loads: pointer to dest load array
2965 * @offset: offset to add
2966 * @shift: shift count to shift the result left
2968 * These values are estimates at best, so no need for locking.
2970 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2972 loads[0] = (avenrun[0] + offset) << shift;
2973 loads[1] = (avenrun[1] + offset) << shift;
2974 loads[2] = (avenrun[2] + offset) << shift;
2977 static unsigned long
2978 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2981 load += active * (FIXED_1 - exp);
2982 return load >> FSHIFT;
2986 * calc_load - update the avenrun load estimates 10 ticks after the
2987 * CPUs have updated calc_load_tasks.
2989 void calc_global_load(void)
2991 unsigned long upd = calc_load_update + 10;
2994 if (time_before(jiffies, upd))
2997 active = atomic_long_read(&calc_load_tasks);
2998 active = active > 0 ? active * FIXED_1 : 0;
3000 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3001 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3002 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3004 calc_load_update += LOAD_FREQ;
3008 * Either called from update_cpu_load() or from a cpu going idle
3010 static void calc_load_account_active(struct rq *this_rq)
3012 long nr_active, delta;
3014 nr_active = this_rq->nr_running;
3015 nr_active += (long) this_rq->nr_uninterruptible;
3017 if (nr_active != this_rq->calc_load_active) {
3018 delta = nr_active - this_rq->calc_load_active;
3019 this_rq->calc_load_active = nr_active;
3020 atomic_long_add(delta, &calc_load_tasks);
3025 * Update rq->cpu_load[] statistics. This function is usually called every
3026 * scheduler tick (TICK_NSEC).
3028 static void update_cpu_load(struct rq *this_rq)
3030 unsigned long this_load = this_rq->load.weight;
3033 this_rq->nr_load_updates++;
3035 /* Update our load: */
3036 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3037 unsigned long old_load, new_load;
3039 /* scale is effectively 1 << i now, and >> i divides by scale */
3041 old_load = this_rq->cpu_load[i];
3042 new_load = this_load;
3044 * Round up the averaging division if load is increasing. This
3045 * prevents us from getting stuck on 9 if the load is 10, for
3048 if (new_load > old_load)
3049 new_load += scale-1;
3050 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3053 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3054 this_rq->calc_load_update += LOAD_FREQ;
3055 calc_load_account_active(this_rq);
3062 * double_rq_lock - safely lock two runqueues
3064 * Note this does not disable interrupts like task_rq_lock,
3065 * you need to do so manually before calling.
3067 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3068 __acquires(rq1->lock)
3069 __acquires(rq2->lock)
3071 BUG_ON(!irqs_disabled());
3073 raw_spin_lock(&rq1->lock);
3074 __acquire(rq2->lock); /* Fake it out ;) */
3077 raw_spin_lock(&rq1->lock);
3078 raw_spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3080 raw_spin_lock(&rq2->lock);
3081 raw_spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3084 update_rq_clock(rq1);
3085 update_rq_clock(rq2);
3089 * double_rq_unlock - safely unlock two runqueues
3091 * Note this does not restore interrupts like task_rq_unlock,
3092 * you need to do so manually after calling.
3094 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3095 __releases(rq1->lock)
3096 __releases(rq2->lock)
3098 raw_spin_unlock(&rq1->lock);
3100 raw_spin_unlock(&rq2->lock);
3102 __release(rq2->lock);
3106 * If dest_cpu is allowed for this process, migrate the task to it.
3107 * This is accomplished by forcing the cpu_allowed mask to only
3108 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3109 * the cpu_allowed mask is restored.
3111 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3113 struct migration_req req;
3114 unsigned long flags;
3117 rq = task_rq_lock(p, &flags);
3118 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3119 || unlikely(!cpu_active(dest_cpu)))
3122 /* force the process onto the specified CPU */
3123 if (migrate_task(p, dest_cpu, &req)) {
3124 /* Need to wait for migration thread (might exit: take ref). */
3125 struct task_struct *mt = rq->migration_thread;
3127 get_task_struct(mt);
3128 task_rq_unlock(rq, &flags);
3129 wake_up_process(mt);
3130 put_task_struct(mt);
3131 wait_for_completion(&req.done);
3136 task_rq_unlock(rq, &flags);
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 int new_cpu, this_cpu = get_cpu();
3146 new_cpu = select_task_rq(current, SD_BALANCE_EXEC, 0);
3148 if (new_cpu != this_cpu)
3149 sched_migrate_task(current, new_cpu);
3153 * pull_task - move a task from a remote runqueue to the local runqueue.
3154 * Both runqueues must be locked.
3156 static void pull_task(struct rq *src_rq, struct task_struct *p,
3157 struct rq *this_rq, int this_cpu)
3159 deactivate_task(src_rq, p, 0);
3160 set_task_cpu(p, this_cpu);
3161 activate_task(this_rq, p, 0);
3162 check_preempt_curr(this_rq, p, 0);
3166 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3169 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3170 struct sched_domain *sd, enum cpu_idle_type idle,
3173 int tsk_cache_hot = 0;
3175 * We do not migrate tasks that are:
3176 * 1) running (obviously), or
3177 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3178 * 3) are cache-hot on their current CPU.
3180 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3181 schedstat_inc(p, se.nr_failed_migrations_affine);
3186 if (task_running(rq, p)) {
3187 schedstat_inc(p, se.nr_failed_migrations_running);
3192 * Aggressive migration if:
3193 * 1) task is cache cold, or
3194 * 2) too many balance attempts have failed.
3197 tsk_cache_hot = task_hot(p, rq->clock, sd);
3198 if (!tsk_cache_hot ||
3199 sd->nr_balance_failed > sd->cache_nice_tries) {
3200 #ifdef CONFIG_SCHEDSTATS
3201 if (tsk_cache_hot) {
3202 schedstat_inc(sd, lb_hot_gained[idle]);
3203 schedstat_inc(p, se.nr_forced_migrations);
3209 if (tsk_cache_hot) {
3210 schedstat_inc(p, se.nr_failed_migrations_hot);
3216 static unsigned long
3217 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3218 unsigned long max_load_move, struct sched_domain *sd,
3219 enum cpu_idle_type idle, int *all_pinned,
3220 int *this_best_prio, struct rq_iterator *iterator)
3222 int loops = 0, pulled = 0, pinned = 0;
3223 struct task_struct *p;
3224 long rem_load_move = max_load_move;
3226 if (max_load_move == 0)
3232 * Start the load-balancing iterator:
3234 p = iterator->start(iterator->arg);
3236 if (!p || loops++ > sysctl_sched_nr_migrate)
3239 if ((p->se.load.weight >> 1) > rem_load_move ||
3240 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3241 p = iterator->next(iterator->arg);
3245 pull_task(busiest, p, this_rq, this_cpu);
3247 rem_load_move -= p->se.load.weight;
3249 #ifdef CONFIG_PREEMPT
3251 * NEWIDLE balancing is a source of latency, so preemptible kernels
3252 * will stop after the first task is pulled to minimize the critical
3255 if (idle == CPU_NEWLY_IDLE)
3260 * We only want to steal up to the prescribed amount of weighted load.
3262 if (rem_load_move > 0) {
3263 if (p->prio < *this_best_prio)
3264 *this_best_prio = p->prio;
3265 p = iterator->next(iterator->arg);
3270 * Right now, this is one of only two places pull_task() is called,
3271 * so we can safely collect pull_task() stats here rather than
3272 * inside pull_task().
3274 schedstat_add(sd, lb_gained[idle], pulled);
3277 *all_pinned = pinned;
3279 return max_load_move - rem_load_move;
3283 * move_tasks tries to move up to max_load_move weighted load from busiest to
3284 * this_rq, as part of a balancing operation within domain "sd".
3285 * Returns 1 if successful and 0 otherwise.
3287 * Called with both runqueues locked.
3289 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3290 unsigned long max_load_move,
3291 struct sched_domain *sd, enum cpu_idle_type idle,
3294 const struct sched_class *class = sched_class_highest;
3295 unsigned long total_load_moved = 0;
3296 int this_best_prio = this_rq->curr->prio;
3300 class->load_balance(this_rq, this_cpu, busiest,
3301 max_load_move - total_load_moved,
3302 sd, idle, all_pinned, &this_best_prio);
3303 class = class->next;
3305 #ifdef CONFIG_PREEMPT
3307 * NEWIDLE balancing is a source of latency, so preemptible
3308 * kernels will stop after the first task is pulled to minimize
3309 * the critical section.
3311 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3314 } while (class && max_load_move > total_load_moved);
3316 return total_load_moved > 0;
3320 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3321 struct sched_domain *sd, enum cpu_idle_type idle,
3322 struct rq_iterator *iterator)
3324 struct task_struct *p = iterator->start(iterator->arg);
3328 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3329 pull_task(busiest, p, this_rq, this_cpu);
3331 * Right now, this is only the second place pull_task()
3332 * is called, so we can safely collect pull_task()
3333 * stats here rather than inside pull_task().
3335 schedstat_inc(sd, lb_gained[idle]);
3339 p = iterator->next(iterator->arg);
3346 * move_one_task tries to move exactly one task from busiest to this_rq, as
3347 * part of active balancing operations within "domain".
3348 * Returns 1 if successful and 0 otherwise.
3350 * Called with both runqueues locked.
3352 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3353 struct sched_domain *sd, enum cpu_idle_type idle)
3355 const struct sched_class *class;
3357 for_each_class(class) {
3358 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3364 /********** Helpers for find_busiest_group ************************/
3366 * sd_lb_stats - Structure to store the statistics of a sched_domain
3367 * during load balancing.
3369 struct sd_lb_stats {
3370 struct sched_group *busiest; /* Busiest group in this sd */
3371 struct sched_group *this; /* Local group in this sd */
3372 unsigned long total_load; /* Total load of all groups in sd */
3373 unsigned long total_pwr; /* Total power of all groups in sd */
3374 unsigned long avg_load; /* Average load across all groups in sd */
3376 /** Statistics of this group */
3377 unsigned long this_load;
3378 unsigned long this_load_per_task;
3379 unsigned long this_nr_running;
3381 /* Statistics of the busiest group */
3382 unsigned long max_load;
3383 unsigned long busiest_load_per_task;
3384 unsigned long busiest_nr_running;
3386 int group_imb; /* Is there imbalance in this sd */
3387 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3388 int power_savings_balance; /* Is powersave balance needed for this sd */
3389 struct sched_group *group_min; /* Least loaded group in sd */
3390 struct sched_group *group_leader; /* Group which relieves group_min */
3391 unsigned long min_load_per_task; /* load_per_task in group_min */
3392 unsigned long leader_nr_running; /* Nr running of group_leader */
3393 unsigned long min_nr_running; /* Nr running of group_min */
3398 * sg_lb_stats - stats of a sched_group required for load_balancing
3400 struct sg_lb_stats {
3401 unsigned long avg_load; /*Avg load across the CPUs of the group */
3402 unsigned long group_load; /* Total load over the CPUs of the group */
3403 unsigned long sum_nr_running; /* Nr tasks running in the group */
3404 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3405 unsigned long group_capacity;
3406 int group_imb; /* Is there an imbalance in the group ? */
3410 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3411 * @group: The group whose first cpu is to be returned.
3413 static inline unsigned int group_first_cpu(struct sched_group *group)
3415 return cpumask_first(sched_group_cpus(group));
3419 * get_sd_load_idx - Obtain the load index for a given sched domain.
3420 * @sd: The sched_domain whose load_idx is to be obtained.
3421 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3423 static inline int get_sd_load_idx(struct sched_domain *sd,
3424 enum cpu_idle_type idle)
3430 load_idx = sd->busy_idx;
3433 case CPU_NEWLY_IDLE:
3434 load_idx = sd->newidle_idx;
3437 load_idx = sd->idle_idx;
3445 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3447 * init_sd_power_savings_stats - Initialize power savings statistics for
3448 * the given sched_domain, during load balancing.
3450 * @sd: Sched domain whose power-savings statistics are to be initialized.
3451 * @sds: Variable containing the statistics for sd.
3452 * @idle: Idle status of the CPU at which we're performing load-balancing.
3454 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3455 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3458 * Busy processors will not participate in power savings
3461 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3462 sds->power_savings_balance = 0;
3464 sds->power_savings_balance = 1;
3465 sds->min_nr_running = ULONG_MAX;
3466 sds->leader_nr_running = 0;
3471 * update_sd_power_savings_stats - Update the power saving stats for a
3472 * sched_domain while performing load balancing.
3474 * @group: sched_group belonging to the sched_domain under consideration.
3475 * @sds: Variable containing the statistics of the sched_domain
3476 * @local_group: Does group contain the CPU for which we're performing
3478 * @sgs: Variable containing the statistics of the group.
3480 static inline void update_sd_power_savings_stats(struct sched_group *group,
3481 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3484 if (!sds->power_savings_balance)
3488 * If the local group is idle or completely loaded
3489 * no need to do power savings balance at this domain
3491 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3492 !sds->this_nr_running))
3493 sds->power_savings_balance = 0;
3496 * If a group is already running at full capacity or idle,
3497 * don't include that group in power savings calculations
3499 if (!sds->power_savings_balance ||
3500 sgs->sum_nr_running >= sgs->group_capacity ||
3501 !sgs->sum_nr_running)
3505 * Calculate the group which has the least non-idle load.
3506 * This is the group from where we need to pick up the load
3509 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3510 (sgs->sum_nr_running == sds->min_nr_running &&
3511 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3512 sds->group_min = group;
3513 sds->min_nr_running = sgs->sum_nr_running;
3514 sds->min_load_per_task = sgs->sum_weighted_load /
3515 sgs->sum_nr_running;
3519 * Calculate the group which is almost near its
3520 * capacity but still has some space to pick up some load
3521 * from other group and save more power
3523 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3526 if (sgs->sum_nr_running > sds->leader_nr_running ||
3527 (sgs->sum_nr_running == sds->leader_nr_running &&
3528 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3529 sds->group_leader = group;
3530 sds->leader_nr_running = sgs->sum_nr_running;
3535 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3536 * @sds: Variable containing the statistics of the sched_domain
3537 * under consideration.
3538 * @this_cpu: Cpu at which we're currently performing load-balancing.
3539 * @imbalance: Variable to store the imbalance.
3542 * Check if we have potential to perform some power-savings balance.
3543 * If yes, set the busiest group to be the least loaded group in the
3544 * sched_domain, so that it's CPUs can be put to idle.
3546 * Returns 1 if there is potential to perform power-savings balance.
3549 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3550 int this_cpu, unsigned long *imbalance)
3552 if (!sds->power_savings_balance)
3555 if (sds->this != sds->group_leader ||
3556 sds->group_leader == sds->group_min)
3559 *imbalance = sds->min_load_per_task;
3560 sds->busiest = sds->group_min;
3565 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3566 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3567 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3572 static inline void update_sd_power_savings_stats(struct sched_group *group,
3573 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3578 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3579 int this_cpu, unsigned long *imbalance)
3583 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3586 unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3588 return SCHED_LOAD_SCALE;
3591 unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3593 return default_scale_freq_power(sd, cpu);
3596 unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
3598 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3599 unsigned long smt_gain = sd->smt_gain;
3606 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3608 return default_scale_smt_power(sd, cpu);
3611 unsigned long scale_rt_power(int cpu)
3613 struct rq *rq = cpu_rq(cpu);
3614 u64 total, available;
3616 sched_avg_update(rq);
3618 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3619 available = total - rq->rt_avg;
3621 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3622 total = SCHED_LOAD_SCALE;
3624 total >>= SCHED_LOAD_SHIFT;
3626 return div_u64(available, total);
3629 static void update_cpu_power(struct sched_domain *sd, int cpu)
3631 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3632 unsigned long power = SCHED_LOAD_SCALE;
3633 struct sched_group *sdg = sd->groups;
3635 if (sched_feat(ARCH_POWER))
3636 power *= arch_scale_freq_power(sd, cpu);
3638 power *= default_scale_freq_power(sd, cpu);
3640 power >>= SCHED_LOAD_SHIFT;
3642 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3643 if (sched_feat(ARCH_POWER))
3644 power *= arch_scale_smt_power(sd, cpu);
3646 power *= default_scale_smt_power(sd, cpu);
3648 power >>= SCHED_LOAD_SHIFT;
3651 power *= scale_rt_power(cpu);
3652 power >>= SCHED_LOAD_SHIFT;
3657 sdg->cpu_power = power;
3660 static void update_group_power(struct sched_domain *sd, int cpu)
3662 struct sched_domain *child = sd->child;
3663 struct sched_group *group, *sdg = sd->groups;
3664 unsigned long power;
3667 update_cpu_power(sd, cpu);
3673 group = child->groups;
3675 power += group->cpu_power;
3676 group = group->next;
3677 } while (group != child->groups);
3679 sdg->cpu_power = power;
3683 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3684 * @sd: The sched_domain whose statistics are to be updated.
3685 * @group: sched_group whose statistics are to be updated.
3686 * @this_cpu: Cpu for which load balance is currently performed.
3687 * @idle: Idle status of this_cpu
3688 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3689 * @sd_idle: Idle status of the sched_domain containing group.
3690 * @local_group: Does group contain this_cpu.
3691 * @cpus: Set of cpus considered for load balancing.
3692 * @balance: Should we balance.
3693 * @sgs: variable to hold the statistics for this group.
3695 static inline void update_sg_lb_stats(struct sched_domain *sd,
3696 struct sched_group *group, int this_cpu,
3697 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3698 int local_group, const struct cpumask *cpus,
3699 int *balance, struct sg_lb_stats *sgs)
3701 unsigned long load, max_cpu_load, min_cpu_load;
3703 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3704 unsigned long sum_avg_load_per_task;
3705 unsigned long avg_load_per_task;
3708 balance_cpu = group_first_cpu(group);
3709 if (balance_cpu == this_cpu)
3710 update_group_power(sd, this_cpu);
3713 /* Tally up the load of all CPUs in the group */
3714 sum_avg_load_per_task = avg_load_per_task = 0;
3716 min_cpu_load = ~0UL;
3718 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3719 struct rq *rq = cpu_rq(i);
3721 if (*sd_idle && rq->nr_running)
3724 /* Bias balancing toward cpus of our domain */
3726 if (idle_cpu(i) && !first_idle_cpu) {
3731 load = target_load(i, load_idx);
3733 load = source_load(i, load_idx);
3734 if (load > max_cpu_load)
3735 max_cpu_load = load;
3736 if (min_cpu_load > load)
3737 min_cpu_load = load;
3740 sgs->group_load += load;
3741 sgs->sum_nr_running += rq->nr_running;
3742 sgs->sum_weighted_load += weighted_cpuload(i);
3744 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3748 * First idle cpu or the first cpu(busiest) in this sched group
3749 * is eligible for doing load balancing at this and above
3750 * domains. In the newly idle case, we will allow all the cpu's
3751 * to do the newly idle load balance.
3753 if (idle != CPU_NEWLY_IDLE && local_group &&
3754 balance_cpu != this_cpu && balance) {
3759 /* Adjust by relative CPU power of the group */
3760 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3764 * Consider the group unbalanced when the imbalance is larger
3765 * than the average weight of two tasks.
3767 * APZ: with cgroup the avg task weight can vary wildly and
3768 * might not be a suitable number - should we keep a
3769 * normalized nr_running number somewhere that negates
3772 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3775 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3778 sgs->group_capacity =
3779 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3783 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3784 * @sd: sched_domain whose statistics are to be updated.
3785 * @this_cpu: Cpu for which load balance is currently performed.
3786 * @idle: Idle status of this_cpu
3787 * @sd_idle: Idle status of the sched_domain containing group.
3788 * @cpus: Set of cpus considered for load balancing.
3789 * @balance: Should we balance.
3790 * @sds: variable to hold the statistics for this sched_domain.
3792 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3793 enum cpu_idle_type idle, int *sd_idle,
3794 const struct cpumask *cpus, int *balance,
3795 struct sd_lb_stats *sds)
3797 struct sched_domain *child = sd->child;
3798 struct sched_group *group = sd->groups;
3799 struct sg_lb_stats sgs;
3800 int load_idx, prefer_sibling = 0;
3802 if (child && child->flags & SD_PREFER_SIBLING)
3805 init_sd_power_savings_stats(sd, sds, idle);
3806 load_idx = get_sd_load_idx(sd, idle);
3811 local_group = cpumask_test_cpu(this_cpu,
3812 sched_group_cpus(group));
3813 memset(&sgs, 0, sizeof(sgs));
3814 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3815 local_group, cpus, balance, &sgs);
3817 if (local_group && balance && !(*balance))
3820 sds->total_load += sgs.group_load;
3821 sds->total_pwr += group->cpu_power;
3824 * In case the child domain prefers tasks go to siblings
3825 * first, lower the group capacity to one so that we'll try
3826 * and move all the excess tasks away.
3829 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3832 sds->this_load = sgs.avg_load;
3834 sds->this_nr_running = sgs.sum_nr_running;
3835 sds->this_load_per_task = sgs.sum_weighted_load;
3836 } else if (sgs.avg_load > sds->max_load &&
3837 (sgs.sum_nr_running > sgs.group_capacity ||
3839 sds->max_load = sgs.avg_load;
3840 sds->busiest = group;
3841 sds->busiest_nr_running = sgs.sum_nr_running;
3842 sds->busiest_load_per_task = sgs.sum_weighted_load;
3843 sds->group_imb = sgs.group_imb;
3846 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3847 group = group->next;
3848 } while (group != sd->groups);
3852 * fix_small_imbalance - Calculate the minor imbalance that exists
3853 * amongst the groups of a sched_domain, during
3855 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3856 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3857 * @imbalance: Variable to store the imbalance.
3859 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3860 int this_cpu, unsigned long *imbalance)
3862 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3863 unsigned int imbn = 2;
3865 if (sds->this_nr_running) {
3866 sds->this_load_per_task /= sds->this_nr_running;
3867 if (sds->busiest_load_per_task >
3868 sds->this_load_per_task)
3871 sds->this_load_per_task =
3872 cpu_avg_load_per_task(this_cpu);
3874 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3875 sds->busiest_load_per_task * imbn) {
3876 *imbalance = sds->busiest_load_per_task;
3881 * OK, we don't have enough imbalance to justify moving tasks,
3882 * however we may be able to increase total CPU power used by
3886 pwr_now += sds->busiest->cpu_power *
3887 min(sds->busiest_load_per_task, sds->max_load);
3888 pwr_now += sds->this->cpu_power *
3889 min(sds->this_load_per_task, sds->this_load);
3890 pwr_now /= SCHED_LOAD_SCALE;
3892 /* Amount of load we'd subtract */
3893 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3894 sds->busiest->cpu_power;
3895 if (sds->max_load > tmp)
3896 pwr_move += sds->busiest->cpu_power *
3897 min(sds->busiest_load_per_task, sds->max_load - tmp);
3899 /* Amount of load we'd add */
3900 if (sds->max_load * sds->busiest->cpu_power <
3901 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3902 tmp = (sds->max_load * sds->busiest->cpu_power) /
3903 sds->this->cpu_power;
3905 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3906 sds->this->cpu_power;
3907 pwr_move += sds->this->cpu_power *
3908 min(sds->this_load_per_task, sds->this_load + tmp);
3909 pwr_move /= SCHED_LOAD_SCALE;
3911 /* Move if we gain throughput */
3912 if (pwr_move > pwr_now)
3913 *imbalance = sds->busiest_load_per_task;
3917 * calculate_imbalance - Calculate the amount of imbalance present within the
3918 * groups of a given sched_domain during load balance.
3919 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3920 * @this_cpu: Cpu for which currently load balance is being performed.
3921 * @imbalance: The variable to store the imbalance.
3923 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3924 unsigned long *imbalance)
3926 unsigned long max_pull;
3928 * In the presence of smp nice balancing, certain scenarios can have
3929 * max load less than avg load(as we skip the groups at or below
3930 * its cpu_power, while calculating max_load..)
3932 if (sds->max_load < sds->avg_load) {
3934 return fix_small_imbalance(sds, this_cpu, imbalance);
3937 /* Don't want to pull so many tasks that a group would go idle */
3938 max_pull = min(sds->max_load - sds->avg_load,
3939 sds->max_load - sds->busiest_load_per_task);
3941 /* How much load to actually move to equalise the imbalance */
3942 *imbalance = min(max_pull * sds->busiest->cpu_power,
3943 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3947 * if *imbalance is less than the average load per runnable task
3948 * there is no gaurantee that any tasks will be moved so we'll have
3949 * a think about bumping its value to force at least one task to be
3952 if (*imbalance < sds->busiest_load_per_task)
3953 return fix_small_imbalance(sds, this_cpu, imbalance);
3956 /******* find_busiest_group() helpers end here *********************/
3959 * find_busiest_group - Returns the busiest group within the sched_domain
3960 * if there is an imbalance. If there isn't an imbalance, and
3961 * the user has opted for power-savings, it returns a group whose
3962 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3963 * such a group exists.
3965 * Also calculates the amount of weighted load which should be moved
3966 * to restore balance.
3968 * @sd: The sched_domain whose busiest group is to be returned.
3969 * @this_cpu: The cpu for which load balancing is currently being performed.
3970 * @imbalance: Variable which stores amount of weighted load which should
3971 * be moved to restore balance/put a group to idle.
3972 * @idle: The idle status of this_cpu.
3973 * @sd_idle: The idleness of sd
3974 * @cpus: The set of CPUs under consideration for load-balancing.
3975 * @balance: Pointer to a variable indicating if this_cpu
3976 * is the appropriate cpu to perform load balancing at this_level.
3978 * Returns: - the busiest group if imbalance exists.
3979 * - If no imbalance and user has opted for power-savings balance,
3980 * return the least loaded group whose CPUs can be
3981 * put to idle by rebalancing its tasks onto our group.
3983 static struct sched_group *
3984 find_busiest_group(struct sched_domain *sd, int this_cpu,
3985 unsigned long *imbalance, enum cpu_idle_type idle,
3986 int *sd_idle, const struct cpumask *cpus, int *balance)
3988 struct sd_lb_stats sds;
3990 memset(&sds, 0, sizeof(sds));
3993 * Compute the various statistics relavent for load balancing at
3996 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3999 /* Cases where imbalance does not exist from POV of this_cpu */
4000 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4002 * 2) There is no busy sibling group to pull from.
4003 * 3) This group is the busiest group.
4004 * 4) This group is more busy than the avg busieness at this
4006 * 5) The imbalance is within the specified limit.
4007 * 6) Any rebalance would lead to ping-pong
4009 if (balance && !(*balance))
4012 if (!sds.busiest || sds.busiest_nr_running == 0)
4015 if (sds.this_load >= sds.max_load)
4018 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4020 if (sds.this_load >= sds.avg_load)
4023 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4026 sds.busiest_load_per_task /= sds.busiest_nr_running;
4028 sds.busiest_load_per_task =
4029 min(sds.busiest_load_per_task, sds.avg_load);
4032 * We're trying to get all the cpus to the average_load, so we don't
4033 * want to push ourselves above the average load, nor do we wish to
4034 * reduce the max loaded cpu below the average load, as either of these
4035 * actions would just result in more rebalancing later, and ping-pong
4036 * tasks around. Thus we look for the minimum possible imbalance.
4037 * Negative imbalances (*we* are more loaded than anyone else) will
4038 * be counted as no imbalance for these purposes -- we can't fix that
4039 * by pulling tasks to us. Be careful of negative numbers as they'll
4040 * appear as very large values with unsigned longs.
4042 if (sds.max_load <= sds.busiest_load_per_task)
4045 /* Looks like there is an imbalance. Compute it */
4046 calculate_imbalance(&sds, this_cpu, imbalance);
4051 * There is no obvious imbalance. But check if we can do some balancing
4054 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4062 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4065 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4066 unsigned long imbalance, const struct cpumask *cpus)
4068 struct rq *busiest = NULL, *rq;
4069 unsigned long max_load = 0;
4072 for_each_cpu(i, sched_group_cpus(group)) {
4073 unsigned long power = power_of(i);
4074 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4077 if (!cpumask_test_cpu(i, cpus))
4081 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4084 if (capacity && rq->nr_running == 1 && wl > imbalance)
4087 if (wl > max_load) {
4097 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4098 * so long as it is large enough.
4100 #define MAX_PINNED_INTERVAL 512
4102 /* Working cpumask for load_balance and load_balance_newidle. */
4103 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4106 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4107 * tasks if there is an imbalance.
4109 static int load_balance(int this_cpu, struct rq *this_rq,
4110 struct sched_domain *sd, enum cpu_idle_type idle,
4113 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4114 struct sched_group *group;
4115 unsigned long imbalance;
4117 unsigned long flags;
4118 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4120 cpumask_copy(cpus, cpu_active_mask);
4123 * When power savings policy is enabled for the parent domain, idle
4124 * sibling can pick up load irrespective of busy siblings. In this case,
4125 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4126 * portraying it as CPU_NOT_IDLE.
4128 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4129 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4132 schedstat_inc(sd, lb_count[idle]);
4136 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4143 schedstat_inc(sd, lb_nobusyg[idle]);
4147 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4149 schedstat_inc(sd, lb_nobusyq[idle]);
4153 BUG_ON(busiest == this_rq);
4155 schedstat_add(sd, lb_imbalance[idle], imbalance);
4158 if (busiest->nr_running > 1) {
4160 * Attempt to move tasks. If find_busiest_group has found
4161 * an imbalance but busiest->nr_running <= 1, the group is
4162 * still unbalanced. ld_moved simply stays zero, so it is
4163 * correctly treated as an imbalance.
4165 local_irq_save(flags);
4166 double_rq_lock(this_rq, busiest);
4167 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4168 imbalance, sd, idle, &all_pinned);
4169 double_rq_unlock(this_rq, busiest);
4170 local_irq_restore(flags);
4173 * some other cpu did the load balance for us.
4175 if (ld_moved && this_cpu != smp_processor_id())
4176 resched_cpu(this_cpu);
4178 /* All tasks on this runqueue were pinned by CPU affinity */
4179 if (unlikely(all_pinned)) {
4180 cpumask_clear_cpu(cpu_of(busiest), cpus);
4181 if (!cpumask_empty(cpus))
4188 schedstat_inc(sd, lb_failed[idle]);
4189 sd->nr_balance_failed++;
4191 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4193 raw_spin_lock_irqsave(&busiest->lock, flags);
4195 /* don't kick the migration_thread, if the curr
4196 * task on busiest cpu can't be moved to this_cpu
4198 if (!cpumask_test_cpu(this_cpu,
4199 &busiest->curr->cpus_allowed)) {
4200 raw_spin_unlock_irqrestore(&busiest->lock,
4203 goto out_one_pinned;
4206 if (!busiest->active_balance) {
4207 busiest->active_balance = 1;
4208 busiest->push_cpu = this_cpu;
4211 raw_spin_unlock_irqrestore(&busiest->lock, flags);
4213 wake_up_process(busiest->migration_thread);
4216 * We've kicked active balancing, reset the failure
4219 sd->nr_balance_failed = sd->cache_nice_tries+1;
4222 sd->nr_balance_failed = 0;
4224 if (likely(!active_balance)) {
4225 /* We were unbalanced, so reset the balancing interval */
4226 sd->balance_interval = sd->min_interval;
4229 * If we've begun active balancing, start to back off. This
4230 * case may not be covered by the all_pinned logic if there
4231 * is only 1 task on the busy runqueue (because we don't call
4234 if (sd->balance_interval < sd->max_interval)
4235 sd->balance_interval *= 2;
4238 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4239 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4245 schedstat_inc(sd, lb_balanced[idle]);
4247 sd->nr_balance_failed = 0;
4250 /* tune up the balancing interval */
4251 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4252 (sd->balance_interval < sd->max_interval))
4253 sd->balance_interval *= 2;
4255 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4256 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4267 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4268 * tasks if there is an imbalance.
4270 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4271 * this_rq is locked.
4274 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4276 struct sched_group *group;
4277 struct rq *busiest = NULL;
4278 unsigned long imbalance;
4282 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4284 cpumask_copy(cpus, cpu_active_mask);
4287 * When power savings policy is enabled for the parent domain, idle
4288 * sibling can pick up load irrespective of busy siblings. In this case,
4289 * let the state of idle sibling percolate up as IDLE, instead of
4290 * portraying it as CPU_NOT_IDLE.
4292 if (sd->flags & SD_SHARE_CPUPOWER &&
4293 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4296 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4298 update_shares_locked(this_rq, sd);
4299 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4300 &sd_idle, cpus, NULL);
4302 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4306 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4308 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4312 BUG_ON(busiest == this_rq);
4314 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4317 if (busiest->nr_running > 1) {
4318 /* Attempt to move tasks */
4319 double_lock_balance(this_rq, busiest);
4320 /* this_rq->clock is already updated */
4321 update_rq_clock(busiest);
4322 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4323 imbalance, sd, CPU_NEWLY_IDLE,
4325 double_unlock_balance(this_rq, busiest);
4327 if (unlikely(all_pinned)) {
4328 cpumask_clear_cpu(cpu_of(busiest), cpus);
4329 if (!cpumask_empty(cpus))
4335 int active_balance = 0;
4337 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4338 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4339 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4342 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4345 if (sd->nr_balance_failed++ < 2)
4349 * The only task running in a non-idle cpu can be moved to this
4350 * cpu in an attempt to completely freeup the other CPU
4351 * package. The same method used to move task in load_balance()
4352 * have been extended for load_balance_newidle() to speedup
4353 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4355 * The package power saving logic comes from
4356 * find_busiest_group(). If there are no imbalance, then
4357 * f_b_g() will return NULL. However when sched_mc={1,2} then
4358 * f_b_g() will select a group from which a running task may be
4359 * pulled to this cpu in order to make the other package idle.
4360 * If there is no opportunity to make a package idle and if
4361 * there are no imbalance, then f_b_g() will return NULL and no
4362 * action will be taken in load_balance_newidle().
4364 * Under normal task pull operation due to imbalance, there
4365 * will be more than one task in the source run queue and
4366 * move_tasks() will succeed. ld_moved will be true and this
4367 * active balance code will not be triggered.
4370 /* Lock busiest in correct order while this_rq is held */
4371 double_lock_balance(this_rq, busiest);
4374 * don't kick the migration_thread, if the curr
4375 * task on busiest cpu can't be moved to this_cpu
4377 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4378 double_unlock_balance(this_rq, busiest);
4383 if (!busiest->active_balance) {
4384 busiest->active_balance = 1;
4385 busiest->push_cpu = this_cpu;
4389 double_unlock_balance(this_rq, busiest);
4391 * Should not call ttwu while holding a rq->lock
4393 raw_spin_unlock(&this_rq->lock);
4395 wake_up_process(busiest->migration_thread);
4396 raw_spin_lock(&this_rq->lock);
4399 sd->nr_balance_failed = 0;
4401 update_shares_locked(this_rq, sd);
4405 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4406 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4407 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4409 sd->nr_balance_failed = 0;
4415 * idle_balance is called by schedule() if this_cpu is about to become
4416 * idle. Attempts to pull tasks from other CPUs.
4418 static void idle_balance(int this_cpu, struct rq *this_rq)
4420 struct sched_domain *sd;
4421 int pulled_task = 0;
4422 unsigned long next_balance = jiffies + HZ;
4424 this_rq->idle_stamp = this_rq->clock;
4426 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4429 for_each_domain(this_cpu, sd) {
4430 unsigned long interval;
4432 if (!(sd->flags & SD_LOAD_BALANCE))
4435 if (sd->flags & SD_BALANCE_NEWIDLE)
4436 /* If we've pulled tasks over stop searching: */
4437 pulled_task = load_balance_newidle(this_cpu, this_rq,
4440 interval = msecs_to_jiffies(sd->balance_interval);
4441 if (time_after(next_balance, sd->last_balance + interval))
4442 next_balance = sd->last_balance + interval;
4444 this_rq->idle_stamp = 0;
4448 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4450 * We are going idle. next_balance may be set based on
4451 * a busy processor. So reset next_balance.
4453 this_rq->next_balance = next_balance;
4458 * active_load_balance is run by migration threads. It pushes running tasks
4459 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4460 * running on each physical CPU where possible, and avoids physical /
4461 * logical imbalances.
4463 * Called with busiest_rq locked.
4465 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4467 int target_cpu = busiest_rq->push_cpu;
4468 struct sched_domain *sd;
4469 struct rq *target_rq;
4471 /* Is there any task to move? */
4472 if (busiest_rq->nr_running <= 1)
4475 target_rq = cpu_rq(target_cpu);
4478 * This condition is "impossible", if it occurs
4479 * we need to fix it. Originally reported by
4480 * Bjorn Helgaas on a 128-cpu setup.
4482 BUG_ON(busiest_rq == target_rq);
4484 /* move a task from busiest_rq to target_rq */
4485 double_lock_balance(busiest_rq, target_rq);
4486 update_rq_clock(busiest_rq);
4487 update_rq_clock(target_rq);
4489 /* Search for an sd spanning us and the target CPU. */
4490 for_each_domain(target_cpu, sd) {
4491 if ((sd->flags & SD_LOAD_BALANCE) &&
4492 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4497 schedstat_inc(sd, alb_count);
4499 if (move_one_task(target_rq, target_cpu, busiest_rq,
4501 schedstat_inc(sd, alb_pushed);
4503 schedstat_inc(sd, alb_failed);
4505 double_unlock_balance(busiest_rq, target_rq);
4510 atomic_t load_balancer;
4511 cpumask_var_t cpu_mask;
4512 cpumask_var_t ilb_grp_nohz_mask;
4513 } nohz ____cacheline_aligned = {
4514 .load_balancer = ATOMIC_INIT(-1),
4517 int get_nohz_load_balancer(void)
4519 return atomic_read(&nohz.load_balancer);
4522 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4524 * lowest_flag_domain - Return lowest sched_domain containing flag.
4525 * @cpu: The cpu whose lowest level of sched domain is to
4527 * @flag: The flag to check for the lowest sched_domain
4528 * for the given cpu.
4530 * Returns the lowest sched_domain of a cpu which contains the given flag.
4532 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4534 struct sched_domain *sd;
4536 for_each_domain(cpu, sd)
4537 if (sd && (sd->flags & flag))
4544 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4545 * @cpu: The cpu whose domains we're iterating over.
4546 * @sd: variable holding the value of the power_savings_sd
4548 * @flag: The flag to filter the sched_domains to be iterated.
4550 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4551 * set, starting from the lowest sched_domain to the highest.
4553 #define for_each_flag_domain(cpu, sd, flag) \
4554 for (sd = lowest_flag_domain(cpu, flag); \
4555 (sd && (sd->flags & flag)); sd = sd->parent)
4558 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4559 * @ilb_group: group to be checked for semi-idleness
4561 * Returns: 1 if the group is semi-idle. 0 otherwise.
4563 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4564 * and atleast one non-idle CPU. This helper function checks if the given
4565 * sched_group is semi-idle or not.
4567 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4569 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4570 sched_group_cpus(ilb_group));
4573 * A sched_group is semi-idle when it has atleast one busy cpu
4574 * and atleast one idle cpu.
4576 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4579 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4585 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4586 * @cpu: The cpu which is nominating a new idle_load_balancer.
4588 * Returns: Returns the id of the idle load balancer if it exists,
4589 * Else, returns >= nr_cpu_ids.
4591 * This algorithm picks the idle load balancer such that it belongs to a
4592 * semi-idle powersavings sched_domain. The idea is to try and avoid
4593 * completely idle packages/cores just for the purpose of idle load balancing
4594 * when there are other idle cpu's which are better suited for that job.
4596 static int find_new_ilb(int cpu)
4598 struct sched_domain *sd;
4599 struct sched_group *ilb_group;
4602 * Have idle load balancer selection from semi-idle packages only
4603 * when power-aware load balancing is enabled
4605 if (!(sched_smt_power_savings || sched_mc_power_savings))
4609 * Optimize for the case when we have no idle CPUs or only one
4610 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4612 if (cpumask_weight(nohz.cpu_mask) < 2)
4615 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4616 ilb_group = sd->groups;
4619 if (is_semi_idle_group(ilb_group))
4620 return cpumask_first(nohz.ilb_grp_nohz_mask);
4622 ilb_group = ilb_group->next;
4624 } while (ilb_group != sd->groups);
4628 return cpumask_first(nohz.cpu_mask);
4630 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4631 static inline int find_new_ilb(int call_cpu)
4633 return cpumask_first(nohz.cpu_mask);
4638 * This routine will try to nominate the ilb (idle load balancing)
4639 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4640 * load balancing on behalf of all those cpus. If all the cpus in the system
4641 * go into this tickless mode, then there will be no ilb owner (as there is
4642 * no need for one) and all the cpus will sleep till the next wakeup event
4645 * For the ilb owner, tick is not stopped. And this tick will be used
4646 * for idle load balancing. ilb owner will still be part of
4649 * While stopping the tick, this cpu will become the ilb owner if there
4650 * is no other owner. And will be the owner till that cpu becomes busy
4651 * or if all cpus in the system stop their ticks at which point
4652 * there is no need for ilb owner.
4654 * When the ilb owner becomes busy, it nominates another owner, during the
4655 * next busy scheduler_tick()
4657 int select_nohz_load_balancer(int stop_tick)
4659 int cpu = smp_processor_id();
4662 cpu_rq(cpu)->in_nohz_recently = 1;
4664 if (!cpu_active(cpu)) {
4665 if (atomic_read(&nohz.load_balancer) != cpu)
4669 * If we are going offline and still the leader,
4672 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4678 cpumask_set_cpu(cpu, nohz.cpu_mask);
4680 /* time for ilb owner also to sleep */
4681 if (cpumask_weight(nohz.cpu_mask) == num_active_cpus()) {
4682 if (atomic_read(&nohz.load_balancer) == cpu)
4683 atomic_set(&nohz.load_balancer, -1);
4687 if (atomic_read(&nohz.load_balancer) == -1) {
4688 /* make me the ilb owner */
4689 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4691 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4694 if (!(sched_smt_power_savings ||
4695 sched_mc_power_savings))
4698 * Check to see if there is a more power-efficient
4701 new_ilb = find_new_ilb(cpu);
4702 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4703 atomic_set(&nohz.load_balancer, -1);
4704 resched_cpu(new_ilb);
4710 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4713 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4715 if (atomic_read(&nohz.load_balancer) == cpu)
4716 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4723 static DEFINE_SPINLOCK(balancing);
4726 * It checks each scheduling domain to see if it is due to be balanced,
4727 * and initiates a balancing operation if so.
4729 * Balancing parameters are set up in arch_init_sched_domains.
4731 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4734 struct rq *rq = cpu_rq(cpu);
4735 unsigned long interval;
4736 struct sched_domain *sd;
4737 /* Earliest time when we have to do rebalance again */
4738 unsigned long next_balance = jiffies + 60*HZ;
4739 int update_next_balance = 0;
4742 for_each_domain(cpu, sd) {
4743 if (!(sd->flags & SD_LOAD_BALANCE))
4746 interval = sd->balance_interval;
4747 if (idle != CPU_IDLE)
4748 interval *= sd->busy_factor;
4750 /* scale ms to jiffies */
4751 interval = msecs_to_jiffies(interval);
4752 if (unlikely(!interval))
4754 if (interval > HZ*NR_CPUS/10)
4755 interval = HZ*NR_CPUS/10;
4757 need_serialize = sd->flags & SD_SERIALIZE;
4759 if (need_serialize) {
4760 if (!spin_trylock(&balancing))
4764 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4765 if (load_balance(cpu, rq, sd, idle, &balance)) {
4767 * We've pulled tasks over so either we're no
4768 * longer idle, or one of our SMT siblings is
4771 idle = CPU_NOT_IDLE;
4773 sd->last_balance = jiffies;
4776 spin_unlock(&balancing);
4778 if (time_after(next_balance, sd->last_balance + interval)) {
4779 next_balance = sd->last_balance + interval;
4780 update_next_balance = 1;
4784 * Stop the load balance at this level. There is another
4785 * CPU in our sched group which is doing load balancing more
4793 * next_balance will be updated only when there is a need.
4794 * When the cpu is attached to null domain for ex, it will not be
4797 if (likely(update_next_balance))
4798 rq->next_balance = next_balance;
4802 * run_rebalance_domains is triggered when needed from the scheduler tick.
4803 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4804 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4806 static void run_rebalance_domains(struct softirq_action *h)
4808 int this_cpu = smp_processor_id();
4809 struct rq *this_rq = cpu_rq(this_cpu);
4810 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4811 CPU_IDLE : CPU_NOT_IDLE;
4813 rebalance_domains(this_cpu, idle);
4817 * If this cpu is the owner for idle load balancing, then do the
4818 * balancing on behalf of the other idle cpus whose ticks are
4821 if (this_rq->idle_at_tick &&
4822 atomic_read(&nohz.load_balancer) == this_cpu) {
4826 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4827 if (balance_cpu == this_cpu)
4831 * If this cpu gets work to do, stop the load balancing
4832 * work being done for other cpus. Next load
4833 * balancing owner will pick it up.
4838 rebalance_domains(balance_cpu, CPU_IDLE);
4840 rq = cpu_rq(balance_cpu);
4841 if (time_after(this_rq->next_balance, rq->next_balance))
4842 this_rq->next_balance = rq->next_balance;
4848 static inline int on_null_domain(int cpu)
4850 return !rcu_dereference(cpu_rq(cpu)->sd);
4854 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4856 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4857 * idle load balancing owner or decide to stop the periodic load balancing,
4858 * if the whole system is idle.
4860 static inline void trigger_load_balance(struct rq *rq, int cpu)
4864 * If we were in the nohz mode recently and busy at the current
4865 * scheduler tick, then check if we need to nominate new idle
4868 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4869 rq->in_nohz_recently = 0;
4871 if (atomic_read(&nohz.load_balancer) == cpu) {
4872 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4873 atomic_set(&nohz.load_balancer, -1);
4876 if (atomic_read(&nohz.load_balancer) == -1) {
4877 int ilb = find_new_ilb(cpu);
4879 if (ilb < nr_cpu_ids)
4885 * If this cpu is idle and doing idle load balancing for all the
4886 * cpus with ticks stopped, is it time for that to stop?
4888 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4889 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4895 * If this cpu is idle and the idle load balancing is done by
4896 * someone else, then no need raise the SCHED_SOFTIRQ
4898 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4899 cpumask_test_cpu(cpu, nohz.cpu_mask))
4902 /* Don't need to rebalance while attached to NULL domain */
4903 if (time_after_eq(jiffies, rq->next_balance) &&
4904 likely(!on_null_domain(cpu)))
4905 raise_softirq(SCHED_SOFTIRQ);
4908 #else /* CONFIG_SMP */
4911 * on UP we do not need to balance between CPUs:
4913 static inline void idle_balance(int cpu, struct rq *rq)
4919 DEFINE_PER_CPU(struct kernel_stat, kstat);
4921 EXPORT_PER_CPU_SYMBOL(kstat);
4924 * Return any ns on the sched_clock that have not yet been accounted in
4925 * @p in case that task is currently running.
4927 * Called with task_rq_lock() held on @rq.
4929 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4933 if (task_current(rq, p)) {
4934 update_rq_clock(rq);
4935 ns = rq->clock - p->se.exec_start;
4943 unsigned long long task_delta_exec(struct task_struct *p)
4945 unsigned long flags;
4949 rq = task_rq_lock(p, &flags);
4950 ns = do_task_delta_exec(p, rq);
4951 task_rq_unlock(rq, &flags);
4957 * Return accounted runtime for the task.
4958 * In case the task is currently running, return the runtime plus current's
4959 * pending runtime that have not been accounted yet.
4961 unsigned long long task_sched_runtime(struct task_struct *p)
4963 unsigned long flags;
4967 rq = task_rq_lock(p, &flags);
4968 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4969 task_rq_unlock(rq, &flags);
4975 * Return sum_exec_runtime for the thread group.
4976 * In case the task is currently running, return the sum plus current's
4977 * pending runtime that have not been accounted yet.
4979 * Note that the thread group might have other running tasks as well,
4980 * so the return value not includes other pending runtime that other
4981 * running tasks might have.
4983 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4985 struct task_cputime totals;
4986 unsigned long flags;
4990 rq = task_rq_lock(p, &flags);
4991 thread_group_cputime(p, &totals);
4992 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4993 task_rq_unlock(rq, &flags);
4999 * Account user cpu time to a process.
5000 * @p: the process that the cpu time gets accounted to
5001 * @cputime: the cpu time spent in user space since the last update
5002 * @cputime_scaled: cputime scaled by cpu frequency
5004 void account_user_time(struct task_struct *p, cputime_t cputime,
5005 cputime_t cputime_scaled)
5007 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5010 /* Add user time to process. */
5011 p->utime = cputime_add(p->utime, cputime);
5012 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5013 account_group_user_time(p, cputime);
5015 /* Add user time to cpustat. */
5016 tmp = cputime_to_cputime64(cputime);
5017 if (TASK_NICE(p) > 0)
5018 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5020 cpustat->user = cputime64_add(cpustat->user, tmp);
5022 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5023 /* Account for user time used */
5024 acct_update_integrals(p);
5028 * Account guest cpu time to a process.
5029 * @p: the process that the cpu time gets accounted to
5030 * @cputime: the cpu time spent in virtual machine since the last update
5031 * @cputime_scaled: cputime scaled by cpu frequency
5033 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5034 cputime_t cputime_scaled)
5037 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5039 tmp = cputime_to_cputime64(cputime);
5041 /* Add guest time to process. */
5042 p->utime = cputime_add(p->utime, cputime);
5043 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5044 account_group_user_time(p, cputime);
5045 p->gtime = cputime_add(p->gtime, cputime);
5047 /* Add guest time to cpustat. */
5048 if (TASK_NICE(p) > 0) {
5049 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5050 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5052 cpustat->user = cputime64_add(cpustat->user, tmp);
5053 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5058 * Account system cpu time to a process.
5059 * @p: the process that the cpu time gets accounted to
5060 * @hardirq_offset: the offset to subtract from hardirq_count()
5061 * @cputime: the cpu time spent in kernel space since the last update
5062 * @cputime_scaled: cputime scaled by cpu frequency
5064 void account_system_time(struct task_struct *p, int hardirq_offset,
5065 cputime_t cputime, cputime_t cputime_scaled)
5067 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5070 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5071 account_guest_time(p, cputime, cputime_scaled);
5075 /* Add system time to process. */
5076 p->stime = cputime_add(p->stime, cputime);
5077 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5078 account_group_system_time(p, cputime);
5080 /* Add system time to cpustat. */
5081 tmp = cputime_to_cputime64(cputime);
5082 if (hardirq_count() - hardirq_offset)
5083 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5084 else if (softirq_count())
5085 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5087 cpustat->system = cputime64_add(cpustat->system, tmp);
5089 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5091 /* Account for system time used */
5092 acct_update_integrals(p);
5096 * Account for involuntary wait time.
5097 * @steal: the cpu time spent in involuntary wait
5099 void account_steal_time(cputime_t cputime)
5101 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5102 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5104 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5108 * Account for idle time.
5109 * @cputime: the cpu time spent in idle wait
5111 void account_idle_time(cputime_t cputime)
5113 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5114 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5115 struct rq *rq = this_rq();
5117 if (atomic_read(&rq->nr_iowait) > 0)
5118 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5120 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5123 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5126 * Account a single tick of cpu time.
5127 * @p: the process that the cpu time gets accounted to
5128 * @user_tick: indicates if the tick is a user or a system tick
5130 void account_process_tick(struct task_struct *p, int user_tick)
5132 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
5133 struct rq *rq = this_rq();
5136 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
5137 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5138 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
5141 account_idle_time(cputime_one_jiffy);
5145 * Account multiple ticks of steal time.
5146 * @p: the process from which the cpu time has been stolen
5147 * @ticks: number of stolen ticks
5149 void account_steal_ticks(unsigned long ticks)
5151 account_steal_time(jiffies_to_cputime(ticks));
5155 * Account multiple ticks of idle time.
5156 * @ticks: number of stolen ticks
5158 void account_idle_ticks(unsigned long ticks)
5160 account_idle_time(jiffies_to_cputime(ticks));
5166 * Use precise platform statistics if available:
5168 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5169 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5175 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5177 struct task_cputime cputime;
5179 thread_group_cputime(p, &cputime);
5181 *ut = cputime.utime;
5182 *st = cputime.stime;
5186 #ifndef nsecs_to_cputime
5187 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5190 void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5192 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
5195 * Use CFS's precise accounting:
5197 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
5202 temp = (u64)(rtime * utime);
5203 do_div(temp, total);
5204 utime = (cputime_t)temp;
5209 * Compare with previous values, to keep monotonicity:
5211 p->prev_utime = max(p->prev_utime, utime);
5212 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
5214 *ut = p->prev_utime;
5215 *st = p->prev_stime;
5219 * Must be called with siglock held.
5221 void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5223 struct signal_struct *sig = p->signal;
5224 struct task_cputime cputime;
5225 cputime_t rtime, utime, total;
5227 thread_group_cputime(p, &cputime);
5229 total = cputime_add(cputime.utime, cputime.stime);
5230 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5235 temp = (u64)(rtime * cputime.utime);
5236 do_div(temp, total);
5237 utime = (cputime_t)temp;
5241 sig->prev_utime = max(sig->prev_utime, utime);
5242 sig->prev_stime = max(sig->prev_stime,
5243 cputime_sub(rtime, sig->prev_utime));
5245 *ut = sig->prev_utime;
5246 *st = sig->prev_stime;
5251 * This function gets called by the timer code, with HZ frequency.
5252 * We call it with interrupts disabled.
5254 * It also gets called by the fork code, when changing the parent's
5257 void scheduler_tick(void)
5259 int cpu = smp_processor_id();
5260 struct rq *rq = cpu_rq(cpu);
5261 struct task_struct *curr = rq->curr;
5265 raw_spin_lock(&rq->lock);
5266 update_rq_clock(rq);
5267 update_cpu_load(rq);
5268 curr->sched_class->task_tick(rq, curr, 0);
5269 raw_spin_unlock(&rq->lock);
5271 perf_event_task_tick(curr, cpu);
5274 rq->idle_at_tick = idle_cpu(cpu);
5275 trigger_load_balance(rq, cpu);
5279 notrace unsigned long get_parent_ip(unsigned long addr)
5281 if (in_lock_functions(addr)) {
5282 addr = CALLER_ADDR2;
5283 if (in_lock_functions(addr))
5284 addr = CALLER_ADDR3;
5289 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5290 defined(CONFIG_PREEMPT_TRACER))
5292 void __kprobes add_preempt_count(int val)
5294 #ifdef CONFIG_DEBUG_PREEMPT
5298 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5301 preempt_count() += val;
5302 #ifdef CONFIG_DEBUG_PREEMPT
5304 * Spinlock count overflowing soon?
5306 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5309 if (preempt_count() == val)
5310 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5312 EXPORT_SYMBOL(add_preempt_count);
5314 void __kprobes sub_preempt_count(int val)
5316 #ifdef CONFIG_DEBUG_PREEMPT
5320 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5323 * Is the spinlock portion underflowing?
5325 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5326 !(preempt_count() & PREEMPT_MASK)))
5330 if (preempt_count() == val)
5331 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5332 preempt_count() -= val;
5334 EXPORT_SYMBOL(sub_preempt_count);
5339 * Print scheduling while atomic bug:
5341 static noinline void __schedule_bug(struct task_struct *prev)
5343 struct pt_regs *regs = get_irq_regs();
5345 pr_err("BUG: scheduling while atomic: %s/%d/0x%08x\n",
5346 prev->comm, prev->pid, preempt_count());
5348 debug_show_held_locks(prev);
5350 if (irqs_disabled())
5351 print_irqtrace_events(prev);
5360 * Various schedule()-time debugging checks and statistics:
5362 static inline void schedule_debug(struct task_struct *prev)
5365 * Test if we are atomic. Since do_exit() needs to call into
5366 * schedule() atomically, we ignore that path for now.
5367 * Otherwise, whine if we are scheduling when we should not be.
5369 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5370 __schedule_bug(prev);
5372 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5374 schedstat_inc(this_rq(), sched_count);
5375 #ifdef CONFIG_SCHEDSTATS
5376 if (unlikely(prev->lock_depth >= 0)) {
5377 schedstat_inc(this_rq(), bkl_count);
5378 schedstat_inc(prev, sched_info.bkl_count);
5383 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5385 if (prev->state == TASK_RUNNING) {
5386 u64 runtime = prev->se.sum_exec_runtime;
5388 runtime -= prev->se.prev_sum_exec_runtime;
5389 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5392 * In order to avoid avg_overlap growing stale when we are
5393 * indeed overlapping and hence not getting put to sleep, grow
5394 * the avg_overlap on preemption.
5396 * We use the average preemption runtime because that
5397 * correlates to the amount of cache footprint a task can
5400 update_avg(&prev->se.avg_overlap, runtime);
5402 prev->sched_class->put_prev_task(rq, prev);
5406 * Pick up the highest-prio task:
5408 static inline struct task_struct *
5409 pick_next_task(struct rq *rq)
5411 const struct sched_class *class;
5412 struct task_struct *p;
5415 * Optimization: we know that if all tasks are in
5416 * the fair class we can call that function directly:
5418 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5419 p = fair_sched_class.pick_next_task(rq);
5424 class = sched_class_highest;
5426 p = class->pick_next_task(rq);
5430 * Will never be NULL as the idle class always
5431 * returns a non-NULL p:
5433 class = class->next;
5438 * schedule() is the main scheduler function.
5440 asmlinkage void __sched schedule(void)
5442 struct task_struct *prev, *next;
5443 unsigned long *switch_count;
5449 cpu = smp_processor_id();
5453 switch_count = &prev->nivcsw;
5455 release_kernel_lock(prev);
5456 need_resched_nonpreemptible:
5458 schedule_debug(prev);
5460 if (sched_feat(HRTICK))
5463 raw_spin_lock_irq(&rq->lock);
5464 update_rq_clock(rq);
5465 clear_tsk_need_resched(prev);
5467 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5468 if (unlikely(signal_pending_state(prev->state, prev)))
5469 prev->state = TASK_RUNNING;
5471 deactivate_task(rq, prev, 1);
5472 switch_count = &prev->nvcsw;
5475 pre_schedule(rq, prev);
5477 if (unlikely(!rq->nr_running))
5478 idle_balance(cpu, rq);
5480 put_prev_task(rq, prev);
5481 next = pick_next_task(rq);
5483 if (likely(prev != next)) {
5484 sched_info_switch(prev, next);
5485 perf_event_task_sched_out(prev, next, cpu);
5491 context_switch(rq, prev, next); /* unlocks the rq */
5493 * the context switch might have flipped the stack from under
5494 * us, hence refresh the local variables.
5496 cpu = smp_processor_id();
5499 raw_spin_unlock_irq(&rq->lock);
5503 if (unlikely(reacquire_kernel_lock(current) < 0))
5504 goto need_resched_nonpreemptible;
5506 preempt_enable_no_resched();
5510 EXPORT_SYMBOL(schedule);
5512 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5514 * Look out! "owner" is an entirely speculative pointer
5515 * access and not reliable.
5517 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5522 if (!sched_feat(OWNER_SPIN))
5525 #ifdef CONFIG_DEBUG_PAGEALLOC
5527 * Need to access the cpu field knowing that
5528 * DEBUG_PAGEALLOC could have unmapped it if
5529 * the mutex owner just released it and exited.
5531 if (probe_kernel_address(&owner->cpu, cpu))
5538 * Even if the access succeeded (likely case),
5539 * the cpu field may no longer be valid.
5541 if (cpu >= nr_cpumask_bits)
5545 * We need to validate that we can do a
5546 * get_cpu() and that we have the percpu area.
5548 if (!cpu_online(cpu))
5555 * Owner changed, break to re-assess state.
5557 if (lock->owner != owner)
5561 * Is that owner really running on that cpu?
5563 if (task_thread_info(rq->curr) != owner || need_resched())
5573 #ifdef CONFIG_PREEMPT
5575 * this is the entry point to schedule() from in-kernel preemption
5576 * off of preempt_enable. Kernel preemptions off return from interrupt
5577 * occur there and call schedule directly.
5579 asmlinkage void __sched preempt_schedule(void)
5581 struct thread_info *ti = current_thread_info();
5584 * If there is a non-zero preempt_count or interrupts are disabled,
5585 * we do not want to preempt the current task. Just return..
5587 if (likely(ti->preempt_count || irqs_disabled()))
5591 add_preempt_count(PREEMPT_ACTIVE);
5593 sub_preempt_count(PREEMPT_ACTIVE);
5596 * Check again in case we missed a preemption opportunity
5597 * between schedule and now.
5600 } while (need_resched());
5602 EXPORT_SYMBOL(preempt_schedule);
5605 * this is the entry point to schedule() from kernel preemption
5606 * off of irq context.
5607 * Note, that this is called and return with irqs disabled. This will
5608 * protect us against recursive calling from irq.
5610 asmlinkage void __sched preempt_schedule_irq(void)
5612 struct thread_info *ti = current_thread_info();
5614 /* Catch callers which need to be fixed */
5615 BUG_ON(ti->preempt_count || !irqs_disabled());
5618 add_preempt_count(PREEMPT_ACTIVE);
5621 local_irq_disable();
5622 sub_preempt_count(PREEMPT_ACTIVE);
5625 * Check again in case we missed a preemption opportunity
5626 * between schedule and now.
5629 } while (need_resched());
5632 #endif /* CONFIG_PREEMPT */
5634 int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
5637 return try_to_wake_up(curr->private, mode, wake_flags);
5639 EXPORT_SYMBOL(default_wake_function);
5642 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5643 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5644 * number) then we wake all the non-exclusive tasks and one exclusive task.
5646 * There are circumstances in which we can try to wake a task which has already
5647 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5648 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5650 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5651 int nr_exclusive, int wake_flags, void *key)
5653 wait_queue_t *curr, *next;
5655 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5656 unsigned flags = curr->flags;
5658 if (curr->func(curr, mode, wake_flags, key) &&
5659 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5665 * __wake_up - wake up threads blocked on a waitqueue.
5667 * @mode: which threads
5668 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5669 * @key: is directly passed to the wakeup function
5671 * It may be assumed that this function implies a write memory barrier before
5672 * changing the task state if and only if any tasks are woken up.
5674 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5675 int nr_exclusive, void *key)
5677 unsigned long flags;
5679 spin_lock_irqsave(&q->lock, flags);
5680 __wake_up_common(q, mode, nr_exclusive, 0, key);
5681 spin_unlock_irqrestore(&q->lock, flags);
5683 EXPORT_SYMBOL(__wake_up);
5686 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5688 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5690 __wake_up_common(q, mode, 1, 0, NULL);
5693 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5695 __wake_up_common(q, mode, 1, 0, key);
5699 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5701 * @mode: which threads
5702 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5703 * @key: opaque value to be passed to wakeup targets
5705 * The sync wakeup differs that the waker knows that it will schedule
5706 * away soon, so while the target thread will be woken up, it will not
5707 * be migrated to another CPU - ie. the two threads are 'synchronized'
5708 * with each other. This can prevent needless bouncing between CPUs.
5710 * On UP it can prevent extra preemption.
5712 * It may be assumed that this function implies a write memory barrier before
5713 * changing the task state if and only if any tasks are woken up.
5715 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5716 int nr_exclusive, void *key)
5718 unsigned long flags;
5719 int wake_flags = WF_SYNC;
5724 if (unlikely(!nr_exclusive))
5727 spin_lock_irqsave(&q->lock, flags);
5728 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
5729 spin_unlock_irqrestore(&q->lock, flags);
5731 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5734 * __wake_up_sync - see __wake_up_sync_key()
5736 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5738 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5740 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5743 * complete: - signals a single thread waiting on this completion
5744 * @x: holds the state of this particular completion
5746 * This will wake up a single thread waiting on this completion. Threads will be
5747 * awakened in the same order in which they were queued.
5749 * See also complete_all(), wait_for_completion() and related routines.
5751 * It may be assumed that this function implies a write memory barrier before
5752 * changing the task state if and only if any tasks are woken up.
5754 void complete(struct completion *x)
5756 unsigned long flags;
5758 spin_lock_irqsave(&x->wait.lock, flags);
5760 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5761 spin_unlock_irqrestore(&x->wait.lock, flags);
5763 EXPORT_SYMBOL(complete);
5766 * complete_all: - signals all threads waiting on this completion
5767 * @x: holds the state of this particular completion
5769 * This will wake up all threads waiting on this particular completion event.
5771 * It may be assumed that this function implies a write memory barrier before
5772 * changing the task state if and only if any tasks are woken up.
5774 void complete_all(struct completion *x)
5776 unsigned long flags;
5778 spin_lock_irqsave(&x->wait.lock, flags);
5779 x->done += UINT_MAX/2;
5780 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5781 spin_unlock_irqrestore(&x->wait.lock, flags);
5783 EXPORT_SYMBOL(complete_all);
5785 static inline long __sched
5786 do_wait_for_common(struct completion *x, long timeout, int state)
5789 DECLARE_WAITQUEUE(wait, current);
5791 wait.flags |= WQ_FLAG_EXCLUSIVE;
5792 __add_wait_queue_tail(&x->wait, &wait);
5794 if (signal_pending_state(state, current)) {
5795 timeout = -ERESTARTSYS;
5798 __set_current_state(state);
5799 spin_unlock_irq(&x->wait.lock);
5800 timeout = schedule_timeout(timeout);
5801 spin_lock_irq(&x->wait.lock);
5802 } while (!x->done && timeout);
5803 __remove_wait_queue(&x->wait, &wait);
5808 return timeout ?: 1;
5812 wait_for_common(struct completion *x, long timeout, int state)
5816 spin_lock_irq(&x->wait.lock);
5817 timeout = do_wait_for_common(x, timeout, state);
5818 spin_unlock_irq(&x->wait.lock);
5823 * wait_for_completion: - waits for completion of a task
5824 * @x: holds the state of this particular completion
5826 * This waits to be signaled for completion of a specific task. It is NOT
5827 * interruptible and there is no timeout.
5829 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5830 * and interrupt capability. Also see complete().
5832 void __sched wait_for_completion(struct completion *x)
5834 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5836 EXPORT_SYMBOL(wait_for_completion);
5839 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5840 * @x: holds the state of this particular completion
5841 * @timeout: timeout value in jiffies
5843 * This waits for either a completion of a specific task to be signaled or for a
5844 * specified timeout to expire. The timeout is in jiffies. It is not
5847 unsigned long __sched
5848 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5850 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5852 EXPORT_SYMBOL(wait_for_completion_timeout);
5855 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5856 * @x: holds the state of this particular completion
5858 * This waits for completion of a specific task to be signaled. It is
5861 int __sched wait_for_completion_interruptible(struct completion *x)
5863 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5864 if (t == -ERESTARTSYS)
5868 EXPORT_SYMBOL(wait_for_completion_interruptible);
5871 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5872 * @x: holds the state of this particular completion
5873 * @timeout: timeout value in jiffies
5875 * This waits for either a completion of a specific task to be signaled or for a
5876 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5878 unsigned long __sched
5879 wait_for_completion_interruptible_timeout(struct completion *x,
5880 unsigned long timeout)
5882 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5884 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5887 * wait_for_completion_killable: - waits for completion of a task (killable)
5888 * @x: holds the state of this particular completion
5890 * This waits to be signaled for completion of a specific task. It can be
5891 * interrupted by a kill signal.
5893 int __sched wait_for_completion_killable(struct completion *x)
5895 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5896 if (t == -ERESTARTSYS)
5900 EXPORT_SYMBOL(wait_for_completion_killable);
5903 * try_wait_for_completion - try to decrement a completion without blocking
5904 * @x: completion structure
5906 * Returns: 0 if a decrement cannot be done without blocking
5907 * 1 if a decrement succeeded.
5909 * If a completion is being used as a counting completion,
5910 * attempt to decrement the counter without blocking. This
5911 * enables us to avoid waiting if the resource the completion
5912 * is protecting is not available.
5914 bool try_wait_for_completion(struct completion *x)
5916 unsigned long flags;
5919 spin_lock_irqsave(&x->wait.lock, flags);
5924 spin_unlock_irqrestore(&x->wait.lock, flags);
5927 EXPORT_SYMBOL(try_wait_for_completion);
5930 * completion_done - Test to see if a completion has any waiters
5931 * @x: completion structure
5933 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5934 * 1 if there are no waiters.
5937 bool completion_done(struct completion *x)
5939 unsigned long flags;
5942 spin_lock_irqsave(&x->wait.lock, flags);
5945 spin_unlock_irqrestore(&x->wait.lock, flags);
5948 EXPORT_SYMBOL(completion_done);
5951 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5953 unsigned long flags;
5956 init_waitqueue_entry(&wait, current);
5958 __set_current_state(state);
5960 spin_lock_irqsave(&q->lock, flags);
5961 __add_wait_queue(q, &wait);
5962 spin_unlock(&q->lock);
5963 timeout = schedule_timeout(timeout);
5964 spin_lock_irq(&q->lock);
5965 __remove_wait_queue(q, &wait);
5966 spin_unlock_irqrestore(&q->lock, flags);
5971 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5973 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5975 EXPORT_SYMBOL(interruptible_sleep_on);
5978 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5980 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5982 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5984 void __sched sleep_on(wait_queue_head_t *q)
5986 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5988 EXPORT_SYMBOL(sleep_on);
5990 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5992 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5994 EXPORT_SYMBOL(sleep_on_timeout);
5996 #ifdef CONFIG_RT_MUTEXES
5999 * rt_mutex_setprio - set the current priority of a task
6001 * @prio: prio value (kernel-internal form)
6003 * This function changes the 'effective' priority of a task. It does
6004 * not touch ->normal_prio like __setscheduler().
6006 * Used by the rt_mutex code to implement priority inheritance logic.
6008 void rt_mutex_setprio(struct task_struct *p, int prio)
6010 unsigned long flags;
6011 int oldprio, on_rq, running;
6013 const struct sched_class *prev_class = p->sched_class;
6015 BUG_ON(prio < 0 || prio > MAX_PRIO);
6017 rq = task_rq_lock(p, &flags);
6018 update_rq_clock(rq);
6021 on_rq = p->se.on_rq;
6022 running = task_current(rq, p);
6024 dequeue_task(rq, p, 0);
6026 p->sched_class->put_prev_task(rq, p);
6029 p->sched_class = &rt_sched_class;
6031 p->sched_class = &fair_sched_class;
6036 p->sched_class->set_curr_task(rq);
6038 enqueue_task(rq, p, 0);
6040 check_class_changed(rq, p, prev_class, oldprio, running);
6042 task_rq_unlock(rq, &flags);
6047 void set_user_nice(struct task_struct *p, long nice)
6049 int old_prio, delta, on_rq;
6050 unsigned long flags;
6053 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6056 * We have to be careful, if called from sys_setpriority(),
6057 * the task might be in the middle of scheduling on another CPU.
6059 rq = task_rq_lock(p, &flags);
6060 update_rq_clock(rq);
6062 * The RT priorities are set via sched_setscheduler(), but we still
6063 * allow the 'normal' nice value to be set - but as expected
6064 * it wont have any effect on scheduling until the task is
6065 * SCHED_FIFO/SCHED_RR:
6067 if (task_has_rt_policy(p)) {
6068 p->static_prio = NICE_TO_PRIO(nice);
6071 on_rq = p->se.on_rq;
6073 dequeue_task(rq, p, 0);
6075 p->static_prio = NICE_TO_PRIO(nice);
6078 p->prio = effective_prio(p);
6079 delta = p->prio - old_prio;
6082 enqueue_task(rq, p, 0);
6084 * If the task increased its priority or is running and
6085 * lowered its priority, then reschedule its CPU:
6087 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6088 resched_task(rq->curr);
6091 task_rq_unlock(rq, &flags);
6093 EXPORT_SYMBOL(set_user_nice);
6096 * can_nice - check if a task can reduce its nice value
6100 int can_nice(const struct task_struct *p, const int nice)
6102 /* convert nice value [19,-20] to rlimit style value [1,40] */
6103 int nice_rlim = 20 - nice;
6105 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6106 capable(CAP_SYS_NICE));
6109 #ifdef __ARCH_WANT_SYS_NICE
6112 * sys_nice - change the priority of the current process.
6113 * @increment: priority increment
6115 * sys_setpriority is a more generic, but much slower function that
6116 * does similar things.
6118 SYSCALL_DEFINE1(nice, int, increment)
6123 * Setpriority might change our priority at the same moment.
6124 * We don't have to worry. Conceptually one call occurs first
6125 * and we have a single winner.
6127 if (increment < -40)
6132 nice = TASK_NICE(current) + increment;
6138 if (increment < 0 && !can_nice(current, nice))
6141 retval = security_task_setnice(current, nice);
6145 set_user_nice(current, nice);
6152 * task_prio - return the priority value of a given task.
6153 * @p: the task in question.
6155 * This is the priority value as seen by users in /proc.
6156 * RT tasks are offset by -200. Normal tasks are centered
6157 * around 0, value goes from -16 to +15.
6159 int task_prio(const struct task_struct *p)
6161 return p->prio - MAX_RT_PRIO;
6165 * task_nice - return the nice value of a given task.
6166 * @p: the task in question.
6168 int task_nice(const struct task_struct *p)
6170 return TASK_NICE(p);
6172 EXPORT_SYMBOL(task_nice);
6175 * idle_cpu - is a given cpu idle currently?
6176 * @cpu: the processor in question.
6178 int idle_cpu(int cpu)
6180 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6184 * idle_task - return the idle task for a given cpu.
6185 * @cpu: the processor in question.
6187 struct task_struct *idle_task(int cpu)
6189 return cpu_rq(cpu)->idle;
6193 * find_process_by_pid - find a process with a matching PID value.
6194 * @pid: the pid in question.
6196 static struct task_struct *find_process_by_pid(pid_t pid)
6198 return pid ? find_task_by_vpid(pid) : current;
6201 /* Actually do priority change: must hold rq lock. */
6203 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6205 BUG_ON(p->se.on_rq);
6208 p->rt_priority = prio;
6209 p->normal_prio = normal_prio(p);
6210 /* we are holding p->pi_lock already */
6211 p->prio = rt_mutex_getprio(p);
6212 if (rt_prio(p->prio))
6213 p->sched_class = &rt_sched_class;
6215 p->sched_class = &fair_sched_class;
6220 * check the target process has a UID that matches the current process's
6222 static bool check_same_owner(struct task_struct *p)
6224 const struct cred *cred = current_cred(), *pcred;
6228 pcred = __task_cred(p);
6229 match = (cred->euid == pcred->euid ||
6230 cred->euid == pcred->uid);
6235 static int __sched_setscheduler(struct task_struct *p, int policy,
6236 struct sched_param *param, bool user)
6238 int retval, oldprio, oldpolicy = -1, on_rq, running;
6239 unsigned long flags;
6240 const struct sched_class *prev_class = p->sched_class;
6244 /* may grab non-irq protected spin_locks */
6245 BUG_ON(in_interrupt());
6247 /* double check policy once rq lock held */
6249 reset_on_fork = p->sched_reset_on_fork;
6250 policy = oldpolicy = p->policy;
6252 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6253 policy &= ~SCHED_RESET_ON_FORK;
6255 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6256 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6257 policy != SCHED_IDLE)
6262 * Valid priorities for SCHED_FIFO and SCHED_RR are
6263 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6264 * SCHED_BATCH and SCHED_IDLE is 0.
6266 if (param->sched_priority < 0 ||
6267 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6268 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6270 if (rt_policy(policy) != (param->sched_priority != 0))
6274 * Allow unprivileged RT tasks to decrease priority:
6276 if (user && !capable(CAP_SYS_NICE)) {
6277 if (rt_policy(policy)) {
6278 unsigned long rlim_rtprio;
6280 if (!lock_task_sighand(p, &flags))
6282 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6283 unlock_task_sighand(p, &flags);
6285 /* can't set/change the rt policy */
6286 if (policy != p->policy && !rlim_rtprio)
6289 /* can't increase priority */
6290 if (param->sched_priority > p->rt_priority &&
6291 param->sched_priority > rlim_rtprio)
6295 * Like positive nice levels, dont allow tasks to
6296 * move out of SCHED_IDLE either:
6298 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6301 /* can't change other user's priorities */
6302 if (!check_same_owner(p))
6305 /* Normal users shall not reset the sched_reset_on_fork flag */
6306 if (p->sched_reset_on_fork && !reset_on_fork)
6311 #ifdef CONFIG_RT_GROUP_SCHED
6313 * Do not allow realtime tasks into groups that have no runtime
6316 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6317 task_group(p)->rt_bandwidth.rt_runtime == 0)
6321 retval = security_task_setscheduler(p, policy, param);
6327 * make sure no PI-waiters arrive (or leave) while we are
6328 * changing the priority of the task:
6330 raw_spin_lock_irqsave(&p->pi_lock, flags);
6332 * To be able to change p->policy safely, the apropriate
6333 * runqueue lock must be held.
6335 rq = __task_rq_lock(p);
6336 /* recheck policy now with rq lock held */
6337 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6338 policy = oldpolicy = -1;
6339 __task_rq_unlock(rq);
6340 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6343 update_rq_clock(rq);
6344 on_rq = p->se.on_rq;
6345 running = task_current(rq, p);
6347 deactivate_task(rq, p, 0);
6349 p->sched_class->put_prev_task(rq, p);
6351 p->sched_reset_on_fork = reset_on_fork;
6354 __setscheduler(rq, p, policy, param->sched_priority);
6357 p->sched_class->set_curr_task(rq);
6359 activate_task(rq, p, 0);
6361 check_class_changed(rq, p, prev_class, oldprio, running);
6363 __task_rq_unlock(rq);
6364 raw_spin_unlock_irqrestore(&p->pi_lock, flags);
6366 rt_mutex_adjust_pi(p);
6372 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6373 * @p: the task in question.
6374 * @policy: new policy.
6375 * @param: structure containing the new RT priority.
6377 * NOTE that the task may be already dead.
6379 int sched_setscheduler(struct task_struct *p, int policy,
6380 struct sched_param *param)
6382 return __sched_setscheduler(p, policy, param, true);
6384 EXPORT_SYMBOL_GPL(sched_setscheduler);
6387 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6388 * @p: the task in question.
6389 * @policy: new policy.
6390 * @param: structure containing the new RT priority.
6392 * Just like sched_setscheduler, only don't bother checking if the
6393 * current context has permission. For example, this is needed in
6394 * stop_machine(): we create temporary high priority worker threads,
6395 * but our caller might not have that capability.
6397 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6398 struct sched_param *param)
6400 return __sched_setscheduler(p, policy, param, false);
6404 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6406 struct sched_param lparam;
6407 struct task_struct *p;
6410 if (!param || pid < 0)
6412 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6417 p = find_process_by_pid(pid);
6419 retval = sched_setscheduler(p, policy, &lparam);
6426 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6427 * @pid: the pid in question.
6428 * @policy: new policy.
6429 * @param: structure containing the new RT priority.
6431 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6432 struct sched_param __user *, param)
6434 /* negative values for policy are not valid */
6438 return do_sched_setscheduler(pid, policy, param);
6442 * sys_sched_setparam - set/change the RT priority of a thread
6443 * @pid: the pid in question.
6444 * @param: structure containing the new RT priority.
6446 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6448 return do_sched_setscheduler(pid, -1, param);
6452 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6453 * @pid: the pid in question.
6455 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6457 struct task_struct *p;
6465 p = find_process_by_pid(pid);
6467 retval = security_task_getscheduler(p);
6470 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6477 * sys_sched_getparam - get the RT priority of a thread
6478 * @pid: the pid in question.
6479 * @param: structure containing the RT priority.
6481 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6483 struct sched_param lp;
6484 struct task_struct *p;
6487 if (!param || pid < 0)
6491 p = find_process_by_pid(pid);
6496 retval = security_task_getscheduler(p);
6500 lp.sched_priority = p->rt_priority;
6504 * This one might sleep, we cannot do it with a spinlock held ...
6506 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6515 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6517 cpumask_var_t cpus_allowed, new_mask;
6518 struct task_struct *p;
6524 p = find_process_by_pid(pid);
6531 /* Prevent p going away */
6535 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6539 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6541 goto out_free_cpus_allowed;
6544 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6547 retval = security_task_setscheduler(p, 0, NULL);
6551 cpuset_cpus_allowed(p, cpus_allowed);
6552 cpumask_and(new_mask, in_mask, cpus_allowed);
6554 retval = set_cpus_allowed_ptr(p, new_mask);
6557 cpuset_cpus_allowed(p, cpus_allowed);
6558 if (!cpumask_subset(new_mask, cpus_allowed)) {
6560 * We must have raced with a concurrent cpuset
6561 * update. Just reset the cpus_allowed to the
6562 * cpuset's cpus_allowed
6564 cpumask_copy(new_mask, cpus_allowed);
6569 free_cpumask_var(new_mask);
6570 out_free_cpus_allowed:
6571 free_cpumask_var(cpus_allowed);
6578 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6579 struct cpumask *new_mask)
6581 if (len < cpumask_size())
6582 cpumask_clear(new_mask);
6583 else if (len > cpumask_size())
6584 len = cpumask_size();
6586 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6590 * sys_sched_setaffinity - set the cpu affinity of a process
6591 * @pid: pid of the process
6592 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6593 * @user_mask_ptr: user-space pointer to the new cpu mask
6595 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6596 unsigned long __user *, user_mask_ptr)
6598 cpumask_var_t new_mask;
6601 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6604 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6606 retval = sched_setaffinity(pid, new_mask);
6607 free_cpumask_var(new_mask);
6611 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6613 struct task_struct *p;
6614 unsigned long flags;
6622 p = find_process_by_pid(pid);
6626 retval = security_task_getscheduler(p);
6630 rq = task_rq_lock(p, &flags);
6631 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6632 task_rq_unlock(rq, &flags);
6642 * sys_sched_getaffinity - get the cpu affinity of a process
6643 * @pid: pid of the process
6644 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6645 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6647 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6648 unsigned long __user *, user_mask_ptr)
6653 if (len < cpumask_size())
6656 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6659 ret = sched_getaffinity(pid, mask);
6661 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6664 ret = cpumask_size();
6666 free_cpumask_var(mask);
6672 * sys_sched_yield - yield the current processor to other threads.
6674 * This function yields the current CPU to other tasks. If there are no
6675 * other threads running on this CPU then this function will return.
6677 SYSCALL_DEFINE0(sched_yield)
6679 struct rq *rq = this_rq_lock();
6681 schedstat_inc(rq, yld_count);
6682 current->sched_class->yield_task(rq);
6685 * Since we are going to call schedule() anyway, there's
6686 * no need to preempt or enable interrupts:
6688 __release(rq->lock);
6689 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6690 do_raw_spin_unlock(&rq->lock);
6691 preempt_enable_no_resched();
6698 static inline int should_resched(void)
6700 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6703 static void __cond_resched(void)
6705 add_preempt_count(PREEMPT_ACTIVE);
6707 sub_preempt_count(PREEMPT_ACTIVE);
6710 int __sched _cond_resched(void)
6712 if (should_resched()) {
6718 EXPORT_SYMBOL(_cond_resched);
6721 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6722 * call schedule, and on return reacquire the lock.
6724 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6725 * operations here to prevent schedule() from being called twice (once via
6726 * spin_unlock(), once by hand).
6728 int __cond_resched_lock(spinlock_t *lock)
6730 int resched = should_resched();
6733 lockdep_assert_held(lock);
6735 if (spin_needbreak(lock) || resched) {
6746 EXPORT_SYMBOL(__cond_resched_lock);
6748 int __sched __cond_resched_softirq(void)
6750 BUG_ON(!in_softirq());
6752 if (should_resched()) {
6760 EXPORT_SYMBOL(__cond_resched_softirq);
6763 * yield - yield the current processor to other threads.
6765 * This is a shortcut for kernel-space yielding - it marks the
6766 * thread runnable and calls sys_sched_yield().
6768 void __sched yield(void)
6770 set_current_state(TASK_RUNNING);
6773 EXPORT_SYMBOL(yield);
6776 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6777 * that process accounting knows that this is a task in IO wait state.
6779 void __sched io_schedule(void)
6781 struct rq *rq = raw_rq();
6783 delayacct_blkio_start();
6784 atomic_inc(&rq->nr_iowait);
6785 current->in_iowait = 1;
6787 current->in_iowait = 0;
6788 atomic_dec(&rq->nr_iowait);
6789 delayacct_blkio_end();
6791 EXPORT_SYMBOL(io_schedule);
6793 long __sched io_schedule_timeout(long timeout)
6795 struct rq *rq = raw_rq();
6798 delayacct_blkio_start();
6799 atomic_inc(&rq->nr_iowait);
6800 current->in_iowait = 1;
6801 ret = schedule_timeout(timeout);
6802 current->in_iowait = 0;
6803 atomic_dec(&rq->nr_iowait);
6804 delayacct_blkio_end();
6809 * sys_sched_get_priority_max - return maximum RT priority.
6810 * @policy: scheduling class.
6812 * this syscall returns the maximum rt_priority that can be used
6813 * by a given scheduling class.
6815 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6822 ret = MAX_USER_RT_PRIO-1;
6834 * sys_sched_get_priority_min - return minimum RT priority.
6835 * @policy: scheduling class.
6837 * this syscall returns the minimum rt_priority that can be used
6838 * by a given scheduling class.
6840 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6858 * sys_sched_rr_get_interval - return the default timeslice of a process.
6859 * @pid: pid of the process.
6860 * @interval: userspace pointer to the timeslice value.
6862 * this syscall writes the default timeslice value of a given process
6863 * into the user-space timespec buffer. A value of '0' means infinity.
6865 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6866 struct timespec __user *, interval)
6868 struct task_struct *p;
6869 unsigned int time_slice;
6870 unsigned long flags;
6880 p = find_process_by_pid(pid);
6884 retval = security_task_getscheduler(p);
6888 rq = task_rq_lock(p, &flags);
6889 time_slice = p->sched_class->get_rr_interval(rq, p);
6890 task_rq_unlock(rq, &flags);
6893 jiffies_to_timespec(time_slice, &t);
6894 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6902 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6904 void sched_show_task(struct task_struct *p)
6906 unsigned long free = 0;
6909 state = p->state ? __ffs(p->state) + 1 : 0;
6910 pr_info("%-13.13s %c", p->comm,
6911 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6912 #if BITS_PER_LONG == 32
6913 if (state == TASK_RUNNING)
6914 pr_cont(" running ");
6916 pr_cont(" %08lx ", thread_saved_pc(p));
6918 if (state == TASK_RUNNING)
6919 pr_cont(" running task ");
6921 pr_cont(" %016lx ", thread_saved_pc(p));
6923 #ifdef CONFIG_DEBUG_STACK_USAGE
6924 free = stack_not_used(p);
6926 pr_cont("%5lu %5d %6d 0x%08lx\n", free,
6927 task_pid_nr(p), task_pid_nr(p->real_parent),
6928 (unsigned long)task_thread_info(p)->flags);
6930 show_stack(p, NULL);
6933 void show_state_filter(unsigned long state_filter)
6935 struct task_struct *g, *p;
6937 #if BITS_PER_LONG == 32
6938 pr_info(" task PC stack pid father\n");
6940 pr_info(" task PC stack pid father\n");
6942 read_lock(&tasklist_lock);
6943 do_each_thread(g, p) {
6945 * reset the NMI-timeout, listing all files on a slow
6946 * console might take alot of time:
6948 touch_nmi_watchdog();
6949 if (!state_filter || (p->state & state_filter))
6951 } while_each_thread(g, p);
6953 touch_all_softlockup_watchdogs();
6955 #ifdef CONFIG_SCHED_DEBUG
6956 sysrq_sched_debug_show();
6958 read_unlock(&tasklist_lock);
6960 * Only show locks if all tasks are dumped:
6963 debug_show_all_locks();
6966 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6968 idle->sched_class = &idle_sched_class;
6972 * init_idle - set up an idle thread for a given CPU
6973 * @idle: task in question
6974 * @cpu: cpu the idle task belongs to
6976 * NOTE: this function does not set the idle thread's NEED_RESCHED
6977 * flag, to make booting more robust.
6979 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6981 struct rq *rq = cpu_rq(cpu);
6982 unsigned long flags;
6984 raw_spin_lock_irqsave(&rq->lock, flags);
6987 idle->se.exec_start = sched_clock();
6989 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6990 __set_task_cpu(idle, cpu);
6992 rq->curr = rq->idle = idle;
6993 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6996 raw_spin_unlock_irqrestore(&rq->lock, flags);
6998 /* Set the preempt count _outside_ the spinlocks! */
6999 #if defined(CONFIG_PREEMPT)
7000 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7002 task_thread_info(idle)->preempt_count = 0;
7005 * The idle tasks have their own, simple scheduling class:
7007 idle->sched_class = &idle_sched_class;
7008 ftrace_graph_init_task(idle);
7012 * In a system that switches off the HZ timer nohz_cpu_mask
7013 * indicates which cpus entered this state. This is used
7014 * in the rcu update to wait only for active cpus. For system
7015 * which do not switch off the HZ timer nohz_cpu_mask should
7016 * always be CPU_BITS_NONE.
7018 cpumask_var_t nohz_cpu_mask;
7021 * Increase the granularity value when there are more CPUs,
7022 * because with more CPUs the 'effective latency' as visible
7023 * to users decreases. But the relationship is not linear,
7024 * so pick a second-best guess by going with the log2 of the
7027 * This idea comes from the SD scheduler of Con Kolivas:
7029 static int get_update_sysctl_factor(void)
7031 unsigned int cpus = min_t(int, num_online_cpus(), 8);
7032 unsigned int factor;
7034 switch (sysctl_sched_tunable_scaling) {
7035 case SCHED_TUNABLESCALING_NONE:
7038 case SCHED_TUNABLESCALING_LINEAR:
7041 case SCHED_TUNABLESCALING_LOG:
7043 factor = 1 + ilog2(cpus);
7050 static void update_sysctl(void)
7052 unsigned int factor = get_update_sysctl_factor();
7054 #define SET_SYSCTL(name) \
7055 (sysctl_##name = (factor) * normalized_sysctl_##name)
7056 SET_SYSCTL(sched_min_granularity);
7057 SET_SYSCTL(sched_latency);
7058 SET_SYSCTL(sched_wakeup_granularity);
7059 SET_SYSCTL(sched_shares_ratelimit);
7063 static inline void sched_init_granularity(void)
7070 * This is how migration works:
7072 * 1) we queue a struct migration_req structure in the source CPU's
7073 * runqueue and wake up that CPU's migration thread.
7074 * 2) we down() the locked semaphore => thread blocks.
7075 * 3) migration thread wakes up (implicitly it forces the migrated
7076 * thread off the CPU)
7077 * 4) it gets the migration request and checks whether the migrated
7078 * task is still in the wrong runqueue.
7079 * 5) if it's in the wrong runqueue then the migration thread removes
7080 * it and puts it into the right queue.
7081 * 6) migration thread up()s the semaphore.
7082 * 7) we wake up and the migration is done.
7086 * Change a given task's CPU affinity. Migrate the thread to a
7087 * proper CPU and schedule it away if the CPU it's executing on
7088 * is removed from the allowed bitmask.
7090 * NOTE: the caller must have a valid reference to the task, the
7091 * task must not exit() & deallocate itself prematurely. The
7092 * call is not atomic; no spinlocks may be held.
7094 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7096 struct migration_req req;
7097 unsigned long flags;
7101 rq = task_rq_lock(p, &flags);
7102 if (!cpumask_intersects(new_mask, cpu_active_mask)) {
7107 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7108 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7113 if (p->sched_class->set_cpus_allowed)
7114 p->sched_class->set_cpus_allowed(p, new_mask);
7116 cpumask_copy(&p->cpus_allowed, new_mask);
7117 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7120 /* Can the task run on the task's current CPU? If so, we're done */
7121 if (cpumask_test_cpu(task_cpu(p), new_mask))
7124 if (migrate_task(p, cpumask_any_and(cpu_active_mask, new_mask), &req)) {
7125 /* Need help from migration thread: drop lock and wait. */
7126 struct task_struct *mt = rq->migration_thread;
7128 get_task_struct(mt);
7129 task_rq_unlock(rq, &flags);
7130 wake_up_process(rq->migration_thread);
7131 put_task_struct(mt);
7132 wait_for_completion(&req.done);
7133 tlb_migrate_finish(p->mm);
7137 task_rq_unlock(rq, &flags);
7141 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7144 * Move (not current) task off this cpu, onto dest cpu. We're doing
7145 * this because either it can't run here any more (set_cpus_allowed()
7146 * away from this CPU, or CPU going down), or because we're
7147 * attempting to rebalance this task on exec (sched_exec).
7149 * So we race with normal scheduler movements, but that's OK, as long
7150 * as the task is no longer on this CPU.
7152 * Returns non-zero if task was successfully migrated.
7154 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7156 struct rq *rq_dest, *rq_src;
7159 if (unlikely(!cpu_active(dest_cpu)))
7162 rq_src = cpu_rq(src_cpu);
7163 rq_dest = cpu_rq(dest_cpu);
7165 double_rq_lock(rq_src, rq_dest);
7166 /* Already moved. */
7167 if (task_cpu(p) != src_cpu)
7169 /* Affinity changed (again). */
7170 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7173 on_rq = p->se.on_rq;
7175 deactivate_task(rq_src, p, 0);
7177 set_task_cpu(p, dest_cpu);
7179 activate_task(rq_dest, p, 0);
7180 check_preempt_curr(rq_dest, p, 0);
7185 double_rq_unlock(rq_src, rq_dest);
7189 #define RCU_MIGRATION_IDLE 0
7190 #define RCU_MIGRATION_NEED_QS 1
7191 #define RCU_MIGRATION_GOT_QS 2
7192 #define RCU_MIGRATION_MUST_SYNC 3
7195 * migration_thread - this is a highprio system thread that performs
7196 * thread migration by bumping thread off CPU then 'pushing' onto
7199 static int migration_thread(void *data)
7202 int cpu = (long)data;
7206 BUG_ON(rq->migration_thread != current);
7208 set_current_state(TASK_INTERRUPTIBLE);
7209 while (!kthread_should_stop()) {
7210 struct migration_req *req;
7211 struct list_head *head;
7213 raw_spin_lock_irq(&rq->lock);
7215 if (cpu_is_offline(cpu)) {
7216 raw_spin_unlock_irq(&rq->lock);
7220 if (rq->active_balance) {
7221 active_load_balance(rq, cpu);
7222 rq->active_balance = 0;
7225 head = &rq->migration_queue;
7227 if (list_empty(head)) {
7228 raw_spin_unlock_irq(&rq->lock);
7230 set_current_state(TASK_INTERRUPTIBLE);
7233 req = list_entry(head->next, struct migration_req, list);
7234 list_del_init(head->next);
7236 if (req->task != NULL) {
7237 raw_spin_unlock(&rq->lock);
7238 __migrate_task(req->task, cpu, req->dest_cpu);
7239 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7240 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7241 raw_spin_unlock(&rq->lock);
7243 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7244 raw_spin_unlock(&rq->lock);
7245 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7249 complete(&req->done);
7251 __set_current_state(TASK_RUNNING);
7256 #ifdef CONFIG_HOTPLUG_CPU
7258 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7262 local_irq_disable();
7263 ret = __migrate_task(p, src_cpu, dest_cpu);
7269 * Figure out where task on dead CPU should go, use force if necessary.
7271 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7274 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7277 /* Look for allowed, online CPU in same node. */
7278 for_each_cpu_and(dest_cpu, nodemask, cpu_active_mask)
7279 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7282 /* Any allowed, online CPU? */
7283 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_active_mask);
7284 if (dest_cpu < nr_cpu_ids)
7287 /* No more Mr. Nice Guy. */
7288 if (dest_cpu >= nr_cpu_ids) {
7289 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7290 dest_cpu = cpumask_any_and(cpu_active_mask, &p->cpus_allowed);
7293 * Don't tell them about moving exiting tasks or
7294 * kernel threads (both mm NULL), since they never
7297 if (p->mm && printk_ratelimit()) {
7298 pr_info("process %d (%s) no longer affine to cpu%d\n",
7299 task_pid_nr(p), p->comm, dead_cpu);
7304 /* It can have affinity changed while we were choosing. */
7305 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7310 * While a dead CPU has no uninterruptible tasks queued at this point,
7311 * it might still have a nonzero ->nr_uninterruptible counter, because
7312 * for performance reasons the counter is not stricly tracking tasks to
7313 * their home CPUs. So we just add the counter to another CPU's counter,
7314 * to keep the global sum constant after CPU-down:
7316 static void migrate_nr_uninterruptible(struct rq *rq_src)
7318 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_active_mask));
7319 unsigned long flags;
7321 local_irq_save(flags);
7322 double_rq_lock(rq_src, rq_dest);
7323 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7324 rq_src->nr_uninterruptible = 0;
7325 double_rq_unlock(rq_src, rq_dest);
7326 local_irq_restore(flags);
7329 /* Run through task list and migrate tasks from the dead cpu. */
7330 static void migrate_live_tasks(int src_cpu)
7332 struct task_struct *p, *t;
7334 read_lock(&tasklist_lock);
7336 do_each_thread(t, p) {
7340 if (task_cpu(p) == src_cpu)
7341 move_task_off_dead_cpu(src_cpu, p);
7342 } while_each_thread(t, p);
7344 read_unlock(&tasklist_lock);
7348 * Schedules idle task to be the next runnable task on current CPU.
7349 * It does so by boosting its priority to highest possible.
7350 * Used by CPU offline code.
7352 void sched_idle_next(void)
7354 int this_cpu = smp_processor_id();
7355 struct rq *rq = cpu_rq(this_cpu);
7356 struct task_struct *p = rq->idle;
7357 unsigned long flags;
7359 /* cpu has to be offline */
7360 BUG_ON(cpu_online(this_cpu));
7363 * Strictly not necessary since rest of the CPUs are stopped by now
7364 * and interrupts disabled on the current cpu.
7366 raw_spin_lock_irqsave(&rq->lock, flags);
7368 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7370 update_rq_clock(rq);
7371 activate_task(rq, p, 0);
7373 raw_spin_unlock_irqrestore(&rq->lock, flags);
7377 * Ensures that the idle task is using init_mm right before its cpu goes
7380 void idle_task_exit(void)
7382 struct mm_struct *mm = current->active_mm;
7384 BUG_ON(cpu_online(smp_processor_id()));
7387 switch_mm(mm, &init_mm, current);
7391 /* called under rq->lock with disabled interrupts */
7392 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7394 struct rq *rq = cpu_rq(dead_cpu);
7396 /* Must be exiting, otherwise would be on tasklist. */
7397 BUG_ON(!p->exit_state);
7399 /* Cannot have done final schedule yet: would have vanished. */
7400 BUG_ON(p->state == TASK_DEAD);
7405 * Drop lock around migration; if someone else moves it,
7406 * that's OK. No task can be added to this CPU, so iteration is
7409 raw_spin_unlock_irq(&rq->lock);
7410 move_task_off_dead_cpu(dead_cpu, p);
7411 raw_spin_lock_irq(&rq->lock);
7416 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7417 static void migrate_dead_tasks(unsigned int dead_cpu)
7419 struct rq *rq = cpu_rq(dead_cpu);
7420 struct task_struct *next;
7423 if (!rq->nr_running)
7425 update_rq_clock(rq);
7426 next = pick_next_task(rq);
7429 next->sched_class->put_prev_task(rq, next);
7430 migrate_dead(dead_cpu, next);
7436 * remove the tasks which were accounted by rq from calc_load_tasks.
7438 static void calc_global_load_remove(struct rq *rq)
7440 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7441 rq->calc_load_active = 0;
7443 #endif /* CONFIG_HOTPLUG_CPU */
7445 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7447 static struct ctl_table sd_ctl_dir[] = {
7449 .procname = "sched_domain",
7455 static struct ctl_table sd_ctl_root[] = {
7457 .procname = "kernel",
7459 .child = sd_ctl_dir,
7464 static struct ctl_table *sd_alloc_ctl_entry(int n)
7466 struct ctl_table *entry =
7467 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7472 static void sd_free_ctl_entry(struct ctl_table **tablep)
7474 struct ctl_table *entry;
7477 * In the intermediate directories, both the child directory and
7478 * procname are dynamically allocated and could fail but the mode
7479 * will always be set. In the lowest directory the names are
7480 * static strings and all have proc handlers.
7482 for (entry = *tablep; entry->mode; entry++) {
7484 sd_free_ctl_entry(&entry->child);
7485 if (entry->proc_handler == NULL)
7486 kfree(entry->procname);
7494 set_table_entry(struct ctl_table *entry,
7495 const char *procname, void *data, int maxlen,
7496 mode_t mode, proc_handler *proc_handler)
7498 entry->procname = procname;
7500 entry->maxlen = maxlen;
7502 entry->proc_handler = proc_handler;
7505 static struct ctl_table *
7506 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7508 struct ctl_table *table = sd_alloc_ctl_entry(13);
7513 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7514 sizeof(long), 0644, proc_doulongvec_minmax);
7515 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7516 sizeof(long), 0644, proc_doulongvec_minmax);
7517 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7518 sizeof(int), 0644, proc_dointvec_minmax);
7519 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7520 sizeof(int), 0644, proc_dointvec_minmax);
7521 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7522 sizeof(int), 0644, proc_dointvec_minmax);
7523 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7524 sizeof(int), 0644, proc_dointvec_minmax);
7525 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7526 sizeof(int), 0644, proc_dointvec_minmax);
7527 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7528 sizeof(int), 0644, proc_dointvec_minmax);
7529 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7530 sizeof(int), 0644, proc_dointvec_minmax);
7531 set_table_entry(&table[9], "cache_nice_tries",
7532 &sd->cache_nice_tries,
7533 sizeof(int), 0644, proc_dointvec_minmax);
7534 set_table_entry(&table[10], "flags", &sd->flags,
7535 sizeof(int), 0644, proc_dointvec_minmax);
7536 set_table_entry(&table[11], "name", sd->name,
7537 CORENAME_MAX_SIZE, 0444, proc_dostring);
7538 /* &table[12] is terminator */
7543 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7545 struct ctl_table *entry, *table;
7546 struct sched_domain *sd;
7547 int domain_num = 0, i;
7550 for_each_domain(cpu, sd)
7552 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7557 for_each_domain(cpu, sd) {
7558 snprintf(buf, 32, "domain%d", i);
7559 entry->procname = kstrdup(buf, GFP_KERNEL);
7561 entry->child = sd_alloc_ctl_domain_table(sd);
7568 static struct ctl_table_header *sd_sysctl_header;
7569 static void register_sched_domain_sysctl(void)
7571 int i, cpu_num = num_possible_cpus();
7572 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7575 WARN_ON(sd_ctl_dir[0].child);
7576 sd_ctl_dir[0].child = entry;
7581 for_each_possible_cpu(i) {
7582 snprintf(buf, 32, "cpu%d", i);
7583 entry->procname = kstrdup(buf, GFP_KERNEL);
7585 entry->child = sd_alloc_ctl_cpu_table(i);
7589 WARN_ON(sd_sysctl_header);
7590 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7593 /* may be called multiple times per register */
7594 static void unregister_sched_domain_sysctl(void)
7596 if (sd_sysctl_header)
7597 unregister_sysctl_table(sd_sysctl_header);
7598 sd_sysctl_header = NULL;
7599 if (sd_ctl_dir[0].child)
7600 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7603 static void register_sched_domain_sysctl(void)
7606 static void unregister_sched_domain_sysctl(void)
7611 static void set_rq_online(struct rq *rq)
7614 const struct sched_class *class;
7616 cpumask_set_cpu(rq->cpu, rq->rd->online);
7619 for_each_class(class) {
7620 if (class->rq_online)
7621 class->rq_online(rq);
7626 static void set_rq_offline(struct rq *rq)
7629 const struct sched_class *class;
7631 for_each_class(class) {
7632 if (class->rq_offline)
7633 class->rq_offline(rq);
7636 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7642 * migration_call - callback that gets triggered when a CPU is added.
7643 * Here we can start up the necessary migration thread for the new CPU.
7645 static int __cpuinit
7646 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7648 struct task_struct *p;
7649 int cpu = (long)hcpu;
7650 unsigned long flags;
7655 case CPU_UP_PREPARE:
7656 case CPU_UP_PREPARE_FROZEN:
7657 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7660 kthread_bind(p, cpu);
7661 /* Must be high prio: stop_machine expects to yield to it. */
7662 rq = task_rq_lock(p, &flags);
7663 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7664 task_rq_unlock(rq, &flags);
7666 cpu_rq(cpu)->migration_thread = p;
7667 rq->calc_load_update = calc_load_update;
7671 case CPU_ONLINE_FROZEN:
7672 /* Strictly unnecessary, as first user will wake it. */
7673 wake_up_process(cpu_rq(cpu)->migration_thread);
7675 /* Update our root-domain */
7677 raw_spin_lock_irqsave(&rq->lock, flags);
7679 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7683 raw_spin_unlock_irqrestore(&rq->lock, flags);
7686 #ifdef CONFIG_HOTPLUG_CPU
7687 case CPU_UP_CANCELED:
7688 case CPU_UP_CANCELED_FROZEN:
7689 if (!cpu_rq(cpu)->migration_thread)
7691 /* Unbind it from offline cpu so it can run. Fall thru. */
7692 kthread_bind(cpu_rq(cpu)->migration_thread,
7693 cpumask_any(cpu_online_mask));
7694 kthread_stop(cpu_rq(cpu)->migration_thread);
7695 put_task_struct(cpu_rq(cpu)->migration_thread);
7696 cpu_rq(cpu)->migration_thread = NULL;
7700 case CPU_DEAD_FROZEN:
7701 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7702 migrate_live_tasks(cpu);
7704 kthread_stop(rq->migration_thread);
7705 put_task_struct(rq->migration_thread);
7706 rq->migration_thread = NULL;
7707 /* Idle task back to normal (off runqueue, low prio) */
7708 raw_spin_lock_irq(&rq->lock);
7709 update_rq_clock(rq);
7710 deactivate_task(rq, rq->idle, 0);
7711 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7712 rq->idle->sched_class = &idle_sched_class;
7713 migrate_dead_tasks(cpu);
7714 raw_spin_unlock_irq(&rq->lock);
7716 migrate_nr_uninterruptible(rq);
7717 BUG_ON(rq->nr_running != 0);
7718 calc_global_load_remove(rq);
7720 * No need to migrate the tasks: it was best-effort if
7721 * they didn't take sched_hotcpu_mutex. Just wake up
7724 raw_spin_lock_irq(&rq->lock);
7725 while (!list_empty(&rq->migration_queue)) {
7726 struct migration_req *req;
7728 req = list_entry(rq->migration_queue.next,
7729 struct migration_req, list);
7730 list_del_init(&req->list);
7731 raw_spin_unlock_irq(&rq->lock);
7732 complete(&req->done);
7733 raw_spin_lock_irq(&rq->lock);
7735 raw_spin_unlock_irq(&rq->lock);
7739 case CPU_DYING_FROZEN:
7740 /* Update our root-domain */
7742 raw_spin_lock_irqsave(&rq->lock, flags);
7744 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7747 raw_spin_unlock_irqrestore(&rq->lock, flags);
7755 * Register at high priority so that task migration (migrate_all_tasks)
7756 * happens before everything else. This has to be lower priority than
7757 * the notifier in the perf_event subsystem, though.
7759 static struct notifier_block __cpuinitdata migration_notifier = {
7760 .notifier_call = migration_call,
7764 static int __init migration_init(void)
7766 void *cpu = (void *)(long)smp_processor_id();
7769 /* Start one for the boot CPU: */
7770 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7771 BUG_ON(err == NOTIFY_BAD);
7772 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7773 register_cpu_notifier(&migration_notifier);
7777 early_initcall(migration_init);
7782 #ifdef CONFIG_SCHED_DEBUG
7784 static __read_mostly int sched_domain_debug_enabled;
7786 static int __init sched_domain_debug_setup(char *str)
7788 sched_domain_debug_enabled = 1;
7792 early_param("sched_debug", sched_domain_debug_setup);
7794 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7795 struct cpumask *groupmask)
7797 struct sched_group *group = sd->groups;
7800 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7801 cpumask_clear(groupmask);
7803 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7805 if (!(sd->flags & SD_LOAD_BALANCE)) {
7806 pr_cont("does not load-balance\n");
7808 pr_err("ERROR: !SD_LOAD_BALANCE domain has parent\n");
7812 pr_cont("span %s level %s\n", str, sd->name);
7814 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7815 pr_err("ERROR: domain->span does not contain CPU%d\n", cpu);
7817 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7818 pr_err("ERROR: domain->groups does not contain CPU%d\n", cpu);
7821 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7825 pr_err("ERROR: group is NULL\n");
7829 if (!group->cpu_power) {
7831 pr_err("ERROR: domain->cpu_power not set\n");
7835 if (!cpumask_weight(sched_group_cpus(group))) {
7837 pr_err("ERROR: empty group\n");
7841 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7843 pr_err("ERROR: repeated CPUs\n");
7847 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7849 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7851 pr_cont(" %s", str);
7852 if (group->cpu_power != SCHED_LOAD_SCALE) {
7853 pr_cont(" (cpu_power = %d)", group->cpu_power);
7856 group = group->next;
7857 } while (group != sd->groups);
7860 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7861 pr_err("ERROR: groups don't span domain->span\n");
7864 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7865 pr_err("ERROR: parent span is not a superset of domain->span\n");
7869 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7871 cpumask_var_t groupmask;
7874 if (!sched_domain_debug_enabled)
7878 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7882 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7884 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7885 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7890 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7897 free_cpumask_var(groupmask);
7899 #else /* !CONFIG_SCHED_DEBUG */
7900 # define sched_domain_debug(sd, cpu) do { } while (0)
7901 #endif /* CONFIG_SCHED_DEBUG */
7903 static int sd_degenerate(struct sched_domain *sd)
7905 if (cpumask_weight(sched_domain_span(sd)) == 1)
7908 /* Following flags need at least 2 groups */
7909 if (sd->flags & (SD_LOAD_BALANCE |
7910 SD_BALANCE_NEWIDLE |
7914 SD_SHARE_PKG_RESOURCES)) {
7915 if (sd->groups != sd->groups->next)
7919 /* Following flags don't use groups */
7920 if (sd->flags & (SD_WAKE_AFFINE))
7927 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7929 unsigned long cflags = sd->flags, pflags = parent->flags;
7931 if (sd_degenerate(parent))
7934 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7937 /* Flags needing groups don't count if only 1 group in parent */
7938 if (parent->groups == parent->groups->next) {
7939 pflags &= ~(SD_LOAD_BALANCE |
7940 SD_BALANCE_NEWIDLE |
7944 SD_SHARE_PKG_RESOURCES);
7945 if (nr_node_ids == 1)
7946 pflags &= ~SD_SERIALIZE;
7948 if (~cflags & pflags)
7954 static void free_rootdomain(struct root_domain *rd)
7956 synchronize_sched();
7958 cpupri_cleanup(&rd->cpupri);
7960 free_cpumask_var(rd->rto_mask);
7961 free_cpumask_var(rd->online);
7962 free_cpumask_var(rd->span);
7966 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7968 struct root_domain *old_rd = NULL;
7969 unsigned long flags;
7971 raw_spin_lock_irqsave(&rq->lock, flags);
7976 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7979 cpumask_clear_cpu(rq->cpu, old_rd->span);
7982 * If we dont want to free the old_rt yet then
7983 * set old_rd to NULL to skip the freeing later
7986 if (!atomic_dec_and_test(&old_rd->refcount))
7990 atomic_inc(&rd->refcount);
7993 cpumask_set_cpu(rq->cpu, rd->span);
7994 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7997 raw_spin_unlock_irqrestore(&rq->lock, flags);
8000 free_rootdomain(old_rd);
8003 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8005 gfp_t gfp = GFP_KERNEL;
8007 memset(rd, 0, sizeof(*rd));
8012 if (!alloc_cpumask_var(&rd->span, gfp))
8014 if (!alloc_cpumask_var(&rd->online, gfp))
8016 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8019 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8024 free_cpumask_var(rd->rto_mask);
8026 free_cpumask_var(rd->online);
8028 free_cpumask_var(rd->span);
8033 static void init_defrootdomain(void)
8035 init_rootdomain(&def_root_domain, true);
8037 atomic_set(&def_root_domain.refcount, 1);
8040 static struct root_domain *alloc_rootdomain(void)
8042 struct root_domain *rd;
8044 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8048 if (init_rootdomain(rd, false) != 0) {
8057 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8058 * hold the hotplug lock.
8061 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8063 struct rq *rq = cpu_rq(cpu);
8064 struct sched_domain *tmp;
8066 /* Remove the sched domains which do not contribute to scheduling. */
8067 for (tmp = sd; tmp; ) {
8068 struct sched_domain *parent = tmp->parent;
8072 if (sd_parent_degenerate(tmp, parent)) {
8073 tmp->parent = parent->parent;
8075 parent->parent->child = tmp;
8080 if (sd && sd_degenerate(sd)) {
8086 sched_domain_debug(sd, cpu);
8088 rq_attach_root(rq, rd);
8089 rcu_assign_pointer(rq->sd, sd);
8092 /* cpus with isolated domains */
8093 static cpumask_var_t cpu_isolated_map;
8095 /* Setup the mask of cpus configured for isolated domains */
8096 static int __init isolated_cpu_setup(char *str)
8098 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8099 cpulist_parse(str, cpu_isolated_map);
8103 __setup("isolcpus=", isolated_cpu_setup);
8106 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8107 * to a function which identifies what group(along with sched group) a CPU
8108 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8109 * (due to the fact that we keep track of groups covered with a struct cpumask).
8111 * init_sched_build_groups will build a circular linked list of the groups
8112 * covered by the given span, and will set each group's ->cpumask correctly,
8113 * and ->cpu_power to 0.
8116 init_sched_build_groups(const struct cpumask *span,
8117 const struct cpumask *cpu_map,
8118 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8119 struct sched_group **sg,
8120 struct cpumask *tmpmask),
8121 struct cpumask *covered, struct cpumask *tmpmask)
8123 struct sched_group *first = NULL, *last = NULL;
8126 cpumask_clear(covered);
8128 for_each_cpu(i, span) {
8129 struct sched_group *sg;
8130 int group = group_fn(i, cpu_map, &sg, tmpmask);
8133 if (cpumask_test_cpu(i, covered))
8136 cpumask_clear(sched_group_cpus(sg));
8139 for_each_cpu(j, span) {
8140 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8143 cpumask_set_cpu(j, covered);
8144 cpumask_set_cpu(j, sched_group_cpus(sg));
8155 #define SD_NODES_PER_DOMAIN 16
8160 * find_next_best_node - find the next node to include in a sched_domain
8161 * @node: node whose sched_domain we're building
8162 * @used_nodes: nodes already in the sched_domain
8164 * Find the next node to include in a given scheduling domain. Simply
8165 * finds the closest node not already in the @used_nodes map.
8167 * Should use nodemask_t.
8169 static int find_next_best_node(int node, nodemask_t *used_nodes)
8171 int i, n, val, min_val, best_node = 0;
8175 for (i = 0; i < nr_node_ids; i++) {
8176 /* Start at @node */
8177 n = (node + i) % nr_node_ids;
8179 if (!nr_cpus_node(n))
8182 /* Skip already used nodes */
8183 if (node_isset(n, *used_nodes))
8186 /* Simple min distance search */
8187 val = node_distance(node, n);
8189 if (val < min_val) {
8195 node_set(best_node, *used_nodes);
8200 * sched_domain_node_span - get a cpumask for a node's sched_domain
8201 * @node: node whose cpumask we're constructing
8202 * @span: resulting cpumask
8204 * Given a node, construct a good cpumask for its sched_domain to span. It
8205 * should be one that prevents unnecessary balancing, but also spreads tasks
8208 static void sched_domain_node_span(int node, struct cpumask *span)
8210 nodemask_t used_nodes;
8213 cpumask_clear(span);
8214 nodes_clear(used_nodes);
8216 cpumask_or(span, span, cpumask_of_node(node));
8217 node_set(node, used_nodes);
8219 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8220 int next_node = find_next_best_node(node, &used_nodes);
8222 cpumask_or(span, span, cpumask_of_node(next_node));
8225 #endif /* CONFIG_NUMA */
8227 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8230 * The cpus mask in sched_group and sched_domain hangs off the end.
8232 * ( See the the comments in include/linux/sched.h:struct sched_group
8233 * and struct sched_domain. )
8235 struct static_sched_group {
8236 struct sched_group sg;
8237 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8240 struct static_sched_domain {
8241 struct sched_domain sd;
8242 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8248 cpumask_var_t domainspan;
8249 cpumask_var_t covered;
8250 cpumask_var_t notcovered;
8252 cpumask_var_t nodemask;
8253 cpumask_var_t this_sibling_map;
8254 cpumask_var_t this_core_map;
8255 cpumask_var_t send_covered;
8256 cpumask_var_t tmpmask;
8257 struct sched_group **sched_group_nodes;
8258 struct root_domain *rd;
8262 sa_sched_groups = 0,
8267 sa_this_sibling_map,
8269 sa_sched_group_nodes,
8279 * SMT sched-domains:
8281 #ifdef CONFIG_SCHED_SMT
8282 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8283 static DEFINE_PER_CPU(struct static_sched_group, sched_groups);
8286 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8287 struct sched_group **sg, struct cpumask *unused)
8290 *sg = &per_cpu(sched_groups, cpu).sg;
8293 #endif /* CONFIG_SCHED_SMT */
8296 * multi-core sched-domains:
8298 #ifdef CONFIG_SCHED_MC
8299 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8300 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8301 #endif /* CONFIG_SCHED_MC */
8303 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8305 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8306 struct sched_group **sg, struct cpumask *mask)
8310 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8311 group = cpumask_first(mask);
8313 *sg = &per_cpu(sched_group_core, group).sg;
8316 #elif defined(CONFIG_SCHED_MC)
8318 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8319 struct sched_group **sg, struct cpumask *unused)
8322 *sg = &per_cpu(sched_group_core, cpu).sg;
8327 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8328 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8331 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8332 struct sched_group **sg, struct cpumask *mask)
8335 #ifdef CONFIG_SCHED_MC
8336 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8337 group = cpumask_first(mask);
8338 #elif defined(CONFIG_SCHED_SMT)
8339 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8340 group = cpumask_first(mask);
8345 *sg = &per_cpu(sched_group_phys, group).sg;
8351 * The init_sched_build_groups can't handle what we want to do with node
8352 * groups, so roll our own. Now each node has its own list of groups which
8353 * gets dynamically allocated.
8355 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8356 static struct sched_group ***sched_group_nodes_bycpu;
8358 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8359 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8361 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8362 struct sched_group **sg,
8363 struct cpumask *nodemask)
8367 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8368 group = cpumask_first(nodemask);
8371 *sg = &per_cpu(sched_group_allnodes, group).sg;
8375 static void init_numa_sched_groups_power(struct sched_group *group_head)
8377 struct sched_group *sg = group_head;
8383 for_each_cpu(j, sched_group_cpus(sg)) {
8384 struct sched_domain *sd;
8386 sd = &per_cpu(phys_domains, j).sd;
8387 if (j != group_first_cpu(sd->groups)) {
8389 * Only add "power" once for each
8395 sg->cpu_power += sd->groups->cpu_power;
8398 } while (sg != group_head);
8401 static int build_numa_sched_groups(struct s_data *d,
8402 const struct cpumask *cpu_map, int num)
8404 struct sched_domain *sd;
8405 struct sched_group *sg, *prev;
8408 cpumask_clear(d->covered);
8409 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8410 if (cpumask_empty(d->nodemask)) {
8411 d->sched_group_nodes[num] = NULL;
8415 sched_domain_node_span(num, d->domainspan);
8416 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8418 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8421 pr_warning("Can not alloc domain group for node %d\n", num);
8424 d->sched_group_nodes[num] = sg;
8426 for_each_cpu(j, d->nodemask) {
8427 sd = &per_cpu(node_domains, j).sd;
8432 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8434 cpumask_or(d->covered, d->covered, d->nodemask);
8437 for (j = 0; j < nr_node_ids; j++) {
8438 n = (num + j) % nr_node_ids;
8439 cpumask_complement(d->notcovered, d->covered);
8440 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8441 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8442 if (cpumask_empty(d->tmpmask))
8444 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8445 if (cpumask_empty(d->tmpmask))
8447 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8450 pr_warning("Can not alloc domain group for node %d\n",
8455 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8456 sg->next = prev->next;
8457 cpumask_or(d->covered, d->covered, d->tmpmask);
8464 #endif /* CONFIG_NUMA */
8467 /* Free memory allocated for various sched_group structures */
8468 static void free_sched_groups(const struct cpumask *cpu_map,
8469 struct cpumask *nodemask)
8473 for_each_cpu(cpu, cpu_map) {
8474 struct sched_group **sched_group_nodes
8475 = sched_group_nodes_bycpu[cpu];
8477 if (!sched_group_nodes)
8480 for (i = 0; i < nr_node_ids; i++) {
8481 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8483 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8484 if (cpumask_empty(nodemask))
8494 if (oldsg != sched_group_nodes[i])
8497 kfree(sched_group_nodes);
8498 sched_group_nodes_bycpu[cpu] = NULL;
8501 #else /* !CONFIG_NUMA */
8502 static void free_sched_groups(const struct cpumask *cpu_map,
8503 struct cpumask *nodemask)
8506 #endif /* CONFIG_NUMA */
8509 * Initialize sched groups cpu_power.
8511 * cpu_power indicates the capacity of sched group, which is used while
8512 * distributing the load between different sched groups in a sched domain.
8513 * Typically cpu_power for all the groups in a sched domain will be same unless
8514 * there are asymmetries in the topology. If there are asymmetries, group
8515 * having more cpu_power will pickup more load compared to the group having
8518 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8520 struct sched_domain *child;
8521 struct sched_group *group;
8525 WARN_ON(!sd || !sd->groups);
8527 if (cpu != group_first_cpu(sd->groups))
8532 sd->groups->cpu_power = 0;
8535 power = SCHED_LOAD_SCALE;
8536 weight = cpumask_weight(sched_domain_span(sd));
8538 * SMT siblings share the power of a single core.
8539 * Usually multiple threads get a better yield out of
8540 * that one core than a single thread would have,
8541 * reflect that in sd->smt_gain.
8543 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8544 power *= sd->smt_gain;
8546 power >>= SCHED_LOAD_SHIFT;
8548 sd->groups->cpu_power += power;
8553 * Add cpu_power of each child group to this groups cpu_power.
8555 group = child->groups;
8557 sd->groups->cpu_power += group->cpu_power;
8558 group = group->next;
8559 } while (group != child->groups);
8563 * Initializers for schedule domains
8564 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8567 #ifdef CONFIG_SCHED_DEBUG
8568 # define SD_INIT_NAME(sd, type) sd->name = #type
8570 # define SD_INIT_NAME(sd, type) do { } while (0)
8573 #define SD_INIT(sd, type) sd_init_##type(sd)
8575 #define SD_INIT_FUNC(type) \
8576 static noinline void sd_init_##type(struct sched_domain *sd) \
8578 memset(sd, 0, sizeof(*sd)); \
8579 *sd = SD_##type##_INIT; \
8580 sd->level = SD_LV_##type; \
8581 SD_INIT_NAME(sd, type); \
8586 SD_INIT_FUNC(ALLNODES)
8589 #ifdef CONFIG_SCHED_SMT
8590 SD_INIT_FUNC(SIBLING)
8592 #ifdef CONFIG_SCHED_MC
8596 static int default_relax_domain_level = -1;
8598 static int __init setup_relax_domain_level(char *str)
8602 val = simple_strtoul(str, NULL, 0);
8603 if (val < SD_LV_MAX)
8604 default_relax_domain_level = val;
8608 __setup("relax_domain_level=", setup_relax_domain_level);
8610 static void set_domain_attribute(struct sched_domain *sd,
8611 struct sched_domain_attr *attr)
8615 if (!attr || attr->relax_domain_level < 0) {
8616 if (default_relax_domain_level < 0)
8619 request = default_relax_domain_level;
8621 request = attr->relax_domain_level;
8622 if (request < sd->level) {
8623 /* turn off idle balance on this domain */
8624 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8626 /* turn on idle balance on this domain */
8627 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
8631 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8632 const struct cpumask *cpu_map)
8635 case sa_sched_groups:
8636 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8637 d->sched_group_nodes = NULL;
8639 free_rootdomain(d->rd); /* fall through */
8641 free_cpumask_var(d->tmpmask); /* fall through */
8642 case sa_send_covered:
8643 free_cpumask_var(d->send_covered); /* fall through */
8644 case sa_this_core_map:
8645 free_cpumask_var(d->this_core_map); /* fall through */
8646 case sa_this_sibling_map:
8647 free_cpumask_var(d->this_sibling_map); /* fall through */
8649 free_cpumask_var(d->nodemask); /* fall through */
8650 case sa_sched_group_nodes:
8652 kfree(d->sched_group_nodes); /* fall through */
8654 free_cpumask_var(d->notcovered); /* fall through */
8656 free_cpumask_var(d->covered); /* fall through */
8658 free_cpumask_var(d->domainspan); /* fall through */
8665 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8666 const struct cpumask *cpu_map)
8669 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8671 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8672 return sa_domainspan;
8673 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8675 /* Allocate the per-node list of sched groups */
8676 d->sched_group_nodes = kcalloc(nr_node_ids,
8677 sizeof(struct sched_group *), GFP_KERNEL);
8678 if (!d->sched_group_nodes) {
8679 pr_warning("Can not alloc sched group node list\n");
8680 return sa_notcovered;
8682 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8684 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8685 return sa_sched_group_nodes;
8686 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8688 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8689 return sa_this_sibling_map;
8690 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8691 return sa_this_core_map;
8692 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8693 return sa_send_covered;
8694 d->rd = alloc_rootdomain();
8696 pr_warning("Cannot alloc root domain\n");
8699 return sa_rootdomain;
8702 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8703 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8705 struct sched_domain *sd = NULL;
8707 struct sched_domain *parent;
8710 if (cpumask_weight(cpu_map) >
8711 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8712 sd = &per_cpu(allnodes_domains, i).sd;
8713 SD_INIT(sd, ALLNODES);
8714 set_domain_attribute(sd, attr);
8715 cpumask_copy(sched_domain_span(sd), cpu_map);
8716 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8721 sd = &per_cpu(node_domains, i).sd;
8723 set_domain_attribute(sd, attr);
8724 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8725 sd->parent = parent;
8728 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8733 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8734 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8735 struct sched_domain *parent, int i)
8737 struct sched_domain *sd;
8738 sd = &per_cpu(phys_domains, i).sd;
8740 set_domain_attribute(sd, attr);
8741 cpumask_copy(sched_domain_span(sd), d->nodemask);
8742 sd->parent = parent;
8745 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8749 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8750 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8751 struct sched_domain *parent, int i)
8753 struct sched_domain *sd = parent;
8754 #ifdef CONFIG_SCHED_MC
8755 sd = &per_cpu(core_domains, i).sd;
8757 set_domain_attribute(sd, attr);
8758 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8759 sd->parent = parent;
8761 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8766 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8767 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8768 struct sched_domain *parent, int i)
8770 struct sched_domain *sd = parent;
8771 #ifdef CONFIG_SCHED_SMT
8772 sd = &per_cpu(cpu_domains, i).sd;
8773 SD_INIT(sd, SIBLING);
8774 set_domain_attribute(sd, attr);
8775 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8776 sd->parent = parent;
8778 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8783 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8784 const struct cpumask *cpu_map, int cpu)
8787 #ifdef CONFIG_SCHED_SMT
8788 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8789 cpumask_and(d->this_sibling_map, cpu_map,
8790 topology_thread_cpumask(cpu));
8791 if (cpu == cpumask_first(d->this_sibling_map))
8792 init_sched_build_groups(d->this_sibling_map, cpu_map,
8794 d->send_covered, d->tmpmask);
8797 #ifdef CONFIG_SCHED_MC
8798 case SD_LV_MC: /* set up multi-core groups */
8799 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8800 if (cpu == cpumask_first(d->this_core_map))
8801 init_sched_build_groups(d->this_core_map, cpu_map,
8803 d->send_covered, d->tmpmask);
8806 case SD_LV_CPU: /* set up physical groups */
8807 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8808 if (!cpumask_empty(d->nodemask))
8809 init_sched_build_groups(d->nodemask, cpu_map,
8811 d->send_covered, d->tmpmask);
8814 case SD_LV_ALLNODES:
8815 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8816 d->send_covered, d->tmpmask);
8825 * Build sched domains for a given set of cpus and attach the sched domains
8826 * to the individual cpus
8828 static int __build_sched_domains(const struct cpumask *cpu_map,
8829 struct sched_domain_attr *attr)
8831 enum s_alloc alloc_state = sa_none;
8833 struct sched_domain *sd;
8839 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8840 if (alloc_state != sa_rootdomain)
8842 alloc_state = sa_sched_groups;
8845 * Set up domains for cpus specified by the cpu_map.
8847 for_each_cpu(i, cpu_map) {
8848 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8851 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8852 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8853 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8854 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8857 for_each_cpu(i, cpu_map) {
8858 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8859 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8862 /* Set up physical groups */
8863 for (i = 0; i < nr_node_ids; i++)
8864 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8867 /* Set up node groups */
8869 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8871 for (i = 0; i < nr_node_ids; i++)
8872 if (build_numa_sched_groups(&d, cpu_map, i))
8876 /* Calculate CPU power for physical packages and nodes */
8877 #ifdef CONFIG_SCHED_SMT
8878 for_each_cpu(i, cpu_map) {
8879 sd = &per_cpu(cpu_domains, i).sd;
8880 init_sched_groups_power(i, sd);
8883 #ifdef CONFIG_SCHED_MC
8884 for_each_cpu(i, cpu_map) {
8885 sd = &per_cpu(core_domains, i).sd;
8886 init_sched_groups_power(i, sd);
8890 for_each_cpu(i, cpu_map) {
8891 sd = &per_cpu(phys_domains, i).sd;
8892 init_sched_groups_power(i, sd);
8896 for (i = 0; i < nr_node_ids; i++)
8897 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8899 if (d.sd_allnodes) {
8900 struct sched_group *sg;
8902 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8904 init_numa_sched_groups_power(sg);
8908 /* Attach the domains */
8909 for_each_cpu(i, cpu_map) {
8910 #ifdef CONFIG_SCHED_SMT
8911 sd = &per_cpu(cpu_domains, i).sd;
8912 #elif defined(CONFIG_SCHED_MC)
8913 sd = &per_cpu(core_domains, i).sd;
8915 sd = &per_cpu(phys_domains, i).sd;
8917 cpu_attach_domain(sd, d.rd, i);
8920 d.sched_group_nodes = NULL; /* don't free this we still need it */
8921 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8925 __free_domain_allocs(&d, alloc_state, cpu_map);
8929 static int build_sched_domains(const struct cpumask *cpu_map)
8931 return __build_sched_domains(cpu_map, NULL);
8934 static cpumask_var_t *doms_cur; /* current sched domains */
8935 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8936 static struct sched_domain_attr *dattr_cur;
8937 /* attribues of custom domains in 'doms_cur' */
8940 * Special case: If a kmalloc of a doms_cur partition (array of
8941 * cpumask) fails, then fallback to a single sched domain,
8942 * as determined by the single cpumask fallback_doms.
8944 static cpumask_var_t fallback_doms;
8947 * arch_update_cpu_topology lets virtualized architectures update the
8948 * cpu core maps. It is supposed to return 1 if the topology changed
8949 * or 0 if it stayed the same.
8951 int __attribute__((weak)) arch_update_cpu_topology(void)
8956 cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8959 cpumask_var_t *doms;
8961 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8964 for (i = 0; i < ndoms; i++) {
8965 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
8966 free_sched_domains(doms, i);
8973 void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
8976 for (i = 0; i < ndoms; i++)
8977 free_cpumask_var(doms[i]);
8982 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8983 * For now this just excludes isolated cpus, but could be used to
8984 * exclude other special cases in the future.
8986 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8990 arch_update_cpu_topology();
8992 doms_cur = alloc_sched_domains(ndoms_cur);
8994 doms_cur = &fallback_doms;
8995 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
8997 err = build_sched_domains(doms_cur[0]);
8998 register_sched_domain_sysctl();
9003 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9004 struct cpumask *tmpmask)
9006 free_sched_groups(cpu_map, tmpmask);
9010 * Detach sched domains from a group of cpus specified in cpu_map
9011 * These cpus will now be attached to the NULL domain
9013 static void detach_destroy_domains(const struct cpumask *cpu_map)
9015 /* Save because hotplug lock held. */
9016 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9019 for_each_cpu(i, cpu_map)
9020 cpu_attach_domain(NULL, &def_root_domain, i);
9021 synchronize_sched();
9022 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9025 /* handle null as "default" */
9026 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9027 struct sched_domain_attr *new, int idx_new)
9029 struct sched_domain_attr tmp;
9036 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9037 new ? (new + idx_new) : &tmp,
9038 sizeof(struct sched_domain_attr));
9042 * Partition sched domains as specified by the 'ndoms_new'
9043 * cpumasks in the array doms_new[] of cpumasks. This compares
9044 * doms_new[] to the current sched domain partitioning, doms_cur[].
9045 * It destroys each deleted domain and builds each new domain.
9047 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9048 * The masks don't intersect (don't overlap.) We should setup one
9049 * sched domain for each mask. CPUs not in any of the cpumasks will
9050 * not be load balanced. If the same cpumask appears both in the
9051 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9054 * The passed in 'doms_new' should be allocated using
9055 * alloc_sched_domains. This routine takes ownership of it and will
9056 * free_sched_domains it when done with it. If the caller failed the
9057 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9058 * and partition_sched_domains() will fallback to the single partition
9059 * 'fallback_doms', it also forces the domains to be rebuilt.
9061 * If doms_new == NULL it will be replaced with cpu_online_mask.
9062 * ndoms_new == 0 is a special case for destroying existing domains,
9063 * and it will not create the default domain.
9065 * Call with hotplug lock held
9067 void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
9068 struct sched_domain_attr *dattr_new)
9073 mutex_lock(&sched_domains_mutex);
9075 /* always unregister in case we don't destroy any domains */
9076 unregister_sched_domain_sysctl();
9078 /* Let architecture update cpu core mappings. */
9079 new_topology = arch_update_cpu_topology();
9081 n = doms_new ? ndoms_new : 0;
9083 /* Destroy deleted domains */
9084 for (i = 0; i < ndoms_cur; i++) {
9085 for (j = 0; j < n && !new_topology; j++) {
9086 if (cpumask_equal(doms_cur[i], doms_new[j])
9087 && dattrs_equal(dattr_cur, i, dattr_new, j))
9090 /* no match - a current sched domain not in new doms_new[] */
9091 detach_destroy_domains(doms_cur[i]);
9096 if (doms_new == NULL) {
9098 doms_new = &fallback_doms;
9099 cpumask_andnot(doms_new[0], cpu_active_mask, cpu_isolated_map);
9100 WARN_ON_ONCE(dattr_new);
9103 /* Build new domains */
9104 for (i = 0; i < ndoms_new; i++) {
9105 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9106 if (cpumask_equal(doms_new[i], doms_cur[j])
9107 && dattrs_equal(dattr_new, i, dattr_cur, j))
9110 /* no match - add a new doms_new */
9111 __build_sched_domains(doms_new[i],
9112 dattr_new ? dattr_new + i : NULL);
9117 /* Remember the new sched domains */
9118 if (doms_cur != &fallback_doms)
9119 free_sched_domains(doms_cur, ndoms_cur);
9120 kfree(dattr_cur); /* kfree(NULL) is safe */
9121 doms_cur = doms_new;
9122 dattr_cur = dattr_new;
9123 ndoms_cur = ndoms_new;
9125 register_sched_domain_sysctl();
9127 mutex_unlock(&sched_domains_mutex);
9130 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9131 static void arch_reinit_sched_domains(void)
9135 /* Destroy domains first to force the rebuild */
9136 partition_sched_domains(0, NULL, NULL);
9138 rebuild_sched_domains();
9142 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9144 unsigned int level = 0;
9146 if (sscanf(buf, "%u", &level) != 1)
9150 * level is always be positive so don't check for
9151 * level < POWERSAVINGS_BALANCE_NONE which is 0
9152 * What happens on 0 or 1 byte write,
9153 * need to check for count as well?
9156 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9160 sched_smt_power_savings = level;
9162 sched_mc_power_savings = level;
9164 arch_reinit_sched_domains();
9169 #ifdef CONFIG_SCHED_MC
9170 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9173 return sprintf(page, "%u\n", sched_mc_power_savings);
9175 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9176 const char *buf, size_t count)
9178 return sched_power_savings_store(buf, count, 0);
9180 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9181 sched_mc_power_savings_show,
9182 sched_mc_power_savings_store);
9185 #ifdef CONFIG_SCHED_SMT
9186 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9189 return sprintf(page, "%u\n", sched_smt_power_savings);
9191 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9192 const char *buf, size_t count)
9194 return sched_power_savings_store(buf, count, 1);
9196 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9197 sched_smt_power_savings_show,
9198 sched_smt_power_savings_store);
9201 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9205 #ifdef CONFIG_SCHED_SMT
9207 err = sysfs_create_file(&cls->kset.kobj,
9208 &attr_sched_smt_power_savings.attr);
9210 #ifdef CONFIG_SCHED_MC
9211 if (!err && mc_capable())
9212 err = sysfs_create_file(&cls->kset.kobj,
9213 &attr_sched_mc_power_savings.attr);
9217 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9219 #ifndef CONFIG_CPUSETS
9221 * Add online and remove offline CPUs from the scheduler domains.
9222 * When cpusets are enabled they take over this function.
9224 static int update_sched_domains(struct notifier_block *nfb,
9225 unsigned long action, void *hcpu)
9229 case CPU_ONLINE_FROZEN:
9230 case CPU_DOWN_PREPARE:
9231 case CPU_DOWN_PREPARE_FROZEN:
9232 case CPU_DOWN_FAILED:
9233 case CPU_DOWN_FAILED_FROZEN:
9234 partition_sched_domains(1, NULL, NULL);
9243 static int update_runtime(struct notifier_block *nfb,
9244 unsigned long action, void *hcpu)
9246 int cpu = (int)(long)hcpu;
9249 case CPU_DOWN_PREPARE:
9250 case CPU_DOWN_PREPARE_FROZEN:
9251 disable_runtime(cpu_rq(cpu));
9254 case CPU_DOWN_FAILED:
9255 case CPU_DOWN_FAILED_FROZEN:
9257 case CPU_ONLINE_FROZEN:
9258 enable_runtime(cpu_rq(cpu));
9266 void __init sched_init_smp(void)
9268 cpumask_var_t non_isolated_cpus;
9270 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9271 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9273 #if defined(CONFIG_NUMA)
9274 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9276 BUG_ON(sched_group_nodes_bycpu == NULL);
9279 mutex_lock(&sched_domains_mutex);
9280 arch_init_sched_domains(cpu_active_mask);
9281 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9282 if (cpumask_empty(non_isolated_cpus))
9283 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9284 mutex_unlock(&sched_domains_mutex);
9287 #ifndef CONFIG_CPUSETS
9288 /* XXX: Theoretical race here - CPU may be hotplugged now */
9289 hotcpu_notifier(update_sched_domains, 0);
9292 /* RT runtime code needs to handle some hotplug events */
9293 hotcpu_notifier(update_runtime, 0);
9297 /* Move init over to a non-isolated CPU */
9298 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9300 sched_init_granularity();
9301 free_cpumask_var(non_isolated_cpus);
9303 init_sched_rt_class();
9306 void __init sched_init_smp(void)
9308 sched_init_granularity();
9310 #endif /* CONFIG_SMP */
9312 const_debug unsigned int sysctl_timer_migration = 1;
9314 int in_sched_functions(unsigned long addr)
9316 return in_lock_functions(addr) ||
9317 (addr >= (unsigned long)__sched_text_start
9318 && addr < (unsigned long)__sched_text_end);
9321 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9323 cfs_rq->tasks_timeline = RB_ROOT;
9324 INIT_LIST_HEAD(&cfs_rq->tasks);
9325 #ifdef CONFIG_FAIR_GROUP_SCHED
9328 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9331 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9333 struct rt_prio_array *array;
9336 array = &rt_rq->active;
9337 for (i = 0; i < MAX_RT_PRIO; i++) {
9338 INIT_LIST_HEAD(array->queue + i);
9339 __clear_bit(i, array->bitmap);
9341 /* delimiter for bitsearch: */
9342 __set_bit(MAX_RT_PRIO, array->bitmap);
9344 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9345 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9347 rt_rq->highest_prio.next = MAX_RT_PRIO;
9351 rt_rq->rt_nr_migratory = 0;
9352 rt_rq->overloaded = 0;
9353 plist_head_init_raw(&rt_rq->pushable_tasks, &rq->lock);
9357 rt_rq->rt_throttled = 0;
9358 rt_rq->rt_runtime = 0;
9359 raw_spin_lock_init(&rt_rq->rt_runtime_lock);
9361 #ifdef CONFIG_RT_GROUP_SCHED
9362 rt_rq->rt_nr_boosted = 0;
9367 #ifdef CONFIG_FAIR_GROUP_SCHED
9368 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9369 struct sched_entity *se, int cpu, int add,
9370 struct sched_entity *parent)
9372 struct rq *rq = cpu_rq(cpu);
9373 tg->cfs_rq[cpu] = cfs_rq;
9374 init_cfs_rq(cfs_rq, rq);
9377 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9380 /* se could be NULL for init_task_group */
9385 se->cfs_rq = &rq->cfs;
9387 se->cfs_rq = parent->my_q;
9390 se->load.weight = tg->shares;
9391 se->load.inv_weight = 0;
9392 se->parent = parent;
9396 #ifdef CONFIG_RT_GROUP_SCHED
9397 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9398 struct sched_rt_entity *rt_se, int cpu, int add,
9399 struct sched_rt_entity *parent)
9401 struct rq *rq = cpu_rq(cpu);
9403 tg->rt_rq[cpu] = rt_rq;
9404 init_rt_rq(rt_rq, rq);
9406 rt_rq->rt_se = rt_se;
9407 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9409 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9411 tg->rt_se[cpu] = rt_se;
9416 rt_se->rt_rq = &rq->rt;
9418 rt_se->rt_rq = parent->my_q;
9420 rt_se->my_q = rt_rq;
9421 rt_se->parent = parent;
9422 INIT_LIST_HEAD(&rt_se->run_list);
9426 void __init sched_init(void)
9429 unsigned long alloc_size = 0, ptr;
9431 #ifdef CONFIG_FAIR_GROUP_SCHED
9432 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9434 #ifdef CONFIG_RT_GROUP_SCHED
9435 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9437 #ifdef CONFIG_USER_SCHED
9440 #ifdef CONFIG_CPUMASK_OFFSTACK
9441 alloc_size += num_possible_cpus() * cpumask_size();
9444 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9446 #ifdef CONFIG_FAIR_GROUP_SCHED
9447 init_task_group.se = (struct sched_entity **)ptr;
9448 ptr += nr_cpu_ids * sizeof(void **);
9450 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9451 ptr += nr_cpu_ids * sizeof(void **);
9453 #ifdef CONFIG_USER_SCHED
9454 root_task_group.se = (struct sched_entity **)ptr;
9455 ptr += nr_cpu_ids * sizeof(void **);
9457 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9458 ptr += nr_cpu_ids * sizeof(void **);
9459 #endif /* CONFIG_USER_SCHED */
9460 #endif /* CONFIG_FAIR_GROUP_SCHED */
9461 #ifdef CONFIG_RT_GROUP_SCHED
9462 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9463 ptr += nr_cpu_ids * sizeof(void **);
9465 init_task_group.rt_rq = (struct rt_rq **)ptr;
9466 ptr += nr_cpu_ids * sizeof(void **);
9468 #ifdef CONFIG_USER_SCHED
9469 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9470 ptr += nr_cpu_ids * sizeof(void **);
9472 root_task_group.rt_rq = (struct rt_rq **)ptr;
9473 ptr += nr_cpu_ids * sizeof(void **);
9474 #endif /* CONFIG_USER_SCHED */
9475 #endif /* CONFIG_RT_GROUP_SCHED */
9476 #ifdef CONFIG_CPUMASK_OFFSTACK
9477 for_each_possible_cpu(i) {
9478 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9479 ptr += cpumask_size();
9481 #endif /* CONFIG_CPUMASK_OFFSTACK */
9485 init_defrootdomain();
9488 init_rt_bandwidth(&def_rt_bandwidth,
9489 global_rt_period(), global_rt_runtime());
9491 #ifdef CONFIG_RT_GROUP_SCHED
9492 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9493 global_rt_period(), global_rt_runtime());
9494 #ifdef CONFIG_USER_SCHED
9495 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9496 global_rt_period(), RUNTIME_INF);
9497 #endif /* CONFIG_USER_SCHED */
9498 #endif /* CONFIG_RT_GROUP_SCHED */
9500 #ifdef CONFIG_GROUP_SCHED
9501 list_add(&init_task_group.list, &task_groups);
9502 INIT_LIST_HEAD(&init_task_group.children);
9504 #ifdef CONFIG_USER_SCHED
9505 INIT_LIST_HEAD(&root_task_group.children);
9506 init_task_group.parent = &root_task_group;
9507 list_add(&init_task_group.siblings, &root_task_group.children);
9508 #endif /* CONFIG_USER_SCHED */
9509 #endif /* CONFIG_GROUP_SCHED */
9511 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9512 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9513 __alignof__(unsigned long));
9515 for_each_possible_cpu(i) {
9519 raw_spin_lock_init(&rq->lock);
9521 rq->calc_load_active = 0;
9522 rq->calc_load_update = jiffies + LOAD_FREQ;
9523 init_cfs_rq(&rq->cfs, rq);
9524 init_rt_rq(&rq->rt, rq);
9525 #ifdef CONFIG_FAIR_GROUP_SCHED
9526 init_task_group.shares = init_task_group_load;
9527 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9528 #ifdef CONFIG_CGROUP_SCHED
9530 * How much cpu bandwidth does init_task_group get?
9532 * In case of task-groups formed thr' the cgroup filesystem, it
9533 * gets 100% of the cpu resources in the system. This overall
9534 * system cpu resource is divided among the tasks of
9535 * init_task_group and its child task-groups in a fair manner,
9536 * based on each entity's (task or task-group's) weight
9537 * (se->load.weight).
9539 * In other words, if init_task_group has 10 tasks of weight
9540 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9541 * then A0's share of the cpu resource is:
9543 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9545 * We achieve this by letting init_task_group's tasks sit
9546 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9548 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9549 #elif defined CONFIG_USER_SCHED
9550 root_task_group.shares = NICE_0_LOAD;
9551 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9553 * In case of task-groups formed thr' the user id of tasks,
9554 * init_task_group represents tasks belonging to root user.
9555 * Hence it forms a sibling of all subsequent groups formed.
9556 * In this case, init_task_group gets only a fraction of overall
9557 * system cpu resource, based on the weight assigned to root
9558 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9559 * by letting tasks of init_task_group sit in a separate cfs_rq
9560 * (init_tg_cfs_rq) and having one entity represent this group of
9561 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9563 init_tg_cfs_entry(&init_task_group,
9564 &per_cpu(init_tg_cfs_rq, i),
9565 &per_cpu(init_sched_entity, i), i, 1,
9566 root_task_group.se[i]);
9569 #endif /* CONFIG_FAIR_GROUP_SCHED */
9571 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9572 #ifdef CONFIG_RT_GROUP_SCHED
9573 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9574 #ifdef CONFIG_CGROUP_SCHED
9575 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9576 #elif defined CONFIG_USER_SCHED
9577 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9578 init_tg_rt_entry(&init_task_group,
9579 &per_cpu(init_rt_rq_var, i),
9580 &per_cpu(init_sched_rt_entity, i), i, 1,
9581 root_task_group.rt_se[i]);
9585 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9586 rq->cpu_load[j] = 0;
9590 rq->post_schedule = 0;
9591 rq->active_balance = 0;
9592 rq->next_balance = jiffies;
9596 rq->migration_thread = NULL;
9598 rq->avg_idle = 2*sysctl_sched_migration_cost;
9599 INIT_LIST_HEAD(&rq->migration_queue);
9600 rq_attach_root(rq, &def_root_domain);
9603 atomic_set(&rq->nr_iowait, 0);
9606 set_load_weight(&init_task);
9608 #ifdef CONFIG_PREEMPT_NOTIFIERS
9609 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9613 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9616 #ifdef CONFIG_RT_MUTEXES
9617 plist_head_init_raw(&init_task.pi_waiters, &init_task.pi_lock);
9621 * The boot idle thread does lazy MMU switching as well:
9623 atomic_inc(&init_mm.mm_count);
9624 enter_lazy_tlb(&init_mm, current);
9627 * Make us the idle thread. Technically, schedule() should not be
9628 * called from this thread, however somewhere below it might be,
9629 * but because we are the idle thread, we just pick up running again
9630 * when this runqueue becomes "idle".
9632 init_idle(current, smp_processor_id());
9634 calc_load_update = jiffies + LOAD_FREQ;
9637 * During early bootup we pretend to be a normal task:
9639 current->sched_class = &fair_sched_class;
9641 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9642 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9645 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9646 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9648 /* May be allocated at isolcpus cmdline parse time */
9649 if (cpu_isolated_map == NULL)
9650 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9655 scheduler_running = 1;
9658 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9659 static inline int preempt_count_equals(int preempt_offset)
9661 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9663 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9666 void __might_sleep(char *file, int line, int preempt_offset)
9669 static unsigned long prev_jiffy; /* ratelimiting */
9671 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9672 system_state != SYSTEM_RUNNING || oops_in_progress)
9674 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9676 prev_jiffy = jiffies;
9678 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9680 pr_err("in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9681 in_atomic(), irqs_disabled(),
9682 current->pid, current->comm);
9684 debug_show_held_locks(current);
9685 if (irqs_disabled())
9686 print_irqtrace_events(current);
9690 EXPORT_SYMBOL(__might_sleep);
9693 #ifdef CONFIG_MAGIC_SYSRQ
9694 static void normalize_task(struct rq *rq, struct task_struct *p)
9698 update_rq_clock(rq);
9699 on_rq = p->se.on_rq;
9701 deactivate_task(rq, p, 0);
9702 __setscheduler(rq, p, SCHED_NORMAL, 0);
9704 activate_task(rq, p, 0);
9705 resched_task(rq->curr);
9709 void normalize_rt_tasks(void)
9711 struct task_struct *g, *p;
9712 unsigned long flags;
9715 read_lock_irqsave(&tasklist_lock, flags);
9716 do_each_thread(g, p) {
9718 * Only normalize user tasks:
9723 p->se.exec_start = 0;
9724 #ifdef CONFIG_SCHEDSTATS
9725 p->se.wait_start = 0;
9726 p->se.sleep_start = 0;
9727 p->se.block_start = 0;
9732 * Renice negative nice level userspace
9735 if (TASK_NICE(p) < 0 && p->mm)
9736 set_user_nice(p, 0);
9740 raw_spin_lock(&p->pi_lock);
9741 rq = __task_rq_lock(p);
9743 normalize_task(rq, p);
9745 __task_rq_unlock(rq);
9746 raw_spin_unlock(&p->pi_lock);
9747 } while_each_thread(g, p);
9749 read_unlock_irqrestore(&tasklist_lock, flags);
9752 #endif /* CONFIG_MAGIC_SYSRQ */
9756 * These functions are only useful for the IA64 MCA handling.
9758 * They can only be called when the whole system has been
9759 * stopped - every CPU needs to be quiescent, and no scheduling
9760 * activity can take place. Using them for anything else would
9761 * be a serious bug, and as a result, they aren't even visible
9762 * under any other configuration.
9766 * curr_task - return the current task for a given cpu.
9767 * @cpu: the processor in question.
9769 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9771 struct task_struct *curr_task(int cpu)
9773 return cpu_curr(cpu);
9777 * set_curr_task - set the current task for a given cpu.
9778 * @cpu: the processor in question.
9779 * @p: the task pointer to set.
9781 * Description: This function must only be used when non-maskable interrupts
9782 * are serviced on a separate stack. It allows the architecture to switch the
9783 * notion of the current task on a cpu in a non-blocking manner. This function
9784 * must be called with all CPU's synchronized, and interrupts disabled, the
9785 * and caller must save the original value of the current task (see
9786 * curr_task() above) and restore that value before reenabling interrupts and
9787 * re-starting the system.
9789 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9791 void set_curr_task(int cpu, struct task_struct *p)
9798 #ifdef CONFIG_FAIR_GROUP_SCHED
9799 static void free_fair_sched_group(struct task_group *tg)
9803 for_each_possible_cpu(i) {
9805 kfree(tg->cfs_rq[i]);
9815 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9817 struct cfs_rq *cfs_rq;
9818 struct sched_entity *se;
9822 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9825 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9829 tg->shares = NICE_0_LOAD;
9831 for_each_possible_cpu(i) {
9834 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9835 GFP_KERNEL, cpu_to_node(i));
9839 se = kzalloc_node(sizeof(struct sched_entity),
9840 GFP_KERNEL, cpu_to_node(i));
9844 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9855 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9857 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9858 &cpu_rq(cpu)->leaf_cfs_rq_list);
9861 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9863 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9865 #else /* !CONFG_FAIR_GROUP_SCHED */
9866 static inline void free_fair_sched_group(struct task_group *tg)
9871 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9876 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9880 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9883 #endif /* CONFIG_FAIR_GROUP_SCHED */
9885 #ifdef CONFIG_RT_GROUP_SCHED
9886 static void free_rt_sched_group(struct task_group *tg)
9890 destroy_rt_bandwidth(&tg->rt_bandwidth);
9892 for_each_possible_cpu(i) {
9894 kfree(tg->rt_rq[i]);
9896 kfree(tg->rt_se[i]);
9904 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9906 struct rt_rq *rt_rq;
9907 struct sched_rt_entity *rt_se;
9911 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9914 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9918 init_rt_bandwidth(&tg->rt_bandwidth,
9919 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9921 for_each_possible_cpu(i) {
9924 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9925 GFP_KERNEL, cpu_to_node(i));
9929 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9930 GFP_KERNEL, cpu_to_node(i));
9934 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9945 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9947 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9948 &cpu_rq(cpu)->leaf_rt_rq_list);
9951 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9953 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9955 #else /* !CONFIG_RT_GROUP_SCHED */
9956 static inline void free_rt_sched_group(struct task_group *tg)
9961 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9966 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9970 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9973 #endif /* CONFIG_RT_GROUP_SCHED */
9975 #ifdef CONFIG_GROUP_SCHED
9976 static void free_sched_group(struct task_group *tg)
9978 free_fair_sched_group(tg);
9979 free_rt_sched_group(tg);
9983 /* allocate runqueue etc for a new task group */
9984 struct task_group *sched_create_group(struct task_group *parent)
9986 struct task_group *tg;
9987 unsigned long flags;
9990 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9992 return ERR_PTR(-ENOMEM);
9994 if (!alloc_fair_sched_group(tg, parent))
9997 if (!alloc_rt_sched_group(tg, parent))
10000 spin_lock_irqsave(&task_group_lock, flags);
10001 for_each_possible_cpu(i) {
10002 register_fair_sched_group(tg, i);
10003 register_rt_sched_group(tg, i);
10005 list_add_rcu(&tg->list, &task_groups);
10007 WARN_ON(!parent); /* root should already exist */
10009 tg->parent = parent;
10010 INIT_LIST_HEAD(&tg->children);
10011 list_add_rcu(&tg->siblings, &parent->children);
10012 spin_unlock_irqrestore(&task_group_lock, flags);
10017 free_sched_group(tg);
10018 return ERR_PTR(-ENOMEM);
10021 /* rcu callback to free various structures associated with a task group */
10022 static void free_sched_group_rcu(struct rcu_head *rhp)
10024 /* now it should be safe to free those cfs_rqs */
10025 free_sched_group(container_of(rhp, struct task_group, rcu));
10028 /* Destroy runqueue etc associated with a task group */
10029 void sched_destroy_group(struct task_group *tg)
10031 unsigned long flags;
10034 spin_lock_irqsave(&task_group_lock, flags);
10035 for_each_possible_cpu(i) {
10036 unregister_fair_sched_group(tg, i);
10037 unregister_rt_sched_group(tg, i);
10039 list_del_rcu(&tg->list);
10040 list_del_rcu(&tg->siblings);
10041 spin_unlock_irqrestore(&task_group_lock, flags);
10043 /* wait for possible concurrent references to cfs_rqs complete */
10044 call_rcu(&tg->rcu, free_sched_group_rcu);
10047 /* change task's runqueue when it moves between groups.
10048 * The caller of this function should have put the task in its new group
10049 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10050 * reflect its new group.
10052 void sched_move_task(struct task_struct *tsk)
10054 int on_rq, running;
10055 unsigned long flags;
10058 rq = task_rq_lock(tsk, &flags);
10060 update_rq_clock(rq);
10062 running = task_current(rq, tsk);
10063 on_rq = tsk->se.on_rq;
10066 dequeue_task(rq, tsk, 0);
10067 if (unlikely(running))
10068 tsk->sched_class->put_prev_task(rq, tsk);
10070 set_task_rq(tsk, task_cpu(tsk));
10072 #ifdef CONFIG_FAIR_GROUP_SCHED
10073 if (tsk->sched_class->moved_group)
10074 tsk->sched_class->moved_group(tsk);
10077 if (unlikely(running))
10078 tsk->sched_class->set_curr_task(rq);
10080 enqueue_task(rq, tsk, 0);
10082 task_rq_unlock(rq, &flags);
10084 #endif /* CONFIG_GROUP_SCHED */
10086 #ifdef CONFIG_FAIR_GROUP_SCHED
10087 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10089 struct cfs_rq *cfs_rq = se->cfs_rq;
10094 dequeue_entity(cfs_rq, se, 0);
10096 se->load.weight = shares;
10097 se->load.inv_weight = 0;
10100 enqueue_entity(cfs_rq, se, 0);
10103 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10105 struct cfs_rq *cfs_rq = se->cfs_rq;
10106 struct rq *rq = cfs_rq->rq;
10107 unsigned long flags;
10109 raw_spin_lock_irqsave(&rq->lock, flags);
10110 __set_se_shares(se, shares);
10111 raw_spin_unlock_irqrestore(&rq->lock, flags);
10114 static DEFINE_MUTEX(shares_mutex);
10116 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10119 unsigned long flags;
10122 * We can't change the weight of the root cgroup.
10127 if (shares < MIN_SHARES)
10128 shares = MIN_SHARES;
10129 else if (shares > MAX_SHARES)
10130 shares = MAX_SHARES;
10132 mutex_lock(&shares_mutex);
10133 if (tg->shares == shares)
10136 spin_lock_irqsave(&task_group_lock, flags);
10137 for_each_possible_cpu(i)
10138 unregister_fair_sched_group(tg, i);
10139 list_del_rcu(&tg->siblings);
10140 spin_unlock_irqrestore(&task_group_lock, flags);
10142 /* wait for any ongoing reference to this group to finish */
10143 synchronize_sched();
10146 * Now we are free to modify the group's share on each cpu
10147 * w/o tripping rebalance_share or load_balance_fair.
10149 tg->shares = shares;
10150 for_each_possible_cpu(i) {
10152 * force a rebalance
10154 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10155 set_se_shares(tg->se[i], shares);
10159 * Enable load balance activity on this group, by inserting it back on
10160 * each cpu's rq->leaf_cfs_rq_list.
10162 spin_lock_irqsave(&task_group_lock, flags);
10163 for_each_possible_cpu(i)
10164 register_fair_sched_group(tg, i);
10165 list_add_rcu(&tg->siblings, &tg->parent->children);
10166 spin_unlock_irqrestore(&task_group_lock, flags);
10168 mutex_unlock(&shares_mutex);
10172 unsigned long sched_group_shares(struct task_group *tg)
10178 #ifdef CONFIG_RT_GROUP_SCHED
10180 * Ensure that the real time constraints are schedulable.
10182 static DEFINE_MUTEX(rt_constraints_mutex);
10184 static unsigned long to_ratio(u64 period, u64 runtime)
10186 if (runtime == RUNTIME_INF)
10189 return div64_u64(runtime << 20, period);
10192 /* Must be called with tasklist_lock held */
10193 static inline int tg_has_rt_tasks(struct task_group *tg)
10195 struct task_struct *g, *p;
10197 do_each_thread(g, p) {
10198 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10200 } while_each_thread(g, p);
10205 struct rt_schedulable_data {
10206 struct task_group *tg;
10211 static int tg_schedulable(struct task_group *tg, void *data)
10213 struct rt_schedulable_data *d = data;
10214 struct task_group *child;
10215 unsigned long total, sum = 0;
10216 u64 period, runtime;
10218 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10219 runtime = tg->rt_bandwidth.rt_runtime;
10222 period = d->rt_period;
10223 runtime = d->rt_runtime;
10226 #ifdef CONFIG_USER_SCHED
10227 if (tg == &root_task_group) {
10228 period = global_rt_period();
10229 runtime = global_rt_runtime();
10234 * Cannot have more runtime than the period.
10236 if (runtime > period && runtime != RUNTIME_INF)
10240 * Ensure we don't starve existing RT tasks.
10242 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10245 total = to_ratio(period, runtime);
10248 * Nobody can have more than the global setting allows.
10250 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10254 * The sum of our children's runtime should not exceed our own.
10256 list_for_each_entry_rcu(child, &tg->children, siblings) {
10257 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10258 runtime = child->rt_bandwidth.rt_runtime;
10260 if (child == d->tg) {
10261 period = d->rt_period;
10262 runtime = d->rt_runtime;
10265 sum += to_ratio(period, runtime);
10274 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10276 struct rt_schedulable_data data = {
10278 .rt_period = period,
10279 .rt_runtime = runtime,
10282 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10285 static int tg_set_bandwidth(struct task_group *tg,
10286 u64 rt_period, u64 rt_runtime)
10290 mutex_lock(&rt_constraints_mutex);
10291 read_lock(&tasklist_lock);
10292 err = __rt_schedulable(tg, rt_period, rt_runtime);
10296 raw_spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10297 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10298 tg->rt_bandwidth.rt_runtime = rt_runtime;
10300 for_each_possible_cpu(i) {
10301 struct rt_rq *rt_rq = tg->rt_rq[i];
10303 raw_spin_lock(&rt_rq->rt_runtime_lock);
10304 rt_rq->rt_runtime = rt_runtime;
10305 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10307 raw_spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10309 read_unlock(&tasklist_lock);
10310 mutex_unlock(&rt_constraints_mutex);
10315 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10317 u64 rt_runtime, rt_period;
10319 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10320 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10321 if (rt_runtime_us < 0)
10322 rt_runtime = RUNTIME_INF;
10324 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10327 long sched_group_rt_runtime(struct task_group *tg)
10331 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10334 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10335 do_div(rt_runtime_us, NSEC_PER_USEC);
10336 return rt_runtime_us;
10339 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10341 u64 rt_runtime, rt_period;
10343 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10344 rt_runtime = tg->rt_bandwidth.rt_runtime;
10346 if (rt_period == 0)
10349 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10352 long sched_group_rt_period(struct task_group *tg)
10356 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10357 do_div(rt_period_us, NSEC_PER_USEC);
10358 return rt_period_us;
10361 static int sched_rt_global_constraints(void)
10363 u64 runtime, period;
10366 if (sysctl_sched_rt_period <= 0)
10369 runtime = global_rt_runtime();
10370 period = global_rt_period();
10373 * Sanity check on the sysctl variables.
10375 if (runtime > period && runtime != RUNTIME_INF)
10378 mutex_lock(&rt_constraints_mutex);
10379 read_lock(&tasklist_lock);
10380 ret = __rt_schedulable(NULL, 0, 0);
10381 read_unlock(&tasklist_lock);
10382 mutex_unlock(&rt_constraints_mutex);
10387 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10389 /* Don't accept realtime tasks when there is no way for them to run */
10390 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10396 #else /* !CONFIG_RT_GROUP_SCHED */
10397 static int sched_rt_global_constraints(void)
10399 unsigned long flags;
10402 if (sysctl_sched_rt_period <= 0)
10406 * There's always some RT tasks in the root group
10407 * -- migration, kstopmachine etc..
10409 if (sysctl_sched_rt_runtime == 0)
10412 raw_spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10413 for_each_possible_cpu(i) {
10414 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10416 raw_spin_lock(&rt_rq->rt_runtime_lock);
10417 rt_rq->rt_runtime = global_rt_runtime();
10418 raw_spin_unlock(&rt_rq->rt_runtime_lock);
10420 raw_spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10424 #endif /* CONFIG_RT_GROUP_SCHED */
10426 int sched_rt_handler(struct ctl_table *table, int write,
10427 void __user *buffer, size_t *lenp,
10431 int old_period, old_runtime;
10432 static DEFINE_MUTEX(mutex);
10434 mutex_lock(&mutex);
10435 old_period = sysctl_sched_rt_period;
10436 old_runtime = sysctl_sched_rt_runtime;
10438 ret = proc_dointvec(table, write, buffer, lenp, ppos);
10440 if (!ret && write) {
10441 ret = sched_rt_global_constraints();
10443 sysctl_sched_rt_period = old_period;
10444 sysctl_sched_rt_runtime = old_runtime;
10446 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10447 def_rt_bandwidth.rt_period =
10448 ns_to_ktime(global_rt_period());
10451 mutex_unlock(&mutex);
10456 #ifdef CONFIG_CGROUP_SCHED
10458 /* return corresponding task_group object of a cgroup */
10459 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10461 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10462 struct task_group, css);
10465 static struct cgroup_subsys_state *
10466 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10468 struct task_group *tg, *parent;
10470 if (!cgrp->parent) {
10471 /* This is early initialization for the top cgroup */
10472 return &init_task_group.css;
10475 parent = cgroup_tg(cgrp->parent);
10476 tg = sched_create_group(parent);
10478 return ERR_PTR(-ENOMEM);
10484 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10486 struct task_group *tg = cgroup_tg(cgrp);
10488 sched_destroy_group(tg);
10492 cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
10494 #ifdef CONFIG_RT_GROUP_SCHED
10495 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10498 /* We don't support RT-tasks being in separate groups */
10499 if (tsk->sched_class != &fair_sched_class)
10506 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10507 struct task_struct *tsk, bool threadgroup)
10509 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10513 struct task_struct *c;
10515 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10516 retval = cpu_cgroup_can_attach_task(cgrp, c);
10528 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10529 struct cgroup *old_cont, struct task_struct *tsk,
10532 sched_move_task(tsk);
10534 struct task_struct *c;
10536 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10537 sched_move_task(c);
10543 #ifdef CONFIG_FAIR_GROUP_SCHED
10544 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10547 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10550 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10552 struct task_group *tg = cgroup_tg(cgrp);
10554 return (u64) tg->shares;
10556 #endif /* CONFIG_FAIR_GROUP_SCHED */
10558 #ifdef CONFIG_RT_GROUP_SCHED
10559 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10562 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10565 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10567 return sched_group_rt_runtime(cgroup_tg(cgrp));
10570 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10573 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10576 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10578 return sched_group_rt_period(cgroup_tg(cgrp));
10580 #endif /* CONFIG_RT_GROUP_SCHED */
10582 static struct cftype cpu_files[] = {
10583 #ifdef CONFIG_FAIR_GROUP_SCHED
10586 .read_u64 = cpu_shares_read_u64,
10587 .write_u64 = cpu_shares_write_u64,
10590 #ifdef CONFIG_RT_GROUP_SCHED
10592 .name = "rt_runtime_us",
10593 .read_s64 = cpu_rt_runtime_read,
10594 .write_s64 = cpu_rt_runtime_write,
10597 .name = "rt_period_us",
10598 .read_u64 = cpu_rt_period_read_uint,
10599 .write_u64 = cpu_rt_period_write_uint,
10604 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10606 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10609 struct cgroup_subsys cpu_cgroup_subsys = {
10611 .create = cpu_cgroup_create,
10612 .destroy = cpu_cgroup_destroy,
10613 .can_attach = cpu_cgroup_can_attach,
10614 .attach = cpu_cgroup_attach,
10615 .populate = cpu_cgroup_populate,
10616 .subsys_id = cpu_cgroup_subsys_id,
10620 #endif /* CONFIG_CGROUP_SCHED */
10622 #ifdef CONFIG_CGROUP_CPUACCT
10625 * CPU accounting code for task groups.
10627 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10628 * (balbir@in.ibm.com).
10631 /* track cpu usage of a group of tasks and its child groups */
10633 struct cgroup_subsys_state css;
10634 /* cpuusage holds pointer to a u64-type object on every cpu */
10636 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10637 struct cpuacct *parent;
10640 struct cgroup_subsys cpuacct_subsys;
10642 /* return cpu accounting group corresponding to this container */
10643 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10645 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10646 struct cpuacct, css);
10649 /* return cpu accounting group to which this task belongs */
10650 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10652 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10653 struct cpuacct, css);
10656 /* create a new cpu accounting group */
10657 static struct cgroup_subsys_state *cpuacct_create(
10658 struct cgroup_subsys *ss, struct cgroup *cgrp)
10660 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10666 ca->cpuusage = alloc_percpu(u64);
10670 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10671 if (percpu_counter_init(&ca->cpustat[i], 0))
10672 goto out_free_counters;
10675 ca->parent = cgroup_ca(cgrp->parent);
10681 percpu_counter_destroy(&ca->cpustat[i]);
10682 free_percpu(ca->cpuusage);
10686 return ERR_PTR(-ENOMEM);
10689 /* destroy an existing cpu accounting group */
10691 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10693 struct cpuacct *ca = cgroup_ca(cgrp);
10696 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10697 percpu_counter_destroy(&ca->cpustat[i]);
10698 free_percpu(ca->cpuusage);
10702 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10704 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10707 #ifndef CONFIG_64BIT
10709 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10711 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10713 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10721 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10723 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10725 #ifndef CONFIG_64BIT
10727 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10729 raw_spin_lock_irq(&cpu_rq(cpu)->lock);
10731 raw_spin_unlock_irq(&cpu_rq(cpu)->lock);
10737 /* return total cpu usage (in nanoseconds) of a group */
10738 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10740 struct cpuacct *ca = cgroup_ca(cgrp);
10741 u64 totalcpuusage = 0;
10744 for_each_present_cpu(i)
10745 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10747 return totalcpuusage;
10750 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10753 struct cpuacct *ca = cgroup_ca(cgrp);
10762 for_each_present_cpu(i)
10763 cpuacct_cpuusage_write(ca, i, 0);
10769 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10770 struct seq_file *m)
10772 struct cpuacct *ca = cgroup_ca(cgroup);
10776 for_each_present_cpu(i) {
10777 percpu = cpuacct_cpuusage_read(ca, i);
10778 seq_printf(m, "%llu ", (unsigned long long) percpu);
10780 seq_printf(m, "\n");
10784 static const char *cpuacct_stat_desc[] = {
10785 [CPUACCT_STAT_USER] = "user",
10786 [CPUACCT_STAT_SYSTEM] = "system",
10789 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10790 struct cgroup_map_cb *cb)
10792 struct cpuacct *ca = cgroup_ca(cgrp);
10795 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10796 s64 val = percpu_counter_read(&ca->cpustat[i]);
10797 val = cputime64_to_clock_t(val);
10798 cb->fill(cb, cpuacct_stat_desc[i], val);
10803 static struct cftype files[] = {
10806 .read_u64 = cpuusage_read,
10807 .write_u64 = cpuusage_write,
10810 .name = "usage_percpu",
10811 .read_seq_string = cpuacct_percpu_seq_read,
10815 .read_map = cpuacct_stats_show,
10819 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10821 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10825 * charge this task's execution time to its accounting group.
10827 * called with rq->lock held.
10829 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10831 struct cpuacct *ca;
10834 if (unlikely(!cpuacct_subsys.active))
10837 cpu = task_cpu(tsk);
10843 for (; ca; ca = ca->parent) {
10844 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10845 *cpuusage += cputime;
10852 * Charge the system/user time to the task's accounting group.
10854 static void cpuacct_update_stats(struct task_struct *tsk,
10855 enum cpuacct_stat_index idx, cputime_t val)
10857 struct cpuacct *ca;
10859 if (unlikely(!cpuacct_subsys.active))
10866 percpu_counter_add(&ca->cpustat[idx], val);
10872 struct cgroup_subsys cpuacct_subsys = {
10874 .create = cpuacct_create,
10875 .destroy = cpuacct_destroy,
10876 .populate = cpuacct_populate,
10877 .subsys_id = cpuacct_subsys_id,
10879 #endif /* CONFIG_CGROUP_CPUACCT */
10883 int rcu_expedited_torture_stats(char *page)
10887 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10889 void synchronize_sched_expedited(void)
10892 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10894 #else /* #ifndef CONFIG_SMP */
10896 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10897 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10899 #define RCU_EXPEDITED_STATE_POST -2
10900 #define RCU_EXPEDITED_STATE_IDLE -1
10902 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10904 int rcu_expedited_torture_stats(char *page)
10909 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10910 for_each_online_cpu(cpu) {
10911 cnt += sprintf(&page[cnt], " %d:%d",
10912 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10914 cnt += sprintf(&page[cnt], "\n");
10917 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10919 static long synchronize_sched_expedited_count;
10922 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10923 * approach to force grace period to end quickly. This consumes
10924 * significant time on all CPUs, and is thus not recommended for
10925 * any sort of common-case code.
10927 * Note that it is illegal to call this function while holding any
10928 * lock that is acquired by a CPU-hotplug notifier. Failing to
10929 * observe this restriction will result in deadlock.
10931 void synchronize_sched_expedited(void)
10934 unsigned long flags;
10935 bool need_full_sync = 0;
10937 struct migration_req *req;
10941 smp_mb(); /* ensure prior mod happens before capturing snap. */
10942 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10944 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10946 if (trycount++ < 10)
10947 udelay(trycount * num_online_cpus());
10949 synchronize_sched();
10952 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10953 smp_mb(); /* ensure test happens before caller kfree */
10958 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10959 for_each_online_cpu(cpu) {
10961 req = &per_cpu(rcu_migration_req, cpu);
10962 init_completion(&req->done);
10964 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10965 raw_spin_lock_irqsave(&rq->lock, flags);
10966 list_add(&req->list, &rq->migration_queue);
10967 raw_spin_unlock_irqrestore(&rq->lock, flags);
10968 wake_up_process(rq->migration_thread);
10970 for_each_online_cpu(cpu) {
10971 rcu_expedited_state = cpu;
10972 req = &per_cpu(rcu_migration_req, cpu);
10974 wait_for_completion(&req->done);
10975 raw_spin_lock_irqsave(&rq->lock, flags);
10976 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10977 need_full_sync = 1;
10978 req->dest_cpu = RCU_MIGRATION_IDLE;
10979 raw_spin_unlock_irqrestore(&rq->lock, flags);
10981 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10982 synchronize_sched_expedited_count++;
10983 mutex_unlock(&rcu_sched_expedited_mutex);
10985 if (need_full_sync)
10986 synchronize_sched();
10988 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10990 #endif /* #else #ifndef CONFIG_SMP */